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 	u8     sl_ipok;
123 	u8     g_mlpath;
124 	__be16 rlid;
125 	__be32 imm_etype_pkey_eec;
126 	__be32 byte_cnt;
127 	__be32 wqe;
128 	u8     opcode;
129 	u8     is_send;
130 	u8     reserved;
131 	u8     owner;
132 };
133 
134 struct mthca_err_cqe {
135 	__be32 my_qpn;
136 	u32    reserved1[3];
137 	u8     syndrome;
138 	u8     vendor_err;
139 	__be16 db_cnt;
140 	u32    reserved2;
141 	__be32 wqe;
142 	u8     opcode;
143 	u8     reserved3[2];
144 	u8     owner;
145 };
146 
147 #define MTHCA_CQ_ENTRY_OWNER_SW      (0 << 7)
148 #define MTHCA_CQ_ENTRY_OWNER_HW      (1 << 7)
149 
150 #define MTHCA_TAVOR_CQ_DB_INC_CI       (1 << 24)
151 #define MTHCA_TAVOR_CQ_DB_REQ_NOT      (2 << 24)
152 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL  (3 << 24)
153 #define MTHCA_TAVOR_CQ_DB_SET_CI       (4 << 24)
154 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_MULT (5 << 24)
155 
156 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL  (1 << 24)
157 #define MTHCA_ARBEL_CQ_DB_REQ_NOT      (2 << 24)
158 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_MULT (3 << 24)
159 
160 static inline struct mthca_cqe *get_cqe_from_buf(struct mthca_cq_buf *buf,
161 						 int entry)
162 {
163 	if (buf->is_direct)
164 		return buf->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
165 	else
166 		return buf->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
167 			+ (entry * MTHCA_CQ_ENTRY_SIZE) % PAGE_SIZE;
168 }
169 
170 static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
171 {
172 	return get_cqe_from_buf(&cq->buf, entry);
173 }
174 
175 static inline struct mthca_cqe *cqe_sw(struct mthca_cqe *cqe)
176 {
177 	return MTHCA_CQ_ENTRY_OWNER_HW & cqe->owner ? NULL : cqe;
178 }
179 
180 static inline struct mthca_cqe *next_cqe_sw(struct mthca_cq *cq)
181 {
182 	return cqe_sw(get_cqe(cq, cq->cons_index & cq->ibcq.cqe));
183 }
184 
185 static inline void set_cqe_hw(struct mthca_cqe *cqe)
186 {
187 	cqe->owner = MTHCA_CQ_ENTRY_OWNER_HW;
188 }
189 
190 static void dump_cqe(struct mthca_dev *dev, void *cqe_ptr)
191 {
192 	__be32 *cqe = cqe_ptr;
193 
194 	(void) cqe;	/* avoid warning if mthca_dbg compiled away... */
195 	mthca_dbg(dev, "CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
196 		  be32_to_cpu(cqe[0]), be32_to_cpu(cqe[1]), be32_to_cpu(cqe[2]),
197 		  be32_to_cpu(cqe[3]), be32_to_cpu(cqe[4]), be32_to_cpu(cqe[5]),
198 		  be32_to_cpu(cqe[6]), be32_to_cpu(cqe[7]));
199 }
200 
201 /*
202  * incr is ignored in native Arbel (mem-free) mode, so cq->cons_index
203  * should be correct before calling update_cons_index().
204  */
205 static inline void update_cons_index(struct mthca_dev *dev, struct mthca_cq *cq,
206 				     int incr)
207 {
208 	if (mthca_is_memfree(dev)) {
209 		*cq->set_ci_db = cpu_to_be32(cq->cons_index);
210 		wmb();
211 	} else {
212 		mthca_write64(MTHCA_TAVOR_CQ_DB_INC_CI | cq->cqn, incr - 1,
213 			      dev->kar + MTHCA_CQ_DOORBELL,
214 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
215 		/*
216 		 * Make sure doorbells don't leak out of CQ spinlock
217 		 * and reach the HCA out of order:
218 		 */
219 		mmiowb();
220 	}
221 }
222 
223 void mthca_cq_completion(struct mthca_dev *dev, u32 cqn)
224 {
225 	struct mthca_cq *cq;
226 
227 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
228 
229 	if (!cq) {
230 		mthca_warn(dev, "Completion event for bogus CQ %08x\n", cqn);
231 		return;
232 	}
233 
234 	++cq->arm_sn;
235 
236 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
237 }
238 
239 void mthca_cq_event(struct mthca_dev *dev, u32 cqn,
240 		    enum ib_event_type event_type)
241 {
242 	struct mthca_cq *cq;
243 	struct ib_event event;
244 
245 	spin_lock(&dev->cq_table.lock);
246 
247 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
248 	if (cq)
249 		++cq->refcount;
250 
251 	spin_unlock(&dev->cq_table.lock);
252 
253 	if (!cq) {
254 		mthca_warn(dev, "Async event for bogus CQ %08x\n", cqn);
255 		return;
256 	}
257 
258 	event.device      = &dev->ib_dev;
259 	event.event       = event_type;
260 	event.element.cq  = &cq->ibcq;
261 	if (cq->ibcq.event_handler)
262 		cq->ibcq.event_handler(&event, cq->ibcq.cq_context);
263 
264 	spin_lock(&dev->cq_table.lock);
265 	if (!--cq->refcount)
266 		wake_up(&cq->wait);
267 	spin_unlock(&dev->cq_table.lock);
268 }
269 
270 static inline int is_recv_cqe(struct mthca_cqe *cqe)
271 {
272 	if ((cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
273 	    MTHCA_ERROR_CQE_OPCODE_MASK)
274 		return !(cqe->opcode & 0x01);
275 	else
276 		return !(cqe->is_send & 0x80);
277 }
278 
279 void mthca_cq_clean(struct mthca_dev *dev, struct mthca_cq *cq, u32 qpn,
280 		    struct mthca_srq *srq)
281 {
282 	struct mthca_cqe *cqe;
283 	u32 prod_index;
284 	int i, nfreed = 0;
285 
286 	spin_lock_irq(&cq->lock);
287 
288 	/*
289 	 * First we need to find the current producer index, so we
290 	 * know where to start cleaning from.  It doesn't matter if HW
291 	 * adds new entries after this loop -- the QP we're worried
292 	 * about is already in RESET, so the new entries won't come
293 	 * from our QP and therefore don't need to be checked.
294 	 */
295 	for (prod_index = cq->cons_index;
296 	     cqe_sw(get_cqe(cq, prod_index & cq->ibcq.cqe));
297 	     ++prod_index)
298 		if (prod_index == cq->cons_index + cq->ibcq.cqe)
299 			break;
300 
301 	if (0)
302 		mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
303 			  qpn, cq->cqn, cq->cons_index, prod_index);
304 
305 	/*
306 	 * Now sweep backwards through the CQ, removing CQ entries
307 	 * that match our QP by copying older entries on top of them.
308 	 */
309 	while ((int) --prod_index - (int) cq->cons_index >= 0) {
310 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
311 		if (cqe->my_qpn == cpu_to_be32(qpn)) {
312 			if (srq && is_recv_cqe(cqe))
313 				mthca_free_srq_wqe(srq, be32_to_cpu(cqe->wqe));
314 			++nfreed;
315 		} else if (nfreed)
316 			memcpy(get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe),
317 			       cqe, MTHCA_CQ_ENTRY_SIZE);
318 	}
319 
320 	if (nfreed) {
321 		for (i = 0; i < nfreed; ++i)
322 			set_cqe_hw(get_cqe(cq, (cq->cons_index + i) & cq->ibcq.cqe));
323 		wmb();
324 		cq->cons_index += nfreed;
325 		update_cons_index(dev, cq, nfreed);
326 	}
327 
328 	spin_unlock_irq(&cq->lock);
329 }
330 
331 void mthca_cq_resize_copy_cqes(struct mthca_cq *cq)
332 {
333 	int i;
334 
335 	/*
336 	 * In Tavor mode, the hardware keeps the consumer and producer
337 	 * indices mod the CQ size.  Since we might be making the CQ
338 	 * bigger, we need to deal with the case where the producer
339 	 * index wrapped around before the CQ was resized.
340 	 */
341 	if (!mthca_is_memfree(to_mdev(cq->ibcq.device)) &&
342 	    cq->ibcq.cqe < cq->resize_buf->cqe) {
343 		cq->cons_index &= cq->ibcq.cqe;
344 		if (cqe_sw(get_cqe(cq, cq->ibcq.cqe)))
345 			cq->cons_index -= cq->ibcq.cqe + 1;
346 	}
347 
348 	for (i = cq->cons_index; cqe_sw(get_cqe(cq, i & cq->ibcq.cqe)); ++i)
349 		memcpy(get_cqe_from_buf(&cq->resize_buf->buf,
350 					i & cq->resize_buf->cqe),
351 		       get_cqe(cq, i & cq->ibcq.cqe), MTHCA_CQ_ENTRY_SIZE);
352 }
353 
354 int mthca_alloc_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int nent)
355 {
356 	int ret;
357 	int i;
358 
359 	ret = mthca_buf_alloc(dev, nent * MTHCA_CQ_ENTRY_SIZE,
360 			      MTHCA_MAX_DIRECT_CQ_SIZE,
361 			      &buf->queue, &buf->is_direct,
362 			      &dev->driver_pd, 1, &buf->mr);
363 	if (ret)
364 		return ret;
365 
366 	for (i = 0; i < nent; ++i)
367 		set_cqe_hw(get_cqe_from_buf(buf, i));
368 
369 	return 0;
370 }
371 
372 void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int cqe)
373 {
374 	mthca_buf_free(dev, (cqe + 1) * MTHCA_CQ_ENTRY_SIZE, &buf->queue,
375 		       buf->is_direct, &buf->mr);
376 }
377 
378 static void handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
379 			     struct mthca_qp *qp, int wqe_index, int is_send,
380 			     struct mthca_err_cqe *cqe,
381 			     struct ib_wc *entry, int *free_cqe)
382 {
383 	int dbd;
384 	__be32 new_wqe;
385 
386 	if (cqe->syndrome == SYNDROME_LOCAL_QP_OP_ERR) {
387 		mthca_dbg(dev, "local QP operation err "
388 			  "(QPN %06x, WQE @ %08x, CQN %06x, index %d)\n",
389 			  be32_to_cpu(cqe->my_qpn), be32_to_cpu(cqe->wqe),
390 			  cq->cqn, cq->cons_index);
391 		dump_cqe(dev, cqe);
392 	}
393 
394 	/*
395 	 * For completions in error, only work request ID, status, vendor error
396 	 * (and freed resource count for RD) have to be set.
397 	 */
398 	switch (cqe->syndrome) {
399 	case SYNDROME_LOCAL_LENGTH_ERR:
400 		entry->status = IB_WC_LOC_LEN_ERR;
401 		break;
402 	case SYNDROME_LOCAL_QP_OP_ERR:
403 		entry->status = IB_WC_LOC_QP_OP_ERR;
404 		break;
405 	case SYNDROME_LOCAL_EEC_OP_ERR:
406 		entry->status = IB_WC_LOC_EEC_OP_ERR;
407 		break;
408 	case SYNDROME_LOCAL_PROT_ERR:
409 		entry->status = IB_WC_LOC_PROT_ERR;
410 		break;
411 	case SYNDROME_WR_FLUSH_ERR:
412 		entry->status = IB_WC_WR_FLUSH_ERR;
413 		break;
414 	case SYNDROME_MW_BIND_ERR:
415 		entry->status = IB_WC_MW_BIND_ERR;
416 		break;
417 	case SYNDROME_BAD_RESP_ERR:
418 		entry->status = IB_WC_BAD_RESP_ERR;
419 		break;
420 	case SYNDROME_LOCAL_ACCESS_ERR:
421 		entry->status = IB_WC_LOC_ACCESS_ERR;
422 		break;
423 	case SYNDROME_REMOTE_INVAL_REQ_ERR:
424 		entry->status = IB_WC_REM_INV_REQ_ERR;
425 		break;
426 	case SYNDROME_REMOTE_ACCESS_ERR:
427 		entry->status = IB_WC_REM_ACCESS_ERR;
428 		break;
429 	case SYNDROME_REMOTE_OP_ERR:
430 		entry->status = IB_WC_REM_OP_ERR;
431 		break;
432 	case SYNDROME_RETRY_EXC_ERR:
433 		entry->status = IB_WC_RETRY_EXC_ERR;
434 		break;
435 	case SYNDROME_RNR_RETRY_EXC_ERR:
436 		entry->status = IB_WC_RNR_RETRY_EXC_ERR;
437 		break;
438 	case SYNDROME_LOCAL_RDD_VIOL_ERR:
439 		entry->status = IB_WC_LOC_RDD_VIOL_ERR;
440 		break;
441 	case SYNDROME_REMOTE_INVAL_RD_REQ_ERR:
442 		entry->status = IB_WC_REM_INV_RD_REQ_ERR;
443 		break;
444 	case SYNDROME_REMOTE_ABORTED_ERR:
445 		entry->status = IB_WC_REM_ABORT_ERR;
446 		break;
447 	case SYNDROME_INVAL_EECN_ERR:
448 		entry->status = IB_WC_INV_EECN_ERR;
449 		break;
450 	case SYNDROME_INVAL_EEC_STATE_ERR:
451 		entry->status = IB_WC_INV_EEC_STATE_ERR;
452 		break;
453 	default:
454 		entry->status = IB_WC_GENERAL_ERR;
455 		break;
456 	}
457 
458 	entry->vendor_err = cqe->vendor_err;
459 
460 	/*
461 	 * Mem-free HCAs always generate one CQE per WQE, even in the
462 	 * error case, so we don't have to check the doorbell count, etc.
463 	 */
464 	if (mthca_is_memfree(dev))
465 		return;
466 
467 	mthca_free_err_wqe(dev, qp, is_send, wqe_index, &dbd, &new_wqe);
468 
469 	/*
470 	 * If we're at the end of the WQE chain, or we've used up our
471 	 * doorbell count, free the CQE.  Otherwise just update it for
472 	 * the next poll operation.
473 	 */
474 	if (!(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
475 		return;
476 
477 	be16_add_cpu(&cqe->db_cnt, -dbd);
478 	cqe->wqe      = new_wqe;
479 	cqe->syndrome = SYNDROME_WR_FLUSH_ERR;
480 
481 	*free_cqe = 0;
482 }
483 
484 static inline int mthca_poll_one(struct mthca_dev *dev,
485 				 struct mthca_cq *cq,
486 				 struct mthca_qp **cur_qp,
487 				 int *freed,
488 				 struct ib_wc *entry)
489 {
490 	struct mthca_wq *wq;
491 	struct mthca_cqe *cqe;
492 	int wqe_index;
493 	int is_error;
494 	int is_send;
495 	int free_cqe = 1;
496 	int err = 0;
497 	u16 checksum;
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   	   = cqe->sl_ipok >> 4;
641 		entry->src_qp 	   = be32_to_cpu(cqe->rqpn) & 0xffffff;
642 		entry->dlid_path_bits = cqe->g_mlpath & 0x7f;
643 		entry->pkey_index  = be32_to_cpu(cqe->imm_etype_pkey_eec) >> 16;
644 		entry->wc_flags   |= cqe->g_mlpath & 0x80 ? IB_WC_GRH : 0;
645 		checksum = (be32_to_cpu(cqe->rqpn) >> 24) |
646 				((be32_to_cpu(cqe->my_ee) >> 16) & 0xff00);
647 		entry->csum_ok = (cqe->sl_ipok & 1 && checksum == 0xffff);
648 	}
649 
650 	entry->status = IB_WC_SUCCESS;
651 
652  out:
653 	if (likely(free_cqe)) {
654 		set_cqe_hw(cqe);
655 		++(*freed);
656 		++cq->cons_index;
657 	}
658 
659 	return err;
660 }
661 
662 int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
663 		  struct ib_wc *entry)
664 {
665 	struct mthca_dev *dev = to_mdev(ibcq->device);
666 	struct mthca_cq *cq = to_mcq(ibcq);
667 	struct mthca_qp *qp = NULL;
668 	unsigned long flags;
669 	int err = 0;
670 	int freed = 0;
671 	int npolled;
672 
673 	spin_lock_irqsave(&cq->lock, flags);
674 
675 	npolled = 0;
676 repoll:
677 	while (npolled < num_entries) {
678 		err = mthca_poll_one(dev, cq, &qp,
679 				     &freed, entry + npolled);
680 		if (err)
681 			break;
682 		++npolled;
683 	}
684 
685 	if (freed) {
686 		wmb();
687 		update_cons_index(dev, cq, freed);
688 	}
689 
690 	/*
691 	 * If a CQ resize is in progress and we discovered that the
692 	 * old buffer is empty, then peek in the new buffer, and if
693 	 * it's not empty, switch to the new buffer and continue
694 	 * polling there.
695 	 */
696 	if (unlikely(err == -EAGAIN && cq->resize_buf &&
697 		     cq->resize_buf->state == CQ_RESIZE_READY)) {
698 		/*
699 		 * In Tavor mode, the hardware keeps the producer
700 		 * index modulo the CQ size.  Since we might be making
701 		 * the CQ bigger, we need to mask our consumer index
702 		 * using the size of the old CQ buffer before looking
703 		 * in the new CQ buffer.
704 		 */
705 		if (!mthca_is_memfree(dev))
706 			cq->cons_index &= cq->ibcq.cqe;
707 
708 		if (cqe_sw(get_cqe_from_buf(&cq->resize_buf->buf,
709 					    cq->cons_index & cq->resize_buf->cqe))) {
710 			struct mthca_cq_buf tbuf;
711 			int tcqe;
712 
713 			tbuf         = cq->buf;
714 			tcqe         = cq->ibcq.cqe;
715 			cq->buf      = cq->resize_buf->buf;
716 			cq->ibcq.cqe = cq->resize_buf->cqe;
717 
718 			cq->resize_buf->buf   = tbuf;
719 			cq->resize_buf->cqe   = tcqe;
720 			cq->resize_buf->state = CQ_RESIZE_SWAPPED;
721 
722 			goto repoll;
723 		}
724 	}
725 
726 	spin_unlock_irqrestore(&cq->lock, flags);
727 
728 	return err == 0 || err == -EAGAIN ? npolled : err;
729 }
730 
731 int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify_flags flags)
732 {
733 	u32 dbhi = ((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
734 		    MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL :
735 		    MTHCA_TAVOR_CQ_DB_REQ_NOT) |
736 		to_mcq(cq)->cqn;
737 
738 	mthca_write64(dbhi, 0xffffffff, to_mdev(cq->device)->kar + MTHCA_CQ_DOORBELL,
739 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(cq->device)->doorbell_lock));
740 
741 	return 0;
742 }
743 
744 int mthca_arbel_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
745 {
746 	struct mthca_cq *cq = to_mcq(ibcq);
747 	__be32 db_rec[2];
748 	u32 dbhi;
749 	u32 sn = cq->arm_sn & 3;
750 
751 	db_rec[0] = cpu_to_be32(cq->cons_index);
752 	db_rec[1] = cpu_to_be32((cq->cqn << 8) | (2 << 5) | (sn << 3) |
753 				((flags & IB_CQ_SOLICITED_MASK) ==
754 				 IB_CQ_SOLICITED ? 1 : 2));
755 
756 	mthca_write_db_rec(db_rec, cq->arm_db);
757 
758 	/*
759 	 * Make sure that the doorbell record in host memory is
760 	 * written before ringing the doorbell via PCI MMIO.
761 	 */
762 	wmb();
763 
764 	dbhi = (sn << 28) |
765 		((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
766 		 MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL :
767 		 MTHCA_ARBEL_CQ_DB_REQ_NOT) | cq->cqn;
768 
769 	mthca_write64(dbhi, cq->cons_index,
770 		      to_mdev(ibcq->device)->kar + MTHCA_CQ_DOORBELL,
771 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->doorbell_lock));
772 
773 	return 0;
774 }
775 
776 int mthca_init_cq(struct mthca_dev *dev, int nent,
777 		  struct mthca_ucontext *ctx, u32 pdn,
778 		  struct mthca_cq *cq)
779 {
780 	struct mthca_mailbox *mailbox;
781 	struct mthca_cq_context *cq_context;
782 	int err = -ENOMEM;
783 	u8 status;
784 
785 	cq->ibcq.cqe  = nent - 1;
786 	cq->is_kernel = !ctx;
787 
788 	cq->cqn = mthca_alloc(&dev->cq_table.alloc);
789 	if (cq->cqn == -1)
790 		return -ENOMEM;
791 
792 	if (mthca_is_memfree(dev)) {
793 		err = mthca_table_get(dev, dev->cq_table.table, cq->cqn);
794 		if (err)
795 			goto err_out;
796 
797 		if (cq->is_kernel) {
798 			cq->arm_sn = 1;
799 
800 			err = -ENOMEM;
801 
802 			cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
803 							     cq->cqn, &cq->set_ci_db);
804 			if (cq->set_ci_db_index < 0)
805 				goto err_out_icm;
806 
807 			cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
808 							  cq->cqn, &cq->arm_db);
809 			if (cq->arm_db_index < 0)
810 				goto err_out_ci;
811 		}
812 	}
813 
814 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
815 	if (IS_ERR(mailbox))
816 		goto err_out_arm;
817 
818 	cq_context = mailbox->buf;
819 
820 	if (cq->is_kernel) {
821 		err = mthca_alloc_cq_buf(dev, &cq->buf, nent);
822 		if (err)
823 			goto err_out_mailbox;
824 	}
825 
826 	spin_lock_init(&cq->lock);
827 	cq->refcount = 1;
828 	init_waitqueue_head(&cq->wait);
829 	mutex_init(&cq->mutex);
830 
831 	memset(cq_context, 0, sizeof *cq_context);
832 	cq_context->flags           = cpu_to_be32(MTHCA_CQ_STATUS_OK      |
833 						  MTHCA_CQ_STATE_DISARMED |
834 						  MTHCA_CQ_FLAG_TR);
835 	cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24);
836 	if (ctx)
837 		cq_context->logsize_usrpage |= cpu_to_be32(ctx->uar.index);
838 	else
839 		cq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
840 	cq_context->error_eqn       = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
841 	cq_context->comp_eqn        = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
842 	cq_context->pd              = cpu_to_be32(pdn);
843 	cq_context->lkey            = cpu_to_be32(cq->buf.mr.ibmr.lkey);
844 	cq_context->cqn             = cpu_to_be32(cq->cqn);
845 
846 	if (mthca_is_memfree(dev)) {
847 		cq_context->ci_db    = cpu_to_be32(cq->set_ci_db_index);
848 		cq_context->state_db = cpu_to_be32(cq->arm_db_index);
849 	}
850 
851 	err = mthca_SW2HW_CQ(dev, mailbox, cq->cqn, &status);
852 	if (err) {
853 		mthca_warn(dev, "SW2HW_CQ failed (%d)\n", err);
854 		goto err_out_free_mr;
855 	}
856 
857 	if (status) {
858 		mthca_warn(dev, "SW2HW_CQ returned status 0x%02x\n",
859 			   status);
860 		err = -EINVAL;
861 		goto err_out_free_mr;
862 	}
863 
864 	spin_lock_irq(&dev->cq_table.lock);
865 	if (mthca_array_set(&dev->cq_table.cq,
866 			    cq->cqn & (dev->limits.num_cqs - 1),
867 			    cq)) {
868 		spin_unlock_irq(&dev->cq_table.lock);
869 		goto err_out_free_mr;
870 	}
871 	spin_unlock_irq(&dev->cq_table.lock);
872 
873 	cq->cons_index = 0;
874 
875 	mthca_free_mailbox(dev, mailbox);
876 
877 	return 0;
878 
879 err_out_free_mr:
880 	if (cq->is_kernel)
881 		mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
882 
883 err_out_mailbox:
884 	mthca_free_mailbox(dev, mailbox);
885 
886 err_out_arm:
887 	if (cq->is_kernel && mthca_is_memfree(dev))
888 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
889 
890 err_out_ci:
891 	if (cq->is_kernel && mthca_is_memfree(dev))
892 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
893 
894 err_out_icm:
895 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
896 
897 err_out:
898 	mthca_free(&dev->cq_table.alloc, cq->cqn);
899 
900 	return err;
901 }
902 
903 static inline int get_cq_refcount(struct mthca_dev *dev, struct mthca_cq *cq)
904 {
905 	int c;
906 
907 	spin_lock_irq(&dev->cq_table.lock);
908 	c = cq->refcount;
909 	spin_unlock_irq(&dev->cq_table.lock);
910 
911 	return c;
912 }
913 
914 void mthca_free_cq(struct mthca_dev *dev,
915 		   struct mthca_cq *cq)
916 {
917 	struct mthca_mailbox *mailbox;
918 	int err;
919 	u8 status;
920 
921 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
922 	if (IS_ERR(mailbox)) {
923 		mthca_warn(dev, "No memory for mailbox to free CQ.\n");
924 		return;
925 	}
926 
927 	err = mthca_HW2SW_CQ(dev, mailbox, cq->cqn, &status);
928 	if (err)
929 		mthca_warn(dev, "HW2SW_CQ failed (%d)\n", err);
930 	else if (status)
931 		mthca_warn(dev, "HW2SW_CQ returned status 0x%02x\n", status);
932 
933 	if (0) {
934 		__be32 *ctx = mailbox->buf;
935 		int j;
936 
937 		printk(KERN_ERR "context for CQN %x (cons index %x, next sw %d)\n",
938 		       cq->cqn, cq->cons_index,
939 		       cq->is_kernel ? !!next_cqe_sw(cq) : 0);
940 		for (j = 0; j < 16; ++j)
941 			printk(KERN_ERR "[%2x] %08x\n", j * 4, be32_to_cpu(ctx[j]));
942 	}
943 
944 	spin_lock_irq(&dev->cq_table.lock);
945 	mthca_array_clear(&dev->cq_table.cq,
946 			  cq->cqn & (dev->limits.num_cqs - 1));
947 	--cq->refcount;
948 	spin_unlock_irq(&dev->cq_table.lock);
949 
950 	if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
951 		synchronize_irq(dev->eq_table.eq[MTHCA_EQ_COMP].msi_x_vector);
952 	else
953 		synchronize_irq(dev->pdev->irq);
954 
955 	wait_event(cq->wait, !get_cq_refcount(dev, cq));
956 
957 	if (cq->is_kernel) {
958 		mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
959 		if (mthca_is_memfree(dev)) {
960 			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM,    cq->arm_db_index);
961 			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
962 		}
963 	}
964 
965 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
966 	mthca_free(&dev->cq_table.alloc, cq->cqn);
967 	mthca_free_mailbox(dev, mailbox);
968 }
969 
970 int mthca_init_cq_table(struct mthca_dev *dev)
971 {
972 	int err;
973 
974 	spin_lock_init(&dev->cq_table.lock);
975 
976 	err = mthca_alloc_init(&dev->cq_table.alloc,
977 			       dev->limits.num_cqs,
978 			       (1 << 24) - 1,
979 			       dev->limits.reserved_cqs);
980 	if (err)
981 		return err;
982 
983 	err = mthca_array_init(&dev->cq_table.cq,
984 			       dev->limits.num_cqs);
985 	if (err)
986 		mthca_alloc_cleanup(&dev->cq_table.alloc);
987 
988 	return err;
989 }
990 
991 void mthca_cleanup_cq_table(struct mthca_dev *dev)
992 {
993 	mthca_array_cleanup(&dev->cq_table.cq, dev->limits.num_cqs);
994 	mthca_alloc_cleanup(&dev->cq_table.alloc);
995 }
996