fs/btrfs/volumes.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/bio.h>
  20 #include <linux/slab.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/blkdev.h>
  23 #include <linux/random.h>
  24 #include <linux/iocontext.h>
  25 #include <linux/capability.h>
  26 #include <linux/ratelimit.h>
  27 #include <linux/kthread.h>
  28 #include <linux/raid/pq.h>
  29 #include <asm/div64.h>
  30 #include "compat.h"
  31 #include "ctree.h"
  32 #include "extent_map.h"
  33 #include "disk-io.h"
  34 #include "transaction.h"
  35 #include "print-tree.h"
  36 #include "volumes.h"
  37 #include "raid56.h"
  38 #include "async-thread.h"
  39 #include "check-integrity.h"
  40 #include "rcu-string.h"
  41 #include "math.h"
  42 #include "dev-replace.h"
  43
  44 static int init_first_rw_device(struct btrfs_trans_handle *trans,
  45                                 struct btrfs_root *root,
  46                                 struct btrfs_device *device);
  47 static int btrfs_relocate_sys_chunks(struct btrfs_root *root);
  48 static void __btrfs_reset_dev_stats(struct btrfs_device *dev);
  49 static void btrfs_dev_stat_print_on_load(struct btrfs_device *device);
  50
  51 static DEFINE_MUTEX(uuid_mutex);
  52 static LIST_HEAD(fs_uuids);
  53
  54 static void lock_chunks(struct btrfs_root *root)
  55 {
  56         mutex_lock(&root->fs_info->chunk_mutex);
  57 }
  58
  59 static void unlock_chunks(struct btrfs_root *root)
  60 {
  61         mutex_unlock(&root->fs_info->chunk_mutex);
  62 }
  63
  64 static void free_fs_devices(struct btrfs_fs_devices *fs_devices)
  65 {
  66         struct btrfs_device *device;
  67         WARN_ON(fs_devices->opened);
  68         while (!list_empty(&fs_devices->devices)) {
  69                 device = list_entry(fs_devices->devices.next,
  70                                     struct btrfs_device, dev_list);
  71                 list_del(&device->dev_list);
  72                 rcu_string_free(device->name);
  73                 kfree(device);
  74         }
  75         kfree(fs_devices);
  76 }
  77
  78 static void btrfs_kobject_uevent(struct block_device *bdev,
  79                                  enum kobject_action action)
  80 {
  81         int ret;
  82
  83         ret = kobject_uevent(&disk_to_dev(bdev->bd_disk)->kobj, action);
  84         if (ret)
  85                 pr_warn("Sending event '%d' to kobject: '%s' (%p): failed\n",
  86                         action,
  87                         kobject_name(&disk_to_dev(bdev->bd_disk)->kobj),
  88                         &disk_to_dev(bdev->bd_disk)->kobj);
  89 }
  90
  91 void btrfs_cleanup_fs_uuids(void)
  92 {
  93         struct btrfs_fs_devices *fs_devices;
  94
  95         while (!list_empty(&fs_uuids)) {
  96                 fs_devices = list_entry(fs_uuids.next,
  97                                         struct btrfs_fs_devices, list);
  98                 list_del(&fs_devices->list);
  99                 free_fs_devices(fs_devices);
 100         }
 101 }
 102
 103 static noinline struct btrfs_device *__find_device(struct list_head *head,
 104                                                    u64 devid, u8 *uuid)
 105 {
 106         struct btrfs_device *dev;
 107
 108         list_for_each_entry(dev, head, dev_list) {
 109                 if (dev->devid == devid &&
 110                     (!uuid || !memcmp(dev->uuid, uuid, BTRFS_UUID_SIZE))) {
 111                         return dev;
 112                 }
 113         }
 114         return NULL;
 115 }
 116
 117 static noinline struct btrfs_fs_devices *find_fsid(u8 *fsid)
 118 {
 119         struct btrfs_fs_devices *fs_devices;
 120
 121         list_for_each_entry(fs_devices, &fs_uuids, list) {
 122                 if (memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE) == 0)
 123                         return fs_devices;
 124         }
 125         return NULL;
 126 }
 127
 128 static int
 129 btrfs_get_bdev_and_sb(const char *device_path, fmode_t flags, void *holder,
 130                       int flush, struct block_device **bdev,
 131                       struct buffer_head **bh)
 132 {
 133         int ret;
 134
 135         *bdev = blkdev_get_by_path(device_path, flags, holder);
 136
 137         if (IS_ERR(*bdev)) {
 138                 ret = PTR_ERR(*bdev);
 139                 printk(KERN_INFO "btrfs: open %s failed\n", device_path);
 140                 goto error;
 141         }
 142
 143         if (flush)
 144                 filemap_write_and_wait((*bdev)->bd_inode->i_mapping);
 145         ret = set_blocksize(*bdev, 4096);
 146         if (ret) {
 147                 blkdev_put(*bdev, flags);
 148                 goto error;
 149         }
 150         invalidate_bdev(*bdev);
 151         *bh = btrfs_read_dev_super(*bdev);
 152         if (!*bh) {
 153                 ret = -EINVAL;
 154                 blkdev_put(*bdev, flags);
 155                 goto error;
 156         }
 157
 158         return 0;
 159
 160 error:
 161         *bdev = NULL;
 162         *bh = NULL;
 163         return ret;
 164 }
 165
 166 static void requeue_list(struct btrfs_pending_bios *pending_bios,
 167                         struct bio *head, struct bio *tail)
 168 {
 169
 170         struct bio *old_head;
 171
 172         old_head = pending_bios->head;
 173         pending_bios->head = head;
 174         if (pending_bios->tail)
 175                 tail->bi_next = old_head;
 176         else
 177                 pending_bios->tail = tail;
 178 }
 179
 180 /*
 181  * we try to collect pending bios for a device so we don't get a large
 182  * number of procs sending bios down to the same device.  This greatly
 183  * improves the schedulers ability to collect and merge the bios.
 184  *
 185  * But, it also turns into a long list of bios to process and that is sure
 186  * to eventually make the worker thread block.  The solution here is to
 187  * make some progress and then put this work struct back at the end of
 188  * the list if the block device is congested.  This way, multiple devices
 189  * can make progress from a single worker thread.
 190  */
 191 static noinline void run_scheduled_bios(struct btrfs_device *device)
 192 {
 193         struct bio *pending;
 194         struct backing_dev_info *bdi;
 195         struct btrfs_fs_info *fs_info;
 196         struct btrfs_pending_bios *pending_bios;
 197         struct bio *tail;
 198         struct bio *cur;
 199         int again = 0;
 200         unsigned long num_run;
 201         unsigned long batch_run = 0;
 202         unsigned long limit;
 203         unsigned long last_waited = 0;
 204         int force_reg = 0;
 205         int sync_pending = 0;
 206         struct blk_plug plug;
 207
 208         /*
 209          * this function runs all the bios we've collected for
 210          * a particular device.  We don't want to wander off to
 211          * another device without first sending all of these down.
 212          * So, setup a plug here and finish it off before we return
 213          */
 214         blk_start_plug(&plug);
 215
 216         bdi = blk_get_backing_dev_info(device->bdev);
 217         fs_info = device->dev_root->fs_info;
 218         limit = btrfs_async_submit_limit(fs_info);
 219         limit = limit * 2 / 3;
 220
 221 loop:
 222         spin_lock(&device->io_lock);
 223
 224 loop_lock:
 225         num_run = 0;
 226
 227         /* take all the bios off the list at once and process them
 228          * later on (without the lock held).  But, remember the
 229          * tail and other pointers so the bios can be properly reinserted
 230          * into the list if we hit congestion
 231          */
 232         if (!force_reg && device->pending_sync_bios.head) {
 233                 pending_bios = &device->pending_sync_bios;
 234                 force_reg = 1;
 235         } else {
 236                 pending_bios = &device->pending_bios;
 237                 force_reg = 0;
 238         }
 239
 240         pending = pending_bios->head;
 241         tail = pending_bios->tail;
 242         WARN_ON(pending && !tail);
 243
 244         /*
 245          * if pending was null this time around, no bios need processing
 246          * at all and we can stop.  Otherwise it'll loop back up again
 247          * and do an additional check so no bios are missed.
 248          *
 249          * device->running_pending is used to synchronize with the
 250          * schedule_bio code.
 251          */
 252         if (device->pending_sync_bios.head == NULL &&
 253             device->pending_bios.head == NULL) {
 254                 again = 0;
 255                 device->running_pending = 0;
 256         } else {
 257                 again = 1;
 258                 device->running_pending = 1;
 259         }
 260
 261         pending_bios->head = NULL;
 262         pending_bios->tail = NULL;
 263
 264         spin_unlock(&device->io_lock);
 265
 266         while (pending) {
 267
 268                 rmb();
 269                 /* we want to work on both lists, but do more bios on the
 270                  * sync list than the regular list
 271                  */
 272                 if ((num_run > 32 &&
 273                     pending_bios != &device->pending_sync_bios &&
 274                     device->pending_sync_bios.head) ||
 275                    (num_run > 64 && pending_bios == &device->pending_sync_bios &&
 276                     device->pending_bios.head)) {
 277                         spin_lock(&device->io_lock);
 278                         requeue_list(pending_bios, pending, tail);
 279                         goto loop_lock;
 280                 }
 281
 282                 cur = pending;
 283                 pending = pending->bi_next;
 284                 cur->bi_next = NULL;
 285
 286                 if (atomic_dec_return(&fs_info->nr_async_bios) < limit &&
 287                     waitqueue_active(&fs_info->async_submit_wait))
 288                         wake_up(&fs_info->async_submit_wait);
 289
 290                 BUG_ON(atomic_read(&cur->bi_cnt) == 0);
 291
 292                 /*
 293                  * if we're doing the sync list, record that our
 294                  * plug has some sync requests on it
 295                  *
 296                  * If we're doing the regular list and there are
 297                  * sync requests sitting around, unplug before
 298                  * we add more
 299                  */
 300                 if (pending_bios == &device->pending_sync_bios) {
 301                         sync_pending = 1;
 302                 } else if (sync_pending) {
 303                         blk_finish_plug(&plug);
 304                         blk_start_plug(&plug);
 305                         sync_pending = 0;
 306                 }
 307
 308                 btrfsic_submit_bio(cur->bi_rw, cur);
 309                 num_run++;
 310                 batch_run++;
 311                 if (need_resched())
 312                         cond_resched();
 313
 314                 /*
 315                  * we made progress, there is more work to do and the bdi
 316                  * is now congested.  Back off and let other work structs
 317                  * run instead
 318                  */
 319                 if (pending && bdi_write_congested(bdi) && batch_run > 8 &&
 320                     fs_info->fs_devices->open_devices > 1) {
 321                         struct io_context *ioc;
 322
 323                         ioc = current->io_context;
 324
 325                         /*
 326                          * the main goal here is that we don't want to
 327                          * block if we're going to be able to submit
 328                          * more requests without blocking.
 329                          *
 330                          * This code does two great things, it pokes into
 331                          * the elevator code from a filesystem _and_
 332                          * it makes assumptions about how batching works.
 333                          */
 334                         if (ioc && ioc->nr_batch_requests > 0 &&
 335                             time_before(jiffies, ioc->last_waited + HZ/50UL) &&
 336                             (last_waited == 0 ||
 337                              ioc->last_waited == last_waited)) {
 338                                 /*
 339                                  * we want to go through our batch of
 340                                  * requests and stop.  So, we copy out
 341                                  * the ioc->last_waited time and test
 342                                  * against it before looping
 343                                  */
 344                                 last_waited = ioc->last_waited;
 345                                 if (need_resched())
 346                                         cond_resched();
 347                                 continue;
 348                         }
 349                         spin_lock(&device->io_lock);
 350                         requeue_list(pending_bios, pending, tail);
 351                         device->running_pending = 1;
 352
 353                         spin_unlock(&device->io_lock);
 354                         btrfs_requeue_work(&device->work);
 355                         goto done;
 356                 }
 357                 /* unplug every 64 requests just for good measure */
 358                 if (batch_run % 64 == 0) {
 359                         blk_finish_plug(&plug);
 360                         blk_start_plug(&plug);
 361                         sync_pending = 0;
 362                 }
 363         }
 364
 365         cond_resched();
 366         if (again)
 367                 goto loop;
 368
 369         spin_lock(&device->io_lock);
 370         if (device->pending_bios.head || device->pending_sync_bios.head)
 371                 goto loop_lock;
 372         spin_unlock(&device->io_lock);
 373
 374 done:
 375         blk_finish_plug(&plug);
 376 }
 377
 378 static void pending_bios_fn(struct btrfs_work *work)
 379 {
 380         struct btrfs_device *device;
 381
 382         device = container_of(work, struct btrfs_device, work);
 383         run_scheduled_bios(device);
 384 }
 385
 386 static noinline int device_list_add(const char *path,
 387                            struct btrfs_super_block *disk_super,
 388                            u64 devid, struct btrfs_fs_devices **fs_devices_ret)
 389 {
 390         struct btrfs_device *device;
 391         struct btrfs_fs_devices *fs_devices;
 392         struct rcu_string *name;
 393         u64 found_transid = btrfs_super_generation(disk_super);
 394
 395         fs_devices = find_fsid(disk_super->fsid);
 396         if (!fs_devices) {
 397                 fs_devices = kzalloc(sizeof(*fs_devices), GFP_NOFS);
 398                 if (!fs_devices)
 399                         return -ENOMEM;
 400                 INIT_LIST_HEAD(&fs_devices->devices);
 401                 INIT_LIST_HEAD(&fs_devices->alloc_list);
 402                 list_add(&fs_devices->list, &fs_uuids);
 403                 memcpy(fs_devices->fsid, disk_super->fsid, BTRFS_FSID_SIZE);
 404                 fs_devices->latest_devid = devid;
 405                 fs_devices->latest_trans = found_transid;
 406                 mutex_init(&fs_devices->device_list_mutex);
 407                 device = NULL;
 408         } else {
 409                 device = __find_device(&fs_devices->devices, devid,
 410                                        disk_super->dev_item.uuid);
 411         }
 412         if (!device) {
 413                 if (fs_devices->opened)
 414                         return -EBUSY;
 415
 416                 device = kzalloc(sizeof(*device), GFP_NOFS);
 417                 if (!device) {
 418                         /* we can safely leave the fs_devices entry around */
 419                         return -ENOMEM;
 420                 }
 421                 device->devid = devid;
 422                 device->dev_stats_valid = 0;
 423                 device->work.func = pending_bios_fn;
 424                 memcpy(device->uuid, disk_super->dev_item.uuid,
 425                        BTRFS_UUID_SIZE);
 426                 spin_lock_init(&device->io_lock);
 427
 428                 name = rcu_string_strdup(path, GFP_NOFS);
 429                 if (!name) {
 430                         kfree(device);
 431                         return -ENOMEM;
 432                 }
 433                 rcu_assign_pointer(device->name, name);
 434                 INIT_LIST_HEAD(&device->dev_alloc_list);
 435
 436                 /* init readahead state */
 437                 spin_lock_init(&device->reada_lock);
 438                 device->reada_curr_zone = NULL;
 439                 atomic_set(&device->reada_in_flight, 0);
 440                 device->reada_next = 0;
 441                 INIT_RADIX_TREE(&device->reada_zones, GFP_NOFS & ~__GFP_WAIT);
 442                 INIT_RADIX_TREE(&device->reada_extents, GFP_NOFS & ~__GFP_WAIT);
 443
 444                 mutex_lock(&fs_devices->device_list_mutex);
 445                 list_add_rcu(&device->dev_list, &fs_devices->devices);
 446                 mutex_unlock(&fs_devices->device_list_mutex);
 447
 448                 device->fs_devices = fs_devices;
 449                 fs_devices->num_devices++;
 450         } else if (!device->name || strcmp(device->name->str, path)) {
 451                 name = rcu_string_strdup(path, GFP_NOFS);
 452                 if (!name)
 453                         return -ENOMEM;
 454                 rcu_string_free(device->name);
 455                 rcu_assign_pointer(device->name, name);
 456                 if (device->missing) {
 457                         fs_devices->missing_devices--;
 458                         device->missing = 0;
 459                 }
 460         }
 461
 462         if (found_transid > fs_devices->latest_trans) {
 463                 fs_devices->latest_devid = devid;
 464                 fs_devices->latest_trans = found_transid;
 465         }
 466         *fs_devices_ret = fs_devices;
 467         return 0;
 468 }
 469
 470 static struct btrfs_fs_devices *clone_fs_devices(struct btrfs_fs_devices *orig)
 471 {
 472         struct btrfs_fs_devices *fs_devices;
 473         struct btrfs_device *device;
 474         struct btrfs_device *orig_dev;
 475
 476         fs_devices = kzalloc(sizeof(*fs_devices), GFP_NOFS);
 477         if (!fs_devices)
 478                 return ERR_PTR(-ENOMEM);
 479
 480         INIT_LIST_HEAD(&fs_devices->devices);
 481         INIT_LIST_HEAD(&fs_devices->alloc_list);
 482         INIT_LIST_HEAD(&fs_devices->list);
 483         mutex_init(&fs_devices->device_list_mutex);
 484         fs_devices->latest_devid = orig->latest_devid;
 485         fs_devices->latest_trans = orig->latest_trans;
 486         fs_devices->total_devices = orig->total_devices;
 487         memcpy(fs_devices->fsid, orig->fsid, sizeof(fs_devices->fsid));
 488
 489         /* We have held the volume lock, it is safe to get the devices. */
 490         list_for_each_entry(orig_dev, &orig->devices, dev_list) {
 491                 struct rcu_string *name;
 492
 493                 device = kzalloc(sizeof(*device), GFP_NOFS);
 494                 if (!device)
 495                         goto error;
 496
 497                 /*
 498                  * This is ok to do without rcu read locked because we hold the
 499                  * uuid mutex so nothing we touch in here is going to disappear.
 500                  */
 501                 name = rcu_string_strdup(orig_dev->name->str, GFP_NOFS);
 502                 if (!name) {
 503                         kfree(device);
 504                         goto error;
 505                 }
 506                 rcu_assign_pointer(device->name, name);
 507
 508                 device->devid = orig_dev->devid;
 509                 device->work.func = pending_bios_fn;
 510                 memcpy(device->uuid, orig_dev->uuid, sizeof(device->uuid));
 511                 spin_lock_init(&device->io_lock);
 512                 INIT_LIST_HEAD(&device->dev_list);
 513                 INIT_LIST_HEAD(&device->dev_alloc_list);
 514
 515                 list_add(&device->dev_list, &fs_devices->devices);
 516                 device->fs_devices = fs_devices;
 517                 fs_devices->num_devices++;
 518         }
 519         return fs_devices;
 520 error:
 521         free_fs_devices(fs_devices);
 522         return ERR_PTR(-ENOMEM);
 523 }
 524
 525 void btrfs_close_extra_devices(struct btrfs_fs_info *fs_info,
 526                                struct btrfs_fs_devices *fs_devices, int step)
 527 {
 528         struct btrfs_device *device, *next;
 529
 530         struct block_device *latest_bdev = NULL;
 531         u64 latest_devid = 0;
 532         u64 latest_transid = 0;
 533
 534         mutex_lock(&uuid_mutex);
 535 again:
 536         /* This is the initialized path, it is safe to release the devices. */
 537         list_for_each_entry_safe(device, next, &fs_devices->devices, dev_list) {
 538                 if (device->in_fs_metadata) {
 539                         if (!device->is_tgtdev_for_dev_replace &&
 540                             (!latest_transid ||
 541                              device->generation > latest_transid)) {
 542                                 latest_devid = device->devid;
 543                                 latest_transid = device->generation;
 544                                 latest_bdev = device->bdev;
 545                         }
 546                         continue;
 547                 }
 548
 549                 if (device->devid == BTRFS_DEV_REPLACE_DEVID) {
 550                         /*
 551                          * In the first step, keep the device which has
 552                          * the correct fsid and the devid that is used
 553                          * for the dev_replace procedure.
 554                          * In the second step, the dev_replace state is
 555                          * read from the device tree and it is known
 556                          * whether the procedure is really active or
 557                          * not, which means whether this device is
 558                          * used or whether it should be removed.
 559                          */
 560                         if (step == 0 || device->is_tgtdev_for_dev_replace) {
 561                                 continue;
 562                         }
 563                 }
 564                 if (device->bdev) {
 565                         blkdev_put(device->bdev, device->mode);
 566                         device->bdev = NULL;
 567                         fs_devices->open_devices--;
 568                 }
 569                 if (device->writeable) {
 570                         list_del_init(&device->dev_alloc_list);
 571                         device->writeable = 0;
 572                         if (!device->is_tgtdev_for_dev_replace)
 573                                 fs_devices->rw_devices--;
 574                 }
 575                 list_del_init(&device->dev_list);
 576                 fs_devices->num_devices--;
 577                 rcu_string_free(device->name);
 578                 kfree(device);
 579         }
 580
 581         if (fs_devices->seed) {
 582                 fs_devices = fs_devices->seed;
 583                 goto again;
 584         }
 585
 586         fs_devices->latest_bdev = latest_bdev;
 587         fs_devices->latest_devid = latest_devid;
 588         fs_devices->latest_trans = latest_transid;
 589
 590         mutex_unlock(&uuid_mutex);
 591 }
 592
 593 static void __free_device(struct work_struct *work)
 594 {
 595         struct btrfs_device *device;
 596
 597         device = container_of(work, struct btrfs_device, rcu_work);
 598
 599         if (device->bdev)
 600                 blkdev_put(device->bdev, device->mode);
 601
 602         rcu_string_free(device->name);
 603         kfree(device);
 604 }
 605
 606 static void free_device(struct rcu_head *head)
 607 {
 608         struct btrfs_device *device;
 609
 610         device = container_of(head, struct btrfs_device, rcu);
 611
 612         INIT_WORK(&device->rcu_work, __free_device);
 613         schedule_work(&device->rcu_work);
 614 }
 615
 616 static int __btrfs_close_devices(struct btrfs_fs_devices *fs_devices)
 617 {
 618         struct btrfs_device *device;
 619
 620         if (--fs_devices->opened > 0)
 621                 return 0;
 622
 623         mutex_lock(&fs_devices->device_list_mutex);
 624         list_for_each_entry(device, &fs_devices->devices, dev_list) {
 625                 struct btrfs_device *new_device;
 626                 struct rcu_string *name;
 627
 628                 if (device->bdev)
 629                         fs_devices->open_devices--;
 630
 631                 if (device->writeable && !device->is_tgtdev_for_dev_replace) {
 632                         list_del_init(&device->dev_alloc_list);
 633                         fs_devices->rw_devices--;
 634                 }
 635
 636                 if (device->can_discard)
 637                         fs_devices->num_can_discard--;
 638
 639                 new_device = kmalloc(sizeof(*new_device), GFP_NOFS);
 640                 BUG_ON(!new_device); /* -ENOMEM */
 641                 memcpy(new_device, device, sizeof(*new_device));
 642
 643                 /* Safe because we are under uuid_mutex */
 644                 if (device->name) {
 645                         name = rcu_string_strdup(device->name->str, GFP_NOFS);
 646                         BUG_ON(device->name && !name); /* -ENOMEM */
 647                         rcu_assign_pointer(new_device->name, name);
 648                 }
 649                 new_device->bdev = NULL;
 650                 new_device->writeable = 0;
 651                 new_device->in_fs_metadata = 0;
 652                 new_device->can_discard = 0;
 653                 spin_lock_init(&new_device->io_lock);
 654                 list_replace_rcu(&device->dev_list, &new_device->dev_list);
 655
 656                 call_rcu(&device->rcu, free_device);
 657         }
 658         mutex_unlock(&fs_devices->device_list_mutex);
 659
 660         WARN_ON(fs_devices->open_devices);
 661         WARN_ON(fs_devices->rw_devices);
 662         fs_devices->opened = 0;
 663         fs_devices->seeding = 0;
 664
 665         return 0;
 666 }
 667
 668 int btrfs_close_devices(struct btrfs_fs_devices *fs_devices)
 669 {
 670         struct btrfs_fs_devices *seed_devices = NULL;
 671         int ret;
 672
 673         mutex_lock(&uuid_mutex);
 674         ret = __btrfs_close_devices(fs_devices);
 675         if (!fs_devices->opened) {
 676                 seed_devices = fs_devices->seed;
 677                 fs_devices->seed = NULL;
 678         }
 679         mutex_unlock(&uuid_mutex);
 680
 681         while (seed_devices) {
 682                 fs_devices = seed_devices;
 683                 seed_devices = fs_devices->seed;
 684                 __btrfs_close_devices(fs_devices);
 685                 free_fs_devices(fs_devices);
 686         }
 687         /*
 688          * Wait for rcu kworkers under __btrfs_close_devices
 689          * to finish all blkdev_puts so device is really
 690          * free when umount is done.
 691          */
 692         rcu_barrier();
 693         return ret;
 694 }
 695
 696 static int __btrfs_open_devices(struct btrfs_fs_devices *fs_devices,
 697                                 fmode_t flags, void *holder)
 698 {
 699         struct request_queue *q;
 700         struct block_device *bdev;
 701         struct list_head *head = &fs_devices->devices;
 702         struct btrfs_device *device;
 703         struct block_device *latest_bdev = NULL;
 704         struct buffer_head *bh;
 705         struct btrfs_super_block *disk_super;
 706         u64 latest_devid = 0;
 707         u64 latest_transid = 0;
 708         u64 devid;
 709         int seeding = 1;
 710         int ret = 0;
 711
 712         flags |= FMODE_EXCL;
 713
 714         list_for_each_entry(device, head, dev_list) {
 715                 if (device->bdev)
 716                         continue;
 717                 if (!device->name)
 718                         continue;
 719
 720                 /* Just open everything we can; ignore failures here */
 721                 if (btrfs_get_bdev_and_sb(device->name->str, flags, holder, 1,
 722                                             &bdev, &bh))
 723                         continue;
 724
 725                 disk_super = (struct btrfs_super_block *)bh->b_data;
 726                 devid = btrfs_stack_device_id(&disk_super->dev_item);
 727                 if (devid != device->devid)
 728                         goto error_brelse;
 729
 730                 if (memcmp(device->uuid, disk_super->dev_item.uuid,
 731                            BTRFS_UUID_SIZE))
 732                         goto error_brelse;
 733
 734                 device->generation = btrfs_super_generation(disk_super);
 735                 if (!latest_transid || device->generation > latest_transid) {
 736                         latest_devid = devid;
 737                         latest_transid = device->generation;
 738                         latest_bdev = bdev;
 739                 }
 740
 741                 if (btrfs_super_flags(disk_super) & BTRFS_SUPER_FLAG_SEEDING) {
 742                         device->writeable = 0;
 743                 } else {
 744                         device->writeable = !bdev_read_only(bdev);
 745                         seeding = 0;
 746                 }
 747
 748                 q = bdev_get_queue(bdev);
 749                 if (blk_queue_discard(q)) {
 750                         device->can_discard = 1;
 751                         fs_devices->num_can_discard++;
 752                 }
 753
 754                 device->bdev = bdev;
 755                 device->in_fs_metadata = 0;
 756                 device->mode = flags;
 757
 758                 if (!blk_queue_nonrot(bdev_get_queue(bdev)))
 759                         fs_devices->rotating = 1;
 760
 761                 fs_devices->open_devices++;
 762                 if (device->writeable && !device->is_tgtdev_for_dev_replace) {
 763                         fs_devices->rw_devices++;
 764                         list_add(&device->dev_alloc_list,
 765                                  &fs_devices->alloc_list);
 766                 }
 767                 brelse(bh);
 768                 continue;
 769
 770 error_brelse:
 771                 brelse(bh);
 772                 blkdev_put(bdev, flags);
 773                 continue;
 774         }
 775         if (fs_devices->open_devices == 0) {
 776                 ret = -EINVAL;
 777                 goto out;
 778         }
 779         fs_devices->seeding = seeding;
 780         fs_devices->opened = 1;
 781         fs_devices->latest_bdev = latest_bdev;
 782         fs_devices->latest_devid = latest_devid;
 783         fs_devices->latest_trans = latest_transid;
 784         fs_devices->total_rw_bytes = 0;
 785 out:
 786         return ret;
 787 }
 788
 789 int btrfs_open_devices(struct btrfs_fs_devices *fs_devices,
 790                        fmode_t flags, void *holder)
 791 {
 792         int ret;
 793
 794         mutex_lock(&uuid_mutex);
 795         if (fs_devices->opened) {
 796                 fs_devices->opened++;
 797                 ret = 0;
 798         } else {
 799                 ret = __btrfs_open_devices(fs_devices, flags, holder);
 800         }
 801         mutex_unlock(&uuid_mutex);
 802         return ret;
 803 }
 804
 805 /*
 806  * Look for a btrfs signature on a device. This may be called out of the mount path
 807  * and we are not allowed to call set_blocksize during the scan. The superblock
 808  * is read via pagecache
 809  */
 810 int btrfs_scan_one_device(const char *path, fmode_t flags, void *holder,
 811                           struct btrfs_fs_devices **fs_devices_ret)
 812 {
 813         struct btrfs_super_block *disk_super;
 814         struct block_device *bdev;
 815         struct page *page;
 816         void *p;
 817         int ret = -EINVAL;
 818         u64 devid;
 819         u64 transid;
 820         u64 total_devices;
 821         u64 bytenr;
 822         pgoff_t index;
 823
 824         /*
 825          * we would like to check all the supers, but that would make
 826          * a btrfs mount succeed after a mkfs from a different FS.
 827          * So, we need to add a special mount option to scan for
 828          * later supers, using BTRFS_SUPER_MIRROR_MAX instead
 829          */
 830         bytenr = btrfs_sb_offset(0);
 831         flags |= FMODE_EXCL;
 832         mutex_lock(&uuid_mutex);
 833
 834         bdev = blkdev_get_by_path(path, flags, holder);
 835
 836         if (IS_ERR(bdev)) {
 837                 ret = PTR_ERR(bdev);
 838                 goto error;
 839         }
 840
 841         /* make sure our super fits in the device */
 842         if (bytenr + PAGE_CACHE_SIZE >= i_size_read(bdev->bd_inode))
 843                 goto error_bdev_put;
 844
 845         /* make sure our super fits in the page */
 846         if (sizeof(*disk_super) > PAGE_CACHE_SIZE)
 847                 goto error_bdev_put;
 848
 849         /* make sure our super doesn't straddle pages on disk */
 850         index = bytenr >> PAGE_CACHE_SHIFT;
 851         if ((bytenr + sizeof(*disk_super) - 1) >> PAGE_CACHE_SHIFT != index)
 852                 goto error_bdev_put;
 853
 854         /* pull in the page with our super */
 855         page = read_cache_page_gfp(bdev->bd_inode->i_mapping,
 856                                    index, GFP_NOFS);
 857
 858         if (IS_ERR_OR_NULL(page))
 859                 goto error_bdev_put;
 860
 861         p = kmap(page);
 862
 863         /* align our pointer to the offset of the super block */
 864         disk_super = p + (bytenr & ~PAGE_CACHE_MASK);
 865
 866         if (btrfs_super_bytenr(disk_super) != bytenr ||
 867             disk_super->magic != cpu_to_le64(BTRFS_MAGIC))
 868                 goto error_unmap;
 869
 870         devid = btrfs_stack_device_id(&disk_super->dev_item);
 871         transid = btrfs_super_generation(disk_super);
 872         total_devices = btrfs_super_num_devices(disk_super);
 873
 874         if (disk_super->label[0]) {
 875                 if (disk_super->label[BTRFS_LABEL_SIZE - 1])
 876                         disk_super->label[BTRFS_LABEL_SIZE - 1] = '\0';
 877                 printk(KERN_INFO "device label %s ", disk_super->label);
 878         } else {
 879                 printk(KERN_INFO "device fsid %pU ", disk_super->fsid);
 880         }
 881
 882         printk(KERN_CONT "devid %llu transid %llu %s\n",
 883                (unsigned long long)devid, (unsigned long long)transid, path);
 884
 885         ret = device_list_add(path, disk_super, devid, fs_devices_ret);
 886         if (!ret && fs_devices_ret)
 887                 (*fs_devices_ret)->total_devices = total_devices;
 888
 889 error_unmap:
 890         kunmap(page);
 891         page_cache_release(page);
 892
 893 error_bdev_put:
 894         blkdev_put(bdev, flags);
 895 error:
 896         mutex_unlock(&uuid_mutex);
 897         return ret;
 898 }
 899
 900 /* helper to account the used device space in the range */
 901 int btrfs_account_dev_extents_size(struct btrfs_device *device, u64 start,
 902                                    u64 end, u64 *length)
 903 {
 904         struct btrfs_key key;
 905         struct btrfs_root *root = device->dev_root;
 906         struct btrfs_dev_extent *dev_extent;
 907         struct btrfs_path *path;
 908         u64 extent_end;
 909         int ret;
 910         int slot;
 911         struct extent_buffer *l;
 912
 913         *length = 0;
 914
 915         if (start >= device->total_bytes || device->is_tgtdev_for_dev_replace)
 916                 return 0;
 917
 918         path = btrfs_alloc_path();
 919         if (!path)
 920                 return -ENOMEM;
 921         path->reada = 2;
 922
 923         key.objectid = device->devid;
 924         key.offset = start;
 925         key.type = BTRFS_DEV_EXTENT_KEY;
 926
 927         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 928         if (ret < 0)
 929                 goto out;
 930         if (ret > 0) {
 931                 ret = btrfs_previous_item(root, path, key.objectid, key.type);
 932                 if (ret < 0)
 933                         goto out;
 934         }
 935
 936         while (1) {
 937                 l = path->nodes[0];
 938                 slot = path->slots[0];
 939                 if (slot >= btrfs_header_nritems(l)) {
 940                         ret = btrfs_next_leaf(root, path);
 941                         if (ret == 0)
 942                                 continue;
 943                         if (ret < 0)
 944                                 goto out;
 945
 946                         break;
 947                 }
 948                 btrfs_item_key_to_cpu(l, &key, slot);
 949
 950                 if (key.objectid < device->devid)
 951                         goto next;
 952
 953                 if (key.objectid > device->devid)
 954                         break;
 955
 956                 if (btrfs_key_type(&key) != BTRFS_DEV_EXTENT_KEY)
 957                         goto next;
 958
 959                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
 960                 extent_end = key.offset + btrfs_dev_extent_length(l,
 961                                                                   dev_extent);
 962                 if (key.offset <= start && extent_end > end) {
 963                         *length = end - start + 1;
 964                         break;
 965                 } else if (key.offset <= start && extent_end > start)
 966                         *length += extent_end - start;
 967                 else if (key.offset > start && extent_end <= end)
 968                         *length += extent_end - key.offset;
 969                 else if (key.offset > start && key.offset <= end) {
 970                         *length += end - key.offset + 1;
 971                         break;
 972                 } else if (key.offset > end)
 973                         break;
 974
 975 next:
 976                 path->slots[0]++;
 977         }
 978         ret = 0;
 979 out:
 980         btrfs_free_path(path);
 981         return ret;
 982 }
 983
 984 /*
 985  * find_free_dev_extent - find free space in the specified device
 986  * @device:     the device which we search the free space in
 987  * @num_bytes:  the size of the free space that we need
 988  * @start:      store the start of the free space.
 989  * @len:        the size of the free space. that we find, or the size of the max
 990  *              free space if we don't find suitable free space
 991  *
 992  * this uses a pretty simple search, the expectation is that it is
 993  * called very infrequently and that a given device has a small number
 994  * of extents
 995  *
 996  * @start is used to store the start of the free space if we find. But if we
 997  * don't find suitable free space, it will be used to store the start position
 998  * of the max free space.
 999  *
1000  * @len is used to store the size of the free space that we find.
1001  * But if we don't find suitable free space, it is used to store the size of
1002  * the max free space.
1003  */
1004 int find_free_dev_extent(struct btrfs_device *device, u64 num_bytes,
1005                          u64 *start, u64 *len)
1006 {
1007         struct btrfs_key key;
1008         struct btrfs_root *root = device->dev_root;
1009         struct btrfs_dev_extent *dev_extent;
1010         struct btrfs_path *path;
1011         u64 hole_size;
1012         u64 max_hole_start;
1013         u64 max_hole_size;
1014         u64 extent_end;
1015         u64 search_start;
1016         u64 search_end = device->total_bytes;
1017         int ret;
1018         int slot;
1019         struct extent_buffer *l;
1020
1021         /* FIXME use last free of some kind */
1022
1023         /* we don't want to overwrite the superblock on the drive,
1024          * so we make sure to start at an offset of at least 1MB
1025          */
1026         search_start = max(root->fs_info->alloc_start, 1024ull * 1024);
1027
1028         max_hole_start = search_start;
1029         max_hole_size = 0;
1030         hole_size = 0;
1031
1032         if (search_start >= search_end || device->is_tgtdev_for_dev_replace) {
1033                 ret = -ENOSPC;
1034                 goto error;
1035         }
1036
1037         path = btrfs_alloc_path();
1038         if (!path) {
1039                 ret = -ENOMEM;
1040                 goto error;
1041         }
1042         path->reada = 2;
1043
1044         key.objectid = device->devid;
1045         key.offset = search_start;
1046         key.type = BTRFS_DEV_EXTENT_KEY;
1047
1048         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1049         if (ret < 0)
1050                 goto out;
1051         if (ret > 0) {
1052                 ret = btrfs_previous_item(root, path, key.objectid, key.type);
1053                 if (ret < 0)
1054                         goto out;
1055         }
1056
1057         while (1) {
1058                 l = path->nodes[0];
1059                 slot = path->slots[0];
1060                 if (slot >= btrfs_header_nritems(l)) {
1061                         ret = btrfs_next_leaf(root, path);
1062                         if (ret == 0)
1063                                 continue;
1064                         if (ret < 0)
1065                                 goto out;
1066
1067                         break;
1068                 }
1069                 btrfs_item_key_to_cpu(l, &key, slot);
1070
1071                 if (key.objectid < device->devid)
1072                         goto next;
1073
1074                 if (key.objectid > device->devid)
1075                         break;
1076
1077                 if (btrfs_key_type(&key) != BTRFS_DEV_EXTENT_KEY)
1078                         goto next;
1079
1080                 if (key.offset > search_start) {
1081                         hole_size = key.offset - search_start;
1082
1083                         if (hole_size > max_hole_size) {
1084                                 max_hole_start = search_start;
1085                                 max_hole_size = hole_size;
1086                         }
1087
1088                         /*
1089                          * If this free space is greater than which we need,
1090                          * it must be the max free space that we have found
1091                          * until now, so max_hole_start must point to the start
1092                          * of this free space and the length of this free space
1093                          * is stored in max_hole_size. Thus, we return
1094                          * max_hole_start and max_hole_size and go back to the
1095                          * caller.
1096                          */
1097                         if (hole_size >= num_bytes) {
1098                                 ret = 0;
1099                                 goto out;
1100                         }
1101                 }
1102
1103                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
1104                 extent_end = key.offset + btrfs_dev_extent_length(l,
1105                                                                   dev_extent);
1106                 if (extent_end > search_start)
1107                         search_start = extent_end;
1108 next:
1109                 path->slots[0]++;
1110                 cond_resched();
1111         }
1112
1113         /*
1114          * At this point, search_start should be the end of
1115          * allocated dev extents, and when shrinking the device,
1116          * search_end may be smaller than search_start.
1117          */
1118         if (search_end > search_start)
1119                 hole_size = search_end - search_start;
1120
1121         if (hole_size > max_hole_size) {
1122                 max_hole_start = search_start;
1123                 max_hole_size = hole_size;
1124         }
1125
1126         /* See above. */
1127         if (hole_size < num_bytes)
1128                 ret = -ENOSPC;
1129         else
1130                 ret = 0;
1131
1132 out:
1133         btrfs_free_path(path);
1134 error:
1135         *start = max_hole_start;
1136         if (len)
1137                 *len = max_hole_size;
1138         return ret;
1139 }
1140
1141 static int btrfs_free_dev_extent(struct btrfs_trans_handle *trans,
1142                           struct btrfs_device *device,
1143                           u64 start)
1144 {
1145         int ret;
1146         struct btrfs_path *path;
1147         struct btrfs_root *root = device->dev_root;
1148         struct btrfs_key key;
1149         struct btrfs_key found_key;
1150         struct extent_buffer *leaf = NULL;
1151         struct btrfs_dev_extent *extent = NULL;
1152
1153         path = btrfs_alloc_path();
1154         if (!path)
1155                 return -ENOMEM;
1156
1157         key.objectid = device->devid;
1158         key.offset = start;
1159         key.type = BTRFS_DEV_EXTENT_KEY;
1160 again:
1161         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1162         if (ret > 0) {
1163                 ret = btrfs_previous_item(root, path, key.objectid,
1164                                           BTRFS_DEV_EXTENT_KEY);
1165                 if (ret)
1166                         goto out;
1167                 leaf = path->nodes[0];
1168                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1169                 extent = btrfs_item_ptr(leaf, path->slots[0],
1170                                         struct btrfs_dev_extent);
1171                 BUG_ON(found_key.offset > start || found_key.offset +
1172                        btrfs_dev_extent_length(leaf, extent) < start);
1173                 key = found_key;
1174                 btrfs_release_path(path);
1175                 goto again;
1176         } else if (ret == 0) {
1177                 leaf = path->nodes[0];
1178                 extent = btrfs_item_ptr(leaf, path->slots[0],
1179                                         struct btrfs_dev_extent);
1180         } else {
1181                 btrfs_error(root->fs_info, ret, "Slot search failed");
1182                 goto out;
1183         }
1184
1185         if (device->bytes_used > 0) {
1186                 u64 len = btrfs_dev_extent_length(leaf, extent);
1187                 device->bytes_used -= len;
1188                 spin_lock(&root->fs_info->free_chunk_lock);
1189                 root->fs_info->free_chunk_space += len;
1190                 spin_unlock(&root->fs_info->free_chunk_lock);
1191         }
1192         ret = btrfs_del_item(trans, root, path);
1193         if (ret) {
1194                 btrfs_error(root->fs_info, ret,
1195                             "Failed to remove dev extent item");
1196         }
1197 out:
1198         btrfs_free_path(path);
1199         return ret;
1200 }
1201
1202 int btrfs_alloc_dev_extent(struct btrfs_trans_handle *trans,
1203                            struct btrfs_device *device,
1204                            u64 chunk_tree, u64 chunk_objectid,
1205                            u64 chunk_offset, u64 start, u64 num_bytes)
1206 {
1207         int ret;
1208         struct btrfs_path *path;
1209         struct btrfs_root *root = device->dev_root;
1210         struct btrfs_dev_extent *extent;
1211         struct extent_buffer *leaf;
1212         struct btrfs_key key;
1213
1214         WARN_ON(!device->in_fs_metadata);
1215         WARN_ON(device->is_tgtdev_for_dev_replace);
1216         path = btrfs_alloc_path();
1217         if (!path)
1218                 return -ENOMEM;
1219
1220         key.objectid = device->devid;
1221         key.offset = start;
1222         key.type = BTRFS_DEV_EXTENT_KEY;
1223         ret = btrfs_insert_empty_item(trans, root, path, &key,
1224                                       sizeof(*extent));
1225         if (ret)
1226                 goto out;
1227
1228         leaf = path->nodes[0];
1229         extent = btrfs_item_ptr(leaf, path->slots[0],
1230                                 struct btrfs_dev_extent);
1231         btrfs_set_dev_extent_chunk_tree(leaf, extent, chunk_tree);
1232         btrfs_set_dev_extent_chunk_objectid(leaf, extent, chunk_objectid);
1233         btrfs_set_dev_extent_chunk_offset(leaf, extent, chunk_offset);
1234
1235         write_extent_buffer(leaf, root->fs_info->chunk_tree_uuid,
1236                     (unsigned long)btrfs_dev_extent_chunk_tree_uuid(extent),
1237                     BTRFS_UUID_SIZE);
1238
1239         btrfs_set_dev_extent_length(leaf, extent, num_bytes);
1240         btrfs_mark_buffer_dirty(leaf);
1241 out:
1242         btrfs_free_path(path);
1243         return ret;
1244 }
1245
1246 static noinline int find_next_chunk(struct btrfs_root *root,
1247                                     u64 objectid, u64 *offset)
1248 {
1249         struct btrfs_path *path;
1250         int ret;
1251         struct btrfs_key key;
1252         struct btrfs_chunk *chunk;
1253         struct btrfs_key found_key;
1254
1255         path = btrfs_alloc_path();
1256         if (!path)
1257                 return -ENOMEM;
1258
1259         key.objectid = objectid;
1260         key.offset = (u64)-1;
1261         key.type = BTRFS_CHUNK_ITEM_KEY;
1262
1263         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1264         if (ret < 0)
1265                 goto error;
1266
1267         BUG_ON(ret == 0); /* Corruption */
1268
1269         ret = btrfs_previous_item(root, path, 0, BTRFS_CHUNK_ITEM_KEY);
1270         if (ret) {
1271                 *offset = 0;
1272         } else {
1273                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
1274                                       path->slots[0]);
1275                 if (found_key.objectid != objectid)
1276                         *offset = 0;
1277                 else {
1278                         chunk = btrfs_item_ptr(path->nodes[0], path->slots[0],
1279                                                struct btrfs_chunk);
1280                         *offset = found_key.offset +
1281                                 btrfs_chunk_length(path->nodes[0], chunk);
1282                 }
1283         }
1284         ret = 0;
1285 error:
1286         btrfs_free_path(path);
1287         return ret;
1288 }
1289
1290 static noinline int find_next_devid(struct btrfs_root *root, u64 *objectid)
1291 {
1292         int ret;
1293         struct btrfs_key key;
1294         struct btrfs_key found_key;
1295         struct btrfs_path *path;
1296
1297         root = root->fs_info->chunk_root;
1298
1299         path = btrfs_alloc_path();
1300         if (!path)
1301                 return -ENOMEM;
1302
1303         key.objectid = BTRFS_DEV_ITEMS_OBJECTID;
1304         key.type = BTRFS_DEV_ITEM_KEY;
1305         key.offset = (u64)-1;
1306
1307         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1308         if (ret < 0)
1309                 goto error;
1310
1311         BUG_ON(ret == 0); /* Corruption */
1312
1313         ret = btrfs_previous_item(root, path, BTRFS_DEV_ITEMS_OBJECTID,
1314                                   BTRFS_DEV_ITEM_KEY);
1315         if (ret) {
1316                 *objectid = 1;
1317         } else {
1318                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
1319                                       path->slots[0]);
1320                 *objectid = found_key.offset + 1;
1321         }
1322         ret = 0;
1323 error:
1324         btrfs_free_path(path);
1325         return ret;
1326 }
1327
1328 /*
1329  * the device information is stored in the chunk root
1330  * the btrfs_device struct should be fully filled in
1331  */
1332 int btrfs_add_device(struct btrfs_trans_handle *trans,
1333                      struct btrfs_root *root,
1334                      struct btrfs_device *device)
1335 {
1336         int ret;
1337         struct btrfs_path *path;
1338         struct btrfs_dev_item *dev_item;
1339         struct extent_buffer *leaf;
1340         struct btrfs_key key;
1341         unsigned long ptr;
1342
1343         root = root->fs_info->chunk_root;
1344
1345         path = btrfs_alloc_path();
1346         if (!path)
1347                 return -ENOMEM;
1348
1349         key.objectid = BTRFS_DEV_ITEMS_OBJECTID;
1350         key.type = BTRFS_DEV_ITEM_KEY;
1351         key.offset = device->devid;
1352
1353         ret = btrfs_insert_empty_item(trans, root, path, &key,
1354                                       sizeof(*dev_item));
1355         if (ret)
1356                 goto out;
1357
1358         leaf = path->nodes[0];
1359         dev_item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_item);
1360
1361         btrfs_set_device_id(leaf, dev_item, device->devid);
1362         btrfs_set_device_generation(leaf, dev_item, 0);
1363         btrfs_set_device_type(leaf, dev_item, device->type);
1364         btrfs_set_device_io_align(leaf, dev_item, device->io_align);
1365         btrfs_set_device_io_width(leaf, dev_item, device->io_width);
1366         btrfs_set_device_sector_size(leaf, dev_item, device->sector_size);
1367         btrfs_set_device_total_bytes(leaf, dev_item, device->total_bytes);
1368         btrfs_set_device_bytes_used(leaf, dev_item, device->bytes_used);
1369         btrfs_set_device_group(leaf, dev_item, 0);
1370         btrfs_set_device_seek_speed(leaf, dev_item, 0);
1371         btrfs_set_device_bandwidth(leaf, dev_item, 0);
1372         btrfs_set_device_start_offset(leaf, dev_item, 0);
1373
1374         ptr = (unsigned long)btrfs_device_uuid(dev_item);
1375         write_extent_buffer(leaf, device->uuid, ptr, BTRFS_UUID_SIZE);
1376         ptr = (unsigned long)btrfs_device_fsid(dev_item);
1377         write_extent_buffer(leaf, root->fs_info->fsid, ptr, BTRFS_UUID_SIZE);
1378         btrfs_mark_buffer_dirty(leaf);
1379
1380         ret = 0;
1381 out:
1382         btrfs_free_path(path);
1383         return ret;
1384 }
1385
1386 static int btrfs_rm_dev_item(struct btrfs_root *root,
1387                              struct btrfs_device *device)
1388 {
1389         int ret;
1390         struct btrfs_path *path;
1391         struct btrfs_key key;
1392         struct btrfs_trans_handle *trans;
1393
1394         root = root->fs_info->chunk_root;
1395
1396         path = btrfs_alloc_path();
1397         if (!path)
1398                 return -ENOMEM;
1399
1400         trans = btrfs_start_transaction(root, 0);
1401         if (IS_ERR(trans)) {
1402                 btrfs_free_path(path);
1403                 return PTR_ERR(trans);
1404         }
1405         key.objectid = BTRFS_DEV_ITEMS_OBJECTID;
1406         key.type = BTRFS_DEV_ITEM_KEY;
1407         key.offset = device->devid;
1408         lock_chunks(root);
1409
1410         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1411         if (ret < 0)
1412                 goto out;
1413
1414         if (ret > 0) {
1415                 ret = -ENOENT;
1416                 goto out;
1417         }
1418
1419         ret = btrfs_del_item(trans, root, path);
1420         if (ret)
1421                 goto out;
1422 out:
1423         btrfs_free_path(path);
1424         unlock_chunks(root);
1425         btrfs_commit_transaction(trans, root);
1426         return ret;
1427 }
1428
1429 int btrfs_rm_device(struct btrfs_root *root, char *device_path)
1430 {
1431         struct btrfs_device *device;
1432         struct btrfs_device *next_device;
1433         struct block_device *bdev;
1434         struct buffer_head *bh = NULL;
1435         struct btrfs_super_block *disk_super;
1436         struct btrfs_fs_devices *cur_devices;
1437         u64 all_avail;
1438         u64 devid;
1439         u64 num_devices;
1440         u8 *dev_uuid;
1441         unsigned seq;
1442         int ret = 0;
1443         bool clear_super = false;
1444
1445         mutex_lock(&uuid_mutex);
1446
1447         do {
1448                 seq = read_seqbegin(&root->fs_info->profiles_lock);
1449
1450                 all_avail = root->fs_info->avail_data_alloc_bits |
1451                             root->fs_info->avail_system_alloc_bits |
1452                             root->fs_info->avail_metadata_alloc_bits;
1453         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
1454
1455         num_devices = root->fs_info->fs_devices->num_devices;
1456         btrfs_dev_replace_lock(&root->fs_info->dev_replace);
1457         if (btrfs_dev_replace_is_ongoing(&root->fs_info->dev_replace)) {
1458                 WARN_ON(num_devices < 1);
1459                 num_devices--;
1460         }
1461         btrfs_dev_replace_unlock(&root->fs_info->dev_replace);
1462
1463         if ((all_avail & BTRFS_BLOCK_GROUP_RAID10) && num_devices <= 4) {
1464                 printk(KERN_ERR "btrfs: unable to go below four devices "
1465                        "on raid10\n");
1466                 ret = -EINVAL;
1467                 goto out;
1468         }
1469
1470         if ((all_avail & BTRFS_BLOCK_GROUP_RAID1) && num_devices <= 2) {
1471                 printk(KERN_ERR "btrfs: unable to go below two "
1472                        "devices on raid1\n");
1473                 ret = -EINVAL;
1474                 goto out;
1475         }
1476
1477         if ((all_avail & BTRFS_BLOCK_GROUP_RAID5) &&
1478             root->fs_info->fs_devices->rw_devices <= 2) {
1479                 printk(KERN_ERR "btrfs: unable to go below two "
1480                        "devices on raid5\n");
1481                 ret = -EINVAL;
1482                 goto out;
1483         }
1484         if ((all_avail & BTRFS_BLOCK_GROUP_RAID6) &&
1485             root->fs_info->fs_devices->rw_devices <= 3) {
1486                 printk(KERN_ERR "btrfs: unable to go below three "
1487                        "devices on raid6\n");
1488                 ret = -EINVAL;
1489                 goto out;
1490         }
1491
1492         if (strcmp(device_path, "missing") == 0) {
1493                 struct list_head *devices;
1494                 struct btrfs_device *tmp;
1495
1496                 device = NULL;
1497                 devices = &root->fs_info->fs_devices->devices;
1498                 /*
1499                  * It is safe to read the devices since the volume_mutex
1500                  * is held.
1501                  */
1502                 list_for_each_entry(tmp, devices, dev_list) {
1503                         if (tmp->in_fs_metadata &&
1504                             !tmp->is_tgtdev_for_dev_replace &&
1505                             !tmp->bdev) {
1506                                 device = tmp;
1507                                 break;
1508                         }
1509                 }
1510                 bdev = NULL;
1511                 bh = NULL;
1512                 disk_super = NULL;
1513                 if (!device) {
1514                         printk(KERN_ERR "btrfs: no missing devices found to "
1515                                "remove\n");
1516                         goto out;
1517                 }
1518         } else {
1519                 ret = btrfs_get_bdev_and_sb(device_path,
1520                                             FMODE_WRITE | FMODE_EXCL,
1521                                             root->fs_info->bdev_holder, 0,
1522                                             &bdev, &bh);
1523                 if (ret)
1524                         goto out;
1525                 disk_super = (struct btrfs_super_block *)bh->b_data;
1526                 devid = btrfs_stack_device_id(&disk_super->dev_item);
1527                 dev_uuid = disk_super->dev_item.uuid;
1528                 device = btrfs_find_device(root->fs_info, devid, dev_uuid,
1529                                            disk_super->fsid);
1530                 if (!device) {
1531                         ret = -ENOENT;
1532                         goto error_brelse;
1533                 }
1534         }
1535
1536         if (device->is_tgtdev_for_dev_replace) {
1537                 pr_err("btrfs: unable to remove the dev_replace target dev\n");
1538                 ret = -EINVAL;
1539                 goto error_brelse;
1540         }
1541
1542         if (device->writeable && root->fs_info->fs_devices->rw_devices == 1) {
1543                 printk(KERN_ERR "btrfs: unable to remove the only writeable "
1544                        "device\n");
1545                 ret = -EINVAL;
1546                 goto error_brelse;
1547         }
1548
1549         if (device->writeable) {
1550                 lock_chunks(root);
1551                 list_del_init(&device->dev_alloc_list);
1552                 unlock_chunks(root);
1553                 root->fs_info->fs_devices->rw_devices--;
1554                 clear_super = true;
1555         }
1556
1557         ret = btrfs_shrink_device(device, 0);
1558         if (ret)
1559                 goto error_undo;
1560
1561         /*
1562          * TODO: the superblock still includes this device in its num_devices
1563          * counter although write_all_supers() is not locked out. This
1564          * could give a filesystem state which requires a degraded mount.
1565          */
1566         ret = btrfs_rm_dev_item(root->fs_info->chunk_root, device);
1567         if (ret)
1568                 goto error_undo;
1569
1570         spin_lock(&root->fs_info->free_chunk_lock);
1571         root->fs_info->free_chunk_space = device->total_bytes -
1572                 device->bytes_used;
1573         spin_unlock(&root->fs_info->free_chunk_lock);
1574
1575         device->in_fs_metadata = 0;
1576         btrfs_scrub_cancel_dev(root->fs_info, device);
1577
1578         /*
1579          * the device list mutex makes sure that we don't change
1580          * the device list while someone else is writing out all
1581          * the device supers.
1582          */
1583
1584         cur_devices = device->fs_devices;
1585         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
1586         list_del_rcu(&device->dev_list);
1587
1588         device->fs_devices->num_devices--;
1589         device->fs_devices->total_devices--;
1590
1591         if (device->missing)
1592                 root->fs_info->fs_devices->missing_devices--;
1593
1594         next_device = list_entry(root->fs_info->fs_devices->devices.next,
1595                                  struct btrfs_device, dev_list);
1596         if (device->bdev == root->fs_info->sb->s_bdev)
1597                 root->fs_info->sb->s_bdev = next_device->bdev;
1598         if (device->bdev == root->fs_info->fs_devices->latest_bdev)
1599                 root->fs_info->fs_devices->latest_bdev = next_device->bdev;
1600
1601         if (device->bdev)
1602                 device->fs_devices->open_devices--;
1603
1604         call_rcu(&device->rcu, free_device);
1605         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1606
1607         num_devices = btrfs_super_num_devices(root->fs_info->super_copy) - 1;
1608         btrfs_set_super_num_devices(root->fs_info->super_copy, num_devices);
1609
1610         if (cur_devices->open_devices == 0) {
1611                 struct btrfs_fs_devices *fs_devices;
1612                 fs_devices = root->fs_info->fs_devices;
1613                 while (fs_devices) {
1614                         if (fs_devices->seed == cur_devices)
1615                                 break;
1616                         fs_devices = fs_devices->seed;
1617                 }
1618                 fs_devices->seed = cur_devices->seed;
1619                 cur_devices->seed = NULL;
1620                 lock_chunks(root);
1621                 __btrfs_close_devices(cur_devices);
1622                 unlock_chunks(root);
1623                 free_fs_devices(cur_devices);
1624         }
1625
1626         root->fs_info->num_tolerated_disk_barrier_failures =
1627                 btrfs_calc_num_tolerated_disk_barrier_failures(root->fs_info);
1628
1629         /*
1630          * at this point, the device is zero sized.  We want to
1631          * remove it from the devices list and zero out the old super
1632          */
1633         if (clear_super && disk_super) {
1634                 /* make sure this device isn't detected as part of
1635                  * the FS anymore
1636                  */
1637                 memset(&disk_super->magic, 0, sizeof(disk_super->magic));
1638                 set_buffer_dirty(bh);
1639                 sync_dirty_buffer(bh);
1640         }
1641
1642         ret = 0;
1643
1644         /* Notify udev that device has changed */
1645         if (bdev)
1646                 btrfs_kobject_uevent(bdev, KOBJ_CHANGE);
1647
1648 error_brelse:
1649         brelse(bh);
1650         if (bdev)
1651                 blkdev_put(bdev, FMODE_READ | FMODE_EXCL);
1652 out:
1653         mutex_unlock(&uuid_mutex);
1654         return ret;
1655 error_undo:
1656         if (device->writeable) {
1657                 lock_chunks(root);
1658                 list_add(&device->dev_alloc_list,
1659                          &root->fs_info->fs_devices->alloc_list);
1660                 unlock_chunks(root);
1661                 root->fs_info->fs_devices->rw_devices++;
1662         }
1663         goto error_brelse;
1664 }
1665
1666 void btrfs_rm_dev_replace_srcdev(struct btrfs_fs_info *fs_info,
1667                                  struct btrfs_device *srcdev)
1668 {
1669         WARN_ON(!mutex_is_locked(&fs_info->fs_devices->device_list_mutex));
1670         list_del_rcu(&srcdev->dev_list);
1671         list_del_rcu(&srcdev->dev_alloc_list);
1672         fs_info->fs_devices->num_devices--;
1673         if (srcdev->missing) {
1674                 fs_info->fs_devices->missing_devices--;
1675                 fs_info->fs_devices->rw_devices++;
1676         }
1677         if (srcdev->can_discard)
1678                 fs_info->fs_devices->num_can_discard--;
1679         if (srcdev->bdev)
1680                 fs_info->fs_devices->open_devices--;
1681
1682         call_rcu(&srcdev->rcu, free_device);
1683 }
1684
1685 void btrfs_destroy_dev_replace_tgtdev(struct btrfs_fs_info *fs_info,
1686                                       struct btrfs_device *tgtdev)
1687 {
1688         struct btrfs_device *next_device;
1689
1690         WARN_ON(!tgtdev);
1691         mutex_lock(&fs_info->fs_devices->device_list_mutex);
1692         if (tgtdev->bdev) {
1693                 btrfs_scratch_superblock(tgtdev);
1694                 fs_info->fs_devices->open_devices--;
1695         }
1696         fs_info->fs_devices->num_devices--;
1697         if (tgtdev->can_discard)
1698                 fs_info->fs_devices->num_can_discard++;
1699
1700         next_device = list_entry(fs_info->fs_devices->devices.next,
1701                                  struct btrfs_device, dev_list);
1702         if (tgtdev->bdev == fs_info->sb->s_bdev)
1703                 fs_info->sb->s_bdev = next_device->bdev;
1704         if (tgtdev->bdev == fs_info->fs_devices->latest_bdev)
1705                 fs_info->fs_devices->latest_bdev = next_device->bdev;
1706         list_del_rcu(&tgtdev->dev_list);
1707
1708         call_rcu(&tgtdev->rcu, free_device);
1709
1710         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
1711 }
1712
1713 int btrfs_find_device_by_path(struct btrfs_root *root, char *device_path,
1714                               struct btrfs_device **device)
1715 {
1716         int ret = 0;
1717         struct btrfs_super_block *disk_super;
1718         u64 devid;
1719         u8 *dev_uuid;
1720         struct block_device *bdev;
1721         struct buffer_head *bh;
1722
1723         *device = NULL;
1724         ret = btrfs_get_bdev_and_sb(device_path, FMODE_READ,
1725                                     root->fs_info->bdev_holder, 0, &bdev, &bh);
1726         if (ret)
1727                 return ret;
1728         disk_super = (struct btrfs_super_block *)bh->b_data;
1729         devid = btrfs_stack_device_id(&disk_super->dev_item);
1730         dev_uuid = disk_super->dev_item.uuid;
1731         *device = btrfs_find_device(root->fs_info, devid, dev_uuid,
1732                                     disk_super->fsid);
1733         brelse(bh);
1734         if (!*device)
1735                 ret = -ENOENT;
1736         blkdev_put(bdev, FMODE_READ);
1737         return ret;
1738 }
1739
1740 int btrfs_find_device_missing_or_by_path(struct btrfs_root *root,
1741                                          char *device_path,
1742                                          struct btrfs_device **device)
1743 {
1744         *device = NULL;
1745         if (strcmp(device_path, "missing") == 0) {
1746                 struct list_head *devices;
1747                 struct btrfs_device *tmp;
1748
1749                 devices = &root->fs_info->fs_devices->devices;
1750                 /*
1751                  * It is safe to read the devices since the volume_mutex
1752                  * is held by the caller.
1753                  */
1754                 list_for_each_entry(tmp, devices, dev_list) {
1755                         if (tmp->in_fs_metadata && !tmp->bdev) {
1756                                 *device = tmp;
1757                                 break;
1758                         }
1759                 }
1760
1761                 if (!*device) {
1762                         pr_err("btrfs: no missing device found\n");
1763                         return -ENOENT;
1764                 }
1765
1766                 return 0;
1767         } else {
1768                 return btrfs_find_device_by_path(root, device_path, device);
1769         }
1770 }
1771
1772 /*
1773  * does all the dirty work required for changing file system's UUID.
1774  */
1775 static int btrfs_prepare_sprout(struct btrfs_root *root)
1776 {
1777         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
1778         struct btrfs_fs_devices *old_devices;
1779         struct btrfs_fs_devices *seed_devices;
1780         struct btrfs_super_block *disk_super = root->fs_info->super_copy;
1781         struct btrfs_device *device;
1782         u64 super_flags;
1783
1784         BUG_ON(!mutex_is_locked(&uuid_mutex));
1785         if (!fs_devices->seeding)
1786                 return -EINVAL;
1787
1788         seed_devices = kzalloc(sizeof(*fs_devices), GFP_NOFS);
1789         if (!seed_devices)
1790                 return -ENOMEM;
1791
1792         old_devices = clone_fs_devices(fs_devices);
1793         if (IS_ERR(old_devices)) {
1794                 kfree(seed_devices);
1795                 return PTR_ERR(old_devices);
1796         }
1797
1798         list_add(&old_devices->list, &fs_uuids);
1799
1800         memcpy(seed_devices, fs_devices, sizeof(*seed_devices));
1801         seed_devices->opened = 1;
1802         INIT_LIST_HEAD(&seed_devices->devices);
1803         INIT_LIST_HEAD(&seed_devices->alloc_list);
1804         mutex_init(&seed_devices->device_list_mutex);
1805
1806         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
1807         list_splice_init_rcu(&fs_devices->devices, &seed_devices->devices,
1808                               synchronize_rcu);
1809         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1810
1811         list_splice_init(&fs_devices->alloc_list, &seed_devices->alloc_list);
1812         list_for_each_entry(device, &seed_devices->devices, dev_list) {
1813                 device->fs_devices = seed_devices;
1814         }
1815
1816         fs_devices->seeding = 0;
1817         fs_devices->num_devices = 0;
1818         fs_devices->open_devices = 0;
1819         fs_devices->total_devices = 0;
1820         fs_devices->seed = seed_devices;
1821
1822         generate_random_uuid(fs_devices->fsid);
1823         memcpy(root->fs_info->fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1824         memcpy(disk_super->fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1825         super_flags = btrfs_super_flags(disk_super) &
1826                       ~BTRFS_SUPER_FLAG_SEEDING;
1827         btrfs_set_super_flags(disk_super, super_flags);
1828
1829         return 0;
1830 }
1831
1832 /*
1833  * strore the expected generation for seed devices in device items.
1834  */
1835 static int btrfs_finish_sprout(struct btrfs_trans_handle *trans,
1836                                struct btrfs_root *root)
1837 {
1838         struct btrfs_path *path;
1839         struct extent_buffer *leaf;
1840         struct btrfs_dev_item *dev_item;
1841         struct btrfs_device *device;
1842         struct btrfs_key key;
1843         u8 fs_uuid[BTRFS_UUID_SIZE];
1844         u8 dev_uuid[BTRFS_UUID_SIZE];
1845         u64 devid;
1846         int ret;
1847
1848         path = btrfs_alloc_path();
1849         if (!path)
1850                 return -ENOMEM;
1851
1852         root = root->fs_info->chunk_root;
1853         key.objectid = BTRFS_DEV_ITEMS_OBJECTID;
1854         key.offset = 0;
1855         key.type = BTRFS_DEV_ITEM_KEY;
1856
1857         while (1) {
1858                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1859                 if (ret < 0)
1860                         goto error;
1861
1862                 leaf = path->nodes[0];
1863 next_slot:
1864                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1865                         ret = btrfs_next_leaf(root, path);
1866                         if (ret > 0)
1867                                 break;
1868                         if (ret < 0)
1869                                 goto error;
1870                         leaf = path->nodes[0];
1871                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1872                         btrfs_release_path(path);
1873                         continue;
1874                 }
1875
1876                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1877                 if (key.objectid != BTRFS_DEV_ITEMS_OBJECTID ||
1878                     key.type != BTRFS_DEV_ITEM_KEY)
1879                         break;
1880
1881                 dev_item = btrfs_item_ptr(leaf, path->slots[0],
1882                                           struct btrfs_dev_item);
1883                 devid = btrfs_device_id(leaf, dev_item);
1884                 read_extent_buffer(leaf, dev_uuid,
1885                                    (unsigned long)btrfs_device_uuid(dev_item),
1886                                    BTRFS_UUID_SIZE);
1887                 read_extent_buffer(leaf, fs_uuid,
1888                                    (unsigned long)btrfs_device_fsid(dev_item),
1889                                    BTRFS_UUID_SIZE);
1890                 device = btrfs_find_device(root->fs_info, devid, dev_uuid,
1891                                            fs_uuid);
1892                 BUG_ON(!device); /* Logic error */
1893
1894                 if (device->fs_devices->seeding) {
1895                         btrfs_set_device_generation(leaf, dev_item,
1896                                                     device->generation);
1897                         btrfs_mark_buffer_dirty(leaf);
1898                 }
1899
1900                 path->slots[0]++;
1901                 goto next_slot;
1902         }
1903         ret = 0;
1904 error:
1905         btrfs_free_path(path);
1906         return ret;
1907 }
1908
1909 int btrfs_init_new_device(struct btrfs_root *root, char *device_path)
1910 {
1911         struct request_queue *q;
1912         struct btrfs_trans_handle *trans;
1913         struct btrfs_device *device;
1914         struct block_device *bdev;
1915         struct list_head *devices;
1916         struct super_block *sb = root->fs_info->sb;
1917         struct rcu_string *name;
1918         u64 total_bytes;
1919         int seeding_dev = 0;
1920         int ret = 0;
1921
1922         if ((sb->s_flags & MS_RDONLY) && !root->fs_info->fs_devices->seeding)
1923                 return -EROFS;
1924
1925         bdev = blkdev_get_by_path(device_path, FMODE_WRITE | FMODE_EXCL,
1926                                   root->fs_info->bdev_holder);
1927         if (IS_ERR(bdev))
1928                 return PTR_ERR(bdev);
1929
1930         if (root->fs_info->fs_devices->seeding) {
1931                 seeding_dev = 1;
1932                 down_write(&sb->s_umount);
1933                 mutex_lock(&uuid_mutex);
1934         }
1935
1936         filemap_write_and_wait(bdev->bd_inode->i_mapping);
1937
1938         devices = &root->fs_info->fs_devices->devices;
1939
1940         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
1941         list_for_each_entry(device, devices, dev_list) {
1942                 if (device->bdev == bdev) {
1943                         ret = -EEXIST;
1944                         mutex_unlock(
1945                                 &root->fs_info->fs_devices->device_list_mutex);
1946                         goto error;
1947                 }
1948         }
1949         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1950
1951         device = kzalloc(sizeof(*device), GFP_NOFS);
1952         if (!device) {
1953                 /* we can safely leave the fs_devices entry around */
1954                 ret = -ENOMEM;
1955                 goto error;
1956         }
1957
1958         name = rcu_string_strdup(device_path, GFP_NOFS);
1959         if (!name) {
1960                 kfree(device);
1961                 ret = -ENOMEM;
1962                 goto error;
1963         }
1964         rcu_assign_pointer(device->name, name);
1965
1966         ret = find_next_devid(root, &device->devid);
1967         if (ret) {
1968                 rcu_string_free(device->name);
1969                 kfree(device);
1970                 goto error;
1971         }
1972
1973         trans = btrfs_start_transaction(root, 0);
1974         if (IS_ERR(trans)) {
1975                 rcu_string_free(device->name);
1976                 kfree(device);
1977                 ret = PTR_ERR(trans);
1978                 goto error;
1979         }
1980
1981         lock_chunks(root);
1982
1983         q = bdev_get_queue(bdev);
1984         if (blk_queue_discard(q))
1985                 device->can_discard = 1;
1986         device->writeable = 1;
1987         device->work.func = pending_bios_fn;
1988         generate_random_uuid(device->uuid);
1989         spin_lock_init(&device->io_lock);
1990         device->generation = trans->transid;
1991         device->io_width = root->sectorsize;
1992         device->io_align = root->sectorsize;
1993         device->sector_size = root->sectorsize;
1994         device->total_bytes = i_size_read(bdev->bd_inode);
1995         device->disk_total_bytes = device->total_bytes;
1996         device->dev_root = root->fs_info->dev_root;
1997         device->bdev = bdev;
1998         device->in_fs_metadata = 1;
1999         device->is_tgtdev_for_dev_replace = 0;
2000         device->mode = FMODE_EXCL;
2001         set_blocksize(device->bdev, 4096);
2002
2003         if (seeding_dev) {
2004                 sb->s_flags &= ~MS_RDONLY;
2005                 ret = btrfs_prepare_sprout(root);
2006                 BUG_ON(ret); /* -ENOMEM */
2007         }
2008
2009         device->fs_devices = root->fs_info->fs_devices;
2010
2011         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
2012         list_add_rcu(&device->dev_list, &root->fs_info->fs_devices->devices);
2013         list_add(&device->dev_alloc_list,
2014                  &root->fs_info->fs_devices->alloc_list);
2015         root->fs_info->fs_devices->num_devices++;
2016         root->fs_info->fs_devices->open_devices++;
2017         root->fs_info->fs_devices->rw_devices++;
2018         root->fs_info->fs_devices->total_devices++;
2019         if (device->can_discard)
2020                 root->fs_info->fs_devices->num_can_discard++;
2021         root->fs_info->fs_devices->total_rw_bytes += device->total_bytes;
2022
2023         spin_lock(&root->fs_info->free_chunk_lock);
2024         root->fs_info->free_chunk_space += device->total_bytes;
2025         spin_unlock(&root->fs_info->free_chunk_lock);
2026
2027         if (!blk_queue_nonrot(bdev_get_queue(bdev)))
2028                 root->fs_info->fs_devices->rotating = 1;
2029
2030         total_bytes = btrfs_super_total_bytes(root->fs_info->super_copy);
2031         btrfs_set_super_total_bytes(root->fs_info->super_copy,
2032                                     total_bytes + device->total_bytes);
2033
2034         total_bytes = btrfs_super_num_devices(root->fs_info->super_copy);
2035         btrfs_set_super_num_devices(root->fs_info->super_copy,
2036                                     total_bytes + 1);
2037         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2038
2039         if (seeding_dev) {
2040                 ret = init_first_rw_device(trans, root, device);
2041                 if (ret) {
2042                         btrfs_abort_transaction(trans, root, ret);
2043                         goto error_trans;
2044                 }
2045                 ret = btrfs_finish_sprout(trans, root);
2046                 if (ret) {
2047                         btrfs_abort_transaction(trans, root, ret);
2048                         goto error_trans;
2049                 }
2050         } else {
2051                 ret = btrfs_add_device(trans, root, device);
2052                 if (ret) {
2053                         btrfs_abort_transaction(trans, root, ret);
2054                         goto error_trans;
2055                 }
2056         }
2057
2058         /*
2059          * we've got more storage, clear any full flags on the space
2060          * infos
2061          */
2062         btrfs_clear_space_info_full(root->fs_info);
2063
2064         unlock_chunks(root);
2065         root->fs_info->num_tolerated_disk_barrier_failures =
2066                 btrfs_calc_num_tolerated_disk_barrier_failures(root->fs_info);
2067         ret = btrfs_commit_transaction(trans, root);
2068
2069         if (seeding_dev) {
2070                 mutex_unlock(&uuid_mutex);
2071                 up_write(&sb->s_umount);
2072
2073                 if (ret) /* transaction commit */
2074                         return ret;
2075
2076                 ret = btrfs_relocate_sys_chunks(root);
2077                 if (ret < 0)
2078                         btrfs_error(root->fs_info, ret,
2079                                     "Failed to relocate sys chunks after "
2080                                     "device initialization. This can be fixed "
2081                                     "using the \"btrfs balance\" command.");
2082                 trans = btrfs_attach_transaction(root);
2083                 if (IS_ERR(trans)) {
2084                         if (PTR_ERR(trans) == -ENOENT)
2085                                 return 0;
2086                         return PTR_ERR(trans);
2087                 }
2088                 ret = btrfs_commit_transaction(trans, root);
2089         }
2090
2091         return ret;
2092
2093 error_trans:
2094         unlock_chunks(root);
2095         btrfs_end_transaction(trans, root);
2096         rcu_string_free(device->name);
2097         kfree(device);
2098 error:
2099         blkdev_put(bdev, FMODE_EXCL);
2100         if (seeding_dev) {
2101                 mutex_unlock(&uuid_mutex);
2102                 up_write(&sb->s_umount);
2103         }
2104         return ret;
2105 }
2106
2107 int btrfs_init_dev_replace_tgtdev(struct btrfs_root *root, char *device_path,
2108                                   struct btrfs_device **device_out)
2109 {
2110         struct request_queue *q;
2111         struct btrfs_device *device;
2112         struct block_device *bdev;
2113         struct btrfs_fs_info *fs_info = root->fs_info;
2114         struct list_head *devices;
2115         struct rcu_string *name;
2116         int ret = 0;
2117
2118         *device_out = NULL;
2119         if (fs_info->fs_devices->seeding)
2120                 return -EINVAL;
2121
2122         bdev = blkdev_get_by_path(device_path, FMODE_WRITE | FMODE_EXCL,
2123                                   fs_info->bdev_holder);
2124         if (IS_ERR(bdev))
2125                 return PTR_ERR(bdev);
2126
2127         filemap_write_and_wait(bdev->bd_inode->i_mapping);
2128
2129         devices = &fs_info->fs_devices->devices;
2130         list_for_each_entry(device, devices, dev_list) {
2131                 if (device->bdev == bdev) {
2132                         ret = -EEXIST;
2133                         goto error;
2134                 }
2135         }
2136
2137         device = kzalloc(sizeof(*device), GFP_NOFS);
2138         if (!device) {
2139                 ret = -ENOMEM;
2140                 goto error;
2141         }
2142
2143         name = rcu_string_strdup(device_path, GFP_NOFS);
2144         if (!name) {
2145                 kfree(device);
2146                 ret = -ENOMEM;
2147                 goto error;
2148         }
2149         rcu_assign_pointer(device->name, name);
2150
2151         q = bdev_get_queue(bdev);
2152         if (blk_queue_discard(q))
2153                 device->can_discard = 1;
2154         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
2155         device->writeable = 1;
2156         device->work.func = pending_bios_fn;
2157         generate_random_uuid(device->uuid);
2158         device->devid = BTRFS_DEV_REPLACE_DEVID;
2159         spin_lock_init(&device->io_lock);
2160         device->generation = 0;
2161         device->io_width = root->sectorsize;
2162         device->io_align = root->sectorsize;
2163         device->sector_size = root->sectorsize;
2164         device->total_bytes = i_size_read(bdev->bd_inode);
2165         device->disk_total_bytes = device->total_bytes;
2166         device->dev_root = fs_info->dev_root;
2167         device->bdev = bdev;
2168         device->in_fs_metadata = 1;
2169         device->is_tgtdev_for_dev_replace = 1;
2170         device->mode = FMODE_EXCL;
2171         set_blocksize(device->bdev, 4096);
2172         device->fs_devices = fs_info->fs_devices;
2173         list_add(&device->dev_list, &fs_info->fs_devices->devices);
2174         fs_info->fs_devices->num_devices++;
2175         fs_info->fs_devices->open_devices++;
2176         if (device->can_discard)
2177                 fs_info->fs_devices->num_can_discard++;
2178         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2179
2180         *device_out = device;
2181         return ret;
2182
2183 error:
2184         blkdev_put(bdev, FMODE_EXCL);
2185         return ret;
2186 }
2187
2188 void btrfs_init_dev_replace_tgtdev_for_resume(struct btrfs_fs_info *fs_info,
2189                                               struct btrfs_device *tgtdev)
2190 {
2191         WARN_ON(fs_info->fs_devices->rw_devices == 0);
2192         tgtdev->io_width = fs_info->dev_root->sectorsize;
2193         tgtdev->io_align = fs_info->dev_root->sectorsize;
2194         tgtdev->sector_size = fs_info->dev_root->sectorsize;
2195         tgtdev->dev_root = fs_info->dev_root;
2196         tgtdev->in_fs_metadata = 1;
2197 }
2198
2199 static noinline int btrfs_update_device(struct btrfs_trans_handle *trans,
2200                                         struct btrfs_device *device)
2201 {
2202         int ret;
2203         struct btrfs_path *path;
2204         struct btrfs_root *root;
2205         struct btrfs_dev_item *dev_item;
2206         struct extent_buffer *leaf;
2207         struct btrfs_key key;
2208
2209         root = device->dev_root->fs_info->chunk_root;
2210
2211         path = btrfs_alloc_path();
2212         if (!path)
2213                 return -ENOMEM;
2214
2215         key.objectid = BTRFS_DEV_ITEMS_OBJECTID;
2216         key.type = BTRFS_DEV_ITEM_KEY;
2217         key.offset = device->devid;
2218
2219         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2220         if (ret < 0)
2221                 goto out;
2222
2223         if (ret > 0) {
2224                 ret = -ENOENT;
2225                 goto out;
2226         }
2227
2228         leaf = path->nodes[0];
2229         dev_item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_item);
2230
2231         btrfs_set_device_id(leaf, dev_item, device->devid);
2232         btrfs_set_device_type(leaf, dev_item, device->type);
2233         btrfs_set_device_io_align(leaf, dev_item, device->io_align);
2234         btrfs_set_device_io_width(leaf, dev_item, device->io_width);
2235         btrfs_set_device_sector_size(leaf, dev_item, device->sector_size);
2236         btrfs_set_device_total_bytes(leaf, dev_item, device->disk_total_bytes);
2237         btrfs_set_device_bytes_used(leaf, dev_item, device->bytes_used);
2238         btrfs_mark_buffer_dirty(leaf);
2239
2240 out:
2241         btrfs_free_path(path);
2242         return ret;
2243 }
2244
2245 static int __btrfs_grow_device(struct btrfs_trans_handle *trans,
2246                       struct btrfs_device *device, u64 new_size)
2247 {
2248         struct btrfs_super_block *super_copy =
2249                 device->dev_root->fs_info->super_copy;
2250         u64 old_total = btrfs_super_total_bytes(super_copy);
2251         u64 diff = new_size - device->total_bytes;
2252
2253         if (!device->writeable)
2254                 return -EACCES;
2255         if (new_size <= device->total_bytes ||
2256             device->is_tgtdev_for_dev_replace)
2257                 return -EINVAL;
2258
2259         btrfs_set_super_total_bytes(super_copy, old_total + diff);
2260         device->fs_devices->total_rw_bytes += diff;
2261
2262         device->total_bytes = new_size;
2263         device->disk_total_bytes = new_size;
2264         btrfs_clear_space_info_full(device->dev_root->fs_info);
2265
2266         return btrfs_update_device(trans, device);
2267 }
2268
2269 int btrfs_grow_device(struct btrfs_trans_handle *trans,
2270                       struct btrfs_device *device, u64 new_size)
2271 {
2272         int ret;
2273         lock_chunks(device->dev_root);
2274         ret = __btrfs_grow_device(trans, device, new_size);
2275         unlock_chunks(device->dev_root);
2276         return ret;
2277 }
2278
2279 static int btrfs_free_chunk(struct btrfs_trans_handle *trans,
2280                             struct btrfs_root *root,
2281                             u64 chunk_tree, u64 chunk_objectid,
2282                             u64 chunk_offset)
2283 {
2284         int ret;
2285         struct btrfs_path *path;
2286         struct btrfs_key key;
2287
2288         root = root->fs_info->chunk_root;
2289         path = btrfs_alloc_path();
2290         if (!path)
2291                 return -ENOMEM;
2292
2293         key.objectid = chunk_objectid;
2294         key.offset = chunk_offset;
2295         key.type = BTRFS_CHUNK_ITEM_KEY;
2296
2297         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
2298         if (ret < 0)
2299                 goto out;
2300         else if (ret > 0) { /* Logic error or corruption */
2301                 btrfs_error(root->fs_info, -ENOENT,
2302                             "Failed lookup while freeing chunk.");
2303                 ret = -ENOENT;
2304                 goto out;
2305         }
2306
2307         ret = btrfs_del_item(trans, root, path);
2308         if (ret < 0)
2309                 btrfs_error(root->fs_info, ret,
2310                             "Failed to delete chunk item.");
2311 out:
2312         btrfs_free_path(path);
2313         return ret;
2314 }
2315
2316 static int btrfs_del_sys_chunk(struct btrfs_root *root, u64 chunk_objectid, u64
2317                         chunk_offset)
2318 {
2319         struct btrfs_super_block *super_copy = root->fs_info->super_copy;
2320         struct btrfs_disk_key *disk_key;
2321         struct btrfs_chunk *chunk;
2322         u8 *ptr;
2323         int ret = 0;
2324         u32 num_stripes;
2325         u32 array_size;
2326         u32 len = 0;
2327         u32 cur;
2328         struct btrfs_key key;
2329
2330         array_size = btrfs_super_sys_array_size(super_copy);
2331
2332         ptr = super_copy->sys_chunk_array;
2333         cur = 0;
2334
2335         while (cur < array_size) {
2336                 disk_key = (struct btrfs_disk_key *)ptr;
2337                 btrfs_disk_key_to_cpu(&key, disk_key);
2338
2339                 len = sizeof(*disk_key);
2340
2341                 if (key.type == BTRFS_CHUNK_ITEM_KEY) {
2342                         chunk = (struct btrfs_chunk *)(ptr + len);
2343                         num_stripes = btrfs_stack_chunk_num_stripes(chunk);
2344                         len += btrfs_chunk_item_size(num_stripes);
2345                 } else {
2346                         ret = -EIO;
2347                         break;
2348                 }
2349                 if (key.objectid == chunk_objectid &&
2350                     key.offset == chunk_offset) {
2351                         memmove(ptr, ptr + len, array_size - (cur + len));
2352                         array_size -= len;
2353                         btrfs_set_super_sys_array_size(super_copy, array_size);
2354                 } else {
2355                         ptr += len;
2356                         cur += len;
2357                 }
2358         }
2359         return ret;
2360 }
2361
2362 static int btrfs_relocate_chunk(struct btrfs_root *root,
2363                          u64 chunk_tree, u64 chunk_objectid,
2364                          u64 chunk_offset)
2365 {
2366         struct extent_map_tree *em_tree;
2367         struct btrfs_root *extent_root;
2368         struct btrfs_trans_handle *trans;
2369         struct extent_map *em;
2370         struct map_lookup *map;
2371         int ret;
2372         int i;
2373
2374         root = root->fs_info->chunk_root;
2375         extent_root = root->fs_info->extent_root;
2376         em_tree = &root->fs_info->mapping_tree.map_tree;
2377
2378         ret = btrfs_can_relocate(extent_root, chunk_offset);
2379         if (ret)
2380                 return -ENOSPC;
2381
2382         /* step one, relocate all the extents inside this chunk */
2383         ret = btrfs_relocate_block_group(extent_root, chunk_offset);
2384         if (ret)
2385                 return ret;
2386
2387         trans = btrfs_start_transaction(root, 0);
2388         if (IS_ERR(trans)) {
2389                 ret = PTR_ERR(trans);
2390                 btrfs_std_error(root->fs_info, ret);
2391                 return ret;
2392         }
2393
2394         lock_chunks(root);
2395
2396         /*
2397          * step two, delete the device extents and the
2398          * chunk tree entries
2399          */
2400         read_lock(&em_tree->lock);
2401         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
2402         read_unlock(&em_tree->lock);
2403
2404         BUG_ON(!em || em->start > chunk_offset ||
2405                em->start + em->len < chunk_offset);
2406         map = (struct map_lookup *)em->bdev;
2407
2408         for (i = 0; i < map->num_stripes; i++) {
2409                 ret = btrfs_free_dev_extent(trans, map->stripes[i].dev,
2410                                             map->stripes[i].physical);
2411                 BUG_ON(ret);
2412
2413                 if (map->stripes[i].dev) {
2414                         ret = btrfs_update_device(trans, map->stripes[i].dev);
2415                         BUG_ON(ret);
2416                 }
2417         }
2418         ret = btrfs_free_chunk(trans, root, chunk_tree, chunk_objectid,
2419                                chunk_offset);
2420
2421         BUG_ON(ret);
2422
2423         trace_btrfs_chunk_free(root, map, chunk_offset, em->len);
2424
2425         if (map->type & BTRFS_BLOCK_GROUP_SYSTEM) {
2426                 ret = btrfs_del_sys_chunk(root, chunk_objectid, chunk_offset);
2427                 BUG_ON(ret);
2428         }
2429
2430         ret = btrfs_remove_block_group(trans, extent_root, chunk_offset);
2431         BUG_ON(ret);
2432
2433         write_lock(&em_tree->lock);
2434         remove_extent_mapping(em_tree, em);
2435         write_unlock(&em_tree->lock);
2436
2437         kfree(map);
2438         em->bdev = NULL;
2439
2440         /* once for the tree */
2441         free_extent_map(em);
2442         /* once for us */
2443         free_extent_map(em);
2444
2445         unlock_chunks(root);
2446         btrfs_end_transaction(trans, root);
2447         return 0;
2448 }
2449
2450 static int btrfs_relocate_sys_chunks(struct btrfs_root *root)
2451 {
2452         struct btrfs_root *chunk_root = root->fs_info->chunk_root;
2453         struct btrfs_path *path;
2454         struct extent_buffer *leaf;
2455         struct btrfs_chunk *chunk;
2456         struct btrfs_key key;
2457         struct btrfs_key found_key;
2458         u64 chunk_tree = chunk_root->root_key.objectid;
2459         u64 chunk_type;
2460         bool retried = false;
2461         int failed = 0;
2462         int ret;
2463
2464         path = btrfs_alloc_path();
2465         if (!path)
2466                 return -ENOMEM;
2467
2468 again:
2469         key.objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2470         key.offset = (u64)-1;
2471         key.type = BTRFS_CHUNK_ITEM_KEY;
2472
2473         while (1) {
2474                 ret = btrfs_search_slot(NULL, chunk_root, &key, path, 0, 0);
2475                 if (ret < 0)
2476                         goto error;
2477                 BUG_ON(ret == 0); /* Corruption */
2478
2479                 ret = btrfs_previous_item(chunk_root, path, key.objectid,
2480                                           key.type);
2481                 if (ret < 0)
2482                         goto error;
2483                 if (ret > 0)
2484                         break;
2485
2486                 leaf = path->nodes[0];
2487                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2488
2489                 chunk = btrfs_item_ptr(leaf, path->slots[0],
2490                                        struct btrfs_chunk);
2491                 chunk_type = btrfs_chunk_type(leaf, chunk);
2492                 btrfs_release_path(path);
2493
2494                 if (chunk_type & BTRFS_BLOCK_GROUP_SYSTEM) {
2495                         ret = btrfs_relocate_chunk(chunk_root, chunk_tree,
2496                                                    found_key.objectid,
2497                                                    found_key.offset);
2498                         if (ret == -ENOSPC)
2499                                 failed++;
2500                         else if (ret)
2501                                 BUG();
2502                 }
2503
2504                 if (found_key.offset == 0)
2505                         break;
2506                 key.offset = found_key.offset - 1;
2507         }
2508         ret = 0;
2509         if (failed && !retried) {
2510                 failed = 0;
2511                 retried = true;
2512                 goto again;
2513         } else if (failed && retried) {
2514                 WARN_ON(1);
2515                 ret = -ENOSPC;
2516         }
2517 error:
2518         btrfs_free_path(path);
2519         return ret;
2520 }
2521
2522 static int insert_balance_item(struct btrfs_root *root,
2523                                struct btrfs_balance_control *bctl)
2524 {
2525         struct btrfs_trans_handle *trans;
2526         struct btrfs_balance_item *item;
2527         struct btrfs_disk_balance_args disk_bargs;
2528         struct btrfs_path *path;
2529         struct extent_buffer *leaf;
2530         struct btrfs_key key;
2531         int ret, err;
2532
2533         path = btrfs_alloc_path();
2534         if (!path)
2535                 return -ENOMEM;
2536
2537         trans = btrfs_start_transaction(root, 0);
2538         if (IS_ERR(trans)) {
2539                 btrfs_free_path(path);
2540                 return PTR_ERR(trans);
2541         }
2542
2543         key.objectid = BTRFS_BALANCE_OBJECTID;
2544         key.type = BTRFS_BALANCE_ITEM_KEY;
2545         key.offset = 0;
2546
2547         ret = btrfs_insert_empty_item(trans, root, path, &key,
2548                                       sizeof(*item));
2549         if (ret)
2550                 goto out;
2551
2552         leaf = path->nodes[0];
2553         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_balance_item);
2554
2555         memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
2556
2557         btrfs_cpu_balance_args_to_disk(&disk_bargs, &bctl->data);
2558         btrfs_set_balance_data(leaf, item, &disk_bargs);
2559         btrfs_cpu_balance_args_to_disk(&disk_bargs, &bctl->meta);
2560         btrfs_set_balance_meta(leaf, item, &disk_bargs);
2561         btrfs_cpu_balance_args_to_disk(&disk_bargs, &bctl->sys);
2562         btrfs_set_balance_sys(leaf, item, &disk_bargs);
2563
2564         btrfs_set_balance_flags(leaf, item, bctl->flags);
2565
2566         btrfs_mark_buffer_dirty(leaf);
2567 out:
2568         btrfs_free_path(path);
2569         err = btrfs_commit_transaction(trans, root);
2570         if (err && !ret)
2571                 ret = err;
2572         return ret;
2573 }
2574
2575 static int del_balance_item(struct btrfs_root *root)
2576 {
2577         struct btrfs_trans_handle *trans;
2578         struct btrfs_path *path;
2579         struct btrfs_key key;
2580         int ret, err;
2581
2582         path = btrfs_alloc_path();
2583         if (!path)
2584                 return -ENOMEM;
2585
2586         trans = btrfs_start_transaction(root, 0);
2587         if (IS_ERR(trans)) {
2588                 btrfs_free_path(path);
2589                 return PTR_ERR(trans);
2590         }
2591
2592         key.objectid = BTRFS_BALANCE_OBJECTID;
2593         key.type = BTRFS_BALANCE_ITEM_KEY;
2594         key.offset = 0;
2595
2596         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
2597         if (ret < 0)
2598                 goto out;
2599         if (ret > 0) {
2600                 ret = -ENOENT;
2601                 goto out;
2602         }
2603
2604         ret = btrfs_del_item(trans, root, path);
2605 out:
2606         btrfs_free_path(path);
2607         err = btrfs_commit_transaction(trans, root);
2608         if (err && !ret)
2609                 ret = err;
2610         return ret;
2611 }
2612
2613 /*
2614  * This is a heuristic used to reduce the number of chunks balanced on
2615  * resume after balance was interrupted.
2616  */
2617 static void update_balance_args(struct btrfs_balance_control *bctl)
2618 {
2619         /*
2620          * Turn on soft mode for chunk types that were being converted.
2621          */
2622         if (bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT)
2623                 bctl->data.flags |= BTRFS_BALANCE_ARGS_SOFT;
2624         if (bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT)
2625                 bctl->sys.flags |= BTRFS_BALANCE_ARGS_SOFT;
2626         if (bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT)
2627                 bctl->meta.flags |= BTRFS_BALANCE_ARGS_SOFT;
2628
2629         /*
2630          * Turn on usage filter if is not already used.  The idea is
2631          * that chunks that we have already balanced should be
2632          * reasonably full.  Don't do it for chunks that are being
2633          * converted - that will keep us from relocating unconverted
2634          * (albeit full) chunks.
2635          */
2636         if (!(bctl->data.flags & BTRFS_BALANCE_ARGS_USAGE) &&
2637             !(bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT)) {
2638                 bctl->data.flags |= BTRFS_BALANCE_ARGS_USAGE;
2639                 bctl->data.usage = 90;
2640         }
2641         if (!(bctl->sys.flags & BTRFS_BALANCE_ARGS_USAGE) &&
2642             !(bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT)) {
2643                 bctl->sys.flags |= BTRFS_BALANCE_ARGS_USAGE;
2644                 bctl->sys.usage = 90;
2645         }
2646         if (!(bctl->meta.flags & BTRFS_BALANCE_ARGS_USAGE) &&
2647             !(bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT)) {
2648                 bctl->meta.flags |= BTRFS_BALANCE_ARGS_USAGE;
2649                 bctl->meta.usage = 90;
2650         }
2651 }
2652
2653 /*
2654  * Should be called with both balance and volume mutexes held to
2655  * serialize other volume operations (add_dev/rm_dev/resize) with
2656  * restriper.  Same goes for unset_balance_control.
2657  */
2658 static void set_balance_control(struct btrfs_balance_control *bctl)
2659 {
2660         struct btrfs_fs_info *fs_info = bctl->fs_info;
2661
2662         BUG_ON(fs_info->balance_ctl);
2663
2664         spin_lock(&fs_info->balance_lock);
2665         fs_info->balance_ctl = bctl;
2666         spin_unlock(&fs_info->balance_lock);
2667 }
2668
2669 static void unset_balance_control(struct btrfs_fs_info *fs_info)
2670 {
2671         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
2672
2673         BUG_ON(!fs_info->balance_ctl);
2674
2675         spin_lock(&fs_info->balance_lock);
2676         fs_info->balance_ctl = NULL;
2677         spin_unlock(&fs_info->balance_lock);
2678
2679         kfree(bctl);
2680 }
2681
2682 /*
2683  * Balance filters.  Return 1 if chunk should be filtered out
2684  * (should not be balanced).
2685  */
2686 static int chunk_profiles_filter(u64 chunk_type,
2687                                  struct btrfs_balance_args *bargs)
2688 {
2689         chunk_type = chunk_to_extended(chunk_type) &
2690                                 BTRFS_EXTENDED_PROFILE_MASK;
2691
2692         if (bargs->profiles & chunk_type)
2693                 return 0;
2694
2695         return 1;
2696 }
2697
2698 static int chunk_usage_filter(struct btrfs_fs_info *fs_info, u64 chunk_offset,
2699                               struct btrfs_balance_args *bargs)
2700 {
2701         struct btrfs_block_group_cache *cache;
2702         u64 chunk_used, user_thresh;
2703         int ret = 1;
2704
2705         cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2706         chunk_used = btrfs_block_group_used(&cache->item);
2707
2708         if (bargs->usage == 0)
2709                 user_thresh = 1;
2710         else if (bargs->usage > 100)
2711                 user_thresh = cache->key.offset;
2712         else
2713                 user_thresh = div_factor_fine(cache->key.offset,
2714                                               bargs->usage);
2715
2716         if (chunk_used < user_thresh)
2717                 ret = 0;
2718
2719         btrfs_put_block_group(cache);
2720         return ret;
2721 }
2722
2723 static int chunk_devid_filter(struct extent_buffer *leaf,
2724                               struct btrfs_chunk *chunk,
2725                               struct btrfs_balance_args *bargs)
2726 {
2727         struct btrfs_stripe *stripe;
2728         int num_stripes = btrfs_chunk_num_stripes(leaf, chunk);
2729         int i;
2730
2731         for (i = 0; i < num_stripes; i++) {
2732                 stripe = btrfs_stripe_nr(chunk, i);
2733                 if (btrfs_stripe_devid(leaf, stripe) == bargs->devid)
2734                         return 0;
2735         }
2736
2737         return 1;
2738 }
2739
2740 /* [pstart, pend) */
2741 static int chunk_drange_filter(struct extent_buffer *leaf,
2742                                struct btrfs_chunk *chunk,
2743                                u64 chunk_offset,
2744                                struct btrfs_balance_args *bargs)
2745 {
2746         struct btrfs_stripe *stripe;
2747         int num_stripes = btrfs_chunk_num_stripes(leaf, chunk);
2748         u64 stripe_offset;
2749         u64 stripe_length;
2750         int factor;
2751         int i;
2752
2753         if (!(bargs->flags & BTRFS_BALANCE_ARGS_DEVID))
2754                 return 0;
2755
2756         if (btrfs_chunk_type(leaf, chunk) & (BTRFS_BLOCK_GROUP_DUP |
2757              BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10)) {
2758                 factor = num_stripes / 2;
2759         } else if (btrfs_chunk_type(leaf, chunk) & BTRFS_BLOCK_GROUP_RAID5) {
2760                 factor = num_stripes - 1;
2761         } else if (btrfs_chunk_type(leaf, chunk) & BTRFS_BLOCK_GROUP_RAID6) {
2762                 factor = num_stripes - 2;
2763         } else {
2764                 factor = num_stripes;
2765         }
2766
2767         for (i = 0; i < num_stripes; i++) {
2768                 stripe = btrfs_stripe_nr(chunk, i);
2769                 if (btrfs_stripe_devid(leaf, stripe) != bargs->devid)
2770                         continue;
2771
2772                 stripe_offset = btrfs_stripe_offset(leaf, stripe);
2773                 stripe_length = btrfs_chunk_length(leaf, chunk);
2774                 do_div(stripe_length, factor);
2775
2776                 if (stripe_offset < bargs->pend &&
2777                     stripe_offset + stripe_length > bargs->pstart)
2778                         return 0;
2779         }
2780
2781         return 1;
2782 }
2783
2784 /* [vstart, vend) */
2785 static int chunk_vrange_filter(struct extent_buffer *leaf,
2786                                struct btrfs_chunk *chunk,
2787                                u64 chunk_offset,
2788                                struct btrfs_balance_args *bargs)
2789 {
2790         if (chunk_offset < bargs->vend &&
2791             chunk_offset + btrfs_chunk_length(leaf, chunk) > bargs->vstart)
2792                 /* at least part of the chunk is inside this vrange */
2793                 return 0;
2794
2795         return 1;
2796 }
2797
2798 static int chunk_soft_convert_filter(u64 chunk_type,
2799                                      struct btrfs_balance_args *bargs)
2800 {
2801         if (!(bargs->flags & BTRFS_BALANCE_ARGS_CONVERT))
2802                 return 0;
2803
2804         chunk_type = chunk_to_extended(chunk_type) &
2805                                 BTRFS_EXTENDED_PROFILE_MASK;
2806
2807         if (bargs->target == chunk_type)
2808                 return 1;
2809
2810         return 0;
2811 }
2812
2813 static int should_balance_chunk(struct btrfs_root *root,
2814                                 struct extent_buffer *leaf,
2815                                 struct btrfs_chunk *chunk, u64 chunk_offset)
2816 {
2817         struct btrfs_balance_control *bctl = root->fs_info->balance_ctl;
2818         struct btrfs_balance_args *bargs = NULL;
2819         u64 chunk_type = btrfs_chunk_type(leaf, chunk);
2820
2821         /* type filter */
2822         if (!((chunk_type & BTRFS_BLOCK_GROUP_TYPE_MASK) &
2823               (bctl->flags & BTRFS_BALANCE_TYPE_MASK))) {
2824                 return 0;
2825         }
2826
2827         if (chunk_type & BTRFS_BLOCK_GROUP_DATA)
2828                 bargs = &bctl->data;
2829         else if (chunk_type & BTRFS_BLOCK_GROUP_SYSTEM)
2830                 bargs = &bctl->sys;
2831         else if (chunk_type & BTRFS_BLOCK_GROUP_METADATA)
2832                 bargs = &bctl->meta;
2833
2834         /* profiles filter */
2835         if ((bargs->flags & BTRFS_BALANCE_ARGS_PROFILES) &&
2836             chunk_profiles_filter(chunk_type, bargs)) {
2837                 return 0;
2838         }
2839
2840         /* usage filter */
2841         if ((bargs->flags & BTRFS_BALANCE_ARGS_USAGE) &&
2842             chunk_usage_filter(bctl->fs_info, chunk_offset, bargs)) {
2843                 return 0;
2844         }
2845
2846         /* devid filter */
2847         if ((bargs->flags & BTRFS_BALANCE_ARGS_DEVID) &&
2848             chunk_devid_filter(leaf, chunk, bargs)) {
2849                 return 0;
2850         }
2851
2852         /* drange filter, makes sense only with devid filter */
2853         if ((bargs->flags & BTRFS_BALANCE_ARGS_DRANGE) &&
2854             chunk_drange_filter(leaf, chunk, chunk_offset, bargs)) {
2855                 return 0;
2856         }
2857
2858         /* vrange filter */
2859         if ((bargs->flags & BTRFS_BALANCE_ARGS_VRANGE) &&
2860             chunk_vrange_filter(leaf, chunk, chunk_offset, bargs)) {
2861                 return 0;
2862         }
2863
2864         /* soft profile changing mode */
2865         if ((bargs->flags & BTRFS_BALANCE_ARGS_SOFT) &&
2866             chunk_soft_convert_filter(chunk_type, bargs)) {
2867                 return 0;
2868         }
2869
2870         return 1;
2871 }
2872
2873 static int __btrfs_balance(struct btrfs_fs_info *fs_info)
2874 {
2875         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
2876         struct btrfs_root *chunk_root = fs_info->chunk_root;
2877         struct btrfs_root *dev_root = fs_info->dev_root;
2878         struct list_head *devices;
2879         struct btrfs_device *device;
2880         u64 old_size;
2881         u64 size_to_free;
2882         struct btrfs_chunk *chunk;
2883         struct btrfs_path *path;
2884         struct btrfs_key key;
2885         struct btrfs_key found_key;
2886         struct btrfs_trans_handle *trans;
2887         struct extent_buffer *leaf;
2888         int slot;
2889         int ret;
2890         int enospc_errors = 0;
2891         bool counting = true;
2892
2893         /* step one make some room on all the devices */
2894         devices = &fs_info->fs_devices->devices;
2895         list_for_each_entry(device, devices, dev_list) {
2896                 old_size = device->total_bytes;
2897                 size_to_free = div_factor(old_size, 1);
2898                 size_to_free = min(size_to_free, (u64)1 * 1024 * 1024);
2899                 if (!device->writeable ||
2900                     device->total_bytes - device->bytes_used > size_to_free ||
2901                     device->is_tgtdev_for_dev_replace)
2902                         continue;
2903
2904                 ret = btrfs_shrink_device(device, old_size - size_to_free);
2905                 if (ret == -ENOSPC)
2906                         break;
2907                 BUG_ON(ret);
2908
2909                 trans = btrfs_start_transaction(dev_root, 0);
2910                 BUG_ON(IS_ERR(trans));
2911
2912                 ret = btrfs_grow_device(trans, device, old_size);
2913                 BUG_ON(ret);
2914
2915                 btrfs_end_transaction(trans, dev_root);
2916         }
2917
2918         /* step two, relocate all the chunks */
2919         path = btrfs_alloc_path();
2920         if (!path) {
2921                 ret = -ENOMEM;
2922                 goto error;
2923         }
2924
2925         /* zero out stat counters */
2926         spin_lock(&fs_info->balance_lock);
2927         memset(&bctl->stat, 0, sizeof(bctl->stat));
2928         spin_unlock(&fs_info->balance_lock);
2929 again:
2930         key.objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2931         key.offset = (u64)-1;
2932         key.type = BTRFS_CHUNK_ITEM_KEY;
2933
2934         while (1) {
2935                 if ((!counting && atomic_read(&fs_info->balance_pause_req)) ||
2936                     atomic_read(&fs_info->balance_cancel_req)) {
2937                         ret = -ECANCELED;
2938                         goto error;
2939                 }
2940
2941                 ret = btrfs_search_slot(NULL, chunk_root, &key, path, 0, 0);
2942                 if (ret < 0)
2943                         goto error;
2944
2945                 /*
2946                  * this shouldn't happen, it means the last relocate
2947                  * failed
2948                  */
2949                 if (ret == 0)
2950                         BUG(); /* FIXME break ? */
2951
2952                 ret = btrfs_previous_item(chunk_root, path, 0,
2953                                           BTRFS_CHUNK_ITEM_KEY);
2954                 if (ret) {
2955                         ret = 0;
2956                         break;
2957                 }
2958
2959                 leaf = path->nodes[0];
2960                 slot = path->slots[0];
2961                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
2962
2963                 if (found_key.objectid != key.objectid)
2964                         break;
2965
2966                 /* chunk zero is special */
2967                 if (found_key.offset == 0)
2968                         break;
2969
2970                 chunk = btrfs_item_ptr(leaf, slot, struct btrfs_chunk);
2971
2972                 if (!counting) {
2973                         spin_lock(&fs_info->balance_lock);
2974                         bctl->stat.considered++;
2975                         spin_unlock(&fs_info->balance_lock);
2976                 }
2977
2978                 ret = should_balance_chunk(chunk_root, leaf, chunk,
2979                                            found_key.offset);
2980                 btrfs_release_path(path);
2981                 if (!ret)
2982                         goto loop;
2983
2984                 if (counting) {
2985                         spin_lock(&fs_info->balance_lock);
2986                         bctl->stat.expected++;
2987                         spin_unlock(&fs_info->balance_lock);
2988                         goto loop;
2989                 }
2990
2991                 ret = btrfs_relocate_chunk(chunk_root,
2992                                            chunk_root->root_key.objectid,
2993                                            found_key.objectid,
2994                                            found_key.offset);
2995                 if (ret && ret != -ENOSPC)
2996                         goto error;
2997                 if (ret == -ENOSPC) {
2998                         enospc_errors++;
2999                 } else {
3000                         spin_lock(&fs_info->balance_lock);
3001                         bctl->stat.completed++;
3002                         spin_unlock(&fs_info->balance_lock);
3003                 }
3004 loop:
3005                 key.offset = found_key.offset - 1;
3006         }
3007
3008         if (counting) {
3009                 btrfs_release_path(path);
3010                 counting = false;
3011                 goto again;
3012         }
3013 error:
3014         btrfs_free_path(path);
3015         if (enospc_errors) {
3016                 printk(KERN_INFO "btrfs: %d enospc errors during balance\n",
3017                        enospc_errors);
3018                 if (!ret)
3019                         ret = -ENOSPC;
3020         }
3021
3022         return ret;
3023 }
3024
3025 /**
3026  * alloc_profile_is_valid - see if a given profile is valid and reduced
3027  * @flags: profile to validate
3028  * @extended: if true @flags is treated as an extended profile
3029  */
3030 static int alloc_profile_is_valid(u64 flags, int extended)
3031 {
3032         u64 mask = (extended ? BTRFS_EXTENDED_PROFILE_MASK :
3033                                BTRFS_BLOCK_GROUP_PROFILE_MASK);
3034
3035         flags &= ~BTRFS_BLOCK_GROUP_TYPE_MASK;
3036
3037         /* 1) check that all other bits are zeroed */
3038         if (flags & ~mask)
3039                 return 0;
3040
3041         /* 2) see if profile is reduced */
3042         if (flags == 0)
3043                 return !extended; /* "0" is valid for usual profiles */
3044
3045         /* true if exactly one bit set */
3046         return (flags & (flags - 1)) == 0;
3047 }
3048
3049 static inline int balance_need_close(struct btrfs_fs_info *fs_info)
3050 {
3051         /* cancel requested || normal exit path */
3052         return atomic_read(&fs_info->balance_cancel_req) ||
3053                 (atomic_read(&fs_info->balance_pause_req) == 0 &&
3054                  atomic_read(&fs_info->balance_cancel_req) == 0);
3055 }
3056
3057 static void __cancel_balance(struct btrfs_fs_info *fs_info)
3058 {
3059         int ret;
3060
3061         unset_balance_control(fs_info);
3062         ret = del_balance_item(fs_info->tree_root);
3063         if (ret)
3064                 btrfs_std_error(fs_info, ret);
3065
3066         atomic_set(&fs_info->mutually_exclusive_operation_running, 0);
3067 }
3068
3069 void update_ioctl_balance_args(struct btrfs_fs_info *fs_info, int lock,
3070                                struct btrfs_ioctl_balance_args *bargs);
3071
3072 /*
3073  * Should be called with both balance and volume mutexes held
3074  */
3075 int btrfs_balance(struct btrfs_balance_control *bctl,
3076                   struct btrfs_ioctl_balance_args *bargs)
3077 {
3078         struct btrfs_fs_info *fs_info = bctl->fs_info;
3079         u64 allowed;
3080         int mixed = 0;
3081         int ret;
3082         u64 num_devices;
3083         unsigned seq;
3084
3085         if (btrfs_fs_closing(fs_info) ||
3086             atomic_read(&fs_info->balance_pause_req) ||
3087             atomic_read(&fs_info->balance_cancel_req)) {
3088                 ret = -EINVAL;
3089                 goto out;
3090         }
3091
3092         allowed = btrfs_super_incompat_flags(fs_info->super_copy);
3093         if (allowed & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
3094                 mixed = 1;
3095
3096         /*
3097          * In case of mixed groups both data and meta should be picked,
3098          * and identical options should be given for both of them.
3099          */
3100         allowed = BTRFS_BALANCE_DATA | BTRFS_BALANCE_METADATA;
3101         if (mixed && (bctl->flags & allowed)) {
3102                 if (!(bctl->flags & BTRFS_BALANCE_DATA) ||
3103                     !(bctl->flags & BTRFS_BALANCE_METADATA) ||
3104                     memcmp(&bctl->data, &bctl->meta, sizeof(bctl->data))) {
3105                         printk(KERN_ERR "btrfs: with mixed groups data and "
3106                                "metadata balance options must be the same\n");
3107                         ret = -EINVAL;
3108                         goto out;
3109                 }
3110         }
3111
3112         num_devices = fs_info->fs_devices->num_devices;
3113         btrfs_dev_replace_lock(&fs_info->dev_replace);
3114         if (btrfs_dev_replace_is_ongoing(&fs_info->dev_replace)) {
3115                 BUG_ON(num_devices < 1);
3116                 num_devices--;
3117         }
3118         btrfs_dev_replace_unlock(&fs_info->dev_replace);
3119         allowed = BTRFS_AVAIL_ALLOC_BIT_SINGLE;
3120         if (num_devices == 1)
3121                 allowed |= BTRFS_BLOCK_GROUP_DUP;
3122         else if (num_devices < 4)
3123                 allowed |= (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID1);
3124         else
3125                 allowed |= (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID1 |
3126                                 BTRFS_BLOCK_GROUP_RAID10 |
3127                                 BTRFS_BLOCK_GROUP_RAID5 |
3128                                 BTRFS_BLOCK_GROUP_RAID6);
3129
3130         if ((bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) &&
3131             (!alloc_profile_is_valid(bctl->data.target, 1) ||
3132              (bctl->data.target & ~allowed))) {
3133                 printk(KERN_ERR "btrfs: unable to start balance with target "
3134                        "data profile %llu\n",
3135                        (unsigned long long)bctl->data.target);
3136                 ret = -EINVAL;
3137                 goto out;
3138         }
3139         if ((bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) &&
3140             (!alloc_profile_is_valid(bctl->meta.target, 1) ||
3141              (bctl->meta.target & ~allowed))) {
3142                 printk(KERN_ERR "btrfs: unable to start balance with target "
3143                        "metadata profile %llu\n",
3144                        (unsigned long long)bctl->meta.target);
3145                 ret = -EINVAL;
3146                 goto out;
3147         }
3148         if ((bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) &&
3149             (!alloc_profile_is_valid(bctl->sys.target, 1) ||
3150              (bctl->sys.target & ~allowed))) {
3151                 printk(KERN_ERR "btrfs: unable to start balance with target "
3152                        "system profile %llu\n",
3153                        (unsigned long long)bctl->sys.target);
3154                 ret = -EINVAL;
3155                 goto out;
3156         }
3157
3158         /* allow dup'ed data chunks only in mixed mode */
3159         if (!mixed && (bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) &&
3160             (bctl->data.target & BTRFS_BLOCK_GROUP_DUP)) {
3161                 printk(KERN_ERR "btrfs: dup for data is not allowed\n");
3162                 ret = -EINVAL;
3163                 goto out;
3164         }
3165
3166         /* allow to reduce meta or sys integrity only if force set */
3167         allowed = BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3168                         BTRFS_BLOCK_GROUP_RAID10 |
3169                         BTRFS_BLOCK_GROUP_RAID5 |
3170                         BTRFS_BLOCK_GROUP_RAID6;
3171         do {
3172                 seq = read_seqbegin(&fs_info->profiles_lock);
3173
3174                 if (((bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) &&
3175                      (fs_info->avail_system_alloc_bits & allowed) &&
3176                      !(bctl->sys.target & allowed)) ||
3177                     ((bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) &&
3178                      (fs_info->avail_metadata_alloc_bits & allowed) &&
3179                      !(bctl->meta.target & allowed))) {
3180                         if (bctl->flags & BTRFS_BALANCE_FORCE) {
3181                                 printk(KERN_INFO "btrfs: force reducing metadata "
3182                                        "integrity\n");
3183                         } else {
3184                                 printk(KERN_ERR "btrfs: balance will reduce metadata "
3185                                        "integrity, use force if you want this\n");
3186                                 ret = -EINVAL;
3187                                 goto out;
3188                         }
3189                 }
3190         } while (read_seqretry(&fs_info->profiles_lock, seq));
3191
3192         if (bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3193                 int num_tolerated_disk_barrier_failures;
3194                 u64 target = bctl->sys.target;
3195
3196                 num_tolerated_disk_barrier_failures =
3197                         btrfs_calc_num_tolerated_disk_barrier_failures(fs_info);
3198                 if (num_tolerated_disk_barrier_failures > 0 &&
3199                     (target &
3200                      (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3201                       BTRFS_AVAIL_ALLOC_BIT_SINGLE)))
3202                         num_tolerated_disk_barrier_failures = 0;
3203                 else if (num_tolerated_disk_barrier_failures > 1 &&
3204                          (target &
3205                           (BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10)))
3206                         num_tolerated_disk_barrier_failures = 1;
3207
3208                 fs_info->num_tolerated_disk_barrier_failures =
3209                         num_tolerated_disk_barrier_failures;
3210         }
3211
3212         ret = insert_balance_item(fs_info->tree_root, bctl);
3213         if (ret && ret != -EEXIST)
3214                 goto out;
3215
3216         if (!(bctl->flags & BTRFS_BALANCE_RESUME)) {
3217                 BUG_ON(ret == -EEXIST);
3218                 set_balance_control(bctl);
3219         } else {
3220                 BUG_ON(ret != -EEXIST);
3221                 spin_lock(&fs_info->balance_lock);
3222                 update_balance_args(bctl);
3223                 spin_unlock(&fs_info->balance_lock);
3224         }
3225
3226         atomic_inc(&fs_info->balance_running);
3227         mutex_unlock(&fs_info->balance_mutex);
3228
3229         ret = __btrfs_balance(fs_info);
3230
3231         mutex_lock(&fs_info->balance_mutex);
3232         atomic_dec(&fs_info->balance_running);
3233
3234         if (bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3235                 fs_info->num_tolerated_disk_barrier_failures =
3236                         btrfs_calc_num_tolerated_disk_barrier_failures(fs_info);
3237         }
3238
3239         if (bargs) {
3240                 memset(bargs, 0, sizeof(*bargs));
3241                 update_ioctl_balance_args(fs_info, 0, bargs);
3242         }
3243
3244         if ((ret && ret != -ECANCELED && ret != -ENOSPC) ||
3245             balance_need_close(fs_info)) {
3246                 __cancel_balance(fs_info);
3247         }
3248
3249         wake_up(&fs_info->balance_wait_q);
3250
3251         return ret;
3252 out:
3253         if (bctl->flags & BTRFS_BALANCE_RESUME)
3254                 __cancel_balance(fs_info);
3255         else {
3256                 kfree(bctl);
3257                 atomic_set(&fs_info->mutually_exclusive_operation_running, 0);
3258         }
3259         return ret;
3260 }
3261
3262 static int balance_kthread(void *data)
3263 {
3264         struct btrfs_fs_info *fs_info = data;
3265         int ret = 0;
3266
3267         mutex_lock(&fs_info->volume_mutex);
3268         mutex_lock(&fs_info->balance_mutex);
3269
3270         if (fs_info->balance_ctl) {
3271                 printk(KERN_INFO "btrfs: continuing balance\n");
3272                 ret = btrfs_balance(fs_info->balance_ctl, NULL);
3273         }
3274
3275         mutex_unlock(&fs_info->balance_mutex);
3276         mutex_unlock(&fs_info->volume_mutex);
3277
3278         return ret;
3279 }
3280
3281 int btrfs_resume_balance_async(struct btrfs_fs_info *fs_info)
3282 {
3283         struct task_struct *tsk;
3284
3285         spin_lock(&fs_info->balance_lock);
3286         if (!fs_info->balance_ctl) {
3287                 spin_unlock(&fs_info->balance_lock);
3288                 return 0;
3289         }
3290         spin_unlock(&fs_info->balance_lock);
3291
3292         if (btrfs_test_opt(fs_info->tree_root, SKIP_BALANCE)) {
3293                 printk(KERN_INFO "btrfs: force skipping balance\n");
3294                 return 0;
3295         }
3296
3297         tsk = kthread_run(balance_kthread, fs_info, "btrfs-balance");
3298         if (IS_ERR(tsk))
3299                 return PTR_ERR(tsk);
3300
3301         return 0;
3302 }
3303
3304 int btrfs_recover_balance(struct btrfs_fs_info *fs_info)
3305 {
3306         struct btrfs_balance_control *bctl;
3307         struct btrfs_balance_item *item;
3308         struct btrfs_disk_balance_args disk_bargs;
3309         struct btrfs_path *path;
3310         struct extent_buffer *leaf;
3311         struct btrfs_key key;
3312         int ret;
3313
3314         path = btrfs_alloc_path();
3315         if (!path)
3316                 return -ENOMEM;
3317
3318         key.objectid = BTRFS_BALANCE_OBJECTID;
3319         key.type = BTRFS_BALANCE_ITEM_KEY;
3320         key.offset = 0;
3321
3322         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
3323         if (ret < 0)
3324                 goto out;
3325         if (ret > 0) { /* ret = -ENOENT; */
3326                 ret = 0;
3327                 goto out;
3328         }
3329
3330         bctl = kzalloc(sizeof(*bctl), GFP_NOFS);
3331         if (!bctl) {
3332                 ret = -ENOMEM;
3333                 goto out;
3334         }
3335
3336         leaf = path->nodes[0];
3337         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_balance_item);
3338
3339         bctl->fs_info = fs_info;
3340         bctl->flags = btrfs_balance_flags(leaf, item);
3341         bctl->flags |= BTRFS_BALANCE_RESUME;
3342
3343         btrfs_balance_data(leaf, item, &disk_bargs);
3344         btrfs_disk_balance_args_to_cpu(&bctl->data, &disk_bargs);
3345         btrfs_balance_meta(leaf, item, &disk_bargs);
3346         btrfs_disk_balance_args_to_cpu(&bctl->meta, &disk_bargs);
3347         btrfs_balance_sys(leaf, item, &disk_bargs);
3348         btrfs_disk_balance_args_to_cpu(&bctl->sys, &disk_bargs);
3349
3350         WARN_ON(atomic_xchg(&fs_info->mutually_exclusive_operation_running, 1));
3351
3352         mutex_lock(&fs_info->volume_mutex);
3353         mutex_lock(&fs_info->balance_mutex);
3354
3355         set_balance_control(bctl);
3356
3357         mutex_unlock(&fs_info->balance_mutex);
3358         mutex_unlock(&fs_info->volume_mutex);
3359 out:
3360         btrfs_free_path(path);
3361         return ret;
3362 }
3363
3364 int btrfs_pause_balance(struct btrfs_fs_info *fs_info)
3365 {
3366         int ret = 0;
3367
3368         mutex_lock(&fs_info->balance_mutex);
3369         if (!fs_info->balance_ctl) {
3370                 mutex_unlock(&fs_info->balance_mutex);
3371                 return -ENOTCONN;
3372         }
3373
3374         if (atomic_read(&fs_info->balance_running)) {
3375                 atomic_inc(&fs_info->balance_pause_req);
3376                 mutex_unlock(&fs_info->balance_mutex);
3377
3378                 wait_event(fs_info->balance_wait_q,
3379                            atomic_read(&fs_info->balance_running) == 0);
3380
3381                 mutex_lock(&fs_info->balance_mutex);
3382                 /* we are good with balance_ctl ripped off from under us */
3383                 BUG_ON(atomic_read(&fs_info->balance_running));
3384                 atomic_dec(&fs_info->balance_pause_req);
3385         } else {
3386                 ret = -ENOTCONN;
3387         }
3388
3389         mutex_unlock(&fs_info->balance_mutex);
3390         return ret;
3391 }
3392
3393 int btrfs_cancel_balance(struct btrfs_fs_info *fs_info)
3394 {
3395         mutex_lock(&fs_info->balance_mutex);
3396         if (!fs_info->balance_ctl) {
3397                 mutex_unlock(&fs_info->balance_mutex);
3398                 return -ENOTCONN;
3399         }
3400
3401         atomic_inc(&fs_info->balance_cancel_req);
3402         /*
3403          * if we are running just wait and return, balance item is
3404          * deleted in btrfs_balance in this case
3405          */
3406         if (atomic_read(&fs_info->balance_running)) {
3407                 mutex_unlock(&fs_info->balance_mutex);
3408                 wait_event(fs_info->balance_wait_q,
3409                            atomic_read(&fs_info->balance_running) == 0);
3410                 mutex_lock(&fs_info->balance_mutex);
3411         } else {
3412                 /* __cancel_balance needs volume_mutex */
3413                 mutex_unlock(&fs_info->balance_mutex);
3414                 mutex_lock(&fs_info->volume_mutex);
3415                 mutex_lock(&fs_info->balance_mutex);
3416
3417                 if (fs_info->balance_ctl)
3418                         __cancel_balance(fs_info);
3419
3420                 mutex_unlock(&fs_info->volume_mutex);
3421         }
3422
3423         BUG_ON(fs_info->balance_ctl || atomic_read(&fs_info->balance_running));
3424         atomic_dec(&fs_info->balance_cancel_req);
3425         mutex_unlock(&fs_info->balance_mutex);
3426         return 0;
3427 }
3428
3429 /*
3430  * shrinking a device means finding all of the device extents past
3431  * the new size, and then following the back refs to the chunks.
3432  * The chunk relocation code actually frees the device extent
3433  */
3434 int btrfs_shrink_device(struct btrfs_device *device, u64 new_size)
3435 {
3436         struct btrfs_trans_handle *trans;
3437         struct btrfs_root *root = device->dev_root;
3438         struct btrfs_dev_extent *dev_extent = NULL;
3439         struct btrfs_path *path;
3440         u64 length;
3441         u64 chunk_tree;
3442         u64 chunk_objectid;
3443         u64 chunk_offset;
3444         int ret;
3445         int slot;
3446         int failed = 0;
3447         bool retried = false;
3448         struct extent_buffer *l;
3449         struct btrfs_key key;
3450         struct btrfs_super_block *super_copy = root->fs_info->super_copy;
3451         u64 old_total = btrfs_super_total_bytes(super_copy);
3452         u64 old_size = device->total_bytes;
3453         u64 diff = device->total_bytes - new_size;
3454
3455         if (device->is_tgtdev_for_dev_replace)
3456                 return -EINVAL;
3457
3458         path = btrfs_alloc_path();
3459         if (!path)
3460                 return -ENOMEM;
3461
3462         path->reada = 2;
3463
3464         lock_chunks(root);
3465
3466         device->total_bytes = new_size;
3467         if (device->writeable) {
3468                 device->fs_devices->total_rw_bytes -= diff;
3469                 spin_lock(&root->fs_info->free_chunk_lock);
3470                 root->fs_info->free_chunk_space -= diff;
3471                 spin_unlock(&root->fs_info->free_chunk_lock);
3472         }
3473         unlock_chunks(root);
3474
3475 again:
3476         key.objectid = device->devid;
3477         key.offset = (u64)-1;
3478         key.type = BTRFS_DEV_EXTENT_KEY;
3479
3480         do {
3481                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3482                 if (ret < 0)
3483                         goto done;
3484
3485                 ret = btrfs_previous_item(root, path, 0, key.type);
3486                 if (ret < 0)
3487                         goto done;
3488                 if (ret) {
3489                         ret = 0;
3490                         btrfs_release_path(path);
3491                         break;
3492                 }
3493
3494                 l = path->nodes[0];
3495                 slot = path->slots[0];
3496                 btrfs_item_key_to_cpu(l, &key, path->slots[0]);
3497
3498                 if (key.objectid != device->devid) {
3499                         btrfs_release_path(path);
3500                         break;
3501                 }
3502
3503                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3504                 length = btrfs_dev_extent_length(l, dev_extent);
3505
3506                 if (key.offset + length <= new_size) {
3507                         btrfs_release_path(path);
3508                         break;
3509                 }
3510
3511                 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
3512                 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
3513                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3514                 btrfs_release_path(path);
3515
3516                 ret = btrfs_relocate_chunk(root, chunk_tree, chunk_objectid,
3517                                            chunk_offset);
3518                 if (ret && ret != -ENOSPC)
3519                         goto done;
3520                 if (ret == -ENOSPC)
3521                         failed++;
3522         } while (key.offset-- > 0);
3523
3524         if (failed && !retried) {
3525                 failed = 0;
3526                 retried = true;
3527                 goto again;
3528         } else if (failed && retried) {
3529                 ret = -ENOSPC;
3530                 lock_chunks(root);
3531
3532                 device->total_bytes = old_size;
3533                 if (device->writeable)
3534                         device->fs_devices->total_rw_bytes += diff;
3535                 spin_lock(&root->fs_info->free_chunk_lock);
3536                 root->fs_info->free_chunk_space += diff;
3537                 spin_unlock(&root->fs_info->free_chunk_lock);
3538                 unlock_chunks(root);
3539                 goto done;
3540         }
3541
3542         /* Shrinking succeeded, else we would be at "done". */
3543         trans = btrfs_start_transaction(root, 0);
3544         if (IS_ERR(trans)) {
3545                 ret = PTR_ERR(trans);
3546                 goto done;
3547         }
3548
3549         lock_chunks(root);
3550
3551         device->disk_total_bytes = new_size;
3552         /* Now btrfs_update_device() will change the on-disk size. */
3553         ret = btrfs_update_device(trans, device);
3554         if (ret) {
3555                 unlock_chunks(root);
3556                 btrfs_end_transaction(trans, root);
3557                 goto done;
3558         }
3559         WARN_ON(diff > old_total);
3560         btrfs_set_super_total_bytes(super_copy, old_total - diff);
3561         unlock_chunks(root);
3562         btrfs_end_transaction(trans, root);
3563 done:
3564         btrfs_free_path(path);
3565         return ret;
3566 }
3567
3568 static int btrfs_add_system_chunk(struct btrfs_root *root,
3569                            struct btrfs_key *key,
3570                            struct btrfs_chunk *chunk, int item_size)
3571 {
3572         struct btrfs_super_block *super_copy = root->fs_info->super_copy;
3573         struct btrfs_disk_key disk_key;
3574         u32 array_size;
3575         u8 *ptr;
3576
3577         array_size = btrfs_super_sys_array_size(super_copy);
3578         if (array_size + item_size > BTRFS_SYSTEM_CHUNK_ARRAY_SIZE)
3579                 return -EFBIG;
3580
3581         ptr = super_copy->sys_chunk_array + array_size;
3582         btrfs_cpu_key_to_disk(&disk_key, key);
3583         memcpy(ptr, &disk_key, sizeof(disk_key));
3584         ptr += sizeof(disk_key);
3585         memcpy(ptr, chunk, item_size);
3586         item_size += sizeof(disk_key);
3587         btrfs_set_super_sys_array_size(super_copy, array_size + item_size);
3588         return 0;
3589 }
3590
3591 /*
3592  * sort the devices in descending order by max_avail, total_avail
3593  */
3594 static int btrfs_cmp_device_info(const void *a, const void *b)
3595 {
3596         const struct btrfs_device_info *di_a = a;
3597         const struct btrfs_device_info *di_b = b;
3598
3599         if (di_a->max_avail > di_b->max_avail)
3600                 return -1;
3601         if (di_a->max_avail < di_b->max_avail)
3602                 return 1;
3603         if (di_a->total_avail > di_b->total_avail)
3604                 return -1;
3605         if (di_a->total_avail < di_b->total_avail)
3606                 return 1;
3607         return 0;
3608 }
3609
3610 struct btrfs_raid_attr btrfs_raid_array[BTRFS_NR_RAID_TYPES] = {
3611         [BTRFS_RAID_RAID10] = {
3612                 .sub_stripes    = 2,
3613                 .dev_stripes    = 1,
3614                 .devs_max       = 0,    /* 0 == as many as possible */
3615                 .devs_min       = 4,
3616                 .devs_increment = 2,
3617                 .ncopies        = 2,
3618         },
3619         [BTRFS_RAID_RAID1] = {
3620                 .sub_stripes    = 1,
3621                 .dev_stripes    = 1,
3622                 .devs_max       = 2,
3623                 .devs_min       = 2,
3624                 .devs_increment = 2,
3625                 .ncopies        = 2,
3626         },
3627         [BTRFS_RAID_DUP] = {
3628                 .sub_stripes    = 1,
3629                 .dev_stripes    = 2,
3630                 .devs_max       = 1,
3631                 .devs_min       = 1,
3632                 .devs_increment = 1,
3633                 .ncopies        = 2,
3634         },
3635         [BTRFS_RAID_RAID0] = {
3636                 .sub_stripes    = 1,
3637                 .dev_stripes    = 1,
3638                 .devs_max       = 0,
3639                 .devs_min       = 2,
3640                 .devs_increment = 1,
3641                 .ncopies        = 1,
3642         },
3643         [BTRFS_RAID_SINGLE] = {
3644                 .sub_stripes    = 1,
3645                 .dev_stripes    = 1,
3646                 .devs_max       = 1,
3647                 .devs_min       = 1,
3648                 .devs_increment = 1,
3649                 .ncopies        = 1,
3650         },
3651         [BTRFS_RAID_RAID5] = {
3652                 .sub_stripes    = 1,
3653                 .dev_stripes    = 1,
3654                 .devs_max       = 0,
3655                 .devs_min       = 2,
3656                 .devs_increment = 1,
3657                 .ncopies        = 2,
3658         },
3659         [BTRFS_RAID_RAID6] = {
3660                 .sub_stripes    = 1,
3661                 .dev_stripes    = 1,
3662                 .devs_max       = 0,
3663                 .devs_min       = 3,
3664                 .devs_increment = 1,
3665                 .ncopies        = 3,
3666         },
3667 };
3668
3669 static u32 find_raid56_stripe_len(u32 data_devices, u32 dev_stripe_target)
3670 {
3671         /* TODO allow them to set a preferred stripe size */
3672         return 64 * 1024;
3673 }
3674
3675 static void check_raid56_incompat_flag(struct btrfs_fs_info *info, u64 type)
3676 {
3677         if (!(type & (BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6)))
3678                 return;
3679
3680         btrfs_set_fs_incompat(info, RAID56);
3681 }
3682
3683 static int __btrfs_alloc_chunk(struct btrfs_trans_handle *trans,
3684                                struct btrfs_root *extent_root,
3685                                struct map_lookup **map_ret,
3686                                u64 *num_bytes_out, u64 *stripe_size_out,
3687                                u64 start, u64 type)
3688 {
3689         struct btrfs_fs_info *info = extent_root->fs_info;
3690         struct btrfs_fs_devices *fs_devices = info->fs_devices;
3691         struct list_head *cur;
3692         struct map_lookup *map = NULL;
3693         struct extent_map_tree *em_tree;
3694         struct extent_map *em;
3695         struct btrfs_device_info *devices_info = NULL;
3696         u64 total_avail;
3697         int num_stripes;        /* total number of stripes to allocate */
3698         int data_stripes;       /* number of stripes that count for
3699                                    block group size */
3700         int sub_stripes;        /* sub_stripes info for map */
3701         int dev_stripes;        /* stripes per dev */
3702         int devs_max;           /* max devs to use */
3703         int devs_min;           /* min devs needed */
3704         int devs_increment;     /* ndevs has to be a multiple of this */
3705         int ncopies;            /* how many copies to data has */
3706         int ret;
3707         u64 max_stripe_size;
3708         u64 max_chunk_size;
3709         u64 stripe_size;
3710         u64 num_bytes;
3711         u64 raid_stripe_len = BTRFS_STRIPE_LEN;
3712         int ndevs;
3713         int i;
3714         int j;
3715         int index;
3716
3717         BUG_ON(!alloc_profile_is_valid(type, 0));
3718
3719         if (list_empty(&fs_devices->alloc_list))
3720                 return -ENOSPC;
3721
3722         index = __get_raid_index(type);
3723
3724         sub_stripes = btrfs_raid_array[index].sub_stripes;
3725         dev_stripes = btrfs_raid_array[index].dev_stripes;
3726         devs_max = btrfs_raid_array[index].devs_max;
3727         devs_min = btrfs_raid_array[index].devs_min;
3728         devs_increment = btrfs_raid_array[index].devs_increment;
3729         ncopies = btrfs_raid_array[index].ncopies;
3730
3731         if (type & BTRFS_BLOCK_GROUP_DATA) {
3732                 max_stripe_size = 1024 * 1024 * 1024;
3733                 max_chunk_size = 10 * max_stripe_size;
3734         } else if (type & BTRFS_BLOCK_GROUP_METADATA) {
3735                 /* for larger filesystems, use larger metadata chunks */
3736                 if (fs_devices->total_rw_bytes > 50ULL * 1024 * 1024 * 1024)
3737                         max_stripe_size = 1024 * 1024 * 1024;
3738                 else
3739                         max_stripe_size = 256 * 1024 * 1024;
3740                 max_chunk_size = max_stripe_size;
3741         } else if (type & BTRFS_BLOCK_GROUP_SYSTEM) {
3742                 max_stripe_size = 32 * 1024 * 1024;
3743                 max_chunk_size = 2 * max_stripe_size;
3744         } else {
3745                 printk(KERN_ERR "btrfs: invalid chunk type 0x%llx requested\n",
3746                        type);
3747                 BUG_ON(1);
3748         }
3749
3750         /* we don't want a chunk larger than 10% of writeable space */
3751         max_chunk_size = min(div_factor(fs_devices->total_rw_bytes, 1),
3752                              max_chunk_size);
3753
3754         devices_info = kzalloc(sizeof(*devices_info) * fs_devices->rw_devices,
3755                                GFP_NOFS);
3756         if (!devices_info)
3757                 return -ENOMEM;
3758
3759         cur = fs_devices->alloc_list.next;
3760
3761         /*
3762          * in the first pass through the devices list, we gather information
3763          * about the available holes on each device.
3764          */
3765         ndevs = 0;
3766         while (cur != &fs_devices->alloc_list) {
3767                 struct btrfs_device *device;
3768                 u64 max_avail;
3769                 u64 dev_offset;
3770
3771                 device = list_entry(cur, struct btrfs_device, dev_alloc_list);
3772
3773                 cur = cur->next;
3774
3775                 if (!device->writeable) {
3776                         WARN(1, KERN_ERR
3777                                "btrfs: read-only device in alloc_list\n");
3778                         continue;
3779                 }
3780
3781                 if (!device->in_fs_metadata ||
3782                     device->is_tgtdev_for_dev_replace)
3783                         continue;
3784
3785                 if (device->total_bytes > device->bytes_used)
3786                         total_avail = device->total_bytes - device->bytes_used;
3787                 else
3788                         total_avail = 0;
3789
3790                 /* If there is no space on this device, skip it. */
3791                 if (total_avail == 0)
3792                         continue;
3793
3794                 ret = find_free_dev_extent(device,
3795                                            max_stripe_size * dev_stripes,
3796                                            &dev_offset, &max_avail);
3797                 if (ret && ret != -ENOSPC)
3798                         goto error;
3799
3800                 if (ret == 0)
3801                         max_avail = max_stripe_size * dev_stripes;
3802
3803                 if (max_avail < BTRFS_STRIPE_LEN * dev_stripes)
3804                         continue;
3805
3806                 if (ndevs == fs_devices->rw_devices) {
3807                         WARN(1, "%s: found more than %llu devices\n",
3808                              __func__, fs_devices->rw_devices);
3809                         break;
3810                 }
3811                 devices_info[ndevs].dev_offset = dev_offset;
3812                 devices_info[ndevs].max_avail = max_avail;
3813                 devices_info[ndevs].total_avail = total_avail;
3814                 devices_info[ndevs].dev = device;
3815                 ++ndevs;
3816         }
3817
3818         /*
3819          * now sort the devices by hole size / available space
3820          */
3821         sort(devices_info, ndevs, sizeof(struct btrfs_device_info),
3822              btrfs_cmp_device_info, NULL);
3823
3824         /* round down to number of usable stripes */
3825         ndevs -= ndevs % devs_increment;
3826
3827         if (ndevs < devs_increment * sub_stripes || ndevs < devs_min) {
3828                 ret = -ENOSPC;
3829                 goto error;
3830         }
3831
3832         if (devs_max && ndevs > devs_max)
3833                 ndevs = devs_max;
3834         /*
3835          * the primary goal is to maximize the number of stripes, so use as many
3836          * devices as possible, even if the stripes are not maximum sized.
3837          */
3838         stripe_size = devices_info[ndevs-1].max_avail;
3839         num_stripes = ndevs * dev_stripes;
3840
3841         /*
3842          * this will have to be fixed for RAID1 and RAID10 over
3843          * more drives
3844          */
3845         data_stripes = num_stripes / ncopies;
3846
3847         if (type & BTRFS_BLOCK_GROUP_RAID5) {
3848                 raid_stripe_len = find_raid56_stripe_len(ndevs - 1,
3849                                  btrfs_super_stripesize(info->super_copy));
3850                 data_stripes = num_stripes - 1;
3851         }
3852         if (type & BTRFS_BLOCK_GROUP_RAID6) {
3853                 raid_stripe_len = find_raid56_stripe_len(ndevs - 2,
3854                                  btrfs_super_stripesize(info->super_copy));
3855                 data_stripes = num_stripes - 2;
3856         }
3857
3858         /*
3859          * Use the number of data stripes to figure out how big this chunk
3860          * is really going to be in terms of logical address space,
3861          * and compare that answer with the max chunk size
3862          */
3863         if (stripe_size * data_stripes > max_chunk_size) {
3864                 u64 mask = (1ULL << 24) - 1;
3865                 stripe_size = max_chunk_size;
3866                 do_div(stripe_size, data_stripes);
3867
3868                 /* bump the answer up to a 16MB boundary */
3869                 stripe_size = (stripe_size + mask) & ~mask;
3870
3871                 /* but don't go higher than the limits we found
3872                  * while searching for free extents
3873                  */
3874                 if (stripe_size > devices_info[ndevs-1].max_avail)
3875                         stripe_size = devices_info[ndevs-1].max_avail;
3876         }
3877
3878         do_div(stripe_size, dev_stripes);
3879
3880         /* align to BTRFS_STRIPE_LEN */
3881         do_div(stripe_size, raid_stripe_len);
3882         stripe_size *= raid_stripe_len;
3883
3884         map = kmalloc(map_lookup_size(num_stripes), GFP_NOFS);
3885         if (!map) {
3886                 ret = -ENOMEM;
3887                 goto error;
3888         }
3889         map->num_stripes = num_stripes;
3890
3891         for (i = 0; i < ndevs; ++i) {
3892                 for (j = 0; j < dev_stripes; ++j) {
3893                         int s = i * dev_stripes + j;
3894                         map->stripes[s].dev = devices_info[i].dev;
3895                         map->stripes[s].physical = devices_info[i].dev_offset +
3896                                                    j * stripe_size;
3897                 }
3898         }
3899         map->sector_size = extent_root->sectorsize;
3900         map->stripe_len = raid_stripe_len;
3901         map->io_align = raid_stripe_len;
3902         map->io_width = raid_stripe_len;
3903         map->type = type;
3904         map->sub_stripes = sub_stripes;
3905
3906         *map_ret = map;
3907         num_bytes = stripe_size * data_stripes;
3908
3909         *stripe_size_out = stripe_size;
3910         *num_bytes_out = num_bytes;
3911
3912         trace_btrfs_chunk_alloc(info->chunk_root, map, start, num_bytes);
3913
3914         em = alloc_extent_map();
3915         if (!em) {
3916                 ret = -ENOMEM;
3917                 goto error;
3918         }
3919         em->bdev = (struct block_device *)map;
3920         em->start = start;
3921         em->len = num_bytes;
3922         em->block_start = 0;
3923         em->block_len = em->len;
3924
3925         em_tree = &extent_root->fs_info->mapping_tree.map_tree;
3926         write_lock(&em_tree->lock);
3927         ret = add_extent_mapping(em_tree, em, 0);
3928         write_unlock(&em_tree->lock);
3929         if (ret) {
3930                 free_extent_map(em);
3931                 goto error;
3932         }
3933
3934         for (i = 0; i < map->num_stripes; ++i) {
3935                 struct btrfs_device *device;
3936                 u64 dev_offset;
3937
3938                 device = map->stripes[i].dev;
3939                 dev_offset = map->stripes[i].physical;
3940
3941                 ret = btrfs_alloc_dev_extent(trans, device,
3942                                 info->chunk_root->root_key.objectid,
3943                                 BTRFS_FIRST_CHUNK_TREE_OBJECTID,
3944                                 start, dev_offset, stripe_size);
3945                 if (ret)
3946                         goto error_dev_extent;
3947         }
3948
3949         ret = btrfs_make_block_group(trans, extent_root, 0, type,
3950                                      BTRFS_FIRST_CHUNK_TREE_OBJECTID,
3951                                      start, num_bytes);
3952         if (ret) {
3953                 i = map->num_stripes - 1;
3954                 goto error_dev_extent;
3955         }
3956
3957         free_extent_map(em);
3958         check_raid56_incompat_flag(extent_root->fs_info, type);
3959
3960         kfree(devices_info);
3961         return 0;
3962
3963 error_dev_extent:
3964         for (; i >= 0; i--) {
3965                 struct btrfs_device *device;
3966                 int err;
3967
3968                 device = map->stripes[i].dev;
3969                 err = btrfs_free_dev_extent(trans, device, start);
3970                 if (err) {
3971                         btrfs_abort_transaction(trans, extent_root, err);
3972                         break;
3973                 }
3974         }
3975         write_lock(&em_tree->lock);
3976         remove_extent_mapping(em_tree, em);
3977         write_unlock(&em_tree->lock);
3978
3979         /* One for our allocation */
3980         free_extent_map(em);
3981         /* One for the tree reference */
3982         free_extent_map(em);
3983 error:
3984         kfree(map);
3985         kfree(devices_info);
3986         return ret;
3987 }
3988
3989 static int __finish_chunk_alloc(struct btrfs_trans_handle *trans,
3990                                 struct btrfs_root *extent_root,
3991                                 struct map_lookup *map, u64 chunk_offset,
3992                                 u64 chunk_size, u64 stripe_size)
3993 {
3994         u64 dev_offset;
3995         struct btrfs_key key;
3996         struct btrfs_root *chunk_root = extent_root->fs_info->chunk_root;
3997         struct btrfs_device *device;
3998         struct btrfs_chunk *chunk;
3999         struct btrfs_stripe *stripe;
4000         size_t item_size = btrfs_chunk_item_size(map->num_stripes);
4001         int index = 0;
4002         int ret;
4003
4004         chunk = kzalloc(item_size, GFP_NOFS);
4005         if (!chunk)
4006                 return -ENOMEM;
4007
4008         index = 0;
4009         while (index < map->num_stripes) {
4010                 device = map->stripes[index].dev;
4011                 device->bytes_used += stripe_size;
4012                 ret = btrfs_update_device(trans, device);
4013                 if (ret)
4014                         goto out_free;
4015                 index++;
4016         }
4017
4018         spin_lock(&extent_root->fs_info->free_chunk_lock);
4019         extent_root->fs_info->free_chunk_space -= (stripe_size *
4020                                                    map->num_stripes);
4021         spin_unlock(&extent_root->fs_info->free_chunk_lock);
4022
4023         index = 0;
4024         stripe = &chunk->stripe;
4025         while (index < map->num_stripes) {
4026                 device = map->stripes[index].dev;
4027                 dev_offset = map->stripes[index].physical;
4028
4029                 btrfs_set_stack_stripe_devid(stripe, device->devid);
4030                 btrfs_set_stack_stripe_offset(stripe, dev_offset);
4031                 memcpy(stripe->dev_uuid, device->uuid, BTRFS_UUID_SIZE);
4032                 stripe++;
4033                 index++;
4034         }
4035
4036         btrfs_set_stack_chunk_length(chunk, chunk_size);
4037         btrfs_set_stack_chunk_owner(chunk, extent_root->root_key.objectid);
4038         btrfs_set_stack_chunk_stripe_len(chunk, map->stripe_len);
4039         btrfs_set_stack_chunk_type(chunk, map->type);
4040         btrfs_set_stack_chunk_num_stripes(chunk, map->num_stripes);
4041         btrfs_set_stack_chunk_io_align(chunk, map->stripe_len);
4042         btrfs_set_stack_chunk_io_width(chunk, map->stripe_len);
4043         btrfs_set_stack_chunk_sector_size(chunk, extent_root->sectorsize);
4044         btrfs_set_stack_chunk_sub_stripes(chunk, map->sub_stripes);
4045
4046         key.objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
4047         key.type = BTRFS_CHUNK_ITEM_KEY;
4048         key.offset = chunk_offset;
4049
4050         ret = btrfs_insert_item(trans, chunk_root, &key, chunk, item_size);
4051
4052         if (ret == 0 && map->type & BTRFS_BLOCK_GROUP_SYSTEM) {
4053                 /*
4054                  * TODO: Cleanup of inserted chunk root in case of
4055                  * failure.
4056                  */
4057                 ret = btrfs_add_system_chunk(chunk_root, &key, chunk,
4058                                              item_size);
4059         }
4060
4061 out_free:
4062         kfree(chunk);
4063         return ret;
4064 }
4065
4066 /*
4067  * Chunk allocation falls into two parts. The first part does works
4068  * that make the new allocated chunk useable, but not do any operation
4069  * that modifies the chunk tree. The second part does the works that
4070  * require modifying the chunk tree. This division is important for the
4071  * bootstrap process of adding storage to a seed btrfs.
4072  */
4073 int btrfs_alloc_chunk(struct btrfs_trans_handle *trans,
4074                       struct btrfs_root *extent_root, u64 type)
4075 {
4076         u64 chunk_offset;
4077         u64 chunk_size;
4078         u64 stripe_size;
4079         struct map_lookup *map;
4080         struct btrfs_root *chunk_root = extent_root->fs_info->chunk_root;
4081         int ret;
4082
4083         ret = find_next_chunk(chunk_root, BTRFS_FIRST_CHUNK_TREE_OBJECTID,
4084                               &chunk_offset);
4085         if (ret)
4086                 return ret;
4087
4088         ret = __btrfs_alloc_chunk(trans, extent_root, &map, &chunk_size,
4089                                   &stripe_size, chunk_offset, type);
4090         if (ret)
4091                 return ret;
4092
4093         ret = __finish_chunk_alloc(trans, extent_root, map, chunk_offset,
4094                                    chunk_size, stripe_size);
4095         if (ret)
4096                 return ret;
4097         return 0;
4098 }
4099
4100 static noinline int init_first_rw_device(struct btrfs_trans_handle *trans,
4101                                          struct btrfs_root *root,
4102                                          struct btrfs_device *device)
4103 {
4104         u64 chunk_offset;
4105         u64 sys_chunk_offset;
4106         u64 chunk_size;
4107         u64 sys_chunk_size;
4108         u64 stripe_size;
4109         u64 sys_stripe_size;
4110         u64 alloc_profile;
4111         struct map_lookup *map;
4112         struct map_lookup *sys_map;
4113         struct btrfs_fs_info *fs_info = root->fs_info;
4114         struct btrfs_root *extent_root = fs_info->extent_root;
4115         int ret;
4116
4117         ret = find_next_chunk(fs_info->chunk_root,
4118                               BTRFS_FIRST_CHUNK_TREE_OBJECTID, &chunk_offset);
4119         if (ret)
4120                 return ret;
4121
4122         alloc_profile = btrfs_get_alloc_profile(extent_root, 0);
4123         ret = __btrfs_alloc_chunk(trans, extent_root, &map, &chunk_size,
4124                                   &stripe_size, chunk_offset, alloc_profile);
4125         if (ret)
4126                 return ret;
4127
4128         sys_chunk_offset = chunk_offset + chunk_size;
4129
4130         alloc_profile = btrfs_get_alloc_profile(fs_info->chunk_root, 0);
4131         ret = __btrfs_alloc_chunk(trans, extent_root, &sys_map,
4132                                   &sys_chunk_size, &sys_stripe_size,
4133                                   sys_chunk_offset, alloc_profile);
4134         if (ret) {
4135                 btrfs_abort_transaction(trans, root, ret);
4136                 goto out;
4137         }
4138
4139         ret = btrfs_add_device(trans, fs_info->chunk_root, device);
4140         if (ret) {
4141                 btrfs_abort_transaction(trans, root, ret);
4142                 goto out;
4143         }
4144
4145         /*
4146          * Modifying chunk tree needs allocating new blocks from both
4147          * system block group and metadata block group. So we only can
4148          * do operations require modifying the chunk tree after both
4149          * block groups were created.
4150          */
4151         ret = __finish_chunk_alloc(trans, extent_root, map, chunk_offset,
4152                                    chunk_size, stripe_size);
4153         if (ret) {
4154                 btrfs_abort_transaction(trans, root, ret);
4155                 goto out;
4156         }
4157
4158         ret = __finish_chunk_alloc(trans, extent_root, sys_map,
4159                                    sys_chunk_offset, sys_chunk_size,
4160                                    sys_stripe_size);
4161         if (ret)
4162                 btrfs_abort_transaction(trans, root, ret);
4163
4164 out:
4165
4166         return ret;
4167 }
4168
4169 int btrfs_chunk_readonly(struct btrfs_root *root, u64 chunk_offset)
4170 {
4171         struct extent_map *em;
4172         struct map_lookup *map;
4173         struct btrfs_mapping_tree *map_tree = &root->fs_info->mapping_tree;
4174         int readonly = 0;
4175         int i;
4176
4177         read_lock(&map_tree->map_tree.lock);
4178         em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
4179         read_unlock(&map_tree->map_tree.lock);
4180         if (!em)
4181                 return 1;
4182
4183         if (btrfs_test_opt(root, DEGRADED)) {
4184                 free_extent_map(em);
4185                 return 0;
4186         }
4187
4188         map = (struct map_lookup *)em->bdev;
4189         for (i = 0; i < map->num_stripes; i++) {
4190                 if (!map->stripes[i].dev->writeable) {
4191                         readonly = 1;
4192                         break;
4193                 }
4194         }
4195         free_extent_map(em);
4196         return readonly;
4197 }
4198
4199 void btrfs_mapping_init(struct btrfs_mapping_tree *tree)
4200 {
4201         extent_map_tree_init(&tree->map_tree);
4202 }
4203
4204 void btrfs_mapping_tree_free(struct btrfs_mapping_tree *tree)
4205 {
4206         struct extent_map *em;
4207
4208         while (1) {
4209                 write_lock(&tree->map_tree.lock);
4210                 em = lookup_extent_mapping(&tree->map_tree, 0, (u64)-1);
4211                 if (em)
4212                         remove_extent_mapping(&tree->map_tree, em);
4213                 write_unlock(&tree->map_tree.lock);
4214                 if (!em)
4215                         break;
4216                 kfree(em->bdev);
4217                 /* once for us */
4218                 free_extent_map(em);
4219                 /* once for the tree */
4220                 free_extent_map(em);
4221         }
4222 }
4223
4224 int btrfs_num_copies(struct btrfs_fs_info *fs_info, u64 logical, u64 len)
4225 {
4226         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
4227         struct extent_map *em;
4228         struct map_lookup *map;
4229         struct extent_map_tree *em_tree = &map_tree->map_tree;
4230         int ret;
4231
4232         read_lock(&em_tree->lock);
4233         em = lookup_extent_mapping(em_tree, logical, len);
4234         read_unlock(&em_tree->lock);
4235         BUG_ON(!em);
4236
4237         BUG_ON(em->start > logical || em->start + em->len < logical);
4238         map = (struct map_lookup *)em->bdev;
4239         if (map->type & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1))
4240                 ret = map->num_stripes;
4241         else if (map->type & BTRFS_BLOCK_GROUP_RAID10)
4242                 ret = map->sub_stripes;
4243         else if (map->type & BTRFS_BLOCK_GROUP_RAID5)
4244                 ret = 2;
4245         else if (map->type & BTRFS_BLOCK_GROUP_RAID6)
4246                 ret = 3;
4247         else
4248                 ret = 1;
4249         free_extent_map(em);
4250
4251         btrfs_dev_replace_lock(&fs_info->dev_replace);
4252         if (btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))
4253                 ret++;
4254         btrfs_dev_replace_unlock(&fs_info->dev_replace);
4255
4256         return ret;
4257 }
4258
4259 unsigned long btrfs_full_stripe_len(struct btrfs_root *root,
4260                                     struct btrfs_mapping_tree *map_tree,
4261                                     u64 logical)
4262 {
4263         struct extent_map *em;
4264         struct map_lookup *map;
4265         struct extent_map_tree *em_tree = &map_tree->map_tree;
4266         unsigned long len = root->sectorsize;
4267
4268         read_lock(&em_tree->lock);
4269         em = lookup_extent_mapping(em_tree, logical, len);
4270         read_unlock(&em_tree->lock);
4271         BUG_ON(!em);
4272
4273         BUG_ON(em->start > logical || em->start + em->len < logical);
4274         map = (struct map_lookup *)em->bdev;
4275         if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
4276                          BTRFS_BLOCK_GROUP_RAID6)) {
4277                 len = map->stripe_len * nr_data_stripes(map);
4278         }
4279         free_extent_map(em);
4280         return len;
4281 }
4282
4283 int btrfs_is_parity_mirror(struct btrfs_mapping_tree *map_tree,
4284                            u64 logical, u64 len, int mirror_num)
4285 {
4286         struct extent_map *em;
4287         struct map_lookup *map;
4288         struct extent_map_tree *em_tree = &map_tree->map_tree;
4289         int ret = 0;
4290
4291         read_lock(&em_tree->lock);
4292         em = lookup_extent_mapping(em_tree, logical, len);
4293         read_unlock(&em_tree->lock);
4294         BUG_ON(!em);
4295
4296         BUG_ON(em->start > logical || em->start + em->len < logical);
4297         map = (struct map_lookup *)em->bdev;
4298         if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
4299                          BTRFS_BLOCK_GROUP_RAID6))
4300                 ret = 1;
4301         free_extent_map(em);
4302         return ret;
4303 }
4304
4305 static int find_live_mirror(struct btrfs_fs_info *fs_info,
4306                             struct map_lookup *map, int first, int num,
4307                             int optimal, int dev_replace_is_ongoing)
4308 {
4309         int i;
4310         int tolerance;
4311         struct btrfs_device *srcdev;
4312
4313         if (dev_replace_is_ongoing &&
4314             fs_info->dev_replace.cont_reading_from_srcdev_mode ==
4315              BTRFS_DEV_REPLACE_ITEM_CONT_READING_FROM_SRCDEV_MODE_AVOID)
4316                 srcdev = fs_info->dev_replace.srcdev;
4317         else
4318                 srcdev = NULL;
4319
4320         /*
4321          * try to avoid the drive that is the source drive for a
4322          * dev-replace procedure, only choose it if no other non-missing
4323          * mirror is available
4324          */
4325         for (tolerance = 0; tolerance < 2; tolerance++) {
4326                 if (map->stripes[optimal].dev->bdev &&
4327                     (tolerance || map->stripes[optimal].dev != srcdev))
4328                         return optimal;
4329                 for (i = first; i < first + num; i++) {
4330                         if (map->stripes[i].dev->bdev &&
4331                             (tolerance || map->stripes[i].dev != srcdev))
4332                                 return i;
4333                 }
4334         }
4335
4336         /* we couldn't find one that doesn't fail.  Just return something
4337          * and the io error handling code will clean up eventually
4338          */
4339         return optimal;
4340 }
4341
4342 static inline int parity_smaller(u64 a, u64 b)
4343 {
4344         return a > b;
4345 }
4346
4347 /* Bubble-sort the stripe set to put the parity/syndrome stripes last */
4348 static void sort_parity_stripes(struct btrfs_bio *bbio, u64 *raid_map)
4349 {
4350         struct btrfs_bio_stripe s;
4351         int i;
4352         u64 l;
4353         int again = 1;
4354
4355         while (again) {
4356                 again = 0;
4357                 for (i = 0; i < bbio->num_stripes - 1; i++) {
4358                         if (parity_smaller(raid_map[i], raid_map[i+1])) {
4359                                 s = bbio->stripes[i];
4360                                 l = raid_map[i];
4361                                 bbio->stripes[i] = bbio->stripes[i+1];
4362                                 raid_map[i] = raid_map[i+1];
4363                                 bbio->stripes[i+1] = s;
4364                                 raid_map[i+1] = l;
4365                                 again = 1;
4366                         }
4367                 }
4368         }
4369 }
4370
4371 static int __btrfs_map_block(struct btrfs_fs_info *fs_info, int rw,
4372                              u64 logical, u64 *length,
4373                              struct btrfs_bio **bbio_ret,
4374                              int mirror_num, u64 **raid_map_ret)
4375 {
4376         struct extent_map *em;
4377         struct map_lookup *map;
4378         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
4379         struct extent_map_tree *em_tree = &map_tree->map_tree;
4380         u64 offset;
4381         u64 stripe_offset;
4382         u64 stripe_end_offset;
4383         u64 stripe_nr;
4384         u64 stripe_nr_orig;
4385         u64 stripe_nr_end;
4386         u64 stripe_len;
4387         u64 *raid_map = NULL;
4388         int stripe_index;
4389         int i;
4390         int ret = 0;
4391         int num_stripes;
4392         int max_errors = 0;
4393         struct btrfs_bio *bbio = NULL;
4394         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
4395         int dev_replace_is_ongoing = 0;
4396         int num_alloc_stripes;
4397         int patch_the_first_stripe_for_dev_replace = 0;
4398         u64 physical_to_patch_in_first_stripe = 0;
4399         u64 raid56_full_stripe_start = (u64)-1;
4400
4401         read_lock(&em_tree->lock);
4402         em = lookup_extent_mapping(em_tree, logical, *length);
4403         read_unlock(&em_tree->lock);
4404
4405         if (!em) {
4406                 btrfs_crit(fs_info, "unable to find logical %llu len %llu",
4407                         (unsigned long long)logical,
4408                         (unsigned long long)*length);
4409                 BUG();
4410         }
4411
4412         BUG_ON(em->start > logical || em->start + em->len < logical);
4413         map = (struct map_lookup *)em->bdev;
4414         offset = logical - em->start;
4415
4416         if (mirror_num > map->num_stripes)
4417                 mirror_num = 0;
4418
4419         stripe_len = map->stripe_len;
4420         stripe_nr = offset;
4421         /*
4422          * stripe_nr counts the total number of stripes we have to stride
4423          * to get to this block
4424          */
4425         do_div(stripe_nr, stripe_len);
4426
4427         stripe_offset = stripe_nr * stripe_len;
4428         BUG_ON(offset < stripe_offset);
4429
4430         /* stripe_offset is the offset of this block in its stripe*/
4431         stripe_offset = offset - stripe_offset;
4432
4433         /* if we're here for raid56, we need to know the stripe aligned start */
4434         if (map->type & (BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6)) {
4435                 unsigned long full_stripe_len = stripe_len * nr_data_stripes(map);
4436                 raid56_full_stripe_start = offset;
4437
4438                 /* allow a write of a full stripe, but make sure we don't
4439                  * allow straddling of stripes
4440                  */
4441                 do_div(raid56_full_stripe_start, full_stripe_len);
4442                 raid56_full_stripe_start *= full_stripe_len;
4443         }
4444
4445         if (rw & REQ_DISCARD) {
4446                 /* we don't discard raid56 yet */
4447                 if (map->type &
4448                     (BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6)) {
4449                         ret = -EOPNOTSUPP;
4450                         goto out;
4451                 }
4452                 *length = min_t(u64, em->len - offset, *length);
4453         } else if (map->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
4454                 u64 max_len;
4455                 /* For writes to RAID[56], allow a full stripeset across all disks.
4456                    For other RAID types and for RAID[56] reads, just allow a single
4457                    stripe (on a single disk). */
4458                 if (map->type & (BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6) &&
4459                     (rw & REQ_WRITE)) {
4460                         max_len = stripe_len * nr_data_stripes(map) -
4461                                 (offset - raid56_full_stripe_start);
4462                 } else {
4463                         /* we limit the length of each bio to what fits in a stripe */
4464                         max_len = stripe_len - stripe_offset;
4465                 }
4466                 *length = min_t(u64, em->len - offset, max_len);
4467         } else {
4468                 *length = em->len - offset;
4469         }
4470
4471         /* This is for when we're called from btrfs_merge_bio_hook() and all
4472            it cares about is the length */
4473         if (!bbio_ret)
4474                 goto out;
4475
4476         btrfs_dev_replace_lock(dev_replace);
4477         dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(dev_replace);
4478         if (!dev_replace_is_ongoing)
4479                 btrfs_dev_replace_unlock(dev_replace);
4480
4481         if (dev_replace_is_ongoing && mirror_num == map->num_stripes + 1 &&
4482             !(rw & (REQ_WRITE | REQ_DISCARD | REQ_GET_READ_MIRRORS)) &&
4483             dev_replace->tgtdev != NULL) {
4484                 /*
4485                  * in dev-replace case, for repair case (that's the only
4486                  * case where the mirror is selected explicitly when
4487                  * calling btrfs_map_block), blocks left of the left cursor
4488                  * can also be read from the target drive.
4489                  * For REQ_GET_READ_MIRRORS, the target drive is added as
4490                  * the last one to the array of stripes. For READ, it also
4491                  * needs to be supported using the same mirror number.
4492                  * If the requested block is not left of the left cursor,
4493                  * EIO is returned. This can happen because btrfs_num_copies()
4494                  * returns one more in the dev-replace case.
4495                  */
4496                 u64 tmp_length = *length;
4497                 struct btrfs_bio *tmp_bbio = NULL;
4498                 int tmp_num_stripes;
4499                 u64 srcdev_devid = dev_replace->srcdev->devid;
4500                 int index_srcdev = 0;
4501                 int found = 0;
4502                 u64 physical_of_found = 0;
4503
4504                 ret = __btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS,
4505                              logical, &tmp_length, &tmp_bbio, 0, NULL);
4506                 if (ret) {
4507                         WARN_ON(tmp_bbio != NULL);
4508                         goto out;
4509                 }
4510
4511                 tmp_num_stripes = tmp_bbio->num_stripes;
4512                 if (mirror_num > tmp_num_stripes) {
4513                         /*
4514                          * REQ_GET_READ_MIRRORS does not contain this
4515                          * mirror, that means that the requested area
4516                          * is not left of the left cursor
4517                          */
4518                         ret = -EIO;
4519                         kfree(tmp_bbio);
4520                         goto out;
4521                 }
4522
4523                 /*
4524                  * process the rest of the function using the mirror_num
4525                  * of the source drive. Therefore look it up first.
4526                  * At the end, patch the device pointer to the one of the
4527                  * target drive.
4528                  */
4529                 for (i = 0; i < tmp_num_stripes; i++) {
4530                         if (tmp_bbio->stripes[i].dev->devid == srcdev_devid) {
4531                                 /*
4532                                  * In case of DUP, in order to keep it
4533                                  * simple, only add the mirror with the
4534                                  * lowest physical address
4535                                  */
4536                                 if (found &&
4537                                     physical_of_found <=
4538                                      tmp_bbio->stripes[i].physical)
4539                                         continue;
4540                                 index_srcdev = i;
4541                                 found = 1;
4542                                 physical_of_found =
4543                                         tmp_bbio->stripes[i].physical;
4544                         }
4545                 }
4546
4547                 if (found) {
4548                         mirror_num = index_srcdev + 1;
4549                         patch_the_first_stripe_for_dev_replace = 1;
4550                         physical_to_patch_in_first_stripe = physical_of_found;
4551                 } else {
4552                         WARN_ON(1);
4553                         ret = -EIO;
4554                         kfree(tmp_bbio);
4555                         goto out;
4556                 }
4557
4558                 kfree(tmp_bbio);
4559         } else if (mirror_num > map->num_stripes) {
4560                 mirror_num = 0;
4561         }
4562
4563         num_stripes = 1;
4564         stripe_index = 0;
4565         stripe_nr_orig = stripe_nr;
4566         stripe_nr_end = ALIGN(offset + *length, map->stripe_len);
4567         do_div(stripe_nr_end, map->stripe_len);
4568         stripe_end_offset = stripe_nr_end * map->stripe_len -
4569                             (offset + *length);
4570
4571         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
4572                 if (rw & REQ_DISCARD)
4573                         num_stripes = min_t(u64, map->num_stripes,
4574                                             stripe_nr_end - stripe_nr_orig);
4575                 stripe_index = do_div(stripe_nr, map->num_stripes);
4576         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
4577                 if (rw & (REQ_WRITE | REQ_DISCARD | REQ_GET_READ_MIRRORS))
4578                         num_stripes = map->num_stripes;
4579                 else if (mirror_num)
4580                         stripe_index = mirror_num - 1;
4581                 else {
4582                         stripe_index = find_live_mirror(fs_info, map, 0,
4583                                             map->num_stripes,
4584                                             current->pid % map->num_stripes,
4585                                             dev_replace_is_ongoing);
4586                         mirror_num = stripe_index + 1;
4587                 }
4588
4589         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
4590                 if (rw & (REQ_WRITE | REQ_DISCARD | REQ_GET_READ_MIRRORS)) {
4591                         num_stripes = map->num_stripes;
4592                 } else if (mirror_num) {
4593                         stripe_index = mirror_num - 1;
4594                 } else {
4595                         mirror_num = 1;
4596                 }
4597
4598         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
4599                 int factor = map->num_stripes / map->sub_stripes;
4600
4601                 stripe_index = do_div(stripe_nr, factor);
4602                 stripe_index *= map->sub_stripes;
4603
4604                 if (rw & (REQ_WRITE | REQ_GET_READ_MIRRORS))
4605                         num_stripes = map->sub_stripes;
4606                 else if (rw & REQ_DISCARD)
4607                         num_stripes = min_t(u64, map->sub_stripes *
4608                                             (stripe_nr_end - stripe_nr_orig),
4609                                             map->num_stripes);
4610                 else if (mirror_num)
4611                         stripe_index += mirror_num - 1;
4612                 else {
4613                         int old_stripe_index = stripe_index;
4614                         stripe_index = find_live_mirror(fs_info, map,
4615                                               stripe_index,
4616                                               map->sub_stripes, stripe_index +
4617                                               current->pid % map->sub_stripes,
4618                                               dev_replace_is_ongoing);
4619                         mirror_num = stripe_index - old_stripe_index + 1;
4620                 }
4621
4622         } else if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
4623                                 BTRFS_BLOCK_GROUP_RAID6)) {
4624                 u64 tmp;
4625
4626                 if (bbio_ret && ((rw & REQ_WRITE) || mirror_num > 1)
4627                     && raid_map_ret) {
4628                         int i, rot;
4629
4630                         /* push stripe_nr back to the start of the full stripe */
4631                         stripe_nr = raid56_full_stripe_start;
4632                         do_div(stripe_nr, stripe_len);
4633
4634                         stripe_index = do_div(stripe_nr, nr_data_stripes(map));
4635
4636                         /* RAID[56] write or recovery. Return all stripes */
4637                         num_stripes = map->num_stripes;
4638                         max_errors = nr_parity_stripes(map);
4639
4640                         raid_map = kmalloc(sizeof(u64) * num_stripes,
4641                                            GFP_NOFS);
4642                         if (!raid_map) {
4643                                 ret = -ENOMEM;
4644                                 goto out;
4645                         }
4646
4647                         /* Work out the disk rotation on this stripe-set */
4648                         tmp = stripe_nr;
4649                         rot = do_div(tmp, num_stripes);
4650
4651                         /* Fill in the logical address of each stripe */
4652                         tmp = stripe_nr * nr_data_stripes(map);
4653                         for (i = 0; i < nr_data_stripes(map); i++)
4654                                 raid_map[(i+rot) % num_stripes] =
4655                                         em->start + (tmp + i) * map->stripe_len;
4656
4657                         raid_map[(i+rot) % map->num_stripes] = RAID5_P_STRIPE;
4658                         if (map->type & BTRFS_BLOCK_GROUP_RAID6)
4659                                 raid_map[(i+rot+1) % num_stripes] =
4660                                         RAID6_Q_STRIPE;
4661
4662                         *length = map->stripe_len;
4663                         stripe_index = 0;
4664                         stripe_offset = 0;
4665                 } else {
4666                         /*
4667                          * Mirror #0 or #1 means the original data block.
4668                          * Mirror #2 is RAID5 parity block.
4669                          * Mirror #3 is RAID6 Q block.
4670                          */
4671                         stripe_index = do_div(stripe_nr, nr_data_stripes(map));
4672                         if (mirror_num > 1)
4673                                 stripe_index = nr_data_stripes(map) +
4674                                                 mirror_num - 2;
4675
4676                         /* We distribute the parity blocks across stripes */
4677                         tmp = stripe_nr + stripe_index;
4678                         stripe_index = do_div(tmp, map->num_stripes);
4679                 }
4680         } else {
4681                 /*
4682                  * after this do_div call, stripe_nr is the number of stripes
4683                  * on this device we have to walk to find the data, and
4684                  * stripe_index is the number of our device in the stripe array
4685                  */
4686                 stripe_index = do_div(stripe_nr, map->num_stripes);
4687                 mirror_num = stripe_index + 1;
4688         }
4689         BUG_ON(stripe_index >= map->num_stripes);
4690
4691         num_alloc_stripes = num_stripes;
4692         if (dev_replace_is_ongoing) {
4693                 if (rw & (REQ_WRITE | REQ_DISCARD))
4694                         num_alloc_stripes <<= 1;
4695                 if (rw & REQ_GET_READ_MIRRORS)
4696                         num_alloc_stripes++;
4697         }
4698         bbio = kzalloc(btrfs_bio_size(num_alloc_stripes), GFP_NOFS);
4699         if (!bbio) {
4700                 ret = -ENOMEM;
4701                 goto out;
4702         }
4703         atomic_set(&bbio->error, 0);
4704
4705         if (rw & REQ_DISCARD) {
4706                 int factor = 0;
4707                 int sub_stripes = 0;
4708                 u64 stripes_per_dev = 0;
4709                 u32 remaining_stripes = 0;
4710                 u32 last_stripe = 0;
4711
4712                 if (map->type &
4713                     (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID10)) {
4714                         if (map->type & BTRFS_BLOCK_GROUP_RAID0)
4715                                 sub_stripes = 1;
4716                         else
4717                                 sub_stripes = map->sub_stripes;
4718
4719                         factor = map->num_stripes / sub_stripes;
4720                         stripes_per_dev = div_u64_rem(stripe_nr_end -
4721                                                       stripe_nr_orig,
4722                                                       factor,
4723                                                       &remaining_stripes);
4724                         div_u64_rem(stripe_nr_end - 1, factor, &last_stripe);
4725                         last_stripe *= sub_stripes;
4726                 }
4727
4728                 for (i = 0; i < num_stripes; i++) {
4729                         bbio->stripes[i].physical =
4730                                 map->stripes[stripe_index].physical +
4731                                 stripe_offset + stripe_nr * map->stripe_len;
4732                         bbio->stripes[i].dev = map->stripes[stripe_index].dev;
4733
4734                         if (map->type & (BTRFS_BLOCK_GROUP_RAID0 |
4735                                          BTRFS_BLOCK_GROUP_RAID10)) {
4736                                 bbio->stripes[i].length = stripes_per_dev *
4737                                                           map->stripe_len;
4738
4739                                 if (i / sub_stripes < remaining_stripes)
4740                                         bbio->stripes[i].length +=
4741                                                 map->stripe_len;
4742
4743                                 /*
4744                                  * Special for the first stripe and
4745                                  * the last stripe:
4746                                  *
4747                                  * |-------|...|-------|
4748                                  *     |----------|
4749                                  *    off     end_off
4750                                  */
4751                                 if (i < sub_stripes)
4752                                         bbio->stripes[i].length -=
4753                                                 stripe_offset;
4754
4755                                 if (stripe_index >= last_stripe &&
4756                                     stripe_index <= (last_stripe +
4757                                                      sub_stripes - 1))
4758                                         bbio->stripes[i].length -=
4759                                                 stripe_end_offset;
4760
4761                                 if (i == sub_stripes - 1)
4762                                         stripe_offset = 0;
4763                         } else
4764                                 bbio->stripes[i].length = *length;
4765
4766                         stripe_index++;
4767                         if (stripe_index == map->num_stripes) {
4768                                 /* This could only happen for RAID0/10 */
4769                                 stripe_index = 0;
4770                                 stripe_nr++;
4771                         }
4772                 }
4773         } else {
4774                 for (i = 0; i < num_stripes; i++) {
4775                         bbio->stripes[i].physical =
4776                                 map->stripes[stripe_index].physical +
4777                                 stripe_offset +
4778                                 stripe_nr * map->stripe_len;
4779                         bbio->stripes[i].dev =
4780                                 map->stripes[stripe_index].dev;
4781                         stripe_index++;
4782                 }
4783         }
4784
4785         if (rw & (REQ_WRITE | REQ_GET_READ_MIRRORS)) {
4786                 if (map->type & (BTRFS_BLOCK_GROUP_RAID1 |
4787                                  BTRFS_BLOCK_GROUP_RAID10 |
4788                                  BTRFS_BLOCK_GROUP_RAID5 |
4789                                  BTRFS_BLOCK_GROUP_DUP)) {
4790                         max_errors = 1;
4791                 } else if (map->type & BTRFS_BLOCK_GROUP_RAID6) {
4792                         max_errors = 2;
4793                 }
4794         }
4795
4796         if (dev_replace_is_ongoing && (rw & (REQ_WRITE | REQ_DISCARD)) &&
4797             dev_replace->tgtdev != NULL) {
4798                 int index_where_to_add;
4799                 u64 srcdev_devid = dev_replace->srcdev->devid;
4800
4801                 /*
4802                  * duplicate the write operations while the dev replace
4803                  * procedure is running. Since the copying of the old disk
4804                  * to the new disk takes place at run time while the
4805                  * filesystem is mounted writable, the regular write
4806                  * operations to the old disk have to be duplicated to go
4807                  * to the new disk as well.
4808                  * Note that device->missing is handled by the caller, and
4809                  * that the write to the old disk is already set up in the
4810                  * stripes array.
4811                  */
4812                 index_where_to_add = num_stripes;
4813                 for (i = 0; i < num_stripes; i++) {
4814                         if (bbio->stripes[i].dev->devid == srcdev_devid) {
4815                                 /* write to new disk, too */
4816                                 struct btrfs_bio_stripe *new =
4817                                         bbio->stripes + index_where_to_add;
4818                                 struct btrfs_bio_stripe *old =
4819                                         bbio->stripes + i;
4820
4821                                 new->physical = old->physical;
4822                                 new->length = old->length;
4823                                 new->dev = dev_replace->tgtdev;
4824                                 index_where_to_add++;
4825                                 max_errors++;
4826                         }
4827                 }
4828                 num_stripes = index_where_to_add;
4829         } else if (dev_replace_is_ongoing && (rw & REQ_GET_READ_MIRRORS) &&
4830                    dev_replace->tgtdev != NULL) {
4831                 u64 srcdev_devid = dev_replace->srcdev->devid;
4832                 int index_srcdev = 0;
4833                 int found = 0;
4834                 u64 physical_of_found = 0;
4835
4836                 /*
4837                  * During the dev-replace procedure, the target drive can
4838                  * also be used to read data in case it is needed to repair
4839                  * a corrupt block elsewhere. This is possible if the
4840                  * requested area is left of the left cursor. In this area,
4841                  * the target drive is a full copy of the source drive.
4842                  */
4843                 for (i = 0; i < num_stripes; i++) {
4844                         if (bbio->stripes[i].dev->devid == srcdev_devid) {
4845                                 /*
4846                                  * In case of DUP, in order to keep it
4847                                  * simple, only add the mirror with the
4848                                  * lowest physical address
4849                                  */
4850                                 if (found &&
4851                                     physical_of_found <=
4852                                      bbio->stripes[i].physical)
4853                                         continue;
4854                                 index_srcdev = i;
4855                                 found = 1;
4856                                 physical_of_found = bbio->stripes[i].physical;
4857                         }
4858                 }
4859                 if (found) {
4860                         u64 length = map->stripe_len;
4861
4862                         if (physical_of_found + length <=
4863                             dev_replace->cursor_left) {
4864                                 struct btrfs_bio_stripe *tgtdev_stripe =
4865                                         bbio->stripes + num_stripes;
4866
4867                                 tgtdev_stripe->physical = physical_of_found;
4868                                 tgtdev_stripe->length =
4869                                         bbio->stripes[index_srcdev].length;
4870                                 tgtdev_stripe->dev = dev_replace->tgtdev;
4871
4872                                 num_stripes++;
4873                         }
4874                 }
4875         }
4876
4877         *bbio_ret = bbio;
4878         bbio->num_stripes = num_stripes;
4879         bbio->max_errors = max_errors;
4880         bbio->mirror_num = mirror_num;
4881
4882         /*
4883          * this is the case that REQ_READ && dev_replace_is_ongoing &&
4884          * mirror_num == num_stripes + 1 && dev_replace target drive is
4885          * available as a mirror
4886          */
4887         if (patch_the_first_stripe_for_dev_replace && num_stripes > 0) {
4888                 WARN_ON(num_stripes > 1);
4889                 bbio->stripes[0].dev = dev_replace->tgtdev;
4890                 bbio->stripes[0].physical = physical_to_patch_in_first_stripe;
4891                 bbio->mirror_num = map->num_stripes + 1;
4892         }
4893         if (raid_map) {
4894                 sort_parity_stripes(bbio, raid_map);
4895                 *raid_map_ret = raid_map;
4896         }
4897 out:
4898         if (dev_replace_is_ongoing)
4899                 btrfs_dev_replace_unlock(dev_replace);
4900         free_extent_map(em);
4901         return ret;
4902 }
4903
4904 int btrfs_map_block(struct btrfs_fs_info *fs_info, int rw,
4905                       u64 logical, u64 *length,
4906                       struct btrfs_bio **bbio_ret, int mirror_num)
4907 {
4908         return __btrfs_map_block(fs_info, rw, logical, length, bbio_ret,
4909                                  mirror_num, NULL);
4910 }
4911
4912 int btrfs_rmap_block(struct btrfs_mapping_tree *map_tree,
4913                      u64 chunk_start, u64 physical, u64 devid,
4914                      u64 **logical, int *naddrs, int *stripe_len)
4915 {
4916         struct extent_map_tree *em_tree = &map_tree->map_tree;
4917         struct extent_map *em;
4918         struct map_lookup *map;
4919         u64 *buf;
4920         u64 bytenr;
4921         u64 length;
4922         u64 stripe_nr;
4923         u64 rmap_len;
4924         int i, j, nr = 0;
4925
4926         read_lock(&em_tree->lock);
4927         em = lookup_extent_mapping(em_tree, chunk_start, 1);
4928         read_unlock(&em_tree->lock);
4929
4930         if (!em) {
4931                 printk(KERN_ERR "btrfs: couldn't find em for chunk %Lu\n",
4932                        chunk_start);
4933                 return -EIO;
4934         }
4935
4936         if (em->start != chunk_start) {
4937                 printk(KERN_ERR "btrfs: bad chunk start, em=%Lu, wanted=%Lu\n",
4938                        em->start, chunk_start);
4939                 free_extent_map(em);
4940                 return -EIO;
4941         }
4942         map = (struct map_lookup *)em->bdev;
4943
4944         length = em->len;
4945         rmap_len = map->stripe_len;
4946
4947         if (map->type & BTRFS_BLOCK_GROUP_RAID10)
4948                 do_div(length, map->num_stripes / map->sub_stripes);
4949         else if (map->type & BTRFS_BLOCK_GROUP_RAID0)
4950                 do_div(length, map->num_stripes);
4951         else if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
4952                               BTRFS_BLOCK_GROUP_RAID6)) {
4953                 do_div(length, nr_data_stripes(map));
4954                 rmap_len = map->stripe_len * nr_data_stripes(map);
4955         }
4956
4957         buf = kzalloc(sizeof(u64) * map->num_stripes, GFP_NOFS);
4958         BUG_ON(!buf); /* -ENOMEM */
4959
4960         for (i = 0; i < map->num_stripes; i++) {
4961                 if (devid && map->stripes[i].dev->devid != devid)
4962                         continue;
4963                 if (map->stripes[i].physical > physical ||
4964                     map->stripes[i].physical + length <= physical)
4965                         continue;
4966
4967                 stripe_nr = physical - map->stripes[i].physical;
4968                 do_div(stripe_nr, map->stripe_len);
4969
4970                 if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
4971                         stripe_nr = stripe_nr * map->num_stripes + i;
4972                         do_div(stripe_nr, map->sub_stripes);
4973                 } else if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
4974                         stripe_nr = stripe_nr * map->num_stripes + i;
4975                 } /* else if RAID[56], multiply by nr_data_stripes().
4976                    * Alternatively, just use rmap_len below instead of
4977                    * map->stripe_len */
4978
4979                 bytenr = chunk_start + stripe_nr * rmap_len;
4980                 WARN_ON(nr >= map->num_stripes);
4981                 for (j = 0; j < nr; j++) {
4982                         if (buf[j] == bytenr)
4983                                 break;
4984                 }
4985                 if (j == nr) {
4986                         WARN_ON(nr >= map->num_stripes);
4987                         buf[nr++] = bytenr;
4988                 }
4989         }
4990
4991         *logical = buf;
4992         *naddrs = nr;
4993         *stripe_len = rmap_len;
4994
4995         free_extent_map(em);
4996         return 0;
4997 }
4998
4999 static void *merge_stripe_index_into_bio_private(void *bi_private,
5000                                                  unsigned int stripe_index)
5001 {
5002         /*
5003          * with single, dup, RAID0, RAID1 and RAID10, stripe_index is
5004          * at most 1.
5005          * The alternative solution (instead of stealing bits from the
5006          * pointer) would be to allocate an intermediate structure
5007          * that contains the old private pointer plus the stripe_index.
5008          */
5009         BUG_ON((((uintptr_t)bi_private) & 3) != 0);
5010         BUG_ON(stripe_index > 3);
5011         return (void *)(((uintptr_t)bi_private) | stripe_index);
5012 }
5013
5014 static struct btrfs_bio *extract_bbio_from_bio_private(void *bi_private)
5015 {
5016         return (struct btrfs_bio *)(((uintptr_t)bi_private) & ~((uintptr_t)3));
5017 }
5018
5019 static unsigned int extract_stripe_index_from_bio_private(void *bi_private)
5020 {
5021         return (unsigned int)((uintptr_t)bi_private) & 3;
5022 }
5023
5024 static void btrfs_end_bio(struct bio *bio, int err)
5025 {
5026         struct btrfs_bio *bbio = extract_bbio_from_bio_private(bio->bi_private);
5027         int is_orig_bio = 0;
5028
5029         if (err) {
5030                 atomic_inc(&bbio->error);
5031                 if (err == -EIO || err == -EREMOTEIO) {
5032                         unsigned int stripe_index =
5033                                 extract_stripe_index_from_bio_private(
5034                                         bio->bi_private);
5035                         struct btrfs_device *dev;
5036
5037                         BUG_ON(stripe_index >= bbio->num_stripes);
5038                         dev = bbio->stripes[stripe_index].dev;
5039                         if (dev->bdev) {
5040                                 if (bio->bi_rw & WRITE)
5041                                         btrfs_dev_stat_inc(dev,
5042                                                 BTRFS_DEV_STAT_WRITE_ERRS);
5043                                 else
5044                                         btrfs_dev_stat_inc(dev,
5045                                                 BTRFS_DEV_STAT_READ_ERRS);
5046                                 if ((bio->bi_rw & WRITE_FLUSH) == WRITE_FLUSH)
5047                                         btrfs_dev_stat_inc(dev,
5048                                                 BTRFS_DEV_STAT_FLUSH_ERRS);
5049                                 btrfs_dev_stat_print_on_error(dev);
5050                         }
5051                 }
5052         }
5053
5054         if (bio == bbio->orig_bio)
5055                 is_orig_bio = 1;
5056
5057         if (atomic_dec_and_test(&bbio->stripes_pending)) {
5058                 if (!is_orig_bio) {
5059                         bio_put(bio);
5060                         bio = bbio->orig_bio;
5061                 }
5062                 bio->bi_private = bbio->private;
5063                 bio->bi_end_io = bbio->end_io;
5064                 bio->bi_bdev = (struct block_device *)
5065                                         (unsigned long)bbio->mirror_num;
5066                 /* only send an error to the higher layers if it is
5067                  * beyond the tolerance of the btrfs bio
5068                  */
5069                 if (atomic_read(&bbio->error) > bbio->max_errors) {
5070                         err = -EIO;
5071                 } else {
5072                         /*
5073                          * this bio is actually up to date, we didn't
5074                          * go over the max number of errors
5075                          */
5076                         set_bit(BIO_UPTODATE, &bio->bi_flags);
5077                         err = 0;
5078                 }
5079                 kfree(bbio);
5080
5081                 bio_endio(bio, err);
5082         } else if (!is_orig_bio) {
5083                 bio_put(bio);
5084         }
5085 }
5086
5087 struct async_sched {
5088         struct bio *bio;
5089         int rw;
5090         struct btrfs_fs_info *info;
5091         struct btrfs_work work;
5092 };
5093
5094 /*
5095  * see run_scheduled_bios for a description of why bios are collected for
5096  * async submit.
5097  *
5098  * This will add one bio to the pending list for a device and make sure
5099  * the work struct is scheduled.
5100  */
5101 noinline void btrfs_schedule_bio(struct btrfs_root *root,
5102                                  struct btrfs_device *device,
5103                                  int rw, struct bio *bio)
5104 {
5105         int should_queue = 1;
5106         struct btrfs_pending_bios *pending_bios;
5107
5108         if (device->missing || !device->bdev) {
5109                 bio_endio(bio, -EIO);
5110                 return;
5111         }
5112
5113         /* don't bother with additional async steps for reads, right now */
5114         if (!(rw & REQ_WRITE)) {
5115                 bio_get(bio);
5116                 btrfsic_submit_bio(rw, bio);
5117                 bio_put(bio);
5118                 return;
5119         }
5120
5121         /*
5122          * nr_async_bios allows us to reliably return congestion to the
5123          * higher layers.  Otherwise, the async bio makes it appear we have
5124          * made progress against dirty pages when we've really just put it
5125          * on a queue for later
5126          */
5127         atomic_inc(&root->fs_info->nr_async_bios);
5128         WARN_ON(bio->bi_next);
5129         bio->bi_next = NULL;
5130         bio->bi_rw |= rw;
5131
5132         spin_lock(&device->io_lock);
5133         if (bio->bi_rw & REQ_SYNC)
5134                 pending_bios = &device->pending_sync_bios;
5135         else
5136                 pending_bios = &device->pending_bios;
5137
5138         if (pending_bios->tail)
5139                 pending_bios->tail->bi_next = bio;
5140
5141         pending_bios->tail = bio;
5142         if (!pending_bios->head)
5143                 pending_bios->head = bio;
5144         if (device->running_pending)
5145                 should_queue = 0;
5146
5147         spin_unlock(&device->io_lock);
5148
5149         if (should_queue)
5150                 btrfs_queue_worker(&root->fs_info->submit_workers,
5151                                    &device->work);
5152 }
5153
5154 static int bio_size_ok(struct block_device *bdev, struct bio *bio,
5155                        sector_t sector)
5156 {
5157         struct bio_vec *prev;
5158         struct request_queue *q = bdev_get_queue(bdev);
5159         unsigned short max_sectors = queue_max_sectors(q);
5160         struct bvec_merge_data bvm = {
5161                 .bi_bdev = bdev,
5162                 .bi_sector = sector,
5163                 .bi_rw = bio->bi_rw,
5164         };
5165
5166         if (bio->bi_vcnt == 0) {
5167                 WARN_ON(1);
5168                 return 1;
5169         }
5170
5171         prev = &bio->bi_io_vec[bio->bi_vcnt - 1];
5172         if ((bio->bi_size >> 9) > max_sectors)
5173                 return 0;
5174
5175         if (!q->merge_bvec_fn)
5176                 return 1;
5177
5178         bvm.bi_size = bio->bi_size - prev->bv_len;
5179         if (q->merge_bvec_fn(q, &bvm, prev) < prev->bv_len)
5180                 return 0;
5181         return 1;
5182 }
5183
5184 static void submit_stripe_bio(struct btrfs_root *root, struct btrfs_bio *bbio,
5185                               struct bio *bio, u64 physical, int dev_nr,
5186                               int rw, int async)
5187 {
5188         struct btrfs_device *dev = bbio->stripes[dev_nr].dev;
5189
5190         bio->bi_private = bbio;
5191         bio->bi_private = merge_stripe_index_into_bio_private(
5192                         bio->bi_private, (unsigned int)dev_nr);
5193         bio->bi_end_io = btrfs_end_bio;
5194         bio->bi_sector = physical >> 9;
5195 #ifdef DEBUG
5196         {
5197                 struct rcu_string *name;
5198
5199                 rcu_read_lock();
5200                 name = rcu_dereference(dev->name);
5201                 pr_debug("btrfs_map_bio: rw %d, sector=%llu, dev=%lu "
5202                          "(%s id %llu), size=%u\n", rw,
5203                          (u64)bio->bi_sector, (u_long)dev->bdev->bd_dev,
5204                          name->str, dev->devid, bio->bi_size);
5205                 rcu_read_unlock();
5206         }
5207 #endif
5208         bio->bi_bdev = dev->bdev;
5209         if (async)
5210                 btrfs_schedule_bio(root, dev, rw, bio);
5211         else
5212                 btrfsic_submit_bio(rw, bio);
5213 }
5214
5215 static int breakup_stripe_bio(struct btrfs_root *root, struct btrfs_bio *bbio,
5216                               struct bio *first_bio, struct btrfs_device *dev,
5217                               int dev_nr, int rw, int async)
5218 {
5219         struct bio_vec *bvec = first_bio->bi_io_vec;
5220         struct bio *bio;
5221         int nr_vecs = bio_get_nr_vecs(dev->bdev);
5222         u64 physical = bbio->stripes[dev_nr].physical;
5223
5224 again:
5225         bio = btrfs_bio_alloc(dev->bdev, physical >> 9, nr_vecs, GFP_NOFS);
5226         if (!bio)
5227                 return -ENOMEM;
5228
5229         while (bvec <= (first_bio->bi_io_vec + first_bio->bi_vcnt - 1)) {
5230                 if (bio_add_page(bio, bvec->bv_page, bvec->bv_len,
5231                                  bvec->bv_offset) < bvec->bv_len) {
5232                         u64 len = bio->bi_size;
5233
5234                         atomic_inc(&bbio->stripes_pending);
5235                         submit_stripe_bio(root, bbio, bio, physical, dev_nr,
5236                                           rw, async);
5237                         physical += len;
5238                         goto again;
5239                 }
5240                 bvec++;
5241         }
5242
5243         submit_stripe_bio(root, bbio, bio, physical, dev_nr, rw, async);
5244         return 0;
5245 }
5246
5247 static void bbio_error(struct btrfs_bio *bbio, struct bio *bio, u64 logical)
5248 {
5249         atomic_inc(&bbio->error);
5250         if (atomic_dec_and_test(&bbio->stripes_pending)) {
5251                 bio->bi_private = bbio->private;
5252                 bio->bi_end_io = bbio->end_io;
5253                 bio->bi_bdev = (struct block_device *)
5254                         (unsigned long)bbio->mirror_num;
5255                 bio->bi_sector = logical >> 9;
5256                 kfree(bbio);
5257                 bio_endio(bio, -EIO);
5258         }
5259 }
5260
5261 int btrfs_map_bio(struct btrfs_root *root, int rw, struct bio *bio,
5262                   int mirror_num, int async_submit)
5263 {
5264         struct btrfs_device *dev;
5265         struct bio *first_bio = bio;
5266         u64 logical = (u64)bio->bi_sector << 9;
5267         u64 length = 0;
5268         u64 map_length;
5269         u64 *raid_map = NULL;
5270         int ret;
5271         int dev_nr = 0;
5272         int total_devs = 1;
5273         struct btrfs_bio *bbio = NULL;
5274
5275         length = bio->bi_size;
5276         map_length = length;
5277
5278         ret = __btrfs_map_block(root->fs_info, rw, logical, &map_length, &bbio,
5279                               mirror_num, &raid_map);
5280         if (ret) /* -ENOMEM */
5281                 return ret;
5282
5283         total_devs = bbio->num_stripes;
5284         bbio->orig_bio = first_bio;
5285         bbio->private = first_bio->bi_private;
5286         bbio->end_io = first_bio->bi_end_io;
5287         atomic_set(&bbio->stripes_pending, bbio->num_stripes);
5288
5289         if (raid_map) {
5290                 /* In this case, map_length has been set to the length of
5291                    a single stripe; not the whole write */
5292                 if (rw & WRITE) {
5293                         return raid56_parity_write(root, bio, bbio,
5294                                                    raid_map, map_length);
5295                 } else {
5296                         return raid56_parity_recover(root, bio, bbio,
5297                                                      raid_map, map_length,
5298                                                      mirror_num);
5299                 }
5300         }
5301
5302         if (map_length < length) {
5303                 btrfs_crit(root->fs_info, "mapping failed logical %llu bio len %llu len %llu",
5304                         (unsigned long long)logical,
5305                         (unsigned long long)length,
5306                         (unsigned long long)map_length);
5307                 BUG();
5308         }
5309
5310         while (dev_nr < total_devs) {
5311                 dev = bbio->stripes[dev_nr].dev;
5312                 if (!dev || !dev->bdev || (rw & WRITE && !dev->writeable)) {
5313                         bbio_error(bbio, first_bio, logical);
5314                         dev_nr++;
5315                         continue;
5316                 }
5317
5318                 /*
5319                  * Check and see if we're ok with this bio based on it's size
5320                  * and offset with the given device.
5321                  */
5322                 if (!bio_size_ok(dev->bdev, first_bio,
5323                                  bbio->stripes[dev_nr].physical >> 9)) {
5324                         ret = breakup_stripe_bio(root, bbio, first_bio, dev,
5325                                                  dev_nr, rw, async_submit);
5326                         BUG_ON(ret);
5327                         dev_nr++;
5328                         continue;
5329                 }
5330
5331                 if (dev_nr < total_devs - 1) {
5332                         bio = bio_clone(first_bio, GFP_NOFS);
5333                         BUG_ON(!bio); /* -ENOMEM */
5334                 } else {
5335                         bio = first_bio;
5336                 }
5337
5338                 submit_stripe_bio(root, bbio, bio,
5339                                   bbio->stripes[dev_nr].physical, dev_nr, rw,
5340                                   async_submit);
5341                 dev_nr++;
5342         }
5343         return 0;
5344 }
5345
5346 struct btrfs_device *btrfs_find_device(struct btrfs_fs_info *fs_info, u64 devid,
5347                                        u8 *uuid, u8 *fsid)
5348 {
5349         struct btrfs_device *device;
5350         struct btrfs_fs_devices *cur_devices;
5351
5352         cur_devices = fs_info->fs_devices;
5353         while (cur_devices) {
5354                 if (!fsid ||
5355                     !memcmp(cur_devices->fsid, fsid, BTRFS_UUID_SIZE)) {
5356                         device = __find_device(&cur_devices->devices,
5357                                                devid, uuid);
5358                         if (device)
5359                                 return device;
5360                 }
5361                 cur_devices = cur_devices->seed;
5362         }
5363         return NULL;
5364 }
5365
5366 static struct btrfs_device *add_missing_dev(struct btrfs_root *root,
5367                                             u64 devid, u8 *dev_uuid)
5368 {
5369         struct btrfs_device *device;
5370         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
5371
5372         device = kzalloc(sizeof(*device), GFP_NOFS);
5373         if (!device)
5374                 return NULL;
5375         list_add(&device->dev_list,
5376                  &fs_devices->devices);
5377         device->dev_root = root->fs_info->dev_root;
5378         device->devid = devid;
5379         device->work.func = pending_bios_fn;
5380         device->fs_devices = fs_devices;
5381         device->missing = 1;
5382         fs_devices->num_devices++;
5383         fs_devices->missing_devices++;
5384         spin_lock_init(&device->io_lock);
5385         INIT_LIST_HEAD(&device->dev_alloc_list);
5386         memcpy(device->uuid, dev_uuid, BTRFS_UUID_SIZE);
5387         return device;
5388 }
5389
5390 static int read_one_chunk(struct btrfs_root *root, struct btrfs_key *key,
5391                           struct extent_buffer *leaf,
5392                           struct btrfs_chunk *chunk)
5393 {
5394         struct btrfs_mapping_tree *map_tree = &root->fs_info->mapping_tree;
5395         struct map_lookup *map;
5396         struct extent_map *em;
5397         u64 logical;
5398         u64 length;
5399         u64 devid;
5400         u8 uuid[BTRFS_UUID_SIZE];
5401         int num_stripes;
5402         int ret;
5403         int i;
5404
5405         logical = key->offset;
5406         length = btrfs_chunk_length(leaf, chunk);
5407
5408         read_lock(&map_tree->map_tree.lock);
5409         em = lookup_extent_mapping(&map_tree->map_tree, logical, 1);
5410         read_unlock(&map_tree->map_tree.lock);
5411
5412         /* already mapped? */
5413         if (em && em->start <= logical && em->start + em->len > logical) {
5414                 free_extent_map(em);
5415                 return 0;
5416         } else if (em) {
5417                 free_extent_map(em);
5418         }
5419
5420         em = alloc_extent_map();
5421         if (!em)
5422                 return -ENOMEM;
5423         num_stripes = btrfs_chunk_num_stripes(leaf, chunk);
5424         map = kmalloc(map_lookup_size(num_stripes), GFP_NOFS);
5425         if (!map) {
5426                 free_extent_map(em);
5427                 return -ENOMEM;
5428         }
5429
5430         em->bdev = (struct block_device *)map;
5431         em->start = logical;
5432         em->len = length;
5433         em->orig_start = 0;
5434         em->block_start = 0;
5435         em->block_len = em->len;
5436
5437         map->num_stripes = num_stripes;
5438         map->io_width = btrfs_chunk_io_width(leaf, chunk);
5439         map->io_align = btrfs_chunk_io_align(leaf, chunk);
5440         map->sector_size = btrfs_chunk_sector_size(leaf, chunk);
5441         map->stripe_len = btrfs_chunk_stripe_len(leaf, chunk);
5442         map->type = btrfs_chunk_type(leaf, chunk);
5443         map->sub_stripes = btrfs_chunk_sub_stripes(leaf, chunk);
5444         for (i = 0; i < num_stripes; i++) {
5445                 map->stripes[i].physical =
5446                         btrfs_stripe_offset_nr(leaf, chunk, i);
5447                 devid = btrfs_stripe_devid_nr(leaf, chunk, i);
5448                 read_extent_buffer(leaf, uuid, (unsigned long)
5449                                    btrfs_stripe_dev_uuid_nr(chunk, i),
5450                                    BTRFS_UUID_SIZE);
5451                 map->stripes[i].dev = btrfs_find_device(root->fs_info, devid,
5452                                                         uuid, NULL);
5453                 if (!map->stripes[i].dev && !btrfs_test_opt(root, DEGRADED)) {
5454                         kfree(map);
5455                         free_extent_map(em);
5456                         return -EIO;
5457                 }
5458                 if (!map->stripes[i].dev) {
5459                         map->stripes[i].dev =
5460                                 add_missing_dev(root, devid, uuid);
5461                         if (!map->stripes[i].dev) {
5462                                 kfree(map);
5463                                 free_extent_map(em);
5464                                 return -EIO;
5465                         }
5466                 }
5467                 map->stripes[i].dev->in_fs_metadata = 1;
5468         }
5469
5470         write_lock(&map_tree->map_tree.lock);
5471         ret = add_extent_mapping(&map_tree->map_tree, em, 0);
5472         write_unlock(&map_tree->map_tree.lock);
5473         BUG_ON(ret); /* Tree corruption */
5474         free_extent_map(em);
5475
5476         return 0;
5477 }
5478
5479 static void fill_device_from_item(struct extent_buffer *leaf,
5480                                  struct btrfs_dev_item *dev_item,
5481                                  struct btrfs_device *device)
5482 {
5483         unsigned long ptr;
5484
5485         device->devid = btrfs_device_id(leaf, dev_item);
5486         device->disk_total_bytes = btrfs_device_total_bytes(leaf, dev_item);
5487         device->total_bytes = device->disk_total_bytes;
5488         device->bytes_used = btrfs_device_bytes_used(leaf, dev_item);
5489         device->type = btrfs_device_type(leaf, dev_item);
5490         device->io_align = btrfs_device_io_align(leaf, dev_item);
5491         device->io_width = btrfs_device_io_width(leaf, dev_item);
5492         device->sector_size = btrfs_device_sector_size(leaf, dev_item);
5493         WARN_ON(device->devid == BTRFS_DEV_REPLACE_DEVID);
5494         device->is_tgtdev_for_dev_replace = 0;
5495
5496         ptr = (unsigned long)btrfs_device_uuid(dev_item);
5497         read_extent_buffer(leaf, device->uuid, ptr, BTRFS_UUID_SIZE);
5498 }
5499
5500 static int open_seed_devices(struct btrfs_root *root, u8 *fsid)
5501 {
5502         struct btrfs_fs_devices *fs_devices;
5503         int ret;
5504
5505         BUG_ON(!mutex_is_locked(&uuid_mutex));
5506
5507         fs_devices = root->fs_info->fs_devices->seed;
5508         while (fs_devices) {
5509                 if (!memcmp(fs_devices->fsid, fsid, BTRFS_UUID_SIZE)) {
5510                         ret = 0;
5511                         goto out;
5512                 }
5513                 fs_devices = fs_devices->seed;
5514         }
5515
5516         fs_devices = find_fsid(fsid);
5517         if (!fs_devices) {
5518                 ret = -ENOENT;
5519                 goto out;
5520         }
5521
5522         fs_devices = clone_fs_devices(fs_devices);
5523         if (IS_ERR(fs_devices)) {
5524                 ret = PTR_ERR(fs_devices);
5525                 goto out;
5526         }
5527
5528         ret = __btrfs_open_devices(fs_devices, FMODE_READ,
5529                                    root->fs_info->bdev_holder);
5530         if (ret) {
5531                 free_fs_devices(fs_devices);
5532                 goto out;
5533         }
5534
5535         if (!fs_devices->seeding) {
5536                 __btrfs_close_devices(fs_devices);
5537                 free_fs_devices(fs_devices);
5538                 ret = -EINVAL;
5539                 goto out;
5540         }
5541
5542         fs_devices->seed = root->fs_info->fs_devices->seed;
5543         root->fs_info->fs_devices->seed = fs_devices;
5544 out:
5545         return ret;
5546 }
5547
5548 static int read_one_dev(struct btrfs_root *root,
5549                         struct extent_buffer *leaf,
5550                         struct btrfs_dev_item *dev_item)
5551 {
5552         struct btrfs_device *device;
5553         u64 devid;
5554         int ret;
5555         u8 fs_uuid[BTRFS_UUID_SIZE];
5556         u8 dev_uuid[BTRFS_UUID_SIZE];
5557
5558         devid = btrfs_device_id(leaf, dev_item);
5559         read_extent_buffer(leaf, dev_uuid,
5560                            (unsigned long)btrfs_device_uuid(dev_item),
5561                            BTRFS_UUID_SIZE);
5562         read_extent_buffer(leaf, fs_uuid,
5563                            (unsigned long)btrfs_device_fsid(dev_item),
5564                            BTRFS_UUID_SIZE);
5565
5566         if (memcmp(fs_uuid, root->fs_info->fsid, BTRFS_UUID_SIZE)) {
5567                 ret = open_seed_devices(root, fs_uuid);
5568                 if (ret && !btrfs_test_opt(root, DEGRADED))
5569                         return ret;
5570         }
5571
5572         device = btrfs_find_device(root->fs_info, devid, dev_uuid, fs_uuid);
5573         if (!device || !device->bdev) {
5574                 if (!btrfs_test_opt(root, DEGRADED))
5575                         return -EIO;
5576
5577                 if (!device) {
5578                         btrfs_warn(root->fs_info, "devid %llu missing",
5579                                 (unsigned long long)devid);
5580                         device = add_missing_dev(root, devid, dev_uuid);
5581                         if (!device)
5582                                 return -ENOMEM;
5583                 } else if (!device->missing) {
5584                         /*
5585                          * this happens when a device that was properly setup
5586                          * in the device info lists suddenly goes bad.
5587                          * device->bdev is NULL, and so we have to set
5588                          * device->missing to one here
5589                          */
5590                         root->fs_info->fs_devices->missing_devices++;
5591                         device->missing = 1;
5592                 }
5593         }
5594
5595         if (device->fs_devices != root->fs_info->fs_devices) {
5596                 BUG_ON(device->writeable);
5597                 if (device->generation !=
5598                     btrfs_device_generation(leaf, dev_item))
5599                         return -EINVAL;
5600         }
5601
5602         fill_device_from_item(leaf, dev_item, device);
5603         device->dev_root = root->fs_info->dev_root;
5604         device->in_fs_metadata = 1;
5605         if (device->writeable && !device->is_tgtdev_for_dev_replace) {
5606                 device->fs_devices->total_rw_bytes += device->total_bytes;
5607                 spin_lock(&root->fs_info->free_chunk_lock);
5608                 root->fs_info->free_chunk_space += device->total_bytes -
5609                         device->bytes_used;
5610                 spin_unlock(&root->fs_info->free_chunk_lock);
5611         }
5612         ret = 0;
5613         return ret;
5614 }
5615
5616 int btrfs_read_sys_array(struct btrfs_root *root)
5617 {
5618         struct btrfs_super_block *super_copy = root->fs_info->super_copy;
5619         struct extent_buffer *sb;
5620         struct btrfs_disk_key *disk_key;
5621         struct btrfs_chunk *chunk;
5622         u8 *ptr;
5623         unsigned long sb_ptr;
5624         int ret = 0;
5625         u32 num_stripes;
5626         u32 array_size;
5627         u32 len = 0;
5628         u32 cur;
5629         struct btrfs_key key;
5630
5631         sb = btrfs_find_create_tree_block(root, BTRFS_SUPER_INFO_OFFSET,
5632                                           BTRFS_SUPER_INFO_SIZE);
5633         if (!sb)
5634                 return -ENOMEM;
5635         btrfs_set_buffer_uptodate(sb);
5636         btrfs_set_buffer_lockdep_class(root->root_key.objectid, sb, 0);
5637         /*
5638          * The sb extent buffer is artifical and just used to read the system array.
5639          * btrfs_set_buffer_uptodate() call does not properly mark all it's
5640          * pages up-to-date when the page is larger: extent does not cover the
5641          * whole page and consequently check_page_uptodate does not find all
5642          * the page's extents up-to-date (the hole beyond sb),
5643          * write_extent_buffer then triggers a WARN_ON.
5644          *
5645          * Regular short extents go through mark_extent_buffer_dirty/writeback cycle,
5646          * but sb spans only this function. Add an explicit SetPageUptodate call
5647          * to silence the warning eg. on PowerPC 64.
5648          */
5649         if (PAGE_CACHE_SIZE > BTRFS_SUPER_INFO_SIZE)
5650                 SetPageUptodate(sb->pages[0]);
5651
5652         write_extent_buffer(sb, super_copy, 0, BTRFS_SUPER_INFO_SIZE);
5653         array_size = btrfs_super_sys_array_size(super_copy);
5654
5655         ptr = super_copy->sys_chunk_array;
5656         sb_ptr = offsetof(struct btrfs_super_block, sys_chunk_array);
5657         cur = 0;
5658
5659         while (cur < array_size) {
5660                 disk_key = (struct btrfs_disk_key *)ptr;
5661                 btrfs_disk_key_to_cpu(&key, disk_key);
5662
5663                 len = sizeof(*disk_key); ptr += len;
5664                 sb_ptr += len;
5665                 cur += len;
5666
5667                 if (key.type == BTRFS_CHUNK_ITEM_KEY) {
5668                         chunk = (struct btrfs_chunk *)sb_ptr;
5669                         ret = read_one_chunk(root, &key, sb, chunk);
5670                         if (ret)
5671                                 break;
5672                         num_stripes = btrfs_chunk_num_stripes(sb, chunk);
5673                         len = btrfs_chunk_item_size(num_stripes);
5674                 } else {
5675                         ret = -EIO;
5676                         break;
5677                 }
5678                 ptr += len;
5679                 sb_ptr += len;
5680                 cur += len;
5681         }
5682         free_extent_buffer(sb);
5683         return ret;
5684 }
5685
5686 int btrfs_read_chunk_tree(struct btrfs_root *root)
5687 {
5688         struct btrfs_path *path;
5689         struct extent_buffer *leaf;
5690         struct btrfs_key key;
5691         struct btrfs_key found_key;
5692         int ret;
5693         int slot;
5694
5695         root = root->fs_info->chunk_root;
5696
5697         path = btrfs_alloc_path();
5698         if (!path)
5699                 return -ENOMEM;
5700
5701         mutex_lock(&uuid_mutex);
5702         lock_chunks(root);
5703
5704         /* first we search for all of the device items, and then we
5705          * read in all of the chunk items.  This way we can create chunk
5706          * mappings that reference all of the devices that are afound
5707          */
5708         key.objectid = BTRFS_DEV_ITEMS_OBJECTID;
5709         key.offset = 0;
5710         key.type = 0;
5711 again:
5712         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5713         if (ret < 0)
5714                 goto error;
5715         while (1) {
5716                 leaf = path->nodes[0];
5717                 slot = path->slots[0];
5718                 if (slot >= btrfs_header_nritems(leaf)) {
5719                         ret = btrfs_next_leaf(root, path);
5720                         if (ret == 0)
5721                                 continue;
5722                         if (ret < 0)
5723                                 goto error;
5724                         break;
5725                 }
5726                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5727                 if (key.objectid == BTRFS_DEV_ITEMS_OBJECTID) {
5728                         if (found_key.objectid != BTRFS_DEV_ITEMS_OBJECTID)
5729                                 break;
5730                         if (found_key.type == BTRFS_DEV_ITEM_KEY) {
5731                                 struct btrfs_dev_item *dev_item;
5732                                 dev_item = btrfs_item_ptr(leaf, slot,
5733                                                   struct btrfs_dev_item);
5734                                 ret = read_one_dev(root, leaf, dev_item);
5735                                 if (ret)
5736                                         goto error;
5737                         }
5738                 } else if (found_key.type == BTRFS_CHUNK_ITEM_KEY) {
5739                         struct btrfs_chunk *chunk;
5740                         chunk = btrfs_item_ptr(leaf, slot, struct btrfs_chunk);
5741                         ret = read_one_chunk(root, &found_key, leaf, chunk);
5742                         if (ret)
5743                                 goto error;
5744                 }
5745                 path->slots[0]++;
5746         }
5747         if (key.objectid == BTRFS_DEV_ITEMS_OBJECTID) {
5748                 key.objectid = 0;
5749                 btrfs_release_path(path);
5750                 goto again;
5751         }
5752         ret = 0;
5753 error:
5754         unlock_chunks(root);
5755         mutex_unlock(&uuid_mutex);
5756
5757         btrfs_free_path(path);
5758         return ret;
5759 }
5760
5761 static void __btrfs_reset_dev_stats(struct btrfs_device *dev)
5762 {
5763         int i;
5764
5765         for (i = 0; i < BTRFS_DEV_STAT_VALUES_MAX; i++)
5766                 btrfs_dev_stat_reset(dev, i);
5767 }
5768
5769 int btrfs_init_dev_stats(struct btrfs_fs_info *fs_info)
5770 {
5771         struct btrfs_key key;
5772         struct btrfs_key found_key;
5773         struct btrfs_root *dev_root = fs_info->dev_root;
5774         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
5775         struct extent_buffer *eb;
5776         int slot;
5777         int ret = 0;
5778         struct btrfs_device *device;
5779         struct btrfs_path *path = NULL;
5780         int i;
5781
5782         path = btrfs_alloc_path();
5783         if (!path) {
5784                 ret = -ENOMEM;
5785                 goto out;
5786         }
5787
5788         mutex_lock(&fs_devices->device_list_mutex);
5789         list_for_each_entry(device, &fs_devices->devices, dev_list) {
5790                 int item_size;
5791                 struct btrfs_dev_stats_item *ptr;
5792
5793                 key.objectid = 0;
5794                 key.type = BTRFS_DEV_STATS_KEY;
5795                 key.offset = device->devid;
5796                 ret = btrfs_search_slot(NULL, dev_root, &key, path, 0, 0);
5797                 if (ret) {
5798                         __btrfs_reset_dev_stats(device);
5799                         device->dev_stats_valid = 1;
5800                         btrfs_release_path(path);
5801                         continue;
5802                 }
5803                 slot = path->slots[0];
5804                 eb = path->nodes[0];
5805                 btrfs_item_key_to_cpu(eb, &found_key, slot);
5806                 item_size = btrfs_item_size_nr(eb, slot);
5807
5808                 ptr = btrfs_item_ptr(eb, slot,
5809                                      struct btrfs_dev_stats_item);
5810
5811                 for (i = 0; i < BTRFS_DEV_STAT_VALUES_MAX; i++) {
5812                         if (item_size >= (1 + i) * sizeof(__le64))
5813                                 btrfs_dev_stat_set(device, i,
5814                                         btrfs_dev_stats_value(eb, ptr, i));
5815                         else
5816                                 btrfs_dev_stat_reset(device, i);
5817                 }
5818
5819                 device->dev_stats_valid = 1;
5820                 btrfs_dev_stat_print_on_load(device);
5821                 btrfs_release_path(path);
5822         }
5823         mutex_unlock(&fs_devices->device_list_mutex);
5824
5825 out:
5826         btrfs_free_path(path);
5827         return ret < 0 ? ret : 0;
5828 }
5829
5830 static int update_dev_stat_item(struct btrfs_trans_handle *trans,
5831                                 struct btrfs_root *dev_root,
5832                                 struct btrfs_device *device)
5833 {
5834         struct btrfs_path *path;
5835         struct btrfs_key key;
5836         struct extent_buffer *eb;
5837         struct btrfs_dev_stats_item *ptr;
5838         int ret;
5839         int i;
5840
5841         key.objectid = 0;
5842         key.type = BTRFS_DEV_STATS_KEY;
5843         key.offset = device->devid;
5844
5845         path = btrfs_alloc_path();
5846         BUG_ON(!path);
5847         ret = btrfs_search_slot(trans, dev_root, &key, path, -1, 1);
5848         if (ret < 0) {
5849                 printk_in_rcu(KERN_WARNING "btrfs: error %d while searching for dev_stats item for device %s!\n",
5850                               ret, rcu_str_deref(device->name));
5851                 goto out;
5852         }
5853
5854         if (ret == 0 &&
5855             btrfs_item_size_nr(path->nodes[0], path->slots[0]) < sizeof(*ptr)) {
5856                 /* need to delete old one and insert a new one */
5857                 ret = btrfs_del_item(trans, dev_root, path);
5858                 if (ret != 0) {
5859                         printk_in_rcu(KERN_WARNING "btrfs: delete too small dev_stats item for device %s failed %d!\n",
5860                                       rcu_str_deref(device->name), ret);
5861                         goto out;
5862                 }
5863                 ret = 1;
5864         }
5865
5866         if (ret == 1) {
5867                 /* need to insert a new item */
5868                 btrfs_release_path(path);
5869                 ret = btrfs_insert_empty_item(trans, dev_root, path,
5870                                               &key, sizeof(*ptr));
5871                 if (ret < 0) {
5872                         printk_in_rcu(KERN_WARNING "btrfs: insert dev_stats item for device %s failed %d!\n",
5873                                       rcu_str_deref(device->name), ret);
5874                         goto out;
5875                 }
5876         }
5877
5878         eb = path->nodes[0];
5879         ptr = btrfs_item_ptr(eb, path->slots[0], struct btrfs_dev_stats_item);
5880         for (i = 0; i < BTRFS_DEV_STAT_VALUES_MAX; i++)
5881                 btrfs_set_dev_stats_value(eb, ptr, i,
5882                                           btrfs_dev_stat_read(device, i));
5883         btrfs_mark_buffer_dirty(eb);
5884
5885 out:
5886         btrfs_free_path(path);
5887         return ret;
5888 }
5889
5890 /*
5891  * called from commit_transaction. Writes all changed device stats to disk.
5892  */
5893 int btrfs_run_dev_stats(struct btrfs_trans_handle *trans,
5894                         struct btrfs_fs_info *fs_info)
5895 {
5896         struct btrfs_root *dev_root = fs_info->dev_root;
5897         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
5898         struct btrfs_device *device;
5899         int ret = 0;
5900
5901         mutex_lock(&fs_devices->device_list_mutex);
5902         list_for_each_entry(device, &fs_devices->devices, dev_list) {
5903                 if (!device->dev_stats_valid || !device->dev_stats_dirty)
5904                         continue;
5905
5906                 ret = update_dev_stat_item(trans, dev_root, device);
5907                 if (!ret)
5908                         device->dev_stats_dirty = 0;
5909         }
5910         mutex_unlock(&fs_devices->device_list_mutex);
5911
5912         return ret;
5913 }
5914
5915 void btrfs_dev_stat_inc_and_print(struct btrfs_device *dev, int index)
5916 {
5917         btrfs_dev_stat_inc(dev, index);
5918         btrfs_dev_stat_print_on_error(dev);
5919 }
5920
5921 void btrfs_dev_stat_print_on_error(struct btrfs_device *dev)
5922 {
5923         if (!dev->dev_stats_valid)
5924                 return;
5925         printk_ratelimited_in_rcu(KERN_ERR
5926                            "btrfs: bdev %s errs: wr %u, rd %u, flush %u, corrupt %u, gen %u\n",
5927                            rcu_str_deref(dev->name),
5928                            btrfs_dev_stat_read(dev, BTRFS_DEV_STAT_WRITE_ERRS),
5929                            btrfs_dev_stat_read(dev, BTRFS_DEV_STAT_READ_ERRS),
5930                            btrfs_dev_stat_read(dev, BTRFS_DEV_STAT_FLUSH_ERRS),
5931                            btrfs_dev_stat_read(dev,
5932                                                BTRFS_DEV_STAT_CORRUPTION_ERRS),
5933                            btrfs_dev_stat_read(dev,
5934                                                BTRFS_DEV_STAT_GENERATION_ERRS));
5935 }
5936
5937 static void btrfs_dev_stat_print_on_load(struct btrfs_device *dev)
5938 {
5939         int i;
5940
5941         for (i = 0; i < BTRFS_DEV_STAT_VALUES_MAX; i++)
5942                 if (btrfs_dev_stat_read(dev, i) != 0)
5943                         break;
5944         if (i == BTRFS_DEV_STAT_VALUES_MAX)
5945                 return; /* all values == 0, suppress message */
5946
5947         printk_in_rcu(KERN_INFO "btrfs: bdev %s errs: wr %u, rd %u, flush %u, corrupt %u, gen %u\n",
5948                rcu_str_deref(dev->name),
5949                btrfs_dev_stat_read(dev, BTRFS_DEV_STAT_WRITE_ERRS),
5950                btrfs_dev_stat_read(dev, BTRFS_DEV_STAT_READ_ERRS),
5951                btrfs_dev_stat_read(dev, BTRFS_DEV_STAT_FLUSH_ERRS),
5952                btrfs_dev_stat_read(dev, BTRFS_DEV_STAT_CORRUPTION_ERRS),
5953                btrfs_dev_stat_read(dev, BTRFS_DEV_STAT_GENERATION_ERRS));
5954 }
5955
5956 int btrfs_get_dev_stats(struct btrfs_root *root,
5957                         struct btrfs_ioctl_get_dev_stats *stats)
5958 {
5959         struct btrfs_device *dev;
5960         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
5961         int i;
5962
5963         mutex_lock(&fs_devices->device_list_mutex);
5964         dev = btrfs_find_device(root->fs_info, stats->devid, NULL, NULL);
5965         mutex_unlock(&fs_devices->device_list_mutex);
5966
5967         if (!dev) {
5968                 printk(KERN_WARNING
5969                        "btrfs: get dev_stats failed, device not found\n");
5970                 return -ENODEV;
5971         } else if (!dev->dev_stats_valid) {
5972                 printk(KERN_WARNING
5973                        "btrfs: get dev_stats failed, not yet valid\n");
5974                 return -ENODEV;
5975         } else if (stats->flags & BTRFS_DEV_STATS_RESET) {
5976                 for (i = 0; i < BTRFS_DEV_STAT_VALUES_MAX; i++) {
5977                         if (stats->nr_items > i)
5978                                 stats->values[i] =
5979                                         btrfs_dev_stat_read_and_reset(dev, i);
5980                         else
5981                                 btrfs_dev_stat_reset(dev, i);
5982                 }
5983         } else {
5984                 for (i = 0; i < BTRFS_DEV_STAT_VALUES_MAX; i++)
5985                         if (stats->nr_items > i)
5986                                 stats->values[i] = btrfs_dev_stat_read(dev, i);
5987         }
5988         if (stats->nr_items > BTRFS_DEV_STAT_VALUES_MAX)
5989                 stats->nr_items = BTRFS_DEV_STAT_VALUES_MAX;
5990         return 0;
5991 }
5992
5993 int btrfs_scratch_superblock(struct btrfs_device *device)
5994 {
5995         struct buffer_head *bh;
5996         struct btrfs_super_block *disk_super;
5997
5998         bh = btrfs_read_dev_super(device->bdev);
5999         if (!bh)
6000                 return -EINVAL;
6001         disk_super = (struct btrfs_super_block *)bh->b_data;
6002
6003         memset(&disk_super->magic, 0, sizeof(disk_super->magic));
6004         set_buffer_dirty(bh);
6005         sync_dirty_buffer(bh);
6006         brelse(bh);
6007
6008         return 0;
6009 }