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 
37 #include <linux/gfp.h>
38 #include <linux/hardirq.h>
39 #include <linux/sched.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 } __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 	u8     sl_ipok;
122 	u8     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 	if (mthca_is_memfree(dev)) {
208 		*cq->set_ci_db = cpu_to_be32(cq->cons_index);
209 		wmb();
210 	} else {
211 		mthca_write64(MTHCA_TAVOR_CQ_DB_INC_CI | cq->cqn, incr - 1,
212 			      dev->kar + MTHCA_CQ_DOORBELL,
213 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
214 	}
215 }
216 
217 void mthca_cq_completion(struct mthca_dev *dev, u32 cqn)
218 {
219 	struct mthca_cq *cq;
220 
221 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
222 
223 	if (!cq) {
224 		mthca_warn(dev, "Completion event for bogus CQ %08x\n", cqn);
225 		return;
226 	}
227 
228 	++cq->arm_sn;
229 
230 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
231 }
232 
233 void mthca_cq_event(struct mthca_dev *dev, u32 cqn,
234 		    enum ib_event_type event_type)
235 {
236 	struct mthca_cq *cq;
237 	struct ib_event event;
238 
239 	spin_lock(&dev->cq_table.lock);
240 
241 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
242 	if (cq)
243 		++cq->refcount;
244 
245 	spin_unlock(&dev->cq_table.lock);
246 
247 	if (!cq) {
248 		mthca_warn(dev, "Async event for bogus CQ %08x\n", cqn);
249 		return;
250 	}
251 
252 	event.device      = &dev->ib_dev;
253 	event.event       = event_type;
254 	event.element.cq  = &cq->ibcq;
255 	if (cq->ibcq.event_handler)
256 		cq->ibcq.event_handler(&event, cq->ibcq.cq_context);
257 
258 	spin_lock(&dev->cq_table.lock);
259 	if (!--cq->refcount)
260 		wake_up(&cq->wait);
261 	spin_unlock(&dev->cq_table.lock);
262 }
263 
264 static inline int is_recv_cqe(struct mthca_cqe *cqe)
265 {
266 	if ((cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
267 	    MTHCA_ERROR_CQE_OPCODE_MASK)
268 		return !(cqe->opcode & 0x01);
269 	else
270 		return !(cqe->is_send & 0x80);
271 }
272 
273 void mthca_cq_clean(struct mthca_dev *dev, struct mthca_cq *cq, u32 qpn,
274 		    struct mthca_srq *srq)
275 {
276 	struct mthca_cqe *cqe;
277 	u32 prod_index;
278 	int i, nfreed = 0;
279 
280 	spin_lock_irq(&cq->lock);
281 
282 	/*
283 	 * First we need to find the current producer index, so we
284 	 * know where to start cleaning from.  It doesn't matter if HW
285 	 * adds new entries after this loop -- the QP we're worried
286 	 * about is already in RESET, so the new entries won't come
287 	 * from our QP and therefore don't need to be checked.
288 	 */
289 	for (prod_index = cq->cons_index;
290 	     cqe_sw(get_cqe(cq, prod_index & cq->ibcq.cqe));
291 	     ++prod_index)
292 		if (prod_index == cq->cons_index + cq->ibcq.cqe)
293 			break;
294 
295 	if (0)
296 		mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
297 			  qpn, cq->cqn, cq->cons_index, prod_index);
298 
299 	/*
300 	 * Now sweep backwards through the CQ, removing CQ entries
301 	 * that match our QP by copying older entries on top of them.
302 	 */
303 	while ((int) --prod_index - (int) cq->cons_index >= 0) {
304 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
305 		if (cqe->my_qpn == cpu_to_be32(qpn)) {
306 			if (srq && is_recv_cqe(cqe))
307 				mthca_free_srq_wqe(srq, be32_to_cpu(cqe->wqe));
308 			++nfreed;
309 		} else if (nfreed)
310 			memcpy(get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe),
311 			       cqe, MTHCA_CQ_ENTRY_SIZE);
312 	}
313 
314 	if (nfreed) {
315 		for (i = 0; i < nfreed; ++i)
316 			set_cqe_hw(get_cqe(cq, (cq->cons_index + i) & cq->ibcq.cqe));
317 		wmb();
318 		cq->cons_index += nfreed;
319 		update_cons_index(dev, cq, nfreed);
320 	}
321 
322 	spin_unlock_irq(&cq->lock);
323 }
324 
325 void mthca_cq_resize_copy_cqes(struct mthca_cq *cq)
326 {
327 	int i;
328 
329 	/*
330 	 * In Tavor mode, the hardware keeps the consumer and producer
331 	 * indices mod the CQ size.  Since we might be making the CQ
332 	 * bigger, we need to deal with the case where the producer
333 	 * index wrapped around before the CQ was resized.
334 	 */
335 	if (!mthca_is_memfree(to_mdev(cq->ibcq.device)) &&
336 	    cq->ibcq.cqe < cq->resize_buf->cqe) {
337 		cq->cons_index &= cq->ibcq.cqe;
338 		if (cqe_sw(get_cqe(cq, cq->ibcq.cqe)))
339 			cq->cons_index -= cq->ibcq.cqe + 1;
340 	}
341 
342 	for (i = cq->cons_index; cqe_sw(get_cqe(cq, i & cq->ibcq.cqe)); ++i)
343 		memcpy(get_cqe_from_buf(&cq->resize_buf->buf,
344 					i & cq->resize_buf->cqe),
345 		       get_cqe(cq, i & cq->ibcq.cqe), MTHCA_CQ_ENTRY_SIZE);
346 }
347 
348 int mthca_alloc_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int nent)
349 {
350 	int ret;
351 	int i;
352 
353 	ret = mthca_buf_alloc(dev, nent * MTHCA_CQ_ENTRY_SIZE,
354 			      MTHCA_MAX_DIRECT_CQ_SIZE,
355 			      &buf->queue, &buf->is_direct,
356 			      &dev->driver_pd, 1, &buf->mr);
357 	if (ret)
358 		return ret;
359 
360 	for (i = 0; i < nent; ++i)
361 		set_cqe_hw(get_cqe_from_buf(buf, i));
362 
363 	return 0;
364 }
365 
366 void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int cqe)
367 {
368 	mthca_buf_free(dev, (cqe + 1) * MTHCA_CQ_ENTRY_SIZE, &buf->queue,
369 		       buf->is_direct, &buf->mr);
370 }
371 
372 static void handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
373 			     struct mthca_qp *qp, int wqe_index, int is_send,
374 			     struct mthca_err_cqe *cqe,
375 			     struct ib_wc *entry, int *free_cqe)
376 {
377 	int dbd;
378 	__be32 new_wqe;
379 
380 	if (cqe->syndrome == SYNDROME_LOCAL_QP_OP_ERR) {
381 		mthca_dbg(dev, "local QP operation err "
382 			  "(QPN %06x, WQE @ %08x, CQN %06x, index %d)\n",
383 			  be32_to_cpu(cqe->my_qpn), be32_to_cpu(cqe->wqe),
384 			  cq->cqn, cq->cons_index);
385 		dump_cqe(dev, cqe);
386 	}
387 
388 	/*
389 	 * For completions in error, only work request ID, status, vendor error
390 	 * (and freed resource count for RD) have to be set.
391 	 */
392 	switch (cqe->syndrome) {
393 	case SYNDROME_LOCAL_LENGTH_ERR:
394 		entry->status = IB_WC_LOC_LEN_ERR;
395 		break;
396 	case SYNDROME_LOCAL_QP_OP_ERR:
397 		entry->status = IB_WC_LOC_QP_OP_ERR;
398 		break;
399 	case SYNDROME_LOCAL_EEC_OP_ERR:
400 		entry->status = IB_WC_LOC_EEC_OP_ERR;
401 		break;
402 	case SYNDROME_LOCAL_PROT_ERR:
403 		entry->status = IB_WC_LOC_PROT_ERR;
404 		break;
405 	case SYNDROME_WR_FLUSH_ERR:
406 		entry->status = IB_WC_WR_FLUSH_ERR;
407 		break;
408 	case SYNDROME_MW_BIND_ERR:
409 		entry->status = IB_WC_MW_BIND_ERR;
410 		break;
411 	case SYNDROME_BAD_RESP_ERR:
412 		entry->status = IB_WC_BAD_RESP_ERR;
413 		break;
414 	case SYNDROME_LOCAL_ACCESS_ERR:
415 		entry->status = IB_WC_LOC_ACCESS_ERR;
416 		break;
417 	case SYNDROME_REMOTE_INVAL_REQ_ERR:
418 		entry->status = IB_WC_REM_INV_REQ_ERR;
419 		break;
420 	case SYNDROME_REMOTE_ACCESS_ERR:
421 		entry->status = IB_WC_REM_ACCESS_ERR;
422 		break;
423 	case SYNDROME_REMOTE_OP_ERR:
424 		entry->status = IB_WC_REM_OP_ERR;
425 		break;
426 	case SYNDROME_RETRY_EXC_ERR:
427 		entry->status = IB_WC_RETRY_EXC_ERR;
428 		break;
429 	case SYNDROME_RNR_RETRY_EXC_ERR:
430 		entry->status = IB_WC_RNR_RETRY_EXC_ERR;
431 		break;
432 	case SYNDROME_LOCAL_RDD_VIOL_ERR:
433 		entry->status = IB_WC_LOC_RDD_VIOL_ERR;
434 		break;
435 	case SYNDROME_REMOTE_INVAL_RD_REQ_ERR:
436 		entry->status = IB_WC_REM_INV_RD_REQ_ERR;
437 		break;
438 	case SYNDROME_REMOTE_ABORTED_ERR:
439 		entry->status = IB_WC_REM_ABORT_ERR;
440 		break;
441 	case SYNDROME_INVAL_EECN_ERR:
442 		entry->status = IB_WC_INV_EECN_ERR;
443 		break;
444 	case SYNDROME_INVAL_EEC_STATE_ERR:
445 		entry->status = IB_WC_INV_EEC_STATE_ERR;
446 		break;
447 	default:
448 		entry->status = IB_WC_GENERAL_ERR;
449 		break;
450 	}
451 
452 	entry->vendor_err = cqe->vendor_err;
453 
454 	/*
455 	 * Mem-free HCAs always generate one CQE per WQE, even in the
456 	 * error case, so we don't have to check the doorbell count, etc.
457 	 */
458 	if (mthca_is_memfree(dev))
459 		return;
460 
461 	mthca_free_err_wqe(dev, qp, is_send, wqe_index, &dbd, &new_wqe);
462 
463 	/*
464 	 * If we're at the end of the WQE chain, or we've used up our
465 	 * doorbell count, free the CQE.  Otherwise just update it for
466 	 * the next poll operation.
467 	 */
468 	if (!(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
469 		return;
470 
471 	be16_add_cpu(&cqe->db_cnt, -dbd);
472 	cqe->wqe      = new_wqe;
473 	cqe->syndrome = SYNDROME_WR_FLUSH_ERR;
474 
475 	*free_cqe = 0;
476 }
477 
478 static inline int mthca_poll_one(struct mthca_dev *dev,
479 				 struct mthca_cq *cq,
480 				 struct mthca_qp **cur_qp,
481 				 int *freed,
482 				 struct ib_wc *entry)
483 {
484 	struct mthca_wq *wq;
485 	struct mthca_cqe *cqe;
486 	int wqe_index;
487 	int is_error;
488 	int is_send;
489 	int free_cqe = 1;
490 	int err = 0;
491 	u16 checksum;
492 
493 	cqe = next_cqe_sw(cq);
494 	if (!cqe)
495 		return -EAGAIN;
496 
497 	/*
498 	 * Make sure we read CQ entry contents after we've checked the
499 	 * ownership bit.
500 	 */
501 	rmb();
502 
503 	if (0) {
504 		mthca_dbg(dev, "%x/%d: CQE -> QPN %06x, WQE @ %08x\n",
505 			  cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
506 			  be32_to_cpu(cqe->wqe));
507 		dump_cqe(dev, cqe);
508 	}
509 
510 	is_error = (cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
511 		MTHCA_ERROR_CQE_OPCODE_MASK;
512 	is_send  = is_error ? cqe->opcode & 0x01 : cqe->is_send & 0x80;
513 
514 	if (!*cur_qp || be32_to_cpu(cqe->my_qpn) != (*cur_qp)->qpn) {
515 		/*
516 		 * We do not have to take the QP table lock here,
517 		 * because CQs will be locked while QPs are removed
518 		 * from the table.
519 		 */
520 		*cur_qp = mthca_array_get(&dev->qp_table.qp,
521 					  be32_to_cpu(cqe->my_qpn) &
522 					  (dev->limits.num_qps - 1));
523 		if (!*cur_qp) {
524 			mthca_warn(dev, "CQ entry for unknown QP %06x\n",
525 				   be32_to_cpu(cqe->my_qpn) & 0xffffff);
526 			err = -EINVAL;
527 			goto out;
528 		}
529 	}
530 
531 	entry->qp = &(*cur_qp)->ibqp;
532 
533 	if (is_send) {
534 		wq = &(*cur_qp)->sq;
535 		wqe_index = ((be32_to_cpu(cqe->wqe) - (*cur_qp)->send_wqe_offset)
536 			     >> wq->wqe_shift);
537 		entry->wr_id = (*cur_qp)->wrid[wqe_index +
538 					       (*cur_qp)->rq.max];
539 	} else if ((*cur_qp)->ibqp.srq) {
540 		struct mthca_srq *srq = to_msrq((*cur_qp)->ibqp.srq);
541 		u32 wqe = be32_to_cpu(cqe->wqe);
542 		wq = NULL;
543 		wqe_index = wqe >> srq->wqe_shift;
544 		entry->wr_id = srq->wrid[wqe_index];
545 		mthca_free_srq_wqe(srq, wqe);
546 	} else {
547 		s32 wqe;
548 		wq = &(*cur_qp)->rq;
549 		wqe = be32_to_cpu(cqe->wqe);
550 		wqe_index = wqe >> wq->wqe_shift;
551 		/*
552 		 * WQE addr == base - 1 might be reported in receive completion
553 		 * with error instead of (rq size - 1) by Sinai FW 1.0.800 and
554 		 * Arbel FW 5.1.400.  This bug should be fixed in later FW revs.
555 		 */
556 		if (unlikely(wqe_index < 0))
557 			wqe_index = wq->max - 1;
558 		entry->wr_id = (*cur_qp)->wrid[wqe_index];
559 	}
560 
561 	if (wq) {
562 		if (wq->last_comp < wqe_index)
563 			wq->tail += wqe_index - wq->last_comp;
564 		else
565 			wq->tail += wqe_index + wq->max - wq->last_comp;
566 
567 		wq->last_comp = wqe_index;
568 	}
569 
570 	if (is_error) {
571 		handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,
572 				 (struct mthca_err_cqe *) cqe,
573 				 entry, &free_cqe);
574 		goto out;
575 	}
576 
577 	if (is_send) {
578 		entry->wc_flags = 0;
579 		switch (cqe->opcode) {
580 		case MTHCA_OPCODE_RDMA_WRITE:
581 			entry->opcode    = IB_WC_RDMA_WRITE;
582 			break;
583 		case MTHCA_OPCODE_RDMA_WRITE_IMM:
584 			entry->opcode    = IB_WC_RDMA_WRITE;
585 			entry->wc_flags |= IB_WC_WITH_IMM;
586 			break;
587 		case MTHCA_OPCODE_SEND:
588 			entry->opcode    = IB_WC_SEND;
589 			break;
590 		case MTHCA_OPCODE_SEND_IMM:
591 			entry->opcode    = IB_WC_SEND;
592 			entry->wc_flags |= IB_WC_WITH_IMM;
593 			break;
594 		case MTHCA_OPCODE_RDMA_READ:
595 			entry->opcode    = IB_WC_RDMA_READ;
596 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
597 			break;
598 		case MTHCA_OPCODE_ATOMIC_CS:
599 			entry->opcode    = IB_WC_COMP_SWAP;
600 			entry->byte_len  = MTHCA_ATOMIC_BYTE_LEN;
601 			break;
602 		case MTHCA_OPCODE_ATOMIC_FA:
603 			entry->opcode    = IB_WC_FETCH_ADD;
604 			entry->byte_len  = MTHCA_ATOMIC_BYTE_LEN;
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->ex.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->ex.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   	   = cqe->sl_ipok >> 4;
632 		entry->src_qp 	   = be32_to_cpu(cqe->rqpn) & 0xffffff;
633 		entry->dlid_path_bits = cqe->g_mlpath & 0x7f;
634 		entry->pkey_index  = be32_to_cpu(cqe->imm_etype_pkey_eec) >> 16;
635 		entry->wc_flags   |= cqe->g_mlpath & 0x80 ? IB_WC_GRH : 0;
636 		checksum = (be32_to_cpu(cqe->rqpn) >> 24) |
637 				((be32_to_cpu(cqe->my_ee) >> 16) & 0xff00);
638 		entry->wc_flags	  |=  (cqe->sl_ipok & 1 && checksum == 0xffff) ?
639 							IB_WC_IP_CSUM_OK : 0;
640 	}
641 
642 	entry->status = IB_WC_SUCCESS;
643 
644  out:
645 	if (likely(free_cqe)) {
646 		set_cqe_hw(cqe);
647 		++(*freed);
648 		++cq->cons_index;
649 	}
650 
651 	return err;
652 }
653 
654 int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
655 		  struct ib_wc *entry)
656 {
657 	struct mthca_dev *dev = to_mdev(ibcq->device);
658 	struct mthca_cq *cq = to_mcq(ibcq);
659 	struct mthca_qp *qp = NULL;
660 	unsigned long flags;
661 	int err = 0;
662 	int freed = 0;
663 	int npolled;
664 
665 	spin_lock_irqsave(&cq->lock, flags);
666 
667 	npolled = 0;
668 repoll:
669 	while (npolled < num_entries) {
670 		err = mthca_poll_one(dev, cq, &qp,
671 				     &freed, entry + npolled);
672 		if (err)
673 			break;
674 		++npolled;
675 	}
676 
677 	if (freed) {
678 		wmb();
679 		update_cons_index(dev, cq, freed);
680 	}
681 
682 	/*
683 	 * If a CQ resize is in progress and we discovered that the
684 	 * old buffer is empty, then peek in the new buffer, and if
685 	 * it's not empty, switch to the new buffer and continue
686 	 * polling there.
687 	 */
688 	if (unlikely(err == -EAGAIN && cq->resize_buf &&
689 		     cq->resize_buf->state == CQ_RESIZE_READY)) {
690 		/*
691 		 * In Tavor mode, the hardware keeps the producer
692 		 * index modulo the CQ size.  Since we might be making
693 		 * the CQ bigger, we need to mask our consumer index
694 		 * using the size of the old CQ buffer before looking
695 		 * in the new CQ buffer.
696 		 */
697 		if (!mthca_is_memfree(dev))
698 			cq->cons_index &= cq->ibcq.cqe;
699 
700 		if (cqe_sw(get_cqe_from_buf(&cq->resize_buf->buf,
701 					    cq->cons_index & cq->resize_buf->cqe))) {
702 			struct mthca_cq_buf tbuf;
703 			int tcqe;
704 
705 			tbuf         = cq->buf;
706 			tcqe         = cq->ibcq.cqe;
707 			cq->buf      = cq->resize_buf->buf;
708 			cq->ibcq.cqe = cq->resize_buf->cqe;
709 
710 			cq->resize_buf->buf   = tbuf;
711 			cq->resize_buf->cqe   = tcqe;
712 			cq->resize_buf->state = CQ_RESIZE_SWAPPED;
713 
714 			goto repoll;
715 		}
716 	}
717 
718 	spin_unlock_irqrestore(&cq->lock, flags);
719 
720 	return err == 0 || err == -EAGAIN ? npolled : err;
721 }
722 
723 int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify_flags flags)
724 {
725 	u32 dbhi = ((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
726 		    MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL :
727 		    MTHCA_TAVOR_CQ_DB_REQ_NOT) |
728 		to_mcq(cq)->cqn;
729 
730 	mthca_write64(dbhi, 0xffffffff, to_mdev(cq->device)->kar + MTHCA_CQ_DOORBELL,
731 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(cq->device)->doorbell_lock));
732 
733 	return 0;
734 }
735 
736 int mthca_arbel_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
737 {
738 	struct mthca_cq *cq = to_mcq(ibcq);
739 	__be32 db_rec[2];
740 	u32 dbhi;
741 	u32 sn = cq->arm_sn & 3;
742 
743 	db_rec[0] = cpu_to_be32(cq->cons_index);
744 	db_rec[1] = cpu_to_be32((cq->cqn << 8) | (2 << 5) | (sn << 3) |
745 				((flags & IB_CQ_SOLICITED_MASK) ==
746 				 IB_CQ_SOLICITED ? 1 : 2));
747 
748 	mthca_write_db_rec(db_rec, cq->arm_db);
749 
750 	/*
751 	 * Make sure that the doorbell record in host memory is
752 	 * written before ringing the doorbell via PCI MMIO.
753 	 */
754 	wmb();
755 
756 	dbhi = (sn << 28) |
757 		((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
758 		 MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL :
759 		 MTHCA_ARBEL_CQ_DB_REQ_NOT) | cq->cqn;
760 
761 	mthca_write64(dbhi, cq->cons_index,
762 		      to_mdev(ibcq->device)->kar + MTHCA_CQ_DOORBELL,
763 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->doorbell_lock));
764 
765 	return 0;
766 }
767 
768 int mthca_init_cq(struct mthca_dev *dev, int nent,
769 		  struct mthca_ucontext *ctx, u32 pdn,
770 		  struct mthca_cq *cq)
771 {
772 	struct mthca_mailbox *mailbox;
773 	struct mthca_cq_context *cq_context;
774 	int err = -ENOMEM;
775 
776 	cq->ibcq.cqe  = nent - 1;
777 	cq->is_kernel = !ctx;
778 
779 	cq->cqn = mthca_alloc(&dev->cq_table.alloc);
780 	if (cq->cqn == -1)
781 		return -ENOMEM;
782 
783 	if (mthca_is_memfree(dev)) {
784 		err = mthca_table_get(dev, dev->cq_table.table, cq->cqn);
785 		if (err)
786 			goto err_out;
787 
788 		if (cq->is_kernel) {
789 			cq->arm_sn = 1;
790 
791 			err = -ENOMEM;
792 
793 			cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
794 							     cq->cqn, &cq->set_ci_db);
795 			if (cq->set_ci_db_index < 0)
796 				goto err_out_icm;
797 
798 			cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
799 							  cq->cqn, &cq->arm_db);
800 			if (cq->arm_db_index < 0)
801 				goto err_out_ci;
802 		}
803 	}
804 
805 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
806 	if (IS_ERR(mailbox))
807 		goto err_out_arm;
808 
809 	cq_context = mailbox->buf;
810 
811 	if (cq->is_kernel) {
812 		err = mthca_alloc_cq_buf(dev, &cq->buf, nent);
813 		if (err)
814 			goto err_out_mailbox;
815 	}
816 
817 	spin_lock_init(&cq->lock);
818 	cq->refcount = 1;
819 	init_waitqueue_head(&cq->wait);
820 	mutex_init(&cq->mutex);
821 
822 	memset(cq_context, 0, sizeof *cq_context);
823 	cq_context->flags           = cpu_to_be32(MTHCA_CQ_STATUS_OK      |
824 						  MTHCA_CQ_STATE_DISARMED |
825 						  MTHCA_CQ_FLAG_TR);
826 	cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24);
827 	if (ctx)
828 		cq_context->logsize_usrpage |= cpu_to_be32(ctx->uar.index);
829 	else
830 		cq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
831 	cq_context->error_eqn       = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
832 	cq_context->comp_eqn        = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
833 	cq_context->pd              = cpu_to_be32(pdn);
834 	cq_context->lkey            = cpu_to_be32(cq->buf.mr.ibmr.lkey);
835 	cq_context->cqn             = cpu_to_be32(cq->cqn);
836 
837 	if (mthca_is_memfree(dev)) {
838 		cq_context->ci_db    = cpu_to_be32(cq->set_ci_db_index);
839 		cq_context->state_db = cpu_to_be32(cq->arm_db_index);
840 	}
841 
842 	err = mthca_SW2HW_CQ(dev, mailbox, cq->cqn);
843 	if (err) {
844 		mthca_warn(dev, "SW2HW_CQ failed (%d)\n", err);
845 		goto err_out_free_mr;
846 	}
847 
848 	spin_lock_irq(&dev->cq_table.lock);
849 	if (mthca_array_set(&dev->cq_table.cq,
850 			    cq->cqn & (dev->limits.num_cqs - 1),
851 			    cq)) {
852 		spin_unlock_irq(&dev->cq_table.lock);
853 		goto err_out_free_mr;
854 	}
855 	spin_unlock_irq(&dev->cq_table.lock);
856 
857 	cq->cons_index = 0;
858 
859 	mthca_free_mailbox(dev, mailbox);
860 
861 	return 0;
862 
863 err_out_free_mr:
864 	if (cq->is_kernel)
865 		mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
866 
867 err_out_mailbox:
868 	mthca_free_mailbox(dev, mailbox);
869 
870 err_out_arm:
871 	if (cq->is_kernel && mthca_is_memfree(dev))
872 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
873 
874 err_out_ci:
875 	if (cq->is_kernel && mthca_is_memfree(dev))
876 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
877 
878 err_out_icm:
879 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
880 
881 err_out:
882 	mthca_free(&dev->cq_table.alloc, cq->cqn);
883 
884 	return err;
885 }
886 
887 static inline int get_cq_refcount(struct mthca_dev *dev, struct mthca_cq *cq)
888 {
889 	int c;
890 
891 	spin_lock_irq(&dev->cq_table.lock);
892 	c = cq->refcount;
893 	spin_unlock_irq(&dev->cq_table.lock);
894 
895 	return c;
896 }
897 
898 void mthca_free_cq(struct mthca_dev *dev,
899 		   struct mthca_cq *cq)
900 {
901 	struct mthca_mailbox *mailbox;
902 	int err;
903 
904 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
905 	if (IS_ERR(mailbox)) {
906 		mthca_warn(dev, "No memory for mailbox to free CQ.\n");
907 		return;
908 	}
909 
910 	err = mthca_HW2SW_CQ(dev, mailbox, cq->cqn);
911 	if (err)
912 		mthca_warn(dev, "HW2SW_CQ failed (%d)\n", err);
913 
914 	if (0) {
915 		__be32 *ctx = mailbox->buf;
916 		int j;
917 
918 		printk(KERN_ERR "context for CQN %x (cons index %x, next sw %d)\n",
919 		       cq->cqn, cq->cons_index,
920 		       cq->is_kernel ? !!next_cqe_sw(cq) : 0);
921 		for (j = 0; j < 16; ++j)
922 			printk(KERN_ERR "[%2x] %08x\n", j * 4, be32_to_cpu(ctx[j]));
923 	}
924 
925 	spin_lock_irq(&dev->cq_table.lock);
926 	mthca_array_clear(&dev->cq_table.cq,
927 			  cq->cqn & (dev->limits.num_cqs - 1));
928 	--cq->refcount;
929 	spin_unlock_irq(&dev->cq_table.lock);
930 
931 	if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
932 		synchronize_irq(dev->eq_table.eq[MTHCA_EQ_COMP].msi_x_vector);
933 	else
934 		synchronize_irq(dev->pdev->irq);
935 
936 	wait_event(cq->wait, !get_cq_refcount(dev, cq));
937 
938 	if (cq->is_kernel) {
939 		mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
940 		if (mthca_is_memfree(dev)) {
941 			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM,    cq->arm_db_index);
942 			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
943 		}
944 	}
945 
946 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
947 	mthca_free(&dev->cq_table.alloc, cq->cqn);
948 	mthca_free_mailbox(dev, mailbox);
949 }
950 
951 int mthca_init_cq_table(struct mthca_dev *dev)
952 {
953 	int err;
954 
955 	spin_lock_init(&dev->cq_table.lock);
956 
957 	err = mthca_alloc_init(&dev->cq_table.alloc,
958 			       dev->limits.num_cqs,
959 			       (1 << 24) - 1,
960 			       dev->limits.reserved_cqs);
961 	if (err)
962 		return err;
963 
964 	err = mthca_array_init(&dev->cq_table.cq,
965 			       dev->limits.num_cqs);
966 	if (err)
967 		mthca_alloc_cleanup(&dev->cq_table.alloc);
968 
969 	return err;
970 }
971 
972 void mthca_cleanup_cq_table(struct mthca_dev *dev)
973 {
974 	mthca_array_cleanup(&dev->cq_table.cq, dev->limits.num_cqs);
975 	mthca_alloc_cleanup(&dev->cq_table.alloc);
976 }
977