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;
189 if (is_t4(rdev->lldi.adapter_type))
192 return rdev->bar2_kva + bar2_qoffset;
195 static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
196 struct t4_cq *rcq, struct t4_cq *scq,
197 struct c4iw_dev_ucontext *uctx,
198 struct c4iw_wr_wait *wr_waitp)
200 int user = (uctx != &rdev->uctx);
201 struct fw_ri_res_wr *res_wr;
202 struct fw_ri_res *res;
208 wq->sq.qid = c4iw_get_qpid(rdev, uctx);
212 wq->rq.qid = c4iw_get_qpid(rdev, uctx);
219 wq->sq.sw_sq = kzalloc(wq->sq.size * sizeof *wq->sq.sw_sq,
226 wq->rq.sw_rq = kzalloc(wq->rq.size * sizeof *wq->rq.sw_rq,
235 * RQT must be a power of 2 and at least 16 deep.
237 wq->rq.rqt_size = roundup_pow_of_two(max_t(u16, wq->rq.size, 16));
238 wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size);
239 if (!wq->rq.rqt_hwaddr) {
244 ret = alloc_sq(rdev, &wq->sq, user);
247 memset(wq->sq.queue, 0, wq->sq.memsize);
248 dma_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
250 wq->rq.queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev),
251 wq->rq.memsize, &(wq->rq.dma_addr),
257 pr_debug("sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx\n",
259 (unsigned long long)virt_to_phys(wq->sq.queue),
261 (unsigned long long)virt_to_phys(wq->rq.queue));
262 memset(wq->rq.queue, 0, wq->rq.memsize);
263 dma_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
265 wq->db = rdev->lldi.db_reg;
267 wq->sq.bar2_va = c4iw_bar2_addrs(rdev, wq->sq.qid, T4_BAR2_QTYPE_EGRESS,
269 user ? &wq->sq.bar2_pa : NULL);
270 wq->rq.bar2_va = c4iw_bar2_addrs(rdev, wq->rq.qid, T4_BAR2_QTYPE_EGRESS,
272 user ? &wq->rq.bar2_pa : NULL);
275 * User mode must have bar2 access.
277 if (user && (!wq->sq.bar2_pa || !wq->rq.bar2_pa)) {
278 pr_warn("%s: sqid %u or rqid %u not in BAR2 range\n",
279 pci_name(rdev->lldi.pdev), wq->sq.qid, wq->rq.qid);
286 /* build fw_ri_res_wr */
287 wr_len = sizeof *res_wr + 2 * sizeof *res;
289 skb = alloc_skb(wr_len, GFP_KERNEL);
294 set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
296 res_wr = __skb_put_zero(skb, wr_len);
297 res_wr->op_nres = cpu_to_be32(
298 FW_WR_OP_V(FW_RI_RES_WR) |
299 FW_RI_RES_WR_NRES_V(2) |
301 res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
302 res_wr->cookie = (uintptr_t)wr_waitp;
304 res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
305 res->u.sqrq.op = FW_RI_RES_OP_WRITE;
308 * eqsize is the number of 64B entries plus the status page size.
310 eqsize = wq->sq.size * T4_SQ_NUM_SLOTS +
311 rdev->hw_queue.t4_eq_status_entries;
313 res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
314 FW_RI_RES_WR_HOSTFCMODE_V(0) | /* no host cidx updates */
315 FW_RI_RES_WR_CPRIO_V(0) | /* don't keep in chip cache */
316 FW_RI_RES_WR_PCIECHN_V(0) | /* set by uP at ri_init time */
317 (t4_sq_onchip(&wq->sq) ? FW_RI_RES_WR_ONCHIP_F : 0) |
318 FW_RI_RES_WR_IQID_V(scq->cqid));
319 res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
320 FW_RI_RES_WR_DCAEN_V(0) |
321 FW_RI_RES_WR_DCACPU_V(0) |
322 FW_RI_RES_WR_FBMIN_V(2) |
323 (t4_sq_onchip(&wq->sq) ? FW_RI_RES_WR_FBMAX_V(2) :
324 FW_RI_RES_WR_FBMAX_V(3)) |
325 FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
326 FW_RI_RES_WR_CIDXFTHRESH_V(0) |
327 FW_RI_RES_WR_EQSIZE_V(eqsize));
328 res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
329 res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
331 res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
332 res->u.sqrq.op = FW_RI_RES_OP_WRITE;
335 * eqsize is the number of 64B entries plus the status page size.
337 eqsize = wq->rq.size * T4_RQ_NUM_SLOTS +
338 rdev->hw_queue.t4_eq_status_entries;
339 res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
340 FW_RI_RES_WR_HOSTFCMODE_V(0) | /* no host cidx updates */
341 FW_RI_RES_WR_CPRIO_V(0) | /* don't keep in chip cache */
342 FW_RI_RES_WR_PCIECHN_V(0) | /* set by uP at ri_init time */
343 FW_RI_RES_WR_IQID_V(rcq->cqid));
344 res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
345 FW_RI_RES_WR_DCAEN_V(0) |
346 FW_RI_RES_WR_DCACPU_V(0) |
347 FW_RI_RES_WR_FBMIN_V(2) |
348 FW_RI_RES_WR_FBMAX_V(3) |
349 FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
350 FW_RI_RES_WR_CIDXFTHRESH_V(0) |
351 FW_RI_RES_WR_EQSIZE_V(eqsize));
352 res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
353 res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
355 c4iw_init_wr_wait(wr_waitp);
356 ret = c4iw_ref_send_wait(rdev, skb, wr_waitp, 0, wq->sq.qid, __func__);
360 pr_debug("sqid 0x%x rqid 0x%x kdb 0x%p sq_bar2_addr %p rq_bar2_addr %p\n",
361 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)
568 if (wr->num_sge && wr->sg_list[0].length) {
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 void build_tpte_memreg(struct fw_ri_fr_nsmr_tpte_wr *fr,
609 struct ib_reg_wr *wr, struct c4iw_mr *mhp,
612 __be64 *p = (__be64 *)fr->pbl;
614 fr->r2 = cpu_to_be32(0);
615 fr->stag = cpu_to_be32(mhp->ibmr.rkey);
617 fr->tpte.valid_to_pdid = cpu_to_be32(FW_RI_TPTE_VALID_F |
618 FW_RI_TPTE_STAGKEY_V((mhp->ibmr.rkey & FW_RI_TPTE_STAGKEY_M)) |
619 FW_RI_TPTE_STAGSTATE_V(1) |
620 FW_RI_TPTE_STAGTYPE_V(FW_RI_STAG_NSMR) |
621 FW_RI_TPTE_PDID_V(mhp->attr.pdid));
622 fr->tpte.locread_to_qpid = cpu_to_be32(
623 FW_RI_TPTE_PERM_V(c4iw_ib_to_tpt_access(wr->access)) |
624 FW_RI_TPTE_ADDRTYPE_V(FW_RI_VA_BASED_TO) |
625 FW_RI_TPTE_PS_V(ilog2(wr->mr->page_size) - 12));
626 fr->tpte.nosnoop_pbladdr = cpu_to_be32(FW_RI_TPTE_PBLADDR_V(
627 PBL_OFF(&mhp->rhp->rdev, mhp->attr.pbl_addr)>>3));
628 fr->tpte.dca_mwbcnt_pstag = cpu_to_be32(0);
629 fr->tpte.len_hi = cpu_to_be32(0);
630 fr->tpte.len_lo = cpu_to_be32(mhp->ibmr.length);
631 fr->tpte.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
632 fr->tpte.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova & 0xffffffff);
634 p[0] = cpu_to_be64((u64)mhp->mpl[0]);
635 p[1] = cpu_to_be64((u64)mhp->mpl[1]);
637 *len16 = DIV_ROUND_UP(sizeof(*fr), 16);
640 static int build_memreg(struct t4_sq *sq, union t4_wr *wqe,
641 struct ib_reg_wr *wr, struct c4iw_mr *mhp, u8 *len16,
644 struct fw_ri_immd *imdp;
647 int pbllen = roundup(mhp->mpl_len * sizeof(u64), 32);
650 if (mhp->mpl_len > t4_max_fr_depth(dsgl_supported && use_dsgl))
653 wqe->fr.qpbinde_to_dcacpu = 0;
654 wqe->fr.pgsz_shift = ilog2(wr->mr->page_size) - 12;
655 wqe->fr.addr_type = FW_RI_VA_BASED_TO;
656 wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->access);
658 wqe->fr.len_lo = cpu_to_be32(mhp->ibmr.length);
659 wqe->fr.stag = cpu_to_be32(wr->key);
660 wqe->fr.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
661 wqe->fr.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova &
664 if (dsgl_supported && use_dsgl && (pbllen > max_fr_immd)) {
665 struct fw_ri_dsgl *sglp;
667 for (i = 0; i < mhp->mpl_len; i++)
668 mhp->mpl[i] = (__force u64)cpu_to_be64((u64)mhp->mpl[i]);
670 sglp = (struct fw_ri_dsgl *)(&wqe->fr + 1);
671 sglp->op = FW_RI_DATA_DSGL;
673 sglp->nsge = cpu_to_be16(1);
674 sglp->addr0 = cpu_to_be64(mhp->mpl_addr);
675 sglp->len0 = cpu_to_be32(pbllen);
677 *len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*sglp), 16);
679 imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
680 imdp->op = FW_RI_DATA_IMMD;
683 imdp->immdlen = cpu_to_be32(pbllen);
684 p = (__be64 *)(imdp + 1);
686 for (i = 0; i < mhp->mpl_len; i++) {
687 *p = cpu_to_be64((u64)mhp->mpl[i]);
689 if (++p == (__be64 *)&sq->queue[sq->size])
690 p = (__be64 *)sq->queue;
695 if (++p == (__be64 *)&sq->queue[sq->size])
696 p = (__be64 *)sq->queue;
698 *len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*imdp)
704 static int build_inv_stag(union t4_wr *wqe, struct ib_send_wr *wr, u8 *len16)
706 wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
708 *len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
712 static void free_qp_work(struct work_struct *work)
714 struct c4iw_ucontext *ucontext;
716 struct c4iw_dev *rhp;
718 qhp = container_of(work, struct c4iw_qp, free_work);
719 ucontext = qhp->ucontext;
722 pr_debug("qhp %p ucontext %p\n", qhp, ucontext);
723 destroy_qp(&rhp->rdev, &qhp->wq,
724 ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
727 c4iw_put_ucontext(ucontext);
728 c4iw_put_wr_wait(qhp->wr_waitp);
732 static void queue_qp_free(struct kref *kref)
736 qhp = container_of(kref, struct c4iw_qp, kref);
737 pr_debug("qhp %p\n", qhp);
738 queue_work(qhp->rhp->rdev.free_workq, &qhp->free_work);
741 void c4iw_qp_add_ref(struct ib_qp *qp)
743 pr_debug("ib_qp %p\n", qp);
744 kref_get(&to_c4iw_qp(qp)->kref);
747 void c4iw_qp_rem_ref(struct ib_qp *qp)
749 pr_debug("ib_qp %p\n", qp);
750 kref_put(&to_c4iw_qp(qp)->kref, queue_qp_free);
753 static void add_to_fc_list(struct list_head *head, struct list_head *entry)
755 if (list_empty(entry))
756 list_add_tail(entry, head);
759 static int ring_kernel_sq_db(struct c4iw_qp *qhp, u16 inc)
763 spin_lock_irqsave(&qhp->rhp->lock, flags);
764 spin_lock(&qhp->lock);
765 if (qhp->rhp->db_state == NORMAL)
766 t4_ring_sq_db(&qhp->wq, inc, NULL);
768 add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
769 qhp->wq.sq.wq_pidx_inc += inc;
771 spin_unlock(&qhp->lock);
772 spin_unlock_irqrestore(&qhp->rhp->lock, flags);
776 static int ring_kernel_rq_db(struct c4iw_qp *qhp, u16 inc)
780 spin_lock_irqsave(&qhp->rhp->lock, flags);
781 spin_lock(&qhp->lock);
782 if (qhp->rhp->db_state == NORMAL)
783 t4_ring_rq_db(&qhp->wq, inc, NULL);
785 add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
786 qhp->wq.rq.wq_pidx_inc += inc;
788 spin_unlock(&qhp->lock);
789 spin_unlock_irqrestore(&qhp->rhp->lock, flags);
793 static void complete_sq_drain_wr(struct c4iw_qp *qhp, struct ib_send_wr *wr)
795 struct t4_cqe cqe = {};
796 struct c4iw_cq *schp;
800 schp = to_c4iw_cq(qhp->ibqp.send_cq);
803 cqe.u.drain_cookie = wr->wr_id;
804 cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
805 CQE_OPCODE_V(C4IW_DRAIN_OPCODE) |
808 CQE_QPID_V(qhp->wq.sq.qid));
810 spin_lock_irqsave(&schp->lock, flag);
811 cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
812 cq->sw_queue[cq->sw_pidx] = cqe;
814 spin_unlock_irqrestore(&schp->lock, flag);
816 spin_lock_irqsave(&schp->comp_handler_lock, flag);
817 (*schp->ibcq.comp_handler)(&schp->ibcq,
818 schp->ibcq.cq_context);
819 spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
822 static void complete_rq_drain_wr(struct c4iw_qp *qhp, struct ib_recv_wr *wr)
824 struct t4_cqe cqe = {};
825 struct c4iw_cq *rchp;
829 rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
832 cqe.u.drain_cookie = wr->wr_id;
833 cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
834 CQE_OPCODE_V(C4IW_DRAIN_OPCODE) |
837 CQE_QPID_V(qhp->wq.sq.qid));
839 spin_lock_irqsave(&rchp->lock, flag);
840 cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
841 cq->sw_queue[cq->sw_pidx] = cqe;
843 spin_unlock_irqrestore(&rchp->lock, flag);
845 spin_lock_irqsave(&rchp->comp_handler_lock, flag);
846 (*rchp->ibcq.comp_handler)(&rchp->ibcq,
847 rchp->ibcq.cq_context);
848 spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
851 int c4iw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
852 struct ib_send_wr **bad_wr)
856 enum fw_wr_opcodes fw_opcode = 0;
857 enum fw_ri_wr_flags fw_flags;
859 union t4_wr *wqe = NULL;
861 struct t4_swsqe *swsqe;
865 qhp = to_c4iw_qp(ibqp);
866 spin_lock_irqsave(&qhp->lock, flag);
867 if (t4_wq_in_error(&qhp->wq)) {
868 spin_unlock_irqrestore(&qhp->lock, flag);
869 complete_sq_drain_wr(qhp, wr);
872 num_wrs = t4_sq_avail(&qhp->wq);
874 spin_unlock_irqrestore(&qhp->lock, flag);
884 wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
885 qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
888 if (wr->send_flags & IB_SEND_SOLICITED)
889 fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
890 if (wr->send_flags & IB_SEND_SIGNALED || qhp->sq_sig_all)
891 fw_flags |= FW_RI_COMPLETION_FLAG;
892 swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
893 switch (wr->opcode) {
894 case IB_WR_SEND_WITH_INV:
896 if (wr->send_flags & IB_SEND_FENCE)
897 fw_flags |= FW_RI_READ_FENCE_FLAG;
898 fw_opcode = FW_RI_SEND_WR;
899 if (wr->opcode == IB_WR_SEND)
900 swsqe->opcode = FW_RI_SEND;
902 swsqe->opcode = FW_RI_SEND_WITH_INV;
903 err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
905 case IB_WR_RDMA_WRITE:
906 fw_opcode = FW_RI_RDMA_WRITE_WR;
907 swsqe->opcode = FW_RI_RDMA_WRITE;
908 err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
910 case IB_WR_RDMA_READ:
911 case IB_WR_RDMA_READ_WITH_INV:
912 fw_opcode = FW_RI_RDMA_READ_WR;
913 swsqe->opcode = FW_RI_READ_REQ;
914 if (wr->opcode == IB_WR_RDMA_READ_WITH_INV) {
915 c4iw_invalidate_mr(qhp->rhp,
916 wr->sg_list[0].lkey);
917 fw_flags = FW_RI_RDMA_READ_INVALIDATE;
921 err = build_rdma_read(wqe, wr, &len16);
924 swsqe->read_len = wr->sg_list[0].length;
925 if (!qhp->wq.sq.oldest_read)
926 qhp->wq.sq.oldest_read = swsqe;
929 struct c4iw_mr *mhp = to_c4iw_mr(reg_wr(wr)->mr);
931 swsqe->opcode = FW_RI_FAST_REGISTER;
932 if (qhp->rhp->rdev.lldi.fr_nsmr_tpte_wr_support &&
933 !mhp->attr.state && mhp->mpl_len <= 2) {
934 fw_opcode = FW_RI_FR_NSMR_TPTE_WR;
935 build_tpte_memreg(&wqe->fr_tpte, reg_wr(wr),
938 fw_opcode = FW_RI_FR_NSMR_WR;
939 err = build_memreg(&qhp->wq.sq, wqe, reg_wr(wr),
941 qhp->rhp->rdev.lldi.ulptx_memwrite_dsgl);
948 case IB_WR_LOCAL_INV:
949 if (wr->send_flags & IB_SEND_FENCE)
950 fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
951 fw_opcode = FW_RI_INV_LSTAG_WR;
952 swsqe->opcode = FW_RI_LOCAL_INV;
953 err = build_inv_stag(wqe, wr, &len16);
954 c4iw_invalidate_mr(qhp->rhp, wr->ex.invalidate_rkey);
957 pr_warn("%s post of type=%d TBD!\n", __func__,
965 swsqe->idx = qhp->wq.sq.pidx;
967 swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
970 swsqe->wr_id = wr->wr_id;
972 swsqe->sge_ts = cxgb4_read_sge_timestamp(
973 qhp->rhp->rdev.lldi.ports[0]);
974 getnstimeofday(&swsqe->host_ts);
977 init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
979 pr_debug("cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u\n",
980 (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
981 swsqe->opcode, swsqe->read_len);
984 t4_sq_produce(&qhp->wq, len16);
985 idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
987 if (!qhp->rhp->rdev.status_page->db_off) {
988 t4_ring_sq_db(&qhp->wq, idx, wqe);
989 spin_unlock_irqrestore(&qhp->lock, flag);
991 spin_unlock_irqrestore(&qhp->lock, flag);
992 ring_kernel_sq_db(qhp, idx);
997 int c4iw_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
998 struct ib_recv_wr **bad_wr)
1001 struct c4iw_qp *qhp;
1002 union t4_recv_wr *wqe = NULL;
1008 qhp = to_c4iw_qp(ibqp);
1009 spin_lock_irqsave(&qhp->lock, flag);
1010 if (t4_wq_in_error(&qhp->wq)) {
1011 spin_unlock_irqrestore(&qhp->lock, flag);
1012 complete_rq_drain_wr(qhp, wr);
1015 num_wrs = t4_rq_avail(&qhp->wq);
1017 spin_unlock_irqrestore(&qhp->lock, flag);
1022 if (wr->num_sge > T4_MAX_RECV_SGE) {
1027 wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
1028 qhp->wq.rq.wq_pidx *
1031 err = build_rdma_recv(qhp, wqe, wr, &len16);
1039 qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
1041 qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].sge_ts =
1042 cxgb4_read_sge_timestamp(
1043 qhp->rhp->rdev.lldi.ports[0]);
1045 &qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].host_ts);
1048 wqe->recv.opcode = FW_RI_RECV_WR;
1050 wqe->recv.wrid = qhp->wq.rq.pidx;
1051 wqe->recv.r2[0] = 0;
1052 wqe->recv.r2[1] = 0;
1053 wqe->recv.r2[2] = 0;
1054 wqe->recv.len16 = len16;
1055 pr_debug("cookie 0x%llx pidx %u\n",
1056 (unsigned long long)wr->wr_id, qhp->wq.rq.pidx);
1057 t4_rq_produce(&qhp->wq, len16);
1058 idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
1062 if (!qhp->rhp->rdev.status_page->db_off) {
1063 t4_ring_rq_db(&qhp->wq, idx, wqe);
1064 spin_unlock_irqrestore(&qhp->lock, flag);
1066 spin_unlock_irqrestore(&qhp->lock, flag);
1067 ring_kernel_rq_db(qhp, idx);
1072 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
1082 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1087 status = CQE_STATUS(err_cqe);
1088 opcode = CQE_OPCODE(err_cqe);
1089 rqtype = RQ_TYPE(err_cqe);
1090 send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
1091 (opcode == FW_RI_SEND_WITH_SE_INV);
1092 tagged = (opcode == FW_RI_RDMA_WRITE) ||
1093 (rqtype && (opcode == FW_RI_READ_RESP));
1098 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1099 *ecode = RDMAP_CANT_INV_STAG;
1101 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1102 *ecode = RDMAP_INV_STAG;
1106 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1107 if ((opcode == FW_RI_SEND_WITH_INV) ||
1108 (opcode == FW_RI_SEND_WITH_SE_INV))
1109 *ecode = RDMAP_CANT_INV_STAG;
1111 *ecode = RDMAP_STAG_NOT_ASSOC;
1114 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1115 *ecode = RDMAP_STAG_NOT_ASSOC;
1118 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1119 *ecode = RDMAP_ACC_VIOL;
1122 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1123 *ecode = RDMAP_TO_WRAP;
1127 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1128 *ecode = DDPT_BASE_BOUNDS;
1130 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1131 *ecode = RDMAP_BASE_BOUNDS;
1134 case T4_ERR_INVALIDATE_SHARED_MR:
1135 case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
1136 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1137 *ecode = RDMAP_CANT_INV_STAG;
1140 case T4_ERR_ECC_PSTAG:
1141 case T4_ERR_INTERNAL_ERR:
1142 *layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
1145 case T4_ERR_OUT_OF_RQE:
1146 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1147 *ecode = DDPU_INV_MSN_NOBUF;
1149 case T4_ERR_PBL_ADDR_BOUND:
1150 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1151 *ecode = DDPT_BASE_BOUNDS;
1154 *layer_type = LAYER_MPA|DDP_LLP;
1155 *ecode = MPA_CRC_ERR;
1158 *layer_type = LAYER_MPA|DDP_LLP;
1159 *ecode = MPA_MARKER_ERR;
1161 case T4_ERR_PDU_LEN_ERR:
1162 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1163 *ecode = DDPU_MSG_TOOBIG;
1165 case T4_ERR_DDP_VERSION:
1167 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1168 *ecode = DDPT_INV_VERS;
1170 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1171 *ecode = DDPU_INV_VERS;
1174 case T4_ERR_RDMA_VERSION:
1175 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1176 *ecode = RDMAP_INV_VERS;
1179 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1180 *ecode = RDMAP_INV_OPCODE;
1182 case T4_ERR_DDP_QUEUE_NUM:
1183 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1184 *ecode = DDPU_INV_QN;
1187 case T4_ERR_MSN_GAP:
1188 case T4_ERR_MSN_RANGE:
1189 case T4_ERR_IRD_OVERFLOW:
1190 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1191 *ecode = DDPU_INV_MSN_RANGE;
1194 *layer_type = LAYER_DDP|DDP_LOCAL_CATA;
1198 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1199 *ecode = DDPU_INV_MO;
1202 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1208 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
1211 struct fw_ri_wr *wqe;
1212 struct sk_buff *skb;
1213 struct terminate_message *term;
1215 pr_debug("qhp %p qid 0x%x tid %u\n", qhp, qhp->wq.sq.qid,
1218 skb = skb_dequeue(&qhp->ep->com.ep_skb_list);
1222 set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1224 wqe = __skb_put(skb, sizeof(*wqe));
1225 memset(wqe, 0, sizeof *wqe);
1226 wqe->op_compl = cpu_to_be32(FW_WR_OP_V(FW_RI_INIT_WR));
1227 wqe->flowid_len16 = cpu_to_be32(
1228 FW_WR_FLOWID_V(qhp->ep->hwtid) |
1229 FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1231 wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
1232 wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
1233 term = (struct terminate_message *)wqe->u.terminate.termmsg;
1234 if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
1235 term->layer_etype = qhp->attr.layer_etype;
1236 term->ecode = qhp->attr.ecode;
1238 build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
1239 c4iw_ofld_send(&qhp->rhp->rdev, skb);
1243 * Assumes qhp lock is held.
1245 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
1246 struct c4iw_cq *schp)
1249 int rq_flushed, sq_flushed;
1252 pr_debug("qhp %p rchp %p schp %p\n", qhp, rchp, schp);
1254 /* locking hierarchy: cq lock first, then qp lock. */
1255 spin_lock_irqsave(&rchp->lock, flag);
1256 spin_lock(&qhp->lock);
1258 if (qhp->wq.flushed) {
1259 spin_unlock(&qhp->lock);
1260 spin_unlock_irqrestore(&rchp->lock, flag);
1263 qhp->wq.flushed = 1;
1265 c4iw_flush_hw_cq(rchp);
1266 c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
1267 rq_flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
1268 spin_unlock(&qhp->lock);
1269 spin_unlock_irqrestore(&rchp->lock, flag);
1271 /* locking hierarchy: cq lock first, then qp lock. */
1272 spin_lock_irqsave(&schp->lock, flag);
1273 spin_lock(&qhp->lock);
1275 c4iw_flush_hw_cq(schp);
1276 sq_flushed = c4iw_flush_sq(qhp);
1277 spin_unlock(&qhp->lock);
1278 spin_unlock_irqrestore(&schp->lock, flag);
1281 if (t4_clear_cq_armed(&rchp->cq) &&
1282 (rq_flushed || sq_flushed)) {
1283 spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1284 (*rchp->ibcq.comp_handler)(&rchp->ibcq,
1285 rchp->ibcq.cq_context);
1286 spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1289 if (t4_clear_cq_armed(&rchp->cq) && rq_flushed) {
1290 spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1291 (*rchp->ibcq.comp_handler)(&rchp->ibcq,
1292 rchp->ibcq.cq_context);
1293 spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1295 if (t4_clear_cq_armed(&schp->cq) && sq_flushed) {
1296 spin_lock_irqsave(&schp->comp_handler_lock, flag);
1297 (*schp->ibcq.comp_handler)(&schp->ibcq,
1298 schp->ibcq.cq_context);
1299 spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1304 static void flush_qp(struct c4iw_qp *qhp)
1306 struct c4iw_cq *rchp, *schp;
1309 rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1310 schp = to_c4iw_cq(qhp->ibqp.send_cq);
1312 t4_set_wq_in_error(&qhp->wq);
1313 if (qhp->ibqp.uobject) {
1314 t4_set_cq_in_error(&rchp->cq);
1315 spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1316 (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1317 spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1319 t4_set_cq_in_error(&schp->cq);
1320 spin_lock_irqsave(&schp->comp_handler_lock, flag);
1321 (*schp->ibcq.comp_handler)(&schp->ibcq,
1322 schp->ibcq.cq_context);
1323 spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1327 __flush_qp(qhp, rchp, schp);
1330 static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1333 struct fw_ri_wr *wqe;
1335 struct sk_buff *skb;
1337 pr_debug("qhp %p qid 0x%x tid %u\n", qhp, qhp->wq.sq.qid, ep->hwtid);
1339 skb = skb_dequeue(&ep->com.ep_skb_list);
1343 set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
1345 wqe = __skb_put(skb, sizeof(*wqe));
1346 memset(wqe, 0, sizeof *wqe);
1347 wqe->op_compl = cpu_to_be32(
1348 FW_WR_OP_V(FW_RI_INIT_WR) |
1350 wqe->flowid_len16 = cpu_to_be32(
1351 FW_WR_FLOWID_V(ep->hwtid) |
1352 FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1353 wqe->cookie = (uintptr_t)ep->com.wr_waitp;
1355 wqe->u.fini.type = FW_RI_TYPE_FINI;
1357 ret = c4iw_ref_send_wait(&rhp->rdev, skb, ep->com.wr_waitp,
1358 qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1360 pr_debug("ret %d\n", ret);
1364 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1366 pr_debug("p2p_type = %d\n", p2p_type);
1367 memset(&init->u, 0, sizeof init->u);
1369 case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1370 init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1371 init->u.write.stag_sink = cpu_to_be32(1);
1372 init->u.write.to_sink = cpu_to_be64(1);
1373 init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1374 init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1375 sizeof(struct fw_ri_immd),
1378 case FW_RI_INIT_P2PTYPE_READ_REQ:
1379 init->u.write.opcode = FW_RI_RDMA_READ_WR;
1380 init->u.read.stag_src = cpu_to_be32(1);
1381 init->u.read.to_src_lo = cpu_to_be32(1);
1382 init->u.read.stag_sink = cpu_to_be32(1);
1383 init->u.read.to_sink_lo = cpu_to_be32(1);
1384 init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1389 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1391 struct fw_ri_wr *wqe;
1393 struct sk_buff *skb;
1395 pr_debug("qhp %p qid 0x%x tid %u ird %u ord %u\n", qhp,
1396 qhp->wq.sq.qid, qhp->ep->hwtid, qhp->ep->ird, qhp->ep->ord);
1398 skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1403 ret = alloc_ird(rhp, qhp->attr.max_ird);
1405 qhp->attr.max_ird = 0;
1409 set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1411 wqe = __skb_put(skb, sizeof(*wqe));
1412 memset(wqe, 0, sizeof *wqe);
1413 wqe->op_compl = cpu_to_be32(
1414 FW_WR_OP_V(FW_RI_INIT_WR) |
1416 wqe->flowid_len16 = cpu_to_be32(
1417 FW_WR_FLOWID_V(qhp->ep->hwtid) |
1418 FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1420 wqe->cookie = (uintptr_t)qhp->ep->com.wr_waitp;
1422 wqe->u.init.type = FW_RI_TYPE_INIT;
1423 wqe->u.init.mpareqbit_p2ptype =
1424 FW_RI_WR_MPAREQBIT_V(qhp->attr.mpa_attr.initiator) |
1425 FW_RI_WR_P2PTYPE_V(qhp->attr.mpa_attr.p2p_type);
1426 wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1427 if (qhp->attr.mpa_attr.recv_marker_enabled)
1428 wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1429 if (qhp->attr.mpa_attr.xmit_marker_enabled)
1430 wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1431 if (qhp->attr.mpa_attr.crc_enabled)
1432 wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1434 wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1435 FW_RI_QP_RDMA_WRITE_ENABLE |
1436 FW_RI_QP_BIND_ENABLE;
1437 if (!qhp->ibqp.uobject)
1438 wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1439 FW_RI_QP_STAG0_ENABLE;
1440 wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1441 wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1442 wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1443 wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1444 wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1445 wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1446 wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1447 wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1448 wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1449 wqe->u.init.iss = cpu_to_be32(qhp->ep->snd_seq);
1450 wqe->u.init.irs = cpu_to_be32(qhp->ep->rcv_seq);
1451 wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1452 wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1453 rhp->rdev.lldi.vr->rq.start);
1454 if (qhp->attr.mpa_attr.initiator)
1455 build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1457 ret = c4iw_ref_send_wait(&rhp->rdev, skb, qhp->ep->com.wr_waitp,
1458 qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1462 free_ird(rhp, qhp->attr.max_ird);
1464 pr_debug("ret %d\n", ret);
1468 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1469 enum c4iw_qp_attr_mask mask,
1470 struct c4iw_qp_attributes *attrs,
1474 struct c4iw_qp_attributes newattr = qhp->attr;
1479 struct c4iw_ep *ep = NULL;
1481 pr_debug("qhp %p sqid 0x%x rqid 0x%x ep %p state %d -> %d\n",
1482 qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep, qhp->attr.state,
1483 (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1485 mutex_lock(&qhp->mutex);
1487 /* Process attr changes if in IDLE */
1488 if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1489 if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1493 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1494 newattr.enable_rdma_read = attrs->enable_rdma_read;
1495 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1496 newattr.enable_rdma_write = attrs->enable_rdma_write;
1497 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1498 newattr.enable_bind = attrs->enable_bind;
1499 if (mask & C4IW_QP_ATTR_MAX_ORD) {
1500 if (attrs->max_ord > c4iw_max_read_depth) {
1504 newattr.max_ord = attrs->max_ord;
1506 if (mask & C4IW_QP_ATTR_MAX_IRD) {
1507 if (attrs->max_ird > cur_max_read_depth(rhp)) {
1511 newattr.max_ird = attrs->max_ird;
1513 qhp->attr = newattr;
1516 if (mask & C4IW_QP_ATTR_SQ_DB) {
1517 ret = ring_kernel_sq_db(qhp, attrs->sq_db_inc);
1520 if (mask & C4IW_QP_ATTR_RQ_DB) {
1521 ret = ring_kernel_rq_db(qhp, attrs->rq_db_inc);
1525 if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1527 if (qhp->attr.state == attrs->next_state)
1530 switch (qhp->attr.state) {
1531 case C4IW_QP_STATE_IDLE:
1532 switch (attrs->next_state) {
1533 case C4IW_QP_STATE_RTS:
1534 if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1538 if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1542 qhp->attr.mpa_attr = attrs->mpa_attr;
1543 qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1544 qhp->ep = qhp->attr.llp_stream_handle;
1545 set_state(qhp, C4IW_QP_STATE_RTS);
1548 * Ref the endpoint here and deref when we
1549 * disassociate the endpoint from the QP. This
1550 * happens in CLOSING->IDLE transition or *->ERROR
1553 c4iw_get_ep(&qhp->ep->com);
1554 ret = rdma_init(rhp, qhp);
1558 case C4IW_QP_STATE_ERROR:
1559 set_state(qhp, C4IW_QP_STATE_ERROR);
1567 case C4IW_QP_STATE_RTS:
1568 switch (attrs->next_state) {
1569 case C4IW_QP_STATE_CLOSING:
1570 t4_set_wq_in_error(&qhp->wq);
1571 set_state(qhp, C4IW_QP_STATE_CLOSING);
1576 c4iw_get_ep(&qhp->ep->com);
1578 ret = rdma_fini(rhp, qhp, ep);
1582 case C4IW_QP_STATE_TERMINATE:
1583 t4_set_wq_in_error(&qhp->wq);
1584 set_state(qhp, C4IW_QP_STATE_TERMINATE);
1585 qhp->attr.layer_etype = attrs->layer_etype;
1586 qhp->attr.ecode = attrs->ecode;
1589 c4iw_get_ep(&qhp->ep->com);
1593 terminate = qhp->attr.send_term;
1594 ret = rdma_fini(rhp, qhp, ep);
1599 case C4IW_QP_STATE_ERROR:
1600 t4_set_wq_in_error(&qhp->wq);
1601 set_state(qhp, C4IW_QP_STATE_ERROR);
1606 c4iw_get_ep(&qhp->ep->com);
1615 case C4IW_QP_STATE_CLOSING:
1618 * Allow kernel users to move to ERROR for qp draining.
1620 if (!internal && (qhp->ibqp.uobject || attrs->next_state !=
1621 C4IW_QP_STATE_ERROR)) {
1625 switch (attrs->next_state) {
1626 case C4IW_QP_STATE_IDLE:
1628 set_state(qhp, C4IW_QP_STATE_IDLE);
1629 qhp->attr.llp_stream_handle = NULL;
1630 c4iw_put_ep(&qhp->ep->com);
1632 wake_up(&qhp->wait);
1634 case C4IW_QP_STATE_ERROR:
1641 case C4IW_QP_STATE_ERROR:
1642 if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1646 if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1650 set_state(qhp, C4IW_QP_STATE_IDLE);
1652 case C4IW_QP_STATE_TERMINATE:
1660 pr_err("%s in a bad state %d\n", __func__, qhp->attr.state);
1667 pr_debug("disassociating ep %p qpid 0x%x\n", qhp->ep,
1670 /* disassociate the LLP connection */
1671 qhp->attr.llp_stream_handle = NULL;
1675 set_state(qhp, C4IW_QP_STATE_ERROR);
1679 wake_up(&qhp->wait);
1681 mutex_unlock(&qhp->mutex);
1684 post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1687 * If disconnect is 1, then we need to initiate a disconnect
1688 * on the EP. This can be a normal close (RTS->CLOSING) or
1689 * an abnormal close (RTS/CLOSING->ERROR).
1692 c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1694 c4iw_put_ep(&ep->com);
1698 * If free is 1, then we've disassociated the EP from the QP
1699 * and we need to dereference the EP.
1702 c4iw_put_ep(&ep->com);
1703 pr_debug("exit state %d\n", qhp->attr.state);
1707 int c4iw_destroy_qp(struct ib_qp *ib_qp)
1709 struct c4iw_dev *rhp;
1710 struct c4iw_qp *qhp;
1711 struct c4iw_qp_attributes attrs;
1713 qhp = to_c4iw_qp(ib_qp);
1716 attrs.next_state = C4IW_QP_STATE_ERROR;
1717 if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1718 c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1720 c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1721 wait_event(qhp->wait, !qhp->ep);
1723 remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1725 spin_lock_irq(&rhp->lock);
1726 if (!list_empty(&qhp->db_fc_entry))
1727 list_del_init(&qhp->db_fc_entry);
1728 spin_unlock_irq(&rhp->lock);
1729 free_ird(rhp, qhp->attr.max_ird);
1731 c4iw_qp_rem_ref(ib_qp);
1733 pr_debug("ib_qp %p qpid 0x%0x\n", ib_qp, qhp->wq.sq.qid);
1737 struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1738 struct ib_udata *udata)
1740 struct c4iw_dev *rhp;
1741 struct c4iw_qp *qhp;
1742 struct c4iw_pd *php;
1743 struct c4iw_cq *schp;
1744 struct c4iw_cq *rchp;
1745 struct c4iw_create_qp_resp uresp;
1746 unsigned int sqsize, rqsize;
1747 struct c4iw_ucontext *ucontext;
1749 struct c4iw_mm_entry *sq_key_mm, *rq_key_mm = NULL, *sq_db_key_mm;
1750 struct c4iw_mm_entry *rq_db_key_mm = NULL, *ma_sync_key_mm = NULL;
1752 pr_debug("ib_pd %p\n", pd);
1754 if (attrs->qp_type != IB_QPT_RC)
1755 return ERR_PTR(-EINVAL);
1757 php = to_c4iw_pd(pd);
1759 schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1760 rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1762 return ERR_PTR(-EINVAL);
1764 if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1765 return ERR_PTR(-EINVAL);
1767 if (attrs->cap.max_recv_wr > rhp->rdev.hw_queue.t4_max_rq_size)
1768 return ERR_PTR(-E2BIG);
1769 rqsize = attrs->cap.max_recv_wr + 1;
1773 if (attrs->cap.max_send_wr > rhp->rdev.hw_queue.t4_max_sq_size)
1774 return ERR_PTR(-E2BIG);
1775 sqsize = attrs->cap.max_send_wr + 1;
1779 ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1781 qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1783 return ERR_PTR(-ENOMEM);
1785 qhp->wr_waitp = c4iw_alloc_wr_wait(GFP_KERNEL);
1786 if (!qhp->wr_waitp) {
1791 qhp->wq.sq.size = sqsize;
1792 qhp->wq.sq.memsize =
1793 (sqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1794 sizeof(*qhp->wq.sq.queue) + 16 * sizeof(__be64);
1795 qhp->wq.sq.flush_cidx = -1;
1796 qhp->wq.rq.size = rqsize;
1797 qhp->wq.rq.memsize =
1798 (rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1799 sizeof(*qhp->wq.rq.queue);
1802 qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1803 qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1806 ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1807 ucontext ? &ucontext->uctx : &rhp->rdev.uctx,
1810 goto err_free_wr_wait;
1812 attrs->cap.max_recv_wr = rqsize - 1;
1813 attrs->cap.max_send_wr = sqsize - 1;
1814 attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1817 qhp->attr.pd = php->pdid;
1818 qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1819 qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1820 qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1821 qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1822 qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1823 qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1824 qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1825 qhp->attr.state = C4IW_QP_STATE_IDLE;
1826 qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1827 qhp->attr.enable_rdma_read = 1;
1828 qhp->attr.enable_rdma_write = 1;
1829 qhp->attr.enable_bind = 1;
1830 qhp->attr.max_ord = 0;
1831 qhp->attr.max_ird = 0;
1832 qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
1833 spin_lock_init(&qhp->lock);
1834 mutex_init(&qhp->mutex);
1835 init_waitqueue_head(&qhp->wait);
1836 kref_init(&qhp->kref);
1837 INIT_WORK(&qhp->free_work, free_qp_work);
1839 ret = insert_handle(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1841 goto err_destroy_qp;
1843 if (udata && ucontext) {
1844 sq_key_mm = kmalloc(sizeof(*sq_key_mm), GFP_KERNEL);
1847 goto err_remove_handle;
1849 rq_key_mm = kmalloc(sizeof(*rq_key_mm), GFP_KERNEL);
1852 goto err_free_sq_key;
1854 sq_db_key_mm = kmalloc(sizeof(*sq_db_key_mm), GFP_KERNEL);
1855 if (!sq_db_key_mm) {
1857 goto err_free_rq_key;
1859 rq_db_key_mm = kmalloc(sizeof(*rq_db_key_mm), GFP_KERNEL);
1860 if (!rq_db_key_mm) {
1862 goto err_free_sq_db_key;
1864 if (t4_sq_onchip(&qhp->wq.sq)) {
1865 ma_sync_key_mm = kmalloc(sizeof(*ma_sync_key_mm),
1867 if (!ma_sync_key_mm) {
1869 goto err_free_rq_db_key;
1871 uresp.flags = C4IW_QPF_ONCHIP;
1874 uresp.qid_mask = rhp->rdev.qpmask;
1875 uresp.sqid = qhp->wq.sq.qid;
1876 uresp.sq_size = qhp->wq.sq.size;
1877 uresp.sq_memsize = qhp->wq.sq.memsize;
1878 uresp.rqid = qhp->wq.rq.qid;
1879 uresp.rq_size = qhp->wq.rq.size;
1880 uresp.rq_memsize = qhp->wq.rq.memsize;
1881 spin_lock(&ucontext->mmap_lock);
1882 if (ma_sync_key_mm) {
1883 uresp.ma_sync_key = ucontext->key;
1884 ucontext->key += PAGE_SIZE;
1886 uresp.ma_sync_key = 0;
1888 uresp.sq_key = ucontext->key;
1889 ucontext->key += PAGE_SIZE;
1890 uresp.rq_key = ucontext->key;
1891 ucontext->key += PAGE_SIZE;
1892 uresp.sq_db_gts_key = ucontext->key;
1893 ucontext->key += PAGE_SIZE;
1894 uresp.rq_db_gts_key = ucontext->key;
1895 ucontext->key += PAGE_SIZE;
1896 spin_unlock(&ucontext->mmap_lock);
1897 ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1899 goto err_free_ma_sync_key;
1900 sq_key_mm->key = uresp.sq_key;
1901 sq_key_mm->addr = qhp->wq.sq.phys_addr;
1902 sq_key_mm->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1903 insert_mmap(ucontext, sq_key_mm);
1904 rq_key_mm->key = uresp.rq_key;
1905 rq_key_mm->addr = virt_to_phys(qhp->wq.rq.queue);
1906 rq_key_mm->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1907 insert_mmap(ucontext, rq_key_mm);
1908 sq_db_key_mm->key = uresp.sq_db_gts_key;
1909 sq_db_key_mm->addr = (u64)(unsigned long)qhp->wq.sq.bar2_pa;
1910 sq_db_key_mm->len = PAGE_SIZE;
1911 insert_mmap(ucontext, sq_db_key_mm);
1912 rq_db_key_mm->key = uresp.rq_db_gts_key;
1913 rq_db_key_mm->addr = (u64)(unsigned long)qhp->wq.rq.bar2_pa;
1914 rq_db_key_mm->len = PAGE_SIZE;
1915 insert_mmap(ucontext, rq_db_key_mm);
1916 if (ma_sync_key_mm) {
1917 ma_sync_key_mm->key = uresp.ma_sync_key;
1918 ma_sync_key_mm->addr =
1919 (pci_resource_start(rhp->rdev.lldi.pdev, 0) +
1920 PCIE_MA_SYNC_A) & PAGE_MASK;
1921 ma_sync_key_mm->len = PAGE_SIZE;
1922 insert_mmap(ucontext, ma_sync_key_mm);
1925 c4iw_get_ucontext(ucontext);
1926 qhp->ucontext = ucontext;
1928 qhp->ibqp.qp_num = qhp->wq.sq.qid;
1929 INIT_LIST_HEAD(&qhp->db_fc_entry);
1930 pr_debug("sq id %u size %u memsize %zu num_entries %u rq id %u size %u memsize %zu num_entries %u\n",
1931 qhp->wq.sq.qid, qhp->wq.sq.size, qhp->wq.sq.memsize,
1932 attrs->cap.max_send_wr, qhp->wq.rq.qid, qhp->wq.rq.size,
1933 qhp->wq.rq.memsize, attrs->cap.max_recv_wr);
1935 err_free_ma_sync_key:
1936 kfree(ma_sync_key_mm);
1938 kfree(rq_db_key_mm);
1940 kfree(sq_db_key_mm);
1946 remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1948 destroy_qp(&rhp->rdev, &qhp->wq,
1949 ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1951 c4iw_put_wr_wait(qhp->wr_waitp);
1954 return ERR_PTR(ret);
1957 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1958 int attr_mask, struct ib_udata *udata)
1960 struct c4iw_dev *rhp;
1961 struct c4iw_qp *qhp;
1962 enum c4iw_qp_attr_mask mask = 0;
1963 struct c4iw_qp_attributes attrs;
1965 pr_debug("ib_qp %p\n", ibqp);
1967 /* iwarp does not support the RTR state */
1968 if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1969 attr_mask &= ~IB_QP_STATE;
1971 /* Make sure we still have something left to do */
1975 memset(&attrs, 0, sizeof attrs);
1976 qhp = to_c4iw_qp(ibqp);
1979 attrs.next_state = c4iw_convert_state(attr->qp_state);
1980 attrs.enable_rdma_read = (attr->qp_access_flags &
1981 IB_ACCESS_REMOTE_READ) ? 1 : 0;
1982 attrs.enable_rdma_write = (attr->qp_access_flags &
1983 IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1984 attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1987 mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1988 mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1989 (C4IW_QP_ATTR_ENABLE_RDMA_READ |
1990 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1991 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1994 * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
1995 * ringing the queue db when we're in DB_FULL mode.
1996 * Only allow this on T4 devices.
1998 attrs.sq_db_inc = attr->sq_psn;
1999 attrs.rq_db_inc = attr->rq_psn;
2000 mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
2001 mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
2002 if (!is_t4(to_c4iw_qp(ibqp)->rhp->rdev.lldi.adapter_type) &&
2003 (mask & (C4IW_QP_ATTR_SQ_DB|C4IW_QP_ATTR_RQ_DB)))
2006 return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
2009 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
2011 pr_debug("ib_dev %p qpn 0x%x\n", dev, qpn);
2012 return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
2015 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2016 int attr_mask, struct ib_qp_init_attr *init_attr)
2018 struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
2020 memset(attr, 0, sizeof *attr);
2021 memset(init_attr, 0, sizeof *init_attr);
2022 attr->qp_state = to_ib_qp_state(qhp->attr.state);
2023 init_attr->cap.max_send_wr = qhp->attr.sq_num_entries;
2024 init_attr->cap.max_recv_wr = qhp->attr.rq_num_entries;
2025 init_attr->cap.max_send_sge = qhp->attr.sq_max_sges;
2026 init_attr->cap.max_recv_sge = qhp->attr.sq_max_sges;
2027 init_attr->cap.max_inline_data = T4_MAX_SEND_INLINE;
2028 init_attr->sq_sig_type = qhp->sq_sig_all ? IB_SIGNAL_ALL_WR : 0;