xref: /linux/drivers/infiniband/hw/mthca/mthca_cq.c (revision de2fe5e07d58424bc286fff3fd3c1b0bf933cd58)
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, 2006 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     vendor_err;
132 	__be16 db_cnt;
133 	u32    reserved2;
134 	__be32 wqe;
135 	u8     opcode;
136 	u8     reserved3[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_from_buf(struct mthca_cq_buf *buf,
154 						 int entry)
155 {
156 	if (buf->is_direct)
157 		return buf->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
158 	else
159 		return buf->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
160 			+ (entry * MTHCA_CQ_ENTRY_SIZE) % PAGE_SIZE;
161 }
162 
163 static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
164 {
165 	return get_cqe_from_buf(&cq->buf, entry);
166 }
167 
168 static inline struct mthca_cqe *cqe_sw(struct mthca_cqe *cqe)
169 {
170 	return MTHCA_CQ_ENTRY_OWNER_HW & cqe->owner ? NULL : cqe;
171 }
172 
173 static inline struct mthca_cqe *next_cqe_sw(struct mthca_cq *cq)
174 {
175 	return cqe_sw(get_cqe(cq, cq->cons_index & cq->ibcq.cqe));
176 }
177 
178 static inline void set_cqe_hw(struct mthca_cqe *cqe)
179 {
180 	cqe->owner = MTHCA_CQ_ENTRY_OWNER_HW;
181 }
182 
183 static void dump_cqe(struct mthca_dev *dev, void *cqe_ptr)
184 {
185 	__be32 *cqe = cqe_ptr;
186 
187 	(void) cqe;	/* avoid warning if mthca_dbg compiled away... */
188 	mthca_dbg(dev, "CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
189 		  be32_to_cpu(cqe[0]), be32_to_cpu(cqe[1]), be32_to_cpu(cqe[2]),
190 		  be32_to_cpu(cqe[3]), be32_to_cpu(cqe[4]), be32_to_cpu(cqe[5]),
191 		  be32_to_cpu(cqe[6]), be32_to_cpu(cqe[7]));
192 }
193 
194 /*
195  * incr is ignored in native Arbel (mem-free) mode, so cq->cons_index
196  * should be correct before calling update_cons_index().
197  */
198 static inline void update_cons_index(struct mthca_dev *dev, struct mthca_cq *cq,
199 				     int incr)
200 {
201 	__be32 doorbell[2];
202 
203 	if (mthca_is_memfree(dev)) {
204 		*cq->set_ci_db = cpu_to_be32(cq->cons_index);
205 		wmb();
206 	} else {
207 		doorbell[0] = cpu_to_be32(MTHCA_TAVOR_CQ_DB_INC_CI | cq->cqn);
208 		doorbell[1] = cpu_to_be32(incr - 1);
209 
210 		mthca_write64(doorbell,
211 			      dev->kar + MTHCA_CQ_DOORBELL,
212 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
213 	}
214 }
215 
216 void mthca_cq_completion(struct mthca_dev *dev, u32 cqn)
217 {
218 	struct mthca_cq *cq;
219 
220 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
221 
222 	if (!cq) {
223 		mthca_warn(dev, "Completion event for bogus CQ %08x\n", cqn);
224 		return;
225 	}
226 
227 	++cq->arm_sn;
228 
229 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
230 }
231 
232 void mthca_cq_event(struct mthca_dev *dev, u32 cqn,
233 		    enum ib_event_type event_type)
234 {
235 	struct mthca_cq *cq;
236 	struct ib_event event;
237 
238 	spin_lock(&dev->cq_table.lock);
239 
240 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
241 
242 	if (cq)
243 		atomic_inc(&cq->refcount);
244 	spin_unlock(&dev->cq_table.lock);
245 
246 	if (!cq) {
247 		mthca_warn(dev, "Async event for bogus CQ %08x\n", cqn);
248 		return;
249 	}
250 
251 	event.device      = &dev->ib_dev;
252 	event.event       = event_type;
253 	event.element.cq  = &cq->ibcq;
254 	if (cq->ibcq.event_handler)
255 		cq->ibcq.event_handler(&event, cq->ibcq.cq_context);
256 
257 	if (atomic_dec_and_test(&cq->refcount))
258 		wake_up(&cq->wait);
259 }
260 
261 static inline int is_recv_cqe(struct mthca_cqe *cqe)
262 {
263 	if ((cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
264 	    MTHCA_ERROR_CQE_OPCODE_MASK)
265 		return !(cqe->opcode & 0x01);
266 	else
267 		return !(cqe->is_send & 0x80);
268 }
269 
270 void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn,
271 		    struct mthca_srq *srq)
272 {
273 	struct mthca_cq *cq;
274 	struct mthca_cqe *cqe;
275 	u32 prod_index;
276 	int nfreed = 0;
277 
278 	spin_lock_irq(&dev->cq_table.lock);
279 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
280 	if (cq)
281 		atomic_inc(&cq->refcount);
282 	spin_unlock_irq(&dev->cq_table.lock);
283 
284 	if (!cq)
285 		return;
286 
287 	spin_lock_irq(&cq->lock);
288 
289 	/*
290 	 * First we need to find the current producer index, so we
291 	 * know where to start cleaning from.  It doesn't matter if HW
292 	 * adds new entries after this loop -- the QP we're worried
293 	 * about is already in RESET, so the new entries won't come
294 	 * from our QP and therefore don't need to be checked.
295 	 */
296 	for (prod_index = cq->cons_index;
297 	     cqe_sw(get_cqe(cq, prod_index & cq->ibcq.cqe));
298 	     ++prod_index)
299 		if (prod_index == cq->cons_index + cq->ibcq.cqe)
300 			break;
301 
302 	if (0)
303 		mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
304 			  qpn, cqn, cq->cons_index, prod_index);
305 
306 	/*
307 	 * Now sweep backwards through the CQ, removing CQ entries
308 	 * that match our QP by copying older entries on top of them.
309 	 */
310 	while ((int) --prod_index - (int) cq->cons_index >= 0) {
311 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
312 		if (cqe->my_qpn == cpu_to_be32(qpn)) {
313 			if (srq && is_recv_cqe(cqe))
314 				mthca_free_srq_wqe(srq, be32_to_cpu(cqe->wqe));
315 			++nfreed;
316 		} else if (nfreed)
317 			memcpy(get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe),
318 			       cqe, MTHCA_CQ_ENTRY_SIZE);
319 	}
320 
321 	if (nfreed) {
322 		wmb();
323 		cq->cons_index += nfreed;
324 		update_cons_index(dev, cq, nfreed);
325 	}
326 
327 	spin_unlock_irq(&cq->lock);
328 	if (atomic_dec_and_test(&cq->refcount))
329 		wake_up(&cq->wait);
330 }
331 
332 void mthca_cq_resize_copy_cqes(struct mthca_cq *cq)
333 {
334 	int i;
335 
336 	/*
337 	 * In Tavor mode, the hardware keeps the consumer and producer
338 	 * indices mod the CQ size.  Since we might be making the CQ
339 	 * bigger, we need to deal with the case where the producer
340 	 * index wrapped around before the CQ was resized.
341 	 */
342 	if (!mthca_is_memfree(to_mdev(cq->ibcq.device)) &&
343 	    cq->ibcq.cqe < cq->resize_buf->cqe) {
344 		cq->cons_index &= cq->ibcq.cqe;
345 		if (cqe_sw(get_cqe(cq, cq->ibcq.cqe)))
346 			cq->cons_index -= cq->ibcq.cqe + 1;
347 	}
348 
349 	for (i = cq->cons_index; cqe_sw(get_cqe(cq, i & cq->ibcq.cqe)); ++i)
350 		memcpy(get_cqe_from_buf(&cq->resize_buf->buf,
351 					i & cq->resize_buf->cqe),
352 		       get_cqe(cq, i & cq->ibcq.cqe), MTHCA_CQ_ENTRY_SIZE);
353 }
354 
355 int mthca_alloc_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int nent)
356 {
357 	int ret;
358 	int i;
359 
360 	ret = mthca_buf_alloc(dev, nent * MTHCA_CQ_ENTRY_SIZE,
361 			      MTHCA_MAX_DIRECT_CQ_SIZE,
362 			      &buf->queue, &buf->is_direct,
363 			      &dev->driver_pd, 1, &buf->mr);
364 	if (ret)
365 		return ret;
366 
367 	for (i = 0; i < nent; ++i)
368 		set_cqe_hw(get_cqe_from_buf(buf, i));
369 
370 	return 0;
371 }
372 
373 void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int cqe)
374 {
375 	mthca_buf_free(dev, (cqe + 1) * MTHCA_CQ_ENTRY_SIZE, &buf->queue,
376 		       buf->is_direct, &buf->mr);
377 }
378 
379 static void handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
380 			     struct mthca_qp *qp, int wqe_index, int is_send,
381 			     struct mthca_err_cqe *cqe,
382 			     struct ib_wc *entry, int *free_cqe)
383 {
384 	int dbd;
385 	__be32 new_wqe;
386 
387 	if (cqe->syndrome == SYNDROME_LOCAL_QP_OP_ERR) {
388 		mthca_dbg(dev, "local QP operation err "
389 			  "(QPN %06x, WQE @ %08x, CQN %06x, index %d)\n",
390 			  be32_to_cpu(cqe->my_qpn), be32_to_cpu(cqe->wqe),
391 			  cq->cqn, cq->cons_index);
392 		dump_cqe(dev, cqe);
393 	}
394 
395 	/*
396 	 * For completions in error, only work request ID, status, vendor error
397 	 * (and freed resource count for RD) have to be set.
398 	 */
399 	switch (cqe->syndrome) {
400 	case SYNDROME_LOCAL_LENGTH_ERR:
401 		entry->status = IB_WC_LOC_LEN_ERR;
402 		break;
403 	case SYNDROME_LOCAL_QP_OP_ERR:
404 		entry->status = IB_WC_LOC_QP_OP_ERR;
405 		break;
406 	case SYNDROME_LOCAL_EEC_OP_ERR:
407 		entry->status = IB_WC_LOC_EEC_OP_ERR;
408 		break;
409 	case SYNDROME_LOCAL_PROT_ERR:
410 		entry->status = IB_WC_LOC_PROT_ERR;
411 		break;
412 	case SYNDROME_WR_FLUSH_ERR:
413 		entry->status = IB_WC_WR_FLUSH_ERR;
414 		break;
415 	case SYNDROME_MW_BIND_ERR:
416 		entry->status = IB_WC_MW_BIND_ERR;
417 		break;
418 	case SYNDROME_BAD_RESP_ERR:
419 		entry->status = IB_WC_BAD_RESP_ERR;
420 		break;
421 	case SYNDROME_LOCAL_ACCESS_ERR:
422 		entry->status = IB_WC_LOC_ACCESS_ERR;
423 		break;
424 	case SYNDROME_REMOTE_INVAL_REQ_ERR:
425 		entry->status = IB_WC_REM_INV_REQ_ERR;
426 		break;
427 	case SYNDROME_REMOTE_ACCESS_ERR:
428 		entry->status = IB_WC_REM_ACCESS_ERR;
429 		break;
430 	case SYNDROME_REMOTE_OP_ERR:
431 		entry->status = IB_WC_REM_OP_ERR;
432 		break;
433 	case SYNDROME_RETRY_EXC_ERR:
434 		entry->status = IB_WC_RETRY_EXC_ERR;
435 		break;
436 	case SYNDROME_RNR_RETRY_EXC_ERR:
437 		entry->status = IB_WC_RNR_RETRY_EXC_ERR;
438 		break;
439 	case SYNDROME_LOCAL_RDD_VIOL_ERR:
440 		entry->status = IB_WC_LOC_RDD_VIOL_ERR;
441 		break;
442 	case SYNDROME_REMOTE_INVAL_RD_REQ_ERR:
443 		entry->status = IB_WC_REM_INV_RD_REQ_ERR;
444 		break;
445 	case SYNDROME_REMOTE_ABORTED_ERR:
446 		entry->status = IB_WC_REM_ABORT_ERR;
447 		break;
448 	case SYNDROME_INVAL_EECN_ERR:
449 		entry->status = IB_WC_INV_EECN_ERR;
450 		break;
451 	case SYNDROME_INVAL_EEC_STATE_ERR:
452 		entry->status = IB_WC_INV_EEC_STATE_ERR;
453 		break;
454 	default:
455 		entry->status = IB_WC_GENERAL_ERR;
456 		break;
457 	}
458 
459 	entry->vendor_err = cqe->vendor_err;
460 
461 	/*
462 	 * Mem-free HCAs always generate one CQE per WQE, even in the
463 	 * error case, so we don't have to check the doorbell count, etc.
464 	 */
465 	if (mthca_is_memfree(dev))
466 		return;
467 
468 	mthca_free_err_wqe(dev, qp, is_send, wqe_index, &dbd, &new_wqe);
469 
470 	/*
471 	 * If we're at the end of the WQE chain, or we've used up our
472 	 * doorbell count, free the CQE.  Otherwise just update it for
473 	 * the next poll operation.
474 	 */
475 	if (!(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
476 		return;
477 
478 	cqe->db_cnt   = cpu_to_be16(be16_to_cpu(cqe->db_cnt) - dbd);
479 	cqe->wqe      = new_wqe;
480 	cqe->syndrome = SYNDROME_WR_FLUSH_ERR;
481 
482 	*free_cqe = 0;
483 }
484 
485 static inline int mthca_poll_one(struct mthca_dev *dev,
486 				 struct mthca_cq *cq,
487 				 struct mthca_qp **cur_qp,
488 				 int *freed,
489 				 struct ib_wc *entry)
490 {
491 	struct mthca_wq *wq;
492 	struct mthca_cqe *cqe;
493 	int wqe_index;
494 	int is_error;
495 	int is_send;
496 	int free_cqe = 1;
497 	int err = 0;
498 
499 	cqe = next_cqe_sw(cq);
500 	if (!cqe)
501 		return -EAGAIN;
502 
503 	/*
504 	 * Make sure we read CQ entry contents after we've checked the
505 	 * ownership bit.
506 	 */
507 	rmb();
508 
509 	if (0) {
510 		mthca_dbg(dev, "%x/%d: CQE -> QPN %06x, WQE @ %08x\n",
511 			  cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
512 			  be32_to_cpu(cqe->wqe));
513 		dump_cqe(dev, cqe);
514 	}
515 
516 	is_error = (cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
517 		MTHCA_ERROR_CQE_OPCODE_MASK;
518 	is_send  = is_error ? cqe->opcode & 0x01 : cqe->is_send & 0x80;
519 
520 	if (!*cur_qp || be32_to_cpu(cqe->my_qpn) != (*cur_qp)->qpn) {
521 		/*
522 		 * We do not have to take the QP table lock here,
523 		 * because CQs will be locked while QPs are removed
524 		 * from the table.
525 		 */
526 		*cur_qp = mthca_array_get(&dev->qp_table.qp,
527 					  be32_to_cpu(cqe->my_qpn) &
528 					  (dev->limits.num_qps - 1));
529 		if (!*cur_qp) {
530 			mthca_warn(dev, "CQ entry for unknown QP %06x\n",
531 				   be32_to_cpu(cqe->my_qpn) & 0xffffff);
532 			err = -EINVAL;
533 			goto out;
534 		}
535 	}
536 
537 	entry->qp_num = (*cur_qp)->qpn;
538 
539 	if (is_send) {
540 		wq = &(*cur_qp)->sq;
541 		wqe_index = ((be32_to_cpu(cqe->wqe) - (*cur_qp)->send_wqe_offset)
542 			     >> wq->wqe_shift);
543 		entry->wr_id = (*cur_qp)->wrid[wqe_index +
544 					       (*cur_qp)->rq.max];
545 	} else if ((*cur_qp)->ibqp.srq) {
546 		struct mthca_srq *srq = to_msrq((*cur_qp)->ibqp.srq);
547 		u32 wqe = be32_to_cpu(cqe->wqe);
548 		wq = NULL;
549 		wqe_index = wqe >> srq->wqe_shift;
550 		entry->wr_id = srq->wrid[wqe_index];
551 		mthca_free_srq_wqe(srq, wqe);
552 	} else {
553 		wq = &(*cur_qp)->rq;
554 		wqe_index = be32_to_cpu(cqe->wqe) >> wq->wqe_shift;
555 		entry->wr_id = (*cur_qp)->wrid[wqe_index];
556 	}
557 
558 	if (wq) {
559 		if (wq->last_comp < wqe_index)
560 			wq->tail += wqe_index - wq->last_comp;
561 		else
562 			wq->tail += wqe_index + wq->max - wq->last_comp;
563 
564 		wq->last_comp = wqe_index;
565 	}
566 
567 	if (is_error) {
568 		handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,
569 				 (struct mthca_err_cqe *) cqe,
570 				 entry, &free_cqe);
571 		goto out;
572 	}
573 
574 	if (is_send) {
575 		entry->wc_flags = 0;
576 		switch (cqe->opcode) {
577 		case MTHCA_OPCODE_RDMA_WRITE:
578 			entry->opcode    = IB_WC_RDMA_WRITE;
579 			break;
580 		case MTHCA_OPCODE_RDMA_WRITE_IMM:
581 			entry->opcode    = IB_WC_RDMA_WRITE;
582 			entry->wc_flags |= IB_WC_WITH_IMM;
583 			break;
584 		case MTHCA_OPCODE_SEND:
585 			entry->opcode    = IB_WC_SEND;
586 			break;
587 		case MTHCA_OPCODE_SEND_IMM:
588 			entry->opcode    = IB_WC_SEND;
589 			entry->wc_flags |= IB_WC_WITH_IMM;
590 			break;
591 		case MTHCA_OPCODE_RDMA_READ:
592 			entry->opcode    = IB_WC_RDMA_READ;
593 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
594 			break;
595 		case MTHCA_OPCODE_ATOMIC_CS:
596 			entry->opcode    = IB_WC_COMP_SWAP;
597 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
598 			break;
599 		case MTHCA_OPCODE_ATOMIC_FA:
600 			entry->opcode    = IB_WC_FETCH_ADD;
601 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
602 			break;
603 		case MTHCA_OPCODE_BIND_MW:
604 			entry->opcode    = IB_WC_BIND_MW;
605 			break;
606 		default:
607 			entry->opcode    = MTHCA_OPCODE_INVALID;
608 			break;
609 		}
610 	} else {
611 		entry->byte_len = be32_to_cpu(cqe->byte_cnt);
612 		switch (cqe->opcode & 0x1f) {
613 		case IB_OPCODE_SEND_LAST_WITH_IMMEDIATE:
614 		case IB_OPCODE_SEND_ONLY_WITH_IMMEDIATE:
615 			entry->wc_flags = IB_WC_WITH_IMM;
616 			entry->imm_data = cqe->imm_etype_pkey_eec;
617 			entry->opcode = IB_WC_RECV;
618 			break;
619 		case IB_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
620 		case IB_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
621 			entry->wc_flags = IB_WC_WITH_IMM;
622 			entry->imm_data = cqe->imm_etype_pkey_eec;
623 			entry->opcode = IB_WC_RECV_RDMA_WITH_IMM;
624 			break;
625 		default:
626 			entry->wc_flags = 0;
627 			entry->opcode = IB_WC_RECV;
628 			break;
629 		}
630 		entry->slid 	   = be16_to_cpu(cqe->rlid);
631 		entry->sl   	   = be16_to_cpu(cqe->sl_g_mlpath) >> 12;
632 		entry->src_qp 	   = be32_to_cpu(cqe->rqpn) & 0xffffff;
633 		entry->dlid_path_bits = be16_to_cpu(cqe->sl_g_mlpath) & 0x7f;
634 		entry->pkey_index  = be32_to_cpu(cqe->imm_etype_pkey_eec) >> 16;
635 		entry->wc_flags   |= be16_to_cpu(cqe->sl_g_mlpath) & 0x80 ?
636 					IB_WC_GRH : 0;
637 	}
638 
639 	entry->status = IB_WC_SUCCESS;
640 
641  out:
642 	if (likely(free_cqe)) {
643 		set_cqe_hw(cqe);
644 		++(*freed);
645 		++cq->cons_index;
646 	}
647 
648 	return err;
649 }
650 
651 int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
652 		  struct ib_wc *entry)
653 {
654 	struct mthca_dev *dev = to_mdev(ibcq->device);
655 	struct mthca_cq *cq = to_mcq(ibcq);
656 	struct mthca_qp *qp = NULL;
657 	unsigned long flags;
658 	int err = 0;
659 	int freed = 0;
660 	int npolled;
661 
662 	spin_lock_irqsave(&cq->lock, flags);
663 
664 	npolled = 0;
665 repoll:
666 	while (npolled < num_entries) {
667 		err = mthca_poll_one(dev, cq, &qp,
668 				     &freed, entry + npolled);
669 		if (err)
670 			break;
671 		++npolled;
672 	}
673 
674 	if (freed) {
675 		wmb();
676 		update_cons_index(dev, cq, freed);
677 	}
678 
679 	/*
680 	 * If a CQ resize is in progress and we discovered that the
681 	 * old buffer is empty, then peek in the new buffer, and if
682 	 * it's not empty, switch to the new buffer and continue
683 	 * polling there.
684 	 */
685 	if (unlikely(err == -EAGAIN && cq->resize_buf &&
686 		     cq->resize_buf->state == CQ_RESIZE_READY)) {
687 		/*
688 		 * In Tavor mode, the hardware keeps the producer
689 		 * index modulo the CQ size.  Since we might be making
690 		 * the CQ bigger, we need to mask our consumer index
691 		 * using the size of the old CQ buffer before looking
692 		 * in the new CQ buffer.
693 		 */
694 		if (!mthca_is_memfree(dev))
695 			cq->cons_index &= cq->ibcq.cqe;
696 
697 		if (cqe_sw(get_cqe_from_buf(&cq->resize_buf->buf,
698 					    cq->cons_index & cq->resize_buf->cqe))) {
699 			struct mthca_cq_buf tbuf;
700 			int tcqe;
701 
702 			tbuf         = cq->buf;
703 			tcqe         = cq->ibcq.cqe;
704 			cq->buf      = cq->resize_buf->buf;
705 			cq->ibcq.cqe = cq->resize_buf->cqe;
706 
707 			cq->resize_buf->buf   = tbuf;
708 			cq->resize_buf->cqe   = tcqe;
709 			cq->resize_buf->state = CQ_RESIZE_SWAPPED;
710 
711 			goto repoll;
712 		}
713 	}
714 
715 	spin_unlock_irqrestore(&cq->lock, flags);
716 
717 	return err == 0 || err == -EAGAIN ? npolled : err;
718 }
719 
720 int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify notify)
721 {
722 	__be32 doorbell[2];
723 
724 	doorbell[0] = cpu_to_be32((notify == IB_CQ_SOLICITED ?
725 				   MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL :
726 				   MTHCA_TAVOR_CQ_DB_REQ_NOT)      |
727 				  to_mcq(cq)->cqn);
728 	doorbell[1] = (__force __be32) 0xffffffff;
729 
730 	mthca_write64(doorbell,
731 		      to_mdev(cq->device)->kar + MTHCA_CQ_DOORBELL,
732 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(cq->device)->doorbell_lock));
733 
734 	return 0;
735 }
736 
737 int mthca_arbel_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify notify)
738 {
739 	struct mthca_cq *cq = to_mcq(ibcq);
740 	__be32 doorbell[2];
741 	u32 sn;
742 	__be32 ci;
743 
744 	sn = cq->arm_sn & 3;
745 	ci = cpu_to_be32(cq->cons_index);
746 
747 	doorbell[0] = ci;
748 	doorbell[1] = cpu_to_be32((cq->cqn << 8) | (2 << 5) | (sn << 3) |
749 				  (notify == IB_CQ_SOLICITED ? 1 : 2));
750 
751 	mthca_write_db_rec(doorbell, cq->arm_db);
752 
753 	/*
754 	 * Make sure that the doorbell record in host memory is
755 	 * written before ringing the doorbell via PCI MMIO.
756 	 */
757 	wmb();
758 
759 	doorbell[0] = cpu_to_be32((sn << 28)                       |
760 				  (notify == IB_CQ_SOLICITED ?
761 				   MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL :
762 				   MTHCA_ARBEL_CQ_DB_REQ_NOT)      |
763 				  cq->cqn);
764 	doorbell[1] = ci;
765 
766 	mthca_write64(doorbell,
767 		      to_mdev(ibcq->device)->kar + MTHCA_CQ_DOORBELL,
768 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->doorbell_lock));
769 
770 	return 0;
771 }
772 
773 int mthca_init_cq(struct mthca_dev *dev, int nent,
774 		  struct mthca_ucontext *ctx, u32 pdn,
775 		  struct mthca_cq *cq)
776 {
777 	struct mthca_mailbox *mailbox;
778 	struct mthca_cq_context *cq_context;
779 	int err = -ENOMEM;
780 	u8 status;
781 
782 	cq->ibcq.cqe  = nent - 1;
783 	cq->is_kernel = !ctx;
784 
785 	cq->cqn = mthca_alloc(&dev->cq_table.alloc);
786 	if (cq->cqn == -1)
787 		return -ENOMEM;
788 
789 	if (mthca_is_memfree(dev)) {
790 		err = mthca_table_get(dev, dev->cq_table.table, cq->cqn);
791 		if (err)
792 			goto err_out;
793 
794 		if (cq->is_kernel) {
795 			cq->arm_sn = 1;
796 
797 			err = -ENOMEM;
798 
799 			cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
800 							     cq->cqn, &cq->set_ci_db);
801 			if (cq->set_ci_db_index < 0)
802 				goto err_out_icm;
803 
804 			cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
805 							  cq->cqn, &cq->arm_db);
806 			if (cq->arm_db_index < 0)
807 				goto err_out_ci;
808 		}
809 	}
810 
811 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
812 	if (IS_ERR(mailbox))
813 		goto err_out_arm;
814 
815 	cq_context = mailbox->buf;
816 
817 	if (cq->is_kernel) {
818 		err = mthca_alloc_cq_buf(dev, &cq->buf, nent);
819 		if (err)
820 			goto err_out_mailbox;
821 	}
822 
823 	spin_lock_init(&cq->lock);
824 	atomic_set(&cq->refcount, 1);
825 	init_waitqueue_head(&cq->wait);
826 
827 	memset(cq_context, 0, sizeof *cq_context);
828 	cq_context->flags           = cpu_to_be32(MTHCA_CQ_STATUS_OK      |
829 						  MTHCA_CQ_STATE_DISARMED |
830 						  MTHCA_CQ_FLAG_TR);
831 	cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24);
832 	if (ctx)
833 		cq_context->logsize_usrpage |= cpu_to_be32(ctx->uar.index);
834 	else
835 		cq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
836 	cq_context->error_eqn       = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
837 	cq_context->comp_eqn        = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
838 	cq_context->pd              = cpu_to_be32(pdn);
839 	cq_context->lkey            = cpu_to_be32(cq->buf.mr.ibmr.lkey);
840 	cq_context->cqn             = cpu_to_be32(cq->cqn);
841 
842 	if (mthca_is_memfree(dev)) {
843 		cq_context->ci_db    = cpu_to_be32(cq->set_ci_db_index);
844 		cq_context->state_db = cpu_to_be32(cq->arm_db_index);
845 	}
846 
847 	err = mthca_SW2HW_CQ(dev, mailbox, cq->cqn, &status);
848 	if (err) {
849 		mthca_warn(dev, "SW2HW_CQ failed (%d)\n", err);
850 		goto err_out_free_mr;
851 	}
852 
853 	if (status) {
854 		mthca_warn(dev, "SW2HW_CQ returned status 0x%02x\n",
855 			   status);
856 		err = -EINVAL;
857 		goto err_out_free_mr;
858 	}
859 
860 	spin_lock_irq(&dev->cq_table.lock);
861 	if (mthca_array_set(&dev->cq_table.cq,
862 			    cq->cqn & (dev->limits.num_cqs - 1),
863 			    cq)) {
864 		spin_unlock_irq(&dev->cq_table.lock);
865 		goto err_out_free_mr;
866 	}
867 	spin_unlock_irq(&dev->cq_table.lock);
868 
869 	cq->cons_index = 0;
870 
871 	mthca_free_mailbox(dev, mailbox);
872 
873 	return 0;
874 
875 err_out_free_mr:
876 	if (cq->is_kernel)
877 		mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
878 
879 err_out_mailbox:
880 	mthca_free_mailbox(dev, mailbox);
881 
882 err_out_arm:
883 	if (cq->is_kernel && mthca_is_memfree(dev))
884 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
885 
886 err_out_ci:
887 	if (cq->is_kernel && mthca_is_memfree(dev))
888 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
889 
890 err_out_icm:
891 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
892 
893 err_out:
894 	mthca_free(&dev->cq_table.alloc, cq->cqn);
895 
896 	return err;
897 }
898 
899 void mthca_free_cq(struct mthca_dev *dev,
900 		   struct mthca_cq *cq)
901 {
902 	struct mthca_mailbox *mailbox;
903 	int err;
904 	u8 status;
905 
906 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
907 	if (IS_ERR(mailbox)) {
908 		mthca_warn(dev, "No memory for mailbox to free CQ.\n");
909 		return;
910 	}
911 
912 	err = mthca_HW2SW_CQ(dev, mailbox, cq->cqn, &status);
913 	if (err)
914 		mthca_warn(dev, "HW2SW_CQ failed (%d)\n", err);
915 	else if (status)
916 		mthca_warn(dev, "HW2SW_CQ returned status 0x%02x\n", status);
917 
918 	if (0) {
919 		__be32 *ctx = mailbox->buf;
920 		int j;
921 
922 		printk(KERN_ERR "context for CQN %x (cons index %x, next sw %d)\n",
923 		       cq->cqn, cq->cons_index,
924 		       cq->is_kernel ? !!next_cqe_sw(cq) : 0);
925 		for (j = 0; j < 16; ++j)
926 			printk(KERN_ERR "[%2x] %08x\n", j * 4, be32_to_cpu(ctx[j]));
927 	}
928 
929 	spin_lock_irq(&dev->cq_table.lock);
930 	mthca_array_clear(&dev->cq_table.cq,
931 			  cq->cqn & (dev->limits.num_cqs - 1));
932 	spin_unlock_irq(&dev->cq_table.lock);
933 
934 	if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
935 		synchronize_irq(dev->eq_table.eq[MTHCA_EQ_COMP].msi_x_vector);
936 	else
937 		synchronize_irq(dev->pdev->irq);
938 
939 	atomic_dec(&cq->refcount);
940 	wait_event(cq->wait, !atomic_read(&cq->refcount));
941 
942 	if (cq->is_kernel) {
943 		mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
944 		if (mthca_is_memfree(dev)) {
945 			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM,    cq->arm_db_index);
946 			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
947 		}
948 	}
949 
950 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
951 	mthca_free(&dev->cq_table.alloc, cq->cqn);
952 	mthca_free_mailbox(dev, mailbox);
953 }
954 
955 int __devinit mthca_init_cq_table(struct mthca_dev *dev)
956 {
957 	int err;
958 
959 	spin_lock_init(&dev->cq_table.lock);
960 
961 	err = mthca_alloc_init(&dev->cq_table.alloc,
962 			       dev->limits.num_cqs,
963 			       (1 << 24) - 1,
964 			       dev->limits.reserved_cqs);
965 	if (err)
966 		return err;
967 
968 	err = mthca_array_init(&dev->cq_table.cq,
969 			       dev->limits.num_cqs);
970 	if (err)
971 		mthca_alloc_cleanup(&dev->cq_table.alloc);
972 
973 	return err;
974 }
975 
976 void mthca_cleanup_cq_table(struct mthca_dev *dev)
977 {
978 	mthca_array_cleanup(&dev->cq_table.cq, dev->limits.num_cqs);
979 	mthca_alloc_cleanup(&dev->cq_table.alloc);
980 }
981