xref: /linux/drivers/infiniband/hw/mthca/mthca_cq.c (revision 13abf8130139c2ccd4962a7e5a8902be5e6cb5a7)
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 Cisco Systems, Inc. All rights reserved.
5  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
6  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
7  *
8  * This software is available to you under a choice of one of two
9  * licenses.  You may choose to be licensed under the terms of the GNU
10  * General Public License (GPL) Version 2, available from the file
11  * COPYING in the main directory of this source tree, or the
12  * OpenIB.org BSD license below:
13  *
14  *     Redistribution and use in source and binary forms, with or
15  *     without modification, are permitted provided that the following
16  *     conditions are met:
17  *
18  *      - Redistributions of source code must retain the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer.
21  *
22  *      - Redistributions in binary form must reproduce the above
23  *        copyright notice, this list of conditions and the following
24  *        disclaimer in the documentation and/or other materials
25  *        provided with the distribution.
26  *
27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34  * SOFTWARE.
35  *
36  * $Id: mthca_cq.c 1369 2004-12-20 16:17:07Z roland $
37  */
38 
39 #include <linux/init.h>
40 #include <linux/hardirq.h>
41 
42 #include <rdma/ib_pack.h>
43 
44 #include "mthca_dev.h"
45 #include "mthca_cmd.h"
46 #include "mthca_memfree.h"
47 
48 enum {
49 	MTHCA_MAX_DIRECT_CQ_SIZE = 4 * PAGE_SIZE
50 };
51 
52 enum {
53 	MTHCA_CQ_ENTRY_SIZE = 0x20
54 };
55 
56 /*
57  * Must be packed because start is 64 bits but only aligned to 32 bits.
58  */
59 struct mthca_cq_context {
60 	__be32 flags;
61 	__be64 start;
62 	__be32 logsize_usrpage;
63 	__be32 error_eqn;	/* Tavor only */
64 	__be32 comp_eqn;
65 	__be32 pd;
66 	__be32 lkey;
67 	__be32 last_notified_index;
68 	__be32 solicit_producer_index;
69 	__be32 consumer_index;
70 	__be32 producer_index;
71 	__be32 cqn;
72 	__be32 ci_db;		/* Arbel only */
73 	__be32 state_db;	/* Arbel only */
74 	u32    reserved;
75 } __attribute__((packed));
76 
77 #define MTHCA_CQ_STATUS_OK          ( 0 << 28)
78 #define MTHCA_CQ_STATUS_OVERFLOW    ( 9 << 28)
79 #define MTHCA_CQ_STATUS_WRITE_FAIL  (10 << 28)
80 #define MTHCA_CQ_FLAG_TR            ( 1 << 18)
81 #define MTHCA_CQ_FLAG_OI            ( 1 << 17)
82 #define MTHCA_CQ_STATE_DISARMED     ( 0 <<  8)
83 #define MTHCA_CQ_STATE_ARMED        ( 1 <<  8)
84 #define MTHCA_CQ_STATE_ARMED_SOL    ( 4 <<  8)
85 #define MTHCA_EQ_STATE_FIRED        (10 <<  8)
86 
87 enum {
88 	MTHCA_ERROR_CQE_OPCODE_MASK = 0xfe
89 };
90 
91 enum {
92 	SYNDROME_LOCAL_LENGTH_ERR 	 = 0x01,
93 	SYNDROME_LOCAL_QP_OP_ERR  	 = 0x02,
94 	SYNDROME_LOCAL_EEC_OP_ERR 	 = 0x03,
95 	SYNDROME_LOCAL_PROT_ERR   	 = 0x04,
96 	SYNDROME_WR_FLUSH_ERR     	 = 0x05,
97 	SYNDROME_MW_BIND_ERR      	 = 0x06,
98 	SYNDROME_BAD_RESP_ERR     	 = 0x10,
99 	SYNDROME_LOCAL_ACCESS_ERR 	 = 0x11,
100 	SYNDROME_REMOTE_INVAL_REQ_ERR 	 = 0x12,
101 	SYNDROME_REMOTE_ACCESS_ERR 	 = 0x13,
102 	SYNDROME_REMOTE_OP_ERR     	 = 0x14,
103 	SYNDROME_RETRY_EXC_ERR 		 = 0x15,
104 	SYNDROME_RNR_RETRY_EXC_ERR 	 = 0x16,
105 	SYNDROME_LOCAL_RDD_VIOL_ERR 	 = 0x20,
106 	SYNDROME_REMOTE_INVAL_RD_REQ_ERR = 0x21,
107 	SYNDROME_REMOTE_ABORTED_ERR 	 = 0x22,
108 	SYNDROME_INVAL_EECN_ERR 	 = 0x23,
109 	SYNDROME_INVAL_EEC_STATE_ERR 	 = 0x24
110 };
111 
112 struct mthca_cqe {
113 	__be32 my_qpn;
114 	__be32 my_ee;
115 	__be32 rqpn;
116 	__be16 sl_g_mlpath;
117 	__be16 rlid;
118 	__be32 imm_etype_pkey_eec;
119 	__be32 byte_cnt;
120 	__be32 wqe;
121 	u8     opcode;
122 	u8     is_send;
123 	u8     reserved;
124 	u8     owner;
125 };
126 
127 struct mthca_err_cqe {
128 	__be32 my_qpn;
129 	u32    reserved1[3];
130 	u8     syndrome;
131 	u8     reserved2;
132 	__be16 db_cnt;
133 	u32    reserved3;
134 	__be32 wqe;
135 	u8     opcode;
136 	u8     reserved4[2];
137 	u8     owner;
138 };
139 
140 #define MTHCA_CQ_ENTRY_OWNER_SW      (0 << 7)
141 #define MTHCA_CQ_ENTRY_OWNER_HW      (1 << 7)
142 
143 #define MTHCA_TAVOR_CQ_DB_INC_CI       (1 << 24)
144 #define MTHCA_TAVOR_CQ_DB_REQ_NOT      (2 << 24)
145 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL  (3 << 24)
146 #define MTHCA_TAVOR_CQ_DB_SET_CI       (4 << 24)
147 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_MULT (5 << 24)
148 
149 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL  (1 << 24)
150 #define MTHCA_ARBEL_CQ_DB_REQ_NOT      (2 << 24)
151 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_MULT (3 << 24)
152 
153 static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
154 {
155 	if (cq->is_direct)
156 		return cq->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
157 	else
158 		return cq->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
159 			+ (entry * MTHCA_CQ_ENTRY_SIZE) % PAGE_SIZE;
160 }
161 
162 static inline struct mthca_cqe *cqe_sw(struct mthca_cq *cq, int i)
163 {
164 	struct mthca_cqe *cqe = get_cqe(cq, i);
165 	return MTHCA_CQ_ENTRY_OWNER_HW & cqe->owner ? NULL : cqe;
166 }
167 
168 static inline struct mthca_cqe *next_cqe_sw(struct mthca_cq *cq)
169 {
170 	return cqe_sw(cq, cq->cons_index & cq->ibcq.cqe);
171 }
172 
173 static inline void set_cqe_hw(struct mthca_cqe *cqe)
174 {
175 	cqe->owner = MTHCA_CQ_ENTRY_OWNER_HW;
176 }
177 
178 static void dump_cqe(struct mthca_dev *dev, void *cqe_ptr)
179 {
180 	__be32 *cqe = cqe_ptr;
181 
182 	(void) cqe;	/* avoid warning if mthca_dbg compiled away... */
183 	mthca_dbg(dev, "CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
184 		  be32_to_cpu(cqe[0]), be32_to_cpu(cqe[1]), be32_to_cpu(cqe[2]),
185 		  be32_to_cpu(cqe[3]), be32_to_cpu(cqe[4]), be32_to_cpu(cqe[5]),
186 		  be32_to_cpu(cqe[6]), be32_to_cpu(cqe[7]));
187 }
188 
189 /*
190  * incr is ignored in native Arbel (mem-free) mode, so cq->cons_index
191  * should be correct before calling update_cons_index().
192  */
193 static inline void update_cons_index(struct mthca_dev *dev, struct mthca_cq *cq,
194 				     int incr)
195 {
196 	__be32 doorbell[2];
197 
198 	if (mthca_is_memfree(dev)) {
199 		*cq->set_ci_db = cpu_to_be32(cq->cons_index);
200 		wmb();
201 	} else {
202 		doorbell[0] = cpu_to_be32(MTHCA_TAVOR_CQ_DB_INC_CI | cq->cqn);
203 		doorbell[1] = cpu_to_be32(incr - 1);
204 
205 		mthca_write64(doorbell,
206 			      dev->kar + MTHCA_CQ_DOORBELL,
207 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
208 	}
209 }
210 
211 void mthca_cq_event(struct mthca_dev *dev, u32 cqn)
212 {
213 	struct mthca_cq *cq;
214 
215 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
216 
217 	if (!cq) {
218 		mthca_warn(dev, "Completion event for bogus CQ %08x\n", cqn);
219 		return;
220 	}
221 
222 	++cq->arm_sn;
223 
224 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
225 }
226 
227 void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn,
228 		    struct mthca_srq *srq)
229 {
230 	struct mthca_cq *cq;
231 	struct mthca_cqe *cqe;
232 	int prod_index;
233 	int nfreed = 0;
234 
235 	spin_lock_irq(&dev->cq_table.lock);
236 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
237 	if (cq)
238 		atomic_inc(&cq->refcount);
239 	spin_unlock_irq(&dev->cq_table.lock);
240 
241 	if (!cq)
242 		return;
243 
244 	spin_lock_irq(&cq->lock);
245 
246 	/*
247 	 * First we need to find the current producer index, so we
248 	 * know where to start cleaning from.  It doesn't matter if HW
249 	 * adds new entries after this loop -- the QP we're worried
250 	 * about is already in RESET, so the new entries won't come
251 	 * from our QP and therefore don't need to be checked.
252 	 */
253 	for (prod_index = cq->cons_index;
254 	     cqe_sw(cq, prod_index & cq->ibcq.cqe);
255 	     ++prod_index)
256 		if (prod_index == cq->cons_index + cq->ibcq.cqe)
257 			break;
258 
259 	if (0)
260 		mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
261 			  qpn, cqn, cq->cons_index, prod_index);
262 
263 	/*
264 	 * Now sweep backwards through the CQ, removing CQ entries
265 	 * that match our QP by copying older entries on top of them.
266 	 */
267 	while (prod_index > cq->cons_index) {
268 		cqe = get_cqe(cq, (prod_index - 1) & cq->ibcq.cqe);
269 		if (cqe->my_qpn == cpu_to_be32(qpn)) {
270 			if (srq)
271 				mthca_free_srq_wqe(srq, be32_to_cpu(cqe->wqe));
272 			++nfreed;
273 		}
274 		else if (nfreed)
275 			memcpy(get_cqe(cq, (prod_index - 1 + nfreed) &
276 				       cq->ibcq.cqe),
277 			       cqe,
278 			       MTHCA_CQ_ENTRY_SIZE);
279 		--prod_index;
280 	}
281 
282 	if (nfreed) {
283 		wmb();
284 		cq->cons_index += nfreed;
285 		update_cons_index(dev, cq, nfreed);
286 	}
287 
288 	spin_unlock_irq(&cq->lock);
289 	if (atomic_dec_and_test(&cq->refcount))
290 		wake_up(&cq->wait);
291 }
292 
293 static int handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
294 			    struct mthca_qp *qp, int wqe_index, int is_send,
295 			    struct mthca_err_cqe *cqe,
296 			    struct ib_wc *entry, int *free_cqe)
297 {
298 	int err;
299 	int dbd;
300 	__be32 new_wqe;
301 
302 	if (cqe->syndrome == SYNDROME_LOCAL_QP_OP_ERR) {
303 		mthca_dbg(dev, "local QP operation err "
304 			  "(QPN %06x, WQE @ %08x, CQN %06x, index %d)\n",
305 			  be32_to_cpu(cqe->my_qpn), be32_to_cpu(cqe->wqe),
306 			  cq->cqn, cq->cons_index);
307 		dump_cqe(dev, cqe);
308 	}
309 
310 	/*
311 	 * For completions in error, only work request ID, status (and
312 	 * freed resource count for RD) have to be set.
313 	 */
314 	switch (cqe->syndrome) {
315 	case SYNDROME_LOCAL_LENGTH_ERR:
316 		entry->status = IB_WC_LOC_LEN_ERR;
317 		break;
318 	case SYNDROME_LOCAL_QP_OP_ERR:
319 		entry->status = IB_WC_LOC_QP_OP_ERR;
320 		break;
321 	case SYNDROME_LOCAL_EEC_OP_ERR:
322 		entry->status = IB_WC_LOC_EEC_OP_ERR;
323 		break;
324 	case SYNDROME_LOCAL_PROT_ERR:
325 		entry->status = IB_WC_LOC_PROT_ERR;
326 		break;
327 	case SYNDROME_WR_FLUSH_ERR:
328 		entry->status = IB_WC_WR_FLUSH_ERR;
329 		break;
330 	case SYNDROME_MW_BIND_ERR:
331 		entry->status = IB_WC_MW_BIND_ERR;
332 		break;
333 	case SYNDROME_BAD_RESP_ERR:
334 		entry->status = IB_WC_BAD_RESP_ERR;
335 		break;
336 	case SYNDROME_LOCAL_ACCESS_ERR:
337 		entry->status = IB_WC_LOC_ACCESS_ERR;
338 		break;
339 	case SYNDROME_REMOTE_INVAL_REQ_ERR:
340 		entry->status = IB_WC_REM_INV_REQ_ERR;
341 		break;
342 	case SYNDROME_REMOTE_ACCESS_ERR:
343 		entry->status = IB_WC_REM_ACCESS_ERR;
344 		break;
345 	case SYNDROME_REMOTE_OP_ERR:
346 		entry->status = IB_WC_REM_OP_ERR;
347 		break;
348 	case SYNDROME_RETRY_EXC_ERR:
349 		entry->status = IB_WC_RETRY_EXC_ERR;
350 		break;
351 	case SYNDROME_RNR_RETRY_EXC_ERR:
352 		entry->status = IB_WC_RNR_RETRY_EXC_ERR;
353 		break;
354 	case SYNDROME_LOCAL_RDD_VIOL_ERR:
355 		entry->status = IB_WC_LOC_RDD_VIOL_ERR;
356 		break;
357 	case SYNDROME_REMOTE_INVAL_RD_REQ_ERR:
358 		entry->status = IB_WC_REM_INV_RD_REQ_ERR;
359 		break;
360 	case SYNDROME_REMOTE_ABORTED_ERR:
361 		entry->status = IB_WC_REM_ABORT_ERR;
362 		break;
363 	case SYNDROME_INVAL_EECN_ERR:
364 		entry->status = IB_WC_INV_EECN_ERR;
365 		break;
366 	case SYNDROME_INVAL_EEC_STATE_ERR:
367 		entry->status = IB_WC_INV_EEC_STATE_ERR;
368 		break;
369 	default:
370 		entry->status = IB_WC_GENERAL_ERR;
371 		break;
372 	}
373 
374 	/*
375 	 * Mem-free HCAs always generate one CQE per WQE, even in the
376 	 * error case, so we don't have to check the doorbell count, etc.
377 	 */
378 	if (mthca_is_memfree(dev))
379 		return 0;
380 
381 	err = mthca_free_err_wqe(dev, qp, is_send, wqe_index, &dbd, &new_wqe);
382 	if (err)
383 		return err;
384 
385 	/*
386 	 * If we're at the end of the WQE chain, or we've used up our
387 	 * doorbell count, free the CQE.  Otherwise just update it for
388 	 * the next poll operation.
389 	 */
390 	if (!(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
391 		return 0;
392 
393 	cqe->db_cnt   = cpu_to_be16(be16_to_cpu(cqe->db_cnt) - dbd);
394 	cqe->wqe      = new_wqe;
395 	cqe->syndrome = SYNDROME_WR_FLUSH_ERR;
396 
397 	*free_cqe = 0;
398 
399 	return 0;
400 }
401 
402 static inline int mthca_poll_one(struct mthca_dev *dev,
403 				 struct mthca_cq *cq,
404 				 struct mthca_qp **cur_qp,
405 				 int *freed,
406 				 struct ib_wc *entry)
407 {
408 	struct mthca_wq *wq;
409 	struct mthca_cqe *cqe;
410 	int wqe_index;
411 	int is_error;
412 	int is_send;
413 	int free_cqe = 1;
414 	int err = 0;
415 
416 	cqe = next_cqe_sw(cq);
417 	if (!cqe)
418 		return -EAGAIN;
419 
420 	/*
421 	 * Make sure we read CQ entry contents after we've checked the
422 	 * ownership bit.
423 	 */
424 	rmb();
425 
426 	if (0) {
427 		mthca_dbg(dev, "%x/%d: CQE -> QPN %06x, WQE @ %08x\n",
428 			  cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
429 			  be32_to_cpu(cqe->wqe));
430 		dump_cqe(dev, cqe);
431 	}
432 
433 	is_error = (cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
434 		MTHCA_ERROR_CQE_OPCODE_MASK;
435 	is_send  = is_error ? cqe->opcode & 0x01 : cqe->is_send & 0x80;
436 
437 	if (!*cur_qp || be32_to_cpu(cqe->my_qpn) != (*cur_qp)->qpn) {
438 		/*
439 		 * We do not have to take the QP table lock here,
440 		 * because CQs will be locked while QPs are removed
441 		 * from the table.
442 		 */
443 		*cur_qp = mthca_array_get(&dev->qp_table.qp,
444 					  be32_to_cpu(cqe->my_qpn) &
445 					  (dev->limits.num_qps - 1));
446 		if (!*cur_qp) {
447 			mthca_warn(dev, "CQ entry for unknown QP %06x\n",
448 				   be32_to_cpu(cqe->my_qpn) & 0xffffff);
449 			err = -EINVAL;
450 			goto out;
451 		}
452 	}
453 
454 	entry->qp_num = (*cur_qp)->qpn;
455 
456 	if (is_send) {
457 		wq = &(*cur_qp)->sq;
458 		wqe_index = ((be32_to_cpu(cqe->wqe) - (*cur_qp)->send_wqe_offset)
459 			     >> wq->wqe_shift);
460 		entry->wr_id = (*cur_qp)->wrid[wqe_index +
461 					       (*cur_qp)->rq.max];
462 	} else if ((*cur_qp)->ibqp.srq) {
463 		struct mthca_srq *srq = to_msrq((*cur_qp)->ibqp.srq);
464 		u32 wqe = be32_to_cpu(cqe->wqe);
465 		wq = NULL;
466 		wqe_index = wqe >> srq->wqe_shift;
467 		entry->wr_id = srq->wrid[wqe_index];
468 		mthca_free_srq_wqe(srq, wqe);
469 	} else {
470 		wq = &(*cur_qp)->rq;
471 		wqe_index = be32_to_cpu(cqe->wqe) >> wq->wqe_shift;
472 		entry->wr_id = (*cur_qp)->wrid[wqe_index];
473 	}
474 
475 	if (wq) {
476 		if (wq->last_comp < wqe_index)
477 			wq->tail += wqe_index - wq->last_comp;
478 		else
479 			wq->tail += wqe_index + wq->max - wq->last_comp;
480 
481 		wq->last_comp = wqe_index;
482 	}
483 
484 	if (is_error) {
485 		err = handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,
486 				       (struct mthca_err_cqe *) cqe,
487 				       entry, &free_cqe);
488 		goto out;
489 	}
490 
491 	if (is_send) {
492 		entry->wc_flags = 0;
493 		switch (cqe->opcode) {
494 		case MTHCA_OPCODE_RDMA_WRITE:
495 			entry->opcode    = IB_WC_RDMA_WRITE;
496 			break;
497 		case MTHCA_OPCODE_RDMA_WRITE_IMM:
498 			entry->opcode    = IB_WC_RDMA_WRITE;
499 			entry->wc_flags |= IB_WC_WITH_IMM;
500 			break;
501 		case MTHCA_OPCODE_SEND:
502 			entry->opcode    = IB_WC_SEND;
503 			break;
504 		case MTHCA_OPCODE_SEND_IMM:
505 			entry->opcode    = IB_WC_SEND;
506 			entry->wc_flags |= IB_WC_WITH_IMM;
507 			break;
508 		case MTHCA_OPCODE_RDMA_READ:
509 			entry->opcode    = IB_WC_RDMA_READ;
510 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
511 			break;
512 		case MTHCA_OPCODE_ATOMIC_CS:
513 			entry->opcode    = IB_WC_COMP_SWAP;
514 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
515 			break;
516 		case MTHCA_OPCODE_ATOMIC_FA:
517 			entry->opcode    = IB_WC_FETCH_ADD;
518 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
519 			break;
520 		case MTHCA_OPCODE_BIND_MW:
521 			entry->opcode    = IB_WC_BIND_MW;
522 			break;
523 		default:
524 			entry->opcode    = MTHCA_OPCODE_INVALID;
525 			break;
526 		}
527 	} else {
528 		entry->byte_len = be32_to_cpu(cqe->byte_cnt);
529 		switch (cqe->opcode & 0x1f) {
530 		case IB_OPCODE_SEND_LAST_WITH_IMMEDIATE:
531 		case IB_OPCODE_SEND_ONLY_WITH_IMMEDIATE:
532 			entry->wc_flags = IB_WC_WITH_IMM;
533 			entry->imm_data = cqe->imm_etype_pkey_eec;
534 			entry->opcode = IB_WC_RECV;
535 			break;
536 		case IB_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
537 		case IB_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
538 			entry->wc_flags = IB_WC_WITH_IMM;
539 			entry->imm_data = cqe->imm_etype_pkey_eec;
540 			entry->opcode = IB_WC_RECV_RDMA_WITH_IMM;
541 			break;
542 		default:
543 			entry->wc_flags = 0;
544 			entry->opcode = IB_WC_RECV;
545 			break;
546 		}
547 		entry->slid 	   = be16_to_cpu(cqe->rlid);
548 		entry->sl   	   = be16_to_cpu(cqe->sl_g_mlpath) >> 12;
549 		entry->src_qp 	   = be32_to_cpu(cqe->rqpn) & 0xffffff;
550 		entry->dlid_path_bits = be16_to_cpu(cqe->sl_g_mlpath) & 0x7f;
551 		entry->pkey_index  = be32_to_cpu(cqe->imm_etype_pkey_eec) >> 16;
552 		entry->wc_flags   |= be16_to_cpu(cqe->sl_g_mlpath) & 0x80 ?
553 					IB_WC_GRH : 0;
554 	}
555 
556 	entry->status = IB_WC_SUCCESS;
557 
558  out:
559 	if (likely(free_cqe)) {
560 		set_cqe_hw(cqe);
561 		++(*freed);
562 		++cq->cons_index;
563 	}
564 
565 	return err;
566 }
567 
568 int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
569 		  struct ib_wc *entry)
570 {
571 	struct mthca_dev *dev = to_mdev(ibcq->device);
572 	struct mthca_cq *cq = to_mcq(ibcq);
573 	struct mthca_qp *qp = NULL;
574 	unsigned long flags;
575 	int err = 0;
576 	int freed = 0;
577 	int npolled;
578 
579 	spin_lock_irqsave(&cq->lock, flags);
580 
581 	for (npolled = 0; npolled < num_entries; ++npolled) {
582 		err = mthca_poll_one(dev, cq, &qp,
583 				     &freed, entry + npolled);
584 		if (err)
585 			break;
586 	}
587 
588 	if (freed) {
589 		wmb();
590 		update_cons_index(dev, cq, freed);
591 	}
592 
593 	spin_unlock_irqrestore(&cq->lock, flags);
594 
595 	return err == 0 || err == -EAGAIN ? npolled : err;
596 }
597 
598 int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify notify)
599 {
600 	__be32 doorbell[2];
601 
602 	doorbell[0] = cpu_to_be32((notify == IB_CQ_SOLICITED ?
603 				   MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL :
604 				   MTHCA_TAVOR_CQ_DB_REQ_NOT)      |
605 				  to_mcq(cq)->cqn);
606 	doorbell[1] = (__force __be32) 0xffffffff;
607 
608 	mthca_write64(doorbell,
609 		      to_mdev(cq->device)->kar + MTHCA_CQ_DOORBELL,
610 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(cq->device)->doorbell_lock));
611 
612 	return 0;
613 }
614 
615 int mthca_arbel_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify notify)
616 {
617 	struct mthca_cq *cq = to_mcq(ibcq);
618 	__be32 doorbell[2];
619 	u32 sn;
620 	__be32 ci;
621 
622 	sn = cq->arm_sn & 3;
623 	ci = cpu_to_be32(cq->cons_index);
624 
625 	doorbell[0] = ci;
626 	doorbell[1] = cpu_to_be32((cq->cqn << 8) | (2 << 5) | (sn << 3) |
627 				  (notify == IB_CQ_SOLICITED ? 1 : 2));
628 
629 	mthca_write_db_rec(doorbell, cq->arm_db);
630 
631 	/*
632 	 * Make sure that the doorbell record in host memory is
633 	 * written before ringing the doorbell via PCI MMIO.
634 	 */
635 	wmb();
636 
637 	doorbell[0] = cpu_to_be32((sn << 28)                       |
638 				  (notify == IB_CQ_SOLICITED ?
639 				   MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL :
640 				   MTHCA_ARBEL_CQ_DB_REQ_NOT)      |
641 				  cq->cqn);
642 	doorbell[1] = ci;
643 
644 	mthca_write64(doorbell,
645 		      to_mdev(ibcq->device)->kar + MTHCA_CQ_DOORBELL,
646 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->doorbell_lock));
647 
648 	return 0;
649 }
650 
651 static void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq *cq)
652 {
653 	mthca_buf_free(dev, (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE,
654 		       &cq->queue, cq->is_direct, &cq->mr);
655 }
656 
657 int mthca_init_cq(struct mthca_dev *dev, int nent,
658 		  struct mthca_ucontext *ctx, u32 pdn,
659 		  struct mthca_cq *cq)
660 {
661 	int size = nent * MTHCA_CQ_ENTRY_SIZE;
662 	struct mthca_mailbox *mailbox;
663 	struct mthca_cq_context *cq_context;
664 	int err = -ENOMEM;
665 	u8 status;
666 	int i;
667 
668 	might_sleep();
669 
670 	cq->ibcq.cqe  = nent - 1;
671 	cq->is_kernel = !ctx;
672 
673 	cq->cqn = mthca_alloc(&dev->cq_table.alloc);
674 	if (cq->cqn == -1)
675 		return -ENOMEM;
676 
677 	if (mthca_is_memfree(dev)) {
678 		err = mthca_table_get(dev, dev->cq_table.table, cq->cqn);
679 		if (err)
680 			goto err_out;
681 
682 		if (cq->is_kernel) {
683 			cq->arm_sn = 1;
684 
685 			err = -ENOMEM;
686 
687 			cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
688 							     cq->cqn, &cq->set_ci_db);
689 			if (cq->set_ci_db_index < 0)
690 				goto err_out_icm;
691 
692 			cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
693 							  cq->cqn, &cq->arm_db);
694 			if (cq->arm_db_index < 0)
695 				goto err_out_ci;
696 		}
697 	}
698 
699 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
700 	if (IS_ERR(mailbox))
701 		goto err_out_arm;
702 
703 	cq_context = mailbox->buf;
704 
705 	if (cq->is_kernel) {
706 		err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_CQ_SIZE,
707 				      &cq->queue, &cq->is_direct,
708 				      &dev->driver_pd, 1, &cq->mr);
709 		if (err)
710 			goto err_out_mailbox;
711 
712 		for (i = 0; i < nent; ++i)
713 			set_cqe_hw(get_cqe(cq, i));
714 	}
715 
716 	spin_lock_init(&cq->lock);
717 	atomic_set(&cq->refcount, 1);
718 	init_waitqueue_head(&cq->wait);
719 
720 	memset(cq_context, 0, sizeof *cq_context);
721 	cq_context->flags           = cpu_to_be32(MTHCA_CQ_STATUS_OK      |
722 						  MTHCA_CQ_STATE_DISARMED |
723 						  MTHCA_CQ_FLAG_TR);
724 	cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24);
725 	if (ctx)
726 		cq_context->logsize_usrpage |= cpu_to_be32(ctx->uar.index);
727 	else
728 		cq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
729 	cq_context->error_eqn       = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
730 	cq_context->comp_eqn        = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
731 	cq_context->pd              = cpu_to_be32(pdn);
732 	cq_context->lkey            = cpu_to_be32(cq->mr.ibmr.lkey);
733 	cq_context->cqn             = cpu_to_be32(cq->cqn);
734 
735 	if (mthca_is_memfree(dev)) {
736 		cq_context->ci_db    = cpu_to_be32(cq->set_ci_db_index);
737 		cq_context->state_db = cpu_to_be32(cq->arm_db_index);
738 	}
739 
740 	err = mthca_SW2HW_CQ(dev, mailbox, cq->cqn, &status);
741 	if (err) {
742 		mthca_warn(dev, "SW2HW_CQ failed (%d)\n", err);
743 		goto err_out_free_mr;
744 	}
745 
746 	if (status) {
747 		mthca_warn(dev, "SW2HW_CQ returned status 0x%02x\n",
748 			   status);
749 		err = -EINVAL;
750 		goto err_out_free_mr;
751 	}
752 
753 	spin_lock_irq(&dev->cq_table.lock);
754 	if (mthca_array_set(&dev->cq_table.cq,
755 			    cq->cqn & (dev->limits.num_cqs - 1),
756 			    cq)) {
757 		spin_unlock_irq(&dev->cq_table.lock);
758 		goto err_out_free_mr;
759 	}
760 	spin_unlock_irq(&dev->cq_table.lock);
761 
762 	cq->cons_index = 0;
763 
764 	mthca_free_mailbox(dev, mailbox);
765 
766 	return 0;
767 
768 err_out_free_mr:
769 	if (cq->is_kernel)
770 		mthca_free_cq_buf(dev, cq);
771 
772 err_out_mailbox:
773 	mthca_free_mailbox(dev, mailbox);
774 
775 err_out_arm:
776 	if (cq->is_kernel && mthca_is_memfree(dev))
777 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
778 
779 err_out_ci:
780 	if (cq->is_kernel && mthca_is_memfree(dev))
781 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
782 
783 err_out_icm:
784 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
785 
786 err_out:
787 	mthca_free(&dev->cq_table.alloc, cq->cqn);
788 
789 	return err;
790 }
791 
792 void mthca_free_cq(struct mthca_dev *dev,
793 		   struct mthca_cq *cq)
794 {
795 	struct mthca_mailbox *mailbox;
796 	int err;
797 	u8 status;
798 
799 	might_sleep();
800 
801 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
802 	if (IS_ERR(mailbox)) {
803 		mthca_warn(dev, "No memory for mailbox to free CQ.\n");
804 		return;
805 	}
806 
807 	err = mthca_HW2SW_CQ(dev, mailbox, cq->cqn, &status);
808 	if (err)
809 		mthca_warn(dev, "HW2SW_CQ failed (%d)\n", err);
810 	else if (status)
811 		mthca_warn(dev, "HW2SW_CQ returned status 0x%02x\n", status);
812 
813 	if (0) {
814 		__be32 *ctx = mailbox->buf;
815 		int j;
816 
817 		printk(KERN_ERR "context for CQN %x (cons index %x, next sw %d)\n",
818 		       cq->cqn, cq->cons_index,
819 		       cq->is_kernel ? !!next_cqe_sw(cq) : 0);
820 		for (j = 0; j < 16; ++j)
821 			printk(KERN_ERR "[%2x] %08x\n", j * 4, be32_to_cpu(ctx[j]));
822 	}
823 
824 	spin_lock_irq(&dev->cq_table.lock);
825 	mthca_array_clear(&dev->cq_table.cq,
826 			  cq->cqn & (dev->limits.num_cqs - 1));
827 	spin_unlock_irq(&dev->cq_table.lock);
828 
829 	if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
830 		synchronize_irq(dev->eq_table.eq[MTHCA_EQ_COMP].msi_x_vector);
831 	else
832 		synchronize_irq(dev->pdev->irq);
833 
834 	atomic_dec(&cq->refcount);
835 	wait_event(cq->wait, !atomic_read(&cq->refcount));
836 
837 	if (cq->is_kernel) {
838 		mthca_free_cq_buf(dev, cq);
839 		if (mthca_is_memfree(dev)) {
840 			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM,    cq->arm_db_index);
841 			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
842 		}
843 	}
844 
845 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
846 	mthca_free(&dev->cq_table.alloc, cq->cqn);
847 	mthca_free_mailbox(dev, mailbox);
848 }
849 
850 int __devinit mthca_init_cq_table(struct mthca_dev *dev)
851 {
852 	int err;
853 
854 	spin_lock_init(&dev->cq_table.lock);
855 
856 	err = mthca_alloc_init(&dev->cq_table.alloc,
857 			       dev->limits.num_cqs,
858 			       (1 << 24) - 1,
859 			       dev->limits.reserved_cqs);
860 	if (err)
861 		return err;
862 
863 	err = mthca_array_init(&dev->cq_table.cq,
864 			       dev->limits.num_cqs);
865 	if (err)
866 		mthca_alloc_cleanup(&dev->cq_table.alloc);
867 
868 	return err;
869 }
870 
871 void __devexit mthca_cleanup_cq_table(struct mthca_dev *dev)
872 {
873 	mthca_array_cleanup(&dev->cq_table.cq, dev->limits.num_cqs);
874 	mthca_alloc_cleanup(&dev->cq_table.alloc);
875 }
876