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;
191 pr_debug("wq %p cq %p rq.in_use %u skip count %u\n",
192 wq, cq, wq->rq.in_use, count);
194 insert_recv_cqe(wq, cq);
200 static void insert_sq_cqe(struct t4_wq *wq, struct t4_cq *cq,
201 struct t4_swsqe *swcqe)
205 pr_debug("wq %p cq %p sw_cidx %u sw_pidx %u\n",
206 wq, cq, cq->sw_cidx, cq->sw_pidx);
207 memset(&cqe, 0, sizeof(cqe));
208 cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
209 CQE_OPCODE_V(swcqe->opcode) |
212 CQE_QPID_V(wq->sq.qid));
213 CQE_WRID_SQ_IDX(&cqe) = swcqe->idx;
214 cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
215 cq->sw_queue[cq->sw_pidx] = cqe;
219 static void advance_oldest_read(struct t4_wq *wq);
221 int c4iw_flush_sq(struct c4iw_qp *qhp)
224 struct t4_wq *wq = &qhp->wq;
225 struct c4iw_cq *chp = to_c4iw_cq(qhp->ibqp.send_cq);
226 struct t4_cq *cq = &chp->cq;
228 struct t4_swsqe *swsqe;
230 if (wq->sq.flush_cidx == -1)
231 wq->sq.flush_cidx = wq->sq.cidx;
232 idx = wq->sq.flush_cidx;
233 while (idx != wq->sq.pidx) {
234 swsqe = &wq->sq.sw_sq[idx];
236 insert_sq_cqe(wq, cq, swsqe);
237 if (wq->sq.oldest_read == swsqe) {
238 advance_oldest_read(wq);
241 if (++idx == wq->sq.size)
244 wq->sq.flush_cidx += flushed;
245 if (wq->sq.flush_cidx >= wq->sq.size)
246 wq->sq.flush_cidx -= wq->sq.size;
250 static void flush_completed_wrs(struct t4_wq *wq, struct t4_cq *cq)
252 struct t4_swsqe *swsqe;
255 if (wq->sq.flush_cidx == -1)
256 wq->sq.flush_cidx = wq->sq.cidx;
257 cidx = wq->sq.flush_cidx;
259 while (cidx != wq->sq.pidx) {
260 swsqe = &wq->sq.sw_sq[cidx];
261 if (!swsqe->signaled) {
262 if (++cidx == wq->sq.size)
264 } else if (swsqe->complete) {
267 * Insert this completed cqe into the swcq.
269 pr_debug("moving cqe into swcq sq idx %u cq idx %u\n",
271 swsqe->cqe.header |= htonl(CQE_SWCQE_V(1));
272 cq->sw_queue[cq->sw_pidx] = swsqe->cqe;
275 if (++cidx == wq->sq.size)
277 wq->sq.flush_cidx = cidx;
283 static void create_read_req_cqe(struct t4_wq *wq, struct t4_cqe *hw_cqe,
284 struct t4_cqe *read_cqe)
286 read_cqe->u.scqe.cidx = wq->sq.oldest_read->idx;
287 read_cqe->len = htonl(wq->sq.oldest_read->read_len);
288 read_cqe->header = htonl(CQE_QPID_V(CQE_QPID(hw_cqe)) |
289 CQE_SWCQE_V(SW_CQE(hw_cqe)) |
290 CQE_OPCODE_V(FW_RI_READ_REQ) |
292 read_cqe->bits_type_ts = hw_cqe->bits_type_ts;
295 static void advance_oldest_read(struct t4_wq *wq)
298 u32 rptr = wq->sq.oldest_read - wq->sq.sw_sq + 1;
300 if (rptr == wq->sq.size)
302 while (rptr != wq->sq.pidx) {
303 wq->sq.oldest_read = &wq->sq.sw_sq[rptr];
305 if (wq->sq.oldest_read->opcode == FW_RI_READ_REQ)
307 if (++rptr == wq->sq.size)
310 wq->sq.oldest_read = NULL;
314 * Move all CQEs from the HWCQ into the SWCQ.
315 * Deal with out-of-order and/or completions that complete
316 * prior unsignalled WRs.
318 void c4iw_flush_hw_cq(struct c4iw_cq *chp)
320 struct t4_cqe *hw_cqe, *swcqe, read_cqe;
322 struct t4_swsqe *swsqe;
325 pr_debug("cqid 0x%x\n", chp->cq.cqid);
326 ret = t4_next_hw_cqe(&chp->cq, &hw_cqe);
329 * This logic is similar to poll_cq(), but not quite the same
330 * unfortunately. Need to move pertinent HW CQEs to the SW CQ but
331 * also do any translation magic that poll_cq() normally does.
334 qhp = get_qhp(chp->rhp, CQE_QPID(hw_cqe));
337 * drop CQEs with no associated QP
342 if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE)
345 if (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP) {
347 /* If we have reached here because of async
348 * event or other error, and have egress error
351 if (CQE_TYPE(hw_cqe) == 1)
354 /* drop peer2peer RTR reads.
356 if (CQE_WRID_STAG(hw_cqe) == 1)
360 * Eat completions for unsignaled read WRs.
362 if (!qhp->wq.sq.oldest_read->signaled) {
363 advance_oldest_read(&qhp->wq);
368 * Don't write to the HWCQ, create a new read req CQE
369 * in local memory and move it into the swcq.
371 create_read_req_cqe(&qhp->wq, hw_cqe, &read_cqe);
373 advance_oldest_read(&qhp->wq);
376 /* if its a SQ completion, then do the magic to move all the
377 * unsignaled and now in-order completions into the swcq.
379 if (SQ_TYPE(hw_cqe)) {
380 swsqe = &qhp->wq.sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
381 swsqe->cqe = *hw_cqe;
383 flush_completed_wrs(&qhp->wq, &chp->cq);
385 swcqe = &chp->cq.sw_queue[chp->cq.sw_pidx];
387 swcqe->header |= cpu_to_be32(CQE_SWCQE_V(1));
388 t4_swcq_produce(&chp->cq);
391 t4_hwcq_consume(&chp->cq);
392 ret = t4_next_hw_cqe(&chp->cq, &hw_cqe);
396 static int cqe_completes_wr(struct t4_cqe *cqe, struct t4_wq *wq)
398 if (DRAIN_CQE(cqe)) {
399 WARN_ONCE(1, "Unexpected DRAIN CQE qp id %u!\n", wq->sq.qid);
403 if (CQE_OPCODE(cqe) == FW_RI_TERMINATE)
406 if ((CQE_OPCODE(cqe) == FW_RI_RDMA_WRITE) && RQ_TYPE(cqe))
409 if ((CQE_OPCODE(cqe) == FW_RI_READ_RESP) && SQ_TYPE(cqe))
412 if (CQE_SEND_OPCODE(cqe) && RQ_TYPE(cqe) && t4_rq_empty(wq))
417 void c4iw_count_rcqes(struct t4_cq *cq, struct t4_wq *wq, int *count)
423 pr_debug("count zero %d\n", *count);
425 while (ptr != cq->sw_pidx) {
426 cqe = &cq->sw_queue[ptr];
427 if (RQ_TYPE(cqe) && (CQE_OPCODE(cqe) != FW_RI_READ_RESP) &&
428 (CQE_QPID(cqe) == wq->sq.qid) && cqe_completes_wr(cqe, wq))
430 if (++ptr == cq->size)
433 pr_debug("cq %p count %d\n", cq, *count);
440 * check the validity of the first CQE,
441 * supply the wq assicated with the qpid.
443 * credit: cq credit to return to sge.
444 * cqe_flushed: 1 iff the CQE is flushed.
445 * cqe: copy of the polled CQE.
449 * -EAGAIN CQE skipped, try again.
450 * -EOVERFLOW CQ overflow detected.
452 static int poll_cq(struct t4_wq *wq, struct t4_cq *cq, struct t4_cqe *cqe,
453 u8 *cqe_flushed, u64 *cookie, u32 *credit)
456 struct t4_cqe *hw_cqe, read_cqe;
460 ret = t4_next_cqe(cq, &hw_cqe);
464 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",
465 CQE_OVFBIT(hw_cqe), CQE_QPID(hw_cqe),
466 CQE_GENBIT(hw_cqe), CQE_TYPE(hw_cqe), CQE_STATUS(hw_cqe),
467 CQE_OPCODE(hw_cqe), CQE_LEN(hw_cqe), CQE_WRID_HI(hw_cqe),
468 CQE_WRID_LOW(hw_cqe));
471 * skip cqe's not affiliated with a QP.
479 * skip hw cqe's if the wq is flushed.
481 if (wq->flushed && !SW_CQE(hw_cqe)) {
487 * skip TERMINATE cqes...
489 if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE) {
495 * Special cqe for drain WR completions...
497 if (DRAIN_CQE(hw_cqe)) {
498 *cookie = CQE_DRAIN_COOKIE(hw_cqe);
504 * Gotta tweak READ completions:
505 * 1) the cqe doesn't contain the sq_wptr from the wr.
506 * 2) opcode not reflected from the wr.
507 * 3) read_len not reflected from the wr.
508 * 4) cq_type is RQ_TYPE not SQ_TYPE.
510 if (RQ_TYPE(hw_cqe) && (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP)) {
512 /* If we have reached here because of async
513 * event or other error, and have egress error
516 if (CQE_TYPE(hw_cqe) == 1) {
517 if (CQE_STATUS(hw_cqe))
518 t4_set_wq_in_error(wq);
523 /* If this is an unsolicited read response, then the read
524 * was generated by the kernel driver as part of peer-2-peer
525 * connection setup. So ignore the completion.
527 if (CQE_WRID_STAG(hw_cqe) == 1) {
528 if (CQE_STATUS(hw_cqe))
529 t4_set_wq_in_error(wq);
535 * Eat completions for unsignaled read WRs.
537 if (!wq->sq.oldest_read->signaled) {
538 advance_oldest_read(wq);
544 * Don't write to the HWCQ, so create a new read req CQE
547 create_read_req_cqe(wq, hw_cqe, &read_cqe);
549 advance_oldest_read(wq);
552 if (CQE_STATUS(hw_cqe) || t4_wq_in_error(wq)) {
553 *cqe_flushed = (CQE_STATUS(hw_cqe) == T4_ERR_SWFLUSH);
554 t4_set_wq_in_error(wq);
560 if (RQ_TYPE(hw_cqe)) {
563 * HW only validates 4 bits of MSN. So we must validate that
564 * the MSN in the SEND is the next expected MSN. If its not,
565 * then we complete this with T4_ERR_MSN and mark the wq in
569 if (t4_rq_empty(wq)) {
570 t4_set_wq_in_error(wq);
574 if (unlikely(!CQE_STATUS(hw_cqe) &&
575 CQE_WRID_MSN(hw_cqe) != wq->rq.msn)) {
576 t4_set_wq_in_error(wq);
577 hw_cqe->header |= cpu_to_be32(CQE_STATUS_V(T4_ERR_MSN));
583 * If we get here its a send completion.
585 * Handle out of order completion. These get stuffed
586 * in the SW SQ. Then the SW SQ is walked to move any
587 * now in-order completions into the SW CQ. This handles
589 * 1) reaping unsignaled WRs when the first subsequent
590 * signaled WR is completed.
591 * 2) out of order read completions.
593 if (!SW_CQE(hw_cqe) && (CQE_WRID_SQ_IDX(hw_cqe) != wq->sq.cidx)) {
594 struct t4_swsqe *swsqe;
596 pr_debug("out of order completion going in sw_sq at idx %u\n",
597 CQE_WRID_SQ_IDX(hw_cqe));
598 swsqe = &wq->sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
599 swsqe->cqe = *hw_cqe;
609 * Reap the associated WR(s) that are freed up with this
612 if (SQ_TYPE(hw_cqe)) {
613 int idx = CQE_WRID_SQ_IDX(hw_cqe);
616 * Account for any unsignaled completions completed by
617 * this signaled completion. In this case, cidx points
618 * to the first unsignaled one, and idx points to the
619 * signaled one. So adjust in_use based on this delta.
620 * if this is not completing any unsigned wrs, then the
621 * delta will be 0. Handle wrapping also!
623 if (idx < wq->sq.cidx)
624 wq->sq.in_use -= wq->sq.size + idx - wq->sq.cidx;
626 wq->sq.in_use -= idx - wq->sq.cidx;
628 wq->sq.cidx = (uint16_t)idx;
629 pr_debug("completing sq idx %u\n", wq->sq.cidx);
630 *cookie = wq->sq.sw_sq[wq->sq.cidx].wr_id;
632 c4iw_log_wr_stats(wq, hw_cqe);
635 pr_debug("completing rq idx %u\n", wq->rq.cidx);
636 *cookie = wq->rq.sw_rq[wq->rq.cidx].wr_id;
638 c4iw_log_wr_stats(wq, hw_cqe);
645 * Flush any completed cqes that are now in-order.
647 flush_completed_wrs(wq, cq);
650 if (SW_CQE(hw_cqe)) {
651 pr_debug("cq %p cqid 0x%x skip sw cqe cidx %u\n",
652 cq, cq->cqid, cq->sw_cidx);
655 pr_debug("cq %p cqid 0x%x skip hw cqe cidx %u\n",
656 cq, cq->cqid, cq->cidx);
663 * Get one cq entry from c4iw and map it to openib.
668 * -EAGAIN caller must try again
669 * any other -errno fatal error
671 static int c4iw_poll_cq_one(struct c4iw_cq *chp, struct ib_wc *wc)
673 struct c4iw_qp *qhp = NULL;
674 struct t4_cqe uninitialized_var(cqe), *rd_cqe;
681 ret = t4_next_cqe(&chp->cq, &rd_cqe);
686 qhp = get_qhp(chp->rhp, CQE_QPID(rd_cqe));
690 spin_lock(&qhp->lock);
693 ret = poll_cq(wq, &(chp->cq), &cqe, &cqe_flushed, &cookie, &credit);
699 wc->vendor_err = CQE_STATUS(&cqe);
702 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",
704 CQE_TYPE(&cqe), CQE_OPCODE(&cqe),
705 CQE_STATUS(&cqe), CQE_LEN(&cqe),
706 CQE_WRID_HI(&cqe), CQE_WRID_LOW(&cqe),
707 (unsigned long long)cookie);
709 if (CQE_TYPE(&cqe) == 0) {
710 if (!CQE_STATUS(&cqe))
711 wc->byte_len = CQE_LEN(&cqe);
714 wc->opcode = IB_WC_RECV;
715 if (CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_INV ||
716 CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_SE_INV) {
717 wc->ex.invalidate_rkey = CQE_WRID_STAG(&cqe);
718 wc->wc_flags |= IB_WC_WITH_INVALIDATE;
719 c4iw_invalidate_mr(qhp->rhp, wc->ex.invalidate_rkey);
722 switch (CQE_OPCODE(&cqe)) {
723 case FW_RI_RDMA_WRITE:
724 wc->opcode = IB_WC_RDMA_WRITE;
727 wc->opcode = IB_WC_RDMA_READ;
728 wc->byte_len = CQE_LEN(&cqe);
730 case FW_RI_SEND_WITH_INV:
731 case FW_RI_SEND_WITH_SE_INV:
732 wc->opcode = IB_WC_SEND;
733 wc->wc_flags |= IB_WC_WITH_INVALIDATE;
736 case FW_RI_SEND_WITH_SE:
737 wc->opcode = IB_WC_SEND;
740 case FW_RI_LOCAL_INV:
741 wc->opcode = IB_WC_LOCAL_INV;
743 case FW_RI_FAST_REGISTER:
744 wc->opcode = IB_WC_REG_MR;
746 /* Invalidate the MR if the fastreg failed */
747 if (CQE_STATUS(&cqe) != T4_ERR_SUCCESS)
748 c4iw_invalidate_mr(qhp->rhp,
749 CQE_WRID_FR_STAG(&cqe));
752 pr_err("Unexpected opcode %d in the CQE received for QPID=0x%0x\n",
753 CQE_OPCODE(&cqe), CQE_QPID(&cqe));
760 wc->status = IB_WC_WR_FLUSH_ERR;
763 switch (CQE_STATUS(&cqe)) {
765 wc->status = IB_WC_SUCCESS;
768 wc->status = IB_WC_LOC_ACCESS_ERR;
771 wc->status = IB_WC_LOC_PROT_ERR;
775 wc->status = IB_WC_LOC_ACCESS_ERR;
778 wc->status = IB_WC_GENERAL_ERR;
781 wc->status = IB_WC_LOC_LEN_ERR;
783 case T4_ERR_INVALIDATE_SHARED_MR:
784 case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
785 wc->status = IB_WC_MW_BIND_ERR;
789 case T4_ERR_PDU_LEN_ERR:
790 case T4_ERR_OUT_OF_RQE:
791 case T4_ERR_DDP_VERSION:
792 case T4_ERR_RDMA_VERSION:
793 case T4_ERR_DDP_QUEUE_NUM:
797 case T4_ERR_MSN_RANGE:
798 case T4_ERR_IRD_OVERFLOW:
800 case T4_ERR_INTERNAL_ERR:
801 wc->status = IB_WC_FATAL_ERR;
804 wc->status = IB_WC_WR_FLUSH_ERR;
807 pr_err("Unexpected cqe_status 0x%x for QPID=0x%0x\n",
808 CQE_STATUS(&cqe), CQE_QPID(&cqe));
809 wc->status = IB_WC_FATAL_ERR;
814 spin_unlock(&qhp->lock);
818 int c4iw_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
825 chp = to_c4iw_cq(ibcq);
827 spin_lock_irqsave(&chp->lock, flags);
828 for (npolled = 0; npolled < num_entries; ++npolled) {
830 err = c4iw_poll_cq_one(chp, wc + npolled);
831 } while (err == -EAGAIN);
835 spin_unlock_irqrestore(&chp->lock, flags);
836 return !err || err == -ENODATA ? npolled : err;
839 int c4iw_destroy_cq(struct ib_cq *ib_cq)
842 struct c4iw_ucontext *ucontext;
844 pr_debug("ib_cq %p\n", ib_cq);
845 chp = to_c4iw_cq(ib_cq);
847 remove_handle(chp->rhp, &chp->rhp->cqidr, chp->cq.cqid);
848 atomic_dec(&chp->refcnt);
849 wait_event(chp->wait, !atomic_read(&chp->refcnt));
851 ucontext = ib_cq->uobject ? to_c4iw_ucontext(ib_cq->uobject->context)
853 destroy_cq(&chp->rhp->rdev, &chp->cq,
854 ucontext ? &ucontext->uctx : &chp->cq.rdev->uctx,
855 chp->destroy_skb, chp->wr_waitp);
856 c4iw_put_wr_wait(chp->wr_waitp);
861 struct ib_cq *c4iw_create_cq(struct ib_device *ibdev,
862 const struct ib_cq_init_attr *attr,
863 struct ib_ucontext *ib_context,
864 struct ib_udata *udata)
866 int entries = attr->cqe;
867 int vector = attr->comp_vector;
868 struct c4iw_dev *rhp;
870 struct c4iw_create_cq_resp uresp;
871 struct c4iw_ucontext *ucontext = NULL;
873 size_t memsize, hwentries;
874 struct c4iw_mm_entry *mm, *mm2;
876 pr_debug("ib_dev %p entries %d\n", ibdev, entries);
878 return ERR_PTR(-EINVAL);
880 rhp = to_c4iw_dev(ibdev);
882 if (vector >= rhp->rdev.lldi.nciq)
883 return ERR_PTR(-EINVAL);
885 chp = kzalloc(sizeof(*chp), GFP_KERNEL);
887 return ERR_PTR(-ENOMEM);
888 chp->wr_waitp = c4iw_alloc_wr_wait(GFP_KERNEL);
889 if (!chp->wr_waitp) {
893 c4iw_init_wr_wait(chp->wr_waitp);
895 wr_len = sizeof(struct fw_ri_res_wr) + sizeof(struct fw_ri_res);
896 chp->destroy_skb = alloc_skb(wr_len, GFP_KERNEL);
897 if (!chp->destroy_skb) {
899 goto err_free_wr_wait;
903 ucontext = to_c4iw_ucontext(ib_context);
905 /* account for the status page. */
908 /* IQ needs one extra entry to differentiate full vs empty. */
912 * entries must be multiple of 16 for HW.
914 entries = roundup(entries, 16);
917 * Make actual HW queue 2x to avoid cdix_inc overflows.
919 hwentries = min(entries * 2, rhp->rdev.hw_queue.t4_max_iq_size);
922 * Make HW queue at least 64 entries so GTS updates aren't too
928 memsize = hwentries * sizeof *chp->cq.queue;
931 * memsize must be a multiple of the page size if its a user cq.
934 memsize = roundup(memsize, PAGE_SIZE);
935 chp->cq.size = hwentries;
936 chp->cq.memsize = memsize;
937 chp->cq.vector = vector;
939 ret = create_cq(&rhp->rdev, &chp->cq,
940 ucontext ? &ucontext->uctx : &rhp->rdev.uctx,
946 chp->cq.size--; /* status page */
947 chp->ibcq.cqe = entries - 2;
948 spin_lock_init(&chp->lock);
949 spin_lock_init(&chp->comp_handler_lock);
950 atomic_set(&chp->refcnt, 1);
951 init_waitqueue_head(&chp->wait);
952 ret = insert_handle(rhp, &rhp->cqidr, chp, chp->cq.cqid);
958 mm = kmalloc(sizeof *mm, GFP_KERNEL);
960 goto err_remove_handle;
961 mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
965 uresp.qid_mask = rhp->rdev.cqmask;
966 uresp.cqid = chp->cq.cqid;
967 uresp.size = chp->cq.size;
968 uresp.memsize = chp->cq.memsize;
969 spin_lock(&ucontext->mmap_lock);
970 uresp.key = ucontext->key;
971 ucontext->key += PAGE_SIZE;
972 uresp.gts_key = ucontext->key;
973 ucontext->key += PAGE_SIZE;
974 spin_unlock(&ucontext->mmap_lock);
975 ret = ib_copy_to_udata(udata, &uresp,
976 sizeof(uresp) - sizeof(uresp.reserved));
981 mm->addr = virt_to_phys(chp->cq.queue);
982 mm->len = chp->cq.memsize;
983 insert_mmap(ucontext, mm);
985 mm2->key = uresp.gts_key;
986 mm2->addr = chp->cq.bar2_pa;
987 mm2->len = PAGE_SIZE;
988 insert_mmap(ucontext, mm2);
990 pr_debug("cqid 0x%0x chp %p size %u memsize %zu, dma_addr 0x%0llx\n",
991 chp->cq.cqid, chp, chp->cq.size,
992 chp->cq.memsize, (unsigned long long)chp->cq.dma_addr);
999 remove_handle(rhp, &rhp->cqidr, chp->cq.cqid);
1001 destroy_cq(&chp->rhp->rdev, &chp->cq,
1002 ucontext ? &ucontext->uctx : &rhp->rdev.uctx,
1003 chp->destroy_skb, chp->wr_waitp);
1005 kfree_skb(chp->destroy_skb);
1007 c4iw_put_wr_wait(chp->wr_waitp);
1010 return ERR_PTR(ret);
1013 int c4iw_resize_cq(struct ib_cq *cq, int cqe, struct ib_udata *udata)
1018 int c4iw_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
1020 struct c4iw_cq *chp;
1024 chp = to_c4iw_cq(ibcq);
1025 spin_lock_irqsave(&chp->lock, flag);
1027 (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED);
1028 if (flags & IB_CQ_REPORT_MISSED_EVENTS)
1029 ret = t4_cq_notempty(&chp->cq);
1030 spin_unlock_irqrestore(&chp->lock, flag);