xref: /linux/drivers/infiniband/hw/mthca/mthca_cq.c (revision 20d0021394c1b070bf04b22c5bc8fdb437edd4c5)
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  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  *
34  * $Id: mthca_cq.c 1369 2004-12-20 16:17:07Z roland $
35  */
36 
37 #include <linux/init.h>
38 #include <linux/hardirq.h>
39 
40 #include <ib_pack.h>
41 
42 #include "mthca_dev.h"
43 #include "mthca_cmd.h"
44 #include "mthca_memfree.h"
45 
46 enum {
47 	MTHCA_MAX_DIRECT_CQ_SIZE = 4 * PAGE_SIZE
48 };
49 
50 enum {
51 	MTHCA_CQ_ENTRY_SIZE = 0x20
52 };
53 
54 /*
55  * Must be packed because start is 64 bits but only aligned to 32 bits.
56  */
57 struct mthca_cq_context {
58 	u32 flags;
59 	u64 start;
60 	u32 logsize_usrpage;
61 	u32 error_eqn;		/* Tavor only */
62 	u32 comp_eqn;
63 	u32 pd;
64 	u32 lkey;
65 	u32 last_notified_index;
66 	u32 solicit_producer_index;
67 	u32 consumer_index;
68 	u32 producer_index;
69 	u32 cqn;
70 	u32 ci_db;		/* Arbel only */
71 	u32 state_db;		/* Arbel only */
72 	u32 reserved;
73 } __attribute__((packed));
74 
75 #define MTHCA_CQ_STATUS_OK          ( 0 << 28)
76 #define MTHCA_CQ_STATUS_OVERFLOW    ( 9 << 28)
77 #define MTHCA_CQ_STATUS_WRITE_FAIL  (10 << 28)
78 #define MTHCA_CQ_FLAG_TR            ( 1 << 18)
79 #define MTHCA_CQ_FLAG_OI            ( 1 << 17)
80 #define MTHCA_CQ_STATE_DISARMED     ( 0 <<  8)
81 #define MTHCA_CQ_STATE_ARMED        ( 1 <<  8)
82 #define MTHCA_CQ_STATE_ARMED_SOL    ( 4 <<  8)
83 #define MTHCA_EQ_STATE_FIRED        (10 <<  8)
84 
85 enum {
86 	MTHCA_ERROR_CQE_OPCODE_MASK = 0xfe
87 };
88 
89 enum {
90 	SYNDROME_LOCAL_LENGTH_ERR 	 = 0x01,
91 	SYNDROME_LOCAL_QP_OP_ERR  	 = 0x02,
92 	SYNDROME_LOCAL_EEC_OP_ERR 	 = 0x03,
93 	SYNDROME_LOCAL_PROT_ERR   	 = 0x04,
94 	SYNDROME_WR_FLUSH_ERR     	 = 0x05,
95 	SYNDROME_MW_BIND_ERR      	 = 0x06,
96 	SYNDROME_BAD_RESP_ERR     	 = 0x10,
97 	SYNDROME_LOCAL_ACCESS_ERR 	 = 0x11,
98 	SYNDROME_REMOTE_INVAL_REQ_ERR 	 = 0x12,
99 	SYNDROME_REMOTE_ACCESS_ERR 	 = 0x13,
100 	SYNDROME_REMOTE_OP_ERR     	 = 0x14,
101 	SYNDROME_RETRY_EXC_ERR 		 = 0x15,
102 	SYNDROME_RNR_RETRY_EXC_ERR 	 = 0x16,
103 	SYNDROME_LOCAL_RDD_VIOL_ERR 	 = 0x20,
104 	SYNDROME_REMOTE_INVAL_RD_REQ_ERR = 0x21,
105 	SYNDROME_REMOTE_ABORTED_ERR 	 = 0x22,
106 	SYNDROME_INVAL_EECN_ERR 	 = 0x23,
107 	SYNDROME_INVAL_EEC_STATE_ERR 	 = 0x24
108 };
109 
110 struct mthca_cqe {
111 	u32 my_qpn;
112 	u32 my_ee;
113 	u32 rqpn;
114 	u16 sl_g_mlpath;
115 	u16 rlid;
116 	u32 imm_etype_pkey_eec;
117 	u32 byte_cnt;
118 	u32 wqe;
119 	u8  opcode;
120 	u8  is_send;
121 	u8  reserved;
122 	u8  owner;
123 };
124 
125 struct mthca_err_cqe {
126 	u32 my_qpn;
127 	u32 reserved1[3];
128 	u8  syndrome;
129 	u8  reserved2;
130 	u16 db_cnt;
131 	u32 reserved3;
132 	u32 wqe;
133 	u8  opcode;
134 	u8  reserved4[2];
135 	u8  owner;
136 };
137 
138 #define MTHCA_CQ_ENTRY_OWNER_SW      (0 << 7)
139 #define MTHCA_CQ_ENTRY_OWNER_HW      (1 << 7)
140 
141 #define MTHCA_TAVOR_CQ_DB_INC_CI       (1 << 24)
142 #define MTHCA_TAVOR_CQ_DB_REQ_NOT      (2 << 24)
143 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL  (3 << 24)
144 #define MTHCA_TAVOR_CQ_DB_SET_CI       (4 << 24)
145 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_MULT (5 << 24)
146 
147 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL  (1 << 24)
148 #define MTHCA_ARBEL_CQ_DB_REQ_NOT      (2 << 24)
149 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_MULT (3 << 24)
150 
151 static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
152 {
153 	if (cq->is_direct)
154 		return cq->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
155 	else
156 		return cq->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
157 			+ (entry * MTHCA_CQ_ENTRY_SIZE) % PAGE_SIZE;
158 }
159 
160 static inline struct mthca_cqe *cqe_sw(struct mthca_cq *cq, int i)
161 {
162 	struct mthca_cqe *cqe = get_cqe(cq, i);
163 	return MTHCA_CQ_ENTRY_OWNER_HW & cqe->owner ? NULL : cqe;
164 }
165 
166 static inline struct mthca_cqe *next_cqe_sw(struct mthca_cq *cq)
167 {
168 	return cqe_sw(cq, cq->cons_index & cq->ibcq.cqe);
169 }
170 
171 static inline void set_cqe_hw(struct mthca_cqe *cqe)
172 {
173 	cqe->owner = MTHCA_CQ_ENTRY_OWNER_HW;
174 }
175 
176 static void dump_cqe(struct mthca_dev *dev, void *cqe_ptr)
177 {
178 	__be32 *cqe = cqe_ptr;
179 
180 	(void) cqe;	/* avoid warning if mthca_dbg compiled away... */
181 	mthca_dbg(dev, "CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
182 		  be32_to_cpu(cqe[0]), be32_to_cpu(cqe[1]), be32_to_cpu(cqe[2]),
183 		  be32_to_cpu(cqe[3]), be32_to_cpu(cqe[4]), be32_to_cpu(cqe[5]),
184 		  be32_to_cpu(cqe[6]), be32_to_cpu(cqe[7]));
185 }
186 
187 /*
188  * incr is ignored in native Arbel (mem-free) mode, so cq->cons_index
189  * should be correct before calling update_cons_index().
190  */
191 static inline void update_cons_index(struct mthca_dev *dev, struct mthca_cq *cq,
192 				     int incr)
193 {
194 	u32 doorbell[2];
195 
196 	if (mthca_is_memfree(dev)) {
197 		*cq->set_ci_db = cpu_to_be32(cq->cons_index);
198 		wmb();
199 	} else {
200 		doorbell[0] = cpu_to_be32(MTHCA_TAVOR_CQ_DB_INC_CI | cq->cqn);
201 		doorbell[1] = cpu_to_be32(incr - 1);
202 
203 		mthca_write64(doorbell,
204 			      dev->kar + MTHCA_CQ_DOORBELL,
205 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
206 	}
207 }
208 
209 void mthca_cq_event(struct mthca_dev *dev, u32 cqn)
210 {
211 	struct mthca_cq *cq;
212 
213 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
214 
215 	if (!cq) {
216 		mthca_warn(dev, "Completion event for bogus CQ %08x\n", cqn);
217 		return;
218 	}
219 
220 	++cq->arm_sn;
221 
222 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
223 }
224 
225 void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn)
226 {
227 	struct mthca_cq *cq;
228 	struct mthca_cqe *cqe;
229 	int prod_index;
230 	int nfreed = 0;
231 
232 	spin_lock_irq(&dev->cq_table.lock);
233 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
234 	if (cq)
235 		atomic_inc(&cq->refcount);
236 	spin_unlock_irq(&dev->cq_table.lock);
237 
238 	if (!cq)
239 		return;
240 
241 	spin_lock_irq(&cq->lock);
242 
243 	/*
244 	 * First we need to find the current producer index, so we
245 	 * know where to start cleaning from.  It doesn't matter if HW
246 	 * adds new entries after this loop -- the QP we're worried
247 	 * about is already in RESET, so the new entries won't come
248 	 * from our QP and therefore don't need to be checked.
249 	 */
250 	for (prod_index = cq->cons_index;
251 	     cqe_sw(cq, prod_index & cq->ibcq.cqe);
252 	     ++prod_index)
253 		if (prod_index == cq->cons_index + cq->ibcq.cqe)
254 			break;
255 
256 	if (0)
257 		mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
258 			  qpn, cqn, cq->cons_index, prod_index);
259 
260 	/*
261 	 * Now sweep backwards through the CQ, removing CQ entries
262 	 * that match our QP by copying older entries on top of them.
263 	 */
264 	while (prod_index > cq->cons_index) {
265 		cqe = get_cqe(cq, (prod_index - 1) & cq->ibcq.cqe);
266 		if (cqe->my_qpn == cpu_to_be32(qpn))
267 			++nfreed;
268 		else if (nfreed)
269 			memcpy(get_cqe(cq, (prod_index - 1 + nfreed) &
270 				       cq->ibcq.cqe),
271 			       cqe,
272 			       MTHCA_CQ_ENTRY_SIZE);
273 		--prod_index;
274 	}
275 
276 	if (nfreed) {
277 		wmb();
278 		cq->cons_index += nfreed;
279 		update_cons_index(dev, cq, nfreed);
280 	}
281 
282 	spin_unlock_irq(&cq->lock);
283 	if (atomic_dec_and_test(&cq->refcount))
284 		wake_up(&cq->wait);
285 }
286 
287 static int handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
288 			    struct mthca_qp *qp, int wqe_index, int is_send,
289 			    struct mthca_err_cqe *cqe,
290 			    struct ib_wc *entry, int *free_cqe)
291 {
292 	int err;
293 	int dbd;
294 	u32 new_wqe;
295 
296 	if (cqe->syndrome == SYNDROME_LOCAL_QP_OP_ERR) {
297 		mthca_dbg(dev, "local QP operation err "
298 			  "(QPN %06x, WQE @ %08x, CQN %06x, index %d)\n",
299 			  be32_to_cpu(cqe->my_qpn), be32_to_cpu(cqe->wqe),
300 			  cq->cqn, cq->cons_index);
301 		dump_cqe(dev, cqe);
302 	}
303 
304 	/*
305 	 * For completions in error, only work request ID, status (and
306 	 * freed resource count for RD) have to be set.
307 	 */
308 	switch (cqe->syndrome) {
309 	case SYNDROME_LOCAL_LENGTH_ERR:
310 		entry->status = IB_WC_LOC_LEN_ERR;
311 		break;
312 	case SYNDROME_LOCAL_QP_OP_ERR:
313 		entry->status = IB_WC_LOC_QP_OP_ERR;
314 		break;
315 	case SYNDROME_LOCAL_EEC_OP_ERR:
316 		entry->status = IB_WC_LOC_EEC_OP_ERR;
317 		break;
318 	case SYNDROME_LOCAL_PROT_ERR:
319 		entry->status = IB_WC_LOC_PROT_ERR;
320 		break;
321 	case SYNDROME_WR_FLUSH_ERR:
322 		entry->status = IB_WC_WR_FLUSH_ERR;
323 		break;
324 	case SYNDROME_MW_BIND_ERR:
325 		entry->status = IB_WC_MW_BIND_ERR;
326 		break;
327 	case SYNDROME_BAD_RESP_ERR:
328 		entry->status = IB_WC_BAD_RESP_ERR;
329 		break;
330 	case SYNDROME_LOCAL_ACCESS_ERR:
331 		entry->status = IB_WC_LOC_ACCESS_ERR;
332 		break;
333 	case SYNDROME_REMOTE_INVAL_REQ_ERR:
334 		entry->status = IB_WC_REM_INV_REQ_ERR;
335 		break;
336 	case SYNDROME_REMOTE_ACCESS_ERR:
337 		entry->status = IB_WC_REM_ACCESS_ERR;
338 		break;
339 	case SYNDROME_REMOTE_OP_ERR:
340 		entry->status = IB_WC_REM_OP_ERR;
341 		break;
342 	case SYNDROME_RETRY_EXC_ERR:
343 		entry->status = IB_WC_RETRY_EXC_ERR;
344 		break;
345 	case SYNDROME_RNR_RETRY_EXC_ERR:
346 		entry->status = IB_WC_RNR_RETRY_EXC_ERR;
347 		break;
348 	case SYNDROME_LOCAL_RDD_VIOL_ERR:
349 		entry->status = IB_WC_LOC_RDD_VIOL_ERR;
350 		break;
351 	case SYNDROME_REMOTE_INVAL_RD_REQ_ERR:
352 		entry->status = IB_WC_REM_INV_RD_REQ_ERR;
353 		break;
354 	case SYNDROME_REMOTE_ABORTED_ERR:
355 		entry->status = IB_WC_REM_ABORT_ERR;
356 		break;
357 	case SYNDROME_INVAL_EECN_ERR:
358 		entry->status = IB_WC_INV_EECN_ERR;
359 		break;
360 	case SYNDROME_INVAL_EEC_STATE_ERR:
361 		entry->status = IB_WC_INV_EEC_STATE_ERR;
362 		break;
363 	default:
364 		entry->status = IB_WC_GENERAL_ERR;
365 		break;
366 	}
367 
368 	err = mthca_free_err_wqe(dev, qp, is_send, wqe_index, &dbd, &new_wqe);
369 	if (err)
370 		return err;
371 
372 	/*
373 	 * If we're at the end of the WQE chain, or we've used up our
374 	 * doorbell count, free the CQE.  Otherwise just update it for
375 	 * the next poll operation.
376 	 */
377 	if (!(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
378 		return 0;
379 
380 	cqe->db_cnt   = cpu_to_be16(be16_to_cpu(cqe->db_cnt) - dbd);
381 	cqe->wqe      = new_wqe;
382 	cqe->syndrome = SYNDROME_WR_FLUSH_ERR;
383 
384 	*free_cqe = 0;
385 
386 	return 0;
387 }
388 
389 static inline int mthca_poll_one(struct mthca_dev *dev,
390 				 struct mthca_cq *cq,
391 				 struct mthca_qp **cur_qp,
392 				 int *freed,
393 				 struct ib_wc *entry)
394 {
395 	struct mthca_wq *wq;
396 	struct mthca_cqe *cqe;
397 	int wqe_index;
398 	int is_error;
399 	int is_send;
400 	int free_cqe = 1;
401 	int err = 0;
402 
403 	cqe = next_cqe_sw(cq);
404 	if (!cqe)
405 		return -EAGAIN;
406 
407 	/*
408 	 * Make sure we read CQ entry contents after we've checked the
409 	 * ownership bit.
410 	 */
411 	rmb();
412 
413 	if (0) {
414 		mthca_dbg(dev, "%x/%d: CQE -> QPN %06x, WQE @ %08x\n",
415 			  cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
416 			  be32_to_cpu(cqe->wqe));
417 		dump_cqe(dev, cqe);
418 	}
419 
420 	is_error = (cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
421 		MTHCA_ERROR_CQE_OPCODE_MASK;
422 	is_send  = is_error ? cqe->opcode & 0x01 : cqe->is_send & 0x80;
423 
424 	if (!*cur_qp || be32_to_cpu(cqe->my_qpn) != (*cur_qp)->qpn) {
425 		/*
426 		 * We do not have to take the QP table lock here,
427 		 * because CQs will be locked while QPs are removed
428 		 * from the table.
429 		 */
430 		*cur_qp = mthca_array_get(&dev->qp_table.qp,
431 					  be32_to_cpu(cqe->my_qpn) &
432 					  (dev->limits.num_qps - 1));
433 		if (!*cur_qp) {
434 			mthca_warn(dev, "CQ entry for unknown QP %06x\n",
435 				   be32_to_cpu(cqe->my_qpn) & 0xffffff);
436 			err = -EINVAL;
437 			goto out;
438 		}
439 	}
440 
441 	entry->qp_num = (*cur_qp)->qpn;
442 
443 	if (is_send) {
444 		wq = &(*cur_qp)->sq;
445 		wqe_index = ((be32_to_cpu(cqe->wqe) - (*cur_qp)->send_wqe_offset)
446 			     >> wq->wqe_shift);
447 		entry->wr_id = (*cur_qp)->wrid[wqe_index +
448 					       (*cur_qp)->rq.max];
449 	} else {
450 		wq = &(*cur_qp)->rq;
451 		wqe_index = be32_to_cpu(cqe->wqe) >> wq->wqe_shift;
452 		entry->wr_id = (*cur_qp)->wrid[wqe_index];
453 	}
454 
455 	if (wq->last_comp < wqe_index)
456 		wq->tail += wqe_index - wq->last_comp;
457 	else
458 		wq->tail += wqe_index + wq->max - wq->last_comp;
459 
460 	wq->last_comp = wqe_index;
461 
462 	if (0)
463 		mthca_dbg(dev, "%s completion for QP %06x, index %d (nr %d)\n",
464 			  is_send ? "Send" : "Receive",
465 			  (*cur_qp)->qpn, wqe_index, wq->max);
466 
467 	if (is_error) {
468 		err = handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,
469 				       (struct mthca_err_cqe *) cqe,
470 				       entry, &free_cqe);
471 		goto out;
472 	}
473 
474 	if (is_send) {
475 		entry->wc_flags = 0;
476 		switch (cqe->opcode) {
477 		case MTHCA_OPCODE_RDMA_WRITE:
478 			entry->opcode    = IB_WC_RDMA_WRITE;
479 			break;
480 		case MTHCA_OPCODE_RDMA_WRITE_IMM:
481 			entry->opcode    = IB_WC_RDMA_WRITE;
482 			entry->wc_flags |= IB_WC_WITH_IMM;
483 			break;
484 		case MTHCA_OPCODE_SEND:
485 			entry->opcode    = IB_WC_SEND;
486 			break;
487 		case MTHCA_OPCODE_SEND_IMM:
488 			entry->opcode    = IB_WC_SEND;
489 			entry->wc_flags |= IB_WC_WITH_IMM;
490 			break;
491 		case MTHCA_OPCODE_RDMA_READ:
492 			entry->opcode    = IB_WC_RDMA_READ;
493 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
494 			break;
495 		case MTHCA_OPCODE_ATOMIC_CS:
496 			entry->opcode    = IB_WC_COMP_SWAP;
497 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
498 			break;
499 		case MTHCA_OPCODE_ATOMIC_FA:
500 			entry->opcode    = IB_WC_FETCH_ADD;
501 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
502 			break;
503 		case MTHCA_OPCODE_BIND_MW:
504 			entry->opcode    = IB_WC_BIND_MW;
505 			break;
506 		default:
507 			entry->opcode    = MTHCA_OPCODE_INVALID;
508 			break;
509 		}
510 	} else {
511 		entry->byte_len = be32_to_cpu(cqe->byte_cnt);
512 		switch (cqe->opcode & 0x1f) {
513 		case IB_OPCODE_SEND_LAST_WITH_IMMEDIATE:
514 		case IB_OPCODE_SEND_ONLY_WITH_IMMEDIATE:
515 			entry->wc_flags = IB_WC_WITH_IMM;
516 			entry->imm_data = cqe->imm_etype_pkey_eec;
517 			entry->opcode = IB_WC_RECV;
518 			break;
519 		case IB_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
520 		case IB_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
521 			entry->wc_flags = IB_WC_WITH_IMM;
522 			entry->imm_data = cqe->imm_etype_pkey_eec;
523 			entry->opcode = IB_WC_RECV_RDMA_WITH_IMM;
524 			break;
525 		default:
526 			entry->wc_flags = 0;
527 			entry->opcode = IB_WC_RECV;
528 			break;
529 		}
530 		entry->slid 	   = be16_to_cpu(cqe->rlid);
531 		entry->sl   	   = be16_to_cpu(cqe->sl_g_mlpath) >> 12;
532 		entry->src_qp 	   = be32_to_cpu(cqe->rqpn) & 0xffffff;
533 		entry->dlid_path_bits = be16_to_cpu(cqe->sl_g_mlpath) & 0x7f;
534 		entry->pkey_index  = be32_to_cpu(cqe->imm_etype_pkey_eec) >> 16;
535 		entry->wc_flags   |= be16_to_cpu(cqe->sl_g_mlpath) & 0x80 ?
536 					IB_WC_GRH : 0;
537 	}
538 
539 	entry->status = IB_WC_SUCCESS;
540 
541  out:
542 	if (likely(free_cqe)) {
543 		set_cqe_hw(cqe);
544 		++(*freed);
545 		++cq->cons_index;
546 	}
547 
548 	return err;
549 }
550 
551 int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
552 		  struct ib_wc *entry)
553 {
554 	struct mthca_dev *dev = to_mdev(ibcq->device);
555 	struct mthca_cq *cq = to_mcq(ibcq);
556 	struct mthca_qp *qp = NULL;
557 	unsigned long flags;
558 	int err = 0;
559 	int freed = 0;
560 	int npolled;
561 
562 	spin_lock_irqsave(&cq->lock, flags);
563 
564 	for (npolled = 0; npolled < num_entries; ++npolled) {
565 		err = mthca_poll_one(dev, cq, &qp,
566 				     &freed, entry + npolled);
567 		if (err)
568 			break;
569 	}
570 
571 	if (freed) {
572 		wmb();
573 		update_cons_index(dev, cq, freed);
574 	}
575 
576 	spin_unlock_irqrestore(&cq->lock, flags);
577 
578 	return err == 0 || err == -EAGAIN ? npolled : err;
579 }
580 
581 int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify notify)
582 {
583 	u32 doorbell[2];
584 
585 	doorbell[0] = cpu_to_be32((notify == IB_CQ_SOLICITED ?
586 				   MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL :
587 				   MTHCA_TAVOR_CQ_DB_REQ_NOT)      |
588 				  to_mcq(cq)->cqn);
589 	doorbell[1] = 0xffffffff;
590 
591 	mthca_write64(doorbell,
592 		      to_mdev(cq->device)->kar + MTHCA_CQ_DOORBELL,
593 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(cq->device)->doorbell_lock));
594 
595 	return 0;
596 }
597 
598 int mthca_arbel_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify notify)
599 {
600 	struct mthca_cq *cq = to_mcq(ibcq);
601 	u32 doorbell[2];
602 	u32 sn;
603 	u32 ci;
604 
605 	sn = cq->arm_sn & 3;
606 	ci = cpu_to_be32(cq->cons_index);
607 
608 	doorbell[0] = ci;
609 	doorbell[1] = cpu_to_be32((cq->cqn << 8) | (2 << 5) | (sn << 3) |
610 				  (notify == IB_CQ_SOLICITED ? 1 : 2));
611 
612 	mthca_write_db_rec(doorbell, cq->arm_db);
613 
614 	/*
615 	 * Make sure that the doorbell record in host memory is
616 	 * written before ringing the doorbell via PCI MMIO.
617 	 */
618 	wmb();
619 
620 	doorbell[0] = cpu_to_be32((sn << 28)                       |
621 				  (notify == IB_CQ_SOLICITED ?
622 				   MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL :
623 				   MTHCA_ARBEL_CQ_DB_REQ_NOT)      |
624 				  cq->cqn);
625 	doorbell[1] = ci;
626 
627 	mthca_write64(doorbell,
628 		      to_mdev(ibcq->device)->kar + MTHCA_CQ_DOORBELL,
629 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->doorbell_lock));
630 
631 	return 0;
632 }
633 
634 static void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq *cq)
635 {
636 	int i;
637 	int size;
638 
639 	if (cq->is_direct)
640 		dma_free_coherent(&dev->pdev->dev,
641 				  (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE,
642 				  cq->queue.direct.buf,
643 				  pci_unmap_addr(&cq->queue.direct,
644 						 mapping));
645 	else {
646 		size = (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE;
647 		for (i = 0; i < (size + PAGE_SIZE - 1) / PAGE_SIZE; ++i)
648 			if (cq->queue.page_list[i].buf)
649 				dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
650 						  cq->queue.page_list[i].buf,
651 						  pci_unmap_addr(&cq->queue.page_list[i],
652 								 mapping));
653 
654 		kfree(cq->queue.page_list);
655 	}
656 }
657 
658 static int mthca_alloc_cq_buf(struct mthca_dev *dev, int size,
659 			      struct mthca_cq *cq)
660 {
661 	int err = -ENOMEM;
662 	int npages, shift;
663 	u64 *dma_list = NULL;
664 	dma_addr_t t;
665 	int i;
666 
667 	if (size <= MTHCA_MAX_DIRECT_CQ_SIZE) {
668 		cq->is_direct = 1;
669 		npages        = 1;
670 		shift         = get_order(size) + PAGE_SHIFT;
671 
672 		cq->queue.direct.buf = dma_alloc_coherent(&dev->pdev->dev,
673 							  size, &t, GFP_KERNEL);
674 		if (!cq->queue.direct.buf)
675 			return -ENOMEM;
676 
677 		pci_unmap_addr_set(&cq->queue.direct, mapping, t);
678 
679 		memset(cq->queue.direct.buf, 0, size);
680 
681 		while (t & ((1 << shift) - 1)) {
682 			--shift;
683 			npages *= 2;
684 		}
685 
686 		dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
687 		if (!dma_list)
688 			goto err_free;
689 
690 		for (i = 0; i < npages; ++i)
691 			dma_list[i] = t + i * (1 << shift);
692 	} else {
693 		cq->is_direct = 0;
694 		npages        = (size + PAGE_SIZE - 1) / PAGE_SIZE;
695 		shift         = PAGE_SHIFT;
696 
697 		dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
698 		if (!dma_list)
699 			return -ENOMEM;
700 
701 		cq->queue.page_list = kmalloc(npages * sizeof *cq->queue.page_list,
702 					      GFP_KERNEL);
703 		if (!cq->queue.page_list)
704 			goto err_out;
705 
706 		for (i = 0; i < npages; ++i)
707 			cq->queue.page_list[i].buf = NULL;
708 
709 		for (i = 0; i < npages; ++i) {
710 			cq->queue.page_list[i].buf =
711 				dma_alloc_coherent(&dev->pdev->dev, PAGE_SIZE,
712 						   &t, GFP_KERNEL);
713 			if (!cq->queue.page_list[i].buf)
714 				goto err_free;
715 
716 			dma_list[i] = t;
717 			pci_unmap_addr_set(&cq->queue.page_list[i], mapping, t);
718 
719 			memset(cq->queue.page_list[i].buf, 0, PAGE_SIZE);
720 		}
721 	}
722 
723 	err = mthca_mr_alloc_phys(dev, dev->driver_pd.pd_num,
724 				  dma_list, shift, npages,
725 				  0, size,
726 				  MTHCA_MPT_FLAG_LOCAL_WRITE |
727 				  MTHCA_MPT_FLAG_LOCAL_READ,
728 				  &cq->mr);
729 	if (err)
730 		goto err_free;
731 
732 	kfree(dma_list);
733 
734 	return 0;
735 
736 err_free:
737 	mthca_free_cq_buf(dev, cq);
738 
739 err_out:
740 	kfree(dma_list);
741 
742 	return err;
743 }
744 
745 int mthca_init_cq(struct mthca_dev *dev, int nent,
746 		  struct mthca_ucontext *ctx, u32 pdn,
747 		  struct mthca_cq *cq)
748 {
749 	int size = nent * MTHCA_CQ_ENTRY_SIZE;
750 	struct mthca_mailbox *mailbox;
751 	struct mthca_cq_context *cq_context;
752 	int err = -ENOMEM;
753 	u8 status;
754 	int i;
755 
756 	might_sleep();
757 
758 	cq->ibcq.cqe  = nent - 1;
759 	cq->is_kernel = !ctx;
760 
761 	cq->cqn = mthca_alloc(&dev->cq_table.alloc);
762 	if (cq->cqn == -1)
763 		return -ENOMEM;
764 
765 	if (mthca_is_memfree(dev)) {
766 		err = mthca_table_get(dev, dev->cq_table.table, cq->cqn);
767 		if (err)
768 			goto err_out;
769 
770 		if (cq->is_kernel) {
771 			cq->arm_sn = 1;
772 
773 			err = -ENOMEM;
774 
775 			cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
776 							     cq->cqn, &cq->set_ci_db);
777 			if (cq->set_ci_db_index < 0)
778 				goto err_out_icm;
779 
780 			cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
781 							  cq->cqn, &cq->arm_db);
782 			if (cq->arm_db_index < 0)
783 				goto err_out_ci;
784 		}
785 	}
786 
787 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
788 	if (IS_ERR(mailbox))
789 		goto err_out_arm;
790 
791 	cq_context = mailbox->buf;
792 
793 	if (cq->is_kernel) {
794 		err = mthca_alloc_cq_buf(dev, size, cq);
795 		if (err)
796 			goto err_out_mailbox;
797 
798 		for (i = 0; i < nent; ++i)
799 			set_cqe_hw(get_cqe(cq, i));
800 	}
801 
802 	spin_lock_init(&cq->lock);
803 	atomic_set(&cq->refcount, 1);
804 	init_waitqueue_head(&cq->wait);
805 
806 	memset(cq_context, 0, sizeof *cq_context);
807 	cq_context->flags           = cpu_to_be32(MTHCA_CQ_STATUS_OK      |
808 						  MTHCA_CQ_STATE_DISARMED |
809 						  MTHCA_CQ_FLAG_TR);
810 	cq_context->start           = cpu_to_be64(0);
811 	cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24);
812 	if (ctx)
813 		cq_context->logsize_usrpage |= cpu_to_be32(ctx->uar.index);
814 	else
815 		cq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
816 	cq_context->error_eqn       = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
817 	cq_context->comp_eqn        = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
818 	cq_context->pd              = cpu_to_be32(pdn);
819 	cq_context->lkey            = cpu_to_be32(cq->mr.ibmr.lkey);
820 	cq_context->cqn             = cpu_to_be32(cq->cqn);
821 
822 	if (mthca_is_memfree(dev)) {
823 		cq_context->ci_db    = cpu_to_be32(cq->set_ci_db_index);
824 		cq_context->state_db = cpu_to_be32(cq->arm_db_index);
825 	}
826 
827 	err = mthca_SW2HW_CQ(dev, mailbox, cq->cqn, &status);
828 	if (err) {
829 		mthca_warn(dev, "SW2HW_CQ failed (%d)\n", err);
830 		goto err_out_free_mr;
831 	}
832 
833 	if (status) {
834 		mthca_warn(dev, "SW2HW_CQ returned status 0x%02x\n",
835 			   status);
836 		err = -EINVAL;
837 		goto err_out_free_mr;
838 	}
839 
840 	spin_lock_irq(&dev->cq_table.lock);
841 	if (mthca_array_set(&dev->cq_table.cq,
842 			    cq->cqn & (dev->limits.num_cqs - 1),
843 			    cq)) {
844 		spin_unlock_irq(&dev->cq_table.lock);
845 		goto err_out_free_mr;
846 	}
847 	spin_unlock_irq(&dev->cq_table.lock);
848 
849 	cq->cons_index = 0;
850 
851 	mthca_free_mailbox(dev, mailbox);
852 
853 	return 0;
854 
855 err_out_free_mr:
856 	if (cq->is_kernel) {
857 		mthca_free_mr(dev, &cq->mr);
858 		mthca_free_cq_buf(dev, cq);
859 	}
860 
861 err_out_mailbox:
862 	mthca_free_mailbox(dev, mailbox);
863 
864 err_out_arm:
865 	if (cq->is_kernel && mthca_is_memfree(dev))
866 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
867 
868 err_out_ci:
869 	if (cq->is_kernel && mthca_is_memfree(dev))
870 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
871 
872 err_out_icm:
873 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
874 
875 err_out:
876 	mthca_free(&dev->cq_table.alloc, cq->cqn);
877 
878 	return err;
879 }
880 
881 void mthca_free_cq(struct mthca_dev *dev,
882 		   struct mthca_cq *cq)
883 {
884 	struct mthca_mailbox *mailbox;
885 	int err;
886 	u8 status;
887 
888 	might_sleep();
889 
890 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
891 	if (IS_ERR(mailbox)) {
892 		mthca_warn(dev, "No memory for mailbox to free CQ.\n");
893 		return;
894 	}
895 
896 	err = mthca_HW2SW_CQ(dev, mailbox, cq->cqn, &status);
897 	if (err)
898 		mthca_warn(dev, "HW2SW_CQ failed (%d)\n", err);
899 	else if (status)
900 		mthca_warn(dev, "HW2SW_CQ returned status 0x%02x\n", status);
901 
902 	if (0) {
903 		u32 *ctx = mailbox->buf;
904 		int j;
905 
906 		printk(KERN_ERR "context for CQN %x (cons index %x, next sw %d)\n",
907 		       cq->cqn, cq->cons_index,
908 		       cq->is_kernel ? !!next_cqe_sw(cq) : 0);
909 		for (j = 0; j < 16; ++j)
910 			printk(KERN_ERR "[%2x] %08x\n", j * 4, be32_to_cpu(ctx[j]));
911 	}
912 
913 	spin_lock_irq(&dev->cq_table.lock);
914 	mthca_array_clear(&dev->cq_table.cq,
915 			  cq->cqn & (dev->limits.num_cqs - 1));
916 	spin_unlock_irq(&dev->cq_table.lock);
917 
918 	if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
919 		synchronize_irq(dev->eq_table.eq[MTHCA_EQ_COMP].msi_x_vector);
920 	else
921 		synchronize_irq(dev->pdev->irq);
922 
923 	atomic_dec(&cq->refcount);
924 	wait_event(cq->wait, !atomic_read(&cq->refcount));
925 
926 	if (cq->is_kernel) {
927 		mthca_free_mr(dev, &cq->mr);
928 		mthca_free_cq_buf(dev, cq);
929 		if (mthca_is_memfree(dev)) {
930 			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM,    cq->arm_db_index);
931 			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
932 		}
933 	}
934 
935 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
936 	mthca_free(&dev->cq_table.alloc, cq->cqn);
937 	mthca_free_mailbox(dev, mailbox);
938 }
939 
940 int __devinit mthca_init_cq_table(struct mthca_dev *dev)
941 {
942 	int err;
943 
944 	spin_lock_init(&dev->cq_table.lock);
945 
946 	err = mthca_alloc_init(&dev->cq_table.alloc,
947 			       dev->limits.num_cqs,
948 			       (1 << 24) - 1,
949 			       dev->limits.reserved_cqs);
950 	if (err)
951 		return err;
952 
953 	err = mthca_array_init(&dev->cq_table.cq,
954 			       dev->limits.num_cqs);
955 	if (err)
956 		mthca_alloc_cleanup(&dev->cq_table.alloc);
957 
958 	return err;
959 }
960 
961 void __devexit mthca_cleanup_cq_table(struct mthca_dev *dev)
962 {
963 	mthca_array_cleanup(&dev->cq_table.cq, dev->limits.num_cqs);
964 	mthca_alloc_cleanup(&dev->cq_table.alloc);
965 }
966