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
35 static int destroy_cq(struct c4iw_rdev *rdev, struct t4_cq *cq,
36 struct c4iw_dev_ucontext *uctx, struct sk_buff *skb,
37 struct c4iw_wr_wait *wr_waitp)
39 struct fw_ri_res_wr *res_wr;
40 struct fw_ri_res *res;
44 wr_len = sizeof *res_wr + sizeof *res;
45 set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
47 res_wr = __skb_put_zero(skb, wr_len);
48 res_wr->op_nres = cpu_to_be32(
49 FW_WR_OP_V(FW_RI_RES_WR) |
50 FW_RI_RES_WR_NRES_V(1) |
52 res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
53 res_wr->cookie = (uintptr_t)wr_waitp;
55 res->u.cq.restype = FW_RI_RES_TYPE_CQ;
56 res->u.cq.op = FW_RI_RES_OP_RESET;
57 res->u.cq.iqid = cpu_to_be32(cq->cqid);
59 c4iw_init_wr_wait(wr_waitp);
60 ret = c4iw_ref_send_wait(rdev, skb, wr_waitp, 0, 0, __func__);
63 dma_free_coherent(&(rdev->lldi.pdev->dev),
64 cq->memsize, cq->queue,
65 dma_unmap_addr(cq, mapping));
66 c4iw_put_cqid(rdev, cq->cqid, uctx);
70 static int create_cq(struct c4iw_rdev *rdev, struct t4_cq *cq,
71 struct c4iw_dev_ucontext *uctx,
72 struct c4iw_wr_wait *wr_waitp)
74 struct fw_ri_res_wr *res_wr;
75 struct fw_ri_res *res;
77 int user = (uctx != &rdev->uctx);
81 cq->cqid = c4iw_get_cqid(rdev, uctx);
88 cq->sw_queue = kzalloc(cq->memsize, GFP_KERNEL);
94 cq->queue = dma_alloc_coherent(&rdev->lldi.pdev->dev, cq->memsize,
95 &cq->dma_addr, GFP_KERNEL);
100 dma_unmap_addr_set(cq, mapping, cq->dma_addr);
101 memset(cq->queue, 0, cq->memsize);
103 /* build fw_ri_res_wr */
104 wr_len = sizeof *res_wr + sizeof *res;
106 skb = alloc_skb(wr_len, GFP_KERNEL);
111 set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
113 res_wr = __skb_put_zero(skb, wr_len);
114 res_wr->op_nres = cpu_to_be32(
115 FW_WR_OP_V(FW_RI_RES_WR) |
116 FW_RI_RES_WR_NRES_V(1) |
118 res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
119 res_wr->cookie = (uintptr_t)wr_waitp;
121 res->u.cq.restype = FW_RI_RES_TYPE_CQ;
122 res->u.cq.op = FW_RI_RES_OP_WRITE;
123 res->u.cq.iqid = cpu_to_be32(cq->cqid);
124 res->u.cq.iqandst_to_iqandstindex = cpu_to_be32(
125 FW_RI_RES_WR_IQANUS_V(0) |
126 FW_RI_RES_WR_IQANUD_V(1) |
127 FW_RI_RES_WR_IQANDST_F |
128 FW_RI_RES_WR_IQANDSTINDEX_V(
129 rdev->lldi.ciq_ids[cq->vector]));
130 res->u.cq.iqdroprss_to_iqesize = cpu_to_be16(
131 FW_RI_RES_WR_IQDROPRSS_F |
132 FW_RI_RES_WR_IQPCIECH_V(2) |
133 FW_RI_RES_WR_IQINTCNTTHRESH_V(0) |
135 FW_RI_RES_WR_IQESIZE_V(1));
136 res->u.cq.iqsize = cpu_to_be16(cq->size);
137 res->u.cq.iqaddr = cpu_to_be64(cq->dma_addr);
139 c4iw_init_wr_wait(wr_waitp);
140 ret = c4iw_ref_send_wait(rdev, skb, wr_waitp, 0, 0, __func__);
145 cq->gts = rdev->lldi.gts_reg;
148 cq->bar2_va = c4iw_bar2_addrs(rdev, cq->cqid, T4_BAR2_QTYPE_INGRESS,
150 user ? &cq->bar2_pa : NULL);
151 if (user && !cq->bar2_pa) {
152 pr_warn("%s: cqid %u not in BAR2 range\n",
153 pci_name(rdev->lldi.pdev), cq->cqid);
159 dma_free_coherent(&rdev->lldi.pdev->dev, cq->memsize, cq->queue,
160 dma_unmap_addr(cq, mapping));
164 c4iw_put_cqid(rdev, cq->cqid, uctx);
169 static void insert_recv_cqe(struct t4_wq *wq, struct t4_cq *cq)
173 pr_debug("wq %p cq %p sw_cidx %u sw_pidx %u\n",
174 wq, cq, cq->sw_cidx, cq->sw_pidx);
175 memset(&cqe, 0, sizeof(cqe));
176 cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
177 CQE_OPCODE_V(FW_RI_SEND) |
180 CQE_QPID_V(wq->sq.qid));
181 cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
182 cq->sw_queue[cq->sw_pidx] = cqe;
186 int c4iw_flush_rq(struct t4_wq *wq, struct t4_cq *cq, int count)
189 int in_use = wq->rq.in_use - count;
192 pr_debug("wq %p cq %p rq.in_use %u skip count %u\n",
193 wq, cq, wq->rq.in_use, count);
195 insert_recv_cqe(wq, cq);
201 static void insert_sq_cqe(struct t4_wq *wq, struct t4_cq *cq,
202 struct t4_swsqe *swcqe)
206 pr_debug("wq %p cq %p sw_cidx %u sw_pidx %u\n",
207 wq, cq, cq->sw_cidx, cq->sw_pidx);
208 memset(&cqe, 0, sizeof(cqe));
209 cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
210 CQE_OPCODE_V(swcqe->opcode) |
213 CQE_QPID_V(wq->sq.qid));
214 CQE_WRID_SQ_IDX(&cqe) = swcqe->idx;
215 cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
216 cq->sw_queue[cq->sw_pidx] = cqe;
220 static void advance_oldest_read(struct t4_wq *wq);
222 int c4iw_flush_sq(struct c4iw_qp *qhp)
225 struct t4_wq *wq = &qhp->wq;
226 struct c4iw_cq *chp = to_c4iw_cq(qhp->ibqp.send_cq);
227 struct t4_cq *cq = &chp->cq;
229 struct t4_swsqe *swsqe;
231 if (wq->sq.flush_cidx == -1)
232 wq->sq.flush_cidx = wq->sq.cidx;
233 idx = wq->sq.flush_cidx;
234 BUG_ON(idx >= wq->sq.size);
235 while (idx != wq->sq.pidx) {
236 swsqe = &wq->sq.sw_sq[idx];
237 BUG_ON(swsqe->flushed);
239 insert_sq_cqe(wq, cq, swsqe);
240 if (wq->sq.oldest_read == swsqe) {
241 BUG_ON(swsqe->opcode != FW_RI_READ_REQ);
242 advance_oldest_read(wq);
245 if (++idx == wq->sq.size)
248 wq->sq.flush_cidx += flushed;
249 if (wq->sq.flush_cidx >= wq->sq.size)
250 wq->sq.flush_cidx -= wq->sq.size;
254 static void flush_completed_wrs(struct t4_wq *wq, struct t4_cq *cq)
256 struct t4_swsqe *swsqe;
259 if (wq->sq.flush_cidx == -1)
260 wq->sq.flush_cidx = wq->sq.cidx;
261 cidx = wq->sq.flush_cidx;
262 BUG_ON(cidx > wq->sq.size);
264 while (cidx != wq->sq.pidx) {
265 swsqe = &wq->sq.sw_sq[cidx];
266 if (!swsqe->signaled) {
267 if (++cidx == wq->sq.size)
269 } else if (swsqe->complete) {
271 BUG_ON(swsqe->flushed);
274 * Insert this completed cqe into the swcq.
276 pr_debug("moving cqe into swcq sq idx %u cq idx %u\n",
278 swsqe->cqe.header |= htonl(CQE_SWCQE_V(1));
279 cq->sw_queue[cq->sw_pidx] = swsqe->cqe;
282 if (++cidx == wq->sq.size)
284 wq->sq.flush_cidx = cidx;
290 static void create_read_req_cqe(struct t4_wq *wq, struct t4_cqe *hw_cqe,
291 struct t4_cqe *read_cqe)
293 read_cqe->u.scqe.cidx = wq->sq.oldest_read->idx;
294 read_cqe->len = htonl(wq->sq.oldest_read->read_len);
295 read_cqe->header = htonl(CQE_QPID_V(CQE_QPID(hw_cqe)) |
296 CQE_SWCQE_V(SW_CQE(hw_cqe)) |
297 CQE_OPCODE_V(FW_RI_READ_REQ) |
299 read_cqe->bits_type_ts = hw_cqe->bits_type_ts;
302 static void advance_oldest_read(struct t4_wq *wq)
305 u32 rptr = wq->sq.oldest_read - wq->sq.sw_sq + 1;
307 if (rptr == wq->sq.size)
309 while (rptr != wq->sq.pidx) {
310 wq->sq.oldest_read = &wq->sq.sw_sq[rptr];
312 if (wq->sq.oldest_read->opcode == FW_RI_READ_REQ)
314 if (++rptr == wq->sq.size)
317 wq->sq.oldest_read = NULL;
321 * Move all CQEs from the HWCQ into the SWCQ.
322 * Deal with out-of-order and/or completions that complete
323 * prior unsignalled WRs.
325 void c4iw_flush_hw_cq(struct c4iw_cq *chp)
327 struct t4_cqe *hw_cqe, *swcqe, read_cqe;
329 struct t4_swsqe *swsqe;
332 pr_debug("cqid 0x%x\n", chp->cq.cqid);
333 ret = t4_next_hw_cqe(&chp->cq, &hw_cqe);
336 * This logic is similar to poll_cq(), but not quite the same
337 * unfortunately. Need to move pertinent HW CQEs to the SW CQ but
338 * also do any translation magic that poll_cq() normally does.
341 qhp = get_qhp(chp->rhp, CQE_QPID(hw_cqe));
344 * drop CQEs with no associated QP
349 if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE)
352 if (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP) {
354 /* If we have reached here because of async
355 * event or other error, and have egress error
358 if (CQE_TYPE(hw_cqe) == 1)
361 /* drop peer2peer RTR reads.
363 if (CQE_WRID_STAG(hw_cqe) == 1)
367 * Eat completions for unsignaled read WRs.
369 if (!qhp->wq.sq.oldest_read->signaled) {
370 advance_oldest_read(&qhp->wq);
375 * Don't write to the HWCQ, create a new read req CQE
376 * in local memory and move it into the swcq.
378 create_read_req_cqe(&qhp->wq, hw_cqe, &read_cqe);
380 advance_oldest_read(&qhp->wq);
383 /* if its a SQ completion, then do the magic to move all the
384 * unsignaled and now in-order completions into the swcq.
386 if (SQ_TYPE(hw_cqe)) {
387 swsqe = &qhp->wq.sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
388 swsqe->cqe = *hw_cqe;
390 flush_completed_wrs(&qhp->wq, &chp->cq);
392 swcqe = &chp->cq.sw_queue[chp->cq.sw_pidx];
394 swcqe->header |= cpu_to_be32(CQE_SWCQE_V(1));
395 t4_swcq_produce(&chp->cq);
398 t4_hwcq_consume(&chp->cq);
399 ret = t4_next_hw_cqe(&chp->cq, &hw_cqe);
403 static int cqe_completes_wr(struct t4_cqe *cqe, struct t4_wq *wq)
405 if (CQE_OPCODE(cqe) == FW_RI_TERMINATE)
408 if ((CQE_OPCODE(cqe) == FW_RI_RDMA_WRITE) && RQ_TYPE(cqe))
411 if ((CQE_OPCODE(cqe) == FW_RI_READ_RESP) && SQ_TYPE(cqe))
414 if (CQE_SEND_OPCODE(cqe) && RQ_TYPE(cqe) && t4_rq_empty(wq))
419 void c4iw_count_rcqes(struct t4_cq *cq, struct t4_wq *wq, int *count)
425 pr_debug("count zero %d\n", *count);
427 while (ptr != cq->sw_pidx) {
428 cqe = &cq->sw_queue[ptr];
429 if (RQ_TYPE(cqe) && (CQE_OPCODE(cqe) != FW_RI_READ_RESP) &&
430 (CQE_QPID(cqe) == wq->sq.qid) && cqe_completes_wr(cqe, wq))
432 if (++ptr == cq->size)
435 pr_debug("cq %p count %d\n", cq, *count);
442 * check the validity of the first CQE,
443 * supply the wq assicated with the qpid.
445 * credit: cq credit to return to sge.
446 * cqe_flushed: 1 iff the CQE is flushed.
447 * cqe: copy of the polled CQE.
451 * -EAGAIN CQE skipped, try again.
452 * -EOVERFLOW CQ overflow detected.
454 static int poll_cq(struct t4_wq *wq, struct t4_cq *cq, struct t4_cqe *cqe,
455 u8 *cqe_flushed, u64 *cookie, u32 *credit)
458 struct t4_cqe *hw_cqe, read_cqe;
462 ret = t4_next_cqe(cq, &hw_cqe);
466 pr_debug("CQE OVF %u qpid 0x%0x genbit %u type %u status 0x%0x opcode 0x%0x len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x\n",
467 CQE_OVFBIT(hw_cqe), CQE_QPID(hw_cqe),
468 CQE_GENBIT(hw_cqe), CQE_TYPE(hw_cqe), CQE_STATUS(hw_cqe),
469 CQE_OPCODE(hw_cqe), CQE_LEN(hw_cqe), CQE_WRID_HI(hw_cqe),
470 CQE_WRID_LOW(hw_cqe));
473 * skip cqe's not affiliated with a QP.
481 * skip hw cqe's if the wq is flushed.
483 if (wq->flushed && !SW_CQE(hw_cqe)) {
489 * skip TERMINATE cqes...
491 if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE) {
497 * Special cqe for drain WR completions...
499 if (CQE_OPCODE(hw_cqe) == C4IW_DRAIN_OPCODE) {
500 *cookie = CQE_DRAIN_COOKIE(hw_cqe);
506 * Gotta tweak READ completions:
507 * 1) the cqe doesn't contain the sq_wptr from the wr.
508 * 2) opcode not reflected from the wr.
509 * 3) read_len not reflected from the wr.
510 * 4) cq_type is RQ_TYPE not SQ_TYPE.
512 if (RQ_TYPE(hw_cqe) && (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP)) {
514 /* If we have reached here because of async
515 * event or other error, and have egress error
518 if (CQE_TYPE(hw_cqe) == 1) {
519 if (CQE_STATUS(hw_cqe))
520 t4_set_wq_in_error(wq);
525 /* If this is an unsolicited read response, then the read
526 * was generated by the kernel driver as part of peer-2-peer
527 * connection setup. So ignore the completion.
529 if (CQE_WRID_STAG(hw_cqe) == 1) {
530 if (CQE_STATUS(hw_cqe))
531 t4_set_wq_in_error(wq);
537 * Eat completions for unsignaled read WRs.
539 if (!wq->sq.oldest_read->signaled) {
540 advance_oldest_read(wq);
546 * Don't write to the HWCQ, so create a new read req CQE
549 create_read_req_cqe(wq, hw_cqe, &read_cqe);
551 advance_oldest_read(wq);
554 if (CQE_STATUS(hw_cqe) || t4_wq_in_error(wq)) {
555 *cqe_flushed = (CQE_STATUS(hw_cqe) == T4_ERR_SWFLUSH);
556 t4_set_wq_in_error(wq);
562 if (RQ_TYPE(hw_cqe)) {
565 * HW only validates 4 bits of MSN. So we must validate that
566 * the MSN in the SEND is the next expected MSN. If its not,
567 * then we complete this with T4_ERR_MSN and mark the wq in
571 if (t4_rq_empty(wq)) {
572 t4_set_wq_in_error(wq);
576 if (unlikely((CQE_WRID_MSN(hw_cqe) != (wq->rq.msn)))) {
577 t4_set_wq_in_error(wq);
578 hw_cqe->header |= htonl(CQE_STATUS_V(T4_ERR_MSN));
585 * If we get here its a send completion.
587 * Handle out of order completion. These get stuffed
588 * in the SW SQ. Then the SW SQ is walked to move any
589 * now in-order completions into the SW CQ. This handles
591 * 1) reaping unsignaled WRs when the first subsequent
592 * signaled WR is completed.
593 * 2) out of order read completions.
595 if (!SW_CQE(hw_cqe) && (CQE_WRID_SQ_IDX(hw_cqe) != wq->sq.cidx)) {
596 struct t4_swsqe *swsqe;
598 pr_debug("out of order completion going in sw_sq at idx %u\n",
599 CQE_WRID_SQ_IDX(hw_cqe));
600 swsqe = &wq->sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
601 swsqe->cqe = *hw_cqe;
611 * Reap the associated WR(s) that are freed up with this
614 if (SQ_TYPE(hw_cqe)) {
615 int idx = CQE_WRID_SQ_IDX(hw_cqe);
616 BUG_ON(idx >= wq->sq.size);
619 * Account for any unsignaled completions completed by
620 * this signaled completion. In this case, cidx points
621 * to the first unsignaled one, and idx points to the
622 * signaled one. So adjust in_use based on this delta.
623 * if this is not completing any unsigned wrs, then the
624 * delta will be 0. Handle wrapping also!
626 if (idx < wq->sq.cidx)
627 wq->sq.in_use -= wq->sq.size + idx - wq->sq.cidx;
629 wq->sq.in_use -= idx - wq->sq.cidx;
630 BUG_ON(wq->sq.in_use <= 0 && wq->sq.in_use >= wq->sq.size);
632 wq->sq.cidx = (uint16_t)idx;
633 pr_debug("completing sq idx %u\n", wq->sq.cidx);
634 *cookie = wq->sq.sw_sq[wq->sq.cidx].wr_id;
636 c4iw_log_wr_stats(wq, hw_cqe);
639 pr_debug("completing rq idx %u\n", wq->rq.cidx);
640 *cookie = wq->rq.sw_rq[wq->rq.cidx].wr_id;
641 BUG_ON(t4_rq_empty(wq));
643 c4iw_log_wr_stats(wq, hw_cqe);
650 * Flush any completed cqes that are now in-order.
652 flush_completed_wrs(wq, cq);
655 if (SW_CQE(hw_cqe)) {
656 pr_debug("cq %p cqid 0x%x skip sw cqe cidx %u\n",
657 cq, cq->cqid, cq->sw_cidx);
660 pr_debug("cq %p cqid 0x%x skip hw cqe cidx %u\n",
661 cq, cq->cqid, cq->cidx);
668 * Get one cq entry from c4iw and map it to openib.
673 * -EAGAIN caller must try again
674 * any other -errno fatal error
676 static int c4iw_poll_cq_one(struct c4iw_cq *chp, struct ib_wc *wc)
678 struct c4iw_qp *qhp = NULL;
679 struct t4_cqe uninitialized_var(cqe), *rd_cqe;
686 ret = t4_next_cqe(&chp->cq, &rd_cqe);
691 qhp = get_qhp(chp->rhp, CQE_QPID(rd_cqe));
695 spin_lock(&qhp->lock);
698 ret = poll_cq(wq, &(chp->cq), &cqe, &cqe_flushed, &cookie, &credit);
704 wc->vendor_err = CQE_STATUS(&cqe);
707 pr_debug("qpid 0x%x type %d opcode %d status 0x%x len %u wrid hi 0x%x lo 0x%x cookie 0x%llx\n",
709 CQE_TYPE(&cqe), CQE_OPCODE(&cqe),
710 CQE_STATUS(&cqe), CQE_LEN(&cqe),
711 CQE_WRID_HI(&cqe), CQE_WRID_LOW(&cqe),
712 (unsigned long long)cookie);
714 if (CQE_TYPE(&cqe) == 0) {
715 if (!CQE_STATUS(&cqe))
716 wc->byte_len = CQE_LEN(&cqe);
719 wc->opcode = IB_WC_RECV;
720 if (CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_INV ||
721 CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_SE_INV) {
722 wc->ex.invalidate_rkey = CQE_WRID_STAG(&cqe);
723 wc->wc_flags |= IB_WC_WITH_INVALIDATE;
724 c4iw_invalidate_mr(qhp->rhp, wc->ex.invalidate_rkey);
727 switch (CQE_OPCODE(&cqe)) {
728 case FW_RI_RDMA_WRITE:
729 wc->opcode = IB_WC_RDMA_WRITE;
732 wc->opcode = IB_WC_RDMA_READ;
733 wc->byte_len = CQE_LEN(&cqe);
735 case FW_RI_SEND_WITH_INV:
736 case FW_RI_SEND_WITH_SE_INV:
737 wc->opcode = IB_WC_SEND;
738 wc->wc_flags |= IB_WC_WITH_INVALIDATE;
741 case FW_RI_SEND_WITH_SE:
742 wc->opcode = IB_WC_SEND;
745 case FW_RI_LOCAL_INV:
746 wc->opcode = IB_WC_LOCAL_INV;
748 case FW_RI_FAST_REGISTER:
749 wc->opcode = IB_WC_REG_MR;
751 /* Invalidate the MR if the fastreg failed */
752 if (CQE_STATUS(&cqe) != T4_ERR_SUCCESS)
753 c4iw_invalidate_mr(qhp->rhp,
754 CQE_WRID_FR_STAG(&cqe));
756 case C4IW_DRAIN_OPCODE:
757 wc->opcode = IB_WC_SEND;
760 pr_err("Unexpected opcode %d in the CQE received for QPID=0x%0x\n",
761 CQE_OPCODE(&cqe), CQE_QPID(&cqe));
768 wc->status = IB_WC_WR_FLUSH_ERR;
771 switch (CQE_STATUS(&cqe)) {
773 wc->status = IB_WC_SUCCESS;
776 wc->status = IB_WC_LOC_ACCESS_ERR;
779 wc->status = IB_WC_LOC_PROT_ERR;
783 wc->status = IB_WC_LOC_ACCESS_ERR;
786 wc->status = IB_WC_GENERAL_ERR;
789 wc->status = IB_WC_LOC_LEN_ERR;
791 case T4_ERR_INVALIDATE_SHARED_MR:
792 case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
793 wc->status = IB_WC_MW_BIND_ERR;
797 case T4_ERR_PDU_LEN_ERR:
798 case T4_ERR_OUT_OF_RQE:
799 case T4_ERR_DDP_VERSION:
800 case T4_ERR_RDMA_VERSION:
801 case T4_ERR_DDP_QUEUE_NUM:
805 case T4_ERR_MSN_RANGE:
806 case T4_ERR_IRD_OVERFLOW:
808 case T4_ERR_INTERNAL_ERR:
809 wc->status = IB_WC_FATAL_ERR;
812 wc->status = IB_WC_WR_FLUSH_ERR;
815 pr_err("Unexpected cqe_status 0x%x for QPID=0x%0x\n",
816 CQE_STATUS(&cqe), CQE_QPID(&cqe));
817 wc->status = IB_WC_FATAL_ERR;
822 spin_unlock(&qhp->lock);
826 int c4iw_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
833 chp = to_c4iw_cq(ibcq);
835 spin_lock_irqsave(&chp->lock, flags);
836 for (npolled = 0; npolled < num_entries; ++npolled) {
838 err = c4iw_poll_cq_one(chp, wc + npolled);
839 } while (err == -EAGAIN);
843 spin_unlock_irqrestore(&chp->lock, flags);
844 return !err || err == -ENODATA ? npolled : err;
847 int c4iw_destroy_cq(struct ib_cq *ib_cq)
850 struct c4iw_ucontext *ucontext;
852 pr_debug("ib_cq %p\n", ib_cq);
853 chp = to_c4iw_cq(ib_cq);
855 remove_handle(chp->rhp, &chp->rhp->cqidr, chp->cq.cqid);
856 atomic_dec(&chp->refcnt);
857 wait_event(chp->wait, !atomic_read(&chp->refcnt));
859 ucontext = ib_cq->uobject ? to_c4iw_ucontext(ib_cq->uobject->context)
861 destroy_cq(&chp->rhp->rdev, &chp->cq,
862 ucontext ? &ucontext->uctx : &chp->cq.rdev->uctx,
863 chp->destroy_skb, chp->wr_waitp);
864 c4iw_put_wr_wait(chp->wr_waitp);
869 struct ib_cq *c4iw_create_cq(struct ib_device *ibdev,
870 const struct ib_cq_init_attr *attr,
871 struct ib_ucontext *ib_context,
872 struct ib_udata *udata)
874 int entries = attr->cqe;
875 int vector = attr->comp_vector;
876 struct c4iw_dev *rhp;
878 struct c4iw_create_cq_resp uresp;
879 struct c4iw_ucontext *ucontext = NULL;
881 size_t memsize, hwentries;
882 struct c4iw_mm_entry *mm, *mm2;
884 pr_debug("ib_dev %p entries %d\n", ibdev, entries);
886 return ERR_PTR(-EINVAL);
888 rhp = to_c4iw_dev(ibdev);
890 if (vector >= rhp->rdev.lldi.nciq)
891 return ERR_PTR(-EINVAL);
893 chp = kzalloc(sizeof(*chp), GFP_KERNEL);
895 return ERR_PTR(-ENOMEM);
896 chp->wr_waitp = c4iw_alloc_wr_wait(GFP_KERNEL);
897 if (!chp->wr_waitp) {
901 c4iw_init_wr_wait(chp->wr_waitp);
903 wr_len = sizeof(struct fw_ri_res_wr) + sizeof(struct fw_ri_res);
904 chp->destroy_skb = alloc_skb(wr_len, GFP_KERNEL);
905 if (!chp->destroy_skb) {
907 goto err_free_wr_wait;
911 ucontext = to_c4iw_ucontext(ib_context);
913 /* account for the status page. */
916 /* IQ needs one extra entry to differentiate full vs empty. */
920 * entries must be multiple of 16 for HW.
922 entries = roundup(entries, 16);
925 * Make actual HW queue 2x to avoid cdix_inc overflows.
927 hwentries = min(entries * 2, rhp->rdev.hw_queue.t4_max_iq_size);
930 * Make HW queue at least 64 entries so GTS updates aren't too
936 memsize = hwentries * sizeof *chp->cq.queue;
939 * memsize must be a multiple of the page size if its a user cq.
942 memsize = roundup(memsize, PAGE_SIZE);
943 chp->cq.size = hwentries;
944 chp->cq.memsize = memsize;
945 chp->cq.vector = vector;
947 ret = create_cq(&rhp->rdev, &chp->cq,
948 ucontext ? &ucontext->uctx : &rhp->rdev.uctx,
954 chp->cq.size--; /* status page */
955 chp->ibcq.cqe = entries - 2;
956 spin_lock_init(&chp->lock);
957 spin_lock_init(&chp->comp_handler_lock);
958 atomic_set(&chp->refcnt, 1);
959 init_waitqueue_head(&chp->wait);
960 ret = insert_handle(rhp, &rhp->cqidr, chp, chp->cq.cqid);
966 mm = kmalloc(sizeof *mm, GFP_KERNEL);
968 goto err_remove_handle;
969 mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
973 uresp.qid_mask = rhp->rdev.cqmask;
974 uresp.cqid = chp->cq.cqid;
975 uresp.size = chp->cq.size;
976 uresp.memsize = chp->cq.memsize;
977 spin_lock(&ucontext->mmap_lock);
978 uresp.key = ucontext->key;
979 ucontext->key += PAGE_SIZE;
980 uresp.gts_key = ucontext->key;
981 ucontext->key += PAGE_SIZE;
982 spin_unlock(&ucontext->mmap_lock);
983 ret = ib_copy_to_udata(udata, &uresp,
984 sizeof(uresp) - sizeof(uresp.reserved));
989 mm->addr = virt_to_phys(chp->cq.queue);
990 mm->len = chp->cq.memsize;
991 insert_mmap(ucontext, mm);
993 mm2->key = uresp.gts_key;
994 mm2->addr = chp->cq.bar2_pa;
995 mm2->len = PAGE_SIZE;
996 insert_mmap(ucontext, mm2);
998 pr_debug("cqid 0x%0x chp %p size %u memsize %zu, dma_addr 0x%0llx\n",
999 chp->cq.cqid, chp, chp->cq.size,
1000 chp->cq.memsize, (unsigned long long)chp->cq.dma_addr);
1007 remove_handle(rhp, &rhp->cqidr, chp->cq.cqid);
1009 destroy_cq(&chp->rhp->rdev, &chp->cq,
1010 ucontext ? &ucontext->uctx : &rhp->rdev.uctx,
1011 chp->destroy_skb, chp->wr_waitp);
1013 kfree_skb(chp->destroy_skb);
1015 c4iw_put_wr_wait(chp->wr_waitp);
1018 return ERR_PTR(ret);
1021 int c4iw_resize_cq(struct ib_cq *cq, int cqe, struct ib_udata *udata)
1026 int c4iw_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
1028 struct c4iw_cq *chp;
1032 chp = to_c4iw_cq(ibcq);
1033 spin_lock_irqsave(&chp->lock, flag);
1035 (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED);
1036 if (flags & IB_CQ_REPORT_MISSED_EVENTS)
1037 ret = t4_cq_notempty(&chp->cq);
1038 spin_unlock_irqrestore(&chp->lock, flag);