xref: /linux/drivers/infiniband/hw/mthca/mthca_cq.c (revision 9ce7677cfd7cd871adb457c80bea3b581b839641)
1 /*
2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4  * Copyright (c) 2005 Cisco Systems, Inc. All rights reserved.
5  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
6  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
7  *
8  * This software is available to you under a choice of one of two
9  * licenses.  You may choose to be licensed under the terms of the GNU
10  * General Public License (GPL) Version 2, available from the file
11  * COPYING in the main directory of this source tree, or the
12  * OpenIB.org BSD license below:
13  *
14  *     Redistribution and use in source and binary forms, with or
15  *     without modification, are permitted provided that the following
16  *     conditions are met:
17  *
18  *      - Redistributions of source code must retain the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer.
21  *
22  *      - Redistributions in binary form must reproduce the above
23  *        copyright notice, this list of conditions and the following
24  *        disclaimer in the documentation and/or other materials
25  *        provided with the distribution.
26  *
27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34  * SOFTWARE.
35  *
36  * $Id: mthca_cq.c 1369 2004-12-20 16:17:07Z roland $
37  */
38 
39 #include <linux/init.h>
40 #include <linux/hardirq.h>
41 
42 #include <rdma/ib_pack.h>
43 
44 #include "mthca_dev.h"
45 #include "mthca_cmd.h"
46 #include "mthca_memfree.h"
47 
48 enum {
49 	MTHCA_MAX_DIRECT_CQ_SIZE = 4 * PAGE_SIZE
50 };
51 
52 enum {
53 	MTHCA_CQ_ENTRY_SIZE = 0x20
54 };
55 
56 /*
57  * Must be packed because start is 64 bits but only aligned to 32 bits.
58  */
59 struct mthca_cq_context {
60 	__be32 flags;
61 	__be64 start;
62 	__be32 logsize_usrpage;
63 	__be32 error_eqn;	/* Tavor only */
64 	__be32 comp_eqn;
65 	__be32 pd;
66 	__be32 lkey;
67 	__be32 last_notified_index;
68 	__be32 solicit_producer_index;
69 	__be32 consumer_index;
70 	__be32 producer_index;
71 	__be32 cqn;
72 	__be32 ci_db;		/* Arbel only */
73 	__be32 state_db;	/* Arbel only */
74 	u32    reserved;
75 } __attribute__((packed));
76 
77 #define MTHCA_CQ_STATUS_OK          ( 0 << 28)
78 #define MTHCA_CQ_STATUS_OVERFLOW    ( 9 << 28)
79 #define MTHCA_CQ_STATUS_WRITE_FAIL  (10 << 28)
80 #define MTHCA_CQ_FLAG_TR            ( 1 << 18)
81 #define MTHCA_CQ_FLAG_OI            ( 1 << 17)
82 #define MTHCA_CQ_STATE_DISARMED     ( 0 <<  8)
83 #define MTHCA_CQ_STATE_ARMED        ( 1 <<  8)
84 #define MTHCA_CQ_STATE_ARMED_SOL    ( 4 <<  8)
85 #define MTHCA_EQ_STATE_FIRED        (10 <<  8)
86 
87 enum {
88 	MTHCA_ERROR_CQE_OPCODE_MASK = 0xfe
89 };
90 
91 enum {
92 	SYNDROME_LOCAL_LENGTH_ERR 	 = 0x01,
93 	SYNDROME_LOCAL_QP_OP_ERR  	 = 0x02,
94 	SYNDROME_LOCAL_EEC_OP_ERR 	 = 0x03,
95 	SYNDROME_LOCAL_PROT_ERR   	 = 0x04,
96 	SYNDROME_WR_FLUSH_ERR     	 = 0x05,
97 	SYNDROME_MW_BIND_ERR      	 = 0x06,
98 	SYNDROME_BAD_RESP_ERR     	 = 0x10,
99 	SYNDROME_LOCAL_ACCESS_ERR 	 = 0x11,
100 	SYNDROME_REMOTE_INVAL_REQ_ERR 	 = 0x12,
101 	SYNDROME_REMOTE_ACCESS_ERR 	 = 0x13,
102 	SYNDROME_REMOTE_OP_ERR     	 = 0x14,
103 	SYNDROME_RETRY_EXC_ERR 		 = 0x15,
104 	SYNDROME_RNR_RETRY_EXC_ERR 	 = 0x16,
105 	SYNDROME_LOCAL_RDD_VIOL_ERR 	 = 0x20,
106 	SYNDROME_REMOTE_INVAL_RD_REQ_ERR = 0x21,
107 	SYNDROME_REMOTE_ABORTED_ERR 	 = 0x22,
108 	SYNDROME_INVAL_EECN_ERR 	 = 0x23,
109 	SYNDROME_INVAL_EEC_STATE_ERR 	 = 0x24
110 };
111 
112 struct mthca_cqe {
113 	__be32 my_qpn;
114 	__be32 my_ee;
115 	__be32 rqpn;
116 	__be16 sl_g_mlpath;
117 	__be16 rlid;
118 	__be32 imm_etype_pkey_eec;
119 	__be32 byte_cnt;
120 	__be32 wqe;
121 	u8     opcode;
122 	u8     is_send;
123 	u8     reserved;
124 	u8     owner;
125 };
126 
127 struct mthca_err_cqe {
128 	__be32 my_qpn;
129 	u32    reserved1[3];
130 	u8     syndrome;
131 	u8     reserved2;
132 	__be16 db_cnt;
133 	u32    reserved3;
134 	__be32 wqe;
135 	u8     opcode;
136 	u8     reserved4[2];
137 	u8     owner;
138 };
139 
140 #define MTHCA_CQ_ENTRY_OWNER_SW      (0 << 7)
141 #define MTHCA_CQ_ENTRY_OWNER_HW      (1 << 7)
142 
143 #define MTHCA_TAVOR_CQ_DB_INC_CI       (1 << 24)
144 #define MTHCA_TAVOR_CQ_DB_REQ_NOT      (2 << 24)
145 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL  (3 << 24)
146 #define MTHCA_TAVOR_CQ_DB_SET_CI       (4 << 24)
147 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_MULT (5 << 24)
148 
149 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL  (1 << 24)
150 #define MTHCA_ARBEL_CQ_DB_REQ_NOT      (2 << 24)
151 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_MULT (3 << 24)
152 
153 static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
154 {
155 	if (cq->is_direct)
156 		return cq->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
157 	else
158 		return cq->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
159 			+ (entry * MTHCA_CQ_ENTRY_SIZE) % PAGE_SIZE;
160 }
161 
162 static inline struct mthca_cqe *cqe_sw(struct mthca_cq *cq, int i)
163 {
164 	struct mthca_cqe *cqe = get_cqe(cq, i);
165 	return MTHCA_CQ_ENTRY_OWNER_HW & cqe->owner ? NULL : cqe;
166 }
167 
168 static inline struct mthca_cqe *next_cqe_sw(struct mthca_cq *cq)
169 {
170 	return cqe_sw(cq, cq->cons_index & cq->ibcq.cqe);
171 }
172 
173 static inline void set_cqe_hw(struct mthca_cqe *cqe)
174 {
175 	cqe->owner = MTHCA_CQ_ENTRY_OWNER_HW;
176 }
177 
178 static void dump_cqe(struct mthca_dev *dev, void *cqe_ptr)
179 {
180 	__be32 *cqe = cqe_ptr;
181 
182 	(void) cqe;	/* avoid warning if mthca_dbg compiled away... */
183 	mthca_dbg(dev, "CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
184 		  be32_to_cpu(cqe[0]), be32_to_cpu(cqe[1]), be32_to_cpu(cqe[2]),
185 		  be32_to_cpu(cqe[3]), be32_to_cpu(cqe[4]), be32_to_cpu(cqe[5]),
186 		  be32_to_cpu(cqe[6]), be32_to_cpu(cqe[7]));
187 }
188 
189 /*
190  * incr is ignored in native Arbel (mem-free) mode, so cq->cons_index
191  * should be correct before calling update_cons_index().
192  */
193 static inline void update_cons_index(struct mthca_dev *dev, struct mthca_cq *cq,
194 				     int incr)
195 {
196 	__be32 doorbell[2];
197 
198 	if (mthca_is_memfree(dev)) {
199 		*cq->set_ci_db = cpu_to_be32(cq->cons_index);
200 		wmb();
201 	} else {
202 		doorbell[0] = cpu_to_be32(MTHCA_TAVOR_CQ_DB_INC_CI | cq->cqn);
203 		doorbell[1] = cpu_to_be32(incr - 1);
204 
205 		mthca_write64(doorbell,
206 			      dev->kar + MTHCA_CQ_DOORBELL,
207 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
208 	}
209 }
210 
211 void mthca_cq_completion(struct mthca_dev *dev, u32 cqn)
212 {
213 	struct mthca_cq *cq;
214 
215 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
216 
217 	if (!cq) {
218 		mthca_warn(dev, "Completion event for bogus CQ %08x\n", cqn);
219 		return;
220 	}
221 
222 	++cq->arm_sn;
223 
224 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
225 }
226 
227 void mthca_cq_event(struct mthca_dev *dev, u32 cqn,
228 		    enum ib_event_type event_type)
229 {
230 	struct mthca_cq *cq;
231 	struct ib_event event;
232 
233 	spin_lock(&dev->cq_table.lock);
234 
235 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
236 
237 	if (cq)
238 		atomic_inc(&cq->refcount);
239 	spin_unlock(&dev->cq_table.lock);
240 
241 	if (!cq) {
242 		mthca_warn(dev, "Async event for bogus CQ %08x\n", cqn);
243 		return;
244 	}
245 
246 	event.device      = &dev->ib_dev;
247 	event.event       = event_type;
248 	event.element.cq  = &cq->ibcq;
249 	if (cq->ibcq.event_handler)
250 		cq->ibcq.event_handler(&event, cq->ibcq.cq_context);
251 
252 	if (atomic_dec_and_test(&cq->refcount))
253 		wake_up(&cq->wait);
254 }
255 
256 void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn,
257 		    struct mthca_srq *srq)
258 {
259 	struct mthca_cq *cq;
260 	struct mthca_cqe *cqe;
261 	u32 prod_index;
262 	int nfreed = 0;
263 
264 	spin_lock_irq(&dev->cq_table.lock);
265 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
266 	if (cq)
267 		atomic_inc(&cq->refcount);
268 	spin_unlock_irq(&dev->cq_table.lock);
269 
270 	if (!cq)
271 		return;
272 
273 	spin_lock_irq(&cq->lock);
274 
275 	/*
276 	 * First we need to find the current producer index, so we
277 	 * know where to start cleaning from.  It doesn't matter if HW
278 	 * adds new entries after this loop -- the QP we're worried
279 	 * about is already in RESET, so the new entries won't come
280 	 * from our QP and therefore don't need to be checked.
281 	 */
282 	for (prod_index = cq->cons_index;
283 	     cqe_sw(cq, prod_index & cq->ibcq.cqe);
284 	     ++prod_index)
285 		if (prod_index == cq->cons_index + cq->ibcq.cqe)
286 			break;
287 
288 	if (0)
289 		mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
290 			  qpn, cqn, cq->cons_index, prod_index);
291 
292 	/*
293 	 * Now sweep backwards through the CQ, removing CQ entries
294 	 * that match our QP by copying older entries on top of them.
295 	 */
296 	while ((int) --prod_index - (int) cq->cons_index >= 0) {
297 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
298 		if (cqe->my_qpn == cpu_to_be32(qpn)) {
299 			if (srq)
300 				mthca_free_srq_wqe(srq, be32_to_cpu(cqe->wqe));
301 			++nfreed;
302 		} else if (nfreed)
303 			memcpy(get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe),
304 			       cqe, MTHCA_CQ_ENTRY_SIZE);
305 	}
306 
307 	if (nfreed) {
308 		wmb();
309 		cq->cons_index += nfreed;
310 		update_cons_index(dev, cq, nfreed);
311 	}
312 
313 	spin_unlock_irq(&cq->lock);
314 	if (atomic_dec_and_test(&cq->refcount))
315 		wake_up(&cq->wait);
316 }
317 
318 static int handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
319 			    struct mthca_qp *qp, int wqe_index, int is_send,
320 			    struct mthca_err_cqe *cqe,
321 			    struct ib_wc *entry, int *free_cqe)
322 {
323 	int err;
324 	int dbd;
325 	__be32 new_wqe;
326 
327 	if (cqe->syndrome == SYNDROME_LOCAL_QP_OP_ERR) {
328 		mthca_dbg(dev, "local QP operation err "
329 			  "(QPN %06x, WQE @ %08x, CQN %06x, index %d)\n",
330 			  be32_to_cpu(cqe->my_qpn), be32_to_cpu(cqe->wqe),
331 			  cq->cqn, cq->cons_index);
332 		dump_cqe(dev, cqe);
333 	}
334 
335 	/*
336 	 * For completions in error, only work request ID, status (and
337 	 * freed resource count for RD) have to be set.
338 	 */
339 	switch (cqe->syndrome) {
340 	case SYNDROME_LOCAL_LENGTH_ERR:
341 		entry->status = IB_WC_LOC_LEN_ERR;
342 		break;
343 	case SYNDROME_LOCAL_QP_OP_ERR:
344 		entry->status = IB_WC_LOC_QP_OP_ERR;
345 		break;
346 	case SYNDROME_LOCAL_EEC_OP_ERR:
347 		entry->status = IB_WC_LOC_EEC_OP_ERR;
348 		break;
349 	case SYNDROME_LOCAL_PROT_ERR:
350 		entry->status = IB_WC_LOC_PROT_ERR;
351 		break;
352 	case SYNDROME_WR_FLUSH_ERR:
353 		entry->status = IB_WC_WR_FLUSH_ERR;
354 		break;
355 	case SYNDROME_MW_BIND_ERR:
356 		entry->status = IB_WC_MW_BIND_ERR;
357 		break;
358 	case SYNDROME_BAD_RESP_ERR:
359 		entry->status = IB_WC_BAD_RESP_ERR;
360 		break;
361 	case SYNDROME_LOCAL_ACCESS_ERR:
362 		entry->status = IB_WC_LOC_ACCESS_ERR;
363 		break;
364 	case SYNDROME_REMOTE_INVAL_REQ_ERR:
365 		entry->status = IB_WC_REM_INV_REQ_ERR;
366 		break;
367 	case SYNDROME_REMOTE_ACCESS_ERR:
368 		entry->status = IB_WC_REM_ACCESS_ERR;
369 		break;
370 	case SYNDROME_REMOTE_OP_ERR:
371 		entry->status = IB_WC_REM_OP_ERR;
372 		break;
373 	case SYNDROME_RETRY_EXC_ERR:
374 		entry->status = IB_WC_RETRY_EXC_ERR;
375 		break;
376 	case SYNDROME_RNR_RETRY_EXC_ERR:
377 		entry->status = IB_WC_RNR_RETRY_EXC_ERR;
378 		break;
379 	case SYNDROME_LOCAL_RDD_VIOL_ERR:
380 		entry->status = IB_WC_LOC_RDD_VIOL_ERR;
381 		break;
382 	case SYNDROME_REMOTE_INVAL_RD_REQ_ERR:
383 		entry->status = IB_WC_REM_INV_RD_REQ_ERR;
384 		break;
385 	case SYNDROME_REMOTE_ABORTED_ERR:
386 		entry->status = IB_WC_REM_ABORT_ERR;
387 		break;
388 	case SYNDROME_INVAL_EECN_ERR:
389 		entry->status = IB_WC_INV_EECN_ERR;
390 		break;
391 	case SYNDROME_INVAL_EEC_STATE_ERR:
392 		entry->status = IB_WC_INV_EEC_STATE_ERR;
393 		break;
394 	default:
395 		entry->status = IB_WC_GENERAL_ERR;
396 		break;
397 	}
398 
399 	/*
400 	 * Mem-free HCAs always generate one CQE per WQE, even in the
401 	 * error case, so we don't have to check the doorbell count, etc.
402 	 */
403 	if (mthca_is_memfree(dev))
404 		return 0;
405 
406 	err = mthca_free_err_wqe(dev, qp, is_send, wqe_index, &dbd, &new_wqe);
407 	if (err)
408 		return err;
409 
410 	/*
411 	 * If we're at the end of the WQE chain, or we've used up our
412 	 * doorbell count, free the CQE.  Otherwise just update it for
413 	 * the next poll operation.
414 	 */
415 	if (!(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
416 		return 0;
417 
418 	cqe->db_cnt   = cpu_to_be16(be16_to_cpu(cqe->db_cnt) - dbd);
419 	cqe->wqe      = new_wqe;
420 	cqe->syndrome = SYNDROME_WR_FLUSH_ERR;
421 
422 	*free_cqe = 0;
423 
424 	return 0;
425 }
426 
427 static inline int mthca_poll_one(struct mthca_dev *dev,
428 				 struct mthca_cq *cq,
429 				 struct mthca_qp **cur_qp,
430 				 int *freed,
431 				 struct ib_wc *entry)
432 {
433 	struct mthca_wq *wq;
434 	struct mthca_cqe *cqe;
435 	int wqe_index;
436 	int is_error;
437 	int is_send;
438 	int free_cqe = 1;
439 	int err = 0;
440 
441 	cqe = next_cqe_sw(cq);
442 	if (!cqe)
443 		return -EAGAIN;
444 
445 	/*
446 	 * Make sure we read CQ entry contents after we've checked the
447 	 * ownership bit.
448 	 */
449 	rmb();
450 
451 	if (0) {
452 		mthca_dbg(dev, "%x/%d: CQE -> QPN %06x, WQE @ %08x\n",
453 			  cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
454 			  be32_to_cpu(cqe->wqe));
455 		dump_cqe(dev, cqe);
456 	}
457 
458 	is_error = (cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
459 		MTHCA_ERROR_CQE_OPCODE_MASK;
460 	is_send  = is_error ? cqe->opcode & 0x01 : cqe->is_send & 0x80;
461 
462 	if (!*cur_qp || be32_to_cpu(cqe->my_qpn) != (*cur_qp)->qpn) {
463 		/*
464 		 * We do not have to take the QP table lock here,
465 		 * because CQs will be locked while QPs are removed
466 		 * from the table.
467 		 */
468 		*cur_qp = mthca_array_get(&dev->qp_table.qp,
469 					  be32_to_cpu(cqe->my_qpn) &
470 					  (dev->limits.num_qps - 1));
471 		if (!*cur_qp) {
472 			mthca_warn(dev, "CQ entry for unknown QP %06x\n",
473 				   be32_to_cpu(cqe->my_qpn) & 0xffffff);
474 			err = -EINVAL;
475 			goto out;
476 		}
477 	}
478 
479 	entry->qp_num = (*cur_qp)->qpn;
480 
481 	if (is_send) {
482 		wq = &(*cur_qp)->sq;
483 		wqe_index = ((be32_to_cpu(cqe->wqe) - (*cur_qp)->send_wqe_offset)
484 			     >> wq->wqe_shift);
485 		entry->wr_id = (*cur_qp)->wrid[wqe_index +
486 					       (*cur_qp)->rq.max];
487 	} else if ((*cur_qp)->ibqp.srq) {
488 		struct mthca_srq *srq = to_msrq((*cur_qp)->ibqp.srq);
489 		u32 wqe = be32_to_cpu(cqe->wqe);
490 		wq = NULL;
491 		wqe_index = wqe >> srq->wqe_shift;
492 		entry->wr_id = srq->wrid[wqe_index];
493 		mthca_free_srq_wqe(srq, wqe);
494 	} else {
495 		wq = &(*cur_qp)->rq;
496 		wqe_index = be32_to_cpu(cqe->wqe) >> wq->wqe_shift;
497 		entry->wr_id = (*cur_qp)->wrid[wqe_index];
498 	}
499 
500 	if (wq) {
501 		if (wq->last_comp < wqe_index)
502 			wq->tail += wqe_index - wq->last_comp;
503 		else
504 			wq->tail += wqe_index + wq->max - wq->last_comp;
505 
506 		wq->last_comp = wqe_index;
507 	}
508 
509 	if (is_error) {
510 		err = handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,
511 				       (struct mthca_err_cqe *) cqe,
512 				       entry, &free_cqe);
513 		goto out;
514 	}
515 
516 	if (is_send) {
517 		entry->wc_flags = 0;
518 		switch (cqe->opcode) {
519 		case MTHCA_OPCODE_RDMA_WRITE:
520 			entry->opcode    = IB_WC_RDMA_WRITE;
521 			break;
522 		case MTHCA_OPCODE_RDMA_WRITE_IMM:
523 			entry->opcode    = IB_WC_RDMA_WRITE;
524 			entry->wc_flags |= IB_WC_WITH_IMM;
525 			break;
526 		case MTHCA_OPCODE_SEND:
527 			entry->opcode    = IB_WC_SEND;
528 			break;
529 		case MTHCA_OPCODE_SEND_IMM:
530 			entry->opcode    = IB_WC_SEND;
531 			entry->wc_flags |= IB_WC_WITH_IMM;
532 			break;
533 		case MTHCA_OPCODE_RDMA_READ:
534 			entry->opcode    = IB_WC_RDMA_READ;
535 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
536 			break;
537 		case MTHCA_OPCODE_ATOMIC_CS:
538 			entry->opcode    = IB_WC_COMP_SWAP;
539 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
540 			break;
541 		case MTHCA_OPCODE_ATOMIC_FA:
542 			entry->opcode    = IB_WC_FETCH_ADD;
543 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
544 			break;
545 		case MTHCA_OPCODE_BIND_MW:
546 			entry->opcode    = IB_WC_BIND_MW;
547 			break;
548 		default:
549 			entry->opcode    = MTHCA_OPCODE_INVALID;
550 			break;
551 		}
552 	} else {
553 		entry->byte_len = be32_to_cpu(cqe->byte_cnt);
554 		switch (cqe->opcode & 0x1f) {
555 		case IB_OPCODE_SEND_LAST_WITH_IMMEDIATE:
556 		case IB_OPCODE_SEND_ONLY_WITH_IMMEDIATE:
557 			entry->wc_flags = IB_WC_WITH_IMM;
558 			entry->imm_data = cqe->imm_etype_pkey_eec;
559 			entry->opcode = IB_WC_RECV;
560 			break;
561 		case IB_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
562 		case IB_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
563 			entry->wc_flags = IB_WC_WITH_IMM;
564 			entry->imm_data = cqe->imm_etype_pkey_eec;
565 			entry->opcode = IB_WC_RECV_RDMA_WITH_IMM;
566 			break;
567 		default:
568 			entry->wc_flags = 0;
569 			entry->opcode = IB_WC_RECV;
570 			break;
571 		}
572 		entry->slid 	   = be16_to_cpu(cqe->rlid);
573 		entry->sl   	   = be16_to_cpu(cqe->sl_g_mlpath) >> 12;
574 		entry->src_qp 	   = be32_to_cpu(cqe->rqpn) & 0xffffff;
575 		entry->dlid_path_bits = be16_to_cpu(cqe->sl_g_mlpath) & 0x7f;
576 		entry->pkey_index  = be32_to_cpu(cqe->imm_etype_pkey_eec) >> 16;
577 		entry->wc_flags   |= be16_to_cpu(cqe->sl_g_mlpath) & 0x80 ?
578 					IB_WC_GRH : 0;
579 	}
580 
581 	entry->status = IB_WC_SUCCESS;
582 
583  out:
584 	if (likely(free_cqe)) {
585 		set_cqe_hw(cqe);
586 		++(*freed);
587 		++cq->cons_index;
588 	}
589 
590 	return err;
591 }
592 
593 int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
594 		  struct ib_wc *entry)
595 {
596 	struct mthca_dev *dev = to_mdev(ibcq->device);
597 	struct mthca_cq *cq = to_mcq(ibcq);
598 	struct mthca_qp *qp = NULL;
599 	unsigned long flags;
600 	int err = 0;
601 	int freed = 0;
602 	int npolled;
603 
604 	spin_lock_irqsave(&cq->lock, flags);
605 
606 	for (npolled = 0; npolled < num_entries; ++npolled) {
607 		err = mthca_poll_one(dev, cq, &qp,
608 				     &freed, entry + npolled);
609 		if (err)
610 			break;
611 	}
612 
613 	if (freed) {
614 		wmb();
615 		update_cons_index(dev, cq, freed);
616 	}
617 
618 	spin_unlock_irqrestore(&cq->lock, flags);
619 
620 	return err == 0 || err == -EAGAIN ? npolled : err;
621 }
622 
623 int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify notify)
624 {
625 	__be32 doorbell[2];
626 
627 	doorbell[0] = cpu_to_be32((notify == IB_CQ_SOLICITED ?
628 				   MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL :
629 				   MTHCA_TAVOR_CQ_DB_REQ_NOT)      |
630 				  to_mcq(cq)->cqn);
631 	doorbell[1] = (__force __be32) 0xffffffff;
632 
633 	mthca_write64(doorbell,
634 		      to_mdev(cq->device)->kar + MTHCA_CQ_DOORBELL,
635 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(cq->device)->doorbell_lock));
636 
637 	return 0;
638 }
639 
640 int mthca_arbel_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify notify)
641 {
642 	struct mthca_cq *cq = to_mcq(ibcq);
643 	__be32 doorbell[2];
644 	u32 sn;
645 	__be32 ci;
646 
647 	sn = cq->arm_sn & 3;
648 	ci = cpu_to_be32(cq->cons_index);
649 
650 	doorbell[0] = ci;
651 	doorbell[1] = cpu_to_be32((cq->cqn << 8) | (2 << 5) | (sn << 3) |
652 				  (notify == IB_CQ_SOLICITED ? 1 : 2));
653 
654 	mthca_write_db_rec(doorbell, cq->arm_db);
655 
656 	/*
657 	 * Make sure that the doorbell record in host memory is
658 	 * written before ringing the doorbell via PCI MMIO.
659 	 */
660 	wmb();
661 
662 	doorbell[0] = cpu_to_be32((sn << 28)                       |
663 				  (notify == IB_CQ_SOLICITED ?
664 				   MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL :
665 				   MTHCA_ARBEL_CQ_DB_REQ_NOT)      |
666 				  cq->cqn);
667 	doorbell[1] = ci;
668 
669 	mthca_write64(doorbell,
670 		      to_mdev(ibcq->device)->kar + MTHCA_CQ_DOORBELL,
671 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->doorbell_lock));
672 
673 	return 0;
674 }
675 
676 static void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq *cq)
677 {
678 	mthca_buf_free(dev, (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE,
679 		       &cq->queue, cq->is_direct, &cq->mr);
680 }
681 
682 int mthca_init_cq(struct mthca_dev *dev, int nent,
683 		  struct mthca_ucontext *ctx, u32 pdn,
684 		  struct mthca_cq *cq)
685 {
686 	int size = nent * MTHCA_CQ_ENTRY_SIZE;
687 	struct mthca_mailbox *mailbox;
688 	struct mthca_cq_context *cq_context;
689 	int err = -ENOMEM;
690 	u8 status;
691 	int i;
692 
693 	might_sleep();
694 
695 	cq->ibcq.cqe  = nent - 1;
696 	cq->is_kernel = !ctx;
697 
698 	cq->cqn = mthca_alloc(&dev->cq_table.alloc);
699 	if (cq->cqn == -1)
700 		return -ENOMEM;
701 
702 	if (mthca_is_memfree(dev)) {
703 		err = mthca_table_get(dev, dev->cq_table.table, cq->cqn);
704 		if (err)
705 			goto err_out;
706 
707 		if (cq->is_kernel) {
708 			cq->arm_sn = 1;
709 
710 			err = -ENOMEM;
711 
712 			cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
713 							     cq->cqn, &cq->set_ci_db);
714 			if (cq->set_ci_db_index < 0)
715 				goto err_out_icm;
716 
717 			cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
718 							  cq->cqn, &cq->arm_db);
719 			if (cq->arm_db_index < 0)
720 				goto err_out_ci;
721 		}
722 	}
723 
724 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
725 	if (IS_ERR(mailbox))
726 		goto err_out_arm;
727 
728 	cq_context = mailbox->buf;
729 
730 	if (cq->is_kernel) {
731 		err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_CQ_SIZE,
732 				      &cq->queue, &cq->is_direct,
733 				      &dev->driver_pd, 1, &cq->mr);
734 		if (err)
735 			goto err_out_mailbox;
736 
737 		for (i = 0; i < nent; ++i)
738 			set_cqe_hw(get_cqe(cq, i));
739 	}
740 
741 	spin_lock_init(&cq->lock);
742 	atomic_set(&cq->refcount, 1);
743 	init_waitqueue_head(&cq->wait);
744 
745 	memset(cq_context, 0, sizeof *cq_context);
746 	cq_context->flags           = cpu_to_be32(MTHCA_CQ_STATUS_OK      |
747 						  MTHCA_CQ_STATE_DISARMED |
748 						  MTHCA_CQ_FLAG_TR);
749 	cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24);
750 	if (ctx)
751 		cq_context->logsize_usrpage |= cpu_to_be32(ctx->uar.index);
752 	else
753 		cq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
754 	cq_context->error_eqn       = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
755 	cq_context->comp_eqn        = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
756 	cq_context->pd              = cpu_to_be32(pdn);
757 	cq_context->lkey            = cpu_to_be32(cq->mr.ibmr.lkey);
758 	cq_context->cqn             = cpu_to_be32(cq->cqn);
759 
760 	if (mthca_is_memfree(dev)) {
761 		cq_context->ci_db    = cpu_to_be32(cq->set_ci_db_index);
762 		cq_context->state_db = cpu_to_be32(cq->arm_db_index);
763 	}
764 
765 	err = mthca_SW2HW_CQ(dev, mailbox, cq->cqn, &status);
766 	if (err) {
767 		mthca_warn(dev, "SW2HW_CQ failed (%d)\n", err);
768 		goto err_out_free_mr;
769 	}
770 
771 	if (status) {
772 		mthca_warn(dev, "SW2HW_CQ returned status 0x%02x\n",
773 			   status);
774 		err = -EINVAL;
775 		goto err_out_free_mr;
776 	}
777 
778 	spin_lock_irq(&dev->cq_table.lock);
779 	if (mthca_array_set(&dev->cq_table.cq,
780 			    cq->cqn & (dev->limits.num_cqs - 1),
781 			    cq)) {
782 		spin_unlock_irq(&dev->cq_table.lock);
783 		goto err_out_free_mr;
784 	}
785 	spin_unlock_irq(&dev->cq_table.lock);
786 
787 	cq->cons_index = 0;
788 
789 	mthca_free_mailbox(dev, mailbox);
790 
791 	return 0;
792 
793 err_out_free_mr:
794 	if (cq->is_kernel)
795 		mthca_free_cq_buf(dev, cq);
796 
797 err_out_mailbox:
798 	mthca_free_mailbox(dev, mailbox);
799 
800 err_out_arm:
801 	if (cq->is_kernel && mthca_is_memfree(dev))
802 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
803 
804 err_out_ci:
805 	if (cq->is_kernel && mthca_is_memfree(dev))
806 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
807 
808 err_out_icm:
809 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
810 
811 err_out:
812 	mthca_free(&dev->cq_table.alloc, cq->cqn);
813 
814 	return err;
815 }
816 
817 void mthca_free_cq(struct mthca_dev *dev,
818 		   struct mthca_cq *cq)
819 {
820 	struct mthca_mailbox *mailbox;
821 	int err;
822 	u8 status;
823 
824 	might_sleep();
825 
826 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
827 	if (IS_ERR(mailbox)) {
828 		mthca_warn(dev, "No memory for mailbox to free CQ.\n");
829 		return;
830 	}
831 
832 	err = mthca_HW2SW_CQ(dev, mailbox, cq->cqn, &status);
833 	if (err)
834 		mthca_warn(dev, "HW2SW_CQ failed (%d)\n", err);
835 	else if (status)
836 		mthca_warn(dev, "HW2SW_CQ returned status 0x%02x\n", status);
837 
838 	if (0) {
839 		__be32 *ctx = mailbox->buf;
840 		int j;
841 
842 		printk(KERN_ERR "context for CQN %x (cons index %x, next sw %d)\n",
843 		       cq->cqn, cq->cons_index,
844 		       cq->is_kernel ? !!next_cqe_sw(cq) : 0);
845 		for (j = 0; j < 16; ++j)
846 			printk(KERN_ERR "[%2x] %08x\n", j * 4, be32_to_cpu(ctx[j]));
847 	}
848 
849 	spin_lock_irq(&dev->cq_table.lock);
850 	mthca_array_clear(&dev->cq_table.cq,
851 			  cq->cqn & (dev->limits.num_cqs - 1));
852 	spin_unlock_irq(&dev->cq_table.lock);
853 
854 	if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
855 		synchronize_irq(dev->eq_table.eq[MTHCA_EQ_COMP].msi_x_vector);
856 	else
857 		synchronize_irq(dev->pdev->irq);
858 
859 	atomic_dec(&cq->refcount);
860 	wait_event(cq->wait, !atomic_read(&cq->refcount));
861 
862 	if (cq->is_kernel) {
863 		mthca_free_cq_buf(dev, cq);
864 		if (mthca_is_memfree(dev)) {
865 			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM,    cq->arm_db_index);
866 			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
867 		}
868 	}
869 
870 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
871 	mthca_free(&dev->cq_table.alloc, cq->cqn);
872 	mthca_free_mailbox(dev, mailbox);
873 }
874 
875 int __devinit mthca_init_cq_table(struct mthca_dev *dev)
876 {
877 	int err;
878 
879 	spin_lock_init(&dev->cq_table.lock);
880 
881 	err = mthca_alloc_init(&dev->cq_table.alloc,
882 			       dev->limits.num_cqs,
883 			       (1 << 24) - 1,
884 			       dev->limits.reserved_cqs);
885 	if (err)
886 		return err;
887 
888 	err = mthca_array_init(&dev->cq_table.cq,
889 			       dev->limits.num_cqs);
890 	if (err)
891 		mthca_alloc_cleanup(&dev->cq_table.alloc);
892 
893 	return err;
894 }
895 
896 void __devexit mthca_cleanup_cq_table(struct mthca_dev *dev)
897 {
898 	mthca_array_cleanup(&dev->cq_table.cq, dev->limits.num_cqs);
899 	mthca_alloc_cleanup(&dev->cq_table.alloc);
900 }
901