2 * Copyright (c) 2009-2010 Chelsio, Inc. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 #include <linux/module.h>
37 static int db_delay_usecs = 1;
38 module_param(db_delay_usecs, int, 0644);
39 MODULE_PARM_DESC(db_delay_usecs, "Usecs to delay awaiting db fifo to drain");
41 static int ocqp_support = 1;
42 module_param(ocqp_support, int, 0644);
43 MODULE_PARM_DESC(ocqp_support, "Support on-chip SQs (default=1)");
45 int db_fc_threshold = 2000;
46 module_param(db_fc_threshold, int, 0644);
47 MODULE_PARM_DESC(db_fc_threshold, "QP count/threshold that triggers automatic "
48 "db flow control mode (default = 2000)");
50 static void set_state(struct c4iw_qp *qhp, enum c4iw_qp_state state)
53 spin_lock_irqsave(&qhp->lock, flag);
54 qhp->attr.state = state;
55 spin_unlock_irqrestore(&qhp->lock, flag);
58 static void dealloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
60 c4iw_ocqp_pool_free(rdev, sq->dma_addr, sq->memsize);
63 static void dealloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
65 dma_free_coherent(&(rdev->lldi.pdev->dev), sq->memsize, sq->queue,
66 pci_unmap_addr(sq, mapping));
69 static void dealloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
72 dealloc_oc_sq(rdev, sq);
74 dealloc_host_sq(rdev, sq);
77 static int alloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
79 if (!ocqp_support || !t4_ocqp_supported())
81 sq->dma_addr = c4iw_ocqp_pool_alloc(rdev, sq->memsize);
84 sq->phys_addr = rdev->oc_mw_pa + sq->dma_addr -
85 rdev->lldi.vr->ocq.start;
86 sq->queue = (__force union t4_wr *)(rdev->oc_mw_kva + sq->dma_addr -
87 rdev->lldi.vr->ocq.start);
88 sq->flags |= T4_SQ_ONCHIP;
92 static int alloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
94 sq->queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev), sq->memsize,
95 &(sq->dma_addr), GFP_KERNEL);
98 sq->phys_addr = virt_to_phys(sq->queue);
99 pci_unmap_addr_set(sq, mapping, sq->dma_addr);
103 static int alloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq, int user)
107 ret = alloc_oc_sq(rdev, sq);
109 ret = alloc_host_sq(rdev, sq);
113 static int destroy_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
114 struct c4iw_dev_ucontext *uctx)
117 * uP clears EQ contexts when the connection exits rdma mode,
118 * so no need to post a RESET WR for these EQs.
120 dma_free_coherent(&(rdev->lldi.pdev->dev),
121 wq->rq.memsize, wq->rq.queue,
122 dma_unmap_addr(&wq->rq, mapping));
123 dealloc_sq(rdev, &wq->sq);
124 c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
127 c4iw_put_qpid(rdev, wq->rq.qid, uctx);
128 c4iw_put_qpid(rdev, wq->sq.qid, uctx);
132 static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
133 struct t4_cq *rcq, struct t4_cq *scq,
134 struct c4iw_dev_ucontext *uctx)
136 int user = (uctx != &rdev->uctx);
137 struct fw_ri_res_wr *res_wr;
138 struct fw_ri_res *res;
140 struct c4iw_wr_wait wr_wait;
145 wq->sq.qid = c4iw_get_qpid(rdev, uctx);
149 wq->rq.qid = c4iw_get_qpid(rdev, uctx);
156 wq->sq.sw_sq = kzalloc(wq->sq.size * sizeof *wq->sq.sw_sq,
163 wq->rq.sw_rq = kzalloc(wq->rq.size * sizeof *wq->rq.sw_rq,
172 * RQT must be a power of 2.
174 wq->rq.rqt_size = roundup_pow_of_two(wq->rq.size);
175 wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size);
176 if (!wq->rq.rqt_hwaddr) {
181 ret = alloc_sq(rdev, &wq->sq, user);
184 memset(wq->sq.queue, 0, wq->sq.memsize);
185 dma_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
187 wq->rq.queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev),
188 wq->rq.memsize, &(wq->rq.dma_addr),
194 PDBG("%s sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx\n",
195 __func__, wq->sq.queue,
196 (unsigned long long)virt_to_phys(wq->sq.queue),
198 (unsigned long long)virt_to_phys(wq->rq.queue));
199 memset(wq->rq.queue, 0, wq->rq.memsize);
200 dma_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
202 wq->db = rdev->lldi.db_reg;
203 wq->gts = rdev->lldi.gts_reg;
205 wq->sq.udb = (u64)pci_resource_start(rdev->lldi.pdev, 2) +
206 (wq->sq.qid << rdev->qpshift);
207 wq->sq.udb &= PAGE_MASK;
208 wq->rq.udb = (u64)pci_resource_start(rdev->lldi.pdev, 2) +
209 (wq->rq.qid << rdev->qpshift);
210 wq->rq.udb &= PAGE_MASK;
215 /* build fw_ri_res_wr */
216 wr_len = sizeof *res_wr + 2 * sizeof *res;
218 skb = alloc_skb(wr_len, GFP_KERNEL);
223 set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
225 res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
226 memset(res_wr, 0, wr_len);
227 res_wr->op_nres = cpu_to_be32(
228 FW_WR_OP(FW_RI_RES_WR) |
229 V_FW_RI_RES_WR_NRES(2) |
231 res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
232 res_wr->cookie = (unsigned long) &wr_wait;
234 res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
235 res->u.sqrq.op = FW_RI_RES_OP_WRITE;
238 * eqsize is the number of 64B entries plus the status page size.
240 eqsize = wq->sq.size * T4_SQ_NUM_SLOTS + T4_EQ_STATUS_ENTRIES;
242 res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
243 V_FW_RI_RES_WR_HOSTFCMODE(0) | /* no host cidx updates */
244 V_FW_RI_RES_WR_CPRIO(0) | /* don't keep in chip cache */
245 V_FW_RI_RES_WR_PCIECHN(0) | /* set by uP at ri_init time */
246 (t4_sq_onchip(&wq->sq) ? F_FW_RI_RES_WR_ONCHIP : 0) |
247 V_FW_RI_RES_WR_IQID(scq->cqid));
248 res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
249 V_FW_RI_RES_WR_DCAEN(0) |
250 V_FW_RI_RES_WR_DCACPU(0) |
251 V_FW_RI_RES_WR_FBMIN(2) |
252 V_FW_RI_RES_WR_FBMAX(2) |
253 V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
254 V_FW_RI_RES_WR_CIDXFTHRESH(0) |
255 V_FW_RI_RES_WR_EQSIZE(eqsize));
256 res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
257 res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
259 res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
260 res->u.sqrq.op = FW_RI_RES_OP_WRITE;
263 * eqsize is the number of 64B entries plus the status page size.
265 eqsize = wq->rq.size * T4_RQ_NUM_SLOTS + T4_EQ_STATUS_ENTRIES;
266 res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
267 V_FW_RI_RES_WR_HOSTFCMODE(0) | /* no host cidx updates */
268 V_FW_RI_RES_WR_CPRIO(0) | /* don't keep in chip cache */
269 V_FW_RI_RES_WR_PCIECHN(0) | /* set by uP at ri_init time */
270 V_FW_RI_RES_WR_IQID(rcq->cqid));
271 res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
272 V_FW_RI_RES_WR_DCAEN(0) |
273 V_FW_RI_RES_WR_DCACPU(0) |
274 V_FW_RI_RES_WR_FBMIN(2) |
275 V_FW_RI_RES_WR_FBMAX(2) |
276 V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
277 V_FW_RI_RES_WR_CIDXFTHRESH(0) |
278 V_FW_RI_RES_WR_EQSIZE(eqsize));
279 res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
280 res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
282 c4iw_init_wr_wait(&wr_wait);
284 ret = c4iw_ofld_send(rdev, skb);
287 ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, wq->sq.qid, __func__);
291 PDBG("%s sqid 0x%x rqid 0x%x kdb 0x%p squdb 0x%llx rqudb 0x%llx\n",
292 __func__, wq->sq.qid, wq->rq.qid, wq->db,
293 (unsigned long long)wq->sq.udb, (unsigned long long)wq->rq.udb);
297 dma_free_coherent(&(rdev->lldi.pdev->dev),
298 wq->rq.memsize, wq->rq.queue,
299 dma_unmap_addr(&wq->rq, mapping));
301 dealloc_sq(rdev, &wq->sq);
303 c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
309 c4iw_put_qpid(rdev, wq->rq.qid, uctx);
311 c4iw_put_qpid(rdev, wq->sq.qid, uctx);
315 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
316 struct ib_send_wr *wr, int max, u32 *plenp)
323 dstp = (u8 *)immdp->data;
324 for (i = 0; i < wr->num_sge; i++) {
325 if ((plen + wr->sg_list[i].length) > max)
327 srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
328 plen += wr->sg_list[i].length;
329 rem = wr->sg_list[i].length;
331 if (dstp == (u8 *)&sq->queue[sq->size])
332 dstp = (u8 *)sq->queue;
333 if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
336 len = (u8 *)&sq->queue[sq->size] - dstp;
337 memcpy(dstp, srcp, len);
343 len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp);
345 memset(dstp, 0, len);
346 immdp->op = FW_RI_DATA_IMMD;
349 immdp->immdlen = cpu_to_be32(plen);
354 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
355 struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
356 int num_sge, u32 *plenp)
361 __be64 *flitp = (__be64 *)isglp->sge;
363 for (i = 0; i < num_sge; i++) {
364 if ((plen + sg_list[i].length) < plen)
366 plen += sg_list[i].length;
367 *flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
369 if (++flitp == queue_end)
371 *flitp = cpu_to_be64(sg_list[i].addr);
372 if (++flitp == queue_end)
375 *flitp = (__force __be64)0;
376 isglp->op = FW_RI_DATA_ISGL;
378 isglp->nsge = cpu_to_be16(num_sge);
385 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
386 struct ib_send_wr *wr, u8 *len16)
392 if (wr->num_sge > T4_MAX_SEND_SGE)
394 switch (wr->opcode) {
396 if (wr->send_flags & IB_SEND_SOLICITED)
397 wqe->send.sendop_pkd = cpu_to_be32(
398 V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE));
400 wqe->send.sendop_pkd = cpu_to_be32(
401 V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND));
402 wqe->send.stag_inv = 0;
404 case IB_WR_SEND_WITH_INV:
405 if (wr->send_flags & IB_SEND_SOLICITED)
406 wqe->send.sendop_pkd = cpu_to_be32(
407 V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE_INV));
409 wqe->send.sendop_pkd = cpu_to_be32(
410 V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_INV));
411 wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
420 if (wr->send_flags & IB_SEND_INLINE) {
421 ret = build_immd(sq, wqe->send.u.immd_src, wr,
422 T4_MAX_SEND_INLINE, &plen);
425 size = sizeof wqe->send + sizeof(struct fw_ri_immd) +
428 ret = build_isgl((__be64 *)sq->queue,
429 (__be64 *)&sq->queue[sq->size],
430 wqe->send.u.isgl_src,
431 wr->sg_list, wr->num_sge, &plen);
434 size = sizeof wqe->send + sizeof(struct fw_ri_isgl) +
435 wr->num_sge * sizeof(struct fw_ri_sge);
438 wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
439 wqe->send.u.immd_src[0].r1 = 0;
440 wqe->send.u.immd_src[0].r2 = 0;
441 wqe->send.u.immd_src[0].immdlen = 0;
442 size = sizeof wqe->send + sizeof(struct fw_ri_immd);
445 *len16 = DIV_ROUND_UP(size, 16);
446 wqe->send.plen = cpu_to_be32(plen);
450 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
451 struct ib_send_wr *wr, u8 *len16)
457 if (wr->num_sge > T4_MAX_SEND_SGE)
460 wqe->write.stag_sink = cpu_to_be32(wr->wr.rdma.rkey);
461 wqe->write.to_sink = cpu_to_be64(wr->wr.rdma.remote_addr);
463 if (wr->send_flags & IB_SEND_INLINE) {
464 ret = build_immd(sq, wqe->write.u.immd_src, wr,
465 T4_MAX_WRITE_INLINE, &plen);
468 size = sizeof wqe->write + sizeof(struct fw_ri_immd) +
471 ret = build_isgl((__be64 *)sq->queue,
472 (__be64 *)&sq->queue[sq->size],
473 wqe->write.u.isgl_src,
474 wr->sg_list, wr->num_sge, &plen);
477 size = sizeof wqe->write + sizeof(struct fw_ri_isgl) +
478 wr->num_sge * sizeof(struct fw_ri_sge);
481 wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
482 wqe->write.u.immd_src[0].r1 = 0;
483 wqe->write.u.immd_src[0].r2 = 0;
484 wqe->write.u.immd_src[0].immdlen = 0;
485 size = sizeof wqe->write + sizeof(struct fw_ri_immd);
488 *len16 = DIV_ROUND_UP(size, 16);
489 wqe->write.plen = cpu_to_be32(plen);
493 static int build_rdma_read(union t4_wr *wqe, struct ib_send_wr *wr, u8 *len16)
498 wqe->read.stag_src = cpu_to_be32(wr->wr.rdma.rkey);
499 wqe->read.to_src_hi = cpu_to_be32((u32)(wr->wr.rdma.remote_addr
501 wqe->read.to_src_lo = cpu_to_be32((u32)wr->wr.rdma.remote_addr);
502 wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
503 wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
504 wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
506 wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
508 wqe->read.stag_src = cpu_to_be32(2);
509 wqe->read.to_src_hi = 0;
510 wqe->read.to_src_lo = 0;
511 wqe->read.stag_sink = cpu_to_be32(2);
513 wqe->read.to_sink_hi = 0;
514 wqe->read.to_sink_lo = 0;
518 *len16 = DIV_ROUND_UP(sizeof wqe->read, 16);
522 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
523 struct ib_recv_wr *wr, u8 *len16)
527 ret = build_isgl((__be64 *)qhp->wq.rq.queue,
528 (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
529 &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
532 *len16 = DIV_ROUND_UP(sizeof wqe->recv +
533 wr->num_sge * sizeof(struct fw_ri_sge), 16);
537 static int build_fastreg(struct t4_sq *sq, union t4_wr *wqe,
538 struct ib_send_wr *wr, u8 *len16)
541 struct fw_ri_immd *imdp;
544 int pbllen = roundup(wr->wr.fast_reg.page_list_len * sizeof(u64), 32);
547 if (wr->wr.fast_reg.page_list_len > T4_MAX_FR_DEPTH)
550 wqe->fr.qpbinde_to_dcacpu = 0;
551 wqe->fr.pgsz_shift = wr->wr.fast_reg.page_shift - 12;
552 wqe->fr.addr_type = FW_RI_VA_BASED_TO;
553 wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->wr.fast_reg.access_flags);
555 wqe->fr.len_lo = cpu_to_be32(wr->wr.fast_reg.length);
556 wqe->fr.stag = cpu_to_be32(wr->wr.fast_reg.rkey);
557 wqe->fr.va_hi = cpu_to_be32(wr->wr.fast_reg.iova_start >> 32);
558 wqe->fr.va_lo_fbo = cpu_to_be32(wr->wr.fast_reg.iova_start &
560 WARN_ON(pbllen > T4_MAX_FR_IMMD);
561 imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
562 imdp->op = FW_RI_DATA_IMMD;
565 imdp->immdlen = cpu_to_be32(pbllen);
566 p = (__be64 *)(imdp + 1);
568 for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
569 *p = cpu_to_be64((u64)wr->wr.fast_reg.page_list->page_list[i]);
571 if (++p == (__be64 *)&sq->queue[sq->size])
572 p = (__be64 *)sq->queue;
578 if (++p == (__be64 *)&sq->queue[sq->size])
579 p = (__be64 *)sq->queue;
581 *len16 = DIV_ROUND_UP(sizeof wqe->fr + sizeof *imdp + pbllen, 16);
585 static int build_inv_stag(union t4_wr *wqe, struct ib_send_wr *wr,
588 wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
590 *len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
594 void c4iw_qp_add_ref(struct ib_qp *qp)
596 PDBG("%s ib_qp %p\n", __func__, qp);
597 atomic_inc(&(to_c4iw_qp(qp)->refcnt));
600 void c4iw_qp_rem_ref(struct ib_qp *qp)
602 PDBG("%s ib_qp %p\n", __func__, qp);
603 if (atomic_dec_and_test(&(to_c4iw_qp(qp)->refcnt)))
604 wake_up(&(to_c4iw_qp(qp)->wait));
607 int c4iw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
608 struct ib_send_wr **bad_wr)
612 enum fw_wr_opcodes fw_opcode = 0;
613 enum fw_ri_wr_flags fw_flags;
617 struct t4_swsqe *swsqe;
621 qhp = to_c4iw_qp(ibqp);
622 spin_lock_irqsave(&qhp->lock, flag);
623 if (t4_wq_in_error(&qhp->wq)) {
624 spin_unlock_irqrestore(&qhp->lock, flag);
627 num_wrs = t4_sq_avail(&qhp->wq);
629 spin_unlock_irqrestore(&qhp->lock, flag);
638 wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
639 qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
642 if (wr->send_flags & IB_SEND_SOLICITED)
643 fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
644 if (wr->send_flags & IB_SEND_SIGNALED)
645 fw_flags |= FW_RI_COMPLETION_FLAG;
646 swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
647 switch (wr->opcode) {
648 case IB_WR_SEND_WITH_INV:
650 if (wr->send_flags & IB_SEND_FENCE)
651 fw_flags |= FW_RI_READ_FENCE_FLAG;
652 fw_opcode = FW_RI_SEND_WR;
653 if (wr->opcode == IB_WR_SEND)
654 swsqe->opcode = FW_RI_SEND;
656 swsqe->opcode = FW_RI_SEND_WITH_INV;
657 err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
659 case IB_WR_RDMA_WRITE:
660 fw_opcode = FW_RI_RDMA_WRITE_WR;
661 swsqe->opcode = FW_RI_RDMA_WRITE;
662 err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
664 case IB_WR_RDMA_READ:
665 case IB_WR_RDMA_READ_WITH_INV:
666 fw_opcode = FW_RI_RDMA_READ_WR;
667 swsqe->opcode = FW_RI_READ_REQ;
668 if (wr->opcode == IB_WR_RDMA_READ_WITH_INV)
669 fw_flags = FW_RI_RDMA_READ_INVALIDATE;
672 err = build_rdma_read(wqe, wr, &len16);
675 swsqe->read_len = wr->sg_list[0].length;
676 if (!qhp->wq.sq.oldest_read)
677 qhp->wq.sq.oldest_read = swsqe;
679 case IB_WR_FAST_REG_MR:
680 fw_opcode = FW_RI_FR_NSMR_WR;
681 swsqe->opcode = FW_RI_FAST_REGISTER;
682 err = build_fastreg(&qhp->wq.sq, wqe, wr, &len16);
684 case IB_WR_LOCAL_INV:
685 if (wr->send_flags & IB_SEND_FENCE)
686 fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
687 fw_opcode = FW_RI_INV_LSTAG_WR;
688 swsqe->opcode = FW_RI_LOCAL_INV;
689 err = build_inv_stag(wqe, wr, &len16);
692 PDBG("%s post of type=%d TBD!\n", __func__,
700 swsqe->idx = qhp->wq.sq.pidx;
702 swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED);
703 swsqe->wr_id = wr->wr_id;
705 init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
707 PDBG("%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u\n",
708 __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
709 swsqe->opcode, swsqe->read_len);
712 t4_sq_produce(&qhp->wq, len16);
713 idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
715 if (t4_wq_db_enabled(&qhp->wq))
716 t4_ring_sq_db(&qhp->wq, idx);
717 spin_unlock_irqrestore(&qhp->lock, flag);
721 int c4iw_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
722 struct ib_recv_wr **bad_wr)
726 union t4_recv_wr *wqe;
732 qhp = to_c4iw_qp(ibqp);
733 spin_lock_irqsave(&qhp->lock, flag);
734 if (t4_wq_in_error(&qhp->wq)) {
735 spin_unlock_irqrestore(&qhp->lock, flag);
738 num_wrs = t4_rq_avail(&qhp->wq);
740 spin_unlock_irqrestore(&qhp->lock, flag);
744 if (wr->num_sge > T4_MAX_RECV_SGE) {
749 wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
753 err = build_rdma_recv(qhp, wqe, wr, &len16);
761 qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
763 wqe->recv.opcode = FW_RI_RECV_WR;
765 wqe->recv.wrid = qhp->wq.rq.pidx;
769 wqe->recv.len16 = len16;
770 PDBG("%s cookie 0x%llx pidx %u\n", __func__,
771 (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
772 t4_rq_produce(&qhp->wq, len16);
773 idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
777 if (t4_wq_db_enabled(&qhp->wq))
778 t4_ring_rq_db(&qhp->wq, idx);
779 spin_unlock_irqrestore(&qhp->lock, flag);
783 int c4iw_bind_mw(struct ib_qp *qp, struct ib_mw *mw, struct ib_mw_bind *mw_bind)
788 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
798 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
803 status = CQE_STATUS(err_cqe);
804 opcode = CQE_OPCODE(err_cqe);
805 rqtype = RQ_TYPE(err_cqe);
806 send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
807 (opcode == FW_RI_SEND_WITH_SE_INV);
808 tagged = (opcode == FW_RI_RDMA_WRITE) ||
809 (rqtype && (opcode == FW_RI_READ_RESP));
814 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
815 *ecode = RDMAP_CANT_INV_STAG;
817 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
818 *ecode = RDMAP_INV_STAG;
822 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
823 if ((opcode == FW_RI_SEND_WITH_INV) ||
824 (opcode == FW_RI_SEND_WITH_SE_INV))
825 *ecode = RDMAP_CANT_INV_STAG;
827 *ecode = RDMAP_STAG_NOT_ASSOC;
830 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
831 *ecode = RDMAP_STAG_NOT_ASSOC;
834 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
835 *ecode = RDMAP_ACC_VIOL;
838 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
839 *ecode = RDMAP_TO_WRAP;
843 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
844 *ecode = DDPT_BASE_BOUNDS;
846 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
847 *ecode = RDMAP_BASE_BOUNDS;
850 case T4_ERR_INVALIDATE_SHARED_MR:
851 case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
852 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
853 *ecode = RDMAP_CANT_INV_STAG;
856 case T4_ERR_ECC_PSTAG:
857 case T4_ERR_INTERNAL_ERR:
858 *layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
861 case T4_ERR_OUT_OF_RQE:
862 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
863 *ecode = DDPU_INV_MSN_NOBUF;
865 case T4_ERR_PBL_ADDR_BOUND:
866 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
867 *ecode = DDPT_BASE_BOUNDS;
870 *layer_type = LAYER_MPA|DDP_LLP;
871 *ecode = MPA_CRC_ERR;
874 *layer_type = LAYER_MPA|DDP_LLP;
875 *ecode = MPA_MARKER_ERR;
877 case T4_ERR_PDU_LEN_ERR:
878 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
879 *ecode = DDPU_MSG_TOOBIG;
881 case T4_ERR_DDP_VERSION:
883 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
884 *ecode = DDPT_INV_VERS;
886 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
887 *ecode = DDPU_INV_VERS;
890 case T4_ERR_RDMA_VERSION:
891 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
892 *ecode = RDMAP_INV_VERS;
895 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
896 *ecode = RDMAP_INV_OPCODE;
898 case T4_ERR_DDP_QUEUE_NUM:
899 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
900 *ecode = DDPU_INV_QN;
904 case T4_ERR_MSN_RANGE:
905 case T4_ERR_IRD_OVERFLOW:
906 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
907 *ecode = DDPU_INV_MSN_RANGE;
910 *layer_type = LAYER_DDP|DDP_LOCAL_CATA;
914 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
915 *ecode = DDPU_INV_MO;
918 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
924 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
927 struct fw_ri_wr *wqe;
929 struct terminate_message *term;
931 PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
934 skb = alloc_skb(sizeof *wqe, gfp);
937 set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
939 wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
940 memset(wqe, 0, sizeof *wqe);
941 wqe->op_compl = cpu_to_be32(FW_WR_OP(FW_RI_INIT_WR));
942 wqe->flowid_len16 = cpu_to_be32(
943 FW_WR_FLOWID(qhp->ep->hwtid) |
944 FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
946 wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
947 wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
948 term = (struct terminate_message *)wqe->u.terminate.termmsg;
949 if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
950 term->layer_etype = qhp->attr.layer_etype;
951 term->ecode = qhp->attr.ecode;
953 build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
954 c4iw_ofld_send(&qhp->rhp->rdev, skb);
958 * Assumes qhp lock is held.
960 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
961 struct c4iw_cq *schp)
967 PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
969 /* locking hierarchy: cq lock first, then qp lock. */
970 spin_lock_irqsave(&rchp->lock, flag);
971 spin_lock(&qhp->lock);
972 c4iw_flush_hw_cq(&rchp->cq);
973 c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
974 flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
975 spin_unlock(&qhp->lock);
976 spin_unlock_irqrestore(&rchp->lock, flag);
978 spin_lock_irqsave(&rchp->comp_handler_lock, flag);
979 (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
980 spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
983 /* locking hierarchy: cq lock first, then qp lock. */
984 spin_lock_irqsave(&schp->lock, flag);
985 spin_lock(&qhp->lock);
986 c4iw_flush_hw_cq(&schp->cq);
987 c4iw_count_scqes(&schp->cq, &qhp->wq, &count);
988 flushed = c4iw_flush_sq(&qhp->wq, &schp->cq, count);
989 spin_unlock(&qhp->lock);
990 spin_unlock_irqrestore(&schp->lock, flag);
992 spin_lock_irqsave(&schp->comp_handler_lock, flag);
993 (*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context);
994 spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
998 static void flush_qp(struct c4iw_qp *qhp)
1000 struct c4iw_cq *rchp, *schp;
1003 rchp = get_chp(qhp->rhp, qhp->attr.rcq);
1004 schp = get_chp(qhp->rhp, qhp->attr.scq);
1006 if (qhp->ibqp.uobject) {
1007 t4_set_wq_in_error(&qhp->wq);
1008 t4_set_cq_in_error(&rchp->cq);
1009 spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1010 (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1011 spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1013 t4_set_cq_in_error(&schp->cq);
1014 spin_lock_irqsave(&schp->comp_handler_lock, flag);
1015 (*schp->ibcq.comp_handler)(&schp->ibcq,
1016 schp->ibcq.cq_context);
1017 spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1021 __flush_qp(qhp, rchp, schp);
1024 static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1027 struct fw_ri_wr *wqe;
1029 struct sk_buff *skb;
1031 PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1034 skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1037 set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
1039 wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1040 memset(wqe, 0, sizeof *wqe);
1041 wqe->op_compl = cpu_to_be32(
1042 FW_WR_OP(FW_RI_INIT_WR) |
1044 wqe->flowid_len16 = cpu_to_be32(
1045 FW_WR_FLOWID(ep->hwtid) |
1046 FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1047 wqe->cookie = (unsigned long) &ep->com.wr_wait;
1049 wqe->u.fini.type = FW_RI_TYPE_FINI;
1050 ret = c4iw_ofld_send(&rhp->rdev, skb);
1054 ret = c4iw_wait_for_reply(&rhp->rdev, &ep->com.wr_wait, qhp->ep->hwtid,
1055 qhp->wq.sq.qid, __func__);
1057 PDBG("%s ret %d\n", __func__, ret);
1061 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1063 PDBG("%s p2p_type = %d\n", __func__, p2p_type);
1064 memset(&init->u, 0, sizeof init->u);
1066 case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1067 init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1068 init->u.write.stag_sink = cpu_to_be32(1);
1069 init->u.write.to_sink = cpu_to_be64(1);
1070 init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1071 init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1072 sizeof(struct fw_ri_immd),
1075 case FW_RI_INIT_P2PTYPE_READ_REQ:
1076 init->u.write.opcode = FW_RI_RDMA_READ_WR;
1077 init->u.read.stag_src = cpu_to_be32(1);
1078 init->u.read.to_src_lo = cpu_to_be32(1);
1079 init->u.read.stag_sink = cpu_to_be32(1);
1080 init->u.read.to_sink_lo = cpu_to_be32(1);
1081 init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1086 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1088 struct fw_ri_wr *wqe;
1090 struct sk_buff *skb;
1092 PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1095 skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1098 set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1100 wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1101 memset(wqe, 0, sizeof *wqe);
1102 wqe->op_compl = cpu_to_be32(
1103 FW_WR_OP(FW_RI_INIT_WR) |
1105 wqe->flowid_len16 = cpu_to_be32(
1106 FW_WR_FLOWID(qhp->ep->hwtid) |
1107 FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1109 wqe->cookie = (unsigned long) &qhp->ep->com.wr_wait;
1111 wqe->u.init.type = FW_RI_TYPE_INIT;
1112 wqe->u.init.mpareqbit_p2ptype =
1113 V_FW_RI_WR_MPAREQBIT(qhp->attr.mpa_attr.initiator) |
1114 V_FW_RI_WR_P2PTYPE(qhp->attr.mpa_attr.p2p_type);
1115 wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1116 if (qhp->attr.mpa_attr.recv_marker_enabled)
1117 wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1118 if (qhp->attr.mpa_attr.xmit_marker_enabled)
1119 wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1120 if (qhp->attr.mpa_attr.crc_enabled)
1121 wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1123 wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1124 FW_RI_QP_RDMA_WRITE_ENABLE |
1125 FW_RI_QP_BIND_ENABLE;
1126 if (!qhp->ibqp.uobject)
1127 wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1128 FW_RI_QP_STAG0_ENABLE;
1129 wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1130 wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1131 wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1132 wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1133 wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1134 wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1135 wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1136 wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1137 wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1138 wqe->u.init.iss = cpu_to_be32(qhp->ep->snd_seq);
1139 wqe->u.init.irs = cpu_to_be32(qhp->ep->rcv_seq);
1140 wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1141 wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1142 rhp->rdev.lldi.vr->rq.start);
1143 if (qhp->attr.mpa_attr.initiator)
1144 build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1146 ret = c4iw_ofld_send(&rhp->rdev, skb);
1150 ret = c4iw_wait_for_reply(&rhp->rdev, &qhp->ep->com.wr_wait,
1151 qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1153 PDBG("%s ret %d\n", __func__, ret);
1158 * Called by the library when the qp has user dbs disabled due to
1159 * a DB_FULL condition. This function will single-thread all user
1160 * DB rings to avoid overflowing the hw db-fifo.
1162 static int ring_kernel_db(struct c4iw_qp *qhp, u32 qid, u16 inc)
1164 int delay = db_delay_usecs;
1166 mutex_lock(&qhp->rhp->db_mutex);
1170 * The interrupt threshold is dbfifo_int_thresh << 6. So
1171 * make sure we don't cross that and generate an interrupt.
1173 if (cxgb4_dbfifo_count(qhp->rhp->rdev.lldi.ports[0], 1) <
1174 (qhp->rhp->rdev.lldi.dbfifo_int_thresh << 5)) {
1175 writel(QID(qid) | PIDX(inc), qhp->wq.db);
1178 set_current_state(TASK_UNINTERRUPTIBLE);
1179 schedule_timeout(usecs_to_jiffies(delay));
1180 delay = min(delay << 1, 2000);
1182 mutex_unlock(&qhp->rhp->db_mutex);
1186 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1187 enum c4iw_qp_attr_mask mask,
1188 struct c4iw_qp_attributes *attrs,
1192 struct c4iw_qp_attributes newattr = qhp->attr;
1197 struct c4iw_ep *ep = NULL;
1199 PDBG("%s qhp %p sqid 0x%x rqid 0x%x ep %p state %d -> %d\n", __func__,
1200 qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep, qhp->attr.state,
1201 (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1203 mutex_lock(&qhp->mutex);
1205 /* Process attr changes if in IDLE */
1206 if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1207 if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1211 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1212 newattr.enable_rdma_read = attrs->enable_rdma_read;
1213 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1214 newattr.enable_rdma_write = attrs->enable_rdma_write;
1215 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1216 newattr.enable_bind = attrs->enable_bind;
1217 if (mask & C4IW_QP_ATTR_MAX_ORD) {
1218 if (attrs->max_ord > c4iw_max_read_depth) {
1222 newattr.max_ord = attrs->max_ord;
1224 if (mask & C4IW_QP_ATTR_MAX_IRD) {
1225 if (attrs->max_ird > c4iw_max_read_depth) {
1229 newattr.max_ird = attrs->max_ird;
1231 qhp->attr = newattr;
1234 if (mask & C4IW_QP_ATTR_SQ_DB) {
1235 ret = ring_kernel_db(qhp, qhp->wq.sq.qid, attrs->sq_db_inc);
1238 if (mask & C4IW_QP_ATTR_RQ_DB) {
1239 ret = ring_kernel_db(qhp, qhp->wq.rq.qid, attrs->rq_db_inc);
1243 if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1245 if (qhp->attr.state == attrs->next_state)
1248 switch (qhp->attr.state) {
1249 case C4IW_QP_STATE_IDLE:
1250 switch (attrs->next_state) {
1251 case C4IW_QP_STATE_RTS:
1252 if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1256 if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1260 qhp->attr.mpa_attr = attrs->mpa_attr;
1261 qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1262 qhp->ep = qhp->attr.llp_stream_handle;
1263 set_state(qhp, C4IW_QP_STATE_RTS);
1266 * Ref the endpoint here and deref when we
1267 * disassociate the endpoint from the QP. This
1268 * happens in CLOSING->IDLE transition or *->ERROR
1271 c4iw_get_ep(&qhp->ep->com);
1272 ret = rdma_init(rhp, qhp);
1276 case C4IW_QP_STATE_ERROR:
1277 set_state(qhp, C4IW_QP_STATE_ERROR);
1285 case C4IW_QP_STATE_RTS:
1286 switch (attrs->next_state) {
1287 case C4IW_QP_STATE_CLOSING:
1288 BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1289 set_state(qhp, C4IW_QP_STATE_CLOSING);
1294 c4iw_get_ep(&qhp->ep->com);
1296 if (qhp->ibqp.uobject)
1297 t4_set_wq_in_error(&qhp->wq);
1298 ret = rdma_fini(rhp, qhp, ep);
1302 case C4IW_QP_STATE_TERMINATE:
1303 set_state(qhp, C4IW_QP_STATE_TERMINATE);
1304 qhp->attr.layer_etype = attrs->layer_etype;
1305 qhp->attr.ecode = attrs->ecode;
1306 if (qhp->ibqp.uobject)
1307 t4_set_wq_in_error(&qhp->wq);
1312 c4iw_get_ep(&qhp->ep->com);
1314 case C4IW_QP_STATE_ERROR:
1315 set_state(qhp, C4IW_QP_STATE_ERROR);
1316 if (qhp->ibqp.uobject)
1317 t4_set_wq_in_error(&qhp->wq);
1322 c4iw_get_ep(&qhp->ep->com);
1331 case C4IW_QP_STATE_CLOSING:
1336 switch (attrs->next_state) {
1337 case C4IW_QP_STATE_IDLE:
1339 set_state(qhp, C4IW_QP_STATE_IDLE);
1340 qhp->attr.llp_stream_handle = NULL;
1341 c4iw_put_ep(&qhp->ep->com);
1343 wake_up(&qhp->wait);
1345 case C4IW_QP_STATE_ERROR:
1352 case C4IW_QP_STATE_ERROR:
1353 if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1357 if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1361 set_state(qhp, C4IW_QP_STATE_IDLE);
1363 case C4IW_QP_STATE_TERMINATE:
1371 printk(KERN_ERR "%s in a bad state %d\n",
1372 __func__, qhp->attr.state);
1379 PDBG("%s disassociating ep %p qpid 0x%x\n", __func__, qhp->ep,
1382 /* disassociate the LLP connection */
1383 qhp->attr.llp_stream_handle = NULL;
1387 set_state(qhp, C4IW_QP_STATE_ERROR);
1390 wake_up(&qhp->wait);
1394 mutex_unlock(&qhp->mutex);
1397 post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1400 * If disconnect is 1, then we need to initiate a disconnect
1401 * on the EP. This can be a normal close (RTS->CLOSING) or
1402 * an abnormal close (RTS/CLOSING->ERROR).
1405 c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1407 c4iw_put_ep(&ep->com);
1411 * If free is 1, then we've disassociated the EP from the QP
1412 * and we need to dereference the EP.
1415 c4iw_put_ep(&ep->com);
1416 PDBG("%s exit state %d\n", __func__, qhp->attr.state);
1420 static int enable_qp_db(int id, void *p, void *data)
1422 struct c4iw_qp *qp = p;
1424 t4_enable_wq_db(&qp->wq);
1428 int c4iw_destroy_qp(struct ib_qp *ib_qp)
1430 struct c4iw_dev *rhp;
1431 struct c4iw_qp *qhp;
1432 struct c4iw_qp_attributes attrs;
1433 struct c4iw_ucontext *ucontext;
1435 qhp = to_c4iw_qp(ib_qp);
1438 attrs.next_state = C4IW_QP_STATE_ERROR;
1439 if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1440 c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1442 c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1443 wait_event(qhp->wait, !qhp->ep);
1445 spin_lock_irq(&rhp->lock);
1446 remove_handle_nolock(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1448 BUG_ON(rhp->qpcnt < 0);
1449 if (rhp->qpcnt <= db_fc_threshold && rhp->db_state == FLOW_CONTROL) {
1450 rhp->rdev.stats.db_state_transitions++;
1451 rhp->db_state = NORMAL;
1452 idr_for_each(&rhp->qpidr, enable_qp_db, NULL);
1454 spin_unlock_irq(&rhp->lock);
1455 atomic_dec(&qhp->refcnt);
1456 wait_event(qhp->wait, !atomic_read(&qhp->refcnt));
1458 ucontext = ib_qp->uobject ?
1459 to_c4iw_ucontext(ib_qp->uobject->context) : NULL;
1460 destroy_qp(&rhp->rdev, &qhp->wq,
1461 ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1463 PDBG("%s ib_qp %p qpid 0x%0x\n", __func__, ib_qp, qhp->wq.sq.qid);
1468 static int disable_qp_db(int id, void *p, void *data)
1470 struct c4iw_qp *qp = p;
1472 t4_disable_wq_db(&qp->wq);
1476 struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1477 struct ib_udata *udata)
1479 struct c4iw_dev *rhp;
1480 struct c4iw_qp *qhp;
1481 struct c4iw_pd *php;
1482 struct c4iw_cq *schp;
1483 struct c4iw_cq *rchp;
1484 struct c4iw_create_qp_resp uresp;
1486 struct c4iw_ucontext *ucontext;
1488 struct c4iw_mm_entry *mm1, *mm2, *mm3, *mm4, *mm5 = NULL;
1490 PDBG("%s ib_pd %p\n", __func__, pd);
1492 if (attrs->qp_type != IB_QPT_RC)
1493 return ERR_PTR(-EINVAL);
1495 php = to_c4iw_pd(pd);
1497 schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1498 rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1500 return ERR_PTR(-EINVAL);
1502 if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1503 return ERR_PTR(-EINVAL);
1505 rqsize = roundup(attrs->cap.max_recv_wr + 1, 16);
1506 if (rqsize > T4_MAX_RQ_SIZE)
1507 return ERR_PTR(-E2BIG);
1509 sqsize = roundup(attrs->cap.max_send_wr + 1, 16);
1510 if (sqsize > T4_MAX_SQ_SIZE)
1511 return ERR_PTR(-E2BIG);
1513 ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1516 qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1518 return ERR_PTR(-ENOMEM);
1519 qhp->wq.sq.size = sqsize;
1520 qhp->wq.sq.memsize = (sqsize + 1) * sizeof *qhp->wq.sq.queue;
1521 qhp->wq.rq.size = rqsize;
1522 qhp->wq.rq.memsize = (rqsize + 1) * sizeof *qhp->wq.rq.queue;
1525 qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1526 qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1529 PDBG("%s sqsize %u sqmemsize %zu rqsize %u rqmemsize %zu\n",
1530 __func__, sqsize, qhp->wq.sq.memsize, rqsize, qhp->wq.rq.memsize);
1532 ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1533 ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1537 attrs->cap.max_recv_wr = rqsize - 1;
1538 attrs->cap.max_send_wr = sqsize - 1;
1539 attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1542 qhp->attr.pd = php->pdid;
1543 qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1544 qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1545 qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1546 qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1547 qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1548 qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1549 qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1550 qhp->attr.state = C4IW_QP_STATE_IDLE;
1551 qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1552 qhp->attr.enable_rdma_read = 1;
1553 qhp->attr.enable_rdma_write = 1;
1554 qhp->attr.enable_bind = 1;
1555 qhp->attr.max_ord = 1;
1556 qhp->attr.max_ird = 1;
1557 spin_lock_init(&qhp->lock);
1558 mutex_init(&qhp->mutex);
1559 init_waitqueue_head(&qhp->wait);
1560 atomic_set(&qhp->refcnt, 1);
1562 spin_lock_irq(&rhp->lock);
1563 if (rhp->db_state != NORMAL)
1564 t4_disable_wq_db(&qhp->wq);
1565 if (++rhp->qpcnt > db_fc_threshold && rhp->db_state == NORMAL) {
1566 rhp->rdev.stats.db_state_transitions++;
1567 rhp->db_state = FLOW_CONTROL;
1568 idr_for_each(&rhp->qpidr, disable_qp_db, NULL);
1570 ret = insert_handle_nolock(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1571 spin_unlock_irq(&rhp->lock);
1576 mm1 = kmalloc(sizeof *mm1, GFP_KERNEL);
1581 mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
1586 mm3 = kmalloc(sizeof *mm3, GFP_KERNEL);
1591 mm4 = kmalloc(sizeof *mm4, GFP_KERNEL);
1596 if (t4_sq_onchip(&qhp->wq.sq)) {
1597 mm5 = kmalloc(sizeof *mm5, GFP_KERNEL);
1602 uresp.flags = C4IW_QPF_ONCHIP;
1605 uresp.qid_mask = rhp->rdev.qpmask;
1606 uresp.sqid = qhp->wq.sq.qid;
1607 uresp.sq_size = qhp->wq.sq.size;
1608 uresp.sq_memsize = qhp->wq.sq.memsize;
1609 uresp.rqid = qhp->wq.rq.qid;
1610 uresp.rq_size = qhp->wq.rq.size;
1611 uresp.rq_memsize = qhp->wq.rq.memsize;
1612 spin_lock(&ucontext->mmap_lock);
1614 uresp.ma_sync_key = ucontext->key;
1615 ucontext->key += PAGE_SIZE;
1617 uresp.sq_key = ucontext->key;
1618 ucontext->key += PAGE_SIZE;
1619 uresp.rq_key = ucontext->key;
1620 ucontext->key += PAGE_SIZE;
1621 uresp.sq_db_gts_key = ucontext->key;
1622 ucontext->key += PAGE_SIZE;
1623 uresp.rq_db_gts_key = ucontext->key;
1624 ucontext->key += PAGE_SIZE;
1625 spin_unlock(&ucontext->mmap_lock);
1626 ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1629 mm1->key = uresp.sq_key;
1630 mm1->addr = qhp->wq.sq.phys_addr;
1631 mm1->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1632 insert_mmap(ucontext, mm1);
1633 mm2->key = uresp.rq_key;
1634 mm2->addr = virt_to_phys(qhp->wq.rq.queue);
1635 mm2->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1636 insert_mmap(ucontext, mm2);
1637 mm3->key = uresp.sq_db_gts_key;
1638 mm3->addr = qhp->wq.sq.udb;
1639 mm3->len = PAGE_SIZE;
1640 insert_mmap(ucontext, mm3);
1641 mm4->key = uresp.rq_db_gts_key;
1642 mm4->addr = qhp->wq.rq.udb;
1643 mm4->len = PAGE_SIZE;
1644 insert_mmap(ucontext, mm4);
1646 mm5->key = uresp.ma_sync_key;
1647 mm5->addr = (pci_resource_start(rhp->rdev.lldi.pdev, 0)
1648 + A_PCIE_MA_SYNC) & PAGE_MASK;
1649 mm5->len = PAGE_SIZE;
1650 insert_mmap(ucontext, mm5);
1653 qhp->ibqp.qp_num = qhp->wq.sq.qid;
1654 init_timer(&(qhp->timer));
1655 PDBG("%s qhp %p sq_num_entries %d, rq_num_entries %d qpid 0x%0x\n",
1656 __func__, qhp, qhp->attr.sq_num_entries, qhp->attr.rq_num_entries,
1670 remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1672 destroy_qp(&rhp->rdev, &qhp->wq,
1673 ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1676 return ERR_PTR(ret);
1679 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1680 int attr_mask, struct ib_udata *udata)
1682 struct c4iw_dev *rhp;
1683 struct c4iw_qp *qhp;
1684 enum c4iw_qp_attr_mask mask = 0;
1685 struct c4iw_qp_attributes attrs;
1687 PDBG("%s ib_qp %p\n", __func__, ibqp);
1689 /* iwarp does not support the RTR state */
1690 if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1691 attr_mask &= ~IB_QP_STATE;
1693 /* Make sure we still have something left to do */
1697 memset(&attrs, 0, sizeof attrs);
1698 qhp = to_c4iw_qp(ibqp);
1701 attrs.next_state = c4iw_convert_state(attr->qp_state);
1702 attrs.enable_rdma_read = (attr->qp_access_flags &
1703 IB_ACCESS_REMOTE_READ) ? 1 : 0;
1704 attrs.enable_rdma_write = (attr->qp_access_flags &
1705 IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1706 attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1709 mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1710 mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1711 (C4IW_QP_ATTR_ENABLE_RDMA_READ |
1712 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1713 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1716 * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
1717 * ringing the queue db when we're in DB_FULL mode.
1719 attrs.sq_db_inc = attr->sq_psn;
1720 attrs.rq_db_inc = attr->rq_psn;
1721 mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
1722 mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
1724 return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
1727 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
1729 PDBG("%s ib_dev %p qpn 0x%x\n", __func__, dev, qpn);
1730 return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
1733 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1734 int attr_mask, struct ib_qp_init_attr *init_attr)
1736 struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
1738 memset(attr, 0, sizeof *attr);
1739 memset(init_attr, 0, sizeof *init_attr);
1740 attr->qp_state = to_ib_qp_state(qhp->attr.state);