1/*
2 * Copyright (c) 2013, Mellanox Technologies inc.  All rights reserved.
3 *
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:
9 *
10 *     Redistribution and use in source and binary forms, with or
11 *     without modification, are permitted provided that the following
12 *     conditions are met:
13 *
14 *      - Redistributions of source code must retain the above
15 *        copyright notice, this list of conditions and the following
16 *        disclaimer.
17 *
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.
22 *
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
30 * SOFTWARE.
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	/* TBD: waiting decision
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	/* Make sure we read CQ entry contents after we've checked the
430	 * ownership bit.
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		/* We do not have to take the QP table lock here,
450		 * because CQs will be locked while QPs are removed
451		 * from the table.
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	mlx5_cq_arm(&to_mcq(ibcq)->mcq,
576		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
577		    MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT,
578		    to_mdev(ibcq->device)->mdev->priv.uuari.uars[0].map,
579		    MLX5_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->mdev->priv.cq_uar_lock));
580
581	return 0;
582}
583
584static int alloc_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf,
585			int nent, int cqe_size)
586{
587	int err;
588
589	err = mlx5_buf_alloc(dev->mdev, nent * cqe_size,
590			     PAGE_SIZE * 2, &buf->buf);
591	if (err)
592		return err;
593
594	buf->cqe_size = cqe_size;
595	buf->nent = nent;
596
597	return 0;
598}
599
600static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
601			  struct ib_ucontext *context, struct mlx5_ib_cq *cq,
602			  int entries, struct mlx5_create_cq_mbox_in **cqb,
603			  int *cqe_size, int *index, int *inlen)
604{
605	struct mlx5_ib_create_cq ucmd;
606	size_t ucmdlen;
607	int page_shift;
608	int npages;
609	int ncont;
610	int err;
611
612	ucmdlen =
613		(udata->inlen - sizeof(struct ib_uverbs_cmd_hdr) <
614		 sizeof(ucmd)) ? (sizeof(ucmd) -
615				  sizeof(ucmd.reserved)) : sizeof(ucmd);
616
617	if (ib_copy_from_udata(&ucmd, udata, ucmdlen))
618		return -EFAULT;
619
620	if (ucmdlen == sizeof(ucmd) &&
621	    ucmd.reserved != 0)
622		return -EINVAL;
623
624	if (ucmd.cqe_size != 64 && ucmd.cqe_size != 128)
625		return -EINVAL;
626
627	*cqe_size = ucmd.cqe_size;
628
629	cq->buf.umem = ib_umem_get(context, ucmd.buf_addr,
630				   entries * ucmd.cqe_size,
631				   IB_ACCESS_LOCAL_WRITE, 1);
632	if (IS_ERR(cq->buf.umem)) {
633		err = PTR_ERR(cq->buf.umem);
634		return err;
635	}
636
637	err = mlx5_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
638				  &cq->db);
639	if (err)
640		goto err_umem;
641
642	mlx5_ib_cont_pages(cq->buf.umem, ucmd.buf_addr, &npages, &page_shift,
643			   &ncont, NULL);
644	mlx5_ib_dbg(dev, "addr 0x%llx, size %u, npages %d, page_shift %d, ncont %d\n",
645		    ucmd.buf_addr, entries * ucmd.cqe_size, npages, page_shift, ncont);
646
647	*inlen = sizeof(**cqb) + sizeof(*(*cqb)->pas) * ncont;
648	*cqb = mlx5_vzalloc(*inlen);
649	if (!*cqb) {
650		err = -ENOMEM;
651		goto err_db;
652	}
653	mlx5_ib_populate_pas(dev, cq->buf.umem, page_shift, (*cqb)->pas, 0);
654	(*cqb)->ctx.log_pg_sz = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
655
656	*index = to_mucontext(context)->uuari.uars[0].index;
657
658	return 0;
659
660err_db:
661	mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db);
662
663err_umem:
664	ib_umem_release(cq->buf.umem);
665	return err;
666}
667
668static void destroy_cq_user(struct mlx5_ib_cq *cq, struct ib_ucontext *context)
669{
670	mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db);
671	ib_umem_release(cq->buf.umem);
672}
673
674static void init_cq_buf(struct mlx5_ib_cq *cq, struct mlx5_ib_cq_buf *buf)
675{
676	int i;
677	void *cqe;
678	struct mlx5_cqe64 *cqe64;
679
680	for (i = 0; i < buf->nent; i++) {
681		cqe = get_cqe_from_buf(buf, i, buf->cqe_size);
682		cqe64 = buf->cqe_size == 64 ? cqe : cqe + 64;
683		cqe64->op_own = MLX5_CQE_INVALID << 4;
684	}
685}
686
687static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
688			    int entries, int cqe_size,
689			    struct mlx5_create_cq_mbox_in **cqb,
690			    int *index, int *inlen)
691{
692	int err;
693
694	err = mlx5_db_alloc(dev->mdev, &cq->db);
695	if (err)
696		return err;
697
698	cq->mcq.set_ci_db  = cq->db.db;
699	cq->mcq.arm_db     = cq->db.db + 1;
700	*cq->mcq.set_ci_db = 0;
701	*cq->mcq.arm_db    = 0;
702	cq->mcq.cqe_sz = cqe_size;
703
704	err = alloc_cq_buf(dev, &cq->buf, entries, cqe_size);
705	if (err)
706		goto err_db;
707
708	init_cq_buf(cq, &cq->buf);
709
710	*inlen = sizeof(**cqb) + sizeof(*(*cqb)->pas) * cq->buf.buf.npages;
711	*cqb = mlx5_vzalloc(*inlen);
712	if (!*cqb) {
713		err = -ENOMEM;
714		goto err_buf;
715	}
716	mlx5_fill_page_array(&cq->buf.buf, (*cqb)->pas);
717
718	(*cqb)->ctx.log_pg_sz = cq->buf.buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT;
719	*index = dev->mdev->priv.uuari.uars[0].index;
720
721	return 0;
722
723err_buf:
724	free_cq_buf(dev, &cq->buf);
725
726err_db:
727	mlx5_db_free(dev->mdev, &cq->db);
728	return err;
729}
730
731static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
732{
733	free_cq_buf(dev, &cq->buf);
734	mlx5_db_free(dev->mdev, &cq->db);
735}
736
737struct ib_cq *mlx5_ib_create_cq(struct ib_device *ibdev, int entries,
738				int vector, struct ib_ucontext *context,
739				struct ib_udata *udata)
740{
741	struct mlx5_create_cq_mbox_in *cqb = NULL;
742	struct mlx5_ib_dev *dev = to_mdev(ibdev);
743	struct mlx5_ib_cq *cq;
744	int uninitialized_var(index);
745	int uninitialized_var(inlen);
746	int cqe_size;
747	int irqn;
748	int eqn;
749	int err;
750
751	if (entries < 0)
752		return ERR_PTR(-EINVAL);
753
754	entries = roundup_pow_of_two(entries + 1);
755	if (entries > dev->mdev->caps.gen.max_cqes)
756		return ERR_PTR(-EINVAL);
757
758	cq = kzalloc(sizeof(*cq), GFP_KERNEL);
759	if (!cq)
760		return ERR_PTR(-ENOMEM);
761
762	cq->ibcq.cqe = entries - 1;
763	mutex_init(&cq->resize_mutex);
764	spin_lock_init(&cq->lock);
765	cq->resize_buf = NULL;
766	cq->resize_umem = NULL;
767
768	if (context) {
769		err = create_cq_user(dev, udata, context, cq, entries,
770				     &cqb, &cqe_size, &index, &inlen);
771		if (err)
772			goto err_create;
773	} else {
774		/* for now choose 64 bytes till we have a proper interface */
775		cqe_size = 64;
776		err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb,
777				       &index, &inlen);
778		if (err)
779			goto err_create;
780	}
781
782	cq->cqe_size = cqe_size;
783	cqb->ctx.cqe_sz_flags = cqe_sz_to_mlx_sz(cqe_size) << 5;
784	cqb->ctx.log_sz_usr_page = cpu_to_be32((ilog2(entries) << 24) | index);
785	err = mlx5_vector2eqn(dev, vector, &eqn, &irqn);
786	if (err)
787		goto err_cqb;
788
789	cqb->ctx.c_eqn = cpu_to_be16(eqn);
790	cqb->ctx.db_record_addr = cpu_to_be64(cq->db.dma);
791
792	err = mlx5_core_create_cq(dev->mdev, &cq->mcq, cqb, inlen);
793	if (err)
794		goto err_cqb;
795
796	mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
797	cq->mcq.irqn = irqn;
798	cq->mcq.comp  = mlx5_ib_cq_comp;
799	cq->mcq.event = mlx5_ib_cq_event;
800
801	if (context)
802		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) {
803			err = -EFAULT;
804			goto err_cmd;
805		}
806
807
808	mlx5_vfree(cqb);
809	return &cq->ibcq;
810
811err_cmd:
812	mlx5_core_destroy_cq(dev->mdev, &cq->mcq);
813
814err_cqb:
815	mlx5_vfree(cqb);
816	if (context)
817		destroy_cq_user(cq, context);
818	else
819		destroy_cq_kernel(dev, cq);
820
821err_create:
822	kfree(cq);
823
824	return ERR_PTR(err);
825}
826
827
828int mlx5_ib_destroy_cq(struct ib_cq *cq)
829{
830	struct mlx5_ib_dev *dev = to_mdev(cq->device);
831	struct mlx5_ib_cq *mcq = to_mcq(cq);
832	struct ib_ucontext *context = NULL;
833
834	if (cq->uobject)
835		context = cq->uobject->context;
836
837	mlx5_core_destroy_cq(dev->mdev, &mcq->mcq);
838	if (context)
839		destroy_cq_user(mcq, context);
840	else
841		destroy_cq_kernel(dev, mcq);
842
843	kfree(mcq);
844
845	return 0;
846}
847
848static int is_equal_rsn(struct mlx5_cqe64 *cqe64, u32 rsn)
849{
850	return rsn == (ntohl(cqe64->sop_drop_qpn) & 0xffffff);
851}
852
853void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 rsn, struct mlx5_ib_srq *srq)
854{
855	struct mlx5_cqe64 *cqe64, *dest64;
856	void *cqe, *dest;
857	u32 prod_index;
858	int nfreed = 0;
859	u8 owner_bit;
860
861	if (!cq)
862		return;
863
864	/* First we need to find the current producer index, so we
865	 * know where to start cleaning from.  It doesn't matter if HW
866	 * adds new entries after this loop -- the QP we're worried
867	 * about is already in RESET, so the new entries won't come
868	 * from our QP and therefore don't need to be checked.
869	 */
870	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); prod_index++)
871		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
872			break;
873
874	/* Now sweep backwards through the CQ, removing CQ entries
875	 * that match our QP by copying older entries on top of them.
876	 */
877	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
878		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
879		cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
880		if (is_equal_rsn(cqe64, rsn)) {
881			if (srq && (ntohl(cqe64->srqn) & 0xffffff))
882				mlx5_ib_free_srq_wqe(srq, be16_to_cpu(cqe64->wqe_counter));
883			++nfreed;
884		} else if (nfreed) {
885			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
886			dest64 = (cq->mcq.cqe_sz == 64) ? dest : dest + 64;
887			owner_bit = dest64->op_own & MLX5_CQE_OWNER_MASK;
888			memcpy(dest, cqe, cq->mcq.cqe_sz);
889			dest64->op_own = owner_bit |
890				(dest64->op_own & ~MLX5_CQE_OWNER_MASK);
891		}
892	}
893
894	if (nfreed) {
895		cq->mcq.cons_index += nfreed;
896		/* Make sure update of buffer contents is done before
897		 * updating consumer index.
898		 */
899		wmb();
900		mlx5_cq_set_ci(&cq->mcq);
901	}
902}
903
904void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq)
905{
906	if (!cq)
907		return;
908
909	spin_lock_irq(&cq->lock);
910	__mlx5_ib_cq_clean(cq, qpn, srq);
911	spin_unlock_irq(&cq->lock);
912}
913
914int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
915{
916	struct mlx5_modify_cq_mbox_in *in;
917	struct mlx5_ib_dev *dev = to_mdev(cq->device);
918	struct mlx5_ib_cq *mcq = to_mcq(cq);
919	int err;
920	u32 fsel;
921
922	if (!(dev->mdev->caps.gen.flags & MLX5_DEV_CAP_FLAG_CQ_MODER))
923		return -ENOSYS;
924
925	in = kzalloc(sizeof(*in), GFP_KERNEL);
926	if (!in)
927		return -ENOMEM;
928
929	in->cqn = cpu_to_be32(mcq->mcq.cqn);
930	fsel = (MLX5_CQ_MODIFY_PERIOD | MLX5_CQ_MODIFY_COUNT);
931	in->ctx.cq_period = cpu_to_be16(cq_period);
932	in->ctx.cq_max_count = cpu_to_be16(cq_count);
933	in->field_select = cpu_to_be32(fsel);
934	err = mlx5_core_modify_cq(dev->mdev, &mcq->mcq, in, sizeof(*in));
935	kfree(in);
936
937	if (err)
938		mlx5_ib_warn(dev, "modify cq 0x%x failed\n", mcq->mcq.cqn);
939
940	return err;
941}
942
943static int resize_user(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
944		       int entries, struct ib_udata *udata, int *npas,
945		       int *page_shift, int *cqe_size)
946{
947	struct mlx5_ib_resize_cq ucmd;
948	struct ib_umem *umem;
949	int err;
950	int npages;
951	struct ib_ucontext *context = cq->buf.umem->context;
952
953	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
954	if (err)
955		return err;
956
957	if (ucmd.reserved0 || ucmd.reserved1)
958		return -EINVAL;
959
960	umem = ib_umem_get(context, ucmd.buf_addr, entries * ucmd.cqe_size,
961			   IB_ACCESS_LOCAL_WRITE, 1);
962	if (IS_ERR(umem)) {
963		err = PTR_ERR(umem);
964		return err;
965	}
966
967	mlx5_ib_cont_pages(umem, ucmd.buf_addr, &npages, page_shift,
968			   npas, NULL);
969
970	cq->resize_umem = umem;
971	*cqe_size = ucmd.cqe_size;
972
973	return 0;
974}
975
976static void un_resize_user(struct mlx5_ib_cq *cq)
977{
978	ib_umem_release(cq->resize_umem);
979}
980
981static int resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
982			 int entries, int cqe_size)
983{
984	int err;
985
986	cq->resize_buf = kzalloc(sizeof(*cq->resize_buf), GFP_KERNEL);
987	if (!cq->resize_buf)
988		return -ENOMEM;
989
990	err = alloc_cq_buf(dev, cq->resize_buf, entries, cqe_size);
991	if (err)
992		goto ex;
993
994	init_cq_buf(cq, cq->resize_buf);
995
996	return 0;
997
998ex:
999	kfree(cq->resize_buf);
1000	return err;
1001}
1002
1003static void un_resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
1004{
1005	free_cq_buf(dev, cq->resize_buf);
1006	cq->resize_buf = NULL;
1007}
1008
1009static int copy_resize_cqes(struct mlx5_ib_cq *cq)
1010{
1011	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
1012	struct mlx5_cqe64 *scqe64;
1013	struct mlx5_cqe64 *dcqe64;
1014	void *start_cqe;
1015	void *scqe;
1016	void *dcqe;
1017	int ssize;
1018	int dsize;
1019	int i;
1020	u8 sw_own;
1021
1022	ssize = cq->buf.cqe_size;
1023	dsize = cq->resize_buf->cqe_size;
1024	if (ssize != dsize) {
1025		mlx5_ib_warn(dev, "resize from different cqe size is not supported\n");
1026		return -EINVAL;
1027	}
1028
1029	i = cq->mcq.cons_index;
1030	scqe = get_sw_cqe(cq, i);
1031	scqe64 = ssize == 64 ? scqe : scqe + 64;
1032	start_cqe = scqe;
1033	if (!scqe) {
1034		mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1035		return -EINVAL;
1036	}
1037
1038	while ((scqe64->op_own >> 4) != MLX5_CQE_RESIZE_CQ) {
1039		dcqe = get_cqe_from_buf(cq->resize_buf,
1040					(i + 1) & (cq->resize_buf->nent),
1041					dsize);
1042		dcqe64 = dsize == 64 ? dcqe : dcqe + 64;
1043		sw_own = sw_ownership_bit(i + 1, cq->resize_buf->nent);
1044		memcpy(dcqe, scqe, dsize);
1045		dcqe64->op_own = (dcqe64->op_own & ~MLX5_CQE_OWNER_MASK) | sw_own;
1046
1047		++i;
1048		scqe = get_sw_cqe(cq, i);
1049		scqe64 = ssize == 64 ? scqe : scqe + 64;
1050		if (!scqe) {
1051			mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1052			return -EINVAL;
1053		}
1054
1055		if (scqe == start_cqe) {
1056			pr_warn("resize CQ failed to get resize CQE, CQN 0x%x\n",
1057				cq->mcq.cqn);
1058			return -ENOMEM;
1059		}
1060	}
1061	++cq->mcq.cons_index;
1062	return 0;
1063}
1064
1065int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
1066{
1067	struct mlx5_ib_dev *dev = to_mdev(ibcq->device);
1068	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1069	struct mlx5_modify_cq_mbox_in *in;
1070	int err;
1071	int npas;
1072	int page_shift;
1073	int inlen;
1074	int uninitialized_var(cqe_size);
1075	unsigned long flags;
1076
1077	if (!(dev->mdev->caps.gen.flags & MLX5_DEV_CAP_FLAG_RESIZE_CQ)) {
1078		pr_info("Firmware does not support resize CQ\n");
1079		return -ENOSYS;
1080	}
1081
1082	if (entries < 1)
1083		return -EINVAL;
1084
1085	entries = roundup_pow_of_two(entries + 1);
1086	if (entries > dev->mdev->caps.gen.max_cqes + 1)
1087		return -EINVAL;
1088
1089	if (entries == ibcq->cqe + 1)
1090		return 0;
1091
1092	mutex_lock(&cq->resize_mutex);
1093	if (udata) {
1094		err = resize_user(dev, cq, entries, udata, &npas, &page_shift,
1095				  &cqe_size);
1096	} else {
1097		cqe_size = 64;
1098		err = resize_kernel(dev, cq, entries, cqe_size);
1099		if (!err) {
1100			npas = cq->resize_buf->buf.npages;
1101			page_shift = cq->resize_buf->buf.page_shift;
1102		}
1103	}
1104
1105	if (err)
1106		goto ex;
1107
1108	inlen = sizeof(*in) + npas * sizeof(in->pas[0]);
1109	in = mlx5_vzalloc(inlen);
1110	if (!in) {
1111		err = -ENOMEM;
1112		goto ex_resize;
1113	}
1114
1115	if (udata)
1116		mlx5_ib_populate_pas(dev, cq->resize_umem, page_shift,
1117				     in->pas, 0);
1118	else
1119		mlx5_fill_page_array(&cq->resize_buf->buf, in->pas);
1120
1121	in->field_select = cpu_to_be32(MLX5_MODIFY_CQ_MASK_LOG_SIZE  |
1122				       MLX5_MODIFY_CQ_MASK_PG_OFFSET |
1123				       MLX5_MODIFY_CQ_MASK_PG_SIZE);
1124	in->ctx.log_pg_sz = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
1125	in->ctx.cqe_sz_flags = cqe_sz_to_mlx_sz(cqe_size) << 5;
1126	in->ctx.page_offset = 0;
1127	in->ctx.log_sz_usr_page = cpu_to_be32(ilog2(entries) << 24);
1128	in->hdr.opmod = cpu_to_be16(MLX5_CQ_OPMOD_RESIZE);
1129	in->cqn = cpu_to_be32(cq->mcq.cqn);
1130
1131	err = mlx5_core_modify_cq(dev->mdev, &cq->mcq, in, inlen);
1132	if (err)
1133		goto ex_alloc;
1134
1135	if (udata) {
1136		cq->ibcq.cqe = entries - 1;
1137		ib_umem_release(cq->buf.umem);
1138		cq->buf.umem = cq->resize_umem;
1139		cq->resize_umem = NULL;
1140	} else {
1141		struct mlx5_ib_cq_buf tbuf;
1142		int resized = 0;
1143
1144		spin_lock_irqsave(&cq->lock, flags);
1145		if (cq->resize_buf) {
1146			err = copy_resize_cqes(cq);
1147			if (!err) {
1148				tbuf = cq->buf;
1149				cq->buf = *cq->resize_buf;
1150				kfree(cq->resize_buf);
1151				cq->resize_buf = NULL;
1152				resized = 1;
1153			}
1154		}
1155		cq->ibcq.cqe = entries - 1;
1156		spin_unlock_irqrestore(&cq->lock, flags);
1157		if (resized)
1158			free_cq_buf(dev, &tbuf);
1159	}
1160	mutex_unlock(&cq->resize_mutex);
1161
1162	mlx5_vfree(in);
1163	return 0;
1164
1165ex_alloc:
1166	mlx5_vfree(in);
1167
1168ex_resize:
1169	if (udata)
1170		un_resize_user(cq);
1171	else
1172		un_resize_kernel(dev, cq);
1173ex:
1174	mutex_unlock(&cq->resize_mutex);
1175	return err;
1176}
1177
1178int mlx5_ib_get_cqe_size(struct mlx5_ib_dev *dev, struct ib_cq *ibcq)
1179{
1180	struct mlx5_ib_cq *cq;
1181
1182	if (!ibcq)
1183		return 128;
1184
1185	cq = to_mcq(ibcq);
1186	return cq->cqe_size;
1187}
1188