xref: /openbmc/linux/drivers/infiniband/hw/mthca/mthca_cq.c (revision 96de0e252cedffad61b3cb5e05662c591898e69a)
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 #include <linux/sched.h>
41 
42 #include <asm/io.h>
43 
44 #include <rdma/ib_pack.h>
45 
46 #include "mthca_dev.h"
47 #include "mthca_cmd.h"
48 #include "mthca_memfree.h"
49 
50 enum {
51 	MTHCA_MAX_DIRECT_CQ_SIZE = 4 * PAGE_SIZE
52 };
53 
54 enum {
55 	MTHCA_CQ_ENTRY_SIZE = 0x20
56 };
57 
58 enum {
59 	MTHCA_ATOMIC_BYTE_LEN = 8
60 };
61 
62 /*
63  * Must be packed because start is 64 bits but only aligned to 32 bits.
64  */
65 struct mthca_cq_context {
66 	__be32 flags;
67 	__be64 start;
68 	__be32 logsize_usrpage;
69 	__be32 error_eqn;	/* Tavor only */
70 	__be32 comp_eqn;
71 	__be32 pd;
72 	__be32 lkey;
73 	__be32 last_notified_index;
74 	__be32 solicit_producer_index;
75 	__be32 consumer_index;
76 	__be32 producer_index;
77 	__be32 cqn;
78 	__be32 ci_db;		/* Arbel only */
79 	__be32 state_db;	/* Arbel only */
80 	u32    reserved;
81 } __attribute__((packed));
82 
83 #define MTHCA_CQ_STATUS_OK          ( 0 << 28)
84 #define MTHCA_CQ_STATUS_OVERFLOW    ( 9 << 28)
85 #define MTHCA_CQ_STATUS_WRITE_FAIL  (10 << 28)
86 #define MTHCA_CQ_FLAG_TR            ( 1 << 18)
87 #define MTHCA_CQ_FLAG_OI            ( 1 << 17)
88 #define MTHCA_CQ_STATE_DISARMED     ( 0 <<  8)
89 #define MTHCA_CQ_STATE_ARMED        ( 1 <<  8)
90 #define MTHCA_CQ_STATE_ARMED_SOL    ( 4 <<  8)
91 #define MTHCA_EQ_STATE_FIRED        (10 <<  8)
92 
93 enum {
94 	MTHCA_ERROR_CQE_OPCODE_MASK = 0xfe
95 };
96 
97 enum {
98 	SYNDROME_LOCAL_LENGTH_ERR 	 = 0x01,
99 	SYNDROME_LOCAL_QP_OP_ERR  	 = 0x02,
100 	SYNDROME_LOCAL_EEC_OP_ERR 	 = 0x03,
101 	SYNDROME_LOCAL_PROT_ERR   	 = 0x04,
102 	SYNDROME_WR_FLUSH_ERR     	 = 0x05,
103 	SYNDROME_MW_BIND_ERR      	 = 0x06,
104 	SYNDROME_BAD_RESP_ERR     	 = 0x10,
105 	SYNDROME_LOCAL_ACCESS_ERR 	 = 0x11,
106 	SYNDROME_REMOTE_INVAL_REQ_ERR 	 = 0x12,
107 	SYNDROME_REMOTE_ACCESS_ERR 	 = 0x13,
108 	SYNDROME_REMOTE_OP_ERR     	 = 0x14,
109 	SYNDROME_RETRY_EXC_ERR 		 = 0x15,
110 	SYNDROME_RNR_RETRY_EXC_ERR 	 = 0x16,
111 	SYNDROME_LOCAL_RDD_VIOL_ERR 	 = 0x20,
112 	SYNDROME_REMOTE_INVAL_RD_REQ_ERR = 0x21,
113 	SYNDROME_REMOTE_ABORTED_ERR 	 = 0x22,
114 	SYNDROME_INVAL_EECN_ERR 	 = 0x23,
115 	SYNDROME_INVAL_EEC_STATE_ERR 	 = 0x24
116 };
117 
118 struct mthca_cqe {
119 	__be32 my_qpn;
120 	__be32 my_ee;
121 	__be32 rqpn;
122 	__be16 sl_g_mlpath;
123 	__be16 rlid;
124 	__be32 imm_etype_pkey_eec;
125 	__be32 byte_cnt;
126 	__be32 wqe;
127 	u8     opcode;
128 	u8     is_send;
129 	u8     reserved;
130 	u8     owner;
131 };
132 
133 struct mthca_err_cqe {
134 	__be32 my_qpn;
135 	u32    reserved1[3];
136 	u8     syndrome;
137 	u8     vendor_err;
138 	__be16 db_cnt;
139 	u32    reserved2;
140 	__be32 wqe;
141 	u8     opcode;
142 	u8     reserved3[2];
143 	u8     owner;
144 };
145 
146 #define MTHCA_CQ_ENTRY_OWNER_SW      (0 << 7)
147 #define MTHCA_CQ_ENTRY_OWNER_HW      (1 << 7)
148 
149 #define MTHCA_TAVOR_CQ_DB_INC_CI       (1 << 24)
150 #define MTHCA_TAVOR_CQ_DB_REQ_NOT      (2 << 24)
151 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL  (3 << 24)
152 #define MTHCA_TAVOR_CQ_DB_SET_CI       (4 << 24)
153 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_MULT (5 << 24)
154 
155 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL  (1 << 24)
156 #define MTHCA_ARBEL_CQ_DB_REQ_NOT      (2 << 24)
157 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_MULT (3 << 24)
158 
159 static inline struct mthca_cqe *get_cqe_from_buf(struct mthca_cq_buf *buf,
160 						 int entry)
161 {
162 	if (buf->is_direct)
163 		return buf->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
164 	else
165 		return buf->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
166 			+ (entry * MTHCA_CQ_ENTRY_SIZE) % PAGE_SIZE;
167 }
168 
169 static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
170 {
171 	return get_cqe_from_buf(&cq->buf, entry);
172 }
173 
174 static inline struct mthca_cqe *cqe_sw(struct mthca_cqe *cqe)
175 {
176 	return MTHCA_CQ_ENTRY_OWNER_HW & cqe->owner ? NULL : cqe;
177 }
178 
179 static inline struct mthca_cqe *next_cqe_sw(struct mthca_cq *cq)
180 {
181 	return cqe_sw(get_cqe(cq, cq->cons_index & cq->ibcq.cqe));
182 }
183 
184 static inline void set_cqe_hw(struct mthca_cqe *cqe)
185 {
186 	cqe->owner = MTHCA_CQ_ENTRY_OWNER_HW;
187 }
188 
189 static void dump_cqe(struct mthca_dev *dev, void *cqe_ptr)
190 {
191 	__be32 *cqe = cqe_ptr;
192 
193 	(void) cqe;	/* avoid warning if mthca_dbg compiled away... */
194 	mthca_dbg(dev, "CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
195 		  be32_to_cpu(cqe[0]), be32_to_cpu(cqe[1]), be32_to_cpu(cqe[2]),
196 		  be32_to_cpu(cqe[3]), be32_to_cpu(cqe[4]), be32_to_cpu(cqe[5]),
197 		  be32_to_cpu(cqe[6]), be32_to_cpu(cqe[7]));
198 }
199 
200 /*
201  * incr is ignored in native Arbel (mem-free) mode, so cq->cons_index
202  * should be correct before calling update_cons_index().
203  */
204 static inline void update_cons_index(struct mthca_dev *dev, struct mthca_cq *cq,
205 				     int incr)
206 {
207 	__be32 doorbell[2];
208 
209 	if (mthca_is_memfree(dev)) {
210 		*cq->set_ci_db = cpu_to_be32(cq->cons_index);
211 		wmb();
212 	} else {
213 		doorbell[0] = cpu_to_be32(MTHCA_TAVOR_CQ_DB_INC_CI | cq->cqn);
214 		doorbell[1] = cpu_to_be32(incr - 1);
215 
216 		mthca_write64(doorbell,
217 			      dev->kar + MTHCA_CQ_DOORBELL,
218 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
219 		/*
220 		 * Make sure doorbells don't leak out of CQ spinlock
221 		 * and reach the HCA out of order:
222 		 */
223 		mmiowb();
224 	}
225 }
226 
227 void mthca_cq_completion(struct mthca_dev *dev, u32 cqn)
228 {
229 	struct mthca_cq *cq;
230 
231 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
232 
233 	if (!cq) {
234 		mthca_warn(dev, "Completion event for bogus CQ %08x\n", cqn);
235 		return;
236 	}
237 
238 	++cq->arm_sn;
239 
240 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
241 }
242 
243 void mthca_cq_event(struct mthca_dev *dev, u32 cqn,
244 		    enum ib_event_type event_type)
245 {
246 	struct mthca_cq *cq;
247 	struct ib_event event;
248 
249 	spin_lock(&dev->cq_table.lock);
250 
251 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
252 	if (cq)
253 		++cq->refcount;
254 
255 	spin_unlock(&dev->cq_table.lock);
256 
257 	if (!cq) {
258 		mthca_warn(dev, "Async event for bogus CQ %08x\n", cqn);
259 		return;
260 	}
261 
262 	event.device      = &dev->ib_dev;
263 	event.event       = event_type;
264 	event.element.cq  = &cq->ibcq;
265 	if (cq->ibcq.event_handler)
266 		cq->ibcq.event_handler(&event, cq->ibcq.cq_context);
267 
268 	spin_lock(&dev->cq_table.lock);
269 	if (!--cq->refcount)
270 		wake_up(&cq->wait);
271 	spin_unlock(&dev->cq_table.lock);
272 }
273 
274 static inline int is_recv_cqe(struct mthca_cqe *cqe)
275 {
276 	if ((cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
277 	    MTHCA_ERROR_CQE_OPCODE_MASK)
278 		return !(cqe->opcode & 0x01);
279 	else
280 		return !(cqe->is_send & 0x80);
281 }
282 
283 void mthca_cq_clean(struct mthca_dev *dev, struct mthca_cq *cq, u32 qpn,
284 		    struct mthca_srq *srq)
285 {
286 	struct mthca_cqe *cqe;
287 	u32 prod_index;
288 	int i, nfreed = 0;
289 
290 	spin_lock_irq(&cq->lock);
291 
292 	/*
293 	 * First we need to find the current producer index, so we
294 	 * know where to start cleaning from.  It doesn't matter if HW
295 	 * adds new entries after this loop -- the QP we're worried
296 	 * about is already in RESET, so the new entries won't come
297 	 * from our QP and therefore don't need to be checked.
298 	 */
299 	for (prod_index = cq->cons_index;
300 	     cqe_sw(get_cqe(cq, prod_index & cq->ibcq.cqe));
301 	     ++prod_index)
302 		if (prod_index == cq->cons_index + cq->ibcq.cqe)
303 			break;
304 
305 	if (0)
306 		mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
307 			  qpn, cq->cqn, cq->cons_index, prod_index);
308 
309 	/*
310 	 * Now sweep backwards through the CQ, removing CQ entries
311 	 * that match our QP by copying older entries on top of them.
312 	 */
313 	while ((int) --prod_index - (int) cq->cons_index >= 0) {
314 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
315 		if (cqe->my_qpn == cpu_to_be32(qpn)) {
316 			if (srq && is_recv_cqe(cqe))
317 				mthca_free_srq_wqe(srq, be32_to_cpu(cqe->wqe));
318 			++nfreed;
319 		} else if (nfreed)
320 			memcpy(get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe),
321 			       cqe, MTHCA_CQ_ENTRY_SIZE);
322 	}
323 
324 	if (nfreed) {
325 		for (i = 0; i < nfreed; ++i)
326 			set_cqe_hw(get_cqe(cq, (cq->cons_index + i) & cq->ibcq.cqe));
327 		wmb();
328 		cq->cons_index += nfreed;
329 		update_cons_index(dev, cq, nfreed);
330 	}
331 
332 	spin_unlock_irq(&cq->lock);
333 }
334 
335 void mthca_cq_resize_copy_cqes(struct mthca_cq *cq)
336 {
337 	int i;
338 
339 	/*
340 	 * In Tavor mode, the hardware keeps the consumer and producer
341 	 * indices mod the CQ size.  Since we might be making the CQ
342 	 * bigger, we need to deal with the case where the producer
343 	 * index wrapped around before the CQ was resized.
344 	 */
345 	if (!mthca_is_memfree(to_mdev(cq->ibcq.device)) &&
346 	    cq->ibcq.cqe < cq->resize_buf->cqe) {
347 		cq->cons_index &= cq->ibcq.cqe;
348 		if (cqe_sw(get_cqe(cq, cq->ibcq.cqe)))
349 			cq->cons_index -= cq->ibcq.cqe + 1;
350 	}
351 
352 	for (i = cq->cons_index; cqe_sw(get_cqe(cq, i & cq->ibcq.cqe)); ++i)
353 		memcpy(get_cqe_from_buf(&cq->resize_buf->buf,
354 					i & cq->resize_buf->cqe),
355 		       get_cqe(cq, i & cq->ibcq.cqe), MTHCA_CQ_ENTRY_SIZE);
356 }
357 
358 int mthca_alloc_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int nent)
359 {
360 	int ret;
361 	int i;
362 
363 	ret = mthca_buf_alloc(dev, nent * MTHCA_CQ_ENTRY_SIZE,
364 			      MTHCA_MAX_DIRECT_CQ_SIZE,
365 			      &buf->queue, &buf->is_direct,
366 			      &dev->driver_pd, 1, &buf->mr);
367 	if (ret)
368 		return ret;
369 
370 	for (i = 0; i < nent; ++i)
371 		set_cqe_hw(get_cqe_from_buf(buf, i));
372 
373 	return 0;
374 }
375 
376 void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int cqe)
377 {
378 	mthca_buf_free(dev, (cqe + 1) * MTHCA_CQ_ENTRY_SIZE, &buf->queue,
379 		       buf->is_direct, &buf->mr);
380 }
381 
382 static void handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
383 			     struct mthca_qp *qp, int wqe_index, int is_send,
384 			     struct mthca_err_cqe *cqe,
385 			     struct ib_wc *entry, int *free_cqe)
386 {
387 	int dbd;
388 	__be32 new_wqe;
389 
390 	if (cqe->syndrome == SYNDROME_LOCAL_QP_OP_ERR) {
391 		mthca_dbg(dev, "local QP operation err "
392 			  "(QPN %06x, WQE @ %08x, CQN %06x, index %d)\n",
393 			  be32_to_cpu(cqe->my_qpn), be32_to_cpu(cqe->wqe),
394 			  cq->cqn, cq->cons_index);
395 		dump_cqe(dev, cqe);
396 	}
397 
398 	/*
399 	 * For completions in error, only work request ID, status, vendor error
400 	 * (and freed resource count for RD) have to be set.
401 	 */
402 	switch (cqe->syndrome) {
403 	case SYNDROME_LOCAL_LENGTH_ERR:
404 		entry->status = IB_WC_LOC_LEN_ERR;
405 		break;
406 	case SYNDROME_LOCAL_QP_OP_ERR:
407 		entry->status = IB_WC_LOC_QP_OP_ERR;
408 		break;
409 	case SYNDROME_LOCAL_EEC_OP_ERR:
410 		entry->status = IB_WC_LOC_EEC_OP_ERR;
411 		break;
412 	case SYNDROME_LOCAL_PROT_ERR:
413 		entry->status = IB_WC_LOC_PROT_ERR;
414 		break;
415 	case SYNDROME_WR_FLUSH_ERR:
416 		entry->status = IB_WC_WR_FLUSH_ERR;
417 		break;
418 	case SYNDROME_MW_BIND_ERR:
419 		entry->status = IB_WC_MW_BIND_ERR;
420 		break;
421 	case SYNDROME_BAD_RESP_ERR:
422 		entry->status = IB_WC_BAD_RESP_ERR;
423 		break;
424 	case SYNDROME_LOCAL_ACCESS_ERR:
425 		entry->status = IB_WC_LOC_ACCESS_ERR;
426 		break;
427 	case SYNDROME_REMOTE_INVAL_REQ_ERR:
428 		entry->status = IB_WC_REM_INV_REQ_ERR;
429 		break;
430 	case SYNDROME_REMOTE_ACCESS_ERR:
431 		entry->status = IB_WC_REM_ACCESS_ERR;
432 		break;
433 	case SYNDROME_REMOTE_OP_ERR:
434 		entry->status = IB_WC_REM_OP_ERR;
435 		break;
436 	case SYNDROME_RETRY_EXC_ERR:
437 		entry->status = IB_WC_RETRY_EXC_ERR;
438 		break;
439 	case SYNDROME_RNR_RETRY_EXC_ERR:
440 		entry->status = IB_WC_RNR_RETRY_EXC_ERR;
441 		break;
442 	case SYNDROME_LOCAL_RDD_VIOL_ERR:
443 		entry->status = IB_WC_LOC_RDD_VIOL_ERR;
444 		break;
445 	case SYNDROME_REMOTE_INVAL_RD_REQ_ERR:
446 		entry->status = IB_WC_REM_INV_RD_REQ_ERR;
447 		break;
448 	case SYNDROME_REMOTE_ABORTED_ERR:
449 		entry->status = IB_WC_REM_ABORT_ERR;
450 		break;
451 	case SYNDROME_INVAL_EECN_ERR:
452 		entry->status = IB_WC_INV_EECN_ERR;
453 		break;
454 	case SYNDROME_INVAL_EEC_STATE_ERR:
455 		entry->status = IB_WC_INV_EEC_STATE_ERR;
456 		break;
457 	default:
458 		entry->status = IB_WC_GENERAL_ERR;
459 		break;
460 	}
461 
462 	entry->vendor_err = cqe->vendor_err;
463 
464 	/*
465 	 * Mem-free HCAs always generate one CQE per WQE, even in the
466 	 * error case, so we don't have to check the doorbell count, etc.
467 	 */
468 	if (mthca_is_memfree(dev))
469 		return;
470 
471 	mthca_free_err_wqe(dev, qp, is_send, wqe_index, &dbd, &new_wqe);
472 
473 	/*
474 	 * If we're at the end of the WQE chain, or we've used up our
475 	 * doorbell count, free the CQE.  Otherwise just update it for
476 	 * the next poll operation.
477 	 */
478 	if (!(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
479 		return;
480 
481 	cqe->db_cnt   = cpu_to_be16(be16_to_cpu(cqe->db_cnt) - dbd);
482 	cqe->wqe      = new_wqe;
483 	cqe->syndrome = SYNDROME_WR_FLUSH_ERR;
484 
485 	*free_cqe = 0;
486 }
487 
488 static inline int mthca_poll_one(struct mthca_dev *dev,
489 				 struct mthca_cq *cq,
490 				 struct mthca_qp **cur_qp,
491 				 int *freed,
492 				 struct ib_wc *entry)
493 {
494 	struct mthca_wq *wq;
495 	struct mthca_cqe *cqe;
496 	int wqe_index;
497 	int is_error;
498 	int is_send;
499 	int free_cqe = 1;
500 	int err = 0;
501 
502 	cqe = next_cqe_sw(cq);
503 	if (!cqe)
504 		return -EAGAIN;
505 
506 	/*
507 	 * Make sure we read CQ entry contents after we've checked the
508 	 * ownership bit.
509 	 */
510 	rmb();
511 
512 	if (0) {
513 		mthca_dbg(dev, "%x/%d: CQE -> QPN %06x, WQE @ %08x\n",
514 			  cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
515 			  be32_to_cpu(cqe->wqe));
516 		dump_cqe(dev, cqe);
517 	}
518 
519 	is_error = (cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
520 		MTHCA_ERROR_CQE_OPCODE_MASK;
521 	is_send  = is_error ? cqe->opcode & 0x01 : cqe->is_send & 0x80;
522 
523 	if (!*cur_qp || be32_to_cpu(cqe->my_qpn) != (*cur_qp)->qpn) {
524 		/*
525 		 * We do not have to take the QP table lock here,
526 		 * because CQs will be locked while QPs are removed
527 		 * from the table.
528 		 */
529 		*cur_qp = mthca_array_get(&dev->qp_table.qp,
530 					  be32_to_cpu(cqe->my_qpn) &
531 					  (dev->limits.num_qps - 1));
532 		if (!*cur_qp) {
533 			mthca_warn(dev, "CQ entry for unknown QP %06x\n",
534 				   be32_to_cpu(cqe->my_qpn) & 0xffffff);
535 			err = -EINVAL;
536 			goto out;
537 		}
538 	}
539 
540 	entry->qp = &(*cur_qp)->ibqp;
541 
542 	if (is_send) {
543 		wq = &(*cur_qp)->sq;
544 		wqe_index = ((be32_to_cpu(cqe->wqe) - (*cur_qp)->send_wqe_offset)
545 			     >> wq->wqe_shift);
546 		entry->wr_id = (*cur_qp)->wrid[wqe_index +
547 					       (*cur_qp)->rq.max];
548 	} else if ((*cur_qp)->ibqp.srq) {
549 		struct mthca_srq *srq = to_msrq((*cur_qp)->ibqp.srq);
550 		u32 wqe = be32_to_cpu(cqe->wqe);
551 		wq = NULL;
552 		wqe_index = wqe >> srq->wqe_shift;
553 		entry->wr_id = srq->wrid[wqe_index];
554 		mthca_free_srq_wqe(srq, wqe);
555 	} else {
556 		s32 wqe;
557 		wq = &(*cur_qp)->rq;
558 		wqe = be32_to_cpu(cqe->wqe);
559 		wqe_index = wqe >> wq->wqe_shift;
560 		/*
561 		 * WQE addr == base - 1 might be reported in receive completion
562 		 * with error instead of (rq size - 1) by Sinai FW 1.0.800 and
563 		 * Arbel FW 5.1.400.  This bug should be fixed in later FW revs.
564 		 */
565 		if (unlikely(wqe_index < 0))
566 			wqe_index = wq->max - 1;
567 		entry->wr_id = (*cur_qp)->wrid[wqe_index];
568 	}
569 
570 	if (wq) {
571 		if (wq->last_comp < wqe_index)
572 			wq->tail += wqe_index - wq->last_comp;
573 		else
574 			wq->tail += wqe_index + wq->max - wq->last_comp;
575 
576 		wq->last_comp = wqe_index;
577 	}
578 
579 	if (is_error) {
580 		handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,
581 				 (struct mthca_err_cqe *) cqe,
582 				 entry, &free_cqe);
583 		goto out;
584 	}
585 
586 	if (is_send) {
587 		entry->wc_flags = 0;
588 		switch (cqe->opcode) {
589 		case MTHCA_OPCODE_RDMA_WRITE:
590 			entry->opcode    = IB_WC_RDMA_WRITE;
591 			break;
592 		case MTHCA_OPCODE_RDMA_WRITE_IMM:
593 			entry->opcode    = IB_WC_RDMA_WRITE;
594 			entry->wc_flags |= IB_WC_WITH_IMM;
595 			break;
596 		case MTHCA_OPCODE_SEND:
597 			entry->opcode    = IB_WC_SEND;
598 			break;
599 		case MTHCA_OPCODE_SEND_IMM:
600 			entry->opcode    = IB_WC_SEND;
601 			entry->wc_flags |= IB_WC_WITH_IMM;
602 			break;
603 		case MTHCA_OPCODE_RDMA_READ:
604 			entry->opcode    = IB_WC_RDMA_READ;
605 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
606 			break;
607 		case MTHCA_OPCODE_ATOMIC_CS:
608 			entry->opcode    = IB_WC_COMP_SWAP;
609 			entry->byte_len  = MTHCA_ATOMIC_BYTE_LEN;
610 			break;
611 		case MTHCA_OPCODE_ATOMIC_FA:
612 			entry->opcode    = IB_WC_FETCH_ADD;
613 			entry->byte_len  = MTHCA_ATOMIC_BYTE_LEN;
614 			break;
615 		case MTHCA_OPCODE_BIND_MW:
616 			entry->opcode    = IB_WC_BIND_MW;
617 			break;
618 		default:
619 			entry->opcode    = MTHCA_OPCODE_INVALID;
620 			break;
621 		}
622 	} else {
623 		entry->byte_len = be32_to_cpu(cqe->byte_cnt);
624 		switch (cqe->opcode & 0x1f) {
625 		case IB_OPCODE_SEND_LAST_WITH_IMMEDIATE:
626 		case IB_OPCODE_SEND_ONLY_WITH_IMMEDIATE:
627 			entry->wc_flags = IB_WC_WITH_IMM;
628 			entry->imm_data = cqe->imm_etype_pkey_eec;
629 			entry->opcode = IB_WC_RECV;
630 			break;
631 		case IB_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
632 		case IB_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
633 			entry->wc_flags = IB_WC_WITH_IMM;
634 			entry->imm_data = cqe->imm_etype_pkey_eec;
635 			entry->opcode = IB_WC_RECV_RDMA_WITH_IMM;
636 			break;
637 		default:
638 			entry->wc_flags = 0;
639 			entry->opcode = IB_WC_RECV;
640 			break;
641 		}
642 		entry->slid 	   = be16_to_cpu(cqe->rlid);
643 		entry->sl   	   = be16_to_cpu(cqe->sl_g_mlpath) >> 12;
644 		entry->src_qp 	   = be32_to_cpu(cqe->rqpn) & 0xffffff;
645 		entry->dlid_path_bits = be16_to_cpu(cqe->sl_g_mlpath) & 0x7f;
646 		entry->pkey_index  = be32_to_cpu(cqe->imm_etype_pkey_eec) >> 16;
647 		entry->wc_flags   |= be16_to_cpu(cqe->sl_g_mlpath) & 0x80 ?
648 					IB_WC_GRH : 0;
649 	}
650 
651 	entry->status = IB_WC_SUCCESS;
652 
653  out:
654 	if (likely(free_cqe)) {
655 		set_cqe_hw(cqe);
656 		++(*freed);
657 		++cq->cons_index;
658 	}
659 
660 	return err;
661 }
662 
663 int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
664 		  struct ib_wc *entry)
665 {
666 	struct mthca_dev *dev = to_mdev(ibcq->device);
667 	struct mthca_cq *cq = to_mcq(ibcq);
668 	struct mthca_qp *qp = NULL;
669 	unsigned long flags;
670 	int err = 0;
671 	int freed = 0;
672 	int npolled;
673 
674 	spin_lock_irqsave(&cq->lock, flags);
675 
676 	npolled = 0;
677 repoll:
678 	while (npolled < num_entries) {
679 		err = mthca_poll_one(dev, cq, &qp,
680 				     &freed, entry + npolled);
681 		if (err)
682 			break;
683 		++npolled;
684 	}
685 
686 	if (freed) {
687 		wmb();
688 		update_cons_index(dev, cq, freed);
689 	}
690 
691 	/*
692 	 * If a CQ resize is in progress and we discovered that the
693 	 * old buffer is empty, then peek in the new buffer, and if
694 	 * it's not empty, switch to the new buffer and continue
695 	 * polling there.
696 	 */
697 	if (unlikely(err == -EAGAIN && cq->resize_buf &&
698 		     cq->resize_buf->state == CQ_RESIZE_READY)) {
699 		/*
700 		 * In Tavor mode, the hardware keeps the producer
701 		 * index modulo the CQ size.  Since we might be making
702 		 * the CQ bigger, we need to mask our consumer index
703 		 * using the size of the old CQ buffer before looking
704 		 * in the new CQ buffer.
705 		 */
706 		if (!mthca_is_memfree(dev))
707 			cq->cons_index &= cq->ibcq.cqe;
708 
709 		if (cqe_sw(get_cqe_from_buf(&cq->resize_buf->buf,
710 					    cq->cons_index & cq->resize_buf->cqe))) {
711 			struct mthca_cq_buf tbuf;
712 			int tcqe;
713 
714 			tbuf         = cq->buf;
715 			tcqe         = cq->ibcq.cqe;
716 			cq->buf      = cq->resize_buf->buf;
717 			cq->ibcq.cqe = cq->resize_buf->cqe;
718 
719 			cq->resize_buf->buf   = tbuf;
720 			cq->resize_buf->cqe   = tcqe;
721 			cq->resize_buf->state = CQ_RESIZE_SWAPPED;
722 
723 			goto repoll;
724 		}
725 	}
726 
727 	spin_unlock_irqrestore(&cq->lock, flags);
728 
729 	return err == 0 || err == -EAGAIN ? npolled : err;
730 }
731 
732 int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify_flags flags)
733 {
734 	__be32 doorbell[2];
735 
736 	doorbell[0] = cpu_to_be32(((flags & IB_CQ_SOLICITED_MASK) ==
737 				   IB_CQ_SOLICITED ?
738 				   MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL :
739 				   MTHCA_TAVOR_CQ_DB_REQ_NOT)      |
740 				  to_mcq(cq)->cqn);
741 	doorbell[1] = (__force __be32) 0xffffffff;
742 
743 	mthca_write64(doorbell,
744 		      to_mdev(cq->device)->kar + MTHCA_CQ_DOORBELL,
745 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(cq->device)->doorbell_lock));
746 
747 	return 0;
748 }
749 
750 int mthca_arbel_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
751 {
752 	struct mthca_cq *cq = to_mcq(ibcq);
753 	__be32 doorbell[2];
754 	u32 sn;
755 	__be32 ci;
756 
757 	sn = cq->arm_sn & 3;
758 	ci = cpu_to_be32(cq->cons_index);
759 
760 	doorbell[0] = ci;
761 	doorbell[1] = cpu_to_be32((cq->cqn << 8) | (2 << 5) | (sn << 3) |
762 				  ((flags & IB_CQ_SOLICITED_MASK) ==
763 				   IB_CQ_SOLICITED ? 1 : 2));
764 
765 	mthca_write_db_rec(doorbell, cq->arm_db);
766 
767 	/*
768 	 * Make sure that the doorbell record in host memory is
769 	 * written before ringing the doorbell via PCI MMIO.
770 	 */
771 	wmb();
772 
773 	doorbell[0] = cpu_to_be32((sn << 28)                       |
774 				  ((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
775 				   MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL :
776 				   MTHCA_ARBEL_CQ_DB_REQ_NOT)      |
777 				  cq->cqn);
778 	doorbell[1] = ci;
779 
780 	mthca_write64(doorbell,
781 		      to_mdev(ibcq->device)->kar + MTHCA_CQ_DOORBELL,
782 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->doorbell_lock));
783 
784 	return 0;
785 }
786 
787 int mthca_init_cq(struct mthca_dev *dev, int nent,
788 		  struct mthca_ucontext *ctx, u32 pdn,
789 		  struct mthca_cq *cq)
790 {
791 	struct mthca_mailbox *mailbox;
792 	struct mthca_cq_context *cq_context;
793 	int err = -ENOMEM;
794 	u8 status;
795 
796 	cq->ibcq.cqe  = nent - 1;
797 	cq->is_kernel = !ctx;
798 
799 	cq->cqn = mthca_alloc(&dev->cq_table.alloc);
800 	if (cq->cqn == -1)
801 		return -ENOMEM;
802 
803 	if (mthca_is_memfree(dev)) {
804 		err = mthca_table_get(dev, dev->cq_table.table, cq->cqn);
805 		if (err)
806 			goto err_out;
807 
808 		if (cq->is_kernel) {
809 			cq->arm_sn = 1;
810 
811 			err = -ENOMEM;
812 
813 			cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
814 							     cq->cqn, &cq->set_ci_db);
815 			if (cq->set_ci_db_index < 0)
816 				goto err_out_icm;
817 
818 			cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
819 							  cq->cqn, &cq->arm_db);
820 			if (cq->arm_db_index < 0)
821 				goto err_out_ci;
822 		}
823 	}
824 
825 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
826 	if (IS_ERR(mailbox))
827 		goto err_out_arm;
828 
829 	cq_context = mailbox->buf;
830 
831 	if (cq->is_kernel) {
832 		err = mthca_alloc_cq_buf(dev, &cq->buf, nent);
833 		if (err)
834 			goto err_out_mailbox;
835 	}
836 
837 	spin_lock_init(&cq->lock);
838 	cq->refcount = 1;
839 	init_waitqueue_head(&cq->wait);
840 	mutex_init(&cq->mutex);
841 
842 	memset(cq_context, 0, sizeof *cq_context);
843 	cq_context->flags           = cpu_to_be32(MTHCA_CQ_STATUS_OK      |
844 						  MTHCA_CQ_STATE_DISARMED |
845 						  MTHCA_CQ_FLAG_TR);
846 	cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24);
847 	if (ctx)
848 		cq_context->logsize_usrpage |= cpu_to_be32(ctx->uar.index);
849 	else
850 		cq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
851 	cq_context->error_eqn       = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
852 	cq_context->comp_eqn        = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
853 	cq_context->pd              = cpu_to_be32(pdn);
854 	cq_context->lkey            = cpu_to_be32(cq->buf.mr.ibmr.lkey);
855 	cq_context->cqn             = cpu_to_be32(cq->cqn);
856 
857 	if (mthca_is_memfree(dev)) {
858 		cq_context->ci_db    = cpu_to_be32(cq->set_ci_db_index);
859 		cq_context->state_db = cpu_to_be32(cq->arm_db_index);
860 	}
861 
862 	err = mthca_SW2HW_CQ(dev, mailbox, cq->cqn, &status);
863 	if (err) {
864 		mthca_warn(dev, "SW2HW_CQ failed (%d)\n", err);
865 		goto err_out_free_mr;
866 	}
867 
868 	if (status) {
869 		mthca_warn(dev, "SW2HW_CQ returned status 0x%02x\n",
870 			   status);
871 		err = -EINVAL;
872 		goto err_out_free_mr;
873 	}
874 
875 	spin_lock_irq(&dev->cq_table.lock);
876 	if (mthca_array_set(&dev->cq_table.cq,
877 			    cq->cqn & (dev->limits.num_cqs - 1),
878 			    cq)) {
879 		spin_unlock_irq(&dev->cq_table.lock);
880 		goto err_out_free_mr;
881 	}
882 	spin_unlock_irq(&dev->cq_table.lock);
883 
884 	cq->cons_index = 0;
885 
886 	mthca_free_mailbox(dev, mailbox);
887 
888 	return 0;
889 
890 err_out_free_mr:
891 	if (cq->is_kernel)
892 		mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
893 
894 err_out_mailbox:
895 	mthca_free_mailbox(dev, mailbox);
896 
897 err_out_arm:
898 	if (cq->is_kernel && mthca_is_memfree(dev))
899 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
900 
901 err_out_ci:
902 	if (cq->is_kernel && mthca_is_memfree(dev))
903 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
904 
905 err_out_icm:
906 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
907 
908 err_out:
909 	mthca_free(&dev->cq_table.alloc, cq->cqn);
910 
911 	return err;
912 }
913 
914 static inline int get_cq_refcount(struct mthca_dev *dev, struct mthca_cq *cq)
915 {
916 	int c;
917 
918 	spin_lock_irq(&dev->cq_table.lock);
919 	c = cq->refcount;
920 	spin_unlock_irq(&dev->cq_table.lock);
921 
922 	return c;
923 }
924 
925 void mthca_free_cq(struct mthca_dev *dev,
926 		   struct mthca_cq *cq)
927 {
928 	struct mthca_mailbox *mailbox;
929 	int err;
930 	u8 status;
931 
932 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
933 	if (IS_ERR(mailbox)) {
934 		mthca_warn(dev, "No memory for mailbox to free CQ.\n");
935 		return;
936 	}
937 
938 	err = mthca_HW2SW_CQ(dev, mailbox, cq->cqn, &status);
939 	if (err)
940 		mthca_warn(dev, "HW2SW_CQ failed (%d)\n", err);
941 	else if (status)
942 		mthca_warn(dev, "HW2SW_CQ returned status 0x%02x\n", status);
943 
944 	if (0) {
945 		__be32 *ctx = mailbox->buf;
946 		int j;
947 
948 		printk(KERN_ERR "context for CQN %x (cons index %x, next sw %d)\n",
949 		       cq->cqn, cq->cons_index,
950 		       cq->is_kernel ? !!next_cqe_sw(cq) : 0);
951 		for (j = 0; j < 16; ++j)
952 			printk(KERN_ERR "[%2x] %08x\n", j * 4, be32_to_cpu(ctx[j]));
953 	}
954 
955 	spin_lock_irq(&dev->cq_table.lock);
956 	mthca_array_clear(&dev->cq_table.cq,
957 			  cq->cqn & (dev->limits.num_cqs - 1));
958 	--cq->refcount;
959 	spin_unlock_irq(&dev->cq_table.lock);
960 
961 	if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
962 		synchronize_irq(dev->eq_table.eq[MTHCA_EQ_COMP].msi_x_vector);
963 	else
964 		synchronize_irq(dev->pdev->irq);
965 
966 	wait_event(cq->wait, !get_cq_refcount(dev, cq));
967 
968 	if (cq->is_kernel) {
969 		mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
970 		if (mthca_is_memfree(dev)) {
971 			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM,    cq->arm_db_index);
972 			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
973 		}
974 	}
975 
976 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
977 	mthca_free(&dev->cq_table.alloc, cq->cqn);
978 	mthca_free_mailbox(dev, mailbox);
979 }
980 
981 int mthca_init_cq_table(struct mthca_dev *dev)
982 {
983 	int err;
984 
985 	spin_lock_init(&dev->cq_table.lock);
986 
987 	err = mthca_alloc_init(&dev->cq_table.alloc,
988 			       dev->limits.num_cqs,
989 			       (1 << 24) - 1,
990 			       dev->limits.reserved_cqs);
991 	if (err)
992 		return err;
993 
994 	err = mthca_array_init(&dev->cq_table.cq,
995 			       dev->limits.num_cqs);
996 	if (err)
997 		mthca_alloc_cleanup(&dev->cq_table.alloc);
998 
999 	return err;
1000 }
1001 
1002 void mthca_cleanup_cq_table(struct mthca_dev *dev)
1003 {
1004 	mthca_array_cleanup(&dev->cq_table.cq, dev->limits.num_cqs);
1005 	mthca_alloc_cleanup(&dev->cq_table.alloc);
1006 }
1007