xref: /linux/drivers/infiniband/hw/mthca/mthca_cq.c (revision c537b994505099b7197e7d3125b942ecbcc51eb6)
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/hardirq.h>
40 
41 #include <asm/io.h>
42 
43 #include <rdma/ib_pack.h>
44 
45 #include "mthca_dev.h"
46 #include "mthca_cmd.h"
47 #include "mthca_memfree.h"
48 
49 enum {
50 	MTHCA_MAX_DIRECT_CQ_SIZE = 4 * PAGE_SIZE
51 };
52 
53 enum {
54 	MTHCA_CQ_ENTRY_SIZE = 0x20
55 };
56 
57 enum {
58 	MTHCA_ATOMIC_BYTE_LEN = 8
59 };
60 
61 /*
62  * Must be packed because start is 64 bits but only aligned to 32 bits.
63  */
64 struct mthca_cq_context {
65 	__be32 flags;
66 	__be64 start;
67 	__be32 logsize_usrpage;
68 	__be32 error_eqn;	/* Tavor only */
69 	__be32 comp_eqn;
70 	__be32 pd;
71 	__be32 lkey;
72 	__be32 last_notified_index;
73 	__be32 solicit_producer_index;
74 	__be32 consumer_index;
75 	__be32 producer_index;
76 	__be32 cqn;
77 	__be32 ci_db;		/* Arbel only */
78 	__be32 state_db;	/* Arbel only */
79 	u32    reserved;
80 } __attribute__((packed));
81 
82 #define MTHCA_CQ_STATUS_OK          ( 0 << 28)
83 #define MTHCA_CQ_STATUS_OVERFLOW    ( 9 << 28)
84 #define MTHCA_CQ_STATUS_WRITE_FAIL  (10 << 28)
85 #define MTHCA_CQ_FLAG_TR            ( 1 << 18)
86 #define MTHCA_CQ_FLAG_OI            ( 1 << 17)
87 #define MTHCA_CQ_STATE_DISARMED     ( 0 <<  8)
88 #define MTHCA_CQ_STATE_ARMED        ( 1 <<  8)
89 #define MTHCA_CQ_STATE_ARMED_SOL    ( 4 <<  8)
90 #define MTHCA_EQ_STATE_FIRED        (10 <<  8)
91 
92 enum {
93 	MTHCA_ERROR_CQE_OPCODE_MASK = 0xfe
94 };
95 
96 enum {
97 	SYNDROME_LOCAL_LENGTH_ERR 	 = 0x01,
98 	SYNDROME_LOCAL_QP_OP_ERR  	 = 0x02,
99 	SYNDROME_LOCAL_EEC_OP_ERR 	 = 0x03,
100 	SYNDROME_LOCAL_PROT_ERR   	 = 0x04,
101 	SYNDROME_WR_FLUSH_ERR     	 = 0x05,
102 	SYNDROME_MW_BIND_ERR      	 = 0x06,
103 	SYNDROME_BAD_RESP_ERR     	 = 0x10,
104 	SYNDROME_LOCAL_ACCESS_ERR 	 = 0x11,
105 	SYNDROME_REMOTE_INVAL_REQ_ERR 	 = 0x12,
106 	SYNDROME_REMOTE_ACCESS_ERR 	 = 0x13,
107 	SYNDROME_REMOTE_OP_ERR     	 = 0x14,
108 	SYNDROME_RETRY_EXC_ERR 		 = 0x15,
109 	SYNDROME_RNR_RETRY_EXC_ERR 	 = 0x16,
110 	SYNDROME_LOCAL_RDD_VIOL_ERR 	 = 0x20,
111 	SYNDROME_REMOTE_INVAL_RD_REQ_ERR = 0x21,
112 	SYNDROME_REMOTE_ABORTED_ERR 	 = 0x22,
113 	SYNDROME_INVAL_EECN_ERR 	 = 0x23,
114 	SYNDROME_INVAL_EEC_STATE_ERR 	 = 0x24
115 };
116 
117 struct mthca_cqe {
118 	__be32 my_qpn;
119 	__be32 my_ee;
120 	__be32 rqpn;
121 	__be16 sl_g_mlpath;
122 	__be16 rlid;
123 	__be32 imm_etype_pkey_eec;
124 	__be32 byte_cnt;
125 	__be32 wqe;
126 	u8     opcode;
127 	u8     is_send;
128 	u8     reserved;
129 	u8     owner;
130 };
131 
132 struct mthca_err_cqe {
133 	__be32 my_qpn;
134 	u32    reserved1[3];
135 	u8     syndrome;
136 	u8     vendor_err;
137 	__be16 db_cnt;
138 	u32    reserved2;
139 	__be32 wqe;
140 	u8     opcode;
141 	u8     reserved3[2];
142 	u8     owner;
143 };
144 
145 #define MTHCA_CQ_ENTRY_OWNER_SW      (0 << 7)
146 #define MTHCA_CQ_ENTRY_OWNER_HW      (1 << 7)
147 
148 #define MTHCA_TAVOR_CQ_DB_INC_CI       (1 << 24)
149 #define MTHCA_TAVOR_CQ_DB_REQ_NOT      (2 << 24)
150 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL  (3 << 24)
151 #define MTHCA_TAVOR_CQ_DB_SET_CI       (4 << 24)
152 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_MULT (5 << 24)
153 
154 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL  (1 << 24)
155 #define MTHCA_ARBEL_CQ_DB_REQ_NOT      (2 << 24)
156 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_MULT (3 << 24)
157 
158 static inline struct mthca_cqe *get_cqe_from_buf(struct mthca_cq_buf *buf,
159 						 int entry)
160 {
161 	if (buf->is_direct)
162 		return buf->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
163 	else
164 		return buf->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
165 			+ (entry * MTHCA_CQ_ENTRY_SIZE) % PAGE_SIZE;
166 }
167 
168 static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
169 {
170 	return get_cqe_from_buf(&cq->buf, entry);
171 }
172 
173 static inline struct mthca_cqe *cqe_sw(struct mthca_cqe *cqe)
174 {
175 	return MTHCA_CQ_ENTRY_OWNER_HW & cqe->owner ? NULL : cqe;
176 }
177 
178 static inline struct mthca_cqe *next_cqe_sw(struct mthca_cq *cq)
179 {
180 	return cqe_sw(get_cqe(cq, cq->cons_index & cq->ibcq.cqe));
181 }
182 
183 static inline void set_cqe_hw(struct mthca_cqe *cqe)
184 {
185 	cqe->owner = MTHCA_CQ_ENTRY_OWNER_HW;
186 }
187 
188 static void dump_cqe(struct mthca_dev *dev, void *cqe_ptr)
189 {
190 	__be32 *cqe = cqe_ptr;
191 
192 	(void) cqe;	/* avoid warning if mthca_dbg compiled away... */
193 	mthca_dbg(dev, "CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
194 		  be32_to_cpu(cqe[0]), be32_to_cpu(cqe[1]), be32_to_cpu(cqe[2]),
195 		  be32_to_cpu(cqe[3]), be32_to_cpu(cqe[4]), be32_to_cpu(cqe[5]),
196 		  be32_to_cpu(cqe[6]), be32_to_cpu(cqe[7]));
197 }
198 
199 /*
200  * incr is ignored in native Arbel (mem-free) mode, so cq->cons_index
201  * should be correct before calling update_cons_index().
202  */
203 static inline void update_cons_index(struct mthca_dev *dev, struct mthca_cq *cq,
204 				     int incr)
205 {
206 	__be32 doorbell[2];
207 
208 	if (mthca_is_memfree(dev)) {
209 		*cq->set_ci_db = cpu_to_be32(cq->cons_index);
210 		wmb();
211 	} else {
212 		doorbell[0] = cpu_to_be32(MTHCA_TAVOR_CQ_DB_INC_CI | cq->cqn);
213 		doorbell[1] = cpu_to_be32(incr - 1);
214 
215 		mthca_write64(doorbell,
216 			      dev->kar + MTHCA_CQ_DOORBELL,
217 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
218 		/*
219 		 * Make sure doorbells don't leak out of CQ spinlock
220 		 * and reach the HCA out of order:
221 		 */
222 		mmiowb();
223 	}
224 }
225 
226 void mthca_cq_completion(struct mthca_dev *dev, u32 cqn)
227 {
228 	struct mthca_cq *cq;
229 
230 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
231 
232 	if (!cq) {
233 		mthca_warn(dev, "Completion event for bogus CQ %08x\n", cqn);
234 		return;
235 	}
236 
237 	++cq->arm_sn;
238 
239 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
240 }
241 
242 void mthca_cq_event(struct mthca_dev *dev, u32 cqn,
243 		    enum ib_event_type event_type)
244 {
245 	struct mthca_cq *cq;
246 	struct ib_event event;
247 
248 	spin_lock(&dev->cq_table.lock);
249 
250 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
251 	if (cq)
252 		++cq->refcount;
253 
254 	spin_unlock(&dev->cq_table.lock);
255 
256 	if (!cq) {
257 		mthca_warn(dev, "Async event for bogus CQ %08x\n", cqn);
258 		return;
259 	}
260 
261 	event.device      = &dev->ib_dev;
262 	event.event       = event_type;
263 	event.element.cq  = &cq->ibcq;
264 	if (cq->ibcq.event_handler)
265 		cq->ibcq.event_handler(&event, cq->ibcq.cq_context);
266 
267 	spin_lock(&dev->cq_table.lock);
268 	if (!--cq->refcount)
269 		wake_up(&cq->wait);
270 	spin_unlock(&dev->cq_table.lock);
271 }
272 
273 static inline int is_recv_cqe(struct mthca_cqe *cqe)
274 {
275 	if ((cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
276 	    MTHCA_ERROR_CQE_OPCODE_MASK)
277 		return !(cqe->opcode & 0x01);
278 	else
279 		return !(cqe->is_send & 0x80);
280 }
281 
282 void mthca_cq_clean(struct mthca_dev *dev, struct mthca_cq *cq, u32 qpn,
283 		    struct mthca_srq *srq)
284 {
285 	struct mthca_cqe *cqe;
286 	u32 prod_index;
287 	int nfreed = 0;
288 
289 	spin_lock_irq(&cq->lock);
290 
291 	/*
292 	 * First we need to find the current producer index, so we
293 	 * know where to start cleaning from.  It doesn't matter if HW
294 	 * adds new entries after this loop -- the QP we're worried
295 	 * about is already in RESET, so the new entries won't come
296 	 * from our QP and therefore don't need to be checked.
297 	 */
298 	for (prod_index = cq->cons_index;
299 	     cqe_sw(get_cqe(cq, prod_index & cq->ibcq.cqe));
300 	     ++prod_index)
301 		if (prod_index == cq->cons_index + cq->ibcq.cqe)
302 			break;
303 
304 	if (0)
305 		mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
306 			  qpn, cq->cqn, cq->cons_index, prod_index);
307 
308 	/*
309 	 * Now sweep backwards through the CQ, removing CQ entries
310 	 * that match our QP by copying older entries on top of them.
311 	 */
312 	while ((int) --prod_index - (int) cq->cons_index >= 0) {
313 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
314 		if (cqe->my_qpn == cpu_to_be32(qpn)) {
315 			if (srq && is_recv_cqe(cqe))
316 				mthca_free_srq_wqe(srq, be32_to_cpu(cqe->wqe));
317 			++nfreed;
318 		} else if (nfreed)
319 			memcpy(get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe),
320 			       cqe, MTHCA_CQ_ENTRY_SIZE);
321 	}
322 
323 	if (nfreed) {
324 		wmb();
325 		cq->cons_index += nfreed;
326 		update_cons_index(dev, cq, nfreed);
327 	}
328 
329 	spin_unlock_irq(&cq->lock);
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 = &(*cur_qp)->ibqp;
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 		s32 wqe;
554 		wq = &(*cur_qp)->rq;
555 		wqe = be32_to_cpu(cqe->wqe);
556 		wqe_index = wqe >> wq->wqe_shift;
557 		/*
558 		 * WQE addr == base - 1 might be reported in receive completion
559 		 * with error instead of (rq size - 1) by Sinai FW 1.0.800 and
560 		 * Arbel FW 5.1.400.  This bug should be fixed in later FW revs.
561 		 */
562 		if (unlikely(wqe_index < 0))
563 			wqe_index = wq->max - 1;
564 		entry->wr_id = (*cur_qp)->wrid[wqe_index];
565 	}
566 
567 	if (wq) {
568 		if (wq->last_comp < wqe_index)
569 			wq->tail += wqe_index - wq->last_comp;
570 		else
571 			wq->tail += wqe_index + wq->max - wq->last_comp;
572 
573 		wq->last_comp = wqe_index;
574 	}
575 
576 	if (is_error) {
577 		handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,
578 				 (struct mthca_err_cqe *) cqe,
579 				 entry, &free_cqe);
580 		goto out;
581 	}
582 
583 	if (is_send) {
584 		entry->wc_flags = 0;
585 		switch (cqe->opcode) {
586 		case MTHCA_OPCODE_RDMA_WRITE:
587 			entry->opcode    = IB_WC_RDMA_WRITE;
588 			break;
589 		case MTHCA_OPCODE_RDMA_WRITE_IMM:
590 			entry->opcode    = IB_WC_RDMA_WRITE;
591 			entry->wc_flags |= IB_WC_WITH_IMM;
592 			break;
593 		case MTHCA_OPCODE_SEND:
594 			entry->opcode    = IB_WC_SEND;
595 			break;
596 		case MTHCA_OPCODE_SEND_IMM:
597 			entry->opcode    = IB_WC_SEND;
598 			entry->wc_flags |= IB_WC_WITH_IMM;
599 			break;
600 		case MTHCA_OPCODE_RDMA_READ:
601 			entry->opcode    = IB_WC_RDMA_READ;
602 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
603 			break;
604 		case MTHCA_OPCODE_ATOMIC_CS:
605 			entry->opcode    = IB_WC_COMP_SWAP;
606 			entry->byte_len  = MTHCA_ATOMIC_BYTE_LEN;
607 			break;
608 		case MTHCA_OPCODE_ATOMIC_FA:
609 			entry->opcode    = IB_WC_FETCH_ADD;
610 			entry->byte_len  = MTHCA_ATOMIC_BYTE_LEN;
611 			break;
612 		case MTHCA_OPCODE_BIND_MW:
613 			entry->opcode    = IB_WC_BIND_MW;
614 			break;
615 		default:
616 			entry->opcode    = MTHCA_OPCODE_INVALID;
617 			break;
618 		}
619 	} else {
620 		entry->byte_len = be32_to_cpu(cqe->byte_cnt);
621 		switch (cqe->opcode & 0x1f) {
622 		case IB_OPCODE_SEND_LAST_WITH_IMMEDIATE:
623 		case IB_OPCODE_SEND_ONLY_WITH_IMMEDIATE:
624 			entry->wc_flags = IB_WC_WITH_IMM;
625 			entry->imm_data = cqe->imm_etype_pkey_eec;
626 			entry->opcode = IB_WC_RECV;
627 			break;
628 		case IB_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
629 		case IB_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
630 			entry->wc_flags = IB_WC_WITH_IMM;
631 			entry->imm_data = cqe->imm_etype_pkey_eec;
632 			entry->opcode = IB_WC_RECV_RDMA_WITH_IMM;
633 			break;
634 		default:
635 			entry->wc_flags = 0;
636 			entry->opcode = IB_WC_RECV;
637 			break;
638 		}
639 		entry->slid 	   = be16_to_cpu(cqe->rlid);
640 		entry->sl   	   = be16_to_cpu(cqe->sl_g_mlpath) >> 12;
641 		entry->src_qp 	   = be32_to_cpu(cqe->rqpn) & 0xffffff;
642 		entry->dlid_path_bits = be16_to_cpu(cqe->sl_g_mlpath) & 0x7f;
643 		entry->pkey_index  = be32_to_cpu(cqe->imm_etype_pkey_eec) >> 16;
644 		entry->wc_flags   |= be16_to_cpu(cqe->sl_g_mlpath) & 0x80 ?
645 					IB_WC_GRH : 0;
646 	}
647 
648 	entry->status = IB_WC_SUCCESS;
649 
650  out:
651 	if (likely(free_cqe)) {
652 		set_cqe_hw(cqe);
653 		++(*freed);
654 		++cq->cons_index;
655 	}
656 
657 	return err;
658 }
659 
660 int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
661 		  struct ib_wc *entry)
662 {
663 	struct mthca_dev *dev = to_mdev(ibcq->device);
664 	struct mthca_cq *cq = to_mcq(ibcq);
665 	struct mthca_qp *qp = NULL;
666 	unsigned long flags;
667 	int err = 0;
668 	int freed = 0;
669 	int npolled;
670 
671 	spin_lock_irqsave(&cq->lock, flags);
672 
673 	npolled = 0;
674 repoll:
675 	while (npolled < num_entries) {
676 		err = mthca_poll_one(dev, cq, &qp,
677 				     &freed, entry + npolled);
678 		if (err)
679 			break;
680 		++npolled;
681 	}
682 
683 	if (freed) {
684 		wmb();
685 		update_cons_index(dev, cq, freed);
686 	}
687 
688 	/*
689 	 * If a CQ resize is in progress and we discovered that the
690 	 * old buffer is empty, then peek in the new buffer, and if
691 	 * it's not empty, switch to the new buffer and continue
692 	 * polling there.
693 	 */
694 	if (unlikely(err == -EAGAIN && cq->resize_buf &&
695 		     cq->resize_buf->state == CQ_RESIZE_READY)) {
696 		/*
697 		 * In Tavor mode, the hardware keeps the producer
698 		 * index modulo the CQ size.  Since we might be making
699 		 * the CQ bigger, we need to mask our consumer index
700 		 * using the size of the old CQ buffer before looking
701 		 * in the new CQ buffer.
702 		 */
703 		if (!mthca_is_memfree(dev))
704 			cq->cons_index &= cq->ibcq.cqe;
705 
706 		if (cqe_sw(get_cqe_from_buf(&cq->resize_buf->buf,
707 					    cq->cons_index & cq->resize_buf->cqe))) {
708 			struct mthca_cq_buf tbuf;
709 			int tcqe;
710 
711 			tbuf         = cq->buf;
712 			tcqe         = cq->ibcq.cqe;
713 			cq->buf      = cq->resize_buf->buf;
714 			cq->ibcq.cqe = cq->resize_buf->cqe;
715 
716 			cq->resize_buf->buf   = tbuf;
717 			cq->resize_buf->cqe   = tcqe;
718 			cq->resize_buf->state = CQ_RESIZE_SWAPPED;
719 
720 			goto repoll;
721 		}
722 	}
723 
724 	spin_unlock_irqrestore(&cq->lock, flags);
725 
726 	return err == 0 || err == -EAGAIN ? npolled : err;
727 }
728 
729 int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify notify)
730 {
731 	__be32 doorbell[2];
732 
733 	doorbell[0] = cpu_to_be32((notify == IB_CQ_SOLICITED ?
734 				   MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL :
735 				   MTHCA_TAVOR_CQ_DB_REQ_NOT)      |
736 				  to_mcq(cq)->cqn);
737 	doorbell[1] = (__force __be32) 0xffffffff;
738 
739 	mthca_write64(doorbell,
740 		      to_mdev(cq->device)->kar + MTHCA_CQ_DOORBELL,
741 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(cq->device)->doorbell_lock));
742 
743 	return 0;
744 }
745 
746 int mthca_arbel_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify notify)
747 {
748 	struct mthca_cq *cq = to_mcq(ibcq);
749 	__be32 doorbell[2];
750 	u32 sn;
751 	__be32 ci;
752 
753 	sn = cq->arm_sn & 3;
754 	ci = cpu_to_be32(cq->cons_index);
755 
756 	doorbell[0] = ci;
757 	doorbell[1] = cpu_to_be32((cq->cqn << 8) | (2 << 5) | (sn << 3) |
758 				  (notify == IB_CQ_SOLICITED ? 1 : 2));
759 
760 	mthca_write_db_rec(doorbell, cq->arm_db);
761 
762 	/*
763 	 * Make sure that the doorbell record in host memory is
764 	 * written before ringing the doorbell via PCI MMIO.
765 	 */
766 	wmb();
767 
768 	doorbell[0] = cpu_to_be32((sn << 28)                       |
769 				  (notify == IB_CQ_SOLICITED ?
770 				   MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL :
771 				   MTHCA_ARBEL_CQ_DB_REQ_NOT)      |
772 				  cq->cqn);
773 	doorbell[1] = ci;
774 
775 	mthca_write64(doorbell,
776 		      to_mdev(ibcq->device)->kar + MTHCA_CQ_DOORBELL,
777 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->doorbell_lock));
778 
779 	return 0;
780 }
781 
782 int mthca_init_cq(struct mthca_dev *dev, int nent,
783 		  struct mthca_ucontext *ctx, u32 pdn,
784 		  struct mthca_cq *cq)
785 {
786 	struct mthca_mailbox *mailbox;
787 	struct mthca_cq_context *cq_context;
788 	int err = -ENOMEM;
789 	u8 status;
790 
791 	cq->ibcq.cqe  = nent - 1;
792 	cq->is_kernel = !ctx;
793 
794 	cq->cqn = mthca_alloc(&dev->cq_table.alloc);
795 	if (cq->cqn == -1)
796 		return -ENOMEM;
797 
798 	if (mthca_is_memfree(dev)) {
799 		err = mthca_table_get(dev, dev->cq_table.table, cq->cqn);
800 		if (err)
801 			goto err_out;
802 
803 		if (cq->is_kernel) {
804 			cq->arm_sn = 1;
805 
806 			err = -ENOMEM;
807 
808 			cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
809 							     cq->cqn, &cq->set_ci_db);
810 			if (cq->set_ci_db_index < 0)
811 				goto err_out_icm;
812 
813 			cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
814 							  cq->cqn, &cq->arm_db);
815 			if (cq->arm_db_index < 0)
816 				goto err_out_ci;
817 		}
818 	}
819 
820 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
821 	if (IS_ERR(mailbox))
822 		goto err_out_arm;
823 
824 	cq_context = mailbox->buf;
825 
826 	if (cq->is_kernel) {
827 		err = mthca_alloc_cq_buf(dev, &cq->buf, nent);
828 		if (err)
829 			goto err_out_mailbox;
830 	}
831 
832 	spin_lock_init(&cq->lock);
833 	cq->refcount = 1;
834 	init_waitqueue_head(&cq->wait);
835 	mutex_init(&cq->mutex);
836 
837 	memset(cq_context, 0, sizeof *cq_context);
838 	cq_context->flags           = cpu_to_be32(MTHCA_CQ_STATUS_OK      |
839 						  MTHCA_CQ_STATE_DISARMED |
840 						  MTHCA_CQ_FLAG_TR);
841 	cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24);
842 	if (ctx)
843 		cq_context->logsize_usrpage |= cpu_to_be32(ctx->uar.index);
844 	else
845 		cq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
846 	cq_context->error_eqn       = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
847 	cq_context->comp_eqn        = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
848 	cq_context->pd              = cpu_to_be32(pdn);
849 	cq_context->lkey            = cpu_to_be32(cq->buf.mr.ibmr.lkey);
850 	cq_context->cqn             = cpu_to_be32(cq->cqn);
851 
852 	if (mthca_is_memfree(dev)) {
853 		cq_context->ci_db    = cpu_to_be32(cq->set_ci_db_index);
854 		cq_context->state_db = cpu_to_be32(cq->arm_db_index);
855 	}
856 
857 	err = mthca_SW2HW_CQ(dev, mailbox, cq->cqn, &status);
858 	if (err) {
859 		mthca_warn(dev, "SW2HW_CQ failed (%d)\n", err);
860 		goto err_out_free_mr;
861 	}
862 
863 	if (status) {
864 		mthca_warn(dev, "SW2HW_CQ returned status 0x%02x\n",
865 			   status);
866 		err = -EINVAL;
867 		goto err_out_free_mr;
868 	}
869 
870 	spin_lock_irq(&dev->cq_table.lock);
871 	if (mthca_array_set(&dev->cq_table.cq,
872 			    cq->cqn & (dev->limits.num_cqs - 1),
873 			    cq)) {
874 		spin_unlock_irq(&dev->cq_table.lock);
875 		goto err_out_free_mr;
876 	}
877 	spin_unlock_irq(&dev->cq_table.lock);
878 
879 	cq->cons_index = 0;
880 
881 	mthca_free_mailbox(dev, mailbox);
882 
883 	return 0;
884 
885 err_out_free_mr:
886 	if (cq->is_kernel)
887 		mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
888 
889 err_out_mailbox:
890 	mthca_free_mailbox(dev, mailbox);
891 
892 err_out_arm:
893 	if (cq->is_kernel && mthca_is_memfree(dev))
894 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
895 
896 err_out_ci:
897 	if (cq->is_kernel && mthca_is_memfree(dev))
898 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
899 
900 err_out_icm:
901 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
902 
903 err_out:
904 	mthca_free(&dev->cq_table.alloc, cq->cqn);
905 
906 	return err;
907 }
908 
909 static inline int get_cq_refcount(struct mthca_dev *dev, struct mthca_cq *cq)
910 {
911 	int c;
912 
913 	spin_lock_irq(&dev->cq_table.lock);
914 	c = cq->refcount;
915 	spin_unlock_irq(&dev->cq_table.lock);
916 
917 	return c;
918 }
919 
920 void mthca_free_cq(struct mthca_dev *dev,
921 		   struct mthca_cq *cq)
922 {
923 	struct mthca_mailbox *mailbox;
924 	int err;
925 	u8 status;
926 
927 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
928 	if (IS_ERR(mailbox)) {
929 		mthca_warn(dev, "No memory for mailbox to free CQ.\n");
930 		return;
931 	}
932 
933 	err = mthca_HW2SW_CQ(dev, mailbox, cq->cqn, &status);
934 	if (err)
935 		mthca_warn(dev, "HW2SW_CQ failed (%d)\n", err);
936 	else if (status)
937 		mthca_warn(dev, "HW2SW_CQ returned status 0x%02x\n", status);
938 
939 	if (0) {
940 		__be32 *ctx = mailbox->buf;
941 		int j;
942 
943 		printk(KERN_ERR "context for CQN %x (cons index %x, next sw %d)\n",
944 		       cq->cqn, cq->cons_index,
945 		       cq->is_kernel ? !!next_cqe_sw(cq) : 0);
946 		for (j = 0; j < 16; ++j)
947 			printk(KERN_ERR "[%2x] %08x\n", j * 4, be32_to_cpu(ctx[j]));
948 	}
949 
950 	spin_lock_irq(&dev->cq_table.lock);
951 	mthca_array_clear(&dev->cq_table.cq,
952 			  cq->cqn & (dev->limits.num_cqs - 1));
953 	--cq->refcount;
954 	spin_unlock_irq(&dev->cq_table.lock);
955 
956 	if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
957 		synchronize_irq(dev->eq_table.eq[MTHCA_EQ_COMP].msi_x_vector);
958 	else
959 		synchronize_irq(dev->pdev->irq);
960 
961 	wait_event(cq->wait, !get_cq_refcount(dev, cq));
962 
963 	if (cq->is_kernel) {
964 		mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
965 		if (mthca_is_memfree(dev)) {
966 			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM,    cq->arm_db_index);
967 			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
968 		}
969 	}
970 
971 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
972 	mthca_free(&dev->cq_table.alloc, cq->cqn);
973 	mthca_free_mailbox(dev, mailbox);
974 }
975 
976 int mthca_init_cq_table(struct mthca_dev *dev)
977 {
978 	int err;
979 
980 	spin_lock_init(&dev->cq_table.lock);
981 
982 	err = mthca_alloc_init(&dev->cq_table.alloc,
983 			       dev->limits.num_cqs,
984 			       (1 << 24) - 1,
985 			       dev->limits.reserved_cqs);
986 	if (err)
987 		return err;
988 
989 	err = mthca_array_init(&dev->cq_table.cq,
990 			       dev->limits.num_cqs);
991 	if (err)
992 		mthca_alloc_cleanup(&dev->cq_table.alloc);
993 
994 	return err;
995 }
996 
997 void mthca_cleanup_cq_table(struct mthca_dev *dev)
998 {
999 	mthca_array_cleanup(&dev->cq_table.cq, dev->limits.num_cqs);
1000 	mthca_alloc_cleanup(&dev->cq_table.alloc);
1001 }
1002