xref: /linux/drivers/infiniband/hw/mthca/mthca_cq.c (revision 776cfebb430c7b22c208b1b17add97f354d97cab)
1 /*
2  * Copyright (c) 2004, 2005 Topspin Communications.  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  * $Id: mthca_cq.c 1369 2004-12-20 16:17:07Z roland $
33  */
34 
35 #include <linux/init.h>
36 #include <linux/hardirq.h>
37 
38 #include <ib_pack.h>
39 
40 #include "mthca_dev.h"
41 #include "mthca_cmd.h"
42 #include "mthca_memfree.h"
43 
44 enum {
45 	MTHCA_MAX_DIRECT_CQ_SIZE = 4 * PAGE_SIZE
46 };
47 
48 enum {
49 	MTHCA_CQ_ENTRY_SIZE = 0x20
50 };
51 
52 /*
53  * Must be packed because start is 64 bits but only aligned to 32 bits.
54  */
55 struct mthca_cq_context {
56 	u32 flags;
57 	u64 start;
58 	u32 logsize_usrpage;
59 	u32 error_eqn;		/* Tavor only */
60 	u32 comp_eqn;
61 	u32 pd;
62 	u32 lkey;
63 	u32 last_notified_index;
64 	u32 solicit_producer_index;
65 	u32 consumer_index;
66 	u32 producer_index;
67 	u32 cqn;
68 	u32 ci_db;		/* Arbel only */
69 	u32 state_db;		/* Arbel only */
70 	u32 reserved;
71 } __attribute__((packed));
72 
73 #define MTHCA_CQ_STATUS_OK          ( 0 << 28)
74 #define MTHCA_CQ_STATUS_OVERFLOW    ( 9 << 28)
75 #define MTHCA_CQ_STATUS_WRITE_FAIL  (10 << 28)
76 #define MTHCA_CQ_FLAG_TR            ( 1 << 18)
77 #define MTHCA_CQ_FLAG_OI            ( 1 << 17)
78 #define MTHCA_CQ_STATE_DISARMED     ( 0 <<  8)
79 #define MTHCA_CQ_STATE_ARMED        ( 1 <<  8)
80 #define MTHCA_CQ_STATE_ARMED_SOL    ( 4 <<  8)
81 #define MTHCA_EQ_STATE_FIRED        (10 <<  8)
82 
83 enum {
84 	MTHCA_ERROR_CQE_OPCODE_MASK = 0xfe
85 };
86 
87 enum {
88 	SYNDROME_LOCAL_LENGTH_ERR 	 = 0x01,
89 	SYNDROME_LOCAL_QP_OP_ERR  	 = 0x02,
90 	SYNDROME_LOCAL_EEC_OP_ERR 	 = 0x03,
91 	SYNDROME_LOCAL_PROT_ERR   	 = 0x04,
92 	SYNDROME_WR_FLUSH_ERR     	 = 0x05,
93 	SYNDROME_MW_BIND_ERR      	 = 0x06,
94 	SYNDROME_BAD_RESP_ERR     	 = 0x10,
95 	SYNDROME_LOCAL_ACCESS_ERR 	 = 0x11,
96 	SYNDROME_REMOTE_INVAL_REQ_ERR 	 = 0x12,
97 	SYNDROME_REMOTE_ACCESS_ERR 	 = 0x13,
98 	SYNDROME_REMOTE_OP_ERR     	 = 0x14,
99 	SYNDROME_RETRY_EXC_ERR 		 = 0x15,
100 	SYNDROME_RNR_RETRY_EXC_ERR 	 = 0x16,
101 	SYNDROME_LOCAL_RDD_VIOL_ERR 	 = 0x20,
102 	SYNDROME_REMOTE_INVAL_RD_REQ_ERR = 0x21,
103 	SYNDROME_REMOTE_ABORTED_ERR 	 = 0x22,
104 	SYNDROME_INVAL_EECN_ERR 	 = 0x23,
105 	SYNDROME_INVAL_EEC_STATE_ERR 	 = 0x24
106 };
107 
108 struct mthca_cqe {
109 	u32 my_qpn;
110 	u32 my_ee;
111 	u32 rqpn;
112 	u16 sl_g_mlpath;
113 	u16 rlid;
114 	u32 imm_etype_pkey_eec;
115 	u32 byte_cnt;
116 	u32 wqe;
117 	u8  opcode;
118 	u8  is_send;
119 	u8  reserved;
120 	u8  owner;
121 };
122 
123 struct mthca_err_cqe {
124 	u32 my_qpn;
125 	u32 reserved1[3];
126 	u8  syndrome;
127 	u8  reserved2;
128 	u16 db_cnt;
129 	u32 reserved3;
130 	u32 wqe;
131 	u8  opcode;
132 	u8  reserved4[2];
133 	u8  owner;
134 };
135 
136 #define MTHCA_CQ_ENTRY_OWNER_SW      (0 << 7)
137 #define MTHCA_CQ_ENTRY_OWNER_HW      (1 << 7)
138 
139 #define MTHCA_TAVOR_CQ_DB_INC_CI       (1 << 24)
140 #define MTHCA_TAVOR_CQ_DB_REQ_NOT      (2 << 24)
141 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL  (3 << 24)
142 #define MTHCA_TAVOR_CQ_DB_SET_CI       (4 << 24)
143 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_MULT (5 << 24)
144 
145 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL  (1 << 24)
146 #define MTHCA_ARBEL_CQ_DB_REQ_NOT      (2 << 24)
147 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_MULT (3 << 24)
148 
149 static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
150 {
151 	if (cq->is_direct)
152 		return cq->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
153 	else
154 		return cq->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
155 			+ (entry * MTHCA_CQ_ENTRY_SIZE) % PAGE_SIZE;
156 }
157 
158 static inline struct mthca_cqe *cqe_sw(struct mthca_cq *cq, int i)
159 {
160 	struct mthca_cqe *cqe = get_cqe(cq, i);
161 	return MTHCA_CQ_ENTRY_OWNER_HW & cqe->owner ? NULL : cqe;
162 }
163 
164 static inline struct mthca_cqe *next_cqe_sw(struct mthca_cq *cq)
165 {
166 	return cqe_sw(cq, cq->cons_index & cq->ibcq.cqe);
167 }
168 
169 static inline void set_cqe_hw(struct mthca_cqe *cqe)
170 {
171 	cqe->owner = MTHCA_CQ_ENTRY_OWNER_HW;
172 }
173 
174 /*
175  * incr is ignored in native Arbel (mem-free) mode, so cq->cons_index
176  * should be correct before calling update_cons_index().
177  */
178 static inline void update_cons_index(struct mthca_dev *dev, struct mthca_cq *cq,
179 				     int incr)
180 {
181 	u32 doorbell[2];
182 
183 	if (mthca_is_memfree(dev)) {
184 		*cq->set_ci_db = cpu_to_be32(cq->cons_index);
185 		wmb();
186 	} else {
187 		doorbell[0] = cpu_to_be32(MTHCA_TAVOR_CQ_DB_INC_CI | cq->cqn);
188 		doorbell[1] = cpu_to_be32(incr - 1);
189 
190 		mthca_write64(doorbell,
191 			      dev->kar + MTHCA_CQ_DOORBELL,
192 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
193 	}
194 }
195 
196 void mthca_cq_event(struct mthca_dev *dev, u32 cqn)
197 {
198 	struct mthca_cq *cq;
199 
200 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
201 
202 	if (!cq) {
203 		mthca_warn(dev, "Completion event for bogus CQ %08x\n", cqn);
204 		return;
205 	}
206 
207 	++cq->arm_sn;
208 
209 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
210 }
211 
212 void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn)
213 {
214 	struct mthca_cq *cq;
215 	struct mthca_cqe *cqe;
216 	int prod_index;
217 	int nfreed = 0;
218 
219 	spin_lock_irq(&dev->cq_table.lock);
220 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
221 	if (cq)
222 		atomic_inc(&cq->refcount);
223 	spin_unlock_irq(&dev->cq_table.lock);
224 
225 	if (!cq)
226 		return;
227 
228 	spin_lock_irq(&cq->lock);
229 
230 	/*
231 	 * First we need to find the current producer index, so we
232 	 * know where to start cleaning from.  It doesn't matter if HW
233 	 * adds new entries after this loop -- the QP we're worried
234 	 * about is already in RESET, so the new entries won't come
235 	 * from our QP and therefore don't need to be checked.
236 	 */
237 	for (prod_index = cq->cons_index;
238 	     cqe_sw(cq, prod_index & cq->ibcq.cqe);
239 	     ++prod_index)
240 		if (prod_index == cq->cons_index + cq->ibcq.cqe)
241 			break;
242 
243 	if (0)
244 		mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
245 			  qpn, cqn, cq->cons_index, prod_index);
246 
247 	/*
248 	 * Now sweep backwards through the CQ, removing CQ entries
249 	 * that match our QP by copying older entries on top of them.
250 	 */
251 	while (prod_index > cq->cons_index) {
252 		cqe = get_cqe(cq, (prod_index - 1) & cq->ibcq.cqe);
253 		if (cqe->my_qpn == cpu_to_be32(qpn))
254 			++nfreed;
255 		else if (nfreed)
256 			memcpy(get_cqe(cq, (prod_index - 1 + nfreed) &
257 				       cq->ibcq.cqe),
258 			       cqe,
259 			       MTHCA_CQ_ENTRY_SIZE);
260 		--prod_index;
261 	}
262 
263 	if (nfreed) {
264 		wmb();
265 		cq->cons_index += nfreed;
266 		update_cons_index(dev, cq, nfreed);
267 	}
268 
269 	spin_unlock_irq(&cq->lock);
270 	if (atomic_dec_and_test(&cq->refcount))
271 		wake_up(&cq->wait);
272 }
273 
274 static int handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
275 			    struct mthca_qp *qp, int wqe_index, int is_send,
276 			    struct mthca_err_cqe *cqe,
277 			    struct ib_wc *entry, int *free_cqe)
278 {
279 	int err;
280 	int dbd;
281 	u32 new_wqe;
282 
283 	if (1 && cqe->syndrome != SYNDROME_WR_FLUSH_ERR) {
284 		int j;
285 
286 		mthca_dbg(dev, "%x/%d: error CQE -> QPN %06x, WQE @ %08x\n",
287 			  cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
288 			  be32_to_cpu(cqe->wqe));
289 
290 		for (j = 0; j < 8; ++j)
291 			printk(KERN_DEBUG "  [%2x] %08x\n",
292 			       j * 4, be32_to_cpu(((u32 *) cqe)[j]));
293 	}
294 
295 	/*
296 	 * For completions in error, only work request ID, status (and
297 	 * freed resource count for RD) have to be set.
298 	 */
299 	switch (cqe->syndrome) {
300 	case SYNDROME_LOCAL_LENGTH_ERR:
301 		entry->status = IB_WC_LOC_LEN_ERR;
302 		break;
303 	case SYNDROME_LOCAL_QP_OP_ERR:
304 		entry->status = IB_WC_LOC_QP_OP_ERR;
305 		break;
306 	case SYNDROME_LOCAL_EEC_OP_ERR:
307 		entry->status = IB_WC_LOC_EEC_OP_ERR;
308 		break;
309 	case SYNDROME_LOCAL_PROT_ERR:
310 		entry->status = IB_WC_LOC_PROT_ERR;
311 		break;
312 	case SYNDROME_WR_FLUSH_ERR:
313 		entry->status = IB_WC_WR_FLUSH_ERR;
314 		break;
315 	case SYNDROME_MW_BIND_ERR:
316 		entry->status = IB_WC_MW_BIND_ERR;
317 		break;
318 	case SYNDROME_BAD_RESP_ERR:
319 		entry->status = IB_WC_BAD_RESP_ERR;
320 		break;
321 	case SYNDROME_LOCAL_ACCESS_ERR:
322 		entry->status = IB_WC_LOC_ACCESS_ERR;
323 		break;
324 	case SYNDROME_REMOTE_INVAL_REQ_ERR:
325 		entry->status = IB_WC_REM_INV_REQ_ERR;
326 		break;
327 	case SYNDROME_REMOTE_ACCESS_ERR:
328 		entry->status = IB_WC_REM_ACCESS_ERR;
329 		break;
330 	case SYNDROME_REMOTE_OP_ERR:
331 		entry->status = IB_WC_REM_OP_ERR;
332 		break;
333 	case SYNDROME_RETRY_EXC_ERR:
334 		entry->status = IB_WC_RETRY_EXC_ERR;
335 		break;
336 	case SYNDROME_RNR_RETRY_EXC_ERR:
337 		entry->status = IB_WC_RNR_RETRY_EXC_ERR;
338 		break;
339 	case SYNDROME_LOCAL_RDD_VIOL_ERR:
340 		entry->status = IB_WC_LOC_RDD_VIOL_ERR;
341 		break;
342 	case SYNDROME_REMOTE_INVAL_RD_REQ_ERR:
343 		entry->status = IB_WC_REM_INV_RD_REQ_ERR;
344 		break;
345 	case SYNDROME_REMOTE_ABORTED_ERR:
346 		entry->status = IB_WC_REM_ABORT_ERR;
347 		break;
348 	case SYNDROME_INVAL_EECN_ERR:
349 		entry->status = IB_WC_INV_EECN_ERR;
350 		break;
351 	case SYNDROME_INVAL_EEC_STATE_ERR:
352 		entry->status = IB_WC_INV_EEC_STATE_ERR;
353 		break;
354 	default:
355 		entry->status = IB_WC_GENERAL_ERR;
356 		break;
357 	}
358 
359 	err = mthca_free_err_wqe(dev, qp, is_send, wqe_index, &dbd, &new_wqe);
360 	if (err)
361 		return err;
362 
363 	/*
364 	 * If we're at the end of the WQE chain, or we've used up our
365 	 * doorbell count, free the CQE.  Otherwise just update it for
366 	 * the next poll operation.
367 	 */
368 	if (!(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
369 		return 0;
370 
371 	cqe->db_cnt   = cpu_to_be16(be16_to_cpu(cqe->db_cnt) - dbd);
372 	cqe->wqe      = new_wqe;
373 	cqe->syndrome = SYNDROME_WR_FLUSH_ERR;
374 
375 	*free_cqe = 0;
376 
377 	return 0;
378 }
379 
380 static void dump_cqe(struct mthca_cqe *cqe)
381 {
382 	int j;
383 
384 	for (j = 0; j < 8; ++j)
385 		printk(KERN_DEBUG "  [%2x] %08x\n",
386 		       j * 4, be32_to_cpu(((u32 *) cqe)[j]));
387 }
388 
389 static inline int mthca_poll_one(struct mthca_dev *dev,
390 				 struct mthca_cq *cq,
391 				 struct mthca_qp **cur_qp,
392 				 int *freed,
393 				 struct ib_wc *entry)
394 {
395 	struct mthca_wq *wq;
396 	struct mthca_cqe *cqe;
397 	int wqe_index;
398 	int is_error;
399 	int is_send;
400 	int free_cqe = 1;
401 	int err = 0;
402 
403 	cqe = next_cqe_sw(cq);
404 	if (!cqe)
405 		return -EAGAIN;
406 
407 	/*
408 	 * Make sure we read CQ entry contents after we've checked the
409 	 * ownership bit.
410 	 */
411 	rmb();
412 
413 	if (0) {
414 		mthca_dbg(dev, "%x/%d: CQE -> QPN %06x, WQE @ %08x\n",
415 			  cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
416 			  be32_to_cpu(cqe->wqe));
417 
418 		dump_cqe(cqe);
419 	}
420 
421 	is_error = (cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
422 		MTHCA_ERROR_CQE_OPCODE_MASK;
423 	is_send  = is_error ? cqe->opcode & 0x01 : cqe->is_send & 0x80;
424 
425 	if (!*cur_qp || be32_to_cpu(cqe->my_qpn) != (*cur_qp)->qpn) {
426 		/*
427 		 * We do not have to take the QP table lock here,
428 		 * because CQs will be locked while QPs are removed
429 		 * from the table.
430 		 */
431 		*cur_qp = mthca_array_get(&dev->qp_table.qp,
432 					  be32_to_cpu(cqe->my_qpn) &
433 					  (dev->limits.num_qps - 1));
434 		if (!*cur_qp) {
435 			mthca_warn(dev, "CQ entry for unknown QP %06x\n",
436 				   be32_to_cpu(cqe->my_qpn) & 0xffffff);
437 			err = -EINVAL;
438 			goto out;
439 		}
440 	}
441 
442 	entry->qp_num = (*cur_qp)->qpn;
443 
444 	if (is_send) {
445 		wq = &(*cur_qp)->sq;
446 		wqe_index = ((be32_to_cpu(cqe->wqe) - (*cur_qp)->send_wqe_offset)
447 			     >> wq->wqe_shift);
448 		entry->wr_id = (*cur_qp)->wrid[wqe_index +
449 					       (*cur_qp)->rq.max];
450 	} else {
451 		wq = &(*cur_qp)->rq;
452 		wqe_index = be32_to_cpu(cqe->wqe) >> wq->wqe_shift;
453 		entry->wr_id = (*cur_qp)->wrid[wqe_index];
454 	}
455 
456 	if (wq->last_comp < wqe_index)
457 		wq->tail += wqe_index - wq->last_comp;
458 	else
459 		wq->tail += wqe_index + wq->max - wq->last_comp;
460 
461 	wq->last_comp = wqe_index;
462 
463 	if (0)
464 		mthca_dbg(dev, "%s completion for QP %06x, index %d (nr %d)\n",
465 			  is_send ? "Send" : "Receive",
466 			  (*cur_qp)->qpn, wqe_index, wq->max);
467 
468 	if (is_error) {
469 		err = handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,
470 				       (struct mthca_err_cqe *) cqe,
471 				       entry, &free_cqe);
472 		goto out;
473 	}
474 
475 	if (is_send) {
476 		entry->wc_flags = 0;
477 		switch (cqe->opcode) {
478 		case MTHCA_OPCODE_RDMA_WRITE:
479 			entry->opcode    = IB_WC_RDMA_WRITE;
480 			break;
481 		case MTHCA_OPCODE_RDMA_WRITE_IMM:
482 			entry->opcode    = IB_WC_RDMA_WRITE;
483 			entry->wc_flags |= IB_WC_WITH_IMM;
484 			break;
485 		case MTHCA_OPCODE_SEND:
486 			entry->opcode    = IB_WC_SEND;
487 			break;
488 		case MTHCA_OPCODE_SEND_IMM:
489 			entry->opcode    = IB_WC_SEND;
490 			entry->wc_flags |= IB_WC_WITH_IMM;
491 			break;
492 		case MTHCA_OPCODE_RDMA_READ:
493 			entry->opcode    = IB_WC_RDMA_READ;
494 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
495 			break;
496 		case MTHCA_OPCODE_ATOMIC_CS:
497 			entry->opcode    = IB_WC_COMP_SWAP;
498 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
499 			break;
500 		case MTHCA_OPCODE_ATOMIC_FA:
501 			entry->opcode    = IB_WC_FETCH_ADD;
502 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
503 			break;
504 		case MTHCA_OPCODE_BIND_MW:
505 			entry->opcode    = IB_WC_BIND_MW;
506 			break;
507 		default:
508 			entry->opcode    = MTHCA_OPCODE_INVALID;
509 			break;
510 		}
511 	} else {
512 		entry->byte_len = be32_to_cpu(cqe->byte_cnt);
513 		switch (cqe->opcode & 0x1f) {
514 		case IB_OPCODE_SEND_LAST_WITH_IMMEDIATE:
515 		case IB_OPCODE_SEND_ONLY_WITH_IMMEDIATE:
516 			entry->wc_flags = IB_WC_WITH_IMM;
517 			entry->imm_data = cqe->imm_etype_pkey_eec;
518 			entry->opcode = IB_WC_RECV;
519 			break;
520 		case IB_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
521 		case IB_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
522 			entry->wc_flags = IB_WC_WITH_IMM;
523 			entry->imm_data = cqe->imm_etype_pkey_eec;
524 			entry->opcode = IB_WC_RECV_RDMA_WITH_IMM;
525 			break;
526 		default:
527 			entry->wc_flags = 0;
528 			entry->opcode = IB_WC_RECV;
529 			break;
530 		}
531 		entry->slid 	   = be16_to_cpu(cqe->rlid);
532 		entry->sl   	   = be16_to_cpu(cqe->sl_g_mlpath) >> 12;
533 		entry->src_qp 	   = be32_to_cpu(cqe->rqpn) & 0xffffff;
534 		entry->dlid_path_bits = be16_to_cpu(cqe->sl_g_mlpath) & 0x7f;
535 		entry->pkey_index  = be32_to_cpu(cqe->imm_etype_pkey_eec) >> 16;
536 		entry->wc_flags   |= be16_to_cpu(cqe->sl_g_mlpath) & 0x80 ?
537 					IB_WC_GRH : 0;
538 	}
539 
540 	entry->status = IB_WC_SUCCESS;
541 
542  out:
543 	if (likely(free_cqe)) {
544 		set_cqe_hw(cqe);
545 		++(*freed);
546 		++cq->cons_index;
547 	}
548 
549 	return err;
550 }
551 
552 int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
553 		  struct ib_wc *entry)
554 {
555 	struct mthca_dev *dev = to_mdev(ibcq->device);
556 	struct mthca_cq *cq = to_mcq(ibcq);
557 	struct mthca_qp *qp = NULL;
558 	unsigned long flags;
559 	int err = 0;
560 	int freed = 0;
561 	int npolled;
562 
563 	spin_lock_irqsave(&cq->lock, flags);
564 
565 	for (npolled = 0; npolled < num_entries; ++npolled) {
566 		err = mthca_poll_one(dev, cq, &qp,
567 				     &freed, entry + npolled);
568 		if (err)
569 			break;
570 	}
571 
572 	if (freed) {
573 		wmb();
574 		update_cons_index(dev, cq, freed);
575 	}
576 
577 	spin_unlock_irqrestore(&cq->lock, flags);
578 
579 	return err == 0 || err == -EAGAIN ? npolled : err;
580 }
581 
582 int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify notify)
583 {
584 	u32 doorbell[2];
585 
586 	doorbell[0] = cpu_to_be32((notify == IB_CQ_SOLICITED ?
587 				   MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL :
588 				   MTHCA_TAVOR_CQ_DB_REQ_NOT)      |
589 				  to_mcq(cq)->cqn);
590 	doorbell[1] = 0xffffffff;
591 
592 	mthca_write64(doorbell,
593 		      to_mdev(cq->device)->kar + MTHCA_CQ_DOORBELL,
594 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(cq->device)->doorbell_lock));
595 
596 	return 0;
597 }
598 
599 int mthca_arbel_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify notify)
600 {
601 	struct mthca_cq *cq = to_mcq(ibcq);
602 	u32 doorbell[2];
603 	u32 sn;
604 	u32 ci;
605 
606 	sn = cq->arm_sn & 3;
607 	ci = cpu_to_be32(cq->cons_index);
608 
609 	doorbell[0] = ci;
610 	doorbell[1] = cpu_to_be32((cq->cqn << 8) | (2 << 5) | (sn << 3) |
611 				  (notify == IB_CQ_SOLICITED ? 1 : 2));
612 
613 	mthca_write_db_rec(doorbell, cq->arm_db);
614 
615 	/*
616 	 * Make sure that the doorbell record in host memory is
617 	 * written before ringing the doorbell via PCI MMIO.
618 	 */
619 	wmb();
620 
621 	doorbell[0] = cpu_to_be32((sn << 28)                       |
622 				  (notify == IB_CQ_SOLICITED ?
623 				   MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL :
624 				   MTHCA_ARBEL_CQ_DB_REQ_NOT)      |
625 				  cq->cqn);
626 	doorbell[1] = ci;
627 
628 	mthca_write64(doorbell,
629 		      to_mdev(ibcq->device)->kar + MTHCA_CQ_DOORBELL,
630 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->doorbell_lock));
631 
632 	return 0;
633 }
634 
635 static void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq *cq)
636 {
637 	int i;
638 	int size;
639 
640 	if (cq->is_direct)
641 		pci_free_consistent(dev->pdev,
642 				    (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE,
643 				    cq->queue.direct.buf,
644 				    pci_unmap_addr(&cq->queue.direct,
645 						   mapping));
646 	else {
647 		size = (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE;
648 		for (i = 0; i < (size + PAGE_SIZE - 1) / PAGE_SIZE; ++i)
649 			if (cq->queue.page_list[i].buf)
650 				pci_free_consistent(dev->pdev, PAGE_SIZE,
651 						    cq->queue.page_list[i].buf,
652 						    pci_unmap_addr(&cq->queue.page_list[i],
653 								   mapping));
654 
655 		kfree(cq->queue.page_list);
656 	}
657 }
658 
659 static int mthca_alloc_cq_buf(struct mthca_dev *dev, int size,
660 			      struct mthca_cq *cq)
661 {
662 	int err = -ENOMEM;
663 	int npages, shift;
664 	u64 *dma_list = NULL;
665 	dma_addr_t t;
666 	int i;
667 
668 	if (size <= MTHCA_MAX_DIRECT_CQ_SIZE) {
669 		cq->is_direct = 1;
670 		npages        = 1;
671 		shift         = get_order(size) + PAGE_SHIFT;
672 
673 		cq->queue.direct.buf = pci_alloc_consistent(dev->pdev,
674 							    size, &t);
675 		if (!cq->queue.direct.buf)
676 			return -ENOMEM;
677 
678 		pci_unmap_addr_set(&cq->queue.direct, mapping, t);
679 
680 		memset(cq->queue.direct.buf, 0, size);
681 
682 		while (t & ((1 << shift) - 1)) {
683 			--shift;
684 			npages *= 2;
685 		}
686 
687 		dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
688 		if (!dma_list)
689 			goto err_free;
690 
691 		for (i = 0; i < npages; ++i)
692 			dma_list[i] = t + i * (1 << shift);
693 	} else {
694 		cq->is_direct = 0;
695 		npages        = (size + PAGE_SIZE - 1) / PAGE_SIZE;
696 		shift         = PAGE_SHIFT;
697 
698 		dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
699 		if (!dma_list)
700 			return -ENOMEM;
701 
702 		cq->queue.page_list = kmalloc(npages * sizeof *cq->queue.page_list,
703 					      GFP_KERNEL);
704 		if (!cq->queue.page_list)
705 			goto err_out;
706 
707 		for (i = 0; i < npages; ++i)
708 			cq->queue.page_list[i].buf = NULL;
709 
710 		for (i = 0; i < npages; ++i) {
711 			cq->queue.page_list[i].buf =
712 				pci_alloc_consistent(dev->pdev, PAGE_SIZE, &t);
713 			if (!cq->queue.page_list[i].buf)
714 				goto err_free;
715 
716 			dma_list[i] = t;
717 			pci_unmap_addr_set(&cq->queue.page_list[i], mapping, t);
718 
719 			memset(cq->queue.page_list[i].buf, 0, PAGE_SIZE);
720 		}
721 	}
722 
723 	err = mthca_mr_alloc_phys(dev, dev->driver_pd.pd_num,
724 				  dma_list, shift, npages,
725 				  0, size,
726 				  MTHCA_MPT_FLAG_LOCAL_WRITE |
727 				  MTHCA_MPT_FLAG_LOCAL_READ,
728 				  &cq->mr);
729 	if (err)
730 		goto err_free;
731 
732 	kfree(dma_list);
733 
734 	return 0;
735 
736 err_free:
737 	mthca_free_cq_buf(dev, cq);
738 
739 err_out:
740 	kfree(dma_list);
741 
742 	return err;
743 }
744 
745 int mthca_init_cq(struct mthca_dev *dev, int nent,
746 		  struct mthca_cq *cq)
747 {
748 	int size = nent * MTHCA_CQ_ENTRY_SIZE;
749 	void *mailbox = NULL;
750 	struct mthca_cq_context *cq_context;
751 	int err = -ENOMEM;
752 	u8 status;
753 	int i;
754 
755 	might_sleep();
756 
757 	cq->ibcq.cqe = nent - 1;
758 
759 	cq->cqn = mthca_alloc(&dev->cq_table.alloc);
760 	if (cq->cqn == -1)
761 		return -ENOMEM;
762 
763 	if (mthca_is_memfree(dev)) {
764 		cq->arm_sn = 1;
765 
766 		err = mthca_table_get(dev, dev->cq_table.table, cq->cqn);
767 		if (err)
768 			goto err_out;
769 
770 		err = -ENOMEM;
771 
772 		cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
773 						     cq->cqn, &cq->set_ci_db);
774 		if (cq->set_ci_db_index < 0)
775 			goto err_out_icm;
776 
777 		cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
778 						  cq->cqn, &cq->arm_db);
779 		if (cq->arm_db_index < 0)
780 			goto err_out_ci;
781 	}
782 
783 	mailbox = kmalloc(sizeof (struct mthca_cq_context) + MTHCA_CMD_MAILBOX_EXTRA,
784 			  GFP_KERNEL);
785 	if (!mailbox)
786 		goto err_out_mailbox;
787 
788 	cq_context = MAILBOX_ALIGN(mailbox);
789 
790 	err = mthca_alloc_cq_buf(dev, size, cq);
791 	if (err)
792 		goto err_out_mailbox;
793 
794 	for (i = 0; i < nent; ++i)
795 		set_cqe_hw(get_cqe(cq, i));
796 
797 	spin_lock_init(&cq->lock);
798 	atomic_set(&cq->refcount, 1);
799 	init_waitqueue_head(&cq->wait);
800 
801 	memset(cq_context, 0, sizeof *cq_context);
802 	cq_context->flags           = cpu_to_be32(MTHCA_CQ_STATUS_OK      |
803 						  MTHCA_CQ_STATE_DISARMED |
804 						  MTHCA_CQ_FLAG_TR);
805 	cq_context->start           = cpu_to_be64(0);
806 	cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24 |
807 						  dev->driver_uar.index);
808 	cq_context->error_eqn       = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
809 	cq_context->comp_eqn        = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
810 	cq_context->pd              = cpu_to_be32(dev->driver_pd.pd_num);
811 	cq_context->lkey            = cpu_to_be32(cq->mr.ibmr.lkey);
812 	cq_context->cqn             = cpu_to_be32(cq->cqn);
813 
814 	if (mthca_is_memfree(dev)) {
815 		cq_context->ci_db    = cpu_to_be32(cq->set_ci_db_index);
816 		cq_context->state_db = cpu_to_be32(cq->arm_db_index);
817 	}
818 
819 	err = mthca_SW2HW_CQ(dev, cq_context, cq->cqn, &status);
820 	if (err) {
821 		mthca_warn(dev, "SW2HW_CQ failed (%d)\n", err);
822 		goto err_out_free_mr;
823 	}
824 
825 	if (status) {
826 		mthca_warn(dev, "SW2HW_CQ returned status 0x%02x\n",
827 			   status);
828 		err = -EINVAL;
829 		goto err_out_free_mr;
830 	}
831 
832 	spin_lock_irq(&dev->cq_table.lock);
833 	if (mthca_array_set(&dev->cq_table.cq,
834 			    cq->cqn & (dev->limits.num_cqs - 1),
835 			    cq)) {
836 		spin_unlock_irq(&dev->cq_table.lock);
837 		goto err_out_free_mr;
838 	}
839 	spin_unlock_irq(&dev->cq_table.lock);
840 
841 	cq->cons_index = 0;
842 
843 	kfree(mailbox);
844 
845 	return 0;
846 
847 err_out_free_mr:
848 	mthca_free_mr(dev, &cq->mr);
849 	mthca_free_cq_buf(dev, cq);
850 
851 err_out_mailbox:
852 	kfree(mailbox);
853 
854 	if (mthca_is_memfree(dev))
855 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
856 
857 err_out_ci:
858 	if (mthca_is_memfree(dev))
859 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
860 
861 err_out_icm:
862 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
863 
864 err_out:
865 	mthca_free(&dev->cq_table.alloc, cq->cqn);
866 
867 	return err;
868 }
869 
870 void mthca_free_cq(struct mthca_dev *dev,
871 		   struct mthca_cq *cq)
872 {
873 	void *mailbox;
874 	int err;
875 	u8 status;
876 
877 	might_sleep();
878 
879 	mailbox = kmalloc(sizeof (struct mthca_cq_context) + MTHCA_CMD_MAILBOX_EXTRA,
880 			  GFP_KERNEL);
881 	if (!mailbox) {
882 		mthca_warn(dev, "No memory for mailbox to free CQ.\n");
883 		return;
884 	}
885 
886 	err = mthca_HW2SW_CQ(dev, MAILBOX_ALIGN(mailbox), cq->cqn, &status);
887 	if (err)
888 		mthca_warn(dev, "HW2SW_CQ failed (%d)\n", err);
889 	else if (status)
890 		mthca_warn(dev, "HW2SW_CQ returned status 0x%02x\n",
891 			   status);
892 
893 	if (0) {
894 		u32 *ctx = MAILBOX_ALIGN(mailbox);
895 		int j;
896 
897 		printk(KERN_ERR "context for CQN %x (cons index %x, next sw %d)\n",
898 		       cq->cqn, cq->cons_index, !!next_cqe_sw(cq));
899 		for (j = 0; j < 16; ++j)
900 			printk(KERN_ERR "[%2x] %08x\n", j * 4, be32_to_cpu(ctx[j]));
901 	}
902 
903 	spin_lock_irq(&dev->cq_table.lock);
904 	mthca_array_clear(&dev->cq_table.cq,
905 			  cq->cqn & (dev->limits.num_cqs - 1));
906 	spin_unlock_irq(&dev->cq_table.lock);
907 
908 	if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
909 		synchronize_irq(dev->eq_table.eq[MTHCA_EQ_COMP].msi_x_vector);
910 	else
911 		synchronize_irq(dev->pdev->irq);
912 
913 	atomic_dec(&cq->refcount);
914 	wait_event(cq->wait, !atomic_read(&cq->refcount));
915 
916 	mthca_free_mr(dev, &cq->mr);
917 	mthca_free_cq_buf(dev, cq);
918 
919 	if (mthca_is_memfree(dev)) {
920 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM,    cq->arm_db_index);
921 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
922 		mthca_table_put(dev, dev->cq_table.table, cq->cqn);
923 	}
924 
925 	mthca_free(&dev->cq_table.alloc, cq->cqn);
926 	kfree(mailbox);
927 }
928 
929 int __devinit mthca_init_cq_table(struct mthca_dev *dev)
930 {
931 	int err;
932 
933 	spin_lock_init(&dev->cq_table.lock);
934 
935 	err = mthca_alloc_init(&dev->cq_table.alloc,
936 			       dev->limits.num_cqs,
937 			       (1 << 24) - 1,
938 			       dev->limits.reserved_cqs);
939 	if (err)
940 		return err;
941 
942 	err = mthca_array_init(&dev->cq_table.cq,
943 			       dev->limits.num_cqs);
944 	if (err)
945 		mthca_alloc_cleanup(&dev->cq_table.alloc);
946 
947 	return err;
948 }
949 
950 void __devexit mthca_cleanup_cq_table(struct mthca_dev *dev)
951 {
952 	mthca_array_cleanup(&dev->cq_table.cq, dev->limits.num_cqs);
953 	mthca_alloc_cleanup(&dev->cq_table.alloc);
954 }
955