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 = 1000;
46 module_param(db_fc_threshold, int, 0644);
47 MODULE_PARM_DESC(db_fc_threshold,
48 "QP count/threshold that triggers"
49 " automatic db flow control mode (default = 1000)");
51 int db_coalescing_threshold;
52 module_param(db_coalescing_threshold, int, 0644);
53 MODULE_PARM_DESC(db_coalescing_threshold,
54 "QP count/threshold that triggers"
55 " disabling db coalescing (default = 0)");
57 static int max_fr_immd = T4_MAX_FR_IMMD;
58 module_param(max_fr_immd, int, 0644);
59 MODULE_PARM_DESC(max_fr_immd, "fastreg threshold for using DSGL instead of immedate");
61 static int alloc_ird(struct c4iw_dev *dev, u32 ird)
65 spin_lock_irq(&dev->lock);
66 if (ird <= dev->avail_ird)
67 dev->avail_ird -= ird;
70 spin_unlock_irq(&dev->lock);
73 dev_warn(&dev->rdev.lldi.pdev->dev,
74 "device IRD resources exhausted\n");
79 static void free_ird(struct c4iw_dev *dev, int ird)
81 spin_lock_irq(&dev->lock);
82 dev->avail_ird += ird;
83 spin_unlock_irq(&dev->lock);
86 static void set_state(struct c4iw_qp *qhp, enum c4iw_qp_state state)
89 spin_lock_irqsave(&qhp->lock, flag);
90 qhp->attr.state = state;
91 spin_unlock_irqrestore(&qhp->lock, flag);
94 static void dealloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
96 c4iw_ocqp_pool_free(rdev, sq->dma_addr, sq->memsize);
99 static void dealloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
101 dma_free_coherent(&(rdev->lldi.pdev->dev), sq->memsize, sq->queue,
102 pci_unmap_addr(sq, mapping));
105 static void dealloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
107 if (t4_sq_onchip(sq))
108 dealloc_oc_sq(rdev, sq);
110 dealloc_host_sq(rdev, sq);
113 static int alloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
115 if (!ocqp_support || !ocqp_supported(&rdev->lldi))
117 sq->dma_addr = c4iw_ocqp_pool_alloc(rdev, sq->memsize);
120 sq->phys_addr = rdev->oc_mw_pa + sq->dma_addr -
121 rdev->lldi.vr->ocq.start;
122 sq->queue = (__force union t4_wr *)(rdev->oc_mw_kva + sq->dma_addr -
123 rdev->lldi.vr->ocq.start);
124 sq->flags |= T4_SQ_ONCHIP;
128 static int alloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
130 sq->queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev), sq->memsize,
131 &(sq->dma_addr), GFP_KERNEL);
134 sq->phys_addr = virt_to_phys(sq->queue);
135 pci_unmap_addr_set(sq, mapping, sq->dma_addr);
139 static int alloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq, int user)
143 ret = alloc_oc_sq(rdev, sq);
145 ret = alloc_host_sq(rdev, sq);
149 static int destroy_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
150 struct c4iw_dev_ucontext *uctx)
153 * uP clears EQ contexts when the connection exits rdma mode,
154 * so no need to post a RESET WR for these EQs.
156 dma_free_coherent(&(rdev->lldi.pdev->dev),
157 wq->rq.memsize, wq->rq.queue,
158 dma_unmap_addr(&wq->rq, mapping));
159 dealloc_sq(rdev, &wq->sq);
160 c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
163 c4iw_put_qpid(rdev, wq->rq.qid, uctx);
164 c4iw_put_qpid(rdev, wq->sq.qid, uctx);
169 * Determine the BAR2 virtual address and qid. If pbar2_pa is not NULL,
170 * then this is a user mapping so compute the page-aligned physical address
173 void __iomem *c4iw_bar2_addrs(struct c4iw_rdev *rdev, unsigned int qid,
174 enum cxgb4_bar2_qtype qtype,
175 unsigned int *pbar2_qid, u64 *pbar2_pa)
180 ret = cxgb4_bar2_sge_qregs(rdev->lldi.ports[0], qid, qtype,
182 &bar2_qoffset, pbar2_qid);
187 *pbar2_pa = (rdev->bar2_pa + bar2_qoffset) & PAGE_MASK;
188 return rdev->bar2_kva + bar2_qoffset;
191 static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
192 struct t4_cq *rcq, struct t4_cq *scq,
193 struct c4iw_dev_ucontext *uctx)
195 int user = (uctx != &rdev->uctx);
196 struct fw_ri_res_wr *res_wr;
197 struct fw_ri_res *res;
199 struct c4iw_wr_wait wr_wait;
204 wq->sq.qid = c4iw_get_qpid(rdev, uctx);
208 wq->rq.qid = c4iw_get_qpid(rdev, uctx);
215 wq->sq.sw_sq = kzalloc(wq->sq.size * sizeof *wq->sq.sw_sq,
222 wq->rq.sw_rq = kzalloc(wq->rq.size * sizeof *wq->rq.sw_rq,
231 * RQT must be a power of 2 and at least 16 deep.
233 wq->rq.rqt_size = roundup_pow_of_two(max_t(u16, wq->rq.size, 16));
234 wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size);
235 if (!wq->rq.rqt_hwaddr) {
240 ret = alloc_sq(rdev, &wq->sq, user);
243 memset(wq->sq.queue, 0, wq->sq.memsize);
244 dma_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
246 wq->rq.queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev),
247 wq->rq.memsize, &(wq->rq.dma_addr),
253 PDBG("%s sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx\n",
254 __func__, wq->sq.queue,
255 (unsigned long long)virt_to_phys(wq->sq.queue),
257 (unsigned long long)virt_to_phys(wq->rq.queue));
258 memset(wq->rq.queue, 0, wq->rq.memsize);
259 dma_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
261 wq->db = rdev->lldi.db_reg;
263 wq->sq.bar2_va = c4iw_bar2_addrs(rdev, wq->sq.qid, T4_BAR2_QTYPE_EGRESS,
265 user ? &wq->sq.bar2_pa : NULL);
266 wq->rq.bar2_va = c4iw_bar2_addrs(rdev, wq->rq.qid, T4_BAR2_QTYPE_EGRESS,
268 user ? &wq->rq.bar2_pa : NULL);
271 * User mode must have bar2 access.
273 if (user && (!wq->sq.bar2_va || !wq->rq.bar2_va)) {
274 pr_warn(MOD "%s: sqid %u or rqid %u not in BAR2 range.\n",
275 pci_name(rdev->lldi.pdev), wq->sq.qid, wq->rq.qid);
282 /* build fw_ri_res_wr */
283 wr_len = sizeof *res_wr + 2 * sizeof *res;
285 skb = alloc_skb(wr_len, GFP_KERNEL);
290 set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
292 res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
293 memset(res_wr, 0, wr_len);
294 res_wr->op_nres = cpu_to_be32(
295 FW_WR_OP_V(FW_RI_RES_WR) |
296 FW_RI_RES_WR_NRES_V(2) |
298 res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
299 res_wr->cookie = (uintptr_t)&wr_wait;
301 res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
302 res->u.sqrq.op = FW_RI_RES_OP_WRITE;
305 * eqsize is the number of 64B entries plus the status page size.
307 eqsize = wq->sq.size * T4_SQ_NUM_SLOTS +
308 rdev->hw_queue.t4_eq_status_entries;
310 res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
311 FW_RI_RES_WR_HOSTFCMODE_V(0) | /* no host cidx updates */
312 FW_RI_RES_WR_CPRIO_V(0) | /* don't keep in chip cache */
313 FW_RI_RES_WR_PCIECHN_V(0) | /* set by uP at ri_init time */
314 (t4_sq_onchip(&wq->sq) ? FW_RI_RES_WR_ONCHIP_F : 0) |
315 FW_RI_RES_WR_IQID_V(scq->cqid));
316 res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
317 FW_RI_RES_WR_DCAEN_V(0) |
318 FW_RI_RES_WR_DCACPU_V(0) |
319 FW_RI_RES_WR_FBMIN_V(2) |
320 FW_RI_RES_WR_FBMAX_V(2) |
321 FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
322 FW_RI_RES_WR_CIDXFTHRESH_V(0) |
323 FW_RI_RES_WR_EQSIZE_V(eqsize));
324 res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
325 res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
327 res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
328 res->u.sqrq.op = FW_RI_RES_OP_WRITE;
331 * eqsize is the number of 64B entries plus the status page size.
333 eqsize = wq->rq.size * T4_RQ_NUM_SLOTS +
334 rdev->hw_queue.t4_eq_status_entries;
335 res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
336 FW_RI_RES_WR_HOSTFCMODE_V(0) | /* no host cidx updates */
337 FW_RI_RES_WR_CPRIO_V(0) | /* don't keep in chip cache */
338 FW_RI_RES_WR_PCIECHN_V(0) | /* set by uP at ri_init time */
339 FW_RI_RES_WR_IQID_V(rcq->cqid));
340 res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
341 FW_RI_RES_WR_DCAEN_V(0) |
342 FW_RI_RES_WR_DCACPU_V(0) |
343 FW_RI_RES_WR_FBMIN_V(2) |
344 FW_RI_RES_WR_FBMAX_V(2) |
345 FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
346 FW_RI_RES_WR_CIDXFTHRESH_V(0) |
347 FW_RI_RES_WR_EQSIZE_V(eqsize));
348 res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
349 res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
351 c4iw_init_wr_wait(&wr_wait);
353 ret = c4iw_ofld_send(rdev, skb);
356 ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, wq->sq.qid, __func__);
360 PDBG("%s sqid 0x%x rqid 0x%x kdb 0x%p sq_bar2_addr %p rq_bar2_addr %p\n",
361 __func__, wq->sq.qid, wq->rq.qid, wq->db,
362 wq->sq.bar2_va, wq->rq.bar2_va);
366 dma_free_coherent(&(rdev->lldi.pdev->dev),
367 wq->rq.memsize, wq->rq.queue,
368 dma_unmap_addr(&wq->rq, mapping));
370 dealloc_sq(rdev, &wq->sq);
372 c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
378 c4iw_put_qpid(rdev, wq->rq.qid, uctx);
380 c4iw_put_qpid(rdev, wq->sq.qid, uctx);
384 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
385 struct ib_send_wr *wr, int max, u32 *plenp)
392 dstp = (u8 *)immdp->data;
393 for (i = 0; i < wr->num_sge; i++) {
394 if ((plen + wr->sg_list[i].length) > max)
396 srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
397 plen += wr->sg_list[i].length;
398 rem = wr->sg_list[i].length;
400 if (dstp == (u8 *)&sq->queue[sq->size])
401 dstp = (u8 *)sq->queue;
402 if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
405 len = (u8 *)&sq->queue[sq->size] - dstp;
406 memcpy(dstp, srcp, len);
412 len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp);
414 memset(dstp, 0, len);
415 immdp->op = FW_RI_DATA_IMMD;
418 immdp->immdlen = cpu_to_be32(plen);
423 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
424 struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
425 int num_sge, u32 *plenp)
430 __be64 *flitp = (__be64 *)isglp->sge;
432 for (i = 0; i < num_sge; i++) {
433 if ((plen + sg_list[i].length) < plen)
435 plen += sg_list[i].length;
436 *flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
438 if (++flitp == queue_end)
440 *flitp = cpu_to_be64(sg_list[i].addr);
441 if (++flitp == queue_end)
444 *flitp = (__force __be64)0;
445 isglp->op = FW_RI_DATA_ISGL;
447 isglp->nsge = cpu_to_be16(num_sge);
454 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
455 struct ib_send_wr *wr, u8 *len16)
461 if (wr->num_sge > T4_MAX_SEND_SGE)
463 switch (wr->opcode) {
465 if (wr->send_flags & IB_SEND_SOLICITED)
466 wqe->send.sendop_pkd = cpu_to_be32(
467 FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_SE));
469 wqe->send.sendop_pkd = cpu_to_be32(
470 FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND));
471 wqe->send.stag_inv = 0;
473 case IB_WR_SEND_WITH_INV:
474 if (wr->send_flags & IB_SEND_SOLICITED)
475 wqe->send.sendop_pkd = cpu_to_be32(
476 FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_SE_INV));
478 wqe->send.sendop_pkd = cpu_to_be32(
479 FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_INV));
480 wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
491 if (wr->send_flags & IB_SEND_INLINE) {
492 ret = build_immd(sq, wqe->send.u.immd_src, wr,
493 T4_MAX_SEND_INLINE, &plen);
496 size = sizeof wqe->send + sizeof(struct fw_ri_immd) +
499 ret = build_isgl((__be64 *)sq->queue,
500 (__be64 *)&sq->queue[sq->size],
501 wqe->send.u.isgl_src,
502 wr->sg_list, wr->num_sge, &plen);
505 size = sizeof wqe->send + sizeof(struct fw_ri_isgl) +
506 wr->num_sge * sizeof(struct fw_ri_sge);
509 wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
510 wqe->send.u.immd_src[0].r1 = 0;
511 wqe->send.u.immd_src[0].r2 = 0;
512 wqe->send.u.immd_src[0].immdlen = 0;
513 size = sizeof wqe->send + sizeof(struct fw_ri_immd);
516 *len16 = DIV_ROUND_UP(size, 16);
517 wqe->send.plen = cpu_to_be32(plen);
521 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
522 struct ib_send_wr *wr, u8 *len16)
528 if (wr->num_sge > T4_MAX_SEND_SGE)
531 wqe->write.stag_sink = cpu_to_be32(rdma_wr(wr)->rkey);
532 wqe->write.to_sink = cpu_to_be64(rdma_wr(wr)->remote_addr);
534 if (wr->send_flags & IB_SEND_INLINE) {
535 ret = build_immd(sq, wqe->write.u.immd_src, wr,
536 T4_MAX_WRITE_INLINE, &plen);
539 size = sizeof wqe->write + sizeof(struct fw_ri_immd) +
542 ret = build_isgl((__be64 *)sq->queue,
543 (__be64 *)&sq->queue[sq->size],
544 wqe->write.u.isgl_src,
545 wr->sg_list, wr->num_sge, &plen);
548 size = sizeof wqe->write + sizeof(struct fw_ri_isgl) +
549 wr->num_sge * sizeof(struct fw_ri_sge);
552 wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
553 wqe->write.u.immd_src[0].r1 = 0;
554 wqe->write.u.immd_src[0].r2 = 0;
555 wqe->write.u.immd_src[0].immdlen = 0;
556 size = sizeof wqe->write + sizeof(struct fw_ri_immd);
559 *len16 = DIV_ROUND_UP(size, 16);
560 wqe->write.plen = cpu_to_be32(plen);
564 static int build_rdma_read(union t4_wr *wqe, struct ib_send_wr *wr, u8 *len16)
569 wqe->read.stag_src = cpu_to_be32(rdma_wr(wr)->rkey);
570 wqe->read.to_src_hi = cpu_to_be32((u32)(rdma_wr(wr)->remote_addr
572 wqe->read.to_src_lo = cpu_to_be32((u32)rdma_wr(wr)->remote_addr);
573 wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
574 wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
575 wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
577 wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
579 wqe->read.stag_src = cpu_to_be32(2);
580 wqe->read.to_src_hi = 0;
581 wqe->read.to_src_lo = 0;
582 wqe->read.stag_sink = cpu_to_be32(2);
584 wqe->read.to_sink_hi = 0;
585 wqe->read.to_sink_lo = 0;
589 *len16 = DIV_ROUND_UP(sizeof wqe->read, 16);
593 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
594 struct ib_recv_wr *wr, u8 *len16)
598 ret = build_isgl((__be64 *)qhp->wq.rq.queue,
599 (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
600 &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
603 *len16 = DIV_ROUND_UP(sizeof wqe->recv +
604 wr->num_sge * sizeof(struct fw_ri_sge), 16);
608 static int build_fastreg(struct t4_sq *sq, union t4_wr *wqe,
609 struct ib_send_wr *send_wr, u8 *len16, u8 t5dev)
611 struct ib_fast_reg_wr *wr = fast_reg_wr(send_wr);
612 struct fw_ri_immd *imdp;
615 int pbllen = roundup(wr->page_list_len * sizeof(u64), 32);
618 if (wr->page_list_len > t4_max_fr_depth(use_dsgl))
621 wqe->fr.qpbinde_to_dcacpu = 0;
622 wqe->fr.pgsz_shift = wr->page_shift - 12;
623 wqe->fr.addr_type = FW_RI_VA_BASED_TO;
624 wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->access_flags);
626 wqe->fr.len_lo = cpu_to_be32(wr->length);
627 wqe->fr.stag = cpu_to_be32(wr->rkey);
628 wqe->fr.va_hi = cpu_to_be32(wr->iova_start >> 32);
629 wqe->fr.va_lo_fbo = cpu_to_be32(wr->iova_start & 0xffffffff);
631 if (t5dev && use_dsgl && (pbllen > max_fr_immd)) {
632 struct c4iw_fr_page_list *c4pl =
633 to_c4iw_fr_page_list(wr->page_list);
634 struct fw_ri_dsgl *sglp;
636 for (i = 0; i < wr->page_list_len; i++) {
637 wr->page_list->page_list[i] = (__force u64)
638 cpu_to_be64((u64)wr->page_list->page_list[i]);
641 sglp = (struct fw_ri_dsgl *)(&wqe->fr + 1);
642 sglp->op = FW_RI_DATA_DSGL;
644 sglp->nsge = cpu_to_be16(1);
645 sglp->addr0 = cpu_to_be64(c4pl->dma_addr);
646 sglp->len0 = cpu_to_be32(pbllen);
648 *len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*sglp), 16);
650 imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
651 imdp->op = FW_RI_DATA_IMMD;
654 imdp->immdlen = cpu_to_be32(pbllen);
655 p = (__be64 *)(imdp + 1);
657 for (i = 0; i < wr->page_list_len; i++) {
658 *p = cpu_to_be64((u64)wr->page_list->page_list[i]);
660 if (++p == (__be64 *)&sq->queue[sq->size])
661 p = (__be64 *)sq->queue;
667 if (++p == (__be64 *)&sq->queue[sq->size])
668 p = (__be64 *)sq->queue;
670 *len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*imdp)
676 static int build_inv_stag(union t4_wr *wqe, struct ib_send_wr *wr,
679 wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
681 *len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
685 void c4iw_qp_add_ref(struct ib_qp *qp)
687 PDBG("%s ib_qp %p\n", __func__, qp);
688 atomic_inc(&(to_c4iw_qp(qp)->refcnt));
691 void c4iw_qp_rem_ref(struct ib_qp *qp)
693 PDBG("%s ib_qp %p\n", __func__, qp);
694 if (atomic_dec_and_test(&(to_c4iw_qp(qp)->refcnt)))
695 wake_up(&(to_c4iw_qp(qp)->wait));
698 static void add_to_fc_list(struct list_head *head, struct list_head *entry)
700 if (list_empty(entry))
701 list_add_tail(entry, head);
704 static int ring_kernel_sq_db(struct c4iw_qp *qhp, u16 inc)
708 spin_lock_irqsave(&qhp->rhp->lock, flags);
709 spin_lock(&qhp->lock);
710 if (qhp->rhp->db_state == NORMAL)
711 t4_ring_sq_db(&qhp->wq, inc, NULL);
713 add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
714 qhp->wq.sq.wq_pidx_inc += inc;
716 spin_unlock(&qhp->lock);
717 spin_unlock_irqrestore(&qhp->rhp->lock, flags);
721 static int ring_kernel_rq_db(struct c4iw_qp *qhp, u16 inc)
725 spin_lock_irqsave(&qhp->rhp->lock, flags);
726 spin_lock(&qhp->lock);
727 if (qhp->rhp->db_state == NORMAL)
728 t4_ring_rq_db(&qhp->wq, inc, NULL);
730 add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
731 qhp->wq.rq.wq_pidx_inc += inc;
733 spin_unlock(&qhp->lock);
734 spin_unlock_irqrestore(&qhp->rhp->lock, flags);
738 int c4iw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
739 struct ib_send_wr **bad_wr)
743 enum fw_wr_opcodes fw_opcode = 0;
744 enum fw_ri_wr_flags fw_flags;
746 union t4_wr *wqe = NULL;
748 struct t4_swsqe *swsqe;
752 qhp = to_c4iw_qp(ibqp);
753 spin_lock_irqsave(&qhp->lock, flag);
754 if (t4_wq_in_error(&qhp->wq)) {
755 spin_unlock_irqrestore(&qhp->lock, flag);
758 num_wrs = t4_sq_avail(&qhp->wq);
760 spin_unlock_irqrestore(&qhp->lock, flag);
769 wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
770 qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
773 if (wr->send_flags & IB_SEND_SOLICITED)
774 fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
775 if (wr->send_flags & IB_SEND_SIGNALED || qhp->sq_sig_all)
776 fw_flags |= FW_RI_COMPLETION_FLAG;
777 swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
778 switch (wr->opcode) {
779 case IB_WR_SEND_WITH_INV:
781 if (wr->send_flags & IB_SEND_FENCE)
782 fw_flags |= FW_RI_READ_FENCE_FLAG;
783 fw_opcode = FW_RI_SEND_WR;
784 if (wr->opcode == IB_WR_SEND)
785 swsqe->opcode = FW_RI_SEND;
787 swsqe->opcode = FW_RI_SEND_WITH_INV;
788 err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
790 case IB_WR_RDMA_WRITE:
791 fw_opcode = FW_RI_RDMA_WRITE_WR;
792 swsqe->opcode = FW_RI_RDMA_WRITE;
793 err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
795 case IB_WR_RDMA_READ:
796 case IB_WR_RDMA_READ_WITH_INV:
797 fw_opcode = FW_RI_RDMA_READ_WR;
798 swsqe->opcode = FW_RI_READ_REQ;
799 if (wr->opcode == IB_WR_RDMA_READ_WITH_INV)
800 fw_flags = FW_RI_RDMA_READ_INVALIDATE;
803 err = build_rdma_read(wqe, wr, &len16);
806 swsqe->read_len = wr->sg_list[0].length;
807 if (!qhp->wq.sq.oldest_read)
808 qhp->wq.sq.oldest_read = swsqe;
810 case IB_WR_FAST_REG_MR:
811 fw_opcode = FW_RI_FR_NSMR_WR;
812 swsqe->opcode = FW_RI_FAST_REGISTER;
813 err = build_fastreg(&qhp->wq.sq, wqe, wr, &len16,
815 qhp->rhp->rdev.lldi.adapter_type) ?
818 case IB_WR_LOCAL_INV:
819 if (wr->send_flags & IB_SEND_FENCE)
820 fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
821 fw_opcode = FW_RI_INV_LSTAG_WR;
822 swsqe->opcode = FW_RI_LOCAL_INV;
823 err = build_inv_stag(wqe, wr, &len16);
826 PDBG("%s post of type=%d TBD!\n", __func__,
834 swsqe->idx = qhp->wq.sq.pidx;
836 swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
839 swsqe->wr_id = wr->wr_id;
841 swsqe->sge_ts = cxgb4_read_sge_timestamp(
842 qhp->rhp->rdev.lldi.ports[0]);
843 getnstimeofday(&swsqe->host_ts);
846 init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
848 PDBG("%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u\n",
849 __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
850 swsqe->opcode, swsqe->read_len);
853 t4_sq_produce(&qhp->wq, len16);
854 idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
856 if (!qhp->rhp->rdev.status_page->db_off) {
857 t4_ring_sq_db(&qhp->wq, idx, wqe);
858 spin_unlock_irqrestore(&qhp->lock, flag);
860 spin_unlock_irqrestore(&qhp->lock, flag);
861 ring_kernel_sq_db(qhp, idx);
866 int c4iw_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
867 struct ib_recv_wr **bad_wr)
871 union t4_recv_wr *wqe = NULL;
877 qhp = to_c4iw_qp(ibqp);
878 spin_lock_irqsave(&qhp->lock, flag);
879 if (t4_wq_in_error(&qhp->wq)) {
880 spin_unlock_irqrestore(&qhp->lock, flag);
883 num_wrs = t4_rq_avail(&qhp->wq);
885 spin_unlock_irqrestore(&qhp->lock, flag);
889 if (wr->num_sge > T4_MAX_RECV_SGE) {
894 wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
898 err = build_rdma_recv(qhp, wqe, wr, &len16);
906 qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
908 qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].sge_ts =
909 cxgb4_read_sge_timestamp(
910 qhp->rhp->rdev.lldi.ports[0]);
912 &qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].host_ts);
915 wqe->recv.opcode = FW_RI_RECV_WR;
917 wqe->recv.wrid = qhp->wq.rq.pidx;
921 wqe->recv.len16 = len16;
922 PDBG("%s cookie 0x%llx pidx %u\n", __func__,
923 (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
924 t4_rq_produce(&qhp->wq, len16);
925 idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
929 if (!qhp->rhp->rdev.status_page->db_off) {
930 t4_ring_rq_db(&qhp->wq, idx, wqe);
931 spin_unlock_irqrestore(&qhp->lock, flag);
933 spin_unlock_irqrestore(&qhp->lock, flag);
934 ring_kernel_rq_db(qhp, idx);
939 int c4iw_bind_mw(struct ib_qp *qp, struct ib_mw *mw, struct ib_mw_bind *mw_bind)
944 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
954 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
959 status = CQE_STATUS(err_cqe);
960 opcode = CQE_OPCODE(err_cqe);
961 rqtype = RQ_TYPE(err_cqe);
962 send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
963 (opcode == FW_RI_SEND_WITH_SE_INV);
964 tagged = (opcode == FW_RI_RDMA_WRITE) ||
965 (rqtype && (opcode == FW_RI_READ_RESP));
970 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
971 *ecode = RDMAP_CANT_INV_STAG;
973 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
974 *ecode = RDMAP_INV_STAG;
978 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
979 if ((opcode == FW_RI_SEND_WITH_INV) ||
980 (opcode == FW_RI_SEND_WITH_SE_INV))
981 *ecode = RDMAP_CANT_INV_STAG;
983 *ecode = RDMAP_STAG_NOT_ASSOC;
986 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
987 *ecode = RDMAP_STAG_NOT_ASSOC;
990 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
991 *ecode = RDMAP_ACC_VIOL;
994 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
995 *ecode = RDMAP_TO_WRAP;
999 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1000 *ecode = DDPT_BASE_BOUNDS;
1002 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1003 *ecode = RDMAP_BASE_BOUNDS;
1006 case T4_ERR_INVALIDATE_SHARED_MR:
1007 case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
1008 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1009 *ecode = RDMAP_CANT_INV_STAG;
1012 case T4_ERR_ECC_PSTAG:
1013 case T4_ERR_INTERNAL_ERR:
1014 *layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
1017 case T4_ERR_OUT_OF_RQE:
1018 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1019 *ecode = DDPU_INV_MSN_NOBUF;
1021 case T4_ERR_PBL_ADDR_BOUND:
1022 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1023 *ecode = DDPT_BASE_BOUNDS;
1026 *layer_type = LAYER_MPA|DDP_LLP;
1027 *ecode = MPA_CRC_ERR;
1030 *layer_type = LAYER_MPA|DDP_LLP;
1031 *ecode = MPA_MARKER_ERR;
1033 case T4_ERR_PDU_LEN_ERR:
1034 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1035 *ecode = DDPU_MSG_TOOBIG;
1037 case T4_ERR_DDP_VERSION:
1039 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1040 *ecode = DDPT_INV_VERS;
1042 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1043 *ecode = DDPU_INV_VERS;
1046 case T4_ERR_RDMA_VERSION:
1047 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1048 *ecode = RDMAP_INV_VERS;
1051 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1052 *ecode = RDMAP_INV_OPCODE;
1054 case T4_ERR_DDP_QUEUE_NUM:
1055 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1056 *ecode = DDPU_INV_QN;
1059 case T4_ERR_MSN_GAP:
1060 case T4_ERR_MSN_RANGE:
1061 case T4_ERR_IRD_OVERFLOW:
1062 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1063 *ecode = DDPU_INV_MSN_RANGE;
1066 *layer_type = LAYER_DDP|DDP_LOCAL_CATA;
1070 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1071 *ecode = DDPU_INV_MO;
1074 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1080 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
1083 struct fw_ri_wr *wqe;
1084 struct sk_buff *skb;
1085 struct terminate_message *term;
1087 PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1090 skb = alloc_skb(sizeof *wqe, gfp);
1093 set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1095 wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1096 memset(wqe, 0, sizeof *wqe);
1097 wqe->op_compl = cpu_to_be32(FW_WR_OP_V(FW_RI_INIT_WR));
1098 wqe->flowid_len16 = cpu_to_be32(
1099 FW_WR_FLOWID_V(qhp->ep->hwtid) |
1100 FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1102 wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
1103 wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
1104 term = (struct terminate_message *)wqe->u.terminate.termmsg;
1105 if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
1106 term->layer_etype = qhp->attr.layer_etype;
1107 term->ecode = qhp->attr.ecode;
1109 build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
1110 c4iw_ofld_send(&qhp->rhp->rdev, skb);
1114 * Assumes qhp lock is held.
1116 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
1117 struct c4iw_cq *schp)
1120 int rq_flushed, sq_flushed;
1123 PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
1125 /* locking hierarchy: cq lock first, then qp lock. */
1126 spin_lock_irqsave(&rchp->lock, flag);
1127 spin_lock(&qhp->lock);
1129 if (qhp->wq.flushed) {
1130 spin_unlock(&qhp->lock);
1131 spin_unlock_irqrestore(&rchp->lock, flag);
1134 qhp->wq.flushed = 1;
1136 c4iw_flush_hw_cq(rchp);
1137 c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
1138 rq_flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
1139 spin_unlock(&qhp->lock);
1140 spin_unlock_irqrestore(&rchp->lock, flag);
1142 /* locking hierarchy: cq lock first, then qp lock. */
1143 spin_lock_irqsave(&schp->lock, flag);
1144 spin_lock(&qhp->lock);
1146 c4iw_flush_hw_cq(schp);
1147 sq_flushed = c4iw_flush_sq(qhp);
1148 spin_unlock(&qhp->lock);
1149 spin_unlock_irqrestore(&schp->lock, flag);
1152 if (t4_clear_cq_armed(&rchp->cq) &&
1153 (rq_flushed || sq_flushed)) {
1154 spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1155 (*rchp->ibcq.comp_handler)(&rchp->ibcq,
1156 rchp->ibcq.cq_context);
1157 spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1160 if (t4_clear_cq_armed(&rchp->cq) && rq_flushed) {
1161 spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1162 (*rchp->ibcq.comp_handler)(&rchp->ibcq,
1163 rchp->ibcq.cq_context);
1164 spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1166 if (t4_clear_cq_armed(&schp->cq) && sq_flushed) {
1167 spin_lock_irqsave(&schp->comp_handler_lock, flag);
1168 (*schp->ibcq.comp_handler)(&schp->ibcq,
1169 schp->ibcq.cq_context);
1170 spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1175 static void flush_qp(struct c4iw_qp *qhp)
1177 struct c4iw_cq *rchp, *schp;
1180 rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1181 schp = to_c4iw_cq(qhp->ibqp.send_cq);
1183 t4_set_wq_in_error(&qhp->wq);
1184 if (qhp->ibqp.uobject) {
1185 t4_set_cq_in_error(&rchp->cq);
1186 spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1187 (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1188 spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1190 t4_set_cq_in_error(&schp->cq);
1191 spin_lock_irqsave(&schp->comp_handler_lock, flag);
1192 (*schp->ibcq.comp_handler)(&schp->ibcq,
1193 schp->ibcq.cq_context);
1194 spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1198 __flush_qp(qhp, rchp, schp);
1201 static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1204 struct fw_ri_wr *wqe;
1206 struct sk_buff *skb;
1208 PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1211 skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1214 set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
1216 wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1217 memset(wqe, 0, sizeof *wqe);
1218 wqe->op_compl = cpu_to_be32(
1219 FW_WR_OP_V(FW_RI_INIT_WR) |
1221 wqe->flowid_len16 = cpu_to_be32(
1222 FW_WR_FLOWID_V(ep->hwtid) |
1223 FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1224 wqe->cookie = (uintptr_t)&ep->com.wr_wait;
1226 wqe->u.fini.type = FW_RI_TYPE_FINI;
1227 ret = c4iw_ofld_send(&rhp->rdev, skb);
1231 ret = c4iw_wait_for_reply(&rhp->rdev, &ep->com.wr_wait, qhp->ep->hwtid,
1232 qhp->wq.sq.qid, __func__);
1234 PDBG("%s ret %d\n", __func__, ret);
1238 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1240 PDBG("%s p2p_type = %d\n", __func__, p2p_type);
1241 memset(&init->u, 0, sizeof init->u);
1243 case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1244 init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1245 init->u.write.stag_sink = cpu_to_be32(1);
1246 init->u.write.to_sink = cpu_to_be64(1);
1247 init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1248 init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1249 sizeof(struct fw_ri_immd),
1252 case FW_RI_INIT_P2PTYPE_READ_REQ:
1253 init->u.write.opcode = FW_RI_RDMA_READ_WR;
1254 init->u.read.stag_src = cpu_to_be32(1);
1255 init->u.read.to_src_lo = cpu_to_be32(1);
1256 init->u.read.stag_sink = cpu_to_be32(1);
1257 init->u.read.to_sink_lo = cpu_to_be32(1);
1258 init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1263 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1265 struct fw_ri_wr *wqe;
1267 struct sk_buff *skb;
1269 PDBG("%s qhp %p qid 0x%x tid %u ird %u ord %u\n", __func__, qhp,
1270 qhp->wq.sq.qid, qhp->ep->hwtid, qhp->ep->ird, qhp->ep->ord);
1272 skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1277 ret = alloc_ird(rhp, qhp->attr.max_ird);
1279 qhp->attr.max_ird = 0;
1283 set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1285 wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1286 memset(wqe, 0, sizeof *wqe);
1287 wqe->op_compl = cpu_to_be32(
1288 FW_WR_OP_V(FW_RI_INIT_WR) |
1290 wqe->flowid_len16 = cpu_to_be32(
1291 FW_WR_FLOWID_V(qhp->ep->hwtid) |
1292 FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1294 wqe->cookie = (uintptr_t)&qhp->ep->com.wr_wait;
1296 wqe->u.init.type = FW_RI_TYPE_INIT;
1297 wqe->u.init.mpareqbit_p2ptype =
1298 FW_RI_WR_MPAREQBIT_V(qhp->attr.mpa_attr.initiator) |
1299 FW_RI_WR_P2PTYPE_V(qhp->attr.mpa_attr.p2p_type);
1300 wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1301 if (qhp->attr.mpa_attr.recv_marker_enabled)
1302 wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1303 if (qhp->attr.mpa_attr.xmit_marker_enabled)
1304 wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1305 if (qhp->attr.mpa_attr.crc_enabled)
1306 wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1308 wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1309 FW_RI_QP_RDMA_WRITE_ENABLE |
1310 FW_RI_QP_BIND_ENABLE;
1311 if (!qhp->ibqp.uobject)
1312 wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1313 FW_RI_QP_STAG0_ENABLE;
1314 wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1315 wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1316 wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1317 wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1318 wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1319 wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1320 wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1321 wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1322 wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1323 wqe->u.init.iss = cpu_to_be32(qhp->ep->snd_seq);
1324 wqe->u.init.irs = cpu_to_be32(qhp->ep->rcv_seq);
1325 wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1326 wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1327 rhp->rdev.lldi.vr->rq.start);
1328 if (qhp->attr.mpa_attr.initiator)
1329 build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1331 ret = c4iw_ofld_send(&rhp->rdev, skb);
1335 ret = c4iw_wait_for_reply(&rhp->rdev, &qhp->ep->com.wr_wait,
1336 qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1340 free_ird(rhp, qhp->attr.max_ird);
1342 PDBG("%s ret %d\n", __func__, ret);
1346 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1347 enum c4iw_qp_attr_mask mask,
1348 struct c4iw_qp_attributes *attrs,
1352 struct c4iw_qp_attributes newattr = qhp->attr;
1357 struct c4iw_ep *ep = NULL;
1359 PDBG("%s qhp %p sqid 0x%x rqid 0x%x ep %p state %d -> %d\n", __func__,
1360 qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep, qhp->attr.state,
1361 (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1363 mutex_lock(&qhp->mutex);
1365 /* Process attr changes if in IDLE */
1366 if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1367 if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1371 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1372 newattr.enable_rdma_read = attrs->enable_rdma_read;
1373 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1374 newattr.enable_rdma_write = attrs->enable_rdma_write;
1375 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1376 newattr.enable_bind = attrs->enable_bind;
1377 if (mask & C4IW_QP_ATTR_MAX_ORD) {
1378 if (attrs->max_ord > c4iw_max_read_depth) {
1382 newattr.max_ord = attrs->max_ord;
1384 if (mask & C4IW_QP_ATTR_MAX_IRD) {
1385 if (attrs->max_ird > cur_max_read_depth(rhp)) {
1389 newattr.max_ird = attrs->max_ird;
1391 qhp->attr = newattr;
1394 if (mask & C4IW_QP_ATTR_SQ_DB) {
1395 ret = ring_kernel_sq_db(qhp, attrs->sq_db_inc);
1398 if (mask & C4IW_QP_ATTR_RQ_DB) {
1399 ret = ring_kernel_rq_db(qhp, attrs->rq_db_inc);
1403 if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1405 if (qhp->attr.state == attrs->next_state)
1408 switch (qhp->attr.state) {
1409 case C4IW_QP_STATE_IDLE:
1410 switch (attrs->next_state) {
1411 case C4IW_QP_STATE_RTS:
1412 if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1416 if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1420 qhp->attr.mpa_attr = attrs->mpa_attr;
1421 qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1422 qhp->ep = qhp->attr.llp_stream_handle;
1423 set_state(qhp, C4IW_QP_STATE_RTS);
1426 * Ref the endpoint here and deref when we
1427 * disassociate the endpoint from the QP. This
1428 * happens in CLOSING->IDLE transition or *->ERROR
1431 c4iw_get_ep(&qhp->ep->com);
1432 ret = rdma_init(rhp, qhp);
1436 case C4IW_QP_STATE_ERROR:
1437 set_state(qhp, C4IW_QP_STATE_ERROR);
1445 case C4IW_QP_STATE_RTS:
1446 switch (attrs->next_state) {
1447 case C4IW_QP_STATE_CLOSING:
1448 BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1449 t4_set_wq_in_error(&qhp->wq);
1450 set_state(qhp, C4IW_QP_STATE_CLOSING);
1455 c4iw_get_ep(&qhp->ep->com);
1457 ret = rdma_fini(rhp, qhp, ep);
1461 case C4IW_QP_STATE_TERMINATE:
1462 t4_set_wq_in_error(&qhp->wq);
1463 set_state(qhp, C4IW_QP_STATE_TERMINATE);
1464 qhp->attr.layer_etype = attrs->layer_etype;
1465 qhp->attr.ecode = attrs->ecode;
1468 c4iw_get_ep(&qhp->ep->com);
1472 terminate = qhp->attr.send_term;
1473 ret = rdma_fini(rhp, qhp, ep);
1478 case C4IW_QP_STATE_ERROR:
1479 t4_set_wq_in_error(&qhp->wq);
1480 set_state(qhp, C4IW_QP_STATE_ERROR);
1485 c4iw_get_ep(&qhp->ep->com);
1494 case C4IW_QP_STATE_CLOSING:
1499 switch (attrs->next_state) {
1500 case C4IW_QP_STATE_IDLE:
1502 set_state(qhp, C4IW_QP_STATE_IDLE);
1503 qhp->attr.llp_stream_handle = NULL;
1504 c4iw_put_ep(&qhp->ep->com);
1506 wake_up(&qhp->wait);
1508 case C4IW_QP_STATE_ERROR:
1515 case C4IW_QP_STATE_ERROR:
1516 if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1520 if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1524 set_state(qhp, C4IW_QP_STATE_IDLE);
1526 case C4IW_QP_STATE_TERMINATE:
1534 printk(KERN_ERR "%s in a bad state %d\n",
1535 __func__, qhp->attr.state);
1542 PDBG("%s disassociating ep %p qpid 0x%x\n", __func__, qhp->ep,
1545 /* disassociate the LLP connection */
1546 qhp->attr.llp_stream_handle = NULL;
1550 set_state(qhp, C4IW_QP_STATE_ERROR);
1555 wake_up(&qhp->wait);
1557 mutex_unlock(&qhp->mutex);
1560 post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1563 * If disconnect is 1, then we need to initiate a disconnect
1564 * on the EP. This can be a normal close (RTS->CLOSING) or
1565 * an abnormal close (RTS/CLOSING->ERROR).
1568 c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1570 c4iw_put_ep(&ep->com);
1574 * If free is 1, then we've disassociated the EP from the QP
1575 * and we need to dereference the EP.
1578 c4iw_put_ep(&ep->com);
1579 PDBG("%s exit state %d\n", __func__, qhp->attr.state);
1583 int c4iw_destroy_qp(struct ib_qp *ib_qp)
1585 struct c4iw_dev *rhp;
1586 struct c4iw_qp *qhp;
1587 struct c4iw_qp_attributes attrs;
1588 struct c4iw_ucontext *ucontext;
1590 qhp = to_c4iw_qp(ib_qp);
1593 attrs.next_state = C4IW_QP_STATE_ERROR;
1594 if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1595 c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1597 c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1598 wait_event(qhp->wait, !qhp->ep);
1600 remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1601 atomic_dec(&qhp->refcnt);
1602 wait_event(qhp->wait, !atomic_read(&qhp->refcnt));
1604 spin_lock_irq(&rhp->lock);
1605 if (!list_empty(&qhp->db_fc_entry))
1606 list_del_init(&qhp->db_fc_entry);
1607 spin_unlock_irq(&rhp->lock);
1608 free_ird(rhp, qhp->attr.max_ird);
1610 ucontext = ib_qp->uobject ?
1611 to_c4iw_ucontext(ib_qp->uobject->context) : NULL;
1612 destroy_qp(&rhp->rdev, &qhp->wq,
1613 ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1615 PDBG("%s ib_qp %p qpid 0x%0x\n", __func__, ib_qp, qhp->wq.sq.qid);
1620 struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1621 struct ib_udata *udata)
1623 struct c4iw_dev *rhp;
1624 struct c4iw_qp *qhp;
1625 struct c4iw_pd *php;
1626 struct c4iw_cq *schp;
1627 struct c4iw_cq *rchp;
1628 struct c4iw_create_qp_resp uresp;
1629 unsigned int sqsize, rqsize;
1630 struct c4iw_ucontext *ucontext;
1632 struct c4iw_mm_entry *mm1, *mm2, *mm3, *mm4, *mm5 = NULL;
1634 PDBG("%s ib_pd %p\n", __func__, pd);
1636 if (attrs->qp_type != IB_QPT_RC)
1637 return ERR_PTR(-EINVAL);
1639 php = to_c4iw_pd(pd);
1641 schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1642 rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1644 return ERR_PTR(-EINVAL);
1646 if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1647 return ERR_PTR(-EINVAL);
1649 if (attrs->cap.max_recv_wr > rhp->rdev.hw_queue.t4_max_rq_size)
1650 return ERR_PTR(-E2BIG);
1651 rqsize = attrs->cap.max_recv_wr + 1;
1655 if (attrs->cap.max_send_wr > rhp->rdev.hw_queue.t4_max_sq_size)
1656 return ERR_PTR(-E2BIG);
1657 sqsize = attrs->cap.max_send_wr + 1;
1661 ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1663 qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1665 return ERR_PTR(-ENOMEM);
1666 qhp->wq.sq.size = sqsize;
1667 qhp->wq.sq.memsize =
1668 (sqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1669 sizeof(*qhp->wq.sq.queue) + 16 * sizeof(__be64);
1670 qhp->wq.sq.flush_cidx = -1;
1671 qhp->wq.rq.size = rqsize;
1672 qhp->wq.rq.memsize =
1673 (rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1674 sizeof(*qhp->wq.rq.queue);
1677 qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1678 qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1681 ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1682 ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1686 attrs->cap.max_recv_wr = rqsize - 1;
1687 attrs->cap.max_send_wr = sqsize - 1;
1688 attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1691 qhp->attr.pd = php->pdid;
1692 qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1693 qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1694 qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1695 qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1696 qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1697 qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1698 qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1699 qhp->attr.state = C4IW_QP_STATE_IDLE;
1700 qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1701 qhp->attr.enable_rdma_read = 1;
1702 qhp->attr.enable_rdma_write = 1;
1703 qhp->attr.enable_bind = 1;
1704 qhp->attr.max_ord = 0;
1705 qhp->attr.max_ird = 0;
1706 qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
1707 spin_lock_init(&qhp->lock);
1708 mutex_init(&qhp->mutex);
1709 init_waitqueue_head(&qhp->wait);
1710 atomic_set(&qhp->refcnt, 1);
1712 ret = insert_handle(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1717 mm1 = kmalloc(sizeof *mm1, GFP_KERNEL);
1722 mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
1727 mm3 = kmalloc(sizeof *mm3, GFP_KERNEL);
1732 mm4 = kmalloc(sizeof *mm4, GFP_KERNEL);
1737 if (t4_sq_onchip(&qhp->wq.sq)) {
1738 mm5 = kmalloc(sizeof *mm5, GFP_KERNEL);
1743 uresp.flags = C4IW_QPF_ONCHIP;
1746 uresp.qid_mask = rhp->rdev.qpmask;
1747 uresp.sqid = qhp->wq.sq.qid;
1748 uresp.sq_size = qhp->wq.sq.size;
1749 uresp.sq_memsize = qhp->wq.sq.memsize;
1750 uresp.rqid = qhp->wq.rq.qid;
1751 uresp.rq_size = qhp->wq.rq.size;
1752 uresp.rq_memsize = qhp->wq.rq.memsize;
1753 spin_lock(&ucontext->mmap_lock);
1755 uresp.ma_sync_key = ucontext->key;
1756 ucontext->key += PAGE_SIZE;
1758 uresp.ma_sync_key = 0;
1760 uresp.sq_key = ucontext->key;
1761 ucontext->key += PAGE_SIZE;
1762 uresp.rq_key = ucontext->key;
1763 ucontext->key += PAGE_SIZE;
1764 uresp.sq_db_gts_key = ucontext->key;
1765 ucontext->key += PAGE_SIZE;
1766 uresp.rq_db_gts_key = ucontext->key;
1767 ucontext->key += PAGE_SIZE;
1768 spin_unlock(&ucontext->mmap_lock);
1769 ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1772 mm1->key = uresp.sq_key;
1773 mm1->addr = qhp->wq.sq.phys_addr;
1774 mm1->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1775 insert_mmap(ucontext, mm1);
1776 mm2->key = uresp.rq_key;
1777 mm2->addr = virt_to_phys(qhp->wq.rq.queue);
1778 mm2->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1779 insert_mmap(ucontext, mm2);
1780 mm3->key = uresp.sq_db_gts_key;
1781 mm3->addr = (__force unsigned long)qhp->wq.sq.bar2_pa;
1782 mm3->len = PAGE_SIZE;
1783 insert_mmap(ucontext, mm3);
1784 mm4->key = uresp.rq_db_gts_key;
1785 mm4->addr = (__force unsigned long)qhp->wq.rq.bar2_pa;
1786 mm4->len = PAGE_SIZE;
1787 insert_mmap(ucontext, mm4);
1789 mm5->key = uresp.ma_sync_key;
1790 mm5->addr = (pci_resource_start(rhp->rdev.lldi.pdev, 0)
1791 + PCIE_MA_SYNC_A) & PAGE_MASK;
1792 mm5->len = PAGE_SIZE;
1793 insert_mmap(ucontext, mm5);
1796 qhp->ibqp.qp_num = qhp->wq.sq.qid;
1797 init_timer(&(qhp->timer));
1798 INIT_LIST_HEAD(&qhp->db_fc_entry);
1799 PDBG("%s sq id %u size %u memsize %zu num_entries %u "
1800 "rq id %u size %u memsize %zu num_entries %u\n", __func__,
1801 qhp->wq.sq.qid, qhp->wq.sq.size, qhp->wq.sq.memsize,
1802 attrs->cap.max_send_wr, qhp->wq.rq.qid, qhp->wq.rq.size,
1803 qhp->wq.rq.memsize, attrs->cap.max_recv_wr);
1816 remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1818 destroy_qp(&rhp->rdev, &qhp->wq,
1819 ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1822 return ERR_PTR(ret);
1825 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1826 int attr_mask, struct ib_udata *udata)
1828 struct c4iw_dev *rhp;
1829 struct c4iw_qp *qhp;
1830 enum c4iw_qp_attr_mask mask = 0;
1831 struct c4iw_qp_attributes attrs;
1833 PDBG("%s ib_qp %p\n", __func__, ibqp);
1835 /* iwarp does not support the RTR state */
1836 if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1837 attr_mask &= ~IB_QP_STATE;
1839 /* Make sure we still have something left to do */
1843 memset(&attrs, 0, sizeof attrs);
1844 qhp = to_c4iw_qp(ibqp);
1847 attrs.next_state = c4iw_convert_state(attr->qp_state);
1848 attrs.enable_rdma_read = (attr->qp_access_flags &
1849 IB_ACCESS_REMOTE_READ) ? 1 : 0;
1850 attrs.enable_rdma_write = (attr->qp_access_flags &
1851 IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1852 attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1855 mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1856 mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1857 (C4IW_QP_ATTR_ENABLE_RDMA_READ |
1858 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1859 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1862 * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
1863 * ringing the queue db when we're in DB_FULL mode.
1864 * Only allow this on T4 devices.
1866 attrs.sq_db_inc = attr->sq_psn;
1867 attrs.rq_db_inc = attr->rq_psn;
1868 mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
1869 mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
1870 if (!is_t4(to_c4iw_qp(ibqp)->rhp->rdev.lldi.adapter_type) &&
1871 (mask & (C4IW_QP_ATTR_SQ_DB|C4IW_QP_ATTR_RQ_DB)))
1874 return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
1877 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
1879 PDBG("%s ib_dev %p qpn 0x%x\n", __func__, dev, qpn);
1880 return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
1883 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1884 int attr_mask, struct ib_qp_init_attr *init_attr)
1886 struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
1888 memset(attr, 0, sizeof *attr);
1889 memset(init_attr, 0, sizeof *init_attr);
1890 attr->qp_state = to_ib_qp_state(qhp->attr.state);
1891 init_attr->cap.max_send_wr = qhp->attr.sq_num_entries;
1892 init_attr->cap.max_recv_wr = qhp->attr.rq_num_entries;
1893 init_attr->cap.max_send_sge = qhp->attr.sq_max_sges;
1894 init_attr->cap.max_recv_sge = qhp->attr.sq_max_sges;
1895 init_attr->cap.max_inline_data = T4_MAX_SEND_INLINE;
1896 init_attr->sq_sig_type = qhp->sq_sig_all ? IB_SIGNAL_ALL_WR : 0;