1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33#include <linux/kref.h>
34#include <rdma/ib_umem.h>
35#include <rdma/ib_user_verbs.h>
36#include "mlx5_ib.h"
37#include "user.h"
38
39static void mlx5_ib_cq_comp(struct mlx5_core_cq *cq)
40{
41 struct ib_cq *ibcq = &to_mibcq(cq)->ibcq;
42
43 ibcq->comp_handler(ibcq, ibcq->cq_context);
44}
45
46static void mlx5_ib_cq_event(struct mlx5_core_cq *mcq, enum mlx5_event type)
47{
48 struct mlx5_ib_cq *cq = container_of(mcq, struct mlx5_ib_cq, mcq);
49 struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
50 struct ib_cq *ibcq = &cq->ibcq;
51 struct ib_event event;
52
53 if (type != MLX5_EVENT_TYPE_CQ_ERROR) {
54 mlx5_ib_warn(dev, "Unexpected event type %d on CQ %06x\n",
55 type, mcq->cqn);
56 return;
57 }
58
59 if (ibcq->event_handler) {
60 event.device = &dev->ib_dev;
61 event.event = IB_EVENT_CQ_ERR;
62 event.element.cq = ibcq;
63 ibcq->event_handler(&event, ibcq->cq_context);
64 }
65}
66
67static void *get_cqe_from_buf(struct mlx5_ib_cq_buf *buf, int n, int size)
68{
69 return mlx5_buf_offset(&buf->buf, n * size);
70}
71
72static void *get_cqe(struct mlx5_ib_cq *cq, int n)
73{
74 return get_cqe_from_buf(&cq->buf, n, cq->mcq.cqe_sz);
75}
76
77static u8 sw_ownership_bit(int n, int nent)
78{
79 return (n & nent) ? 1 : 0;
80}
81
82static void *get_sw_cqe(struct mlx5_ib_cq *cq, int n)
83{
84 void *cqe = get_cqe(cq, n & cq->ibcq.cqe);
85 struct mlx5_cqe64 *cqe64;
86
87 cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
88
89 if (likely((cqe64->op_own) >> 4 != MLX5_CQE_INVALID) &&
90 !((cqe64->op_own & MLX5_CQE_OWNER_MASK) ^ !!(n & (cq->ibcq.cqe + 1)))) {
91 return cqe;
92 } else {
93 return NULL;
94 }
95}
96
97static void *next_cqe_sw(struct mlx5_ib_cq *cq)
98{
99 return get_sw_cqe(cq, cq->mcq.cons_index);
100}
101
102static enum ib_wc_opcode get_umr_comp(struct mlx5_ib_wq *wq, int idx)
103{
104 switch (wq->wr_data[idx]) {
105 case MLX5_IB_WR_UMR:
106 return 0;
107
108 case IB_WR_LOCAL_INV:
109 return IB_WC_LOCAL_INV;
110
111 case IB_WR_FAST_REG_MR:
112 return IB_WC_FAST_REG_MR;
113
114 default:
115 pr_warn("unknown completion status\n");
116 return 0;
117 }
118}
119
120static void handle_good_req(struct ib_wc *wc, struct mlx5_cqe64 *cqe,
121 struct mlx5_ib_wq *wq, int idx)
122{
123 wc->wc_flags = 0;
124 switch (be32_to_cpu(cqe->sop_drop_qpn) >> 24) {
125 case MLX5_OPCODE_RDMA_WRITE_IMM:
126 wc->wc_flags |= IB_WC_WITH_IMM;
127 case MLX5_OPCODE_RDMA_WRITE:
128 wc->opcode = IB_WC_RDMA_WRITE;
129 break;
130 case MLX5_OPCODE_SEND_IMM:
131 wc->wc_flags |= IB_WC_WITH_IMM;
132 case MLX5_OPCODE_SEND:
133 case MLX5_OPCODE_SEND_INVAL:
134 wc->opcode = IB_WC_SEND;
135 break;
136 case MLX5_OPCODE_RDMA_READ:
137 wc->opcode = IB_WC_RDMA_READ;
138 wc->byte_len = be32_to_cpu(cqe->byte_cnt);
139 break;
140 case MLX5_OPCODE_ATOMIC_CS:
141 wc->opcode = IB_WC_COMP_SWAP;
142 wc->byte_len = 8;
143 break;
144 case MLX5_OPCODE_ATOMIC_FA:
145 wc->opcode = IB_WC_FETCH_ADD;
146 wc->byte_len = 8;
147 break;
148 case MLX5_OPCODE_ATOMIC_MASKED_CS:
149 wc->opcode = IB_WC_MASKED_COMP_SWAP;
150 wc->byte_len = 8;
151 break;
152 case MLX5_OPCODE_ATOMIC_MASKED_FA:
153 wc->opcode = IB_WC_MASKED_FETCH_ADD;
154 wc->byte_len = 8;
155 break;
156 case MLX5_OPCODE_BIND_MW:
157 wc->opcode = IB_WC_BIND_MW;
158 break;
159 case MLX5_OPCODE_UMR:
160 wc->opcode = get_umr_comp(wq, idx);
161 break;
162 }
163}
164
165enum {
166 MLX5_GRH_IN_BUFFER = 1,
167 MLX5_GRH_IN_CQE = 2,
168};
169
170static void handle_responder(struct ib_wc *wc, struct mlx5_cqe64 *cqe,
171 struct mlx5_ib_qp *qp)
172{
173 struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
174 struct mlx5_ib_srq *srq;
175 struct mlx5_ib_wq *wq;
176 u16 wqe_ctr;
177 u8 g;
178
179 if (qp->ibqp.srq || qp->ibqp.xrcd) {
180 struct mlx5_core_srq *msrq = NULL;
181
182 if (qp->ibqp.xrcd) {
183 msrq = mlx5_core_get_srq(dev->mdev,
184 be32_to_cpu(cqe->srqn));
185 srq = to_mibsrq(msrq);
186 } else {
187 srq = to_msrq(qp->ibqp.srq);
188 }
189 if (srq) {
190 wqe_ctr = be16_to_cpu(cqe->wqe_counter);
191 wc->wr_id = srq->wrid[wqe_ctr];
192 mlx5_ib_free_srq_wqe(srq, wqe_ctr);
193 if (msrq && atomic_dec_and_test(&msrq->refcount))
194 complete(&msrq->free);
195 }
196 } else {
197 wq = &qp->rq;
198 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
199 ++wq->tail;
200 }
201 wc->byte_len = be32_to_cpu(cqe->byte_cnt);
202
203 switch (cqe->op_own >> 4) {
204 case MLX5_CQE_RESP_WR_IMM:
205 wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
206 wc->wc_flags = IB_WC_WITH_IMM;
207 wc->ex.imm_data = cqe->imm_inval_pkey;
208 break;
209 case MLX5_CQE_RESP_SEND:
210 wc->opcode = IB_WC_RECV;
211 wc->wc_flags = 0;
212 break;
213 case MLX5_CQE_RESP_SEND_IMM:
214 wc->opcode = IB_WC_RECV;
215 wc->wc_flags = IB_WC_WITH_IMM;
216 wc->ex.imm_data = cqe->imm_inval_pkey;
217 break;
218 case MLX5_CQE_RESP_SEND_INV:
219 wc->opcode = IB_WC_RECV;
220 wc->wc_flags = IB_WC_WITH_INVALIDATE;
221 wc->ex.invalidate_rkey = be32_to_cpu(cqe->imm_inval_pkey);
222 break;
223 }
224 wc->slid = be16_to_cpu(cqe->slid);
225 wc->sl = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0xf;
226 wc->src_qp = be32_to_cpu(cqe->flags_rqpn) & 0xffffff;
227 wc->dlid_path_bits = cqe->ml_path;
228 g = (be32_to_cpu(cqe->flags_rqpn) >> 28) & 3;
229 wc->wc_flags |= g ? IB_WC_GRH : 0;
230 wc->pkey_index = be32_to_cpu(cqe->imm_inval_pkey) & 0xffff;
231}
232
233static void dump_cqe(struct mlx5_ib_dev *dev, struct mlx5_err_cqe *cqe)
234{
235 __be32 *p = (__be32 *)cqe;
236 int i;
237
238 mlx5_ib_warn(dev, "dump error cqe\n");
239 for (i = 0; i < sizeof(*cqe) / 16; i++, p += 4)
240 pr_info("%08x %08x %08x %08x\n", be32_to_cpu(p[0]),
241 be32_to_cpu(p[1]), be32_to_cpu(p[2]),
242 be32_to_cpu(p[3]));
243}
244
245static void mlx5_handle_error_cqe(struct mlx5_ib_dev *dev,
246 struct mlx5_err_cqe *cqe,
247 struct ib_wc *wc)
248{
249 int dump = 1;
250
251 switch (cqe->syndrome) {
252 case MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR:
253 wc->status = IB_WC_LOC_LEN_ERR;
254 break;
255 case MLX5_CQE_SYNDROME_LOCAL_QP_OP_ERR:
256 wc->status = IB_WC_LOC_QP_OP_ERR;
257 break;
258 case MLX5_CQE_SYNDROME_LOCAL_PROT_ERR:
259 wc->status = IB_WC_LOC_PROT_ERR;
260 break;
261 case MLX5_CQE_SYNDROME_WR_FLUSH_ERR:
262 dump = 0;
263 wc->status = IB_WC_WR_FLUSH_ERR;
264 break;
265 case MLX5_CQE_SYNDROME_MW_BIND_ERR:
266 wc->status = IB_WC_MW_BIND_ERR;
267 break;
268 case MLX5_CQE_SYNDROME_BAD_RESP_ERR:
269 wc->status = IB_WC_BAD_RESP_ERR;
270 break;
271 case MLX5_CQE_SYNDROME_LOCAL_ACCESS_ERR:
272 wc->status = IB_WC_LOC_ACCESS_ERR;
273 break;
274 case MLX5_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
275 wc->status = IB_WC_REM_INV_REQ_ERR;
276 break;
277 case MLX5_CQE_SYNDROME_REMOTE_ACCESS_ERR:
278 wc->status = IB_WC_REM_ACCESS_ERR;
279 break;
280 case MLX5_CQE_SYNDROME_REMOTE_OP_ERR:
281 wc->status = IB_WC_REM_OP_ERR;
282 break;
283 case MLX5_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
284 wc->status = IB_WC_RETRY_EXC_ERR;
285 dump = 0;
286 break;
287 case MLX5_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
288 wc->status = IB_WC_RNR_RETRY_EXC_ERR;
289 dump = 0;
290 break;
291 case MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR:
292 wc->status = IB_WC_REM_ABORT_ERR;
293 break;
294 default:
295 wc->status = IB_WC_GENERAL_ERR;
296 break;
297 }
298
299 wc->vendor_err = cqe->vendor_err_synd;
300 if (dump)
301 dump_cqe(dev, cqe);
302}
303
304static int is_atomic_response(struct mlx5_ib_qp *qp, uint16_t idx)
305{
306
307
308 return 0;
309}
310
311static void *mlx5_get_atomic_laddr(struct mlx5_ib_qp *qp, uint16_t idx)
312{
313 struct mlx5_wqe_data_seg *dpseg;
314 void *addr;
315
316 dpseg = mlx5_get_send_wqe(qp, idx) + sizeof(struct mlx5_wqe_ctrl_seg) +
317 sizeof(struct mlx5_wqe_raddr_seg) +
318 sizeof(struct mlx5_wqe_atomic_seg);
319 addr = (void *)(unsigned long)be64_to_cpu(dpseg->addr);
320 return addr;
321}
322
323static void handle_atomic(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
324 uint16_t idx)
325{
326 void *addr;
327 int byte_count;
328 int i;
329
330 if (!is_atomic_response(qp, idx))
331 return;
332
333 byte_count = be32_to_cpu(cqe64->byte_cnt);
334 addr = mlx5_get_atomic_laddr(qp, idx);
335
336 if (byte_count == 4) {
337 *(uint32_t *)addr = be32_to_cpu(*((__be32 *)addr));
338 } else {
339 for (i = 0; i < byte_count; i += 8) {
340 *(uint64_t *)addr = be64_to_cpu(*((__be64 *)addr));
341 addr += 8;
342 }
343 }
344
345 return;
346}
347
348static void handle_atomics(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
349 u16 tail, u16 head)
350{
351 u16 idx;
352
353 do {
354 idx = tail & (qp->sq.wqe_cnt - 1);
355 handle_atomic(qp, cqe64, idx);
356 if (idx == head)
357 break;
358
359 tail = qp->sq.w_list[idx].next;
360 } while (1);
361 tail = qp->sq.w_list[idx].next;
362 qp->sq.last_poll = tail;
363}
364
365static void free_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf)
366{
367 mlx5_buf_free(dev->mdev, &buf->buf);
368}
369
370static void get_sig_err_item(struct mlx5_sig_err_cqe *cqe,
371 struct ib_sig_err *item)
372{
373 u16 syndrome = be16_to_cpu(cqe->syndrome);
374
375#define GUARD_ERR (1 << 13)
376#define APPTAG_ERR (1 << 12)
377#define REFTAG_ERR (1 << 11)
378
379 if (syndrome & GUARD_ERR) {
380 item->err_type = IB_SIG_BAD_GUARD;
381 item->expected = be32_to_cpu(cqe->expected_trans_sig) >> 16;
382 item->actual = be32_to_cpu(cqe->actual_trans_sig) >> 16;
383 } else
384 if (syndrome & REFTAG_ERR) {
385 item->err_type = IB_SIG_BAD_REFTAG;
386 item->expected = be32_to_cpu(cqe->expected_reftag);
387 item->actual = be32_to_cpu(cqe->actual_reftag);
388 } else
389 if (syndrome & APPTAG_ERR) {
390 item->err_type = IB_SIG_BAD_APPTAG;
391 item->expected = be32_to_cpu(cqe->expected_trans_sig) & 0xffff;
392 item->actual = be32_to_cpu(cqe->actual_trans_sig) & 0xffff;
393 } else {
394 pr_err("Got signature completion error with bad syndrome %04x\n",
395 syndrome);
396 }
397
398 item->sig_err_offset = be64_to_cpu(cqe->err_offset);
399 item->key = be32_to_cpu(cqe->mkey);
400}
401
402static int mlx5_poll_one(struct mlx5_ib_cq *cq,
403 struct mlx5_ib_qp **cur_qp,
404 struct ib_wc *wc)
405{
406 struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
407 struct mlx5_err_cqe *err_cqe;
408 struct mlx5_cqe64 *cqe64;
409 struct mlx5_core_qp *mqp;
410 struct mlx5_ib_wq *wq;
411 struct mlx5_sig_err_cqe *sig_err_cqe;
412 struct mlx5_core_mr *mmr;
413 struct mlx5_ib_mr *mr;
414 uint8_t opcode;
415 uint32_t qpn;
416 u16 wqe_ctr;
417 void *cqe;
418 int idx;
419
420repoll:
421 cqe = next_cqe_sw(cq);
422 if (!cqe)
423 return -EAGAIN;
424
425 cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
426
427 ++cq->mcq.cons_index;
428
429
430
431
432 rmb();
433
434 opcode = cqe64->op_own >> 4;
435 if (unlikely(opcode == MLX5_CQE_RESIZE_CQ)) {
436 if (likely(cq->resize_buf)) {
437 free_cq_buf(dev, &cq->buf);
438 cq->buf = *cq->resize_buf;
439 kfree(cq->resize_buf);
440 cq->resize_buf = NULL;
441 goto repoll;
442 } else {
443 mlx5_ib_warn(dev, "unexpected resize cqe\n");
444 }
445 }
446
447 qpn = ntohl(cqe64->sop_drop_qpn) & 0xffffff;
448 if (!*cur_qp || (qpn != (*cur_qp)->ibqp.qp_num)) {
449
450
451
452
453 mqp = __mlx5_qp_lookup(dev->mdev, qpn);
454 if (unlikely(!mqp)) {
455 mlx5_ib_warn(dev, "CQE@CQ %06x for unknown QPN %6x\n",
456 cq->mcq.cqn, qpn);
457 return -EINVAL;
458 }
459
460 *cur_qp = to_mibqp(mqp);
461 }
462
463 wc->qp = &(*cur_qp)->ibqp;
464 switch (opcode) {
465 case MLX5_CQE_REQ:
466 wq = &(*cur_qp)->sq;
467 wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
468 idx = wqe_ctr & (wq->wqe_cnt - 1);
469 handle_good_req(wc, cqe64, wq, idx);
470 handle_atomics(*cur_qp, cqe64, wq->last_poll, idx);
471 wc->wr_id = wq->wrid[idx];
472 wq->tail = wq->wqe_head[idx] + 1;
473 wc->status = IB_WC_SUCCESS;
474 break;
475 case MLX5_CQE_RESP_WR_IMM:
476 case MLX5_CQE_RESP_SEND:
477 case MLX5_CQE_RESP_SEND_IMM:
478 case MLX5_CQE_RESP_SEND_INV:
479 handle_responder(wc, cqe64, *cur_qp);
480 wc->status = IB_WC_SUCCESS;
481 break;
482 case MLX5_CQE_RESIZE_CQ:
483 break;
484 case MLX5_CQE_REQ_ERR:
485 case MLX5_CQE_RESP_ERR:
486 err_cqe = (struct mlx5_err_cqe *)cqe64;
487 mlx5_handle_error_cqe(dev, err_cqe, wc);
488 mlx5_ib_dbg(dev, "%s error cqe on cqn 0x%x:\n",
489 opcode == MLX5_CQE_REQ_ERR ?
490 "Requestor" : "Responder", cq->mcq.cqn);
491 mlx5_ib_dbg(dev, "syndrome 0x%x, vendor syndrome 0x%x\n",
492 err_cqe->syndrome, err_cqe->vendor_err_synd);
493 if (opcode == MLX5_CQE_REQ_ERR) {
494 wq = &(*cur_qp)->sq;
495 wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
496 idx = wqe_ctr & (wq->wqe_cnt - 1);
497 wc->wr_id = wq->wrid[idx];
498 wq->tail = wq->wqe_head[idx] + 1;
499 } else {
500 struct mlx5_ib_srq *srq;
501
502 if ((*cur_qp)->ibqp.srq) {
503 srq = to_msrq((*cur_qp)->ibqp.srq);
504 wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
505 wc->wr_id = srq->wrid[wqe_ctr];
506 mlx5_ib_free_srq_wqe(srq, wqe_ctr);
507 } else {
508 wq = &(*cur_qp)->rq;
509 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
510 ++wq->tail;
511 }
512 }
513 break;
514 case MLX5_CQE_SIG_ERR:
515 sig_err_cqe = (struct mlx5_sig_err_cqe *)cqe64;
516
517 read_lock(&dev->mdev->priv.mr_table.lock);
518 mmr = __mlx5_mr_lookup(dev->mdev,
519 mlx5_base_mkey(be32_to_cpu(sig_err_cqe->mkey)));
520 if (unlikely(!mmr)) {
521 read_unlock(&dev->mdev->priv.mr_table.lock);
522 mlx5_ib_warn(dev, "CQE@CQ %06x for unknown MR %6x\n",
523 cq->mcq.cqn, be32_to_cpu(sig_err_cqe->mkey));
524 return -EINVAL;
525 }
526
527 mr = to_mibmr(mmr);
528 get_sig_err_item(sig_err_cqe, &mr->sig->err_item);
529 mr->sig->sig_err_exists = true;
530 mr->sig->sigerr_count++;
531
532 mlx5_ib_warn(dev, "CQN: 0x%x Got SIGERR on key: 0x%x err_type %x err_offset %llx expected %x actual %x\n",
533 cq->mcq.cqn, mr->sig->err_item.key,
534 mr->sig->err_item.err_type,
535 mr->sig->err_item.sig_err_offset,
536 mr->sig->err_item.expected,
537 mr->sig->err_item.actual);
538
539 read_unlock(&dev->mdev->priv.mr_table.lock);
540 goto repoll;
541 }
542
543 return 0;
544}
545
546int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
547{
548 struct mlx5_ib_cq *cq = to_mcq(ibcq);
549 struct mlx5_ib_qp *cur_qp = NULL;
550 unsigned long flags;
551 int npolled;
552 int err = 0;
553
554 spin_lock_irqsave(&cq->lock, flags);
555
556 for (npolled = 0; npolled < num_entries; npolled++) {
557 err = mlx5_poll_one(cq, &cur_qp, wc + npolled);
558 if (err)
559 break;
560 }
561
562 if (npolled)
563 mlx5_cq_set_ci(&cq->mcq);
564
565 spin_unlock_irqrestore(&cq->lock, flags);
566
567 if (err == 0 || err == -EAGAIN)
568 return npolled;
569 else
570 return err;
571}
572
573int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
574{
575 struct mlx5_core_dev *mdev = to_mdev(ibcq->device)->mdev;
576 void __iomem *uar_page = mdev->priv.uuari.uars[0].map;
577
578 mlx5_cq_arm(&to_mcq(ibcq)->mcq,
579 (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
580 MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT,
581 uar_page,
582 MLX5_GET_DOORBELL_LOCK(&mdev->priv.cq_uar_lock),
583 to_mcq(ibcq)->mcq.cons_index);
584
585 return 0;
586}
587
588static int alloc_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf,
589 int nent, int cqe_size)
590{
591 int err;
592
593 err = mlx5_buf_alloc(dev->mdev, nent * cqe_size,
594 PAGE_SIZE * 2, &buf->buf);
595 if (err)
596 return err;
597
598 buf->cqe_size = cqe_size;
599 buf->nent = nent;
600
601 return 0;
602}
603
604static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
605 struct ib_ucontext *context, struct mlx5_ib_cq *cq,
606 int entries, struct mlx5_create_cq_mbox_in **cqb,
607 int *cqe_size, int *index, int *inlen)
608{
609 struct mlx5_ib_create_cq ucmd;
610 size_t ucmdlen;
611 int page_shift;
612 int npages;
613 int ncont;
614 int err;
615
616 ucmdlen =
617 (udata->inlen - sizeof(struct ib_uverbs_cmd_hdr) <
618 sizeof(ucmd)) ? (sizeof(ucmd) -
619 sizeof(ucmd.reserved)) : sizeof(ucmd);
620
621 if (ib_copy_from_udata(&ucmd, udata, ucmdlen))
622 return -EFAULT;
623
624 if (ucmdlen == sizeof(ucmd) &&
625 ucmd.reserved != 0)
626 return -EINVAL;
627
628 if (ucmd.cqe_size != 64 && ucmd.cqe_size != 128)
629 return -EINVAL;
630
631 *cqe_size = ucmd.cqe_size;
632
633 cq->buf.umem = ib_umem_get(context, ucmd.buf_addr,
634 entries * ucmd.cqe_size,
635 IB_ACCESS_LOCAL_WRITE, 1);
636 if (IS_ERR(cq->buf.umem)) {
637 err = PTR_ERR(cq->buf.umem);
638 return err;
639 }
640
641 err = mlx5_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
642 &cq->db);
643 if (err)
644 goto err_umem;
645
646 mlx5_ib_cont_pages(cq->buf.umem, ucmd.buf_addr, &npages, &page_shift,
647 &ncont, NULL);
648 mlx5_ib_dbg(dev, "addr 0x%llx, size %u, npages %d, page_shift %d, ncont %d\n",
649 ucmd.buf_addr, entries * ucmd.cqe_size, npages, page_shift, ncont);
650
651 *inlen = sizeof(**cqb) + sizeof(*(*cqb)->pas) * ncont;
652 *cqb = mlx5_vzalloc(*inlen);
653 if (!*cqb) {
654 err = -ENOMEM;
655 goto err_db;
656 }
657 mlx5_ib_populate_pas(dev, cq->buf.umem, page_shift, (*cqb)->pas, 0);
658 (*cqb)->ctx.log_pg_sz = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
659
660 *index = to_mucontext(context)->uuari.uars[0].index;
661
662 return 0;
663
664err_db:
665 mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db);
666
667err_umem:
668 ib_umem_release(cq->buf.umem);
669 return err;
670}
671
672static void destroy_cq_user(struct mlx5_ib_cq *cq, struct ib_ucontext *context)
673{
674 mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db);
675 ib_umem_release(cq->buf.umem);
676}
677
678static void init_cq_buf(struct mlx5_ib_cq *cq, struct mlx5_ib_cq_buf *buf)
679{
680 int i;
681 void *cqe;
682 struct mlx5_cqe64 *cqe64;
683
684 for (i = 0; i < buf->nent; i++) {
685 cqe = get_cqe_from_buf(buf, i, buf->cqe_size);
686 cqe64 = buf->cqe_size == 64 ? cqe : cqe + 64;
687 cqe64->op_own = MLX5_CQE_INVALID << 4;
688 }
689}
690
691static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
692 int entries, int cqe_size,
693 struct mlx5_create_cq_mbox_in **cqb,
694 int *index, int *inlen)
695{
696 int err;
697
698 err = mlx5_db_alloc(dev->mdev, &cq->db);
699 if (err)
700 return err;
701
702 cq->mcq.set_ci_db = cq->db.db;
703 cq->mcq.arm_db = cq->db.db + 1;
704 cq->mcq.cqe_sz = cqe_size;
705
706 err = alloc_cq_buf(dev, &cq->buf, entries, cqe_size);
707 if (err)
708 goto err_db;
709
710 init_cq_buf(cq, &cq->buf);
711
712 *inlen = sizeof(**cqb) + sizeof(*(*cqb)->pas) * cq->buf.buf.npages;
713 *cqb = mlx5_vzalloc(*inlen);
714 if (!*cqb) {
715 err = -ENOMEM;
716 goto err_buf;
717 }
718 mlx5_fill_page_array(&cq->buf.buf, (*cqb)->pas);
719
720 (*cqb)->ctx.log_pg_sz = cq->buf.buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT;
721 *index = dev->mdev->priv.uuari.uars[0].index;
722
723 return 0;
724
725err_buf:
726 free_cq_buf(dev, &cq->buf);
727
728err_db:
729 mlx5_db_free(dev->mdev, &cq->db);
730 return err;
731}
732
733static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
734{
735 free_cq_buf(dev, &cq->buf);
736 mlx5_db_free(dev->mdev, &cq->db);
737}
738
739struct ib_cq *mlx5_ib_create_cq(struct ib_device *ibdev, int entries,
740 int vector, struct ib_ucontext *context,
741 struct ib_udata *udata)
742{
743 struct mlx5_create_cq_mbox_in *cqb = NULL;
744 struct mlx5_ib_dev *dev = to_mdev(ibdev);
745 struct mlx5_ib_cq *cq;
746 int uninitialized_var(index);
747 int uninitialized_var(inlen);
748 int cqe_size;
749 int irqn;
750 int eqn;
751 int err;
752
753 if (entries < 0)
754 return ERR_PTR(-EINVAL);
755
756 entries = roundup_pow_of_two(entries + 1);
757 if (entries > dev->mdev->caps.gen.max_cqes)
758 return ERR_PTR(-EINVAL);
759
760 cq = kzalloc(sizeof(*cq), GFP_KERNEL);
761 if (!cq)
762 return ERR_PTR(-ENOMEM);
763
764 cq->ibcq.cqe = entries - 1;
765 mutex_init(&cq->resize_mutex);
766 spin_lock_init(&cq->lock);
767 cq->resize_buf = NULL;
768 cq->resize_umem = NULL;
769
770 if (context) {
771 err = create_cq_user(dev, udata, context, cq, entries,
772 &cqb, &cqe_size, &index, &inlen);
773 if (err)
774 goto err_create;
775 } else {
776
777 cqe_size = 64;
778 err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb,
779 &index, &inlen);
780 if (err)
781 goto err_create;
782 }
783
784 cq->cqe_size = cqe_size;
785 cqb->ctx.cqe_sz_flags = cqe_sz_to_mlx_sz(cqe_size) << 5;
786 cqb->ctx.log_sz_usr_page = cpu_to_be32((ilog2(entries) << 24) | index);
787 err = mlx5_vector2eqn(dev->mdev, vector, &eqn, &irqn);
788 if (err)
789 goto err_cqb;
790
791 cqb->ctx.c_eqn = cpu_to_be16(eqn);
792 cqb->ctx.db_record_addr = cpu_to_be64(cq->db.dma);
793
794 err = mlx5_core_create_cq(dev->mdev, &cq->mcq, cqb, inlen);
795 if (err)
796 goto err_cqb;
797
798 mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
799 cq->mcq.irqn = irqn;
800 cq->mcq.comp = mlx5_ib_cq_comp;
801 cq->mcq.event = mlx5_ib_cq_event;
802
803 if (context)
804 if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) {
805 err = -EFAULT;
806 goto err_cmd;
807 }
808
809
810 kvfree(cqb);
811 return &cq->ibcq;
812
813err_cmd:
814 mlx5_core_destroy_cq(dev->mdev, &cq->mcq);
815
816err_cqb:
817 kvfree(cqb);
818 if (context)
819 destroy_cq_user(cq, context);
820 else
821 destroy_cq_kernel(dev, cq);
822
823err_create:
824 kfree(cq);
825
826 return ERR_PTR(err);
827}
828
829
830int mlx5_ib_destroy_cq(struct ib_cq *cq)
831{
832 struct mlx5_ib_dev *dev = to_mdev(cq->device);
833 struct mlx5_ib_cq *mcq = to_mcq(cq);
834 struct ib_ucontext *context = NULL;
835
836 if (cq->uobject)
837 context = cq->uobject->context;
838
839 mlx5_core_destroy_cq(dev->mdev, &mcq->mcq);
840 if (context)
841 destroy_cq_user(mcq, context);
842 else
843 destroy_cq_kernel(dev, mcq);
844
845 kfree(mcq);
846
847 return 0;
848}
849
850static int is_equal_rsn(struct mlx5_cqe64 *cqe64, u32 rsn)
851{
852 return rsn == (ntohl(cqe64->sop_drop_qpn) & 0xffffff);
853}
854
855void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 rsn, struct mlx5_ib_srq *srq)
856{
857 struct mlx5_cqe64 *cqe64, *dest64;
858 void *cqe, *dest;
859 u32 prod_index;
860 int nfreed = 0;
861 u8 owner_bit;
862
863 if (!cq)
864 return;
865
866
867
868
869
870
871
872 for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); prod_index++)
873 if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
874 break;
875
876
877
878
879 while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
880 cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
881 cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
882 if (is_equal_rsn(cqe64, rsn)) {
883 if (srq && (ntohl(cqe64->srqn) & 0xffffff))
884 mlx5_ib_free_srq_wqe(srq, be16_to_cpu(cqe64->wqe_counter));
885 ++nfreed;
886 } else if (nfreed) {
887 dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
888 dest64 = (cq->mcq.cqe_sz == 64) ? dest : dest + 64;
889 owner_bit = dest64->op_own & MLX5_CQE_OWNER_MASK;
890 memcpy(dest, cqe, cq->mcq.cqe_sz);
891 dest64->op_own = owner_bit |
892 (dest64->op_own & ~MLX5_CQE_OWNER_MASK);
893 }
894 }
895
896 if (nfreed) {
897 cq->mcq.cons_index += nfreed;
898
899
900
901 wmb();
902 mlx5_cq_set_ci(&cq->mcq);
903 }
904}
905
906void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq)
907{
908 if (!cq)
909 return;
910
911 spin_lock_irq(&cq->lock);
912 __mlx5_ib_cq_clean(cq, qpn, srq);
913 spin_unlock_irq(&cq->lock);
914}
915
916int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
917{
918 struct mlx5_modify_cq_mbox_in *in;
919 struct mlx5_ib_dev *dev = to_mdev(cq->device);
920 struct mlx5_ib_cq *mcq = to_mcq(cq);
921 int err;
922 u32 fsel;
923
924 if (!(dev->mdev->caps.gen.flags & MLX5_DEV_CAP_FLAG_CQ_MODER))
925 return -ENOSYS;
926
927 in = kzalloc(sizeof(*in), GFP_KERNEL);
928 if (!in)
929 return -ENOMEM;
930
931 in->cqn = cpu_to_be32(mcq->mcq.cqn);
932 fsel = (MLX5_CQ_MODIFY_PERIOD | MLX5_CQ_MODIFY_COUNT);
933 in->ctx.cq_period = cpu_to_be16(cq_period);
934 in->ctx.cq_max_count = cpu_to_be16(cq_count);
935 in->field_select = cpu_to_be32(fsel);
936 err = mlx5_core_modify_cq(dev->mdev, &mcq->mcq, in, sizeof(*in));
937 kfree(in);
938
939 if (err)
940 mlx5_ib_warn(dev, "modify cq 0x%x failed\n", mcq->mcq.cqn);
941
942 return err;
943}
944
945static int resize_user(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
946 int entries, struct ib_udata *udata, int *npas,
947 int *page_shift, int *cqe_size)
948{
949 struct mlx5_ib_resize_cq ucmd;
950 struct ib_umem *umem;
951 int err;
952 int npages;
953 struct ib_ucontext *context = cq->buf.umem->context;
954
955 err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
956 if (err)
957 return err;
958
959 if (ucmd.reserved0 || ucmd.reserved1)
960 return -EINVAL;
961
962 umem = ib_umem_get(context, ucmd.buf_addr, entries * ucmd.cqe_size,
963 IB_ACCESS_LOCAL_WRITE, 1);
964 if (IS_ERR(umem)) {
965 err = PTR_ERR(umem);
966 return err;
967 }
968
969 mlx5_ib_cont_pages(umem, ucmd.buf_addr, &npages, page_shift,
970 npas, NULL);
971
972 cq->resize_umem = umem;
973 *cqe_size = ucmd.cqe_size;
974
975 return 0;
976}
977
978static void un_resize_user(struct mlx5_ib_cq *cq)
979{
980 ib_umem_release(cq->resize_umem);
981}
982
983static int resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
984 int entries, int cqe_size)
985{
986 int err;
987
988 cq->resize_buf = kzalloc(sizeof(*cq->resize_buf), GFP_KERNEL);
989 if (!cq->resize_buf)
990 return -ENOMEM;
991
992 err = alloc_cq_buf(dev, cq->resize_buf, entries, cqe_size);
993 if (err)
994 goto ex;
995
996 init_cq_buf(cq, cq->resize_buf);
997
998 return 0;
999
1000ex:
1001 kfree(cq->resize_buf);
1002 return err;
1003}
1004
1005static void un_resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
1006{
1007 free_cq_buf(dev, cq->resize_buf);
1008 cq->resize_buf = NULL;
1009}
1010
1011static int copy_resize_cqes(struct mlx5_ib_cq *cq)
1012{
1013 struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
1014 struct mlx5_cqe64 *scqe64;
1015 struct mlx5_cqe64 *dcqe64;
1016 void *start_cqe;
1017 void *scqe;
1018 void *dcqe;
1019 int ssize;
1020 int dsize;
1021 int i;
1022 u8 sw_own;
1023
1024 ssize = cq->buf.cqe_size;
1025 dsize = cq->resize_buf->cqe_size;
1026 if (ssize != dsize) {
1027 mlx5_ib_warn(dev, "resize from different cqe size is not supported\n");
1028 return -EINVAL;
1029 }
1030
1031 i = cq->mcq.cons_index;
1032 scqe = get_sw_cqe(cq, i);
1033 scqe64 = ssize == 64 ? scqe : scqe + 64;
1034 start_cqe = scqe;
1035 if (!scqe) {
1036 mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1037 return -EINVAL;
1038 }
1039
1040 while ((scqe64->op_own >> 4) != MLX5_CQE_RESIZE_CQ) {
1041 dcqe = get_cqe_from_buf(cq->resize_buf,
1042 (i + 1) & (cq->resize_buf->nent),
1043 dsize);
1044 dcqe64 = dsize == 64 ? dcqe : dcqe + 64;
1045 sw_own = sw_ownership_bit(i + 1, cq->resize_buf->nent);
1046 memcpy(dcqe, scqe, dsize);
1047 dcqe64->op_own = (dcqe64->op_own & ~MLX5_CQE_OWNER_MASK) | sw_own;
1048
1049 ++i;
1050 scqe = get_sw_cqe(cq, i);
1051 scqe64 = ssize == 64 ? scqe : scqe + 64;
1052 if (!scqe) {
1053 mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1054 return -EINVAL;
1055 }
1056
1057 if (scqe == start_cqe) {
1058 pr_warn("resize CQ failed to get resize CQE, CQN 0x%x\n",
1059 cq->mcq.cqn);
1060 return -ENOMEM;
1061 }
1062 }
1063 ++cq->mcq.cons_index;
1064 return 0;
1065}
1066
1067int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
1068{
1069 struct mlx5_ib_dev *dev = to_mdev(ibcq->device);
1070 struct mlx5_ib_cq *cq = to_mcq(ibcq);
1071 struct mlx5_modify_cq_mbox_in *in;
1072 int err;
1073 int npas;
1074 int page_shift;
1075 int inlen;
1076 int uninitialized_var(cqe_size);
1077 unsigned long flags;
1078
1079 if (!(dev->mdev->caps.gen.flags & MLX5_DEV_CAP_FLAG_RESIZE_CQ)) {
1080 pr_info("Firmware does not support resize CQ\n");
1081 return -ENOSYS;
1082 }
1083
1084 if (entries < 1)
1085 return -EINVAL;
1086
1087 entries = roundup_pow_of_two(entries + 1);
1088 if (entries > dev->mdev->caps.gen.max_cqes + 1)
1089 return -EINVAL;
1090
1091 if (entries == ibcq->cqe + 1)
1092 return 0;
1093
1094 mutex_lock(&cq->resize_mutex);
1095 if (udata) {
1096 err = resize_user(dev, cq, entries, udata, &npas, &page_shift,
1097 &cqe_size);
1098 } else {
1099 cqe_size = 64;
1100 err = resize_kernel(dev, cq, entries, cqe_size);
1101 if (!err) {
1102 npas = cq->resize_buf->buf.npages;
1103 page_shift = cq->resize_buf->buf.page_shift;
1104 }
1105 }
1106
1107 if (err)
1108 goto ex;
1109
1110 inlen = sizeof(*in) + npas * sizeof(in->pas[0]);
1111 in = mlx5_vzalloc(inlen);
1112 if (!in) {
1113 err = -ENOMEM;
1114 goto ex_resize;
1115 }
1116
1117 if (udata)
1118 mlx5_ib_populate_pas(dev, cq->resize_umem, page_shift,
1119 in->pas, 0);
1120 else
1121 mlx5_fill_page_array(&cq->resize_buf->buf, in->pas);
1122
1123 in->field_select = cpu_to_be32(MLX5_MODIFY_CQ_MASK_LOG_SIZE |
1124 MLX5_MODIFY_CQ_MASK_PG_OFFSET |
1125 MLX5_MODIFY_CQ_MASK_PG_SIZE);
1126 in->ctx.log_pg_sz = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
1127 in->ctx.cqe_sz_flags = cqe_sz_to_mlx_sz(cqe_size) << 5;
1128 in->ctx.page_offset = 0;
1129 in->ctx.log_sz_usr_page = cpu_to_be32(ilog2(entries) << 24);
1130 in->hdr.opmod = cpu_to_be16(MLX5_CQ_OPMOD_RESIZE);
1131 in->cqn = cpu_to_be32(cq->mcq.cqn);
1132
1133 err = mlx5_core_modify_cq(dev->mdev, &cq->mcq, in, inlen);
1134 if (err)
1135 goto ex_alloc;
1136
1137 if (udata) {
1138 cq->ibcq.cqe = entries - 1;
1139 ib_umem_release(cq->buf.umem);
1140 cq->buf.umem = cq->resize_umem;
1141 cq->resize_umem = NULL;
1142 } else {
1143 struct mlx5_ib_cq_buf tbuf;
1144 int resized = 0;
1145
1146 spin_lock_irqsave(&cq->lock, flags);
1147 if (cq->resize_buf) {
1148 err = copy_resize_cqes(cq);
1149 if (!err) {
1150 tbuf = cq->buf;
1151 cq->buf = *cq->resize_buf;
1152 kfree(cq->resize_buf);
1153 cq->resize_buf = NULL;
1154 resized = 1;
1155 }
1156 }
1157 cq->ibcq.cqe = entries - 1;
1158 spin_unlock_irqrestore(&cq->lock, flags);
1159 if (resized)
1160 free_cq_buf(dev, &tbuf);
1161 }
1162 mutex_unlock(&cq->resize_mutex);
1163
1164 kvfree(in);
1165 return 0;
1166
1167ex_alloc:
1168 kvfree(in);
1169
1170ex_resize:
1171 if (udata)
1172 un_resize_user(cq);
1173 else
1174 un_resize_kernel(dev, cq);
1175ex:
1176 mutex_unlock(&cq->resize_mutex);
1177 return err;
1178}
1179
1180int mlx5_ib_get_cqe_size(struct mlx5_ib_dev *dev, struct ib_cq *ibcq)
1181{
1182 struct mlx5_ib_cq *cq;
1183
1184 if (!ibcq)
1185 return 128;
1186
1187 cq = to_mcq(ibcq);
1188 return cq->cqe_size;
1189}
1190
