2 * drivers/video/tegra/host/gk20a/channel_gk20a.c
4 * GK20A Graphics channel
6 * Copyright (c) 2011-2014, NVIDIA CORPORATION. All rights reserved.
8 * This program is free software; you can redistribute it and/or modify it
9 * under the terms and conditions of the GNU General Public License,
10 * version 2, as published by the Free Software Foundation.
12 * This program is distributed in the hope it will be useful, but WITHOUT
13 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
17 * You should have received a copy of the GNU General Public License along with
18 * this program; if not, write to the Free Software Foundation, Inc.,
19 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
22 #include <linux/nvhost.h>
23 #include <linux/list.h>
24 #include <linux/delay.h>
25 #include <linux/highmem.h> /* need for nvmap.h*/
26 #include <trace/events/gk20a.h>
27 #include <linux/scatterlist.h>
28 #include <linux/file.h>
29 #include <linux/anon_inodes.h>
30 #include <linux/dma-buf.h>
32 #include "debug_gk20a.h"
35 #include "dbg_gpu_gk20a.h"
37 #include "hw_ram_gk20a.h"
38 #include "hw_fifo_gk20a.h"
39 #include "hw_pbdma_gk20a.h"
40 #include "hw_ccsr_gk20a.h"
41 #include "hw_ltc_gk20a.h"
43 #define NVMAP_HANDLE_PARAM_SIZE 1
45 static struct channel_gk20a *acquire_unused_channel(struct fifo_gk20a *f);
46 static void release_used_channel(struct fifo_gk20a *f, struct channel_gk20a *c);
48 static void free_priv_cmdbuf(struct channel_gk20a *c,
49 struct priv_cmd_entry *e);
50 static void recycle_priv_cmdbuf(struct channel_gk20a *c);
52 static int channel_gk20a_alloc_priv_cmdbuf(struct channel_gk20a *c);
53 static void channel_gk20a_free_priv_cmdbuf(struct channel_gk20a *c);
55 static int channel_gk20a_commit_userd(struct channel_gk20a *c);
56 static int channel_gk20a_setup_userd(struct channel_gk20a *c);
57 static int channel_gk20a_setup_ramfc(struct channel_gk20a *c,
58 u64 gpfifo_base, u32 gpfifo_entries);
60 static void channel_gk20a_bind(struct channel_gk20a *ch_gk20a);
61 static void channel_gk20a_unbind(struct channel_gk20a *ch_gk20a);
63 static int channel_gk20a_alloc_inst(struct gk20a *g,
64 struct channel_gk20a *ch);
65 static void channel_gk20a_free_inst(struct gk20a *g,
66 struct channel_gk20a *ch);
68 static int channel_gk20a_update_runlist(struct channel_gk20a *c,
70 static void gk20a_free_error_notifiers(struct channel_gk20a *ch);
72 static struct channel_gk20a *acquire_unused_channel(struct fifo_gk20a *f)
74 struct channel_gk20a *ch = NULL;
77 mutex_lock(&f->ch_inuse_mutex);
78 for (chid = 0; chid < f->num_channels; chid++) {
79 if (!f->channel[chid].in_use) {
80 f->channel[chid].in_use = true;
81 ch = &f->channel[chid];
85 mutex_unlock(&f->ch_inuse_mutex);
90 static void release_used_channel(struct fifo_gk20a *f, struct channel_gk20a *c)
92 mutex_lock(&f->ch_inuse_mutex);
93 f->channel[c->hw_chid].in_use = false;
94 mutex_unlock(&f->ch_inuse_mutex);
97 int channel_gk20a_commit_va(struct channel_gk20a *c)
106 inst_ptr = c->inst_block.cpuva;
110 addr = gk20a_mm_iova_addr(c->vm->pdes.sgt->sgl);
111 addr_lo = u64_lo32(addr >> 12);
112 addr_hi = u64_hi32(addr);
114 gk20a_dbg_info("pde pa=0x%llx addr_lo=0x%x addr_hi=0x%x",
115 (u64)addr, addr_lo, addr_hi);
117 gk20a_mem_wr32(inst_ptr, ram_in_page_dir_base_lo_w(),
118 ram_in_page_dir_base_target_vid_mem_f() |
119 ram_in_page_dir_base_vol_true_f() |
120 ram_in_page_dir_base_lo_f(addr_lo));
122 gk20a_mem_wr32(inst_ptr, ram_in_page_dir_base_hi_w(),
123 ram_in_page_dir_base_hi_f(addr_hi));
125 gk20a_mem_wr32(inst_ptr, ram_in_adr_limit_lo_w(),
126 u64_lo32(c->vm->va_limit) | 0xFFF);
128 gk20a_mem_wr32(inst_ptr, ram_in_adr_limit_hi_w(),
129 ram_in_adr_limit_hi_f(u64_hi32(c->vm->va_limit)));
134 static int channel_gk20a_commit_userd(struct channel_gk20a *c)
142 inst_ptr = c->inst_block.cpuva;
146 addr_lo = u64_lo32(c->userd_iova >> ram_userd_base_shift_v());
147 addr_hi = u64_hi32(c->userd_iova);
149 gk20a_dbg_info("channel %d : set ramfc userd 0x%16llx",
150 c->hw_chid, (u64)c->userd_iova);
152 gk20a_mem_wr32(inst_ptr, ram_in_ramfc_w() + ram_fc_userd_w(),
153 pbdma_userd_target_vid_mem_f() |
154 pbdma_userd_addr_f(addr_lo));
156 gk20a_mem_wr32(inst_ptr, ram_in_ramfc_w() + ram_fc_userd_hi_w(),
157 pbdma_userd_target_vid_mem_f() |
158 pbdma_userd_hi_addr_f(addr_hi));
163 static int channel_gk20a_set_schedule_params(struct channel_gk20a *c,
164 u32 timeslice_timeout)
168 int value = timeslice_timeout;
170 inst_ptr = c->inst_block.cpuva;
174 /* disable channel */
175 gk20a_writel(c->g, ccsr_channel_r(c->hw_chid),
176 gk20a_readl(c->g, ccsr_channel_r(c->hw_chid)) |
177 ccsr_channel_enable_clr_true_f());
179 /* preempt the channel */
180 WARN_ON(gk20a_fifo_preempt_channel(c->g, c->hw_chid));
182 /* value field is 8 bits long */
183 while (value >= 1 << 8) {
188 /* time slice register is only 18bits long */
189 if ((value << shift) >= 1<<19) {
190 pr_err("Requested timeslice value is clamped to 18 bits\n");
195 /* set new timeslice */
196 gk20a_mem_wr32(inst_ptr, ram_fc_eng_timeslice_w(),
197 value | (shift << 12) |
198 fifo_eng_timeslice_enable_true_f());
201 gk20a_writel(c->g, ccsr_channel_r(c->hw_chid),
202 gk20a_readl(c->g, ccsr_channel_r(c->hw_chid)) |
203 ccsr_channel_enable_set_true_f());
208 static int channel_gk20a_setup_ramfc(struct channel_gk20a *c,
209 u64 gpfifo_base, u32 gpfifo_entries)
215 inst_ptr = c->inst_block.cpuva;
219 memset(inst_ptr, 0, ram_fc_size_val_v());
221 gk20a_mem_wr32(inst_ptr, ram_fc_gp_base_w(),
222 pbdma_gp_base_offset_f(
223 u64_lo32(gpfifo_base >> pbdma_gp_base_rsvd_s())));
225 gk20a_mem_wr32(inst_ptr, ram_fc_gp_base_hi_w(),
226 pbdma_gp_base_hi_offset_f(u64_hi32(gpfifo_base)) |
227 pbdma_gp_base_hi_limit2_f(ilog2(gpfifo_entries)));
229 gk20a_mem_wr32(inst_ptr, ram_fc_signature_w(),
230 pbdma_signature_hw_valid_f() | pbdma_signature_sw_zero_f());
232 gk20a_mem_wr32(inst_ptr, ram_fc_formats_w(),
233 pbdma_formats_gp_fermi0_f() |
234 pbdma_formats_pb_fermi1_f() |
235 pbdma_formats_mp_fermi0_f());
237 gk20a_mem_wr32(inst_ptr, ram_fc_pb_header_w(),
238 pbdma_pb_header_priv_user_f() |
239 pbdma_pb_header_method_zero_f() |
240 pbdma_pb_header_subchannel_zero_f() |
241 pbdma_pb_header_level_main_f() |
242 pbdma_pb_header_first_true_f() |
243 pbdma_pb_header_type_inc_f());
245 gk20a_mem_wr32(inst_ptr, ram_fc_subdevice_w(),
246 pbdma_subdevice_id_f(1) |
247 pbdma_subdevice_status_active_f() |
248 pbdma_subdevice_channel_dma_enable_f());
250 gk20a_mem_wr32(inst_ptr, ram_fc_target_w(), pbdma_target_engine_sw_f());
252 gk20a_mem_wr32(inst_ptr, ram_fc_acquire_w(),
253 pbdma_acquire_retry_man_2_f() |
254 pbdma_acquire_retry_exp_2_f() |
255 pbdma_acquire_timeout_exp_max_f() |
256 pbdma_acquire_timeout_man_max_f() |
257 pbdma_acquire_timeout_en_disable_f());
259 gk20a_mem_wr32(inst_ptr, ram_fc_eng_timeslice_w(),
260 fifo_eng_timeslice_timeout_128_f() |
261 fifo_eng_timeslice_timescale_3_f() |
262 fifo_eng_timeslice_enable_true_f());
264 gk20a_mem_wr32(inst_ptr, ram_fc_pb_timeslice_w(),
265 fifo_pb_timeslice_timeout_16_f() |
266 fifo_pb_timeslice_timescale_0_f() |
267 fifo_pb_timeslice_enable_true_f());
269 gk20a_mem_wr32(inst_ptr, ram_fc_chid_w(), ram_fc_chid_id_f(c->hw_chid));
274 static int channel_gk20a_setup_userd(struct channel_gk20a *c)
276 BUG_ON(!c->userd_cpu_va);
280 gk20a_mem_wr32(c->userd_cpu_va, ram_userd_put_w(), 0);
281 gk20a_mem_wr32(c->userd_cpu_va, ram_userd_get_w(), 0);
282 gk20a_mem_wr32(c->userd_cpu_va, ram_userd_ref_w(), 0);
283 gk20a_mem_wr32(c->userd_cpu_va, ram_userd_put_hi_w(), 0);
284 gk20a_mem_wr32(c->userd_cpu_va, ram_userd_ref_threshold_w(), 0);
285 gk20a_mem_wr32(c->userd_cpu_va, ram_userd_gp_top_level_get_w(), 0);
286 gk20a_mem_wr32(c->userd_cpu_va, ram_userd_gp_top_level_get_hi_w(), 0);
287 gk20a_mem_wr32(c->userd_cpu_va, ram_userd_get_hi_w(), 0);
288 gk20a_mem_wr32(c->userd_cpu_va, ram_userd_gp_get_w(), 0);
289 gk20a_mem_wr32(c->userd_cpu_va, ram_userd_gp_put_w(), 0);
294 static void channel_gk20a_bind(struct channel_gk20a *ch_gk20a)
296 struct gk20a *g = ch_gk20a->g;
297 struct fifo_gk20a *f = &g->fifo;
298 struct fifo_engine_info_gk20a *engine_info =
299 f->engine_info + ENGINE_GR_GK20A;
301 u32 inst_ptr = ch_gk20a->inst_block.cpu_pa
302 >> ram_in_base_shift_v();
304 gk20a_dbg_info("bind channel %d inst ptr 0x%08x",
305 ch_gk20a->hw_chid, inst_ptr);
307 ch_gk20a->bound = true;
309 gk20a_writel(g, ccsr_channel_r(ch_gk20a->hw_chid),
310 (gk20a_readl(g, ccsr_channel_r(ch_gk20a->hw_chid)) &
311 ~ccsr_channel_runlist_f(~0)) |
312 ccsr_channel_runlist_f(engine_info->runlist_id));
314 gk20a_writel(g, ccsr_channel_inst_r(ch_gk20a->hw_chid),
315 ccsr_channel_inst_ptr_f(inst_ptr) |
316 ccsr_channel_inst_target_vid_mem_f() |
317 ccsr_channel_inst_bind_true_f());
319 gk20a_writel(g, ccsr_channel_r(ch_gk20a->hw_chid),
320 (gk20a_readl(g, ccsr_channel_r(ch_gk20a->hw_chid)) &
321 ~ccsr_channel_enable_set_f(~0)) |
322 ccsr_channel_enable_set_true_f());
325 static void channel_gk20a_unbind(struct channel_gk20a *ch_gk20a)
327 struct gk20a *g = ch_gk20a->g;
332 gk20a_writel(g, ccsr_channel_inst_r(ch_gk20a->hw_chid),
333 ccsr_channel_inst_ptr_f(0) |
334 ccsr_channel_inst_bind_false_f());
336 ch_gk20a->bound = false;
339 * if we are agrressive then we can destroy the syncpt
340 * resource at this point
341 * if not, then it will be destroyed at channel_free()
343 if (ch_gk20a->sync && ch_gk20a->sync->syncpt_aggressive_destroy) {
344 ch_gk20a->sync->destroy(ch_gk20a->sync);
345 ch_gk20a->sync = NULL;
349 static int channel_gk20a_alloc_inst(struct gk20a *g,
350 struct channel_gk20a *ch)
352 struct device *d = dev_from_gk20a(g);
358 ch->inst_block.size = ram_in_alloc_size_v();
359 ch->inst_block.cpuva = dma_alloc_coherent(d,
363 if (!ch->inst_block.cpuva) {
364 gk20a_err(d, "%s: memory allocation failed\n", __func__);
369 ch->inst_block.iova = iova;
370 ch->inst_block.cpu_pa = gk20a_get_phys_from_iova(d,
371 ch->inst_block.iova);
372 if (!ch->inst_block.cpu_pa) {
373 gk20a_err(d, "%s: failed to get physical address\n", __func__);
378 gk20a_dbg_info("channel %d inst block physical addr: 0x%16llx",
379 ch->hw_chid, (u64)ch->inst_block.cpu_pa);
381 gk20a_dbg_fn("done");
385 gk20a_err(d, "fail");
386 channel_gk20a_free_inst(g, ch);
390 static void channel_gk20a_free_inst(struct gk20a *g,
391 struct channel_gk20a *ch)
393 struct device *d = dev_from_gk20a(g);
395 if (ch->inst_block.cpuva)
396 dma_free_coherent(d, ch->inst_block.size,
397 ch->inst_block.cpuva, ch->inst_block.iova);
398 ch->inst_block.cpuva = NULL;
399 ch->inst_block.iova = 0;
400 memset(&ch->inst_block, 0, sizeof(struct inst_desc));
403 static int channel_gk20a_update_runlist(struct channel_gk20a *c, bool add)
405 return gk20a_fifo_update_runlist(c->g, 0, c->hw_chid, add, true);
408 void gk20a_disable_channel_no_update(struct channel_gk20a *ch)
410 /* ensure no fences are pending */
412 ch->sync->set_min_eq_max(ch->sync);
414 /* disable channel */
415 gk20a_writel(ch->g, ccsr_channel_r(ch->hw_chid),
417 ccsr_channel_r(ch->hw_chid)) |
418 ccsr_channel_enable_clr_true_f());
421 int gk20a_wait_channel_idle(struct channel_gk20a *ch)
423 bool channel_idle = false;
424 unsigned long end_jiffies = jiffies +
425 msecs_to_jiffies(gk20a_get_gr_idle_timeout(ch->g));
428 mutex_lock(&ch->jobs_lock);
429 channel_idle = list_empty(&ch->jobs);
430 mutex_unlock(&ch->jobs_lock);
434 usleep_range(1000, 3000);
435 } while (time_before(jiffies, end_jiffies)
436 || !tegra_platform_is_silicon());
439 gk20a_err(dev_from_gk20a(ch->g), "jobs not freed for channel %d\n",
447 void gk20a_disable_channel(struct channel_gk20a *ch,
449 unsigned long finish_timeout)
452 int err = gk20a_channel_finish(ch, finish_timeout);
456 /* disable the channel from hw and increment syncpoints */
457 gk20a_disable_channel_no_update(ch);
459 gk20a_wait_channel_idle(ch);
461 /* preempt the channel */
462 gk20a_fifo_preempt_channel(ch->g, ch->hw_chid);
464 /* remove channel from runlist */
465 channel_gk20a_update_runlist(ch, false);
468 #if defined(CONFIG_GK20A_CYCLE_STATS)
470 static void gk20a_free_cycle_stats_buffer(struct channel_gk20a *ch)
472 /* disable existing cyclestats buffer */
473 mutex_lock(&ch->cyclestate.cyclestate_buffer_mutex);
474 if (ch->cyclestate.cyclestate_buffer_handler) {
475 dma_buf_vunmap(ch->cyclestate.cyclestate_buffer_handler,
476 ch->cyclestate.cyclestate_buffer);
477 dma_buf_put(ch->cyclestate.cyclestate_buffer_handler);
478 ch->cyclestate.cyclestate_buffer_handler = NULL;
479 ch->cyclestate.cyclestate_buffer = NULL;
480 ch->cyclestate.cyclestate_buffer_size = 0;
482 mutex_unlock(&ch->cyclestate.cyclestate_buffer_mutex);
485 static int gk20a_channel_cycle_stats(struct channel_gk20a *ch,
486 struct nvhost_cycle_stats_args *args)
488 struct dma_buf *dmabuf;
489 void *virtual_address;
491 if (args->nvmap_handle && !ch->cyclestate.cyclestate_buffer_handler) {
493 /* set up new cyclestats buffer */
494 dmabuf = dma_buf_get(args->nvmap_handle);
496 return PTR_ERR(dmabuf);
497 virtual_address = dma_buf_vmap(dmabuf);
498 if (!virtual_address)
501 ch->cyclestate.cyclestate_buffer_handler = dmabuf;
502 ch->cyclestate.cyclestate_buffer = virtual_address;
503 ch->cyclestate.cyclestate_buffer_size = dmabuf->size;
506 } else if (!args->nvmap_handle &&
507 ch->cyclestate.cyclestate_buffer_handler) {
508 gk20a_free_cycle_stats_buffer(ch);
511 } else if (!args->nvmap_handle &&
512 !ch->cyclestate.cyclestate_buffer_handler) {
513 /* no requst from GL */
517 pr_err("channel already has cyclestats buffer\n");
523 static int gk20a_init_error_notifier(struct channel_gk20a *ch,
524 struct nvhost_set_error_notifier *args) {
527 struct dma_buf *dmabuf;
530 pr_err("gk20a_init_error_notifier: invalid memory handle\n");
534 dmabuf = dma_buf_get(args->mem);
536 if (ch->error_notifier_ref)
537 gk20a_free_error_notifiers(ch);
539 if (IS_ERR(dmabuf)) {
540 pr_err("Invalid handle: %d\n", args->mem);
544 va = dma_buf_vmap(dmabuf);
547 pr_err("Cannot map notifier handle\n");
551 /* set channel notifiers pointer */
552 ch->error_notifier_ref = dmabuf;
553 ch->error_notifier = va + args->offset;
554 ch->error_notifier_va = va;
555 memset(ch->error_notifier, 0, sizeof(struct nvhost_notification));
559 void gk20a_set_error_notifier(struct channel_gk20a *ch, __u32 error)
561 if (ch->error_notifier_ref) {
562 struct timespec time_data;
564 getnstimeofday(&time_data);
565 nsec = ((u64)time_data.tv_sec) * 1000000000u +
566 (u64)time_data.tv_nsec;
567 ch->error_notifier->time_stamp.nanoseconds[0] =
569 ch->error_notifier->time_stamp.nanoseconds[1] =
571 ch->error_notifier->info32 = error;
572 ch->error_notifier->status = 0xffff;
573 gk20a_err(dev_from_gk20a(ch->g),
574 "error notifier set to %d\n", error);
578 static void gk20a_free_error_notifiers(struct channel_gk20a *ch)
580 if (ch->error_notifier_ref) {
581 dma_buf_vunmap(ch->error_notifier_ref, ch->error_notifier_va);
582 dma_buf_put(ch->error_notifier_ref);
583 ch->error_notifier_ref = 0;
584 ch->error_notifier = 0;
585 ch->error_notifier_va = 0;
589 void gk20a_free_channel(struct channel_gk20a *ch, bool finish)
591 struct gk20a *g = ch->g;
592 struct device *d = dev_from_gk20a(g);
593 struct fifo_gk20a *f = &g->fifo;
594 struct gr_gk20a *gr = &g->gr;
595 struct vm_gk20a *ch_vm = ch->vm;
596 unsigned long timeout = gk20a_get_gr_idle_timeout(g);
597 struct dbg_session_gk20a *dbg_s;
601 /* if engine reset was deferred, perform it now */
602 mutex_lock(&f->deferred_reset_mutex);
603 if (g->fifo.deferred_reset_pending) {
604 gk20a_dbg(gpu_dbg_intr | gpu_dbg_gpu_dbg, "engine reset was"
605 " deferred, running now");
606 fifo_gk20a_finish_mmu_fault_handling(g, g->fifo.mmu_fault_engines);
607 g->fifo.mmu_fault_engines = 0;
608 g->fifo.deferred_reset_pending = false;
610 mutex_unlock(&f->deferred_reset_mutex);
615 if (!gk20a_channel_as_bound(ch))
618 gk20a_dbg_info("freeing bound channel context, timeout=%ld",
621 gk20a_disable_channel(ch, finish && !ch->has_timedout, timeout);
623 gk20a_free_error_notifiers(ch);
625 /* release channel ctx */
626 gk20a_free_channel_ctx(ch);
628 gk20a_gr_flush_channel_tlb(gr);
630 memset(&ch->ramfc, 0, sizeof(struct mem_desc_sub));
633 if (ch->gpfifo.gpu_va)
634 gk20a_gmmu_unmap(ch_vm, ch->gpfifo.gpu_va,
635 ch->gpfifo.size, gk20a_mem_flag_none);
636 if (ch->gpfifo.cpu_va)
637 dma_free_coherent(d, ch->gpfifo.size,
638 ch->gpfifo.cpu_va, ch->gpfifo.iova);
639 ch->gpfifo.cpu_va = NULL;
642 memset(&ch->gpfifo, 0, sizeof(struct gpfifo_desc));
644 #if defined(CONFIG_GK20A_CYCLE_STATS)
645 gk20a_free_cycle_stats_buffer(ch);
648 channel_gk20a_free_priv_cmdbuf(ch);
651 ch->sync->destroy(ch->sync);
655 /* release channel binding to the as_share */
656 gk20a_as_release_share(ch_vm->as_share);
659 channel_gk20a_unbind(ch);
660 channel_gk20a_free_inst(g, ch);
666 /* unlink all debug sessions */
667 mutex_lock(&ch->dbg_s_lock);
669 list_for_each_entry(dbg_s, &ch->dbg_s_list, dbg_s_list_node) {
671 list_del_init(&dbg_s->dbg_s_list_node);
674 mutex_unlock(&ch->dbg_s_lock);
677 release_used_channel(f, ch);
680 int gk20a_channel_release(struct inode *inode, struct file *filp)
682 struct channel_gk20a *ch = (struct channel_gk20a *)filp->private_data;
683 struct gk20a *g = ch->g;
686 trace_gk20a_channel_release(dev_name(&g->dev->dev));
688 err = gk20a_busy(ch->g->dev);
690 gk20a_err(dev_from_gk20a(g), "failed to release channel %d",
694 gk20a_free_channel(ch, true);
695 gk20a_idle(ch->g->dev);
698 filp->private_data = NULL;
702 static struct channel_gk20a *gk20a_open_new_channel(struct gk20a *g)
704 struct fifo_gk20a *f = &g->fifo;
705 struct channel_gk20a *ch;
707 ch = acquire_unused_channel(f);
709 /* TBD: we want to make this virtualizable */
710 gk20a_err(dev_from_gk20a(g), "out of hw chids");
716 if (channel_gk20a_alloc_inst(g, ch)) {
718 gk20a_err(dev_from_gk20a(g),
719 "failed to open gk20a channel, out of inst mem");
723 g->ops.fifo.bind_channel(ch);
724 ch->pid = current->pid;
726 /* reset timeout counter and update timestamp */
727 ch->timeout_accumulated_ms = 0;
728 ch->timeout_gpfifo_get = 0;
729 /* set gr host default timeout */
730 ch->timeout_ms_max = gk20a_get_gr_idle_timeout(g);
731 ch->timeout_debug_dump = true;
732 ch->has_timedout = false;
735 /* The channel is *not* runnable at this point. It still needs to have
736 * an address space bound and allocate a gpfifo and grctx. */
738 init_waitqueue_head(&ch->notifier_wq);
739 init_waitqueue_head(&ch->semaphore_wq);
740 init_waitqueue_head(&ch->submit_wq);
745 static int __gk20a_channel_open(struct gk20a *g, struct file *filp)
748 struct channel_gk20a *ch;
750 trace_gk20a_channel_open(dev_name(&g->dev->dev));
752 err = gk20a_get_client(g);
754 gk20a_err(dev_from_gk20a(g),
755 "failed to get client ref");
759 err = gk20a_busy(g->dev);
762 gk20a_err(dev_from_gk20a(g), "failed to power on, %d", err);
765 ch = gk20a_open_new_channel(g);
769 gk20a_err(dev_from_gk20a(g),
774 filp->private_data = ch;
778 int gk20a_channel_open(struct inode *inode, struct file *filp)
780 struct gk20a *g = container_of(inode->i_cdev,
781 struct gk20a, channel.cdev);
782 return __gk20a_channel_open(g, filp);
785 /* allocate private cmd buffer.
786 used for inserting commands before/after user submitted buffers. */
787 static int channel_gk20a_alloc_priv_cmdbuf(struct channel_gk20a *c)
789 struct device *d = dev_from_gk20a(c->g);
790 struct vm_gk20a *ch_vm = c->vm;
791 struct priv_cmd_queue *q = &c->priv_cmd_q;
792 struct priv_cmd_entry *e;
795 struct sg_table *sgt;
798 /* Kernel can insert gpfifos before and after user gpfifos.
799 Before user gpfifos, kernel inserts fence_wait, which takes
800 syncpoint_a (2 dwords) + syncpoint_b (2 dwords) = 4 dwords.
801 After user gpfifos, kernel inserts fence_get, which takes
802 wfi (2 dwords) + syncpoint_a (2 dwords) + syncpoint_b (2 dwords)
804 Worse case if kernel adds both of them for every user gpfifo,
805 max size of priv_cmdbuf is :
806 (gpfifo entry number * (2 / 3) * (4 + 6) * 4 bytes */
807 size = roundup_pow_of_two(
808 c->gpfifo.entry_num * 2 * 10 * sizeof(u32) / 3);
810 q->mem.base_cpuva = dma_alloc_coherent(d, size,
813 if (!q->mem.base_cpuva) {
814 gk20a_err(d, "%s: memory allocation failed\n", __func__);
819 q->mem.base_iova = iova;
822 err = gk20a_get_sgtable(d, &sgt,
823 q->mem.base_cpuva, q->mem.base_iova, size);
825 gk20a_err(d, "%s: failed to create sg table\n", __func__);
829 memset(q->mem.base_cpuva, 0, size);
831 q->base_gpuva = gk20a_gmmu_map(ch_vm, &sgt,
834 gk20a_mem_flag_none);
835 if (!q->base_gpuva) {
836 gk20a_err(d, "ch %d : failed to map gpu va"
837 "for priv cmd buffer", c->hw_chid);
842 q->size = q->mem.size / sizeof (u32);
844 INIT_LIST_HEAD(&q->head);
845 INIT_LIST_HEAD(&q->free);
847 /* pre-alloc 25% of priv cmdbuf entries and put them on free list */
848 for (i = 0; i < q->size / 4; i++) {
849 e = kzalloc(sizeof(struct priv_cmd_entry), GFP_KERNEL);
851 gk20a_err(d, "ch %d: fail to pre-alloc cmd entry",
857 list_add(&e->list, &q->free);
860 gk20a_free_sgtable(&sgt);
865 gk20a_free_sgtable(&sgt);
867 channel_gk20a_free_priv_cmdbuf(c);
871 static void channel_gk20a_free_priv_cmdbuf(struct channel_gk20a *c)
873 struct device *d = dev_from_gk20a(c->g);
874 struct vm_gk20a *ch_vm = c->vm;
875 struct priv_cmd_queue *q = &c->priv_cmd_q;
876 struct priv_cmd_entry *e;
877 struct list_head *pos, *tmp, *head;
883 gk20a_gmmu_unmap(ch_vm, q->base_gpuva,
884 q->mem.size, gk20a_mem_flag_none);
885 if (q->mem.base_cpuva)
886 dma_free_coherent(d, q->mem.size,
887 q->mem.base_cpuva, q->mem.base_iova);
888 q->mem.base_cpuva = NULL;
889 q->mem.base_iova = 0;
893 list_for_each_safe(pos, tmp, head) {
894 e = container_of(pos, struct priv_cmd_entry, list);
895 free_priv_cmdbuf(c, e);
900 list_for_each_safe(pos, tmp, head) {
901 e = container_of(pos, struct priv_cmd_entry, list);
902 e->pre_alloc = false;
903 free_priv_cmdbuf(c, e);
906 memset(q, 0, sizeof(struct priv_cmd_queue));
909 /* allocate a cmd buffer with given size. size is number of u32 entries */
910 int gk20a_channel_alloc_priv_cmdbuf(struct channel_gk20a *c, u32 orig_size,
911 struct priv_cmd_entry **entry)
913 struct priv_cmd_queue *q = &c->priv_cmd_q;
914 struct priv_cmd_entry *e;
915 struct list_head *node;
917 u32 size = orig_size;
918 bool no_retry = false;
920 gk20a_dbg_fn("size %d", orig_size);
924 /* if free space in the end is less than requested, increase the size
925 * to make the real allocated space start from beginning. */
926 if (q->put + size > q->size)
927 size = orig_size + (q->size - q->put);
929 gk20a_dbg_info("ch %d: priv cmd queue get:put %d:%d",
930 c->hw_chid, q->get, q->put);
933 free_count = (q->size - (q->put - q->get) - 1) % q->size;
935 if (size > free_count) {
937 recycle_priv_cmdbuf(c);
944 if (unlikely(list_empty(&q->free))) {
946 gk20a_dbg_info("ch %d: run out of pre-alloc entries",
949 e = kzalloc(sizeof(struct priv_cmd_entry), GFP_KERNEL);
951 gk20a_err(dev_from_gk20a(c->g),
952 "ch %d: fail to allocate priv cmd entry",
959 e = container_of(node, struct priv_cmd_entry, list);
963 e->gp_get = c->gpfifo.get;
964 e->gp_put = c->gpfifo.put;
965 e->gp_wrap = c->gpfifo.wrap;
967 /* if we have increased size to skip free space in the end, set put
968 to beginning of cmd buffer (0) + size */
969 if (size != orig_size) {
970 e->ptr = q->mem.base_cpuva;
971 e->gva = q->base_gpuva;
974 e->ptr = q->mem.base_cpuva + q->put;
975 e->gva = q->base_gpuva + q->put * sizeof(u32);
976 q->put = (q->put + orig_size) & (q->size - 1);
979 /* we already handled q->put + size > q->size so BUG_ON this */
980 BUG_ON(q->put > q->size);
982 /* add new entry to head since we free from head */
983 list_add(&e->list, &q->head);
987 gk20a_dbg_fn("done");
992 /* Don't call this to free an explict cmd entry.
993 * It doesn't update priv_cmd_queue get/put */
994 static void free_priv_cmdbuf(struct channel_gk20a *c,
995 struct priv_cmd_entry *e)
997 struct priv_cmd_queue *q = &c->priv_cmd_q;
1004 if (unlikely(!e->pre_alloc))
1007 memset(e, 0, sizeof(struct priv_cmd_entry));
1008 e->pre_alloc = true;
1009 list_add(&e->list, &q->free);
1013 /* free entries if they're no longer being used */
1014 static void recycle_priv_cmdbuf(struct channel_gk20a *c)
1016 struct priv_cmd_queue *q = &c->priv_cmd_q;
1017 struct priv_cmd_entry *e, *tmp;
1018 struct list_head *head = &q->head;
1019 bool wrap_around, found = false;
1023 /* Find the most recent free entry. Free it and everything before it */
1024 list_for_each_entry(e, head, list) {
1026 gk20a_dbg_info("ch %d: cmd entry get:put:wrap %d:%d:%d "
1027 "curr get:put:wrap %d:%d:%d",
1028 c->hw_chid, e->gp_get, e->gp_put, e->gp_wrap,
1029 c->gpfifo.get, c->gpfifo.put, c->gpfifo.wrap);
1031 wrap_around = (c->gpfifo.wrap != e->gp_wrap);
1032 if (e->gp_get < e->gp_put) {
1033 if (c->gpfifo.get >= e->gp_put ||
1038 e->gp_get = c->gpfifo.get;
1039 } else if (e->gp_get > e->gp_put) {
1041 c->gpfifo.get >= e->gp_put) {
1045 e->gp_get = c->gpfifo.get;
1050 q->get = (e->ptr - q->mem.base_cpuva) + e->size;
1052 gk20a_dbg_info("no free entry recycled");
1056 list_for_each_entry_safe_continue(e, tmp, head, list) {
1057 free_priv_cmdbuf(c, e);
1060 gk20a_dbg_fn("done");
1064 static int gk20a_alloc_channel_gpfifo(struct channel_gk20a *c,
1065 struct nvhost_alloc_gpfifo_args *args)
1067 struct gk20a *g = c->g;
1068 struct device *d = dev_from_gk20a(g);
1069 struct vm_gk20a *ch_vm;
1072 struct sg_table *sgt;
1075 /* Kernel can insert one extra gpfifo entry before user submitted gpfifos
1076 and another one after, for internal usage. Triple the requested size. */
1077 gpfifo_size = roundup_pow_of_two(args->num_entries * 3);
1079 if (args->flags & NVHOST_ALLOC_GPFIFO_FLAGS_VPR_ENABLED)
1082 /* an address space needs to have been bound at this point. */
1083 if (!gk20a_channel_as_bound(c)) {
1085 "not bound to an address space at time of gpfifo"
1086 " allocation. Attempting to create and bind to"
1092 c->cmds_pending = false;
1093 c->last_submit_fence.valid = false;
1095 c->ramfc.offset = 0;
1096 c->ramfc.size = ram_in_ramfc_s() / 8;
1098 if (c->gpfifo.cpu_va) {
1099 gk20a_err(d, "channel %d :"
1100 "gpfifo already allocated", c->hw_chid);
1104 c->gpfifo.size = gpfifo_size * sizeof(struct gpfifo);
1105 c->gpfifo.cpu_va = (struct gpfifo *)dma_alloc_coherent(d,
1109 if (!c->gpfifo.cpu_va) {
1110 gk20a_err(d, "%s: memory allocation failed\n", __func__);
1115 c->gpfifo.iova = iova;
1116 c->gpfifo.entry_num = gpfifo_size;
1118 c->gpfifo.get = c->gpfifo.put = 0;
1120 err = gk20a_get_sgtable(d, &sgt,
1121 c->gpfifo.cpu_va, c->gpfifo.iova, c->gpfifo.size);
1123 gk20a_err(d, "%s: failed to allocate sg table\n", __func__);
1127 c->gpfifo.gpu_va = gk20a_gmmu_map(ch_vm,
1131 gk20a_mem_flag_none);
1132 if (!c->gpfifo.gpu_va) {
1133 gk20a_err(d, "channel %d : failed to map"
1134 " gpu_va for gpfifo", c->hw_chid);
1139 gk20a_dbg_info("channel %d : gpfifo_base 0x%016llx, size %d",
1140 c->hw_chid, c->gpfifo.gpu_va, c->gpfifo.entry_num);
1142 channel_gk20a_setup_ramfc(c, c->gpfifo.gpu_va, c->gpfifo.entry_num);
1144 channel_gk20a_setup_userd(c);
1145 channel_gk20a_commit_userd(c);
1147 /* TBD: setup engine contexts */
1149 err = channel_gk20a_alloc_priv_cmdbuf(c);
1151 goto clean_up_unmap;
1153 err = channel_gk20a_update_runlist(c, true);
1155 goto clean_up_unmap;
1157 gk20a_free_sgtable(&sgt);
1159 gk20a_dbg_fn("done");
1163 gk20a_gmmu_unmap(ch_vm, c->gpfifo.gpu_va,
1164 c->gpfifo.size, gk20a_mem_flag_none);
1166 gk20a_free_sgtable(&sgt);
1168 dma_free_coherent(d, c->gpfifo.size,
1169 c->gpfifo.cpu_va, c->gpfifo.iova);
1170 c->gpfifo.cpu_va = NULL;
1172 memset(&c->gpfifo, 0, sizeof(struct gpfifo_desc));
1173 gk20a_err(d, "fail");
1177 static inline int wfi_cmd_size(void)
1181 void add_wfi_cmd(struct priv_cmd_entry *cmd, int *i)
1184 cmd->ptr[(*i)++] = 0x2001001E;
1185 /* handle, ignored */
1186 cmd->ptr[(*i)++] = 0x00000000;
1189 static inline bool check_gp_put(struct gk20a *g,
1190 struct channel_gk20a *c)
1193 /* gp_put changed unexpectedly since last update? */
1194 put = gk20a_bar1_readl(g,
1195 c->userd_gpu_va + 4 * ram_userd_gp_put_w());
1196 if (c->gpfifo.put != put) {
1197 /*TBD: BUG_ON/teardown on this*/
1198 gk20a_err(dev_from_gk20a(g), "gp_put changed unexpectedly "
1199 "since last update");
1200 c->gpfifo.put = put;
1201 return false; /* surprise! */
1203 return true; /* checked out ok */
1206 /* Update with this periodically to determine how the gpfifo is draining. */
1207 static inline u32 update_gp_get(struct gk20a *g,
1208 struct channel_gk20a *c)
1210 u32 new_get = gk20a_bar1_readl(g,
1211 c->userd_gpu_va + sizeof(u32) * ram_userd_gp_get_w());
1212 if (new_get < c->gpfifo.get)
1213 c->gpfifo.wrap = !c->gpfifo.wrap;
1214 c->gpfifo.get = new_get;
1218 static inline u32 gp_free_count(struct channel_gk20a *c)
1220 return (c->gpfifo.entry_num - (c->gpfifo.put - c->gpfifo.get) - 1) %
1221 c->gpfifo.entry_num;
1224 bool gk20a_channel_update_and_check_timeout(struct channel_gk20a *ch,
1225 u32 timeout_delta_ms)
1227 u32 gpfifo_get = update_gp_get(ch->g, ch);
1228 /* Count consequent timeout isr */
1229 if (gpfifo_get == ch->timeout_gpfifo_get) {
1230 /* we didn't advance since previous channel timeout check */
1231 ch->timeout_accumulated_ms += timeout_delta_ms;
1233 /* first timeout isr encountered */
1234 ch->timeout_accumulated_ms = timeout_delta_ms;
1237 ch->timeout_gpfifo_get = gpfifo_get;
1239 return ch->g->timeouts_enabled &&
1240 ch->timeout_accumulated_ms > ch->timeout_ms_max;
1244 /* Issue a syncpoint increment *preceded* by a wait-for-idle
1245 * command. All commands on the channel will have been
1246 * consumed at the time the fence syncpoint increment occurs.
1248 static int gk20a_channel_submit_wfi(struct channel_gk20a *c)
1250 struct priv_cmd_entry *cmd = NULL;
1251 struct gk20a *g = c->g;
1255 if (c->has_timedout)
1259 c->sync = gk20a_channel_sync_create(c);
1264 update_gp_get(g, c);
1265 free_count = gp_free_count(c);
1266 if (unlikely(!free_count)) {
1267 gk20a_err(dev_from_gk20a(g),
1268 "not enough gpfifo space");
1272 err = c->sync->incr_wfi(c->sync, &cmd, &c->last_submit_fence);
1276 WARN_ON(!c->last_submit_fence.wfi);
1278 c->gpfifo.cpu_va[c->gpfifo.put].entry0 = u64_lo32(cmd->gva);
1279 c->gpfifo.cpu_va[c->gpfifo.put].entry1 = u64_hi32(cmd->gva) |
1280 pbdma_gp_entry1_length_f(cmd->size);
1282 c->gpfifo.put = (c->gpfifo.put + 1) & (c->gpfifo.entry_num - 1);
1285 cmd->gp_put = c->gpfifo.put;
1287 gk20a_bar1_writel(g,
1288 c->userd_gpu_va + 4 * ram_userd_gp_put_w(),
1291 gk20a_dbg_info("post-submit put %d, get %d, size %d",
1292 c->gpfifo.put, c->gpfifo.get, c->gpfifo.entry_num);
1297 static u32 get_gp_free_count(struct channel_gk20a *c)
1299 update_gp_get(c->g, c);
1300 return gp_free_count(c);
1303 static void trace_write_pushbuffer(struct channel_gk20a *c, struct gpfifo *g)
1308 struct dma_buf *dmabuf = NULL;
1310 if (gk20a_debug_trace_cmdbuf) {
1311 u64 gpu_va = (u64)g->entry0 |
1312 (u64)((u64)pbdma_gp_entry1_get_hi_v(g->entry1) << 32);
1315 words = pbdma_gp_entry1_length_v(g->entry1);
1316 err = gk20a_vm_find_buffer(c->vm, gpu_va, &dmabuf, &offset);
1318 mem = dma_buf_vmap(dmabuf);
1324 * Write in batches of 128 as there seems to be a limit
1325 * of how much you can output to ftrace at once.
1327 for (i = 0; i < words; i += 128U) {
1328 trace_gk20a_push_cmdbuf(
1331 min(words - i, 128U),
1332 offset + i * sizeof(u32),
1335 dma_buf_vunmap(dmabuf, mem);
1339 static int gk20a_channel_add_job(struct channel_gk20a *c,
1340 struct gk20a_channel_fence *fence)
1342 struct vm_gk20a *vm = c->vm;
1343 struct channel_gk20a_job *job = NULL;
1344 struct mapped_buffer_node **mapped_buffers = NULL;
1345 int err = 0, num_mapped_buffers;
1347 /* job needs reference to this vm */
1350 err = gk20a_vm_get_buffers(vm, &mapped_buffers, &num_mapped_buffers);
1356 job = kzalloc(sizeof(*job), GFP_KERNEL);
1358 gk20a_vm_put_buffers(vm, mapped_buffers, num_mapped_buffers);
1363 job->num_mapped_buffers = num_mapped_buffers;
1364 job->mapped_buffers = mapped_buffers;
1365 job->fence = *fence;
1367 mutex_lock(&c->jobs_lock);
1368 list_add_tail(&job->list, &c->jobs);
1369 mutex_unlock(&c->jobs_lock);
1374 void gk20a_channel_update(struct channel_gk20a *c, int nr_completed)
1376 struct vm_gk20a *vm = c->vm;
1377 struct channel_gk20a_job *job, *n;
1379 wake_up(&c->submit_wq);
1381 mutex_lock(&c->jobs_lock);
1382 list_for_each_entry_safe(job, n, &c->jobs, list) {
1383 bool completed = WARN_ON(!c->sync) ||
1384 c->sync->is_expired(c->sync, &job->fence);
1388 gk20a_vm_put_buffers(vm, job->mapped_buffers,
1389 job->num_mapped_buffers);
1391 /* job is done. release its reference to vm */
1394 list_del_init(&job->list);
1396 gk20a_idle(c->g->dev);
1398 mutex_unlock(&c->jobs_lock);
1401 static int gk20a_submit_channel_gpfifo(struct channel_gk20a *c,
1402 struct nvhost_gpfifo *gpfifo,
1404 struct nvhost_fence *fence,
1407 struct gk20a *g = c->g;
1408 struct device *d = dev_from_gk20a(g);
1411 struct priv_cmd_entry *wait_cmd = NULL;
1412 struct priv_cmd_entry *incr_cmd = NULL;
1413 /* we might need two extra gpfifo entries - one for pre fence
1414 * and one for post fence. */
1415 const int extra_entries = 2;
1416 bool need_wfi = !(flags & NVHOST_SUBMIT_GPFIFO_FLAGS_SUPPRESS_WFI);
1418 if (c->has_timedout)
1421 if ((flags & (NVHOST_SUBMIT_GPFIFO_FLAGS_FENCE_WAIT |
1422 NVHOST_SUBMIT_GPFIFO_FLAGS_FENCE_GET)) &&
1427 c->sync = gk20a_channel_sync_create(c);
1432 #ifdef CONFIG_DEBUG_FS
1433 /* update debug settings */
1434 if (g->ops.ltc.sync_debugfs)
1435 g->ops.ltc.sync_debugfs(g);
1438 gk20a_dbg_info("channel %d", c->hw_chid);
1440 /* gk20a_channel_update releases this ref. */
1441 err = gk20a_busy(g->dev);
1443 gk20a_err(d, "failed to host gk20a to submit gpfifo");
1447 trace_gk20a_channel_submit_gpfifo(c->g->dev->name,
1451 fence->syncpt_id, fence->value);
1453 update_gp_get(g, c);
1455 gk20a_dbg_info("pre-submit put %d, get %d, size %d",
1456 c->gpfifo.put, c->gpfifo.get, c->gpfifo.entry_num);
1458 /* Invalidate tlb if it's dirty... */
1459 /* TBD: this should be done in the cmd stream, not with PRIs. */
1460 /* We don't know what context is currently running... */
1461 /* Note also: there can be more than one context associated with the */
1462 /* address space (vm). */
1463 gk20a_mm_tlb_invalidate(c->vm);
1465 /* Make sure we have enough space for gpfifo entries. If not,
1466 * wait for signals from completed submits */
1467 if (gp_free_count(c) < num_entries + extra_entries) {
1468 err = wait_event_interruptible(c->submit_wq,
1469 get_gp_free_count(c) >= num_entries + extra_entries ||
1473 if (c->has_timedout) {
1479 gk20a_err(d, "not enough gpfifo space");
1485 * optionally insert syncpt wait in the beginning of gpfifo submission
1486 * when user requested and the wait hasn't expired.
1487 * validate that the id makes sense, elide if not
1488 * the only reason this isn't being unceremoniously killed is to
1489 * keep running some tests which trigger this condition
1491 if (flags & NVHOST_SUBMIT_GPFIFO_FLAGS_FENCE_WAIT) {
1492 if (flags & NVHOST_SUBMIT_GPFIFO_FLAGS_SYNC_FENCE)
1493 err = c->sync->wait_fd(c->sync, fence->syncpt_id,
1496 err = c->sync->wait_syncpt(c->sync, fence->syncpt_id,
1497 fence->value, &wait_cmd);
1503 /* always insert syncpt increment at end of gpfifo submission
1504 to keep track of method completion for idle railgating */
1505 if (flags & NVHOST_SUBMIT_GPFIFO_FLAGS_FENCE_GET &&
1506 flags & NVHOST_SUBMIT_GPFIFO_FLAGS_SYNC_FENCE)
1507 err = c->sync->incr_user_fd(c->sync, &incr_cmd,
1508 &c->last_submit_fence,
1511 else if (flags & NVHOST_SUBMIT_GPFIFO_FLAGS_FENCE_GET)
1512 err = c->sync->incr_user_syncpt(c->sync, &incr_cmd,
1513 &c->last_submit_fence,
1518 err = c->sync->incr(c->sync, &incr_cmd,
1519 &c->last_submit_fence);
1524 c->gpfifo.cpu_va[c->gpfifo.put].entry0 =
1525 u64_lo32(wait_cmd->gva);
1526 c->gpfifo.cpu_va[c->gpfifo.put].entry1 =
1527 u64_hi32(wait_cmd->gva) |
1528 pbdma_gp_entry1_length_f(wait_cmd->size);
1529 trace_gk20a_push_cmdbuf(c->g->dev->name,
1530 0, wait_cmd->size, 0, wait_cmd->ptr);
1532 c->gpfifo.put = (c->gpfifo.put + 1) &
1533 (c->gpfifo.entry_num - 1);
1536 wait_cmd->gp_put = c->gpfifo.put;
1539 for (i = 0; i < num_entries; i++) {
1540 c->gpfifo.cpu_va[c->gpfifo.put].entry0 =
1541 gpfifo[i].entry0; /* cmd buf va low 32 */
1542 c->gpfifo.cpu_va[c->gpfifo.put].entry1 =
1543 gpfifo[i].entry1; /* cmd buf va high 32 | words << 10 */
1544 trace_write_pushbuffer(c, &c->gpfifo.cpu_va[c->gpfifo.put]);
1545 c->gpfifo.put = (c->gpfifo.put + 1) &
1546 (c->gpfifo.entry_num - 1);
1550 c->gpfifo.cpu_va[c->gpfifo.put].entry0 =
1551 u64_lo32(incr_cmd->gva);
1552 c->gpfifo.cpu_va[c->gpfifo.put].entry1 =
1553 u64_hi32(incr_cmd->gva) |
1554 pbdma_gp_entry1_length_f(incr_cmd->size);
1555 trace_gk20a_push_cmdbuf(c->g->dev->name,
1556 0, incr_cmd->size, 0, incr_cmd->ptr);
1558 c->gpfifo.put = (c->gpfifo.put + 1) &
1559 (c->gpfifo.entry_num - 1);
1562 incr_cmd->gp_put = c->gpfifo.put;
1565 trace_gk20a_channel_submitted_gpfifo(c->g->dev->name,
1569 fence->syncpt_id, fence->value);
1571 /* TODO! Check for errors... */
1572 gk20a_channel_add_job(c, &c->last_submit_fence);
1574 c->cmds_pending = true;
1575 gk20a_bar1_writel(g,
1576 c->userd_gpu_va + 4 * ram_userd_gp_put_w(),
1579 gk20a_dbg_info("post-submit put %d, get %d, size %d",
1580 c->gpfifo.put, c->gpfifo.get, c->gpfifo.entry_num);
1582 gk20a_dbg_fn("done");
1586 gk20a_err(d, "fail");
1587 free_priv_cmdbuf(c, wait_cmd);
1588 free_priv_cmdbuf(c, incr_cmd);
1593 void gk20a_remove_channel_support(struct channel_gk20a *c)
1598 int gk20a_init_channel_support(struct gk20a *g, u32 chid)
1600 struct channel_gk20a *c = g->fifo.channel+chid;
1605 c->remove_support = gk20a_remove_channel_support;
1606 mutex_init(&c->jobs_lock);
1607 INIT_LIST_HEAD(&c->jobs);
1608 #if defined(CONFIG_GK20A_CYCLE_STATS)
1609 mutex_init(&c->cyclestate.cyclestate_buffer_mutex);
1611 INIT_LIST_HEAD(&c->dbg_s_list);
1612 mutex_init(&c->dbg_s_lock);
1617 int gk20a_channel_finish(struct channel_gk20a *ch, unsigned long timeout)
1621 if (!ch->cmds_pending)
1624 /* Do not wait for a timedout channel */
1625 if (ch->has_timedout)
1628 if (!(ch->last_submit_fence.valid && ch->last_submit_fence.wfi)) {
1629 gk20a_dbg_fn("issuing wfi, incr to finish the channel");
1630 err = gk20a_channel_submit_wfi(ch);
1635 BUG_ON(!(ch->last_submit_fence.valid && ch->last_submit_fence.wfi));
1637 gk20a_dbg_fn("waiting for channel to finish thresh:%d",
1638 ch->last_submit_fence.thresh);
1640 err = ch->sync->wait_cpu(ch->sync, &ch->last_submit_fence, timeout);
1642 dev_warn(dev_from_gk20a(ch->g),
1643 "timed out waiting for gk20a channel to finish");
1645 ch->cmds_pending = false;
1650 static int gk20a_channel_wait_semaphore(struct channel_gk20a *ch,
1651 ulong id, u32 offset,
1652 u32 payload, long timeout)
1654 struct platform_device *pdev = ch->g->dev;
1655 struct dma_buf *dmabuf;
1661 /* do not wait if channel has timed out */
1662 if (ch->has_timedout)
1665 dmabuf = dma_buf_get(id);
1666 if (IS_ERR(dmabuf)) {
1667 gk20a_err(&pdev->dev, "invalid notifier nvmap handle 0x%lx",
1672 data = dma_buf_kmap(dmabuf, offset >> PAGE_SHIFT);
1674 gk20a_err(&pdev->dev, "failed to map notifier memory");
1679 semaphore = data + (offset & ~PAGE_MASK);
1681 remain = wait_event_interruptible_timeout(
1683 *semaphore == payload || ch->has_timedout,
1686 if (remain == 0 && *semaphore != payload)
1688 else if (remain < 0)
1691 dma_buf_kunmap(dmabuf, offset >> PAGE_SHIFT, data);
1693 dma_buf_put(dmabuf);
1697 static int gk20a_channel_wait(struct channel_gk20a *ch,
1698 struct nvhost_wait_args *args)
1700 struct device *d = dev_from_gk20a(ch->g);
1701 struct dma_buf *dmabuf;
1702 struct notification *notif;
1707 unsigned long timeout;
1708 int remain, ret = 0;
1712 if (ch->has_timedout)
1715 if (args->timeout == NVHOST_NO_TIMEOUT)
1716 timeout = MAX_SCHEDULE_TIMEOUT;
1718 timeout = (u32)msecs_to_jiffies(args->timeout);
1720 switch (args->type) {
1721 case NVHOST_WAIT_TYPE_NOTIFIER:
1722 id = args->condition.notifier.nvmap_handle;
1723 offset = args->condition.notifier.offset;
1725 dmabuf = dma_buf_get(id);
1726 if (IS_ERR(dmabuf)) {
1727 gk20a_err(d, "invalid notifier nvmap handle 0x%lx",
1732 notif = dma_buf_vmap(dmabuf);
1734 gk20a_err(d, "failed to map notifier memory");
1738 notif = (struct notification *)((uintptr_t)notif + offset);
1740 /* user should set status pending before
1741 * calling this ioctl */
1742 remain = wait_event_interruptible_timeout(
1744 notif->status == 0 || ch->has_timedout,
1747 if (remain == 0 && notif->status != 0) {
1749 goto notif_clean_up;
1750 } else if (remain < 0) {
1752 goto notif_clean_up;
1755 /* TBD: fill in correct information */
1756 jiffies = get_jiffies_64();
1757 jiffies_to_timespec(jiffies, &tv);
1758 notif->timestamp.nanoseconds[0] = tv.tv_nsec;
1759 notif->timestamp.nanoseconds[1] = tv.tv_sec;
1760 notif->info32 = 0xDEADBEEF; /* should be object name */
1761 notif->info16 = ch->hw_chid; /* should be method offset */
1764 dma_buf_vunmap(dmabuf, notif);
1767 case NVHOST_WAIT_TYPE_SEMAPHORE:
1768 ret = gk20a_channel_wait_semaphore(ch,
1769 args->condition.semaphore.nvmap_handle,
1770 args->condition.semaphore.offset,
1771 args->condition.semaphore.payload,
1784 static int gk20a_channel_set_priority(struct channel_gk20a *ch,
1787 u32 timeslice_timeout;
1788 /* set priority of graphics channel */
1790 case NVHOST_PRIORITY_LOW:
1791 /* 64 << 3 = 512us */
1792 timeslice_timeout = 64;
1794 case NVHOST_PRIORITY_MEDIUM:
1795 /* 128 << 3 = 1024us */
1796 timeslice_timeout = 128;
1798 case NVHOST_PRIORITY_HIGH:
1799 /* 255 << 3 = 2048us */
1800 timeslice_timeout = 255;
1803 pr_err("Unsupported priority");
1806 channel_gk20a_set_schedule_params(ch,
1811 static int gk20a_channel_zcull_bind(struct channel_gk20a *ch,
1812 struct nvhost_zcull_bind_args *args)
1814 struct gk20a *g = ch->g;
1815 struct gr_gk20a *gr = &g->gr;
1819 return gr_gk20a_bind_ctxsw_zcull(g, gr, ch,
1820 args->gpu_va, args->mode);
1823 /* in this context the "channel" is the host1x channel which
1824 * maps to *all* gk20a channels */
1825 int gk20a_channel_suspend(struct gk20a *g)
1827 struct fifo_gk20a *f = &g->fifo;
1829 bool channels_in_use = false;
1834 /* wait for engine idle */
1835 err = gk20a_fifo_wait_engine_idle(g);
1839 for (chid = 0; chid < f->num_channels; chid++) {
1840 if (f->channel[chid].in_use) {
1842 gk20a_dbg_info("suspend channel %d", chid);
1843 /* disable channel */
1844 gk20a_writel(g, ccsr_channel_r(chid),
1845 gk20a_readl(g, ccsr_channel_r(chid)) |
1846 ccsr_channel_enable_clr_true_f());
1847 /* preempt the channel */
1848 gk20a_fifo_preempt_channel(g, chid);
1850 channels_in_use = true;
1854 if (channels_in_use) {
1855 gk20a_fifo_update_runlist(g, 0, ~0, false, true);
1857 for (chid = 0; chid < f->num_channels; chid++) {
1858 if (f->channel[chid].in_use)
1859 channel_gk20a_unbind(&f->channel[chid]);
1863 gk20a_dbg_fn("done");
1867 /* in this context the "channel" is the host1x channel which
1868 * maps to *all* gk20a channels */
1869 int gk20a_channel_resume(struct gk20a *g)
1871 struct fifo_gk20a *f = &g->fifo;
1873 bool channels_in_use = false;
1877 for (chid = 0; chid < f->num_channels; chid++) {
1878 if (f->channel[chid].in_use) {
1879 gk20a_dbg_info("resume channel %d", chid);
1880 g->ops.fifo.bind_channel(&f->channel[chid]);
1881 channels_in_use = true;
1885 if (channels_in_use)
1886 gk20a_fifo_update_runlist(g, 0, ~0, true, true);
1888 gk20a_dbg_fn("done");
1892 void gk20a_channel_semaphore_wakeup(struct gk20a *g)
1894 struct fifo_gk20a *f = &g->fifo;
1899 for (chid = 0; chid < f->num_channels; chid++) {
1900 struct channel_gk20a *c = g->fifo.channel+chid;
1902 wake_up_interruptible_all(&c->semaphore_wq);
1906 static int gk20a_ioctl_channel_submit_gpfifo(
1907 struct channel_gk20a *ch,
1908 struct nvhost_submit_gpfifo_args *args)
1916 if (ch->has_timedout)
1919 size = args->num_entries * sizeof(struct nvhost_gpfifo);
1921 gpfifo = kzalloc(size, GFP_KERNEL);
1925 if (copy_from_user(gpfifo,
1926 (void __user *)(uintptr_t)args->gpfifo, size)) {
1931 ret = gk20a_submit_channel_gpfifo(ch, gpfifo, args->num_entries,
1932 &args->fence, args->flags);
1939 void gk20a_init_channel(struct gpu_ops *gops)
1941 gops->fifo.bind_channel = channel_gk20a_bind;
1944 long gk20a_channel_ioctl(struct file *filp,
1945 unsigned int cmd, unsigned long arg)
1947 struct channel_gk20a *ch = filp->private_data;
1948 struct platform_device *dev = ch->g->dev;
1949 u8 buf[NVHOST_IOCTL_CHANNEL_MAX_ARG_SIZE];
1952 if ((_IOC_TYPE(cmd) != NVHOST_IOCTL_MAGIC) ||
1953 (_IOC_NR(cmd) == 0) ||
1954 (_IOC_NR(cmd) > NVHOST_IOCTL_CHANNEL_LAST) ||
1955 (_IOC_SIZE(cmd) > NVHOST_IOCTL_CHANNEL_MAX_ARG_SIZE))
1958 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1959 if (copy_from_user(buf, (void __user *)arg, _IOC_SIZE(cmd)))
1964 case NVHOST_IOCTL_CHANNEL_OPEN:
1970 err = get_unused_fd_flags(O_RDWR);
1975 name = kasprintf(GFP_KERNEL, "nvhost-%s-fd%d",
1976 dev_name(&dev->dev), fd);
1983 file = anon_inode_getfile(name, filp->f_op, NULL, O_RDWR);
1986 err = PTR_ERR(file);
1990 fd_install(fd, file);
1992 err = __gk20a_channel_open(ch->g, file);
1999 ((struct nvhost_channel_open_args *)buf)->channel_fd = fd;
2002 case NVHOST_IOCTL_CHANNEL_SET_NVMAP_FD:
2004 case NVHOST_IOCTL_CHANNEL_ALLOC_OBJ_CTX:
2005 err = gk20a_busy(dev);
2008 "%s: failed to host gk20a for ioctl cmd: 0x%x",
2012 err = gk20a_alloc_obj_ctx(ch,
2013 (struct nvhost_alloc_obj_ctx_args *)buf);
2016 case NVHOST_IOCTL_CHANNEL_FREE_OBJ_CTX:
2017 err = gk20a_busy(dev);
2020 "%s: failed to host gk20a for ioctl cmd: 0x%x",
2024 err = gk20a_free_obj_ctx(ch,
2025 (struct nvhost_free_obj_ctx_args *)buf);
2028 case NVHOST_IOCTL_CHANNEL_ALLOC_GPFIFO:
2029 err = gk20a_busy(dev);
2032 "%s: failed to host gk20a for ioctl cmd: 0x%x",
2036 err = gk20a_alloc_channel_gpfifo(ch,
2037 (struct nvhost_alloc_gpfifo_args *)buf);
2040 case NVHOST_IOCTL_CHANNEL_SUBMIT_GPFIFO:
2041 err = gk20a_ioctl_channel_submit_gpfifo(ch,
2042 (struct nvhost_submit_gpfifo_args *)buf);
2044 case NVHOST_IOCTL_CHANNEL_WAIT:
2045 err = gk20a_busy(dev);
2048 "%s: failed to host gk20a for ioctl cmd: 0x%x",
2052 err = gk20a_channel_wait(ch,
2053 (struct nvhost_wait_args *)buf);
2056 case NVHOST_IOCTL_CHANNEL_ZCULL_BIND:
2057 err = gk20a_busy(dev);
2060 "%s: failed to host gk20a for ioctl cmd: 0x%x",
2064 err = gk20a_channel_zcull_bind(ch,
2065 (struct nvhost_zcull_bind_args *)buf);
2068 case NVHOST_IOCTL_CHANNEL_SET_ERROR_NOTIFIER:
2069 err = gk20a_busy(dev);
2072 "%s: failed to host gk20a for ioctl cmd: 0x%x",
2076 err = gk20a_init_error_notifier(ch,
2077 (struct nvhost_set_error_notifier *)buf);
2080 #ifdef CONFIG_GK20A_CYCLE_STATS
2081 case NVHOST_IOCTL_CHANNEL_CYCLE_STATS:
2082 err = gk20a_busy(dev);
2085 "%s: failed to host gk20a for ioctl cmd: 0x%x",
2089 err = gk20a_channel_cycle_stats(ch,
2090 (struct nvhost_cycle_stats_args *)buf);
2094 case NVHOST_IOCTL_CHANNEL_SET_TIMEOUT:
2097 (u32)((struct nvhost_set_timeout_args *)buf)->timeout;
2098 gk20a_dbg(gpu_dbg_gpu_dbg, "setting timeout (%d ms) for chid %d",
2099 timeout, ch->hw_chid);
2100 ch->timeout_ms_max = timeout;
2103 case NVHOST_IOCTL_CHANNEL_SET_TIMEOUT_EX:
2106 (u32)((struct nvhost_set_timeout_args *)buf)->timeout;
2107 bool timeout_debug_dump = !((u32)
2108 ((struct nvhost_set_timeout_ex_args *)buf)->flags &
2109 (1 << NVHOST_TIMEOUT_FLAG_DISABLE_DUMP));
2110 gk20a_dbg(gpu_dbg_gpu_dbg, "setting timeout (%d ms) for chid %d",
2111 timeout, ch->hw_chid);
2112 ch->timeout_ms_max = timeout;
2113 ch->timeout_debug_dump = timeout_debug_dump;
2116 case NVHOST_IOCTL_CHANNEL_GET_TIMEDOUT:
2117 ((struct nvhost_get_param_args *)buf)->value =
2120 case NVHOST_IOCTL_CHANNEL_SET_PRIORITY:
2121 err = gk20a_busy(dev);
2124 "%s: failed to host gk20a for ioctl cmd: 0x%x",
2128 gk20a_channel_set_priority(ch,
2129 ((struct nvhost_set_priority_args *)buf)->priority);
2133 dev_err(&dev->dev, "unrecognized ioctl cmd: 0x%x", cmd);
2138 if ((err == 0) && (_IOC_DIR(cmd) & _IOC_READ))
2139 err = copy_to_user((void __user *)arg, buf, _IOC_SIZE(cmd));