[IB] mthca: report asynchronous CQ events
[safe/jmp/linux-2.6] / drivers / infiniband / hw / mthca / mthca_cq.c
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_completion(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_event(struct mthca_dev *dev, u32 cqn,
228                     enum ib_event_type event_type)
229 {
230         struct mthca_cq *cq;
231         struct ib_event event;
232
233         spin_lock(&dev->cq_table.lock);
234
235         cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
236
237         if (cq)
238                 atomic_inc(&cq->refcount);
239         spin_unlock(&dev->cq_table.lock);
240
241         if (!cq) {
242                 mthca_warn(dev, "Async event for bogus CQ %08x\n", cqn);
243                 return;
244         }
245
246         event.device      = &dev->ib_dev;
247         event.event       = event_type;
248         event.element.cq  = &cq->ibcq;
249         if (cq->ibcq.event_handler)
250                 cq->ibcq.event_handler(&event, cq->ibcq.cq_context);
251
252         if (atomic_dec_and_test(&cq->refcount))
253                 wake_up(&cq->wait);
254 }
255
256 void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn,
257                     struct mthca_srq *srq)
258 {
259         struct mthca_cq *cq;
260         struct mthca_cqe *cqe;
261         int prod_index;
262         int nfreed = 0;
263
264         spin_lock_irq(&dev->cq_table.lock);
265         cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
266         if (cq)
267                 atomic_inc(&cq->refcount);
268         spin_unlock_irq(&dev->cq_table.lock);
269
270         if (!cq)
271                 return;
272
273         spin_lock_irq(&cq->lock);
274
275         /*
276          * First we need to find the current producer index, so we
277          * know where to start cleaning from.  It doesn't matter if HW
278          * adds new entries after this loop -- the QP we're worried
279          * about is already in RESET, so the new entries won't come
280          * from our QP and therefore don't need to be checked.
281          */
282         for (prod_index = cq->cons_index;
283              cqe_sw(cq, prod_index & cq->ibcq.cqe);
284              ++prod_index)
285                 if (prod_index == cq->cons_index + cq->ibcq.cqe)
286                         break;
287
288         if (0)
289                 mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
290                           qpn, cqn, cq->cons_index, prod_index);
291
292         /*
293          * Now sweep backwards through the CQ, removing CQ entries
294          * that match our QP by copying older entries on top of them.
295          */
296         while (prod_index > cq->cons_index) {
297                 cqe = get_cqe(cq, (prod_index - 1) & cq->ibcq.cqe);
298                 if (cqe->my_qpn == cpu_to_be32(qpn)) {
299                         if (srq)
300                                 mthca_free_srq_wqe(srq, be32_to_cpu(cqe->wqe));
301                         ++nfreed;
302                 }
303                 else if (nfreed)
304                         memcpy(get_cqe(cq, (prod_index - 1 + nfreed) &
305                                        cq->ibcq.cqe),
306                                cqe,
307                                MTHCA_CQ_ENTRY_SIZE);
308                 --prod_index;
309         }
310
311         if (nfreed) {
312                 wmb();
313                 cq->cons_index += nfreed;
314                 update_cons_index(dev, cq, nfreed);
315         }
316
317         spin_unlock_irq(&cq->lock);
318         if (atomic_dec_and_test(&cq->refcount))
319                 wake_up(&cq->wait);
320 }
321
322 static int handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
323                             struct mthca_qp *qp, int wqe_index, int is_send,
324                             struct mthca_err_cqe *cqe,
325                             struct ib_wc *entry, int *free_cqe)
326 {
327         int err;
328         int dbd;
329         __be32 new_wqe;
330
331         if (cqe->syndrome == SYNDROME_LOCAL_QP_OP_ERR) {
332                 mthca_dbg(dev, "local QP operation err "
333                           "(QPN %06x, WQE @ %08x, CQN %06x, index %d)\n",
334                           be32_to_cpu(cqe->my_qpn), be32_to_cpu(cqe->wqe),
335                           cq->cqn, cq->cons_index);
336                 dump_cqe(dev, cqe);
337         }
338
339         /*
340          * For completions in error, only work request ID, status (and
341          * freed resource count for RD) have to be set.
342          */
343         switch (cqe->syndrome) {
344         case SYNDROME_LOCAL_LENGTH_ERR:
345                 entry->status = IB_WC_LOC_LEN_ERR;
346                 break;
347         case SYNDROME_LOCAL_QP_OP_ERR:
348                 entry->status = IB_WC_LOC_QP_OP_ERR;
349                 break;
350         case SYNDROME_LOCAL_EEC_OP_ERR:
351                 entry->status = IB_WC_LOC_EEC_OP_ERR;
352                 break;
353         case SYNDROME_LOCAL_PROT_ERR:
354                 entry->status = IB_WC_LOC_PROT_ERR;
355                 break;
356         case SYNDROME_WR_FLUSH_ERR:
357                 entry->status = IB_WC_WR_FLUSH_ERR;
358                 break;
359         case SYNDROME_MW_BIND_ERR:
360                 entry->status = IB_WC_MW_BIND_ERR;
361                 break;
362         case SYNDROME_BAD_RESP_ERR:
363                 entry->status = IB_WC_BAD_RESP_ERR;
364                 break;
365         case SYNDROME_LOCAL_ACCESS_ERR:
366                 entry->status = IB_WC_LOC_ACCESS_ERR;
367                 break;
368         case SYNDROME_REMOTE_INVAL_REQ_ERR:
369                 entry->status = IB_WC_REM_INV_REQ_ERR;
370                 break;
371         case SYNDROME_REMOTE_ACCESS_ERR:
372                 entry->status = IB_WC_REM_ACCESS_ERR;
373                 break;
374         case SYNDROME_REMOTE_OP_ERR:
375                 entry->status = IB_WC_REM_OP_ERR;
376                 break;
377         case SYNDROME_RETRY_EXC_ERR:
378                 entry->status = IB_WC_RETRY_EXC_ERR;
379                 break;
380         case SYNDROME_RNR_RETRY_EXC_ERR:
381                 entry->status = IB_WC_RNR_RETRY_EXC_ERR;
382                 break;
383         case SYNDROME_LOCAL_RDD_VIOL_ERR:
384                 entry->status = IB_WC_LOC_RDD_VIOL_ERR;
385                 break;
386         case SYNDROME_REMOTE_INVAL_RD_REQ_ERR:
387                 entry->status = IB_WC_REM_INV_RD_REQ_ERR;
388                 break;
389         case SYNDROME_REMOTE_ABORTED_ERR:
390                 entry->status = IB_WC_REM_ABORT_ERR;
391                 break;
392         case SYNDROME_INVAL_EECN_ERR:
393                 entry->status = IB_WC_INV_EECN_ERR;
394                 break;
395         case SYNDROME_INVAL_EEC_STATE_ERR:
396                 entry->status = IB_WC_INV_EEC_STATE_ERR;
397                 break;
398         default:
399                 entry->status = IB_WC_GENERAL_ERR;
400                 break;
401         }
402
403         /*
404          * Mem-free HCAs always generate one CQE per WQE, even in the
405          * error case, so we don't have to check the doorbell count, etc.
406          */
407         if (mthca_is_memfree(dev))
408                 return 0;
409
410         err = mthca_free_err_wqe(dev, qp, is_send, wqe_index, &dbd, &new_wqe);
411         if (err)
412                 return err;
413
414         /*
415          * If we're at the end of the WQE chain, or we've used up our
416          * doorbell count, free the CQE.  Otherwise just update it for
417          * the next poll operation.
418          */
419         if (!(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
420                 return 0;
421
422         cqe->db_cnt   = cpu_to_be16(be16_to_cpu(cqe->db_cnt) - dbd);
423         cqe->wqe      = new_wqe;
424         cqe->syndrome = SYNDROME_WR_FLUSH_ERR;
425
426         *free_cqe = 0;
427
428         return 0;
429 }
430
431 static inline int mthca_poll_one(struct mthca_dev *dev,
432                                  struct mthca_cq *cq,
433                                  struct mthca_qp **cur_qp,
434                                  int *freed,
435                                  struct ib_wc *entry)
436 {
437         struct mthca_wq *wq;
438         struct mthca_cqe *cqe;
439         int wqe_index;
440         int is_error;
441         int is_send;
442         int free_cqe = 1;
443         int err = 0;
444
445         cqe = next_cqe_sw(cq);
446         if (!cqe)
447                 return -EAGAIN;
448
449         /*
450          * Make sure we read CQ entry contents after we've checked the
451          * ownership bit.
452          */
453         rmb();
454
455         if (0) {
456                 mthca_dbg(dev, "%x/%d: CQE -> QPN %06x, WQE @ %08x\n",
457                           cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
458                           be32_to_cpu(cqe->wqe));
459                 dump_cqe(dev, cqe);
460         }
461
462         is_error = (cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
463                 MTHCA_ERROR_CQE_OPCODE_MASK;
464         is_send  = is_error ? cqe->opcode & 0x01 : cqe->is_send & 0x80;
465
466         if (!*cur_qp || be32_to_cpu(cqe->my_qpn) != (*cur_qp)->qpn) {
467                 /*
468                  * We do not have to take the QP table lock here,
469                  * because CQs will be locked while QPs are removed
470                  * from the table.
471                  */
472                 *cur_qp = mthca_array_get(&dev->qp_table.qp,
473                                           be32_to_cpu(cqe->my_qpn) &
474                                           (dev->limits.num_qps - 1));
475                 if (!*cur_qp) {
476                         mthca_warn(dev, "CQ entry for unknown QP %06x\n",
477                                    be32_to_cpu(cqe->my_qpn) & 0xffffff);
478                         err = -EINVAL;
479                         goto out;
480                 }
481         }
482
483         entry->qp_num = (*cur_qp)->qpn;
484
485         if (is_send) {
486                 wq = &(*cur_qp)->sq;
487                 wqe_index = ((be32_to_cpu(cqe->wqe) - (*cur_qp)->send_wqe_offset)
488                              >> wq->wqe_shift);
489                 entry->wr_id = (*cur_qp)->wrid[wqe_index +
490                                                (*cur_qp)->rq.max];
491         } else if ((*cur_qp)->ibqp.srq) {
492                 struct mthca_srq *srq = to_msrq((*cur_qp)->ibqp.srq);
493                 u32 wqe = be32_to_cpu(cqe->wqe);
494                 wq = NULL;
495                 wqe_index = wqe >> srq->wqe_shift;
496                 entry->wr_id = srq->wrid[wqe_index];
497                 mthca_free_srq_wqe(srq, wqe);
498         } else {
499                 wq = &(*cur_qp)->rq;
500                 wqe_index = be32_to_cpu(cqe->wqe) >> wq->wqe_shift;
501                 entry->wr_id = (*cur_qp)->wrid[wqe_index];
502         }
503
504         if (wq) {
505                 if (wq->last_comp < wqe_index)
506                         wq->tail += wqe_index - wq->last_comp;
507                 else
508                         wq->tail += wqe_index + wq->max - wq->last_comp;
509
510                 wq->last_comp = wqe_index;
511         }
512
513         if (is_error) {
514                 err = handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,
515                                        (struct mthca_err_cqe *) cqe,
516                                        entry, &free_cqe);
517                 goto out;
518         }
519
520         if (is_send) {
521                 entry->wc_flags = 0;
522                 switch (cqe->opcode) {
523                 case MTHCA_OPCODE_RDMA_WRITE:
524                         entry->opcode    = IB_WC_RDMA_WRITE;
525                         break;
526                 case MTHCA_OPCODE_RDMA_WRITE_IMM:
527                         entry->opcode    = IB_WC_RDMA_WRITE;
528                         entry->wc_flags |= IB_WC_WITH_IMM;
529                         break;
530                 case MTHCA_OPCODE_SEND:
531                         entry->opcode    = IB_WC_SEND;
532                         break;
533                 case MTHCA_OPCODE_SEND_IMM:
534                         entry->opcode    = IB_WC_SEND;
535                         entry->wc_flags |= IB_WC_WITH_IMM;
536                         break;
537                 case MTHCA_OPCODE_RDMA_READ:
538                         entry->opcode    = IB_WC_RDMA_READ;
539                         entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
540                         break;
541                 case MTHCA_OPCODE_ATOMIC_CS:
542                         entry->opcode    = IB_WC_COMP_SWAP;
543                         entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
544                         break;
545                 case MTHCA_OPCODE_ATOMIC_FA:
546                         entry->opcode    = IB_WC_FETCH_ADD;
547                         entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
548                         break;
549                 case MTHCA_OPCODE_BIND_MW:
550                         entry->opcode    = IB_WC_BIND_MW;
551                         break;
552                 default:
553                         entry->opcode    = MTHCA_OPCODE_INVALID;
554                         break;
555                 }
556         } else {
557                 entry->byte_len = be32_to_cpu(cqe->byte_cnt);
558                 switch (cqe->opcode & 0x1f) {
559                 case IB_OPCODE_SEND_LAST_WITH_IMMEDIATE:
560                 case IB_OPCODE_SEND_ONLY_WITH_IMMEDIATE:
561                         entry->wc_flags = IB_WC_WITH_IMM;
562                         entry->imm_data = cqe->imm_etype_pkey_eec;
563                         entry->opcode = IB_WC_RECV;
564                         break;
565                 case IB_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
566                 case IB_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
567                         entry->wc_flags = IB_WC_WITH_IMM;
568                         entry->imm_data = cqe->imm_etype_pkey_eec;
569                         entry->opcode = IB_WC_RECV_RDMA_WITH_IMM;
570                         break;
571                 default:
572                         entry->wc_flags = 0;
573                         entry->opcode = IB_WC_RECV;
574                         break;
575                 }
576                 entry->slid        = be16_to_cpu(cqe->rlid);
577                 entry->sl          = be16_to_cpu(cqe->sl_g_mlpath) >> 12;
578                 entry->src_qp      = be32_to_cpu(cqe->rqpn) & 0xffffff;
579                 entry->dlid_path_bits = be16_to_cpu(cqe->sl_g_mlpath) & 0x7f;
580                 entry->pkey_index  = be32_to_cpu(cqe->imm_etype_pkey_eec) >> 16;
581                 entry->wc_flags   |= be16_to_cpu(cqe->sl_g_mlpath) & 0x80 ?
582                                         IB_WC_GRH : 0;
583         }
584
585         entry->status = IB_WC_SUCCESS;
586
587  out:
588         if (likely(free_cqe)) {
589                 set_cqe_hw(cqe);
590                 ++(*freed);
591                 ++cq->cons_index;
592         }
593
594         return err;
595 }
596
597 int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
598                   struct ib_wc *entry)
599 {
600         struct mthca_dev *dev = to_mdev(ibcq->device);
601         struct mthca_cq *cq = to_mcq(ibcq);
602         struct mthca_qp *qp = NULL;
603         unsigned long flags;
604         int err = 0;
605         int freed = 0;
606         int npolled;
607
608         spin_lock_irqsave(&cq->lock, flags);
609
610         for (npolled = 0; npolled < num_entries; ++npolled) {
611                 err = mthca_poll_one(dev, cq, &qp,
612                                      &freed, entry + npolled);
613                 if (err)
614                         break;
615         }
616
617         if (freed) {
618                 wmb();
619                 update_cons_index(dev, cq, freed);
620         }
621
622         spin_unlock_irqrestore(&cq->lock, flags);
623
624         return err == 0 || err == -EAGAIN ? npolled : err;
625 }
626
627 int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify notify)
628 {
629         __be32 doorbell[2];
630
631         doorbell[0] = cpu_to_be32((notify == IB_CQ_SOLICITED ?
632                                    MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL :
633                                    MTHCA_TAVOR_CQ_DB_REQ_NOT)      |
634                                   to_mcq(cq)->cqn);
635         doorbell[1] = (__force __be32) 0xffffffff;
636
637         mthca_write64(doorbell,
638                       to_mdev(cq->device)->kar + MTHCA_CQ_DOORBELL,
639                       MTHCA_GET_DOORBELL_LOCK(&to_mdev(cq->device)->doorbell_lock));
640
641         return 0;
642 }
643
644 int mthca_arbel_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify notify)
645 {
646         struct mthca_cq *cq = to_mcq(ibcq);
647         __be32 doorbell[2];
648         u32 sn;
649         __be32 ci;
650
651         sn = cq->arm_sn & 3;
652         ci = cpu_to_be32(cq->cons_index);
653
654         doorbell[0] = ci;
655         doorbell[1] = cpu_to_be32((cq->cqn << 8) | (2 << 5) | (sn << 3) |
656                                   (notify == IB_CQ_SOLICITED ? 1 : 2));
657
658         mthca_write_db_rec(doorbell, cq->arm_db);
659
660         /*
661          * Make sure that the doorbell record in host memory is
662          * written before ringing the doorbell via PCI MMIO.
663          */
664         wmb();
665
666         doorbell[0] = cpu_to_be32((sn << 28)                       |
667                                   (notify == IB_CQ_SOLICITED ?
668                                    MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL :
669                                    MTHCA_ARBEL_CQ_DB_REQ_NOT)      |
670                                   cq->cqn);
671         doorbell[1] = ci;
672
673         mthca_write64(doorbell,
674                       to_mdev(ibcq->device)->kar + MTHCA_CQ_DOORBELL,
675                       MTHCA_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->doorbell_lock));
676
677         return 0;
678 }
679
680 static void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq *cq)
681 {
682         mthca_buf_free(dev, (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE,
683                        &cq->queue, cq->is_direct, &cq->mr);
684 }
685
686 int mthca_init_cq(struct mthca_dev *dev, int nent,
687                   struct mthca_ucontext *ctx, u32 pdn,
688                   struct mthca_cq *cq)
689 {
690         int size = nent * MTHCA_CQ_ENTRY_SIZE;
691         struct mthca_mailbox *mailbox;
692         struct mthca_cq_context *cq_context;
693         int err = -ENOMEM;
694         u8 status;
695         int i;
696
697         might_sleep();
698
699         cq->ibcq.cqe  = nent - 1;
700         cq->is_kernel = !ctx;
701
702         cq->cqn = mthca_alloc(&dev->cq_table.alloc);
703         if (cq->cqn == -1)
704                 return -ENOMEM;
705
706         if (mthca_is_memfree(dev)) {
707                 err = mthca_table_get(dev, dev->cq_table.table, cq->cqn);
708                 if (err)
709                         goto err_out;
710
711                 if (cq->is_kernel) {
712                         cq->arm_sn = 1;
713
714                         err = -ENOMEM;
715
716                         cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
717                                                              cq->cqn, &cq->set_ci_db);
718                         if (cq->set_ci_db_index < 0)
719                                 goto err_out_icm;
720
721                         cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
722                                                           cq->cqn, &cq->arm_db);
723                         if (cq->arm_db_index < 0)
724                                 goto err_out_ci;
725                 }
726         }
727
728         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
729         if (IS_ERR(mailbox))
730                 goto err_out_arm;
731
732         cq_context = mailbox->buf;
733
734         if (cq->is_kernel) {
735                 err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_CQ_SIZE,
736                                       &cq->queue, &cq->is_direct,
737                                       &dev->driver_pd, 1, &cq->mr);
738                 if (err)
739                         goto err_out_mailbox;
740
741                 for (i = 0; i < nent; ++i)
742                         set_cqe_hw(get_cqe(cq, i));
743         }
744
745         spin_lock_init(&cq->lock);
746         atomic_set(&cq->refcount, 1);
747         init_waitqueue_head(&cq->wait);
748
749         memset(cq_context, 0, sizeof *cq_context);
750         cq_context->flags           = cpu_to_be32(MTHCA_CQ_STATUS_OK      |
751                                                   MTHCA_CQ_STATE_DISARMED |
752                                                   MTHCA_CQ_FLAG_TR);
753         cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24);
754         if (ctx)
755                 cq_context->logsize_usrpage |= cpu_to_be32(ctx->uar.index);
756         else
757                 cq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
758         cq_context->error_eqn       = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
759         cq_context->comp_eqn        = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
760         cq_context->pd              = cpu_to_be32(pdn);
761         cq_context->lkey            = cpu_to_be32(cq->mr.ibmr.lkey);
762         cq_context->cqn             = cpu_to_be32(cq->cqn);
763
764         if (mthca_is_memfree(dev)) {
765                 cq_context->ci_db    = cpu_to_be32(cq->set_ci_db_index);
766                 cq_context->state_db = cpu_to_be32(cq->arm_db_index);
767         }
768
769         err = mthca_SW2HW_CQ(dev, mailbox, cq->cqn, &status);
770         if (err) {
771                 mthca_warn(dev, "SW2HW_CQ failed (%d)\n", err);
772                 goto err_out_free_mr;
773         }
774
775         if (status) {
776                 mthca_warn(dev, "SW2HW_CQ returned status 0x%02x\n",
777                            status);
778                 err = -EINVAL;
779                 goto err_out_free_mr;
780         }
781
782         spin_lock_irq(&dev->cq_table.lock);
783         if (mthca_array_set(&dev->cq_table.cq,
784                             cq->cqn & (dev->limits.num_cqs - 1),
785                             cq)) {
786                 spin_unlock_irq(&dev->cq_table.lock);
787                 goto err_out_free_mr;
788         }
789         spin_unlock_irq(&dev->cq_table.lock);
790
791         cq->cons_index = 0;
792
793         mthca_free_mailbox(dev, mailbox);
794
795         return 0;
796
797 err_out_free_mr:
798         if (cq->is_kernel)
799                 mthca_free_cq_buf(dev, cq);
800
801 err_out_mailbox:
802         mthca_free_mailbox(dev, mailbox);
803
804 err_out_arm:
805         if (cq->is_kernel && mthca_is_memfree(dev))
806                 mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
807
808 err_out_ci:
809         if (cq->is_kernel && mthca_is_memfree(dev))
810                 mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
811
812 err_out_icm:
813         mthca_table_put(dev, dev->cq_table.table, cq->cqn);
814
815 err_out:
816         mthca_free(&dev->cq_table.alloc, cq->cqn);
817
818         return err;
819 }
820
821 void mthca_free_cq(struct mthca_dev *dev,
822                    struct mthca_cq *cq)
823 {
824         struct mthca_mailbox *mailbox;
825         int err;
826         u8 status;
827
828         might_sleep();
829
830         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
831         if (IS_ERR(mailbox)) {
832                 mthca_warn(dev, "No memory for mailbox to free CQ.\n");
833                 return;
834         }
835
836         err = mthca_HW2SW_CQ(dev, mailbox, cq->cqn, &status);
837         if (err)
838                 mthca_warn(dev, "HW2SW_CQ failed (%d)\n", err);
839         else if (status)
840                 mthca_warn(dev, "HW2SW_CQ returned status 0x%02x\n", status);
841
842         if (0) {
843                 __be32 *ctx = mailbox->buf;
844                 int j;
845
846                 printk(KERN_ERR "context for CQN %x (cons index %x, next sw %d)\n",
847                        cq->cqn, cq->cons_index,
848                        cq->is_kernel ? !!next_cqe_sw(cq) : 0);
849                 for (j = 0; j < 16; ++j)
850                         printk(KERN_ERR "[%2x] %08x\n", j * 4, be32_to_cpu(ctx[j]));
851         }
852
853         spin_lock_irq(&dev->cq_table.lock);
854         mthca_array_clear(&dev->cq_table.cq,
855                           cq->cqn & (dev->limits.num_cqs - 1));
856         spin_unlock_irq(&dev->cq_table.lock);
857
858         if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
859                 synchronize_irq(dev->eq_table.eq[MTHCA_EQ_COMP].msi_x_vector);
860         else
861                 synchronize_irq(dev->pdev->irq);
862
863         atomic_dec(&cq->refcount);
864         wait_event(cq->wait, !atomic_read(&cq->refcount));
865
866         if (cq->is_kernel) {
867                 mthca_free_cq_buf(dev, cq);
868                 if (mthca_is_memfree(dev)) {
869                         mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM,    cq->arm_db_index);
870                         mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
871                 }
872         }
873
874         mthca_table_put(dev, dev->cq_table.table, cq->cqn);
875         mthca_free(&dev->cq_table.alloc, cq->cqn);
876         mthca_free_mailbox(dev, mailbox);
877 }
878
879 int __devinit mthca_init_cq_table(struct mthca_dev *dev)
880 {
881         int err;
882
883         spin_lock_init(&dev->cq_table.lock);
884
885         err = mthca_alloc_init(&dev->cq_table.alloc,
886                                dev->limits.num_cqs,
887                                (1 << 24) - 1,
888                                dev->limits.reserved_cqs);
889         if (err)
890                 return err;
891
892         err = mthca_array_init(&dev->cq_table.cq,
893                                dev->limits.num_cqs);
894         if (err)
895                 mthca_alloc_cleanup(&dev->cq_table.alloc);
896
897         return err;
898 }
899
900 void __devexit mthca_cleanup_cq_table(struct mthca_dev *dev)
901 {
902         mthca_array_cleanup(&dev->cq_table.cq, dev->limits.num_cqs);
903         mthca_alloc_cleanup(&dev->cq_table.alloc);
904 }