xref: /linux/drivers/infiniband/hw/mthca/mthca_cq.c (revision d67b569f5f620c0fb95d5212642746b7ba9d29e4)
1 /*
2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  *
33  * $Id: mthca_cq.c 1369 2004-12-20 16:17:07Z roland $
34  */
35 
36 #include <linux/init.h>
37 #include <linux/hardirq.h>
38 
39 #include <ib_pack.h>
40 
41 #include "mthca_dev.h"
42 #include "mthca_cmd.h"
43 #include "mthca_memfree.h"
44 
45 enum {
46 	MTHCA_MAX_DIRECT_CQ_SIZE = 4 * PAGE_SIZE
47 };
48 
49 enum {
50 	MTHCA_CQ_ENTRY_SIZE = 0x20
51 };
52 
53 /*
54  * Must be packed because start is 64 bits but only aligned to 32 bits.
55  */
56 struct mthca_cq_context {
57 	u32 flags;
58 	u64 start;
59 	u32 logsize_usrpage;
60 	u32 error_eqn;		/* Tavor only */
61 	u32 comp_eqn;
62 	u32 pd;
63 	u32 lkey;
64 	u32 last_notified_index;
65 	u32 solicit_producer_index;
66 	u32 consumer_index;
67 	u32 producer_index;
68 	u32 cqn;
69 	u32 ci_db;		/* Arbel only */
70 	u32 state_db;		/* Arbel only */
71 	u32 reserved;
72 } __attribute__((packed));
73 
74 #define MTHCA_CQ_STATUS_OK          ( 0 << 28)
75 #define MTHCA_CQ_STATUS_OVERFLOW    ( 9 << 28)
76 #define MTHCA_CQ_STATUS_WRITE_FAIL  (10 << 28)
77 #define MTHCA_CQ_FLAG_TR            ( 1 << 18)
78 #define MTHCA_CQ_FLAG_OI            ( 1 << 17)
79 #define MTHCA_CQ_STATE_DISARMED     ( 0 <<  8)
80 #define MTHCA_CQ_STATE_ARMED        ( 1 <<  8)
81 #define MTHCA_CQ_STATE_ARMED_SOL    ( 4 <<  8)
82 #define MTHCA_EQ_STATE_FIRED        (10 <<  8)
83 
84 enum {
85 	MTHCA_ERROR_CQE_OPCODE_MASK = 0xfe
86 };
87 
88 enum {
89 	SYNDROME_LOCAL_LENGTH_ERR 	 = 0x01,
90 	SYNDROME_LOCAL_QP_OP_ERR  	 = 0x02,
91 	SYNDROME_LOCAL_EEC_OP_ERR 	 = 0x03,
92 	SYNDROME_LOCAL_PROT_ERR   	 = 0x04,
93 	SYNDROME_WR_FLUSH_ERR     	 = 0x05,
94 	SYNDROME_MW_BIND_ERR      	 = 0x06,
95 	SYNDROME_BAD_RESP_ERR     	 = 0x10,
96 	SYNDROME_LOCAL_ACCESS_ERR 	 = 0x11,
97 	SYNDROME_REMOTE_INVAL_REQ_ERR 	 = 0x12,
98 	SYNDROME_REMOTE_ACCESS_ERR 	 = 0x13,
99 	SYNDROME_REMOTE_OP_ERR     	 = 0x14,
100 	SYNDROME_RETRY_EXC_ERR 		 = 0x15,
101 	SYNDROME_RNR_RETRY_EXC_ERR 	 = 0x16,
102 	SYNDROME_LOCAL_RDD_VIOL_ERR 	 = 0x20,
103 	SYNDROME_REMOTE_INVAL_RD_REQ_ERR = 0x21,
104 	SYNDROME_REMOTE_ABORTED_ERR 	 = 0x22,
105 	SYNDROME_INVAL_EECN_ERR 	 = 0x23,
106 	SYNDROME_INVAL_EEC_STATE_ERR 	 = 0x24
107 };
108 
109 struct mthca_cqe {
110 	u32 my_qpn;
111 	u32 my_ee;
112 	u32 rqpn;
113 	u16 sl_g_mlpath;
114 	u16 rlid;
115 	u32 imm_etype_pkey_eec;
116 	u32 byte_cnt;
117 	u32 wqe;
118 	u8  opcode;
119 	u8  is_send;
120 	u8  reserved;
121 	u8  owner;
122 };
123 
124 struct mthca_err_cqe {
125 	u32 my_qpn;
126 	u32 reserved1[3];
127 	u8  syndrome;
128 	u8  reserved2;
129 	u16 db_cnt;
130 	u32 reserved3;
131 	u32 wqe;
132 	u8  opcode;
133 	u8  reserved4[2];
134 	u8  owner;
135 };
136 
137 #define MTHCA_CQ_ENTRY_OWNER_SW      (0 << 7)
138 #define MTHCA_CQ_ENTRY_OWNER_HW      (1 << 7)
139 
140 #define MTHCA_TAVOR_CQ_DB_INC_CI       (1 << 24)
141 #define MTHCA_TAVOR_CQ_DB_REQ_NOT      (2 << 24)
142 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL  (3 << 24)
143 #define MTHCA_TAVOR_CQ_DB_SET_CI       (4 << 24)
144 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_MULT (5 << 24)
145 
146 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL  (1 << 24)
147 #define MTHCA_ARBEL_CQ_DB_REQ_NOT      (2 << 24)
148 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_MULT (3 << 24)
149 
150 static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
151 {
152 	if (cq->is_direct)
153 		return cq->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
154 	else
155 		return cq->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
156 			+ (entry * MTHCA_CQ_ENTRY_SIZE) % PAGE_SIZE;
157 }
158 
159 static inline struct mthca_cqe *cqe_sw(struct mthca_cq *cq, int i)
160 {
161 	struct mthca_cqe *cqe = get_cqe(cq, i);
162 	return MTHCA_CQ_ENTRY_OWNER_HW & cqe->owner ? NULL : cqe;
163 }
164 
165 static inline struct mthca_cqe *next_cqe_sw(struct mthca_cq *cq)
166 {
167 	return cqe_sw(cq, cq->cons_index & cq->ibcq.cqe);
168 }
169 
170 static inline void set_cqe_hw(struct mthca_cqe *cqe)
171 {
172 	cqe->owner = MTHCA_CQ_ENTRY_OWNER_HW;
173 }
174 
175 static void dump_cqe(struct mthca_dev *dev, void *cqe_ptr)
176 {
177 	__be32 *cqe = cqe_ptr;
178 
179 	(void) cqe;	/* avoid warning if mthca_dbg compiled away... */
180 	mthca_dbg(dev, "CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
181 		  be32_to_cpu(cqe[0]), be32_to_cpu(cqe[1]), be32_to_cpu(cqe[2]),
182 		  be32_to_cpu(cqe[3]), be32_to_cpu(cqe[4]), be32_to_cpu(cqe[5]),
183 		  be32_to_cpu(cqe[6]), be32_to_cpu(cqe[7]));
184 }
185 
186 /*
187  * incr is ignored in native Arbel (mem-free) mode, so cq->cons_index
188  * should be correct before calling update_cons_index().
189  */
190 static inline void update_cons_index(struct mthca_dev *dev, struct mthca_cq *cq,
191 				     int incr)
192 {
193 	u32 doorbell[2];
194 
195 	if (mthca_is_memfree(dev)) {
196 		*cq->set_ci_db = cpu_to_be32(cq->cons_index);
197 		wmb();
198 	} else {
199 		doorbell[0] = cpu_to_be32(MTHCA_TAVOR_CQ_DB_INC_CI | cq->cqn);
200 		doorbell[1] = cpu_to_be32(incr - 1);
201 
202 		mthca_write64(doorbell,
203 			      dev->kar + MTHCA_CQ_DOORBELL,
204 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
205 	}
206 }
207 
208 void mthca_cq_event(struct mthca_dev *dev, u32 cqn)
209 {
210 	struct mthca_cq *cq;
211 
212 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
213 
214 	if (!cq) {
215 		mthca_warn(dev, "Completion event for bogus CQ %08x\n", cqn);
216 		return;
217 	}
218 
219 	++cq->arm_sn;
220 
221 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
222 }
223 
224 void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn)
225 {
226 	struct mthca_cq *cq;
227 	struct mthca_cqe *cqe;
228 	int prod_index;
229 	int nfreed = 0;
230 
231 	spin_lock_irq(&dev->cq_table.lock);
232 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
233 	if (cq)
234 		atomic_inc(&cq->refcount);
235 	spin_unlock_irq(&dev->cq_table.lock);
236 
237 	if (!cq)
238 		return;
239 
240 	spin_lock_irq(&cq->lock);
241 
242 	/*
243 	 * First we need to find the current producer index, so we
244 	 * know where to start cleaning from.  It doesn't matter if HW
245 	 * adds new entries after this loop -- the QP we're worried
246 	 * about is already in RESET, so the new entries won't come
247 	 * from our QP and therefore don't need to be checked.
248 	 */
249 	for (prod_index = cq->cons_index;
250 	     cqe_sw(cq, prod_index & cq->ibcq.cqe);
251 	     ++prod_index)
252 		if (prod_index == cq->cons_index + cq->ibcq.cqe)
253 			break;
254 
255 	if (0)
256 		mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
257 			  qpn, cqn, cq->cons_index, prod_index);
258 
259 	/*
260 	 * Now sweep backwards through the CQ, removing CQ entries
261 	 * that match our QP by copying older entries on top of them.
262 	 */
263 	while (prod_index > cq->cons_index) {
264 		cqe = get_cqe(cq, (prod_index - 1) & cq->ibcq.cqe);
265 		if (cqe->my_qpn == cpu_to_be32(qpn))
266 			++nfreed;
267 		else if (nfreed)
268 			memcpy(get_cqe(cq, (prod_index - 1 + nfreed) &
269 				       cq->ibcq.cqe),
270 			       cqe,
271 			       MTHCA_CQ_ENTRY_SIZE);
272 		--prod_index;
273 	}
274 
275 	if (nfreed) {
276 		wmb();
277 		cq->cons_index += nfreed;
278 		update_cons_index(dev, cq, nfreed);
279 	}
280 
281 	spin_unlock_irq(&cq->lock);
282 	if (atomic_dec_and_test(&cq->refcount))
283 		wake_up(&cq->wait);
284 }
285 
286 static int handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
287 			    struct mthca_qp *qp, int wqe_index, int is_send,
288 			    struct mthca_err_cqe *cqe,
289 			    struct ib_wc *entry, int *free_cqe)
290 {
291 	int err;
292 	int dbd;
293 	u32 new_wqe;
294 
295 	if (cqe->syndrome == SYNDROME_LOCAL_QP_OP_ERR) {
296 		mthca_dbg(dev, "local QP operation err "
297 			  "(QPN %06x, WQE @ %08x, CQN %06x, index %d)\n",
298 			  be32_to_cpu(cqe->my_qpn), be32_to_cpu(cqe->wqe),
299 			  cq->cqn, cq->cons_index);
300 		dump_cqe(dev, cqe);
301 	}
302 
303 	/*
304 	 * For completions in error, only work request ID, status (and
305 	 * freed resource count for RD) have to be set.
306 	 */
307 	switch (cqe->syndrome) {
308 	case SYNDROME_LOCAL_LENGTH_ERR:
309 		entry->status = IB_WC_LOC_LEN_ERR;
310 		break;
311 	case SYNDROME_LOCAL_QP_OP_ERR:
312 		entry->status = IB_WC_LOC_QP_OP_ERR;
313 		break;
314 	case SYNDROME_LOCAL_EEC_OP_ERR:
315 		entry->status = IB_WC_LOC_EEC_OP_ERR;
316 		break;
317 	case SYNDROME_LOCAL_PROT_ERR:
318 		entry->status = IB_WC_LOC_PROT_ERR;
319 		break;
320 	case SYNDROME_WR_FLUSH_ERR:
321 		entry->status = IB_WC_WR_FLUSH_ERR;
322 		break;
323 	case SYNDROME_MW_BIND_ERR:
324 		entry->status = IB_WC_MW_BIND_ERR;
325 		break;
326 	case SYNDROME_BAD_RESP_ERR:
327 		entry->status = IB_WC_BAD_RESP_ERR;
328 		break;
329 	case SYNDROME_LOCAL_ACCESS_ERR:
330 		entry->status = IB_WC_LOC_ACCESS_ERR;
331 		break;
332 	case SYNDROME_REMOTE_INVAL_REQ_ERR:
333 		entry->status = IB_WC_REM_INV_REQ_ERR;
334 		break;
335 	case SYNDROME_REMOTE_ACCESS_ERR:
336 		entry->status = IB_WC_REM_ACCESS_ERR;
337 		break;
338 	case SYNDROME_REMOTE_OP_ERR:
339 		entry->status = IB_WC_REM_OP_ERR;
340 		break;
341 	case SYNDROME_RETRY_EXC_ERR:
342 		entry->status = IB_WC_RETRY_EXC_ERR;
343 		break;
344 	case SYNDROME_RNR_RETRY_EXC_ERR:
345 		entry->status = IB_WC_RNR_RETRY_EXC_ERR;
346 		break;
347 	case SYNDROME_LOCAL_RDD_VIOL_ERR:
348 		entry->status = IB_WC_LOC_RDD_VIOL_ERR;
349 		break;
350 	case SYNDROME_REMOTE_INVAL_RD_REQ_ERR:
351 		entry->status = IB_WC_REM_INV_RD_REQ_ERR;
352 		break;
353 	case SYNDROME_REMOTE_ABORTED_ERR:
354 		entry->status = IB_WC_REM_ABORT_ERR;
355 		break;
356 	case SYNDROME_INVAL_EECN_ERR:
357 		entry->status = IB_WC_INV_EECN_ERR;
358 		break;
359 	case SYNDROME_INVAL_EEC_STATE_ERR:
360 		entry->status = IB_WC_INV_EEC_STATE_ERR;
361 		break;
362 	default:
363 		entry->status = IB_WC_GENERAL_ERR;
364 		break;
365 	}
366 
367 	err = mthca_free_err_wqe(dev, qp, is_send, wqe_index, &dbd, &new_wqe);
368 	if (err)
369 		return err;
370 
371 	/*
372 	 * If we're at the end of the WQE chain, or we've used up our
373 	 * doorbell count, free the CQE.  Otherwise just update it for
374 	 * the next poll operation.
375 	 */
376 	if (!(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
377 		return 0;
378 
379 	cqe->db_cnt   = cpu_to_be16(be16_to_cpu(cqe->db_cnt) - dbd);
380 	cqe->wqe      = new_wqe;
381 	cqe->syndrome = SYNDROME_WR_FLUSH_ERR;
382 
383 	*free_cqe = 0;
384 
385 	return 0;
386 }
387 
388 static inline int mthca_poll_one(struct mthca_dev *dev,
389 				 struct mthca_cq *cq,
390 				 struct mthca_qp **cur_qp,
391 				 int *freed,
392 				 struct ib_wc *entry)
393 {
394 	struct mthca_wq *wq;
395 	struct mthca_cqe *cqe;
396 	int wqe_index;
397 	int is_error;
398 	int is_send;
399 	int free_cqe = 1;
400 	int err = 0;
401 
402 	cqe = next_cqe_sw(cq);
403 	if (!cqe)
404 		return -EAGAIN;
405 
406 	/*
407 	 * Make sure we read CQ entry contents after we've checked the
408 	 * ownership bit.
409 	 */
410 	rmb();
411 
412 	if (0) {
413 		mthca_dbg(dev, "%x/%d: CQE -> QPN %06x, WQE @ %08x\n",
414 			  cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
415 			  be32_to_cpu(cqe->wqe));
416 		dump_cqe(dev, cqe);
417 	}
418 
419 	is_error = (cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
420 		MTHCA_ERROR_CQE_OPCODE_MASK;
421 	is_send  = is_error ? cqe->opcode & 0x01 : cqe->is_send & 0x80;
422 
423 	if (!*cur_qp || be32_to_cpu(cqe->my_qpn) != (*cur_qp)->qpn) {
424 		/*
425 		 * We do not have to take the QP table lock here,
426 		 * because CQs will be locked while QPs are removed
427 		 * from the table.
428 		 */
429 		*cur_qp = mthca_array_get(&dev->qp_table.qp,
430 					  be32_to_cpu(cqe->my_qpn) &
431 					  (dev->limits.num_qps - 1));
432 		if (!*cur_qp) {
433 			mthca_warn(dev, "CQ entry for unknown QP %06x\n",
434 				   be32_to_cpu(cqe->my_qpn) & 0xffffff);
435 			err = -EINVAL;
436 			goto out;
437 		}
438 	}
439 
440 	entry->qp_num = (*cur_qp)->qpn;
441 
442 	if (is_send) {
443 		wq = &(*cur_qp)->sq;
444 		wqe_index = ((be32_to_cpu(cqe->wqe) - (*cur_qp)->send_wqe_offset)
445 			     >> wq->wqe_shift);
446 		entry->wr_id = (*cur_qp)->wrid[wqe_index +
447 					       (*cur_qp)->rq.max];
448 	} else {
449 		wq = &(*cur_qp)->rq;
450 		wqe_index = be32_to_cpu(cqe->wqe) >> wq->wqe_shift;
451 		entry->wr_id = (*cur_qp)->wrid[wqe_index];
452 	}
453 
454 	if (wq->last_comp < wqe_index)
455 		wq->tail += wqe_index - wq->last_comp;
456 	else
457 		wq->tail += wqe_index + wq->max - wq->last_comp;
458 
459 	wq->last_comp = wqe_index;
460 
461 	if (0)
462 		mthca_dbg(dev, "%s completion for QP %06x, index %d (nr %d)\n",
463 			  is_send ? "Send" : "Receive",
464 			  (*cur_qp)->qpn, wqe_index, wq->max);
465 
466 	if (is_error) {
467 		err = handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,
468 				       (struct mthca_err_cqe *) cqe,
469 				       entry, &free_cqe);
470 		goto out;
471 	}
472 
473 	if (is_send) {
474 		entry->wc_flags = 0;
475 		switch (cqe->opcode) {
476 		case MTHCA_OPCODE_RDMA_WRITE:
477 			entry->opcode    = IB_WC_RDMA_WRITE;
478 			break;
479 		case MTHCA_OPCODE_RDMA_WRITE_IMM:
480 			entry->opcode    = IB_WC_RDMA_WRITE;
481 			entry->wc_flags |= IB_WC_WITH_IMM;
482 			break;
483 		case MTHCA_OPCODE_SEND:
484 			entry->opcode    = IB_WC_SEND;
485 			break;
486 		case MTHCA_OPCODE_SEND_IMM:
487 			entry->opcode    = IB_WC_SEND;
488 			entry->wc_flags |= IB_WC_WITH_IMM;
489 			break;
490 		case MTHCA_OPCODE_RDMA_READ:
491 			entry->opcode    = IB_WC_RDMA_READ;
492 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
493 			break;
494 		case MTHCA_OPCODE_ATOMIC_CS:
495 			entry->opcode    = IB_WC_COMP_SWAP;
496 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
497 			break;
498 		case MTHCA_OPCODE_ATOMIC_FA:
499 			entry->opcode    = IB_WC_FETCH_ADD;
500 			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
501 			break;
502 		case MTHCA_OPCODE_BIND_MW:
503 			entry->opcode    = IB_WC_BIND_MW;
504 			break;
505 		default:
506 			entry->opcode    = MTHCA_OPCODE_INVALID;
507 			break;
508 		}
509 	} else {
510 		entry->byte_len = be32_to_cpu(cqe->byte_cnt);
511 		switch (cqe->opcode & 0x1f) {
512 		case IB_OPCODE_SEND_LAST_WITH_IMMEDIATE:
513 		case IB_OPCODE_SEND_ONLY_WITH_IMMEDIATE:
514 			entry->wc_flags = IB_WC_WITH_IMM;
515 			entry->imm_data = cqe->imm_etype_pkey_eec;
516 			entry->opcode = IB_WC_RECV;
517 			break;
518 		case IB_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
519 		case IB_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
520 			entry->wc_flags = IB_WC_WITH_IMM;
521 			entry->imm_data = cqe->imm_etype_pkey_eec;
522 			entry->opcode = IB_WC_RECV_RDMA_WITH_IMM;
523 			break;
524 		default:
525 			entry->wc_flags = 0;
526 			entry->opcode = IB_WC_RECV;
527 			break;
528 		}
529 		entry->slid 	   = be16_to_cpu(cqe->rlid);
530 		entry->sl   	   = be16_to_cpu(cqe->sl_g_mlpath) >> 12;
531 		entry->src_qp 	   = be32_to_cpu(cqe->rqpn) & 0xffffff;
532 		entry->dlid_path_bits = be16_to_cpu(cqe->sl_g_mlpath) & 0x7f;
533 		entry->pkey_index  = be32_to_cpu(cqe->imm_etype_pkey_eec) >> 16;
534 		entry->wc_flags   |= be16_to_cpu(cqe->sl_g_mlpath) & 0x80 ?
535 					IB_WC_GRH : 0;
536 	}
537 
538 	entry->status = IB_WC_SUCCESS;
539 
540  out:
541 	if (likely(free_cqe)) {
542 		set_cqe_hw(cqe);
543 		++(*freed);
544 		++cq->cons_index;
545 	}
546 
547 	return err;
548 }
549 
550 int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
551 		  struct ib_wc *entry)
552 {
553 	struct mthca_dev *dev = to_mdev(ibcq->device);
554 	struct mthca_cq *cq = to_mcq(ibcq);
555 	struct mthca_qp *qp = NULL;
556 	unsigned long flags;
557 	int err = 0;
558 	int freed = 0;
559 	int npolled;
560 
561 	spin_lock_irqsave(&cq->lock, flags);
562 
563 	for (npolled = 0; npolled < num_entries; ++npolled) {
564 		err = mthca_poll_one(dev, cq, &qp,
565 				     &freed, entry + npolled);
566 		if (err)
567 			break;
568 	}
569 
570 	if (freed) {
571 		wmb();
572 		update_cons_index(dev, cq, freed);
573 	}
574 
575 	spin_unlock_irqrestore(&cq->lock, flags);
576 
577 	return err == 0 || err == -EAGAIN ? npolled : err;
578 }
579 
580 int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify notify)
581 {
582 	u32 doorbell[2];
583 
584 	doorbell[0] = cpu_to_be32((notify == IB_CQ_SOLICITED ?
585 				   MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL :
586 				   MTHCA_TAVOR_CQ_DB_REQ_NOT)      |
587 				  to_mcq(cq)->cqn);
588 	doorbell[1] = 0xffffffff;
589 
590 	mthca_write64(doorbell,
591 		      to_mdev(cq->device)->kar + MTHCA_CQ_DOORBELL,
592 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(cq->device)->doorbell_lock));
593 
594 	return 0;
595 }
596 
597 int mthca_arbel_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify notify)
598 {
599 	struct mthca_cq *cq = to_mcq(ibcq);
600 	u32 doorbell[2];
601 	u32 sn;
602 	u32 ci;
603 
604 	sn = cq->arm_sn & 3;
605 	ci = cpu_to_be32(cq->cons_index);
606 
607 	doorbell[0] = ci;
608 	doorbell[1] = cpu_to_be32((cq->cqn << 8) | (2 << 5) | (sn << 3) |
609 				  (notify == IB_CQ_SOLICITED ? 1 : 2));
610 
611 	mthca_write_db_rec(doorbell, cq->arm_db);
612 
613 	/*
614 	 * Make sure that the doorbell record in host memory is
615 	 * written before ringing the doorbell via PCI MMIO.
616 	 */
617 	wmb();
618 
619 	doorbell[0] = cpu_to_be32((sn << 28)                       |
620 				  (notify == IB_CQ_SOLICITED ?
621 				   MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL :
622 				   MTHCA_ARBEL_CQ_DB_REQ_NOT)      |
623 				  cq->cqn);
624 	doorbell[1] = ci;
625 
626 	mthca_write64(doorbell,
627 		      to_mdev(ibcq->device)->kar + MTHCA_CQ_DOORBELL,
628 		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->doorbell_lock));
629 
630 	return 0;
631 }
632 
633 static void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq *cq)
634 {
635 	int i;
636 	int size;
637 
638 	if (cq->is_direct)
639 		dma_free_coherent(&dev->pdev->dev,
640 				  (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE,
641 				  cq->queue.direct.buf,
642 				  pci_unmap_addr(&cq->queue.direct,
643 						 mapping));
644 	else {
645 		size = (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE;
646 		for (i = 0; i < (size + PAGE_SIZE - 1) / PAGE_SIZE; ++i)
647 			if (cq->queue.page_list[i].buf)
648 				dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
649 						  cq->queue.page_list[i].buf,
650 						  pci_unmap_addr(&cq->queue.page_list[i],
651 								 mapping));
652 
653 		kfree(cq->queue.page_list);
654 	}
655 }
656 
657 static int mthca_alloc_cq_buf(struct mthca_dev *dev, int size,
658 			      struct mthca_cq *cq)
659 {
660 	int err = -ENOMEM;
661 	int npages, shift;
662 	u64 *dma_list = NULL;
663 	dma_addr_t t;
664 	int i;
665 
666 	if (size <= MTHCA_MAX_DIRECT_CQ_SIZE) {
667 		cq->is_direct = 1;
668 		npages        = 1;
669 		shift         = get_order(size) + PAGE_SHIFT;
670 
671 		cq->queue.direct.buf = dma_alloc_coherent(&dev->pdev->dev,
672 							  size, &t, GFP_KERNEL);
673 		if (!cq->queue.direct.buf)
674 			return -ENOMEM;
675 
676 		pci_unmap_addr_set(&cq->queue.direct, mapping, t);
677 
678 		memset(cq->queue.direct.buf, 0, size);
679 
680 		while (t & ((1 << shift) - 1)) {
681 			--shift;
682 			npages *= 2;
683 		}
684 
685 		dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
686 		if (!dma_list)
687 			goto err_free;
688 
689 		for (i = 0; i < npages; ++i)
690 			dma_list[i] = t + i * (1 << shift);
691 	} else {
692 		cq->is_direct = 0;
693 		npages        = (size + PAGE_SIZE - 1) / PAGE_SIZE;
694 		shift         = PAGE_SHIFT;
695 
696 		dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
697 		if (!dma_list)
698 			return -ENOMEM;
699 
700 		cq->queue.page_list = kmalloc(npages * sizeof *cq->queue.page_list,
701 					      GFP_KERNEL);
702 		if (!cq->queue.page_list)
703 			goto err_out;
704 
705 		for (i = 0; i < npages; ++i)
706 			cq->queue.page_list[i].buf = NULL;
707 
708 		for (i = 0; i < npages; ++i) {
709 			cq->queue.page_list[i].buf =
710 				dma_alloc_coherent(&dev->pdev->dev, PAGE_SIZE,
711 						   &t, GFP_KERNEL);
712 			if (!cq->queue.page_list[i].buf)
713 				goto err_free;
714 
715 			dma_list[i] = t;
716 			pci_unmap_addr_set(&cq->queue.page_list[i], mapping, t);
717 
718 			memset(cq->queue.page_list[i].buf, 0, PAGE_SIZE);
719 		}
720 	}
721 
722 	err = mthca_mr_alloc_phys(dev, dev->driver_pd.pd_num,
723 				  dma_list, shift, npages,
724 				  0, size,
725 				  MTHCA_MPT_FLAG_LOCAL_WRITE |
726 				  MTHCA_MPT_FLAG_LOCAL_READ,
727 				  &cq->mr);
728 	if (err)
729 		goto err_free;
730 
731 	kfree(dma_list);
732 
733 	return 0;
734 
735 err_free:
736 	mthca_free_cq_buf(dev, cq);
737 
738 err_out:
739 	kfree(dma_list);
740 
741 	return err;
742 }
743 
744 int mthca_init_cq(struct mthca_dev *dev, int nent,
745 		  struct mthca_cq *cq)
746 {
747 	int size = nent * MTHCA_CQ_ENTRY_SIZE;
748 	struct mthca_mailbox *mailbox;
749 	struct mthca_cq_context *cq_context;
750 	int err = -ENOMEM;
751 	u8 status;
752 	int i;
753 
754 	might_sleep();
755 
756 	cq->ibcq.cqe = nent - 1;
757 
758 	cq->cqn = mthca_alloc(&dev->cq_table.alloc);
759 	if (cq->cqn == -1)
760 		return -ENOMEM;
761 
762 	if (mthca_is_memfree(dev)) {
763 		cq->arm_sn = 1;
764 
765 		err = mthca_table_get(dev, dev->cq_table.table, cq->cqn);
766 		if (err)
767 			goto err_out;
768 
769 		err = -ENOMEM;
770 
771 		cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
772 						     cq->cqn, &cq->set_ci_db);
773 		if (cq->set_ci_db_index < 0)
774 			goto err_out_icm;
775 
776 		cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
777 						  cq->cqn, &cq->arm_db);
778 		if (cq->arm_db_index < 0)
779 			goto err_out_ci;
780 	}
781 
782 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
783 	if (IS_ERR(mailbox))
784 		goto err_out_arm;
785 
786 	cq_context = mailbox->buf;
787 
788 	err = mthca_alloc_cq_buf(dev, size, cq);
789 	if (err)
790 		goto err_out_mailbox;
791 
792 	for (i = 0; i < nent; ++i)
793 		set_cqe_hw(get_cqe(cq, i));
794 
795 	spin_lock_init(&cq->lock);
796 	atomic_set(&cq->refcount, 1);
797 	init_waitqueue_head(&cq->wait);
798 
799 	memset(cq_context, 0, sizeof *cq_context);
800 	cq_context->flags           = cpu_to_be32(MTHCA_CQ_STATUS_OK      |
801 						  MTHCA_CQ_STATE_DISARMED |
802 						  MTHCA_CQ_FLAG_TR);
803 	cq_context->start           = cpu_to_be64(0);
804 	cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24 |
805 						  dev->driver_uar.index);
806 	cq_context->error_eqn       = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
807 	cq_context->comp_eqn        = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
808 	cq_context->pd              = cpu_to_be32(dev->driver_pd.pd_num);
809 	cq_context->lkey            = cpu_to_be32(cq->mr.ibmr.lkey);
810 	cq_context->cqn             = cpu_to_be32(cq->cqn);
811 
812 	if (mthca_is_memfree(dev)) {
813 		cq_context->ci_db    = cpu_to_be32(cq->set_ci_db_index);
814 		cq_context->state_db = cpu_to_be32(cq->arm_db_index);
815 	}
816 
817 	err = mthca_SW2HW_CQ(dev, mailbox, cq->cqn, &status);
818 	if (err) {
819 		mthca_warn(dev, "SW2HW_CQ failed (%d)\n", err);
820 		goto err_out_free_mr;
821 	}
822 
823 	if (status) {
824 		mthca_warn(dev, "SW2HW_CQ returned status 0x%02x\n",
825 			   status);
826 		err = -EINVAL;
827 		goto err_out_free_mr;
828 	}
829 
830 	spin_lock_irq(&dev->cq_table.lock);
831 	if (mthca_array_set(&dev->cq_table.cq,
832 			    cq->cqn & (dev->limits.num_cqs - 1),
833 			    cq)) {
834 		spin_unlock_irq(&dev->cq_table.lock);
835 		goto err_out_free_mr;
836 	}
837 	spin_unlock_irq(&dev->cq_table.lock);
838 
839 	cq->cons_index = 0;
840 
841 	mthca_free_mailbox(dev, mailbox);
842 
843 	return 0;
844 
845 err_out_free_mr:
846 	mthca_free_mr(dev, &cq->mr);
847 	mthca_free_cq_buf(dev, cq);
848 
849 err_out_mailbox:
850 	mthca_free_mailbox(dev, mailbox);
851 
852 err_out_arm:
853 	if (mthca_is_memfree(dev))
854 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
855 
856 err_out_ci:
857 	if (mthca_is_memfree(dev))
858 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
859 
860 err_out_icm:
861 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
862 
863 err_out:
864 	mthca_free(&dev->cq_table.alloc, cq->cqn);
865 
866 	return err;
867 }
868 
869 void mthca_free_cq(struct mthca_dev *dev,
870 		   struct mthca_cq *cq)
871 {
872 	struct mthca_mailbox *mailbox;
873 	int err;
874 	u8 status;
875 
876 	might_sleep();
877 
878 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
879 	if (IS_ERR(mailbox)) {
880 		mthca_warn(dev, "No memory for mailbox to free CQ.\n");
881 		return;
882 	}
883 
884 	err = mthca_HW2SW_CQ(dev, mailbox, cq->cqn, &status);
885 	if (err)
886 		mthca_warn(dev, "HW2SW_CQ failed (%d)\n", err);
887 	else if (status)
888 		mthca_warn(dev, "HW2SW_CQ returned status 0x%02x\n", status);
889 
890 	if (0) {
891 		u32 *ctx = mailbox->buf;
892 		int j;
893 
894 		printk(KERN_ERR "context for CQN %x (cons index %x, next sw %d)\n",
895 		       cq->cqn, cq->cons_index, !!next_cqe_sw(cq));
896 		for (j = 0; j < 16; ++j)
897 			printk(KERN_ERR "[%2x] %08x\n", j * 4, be32_to_cpu(ctx[j]));
898 	}
899 
900 	spin_lock_irq(&dev->cq_table.lock);
901 	mthca_array_clear(&dev->cq_table.cq,
902 			  cq->cqn & (dev->limits.num_cqs - 1));
903 	spin_unlock_irq(&dev->cq_table.lock);
904 
905 	if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
906 		synchronize_irq(dev->eq_table.eq[MTHCA_EQ_COMP].msi_x_vector);
907 	else
908 		synchronize_irq(dev->pdev->irq);
909 
910 	atomic_dec(&cq->refcount);
911 	wait_event(cq->wait, !atomic_read(&cq->refcount));
912 
913 	mthca_free_mr(dev, &cq->mr);
914 	mthca_free_cq_buf(dev, cq);
915 
916 	if (mthca_is_memfree(dev)) {
917 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM,    cq->arm_db_index);
918 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
919 	}
920 
921 	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
922 	mthca_free(&dev->cq_table.alloc, cq->cqn);
923 	mthca_free_mailbox(dev, mailbox);
924 }
925 
926 int __devinit mthca_init_cq_table(struct mthca_dev *dev)
927 {
928 	int err;
929 
930 	spin_lock_init(&dev->cq_table.lock);
931 
932 	err = mthca_alloc_init(&dev->cq_table.alloc,
933 			       dev->limits.num_cqs,
934 			       (1 << 24) - 1,
935 			       dev->limits.reserved_cqs);
936 	if (err)
937 		return err;
938 
939 	err = mthca_array_init(&dev->cq_table.cq,
940 			       dev->limits.num_cqs);
941 	if (err)
942 		mthca_alloc_cleanup(&dev->cq_table.alloc);
943 
944 	return err;
945 }
946 
947 void __devexit mthca_cleanup_cq_table(struct mthca_dev *dev)
948 {
949 	mthca_array_cleanup(&dev->cq_table.cq, dev->limits.num_cqs);
950 	mthca_alloc_cleanup(&dev->cq_table.alloc);
951 }
952