Merge branch 'master' of git://oak/home/sfr/kernels/iseries/work
[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, 2006 Cisco Systems, Inc. All rights reserved.
5  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
6  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
7  *
8  * This software is available to you under a choice of one of two
9  * licenses.  You may choose to be licensed under the terms of the GNU
10  * General Public License (GPL) Version 2, available from the file
11  * COPYING in the main directory of this source tree, or the
12  * OpenIB.org BSD license below:
13  *
14  *     Redistribution and use in source and binary forms, with or
15  *     without modification, are permitted provided that the following
16  *     conditions are met:
17  *
18  *      - Redistributions of source code must retain the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer.
21  *
22  *      - Redistributions in binary form must reproduce the above
23  *        copyright notice, this list of conditions and the following
24  *        disclaimer in the documentation and/or other materials
25  *        provided with the distribution.
26  *
27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34  * SOFTWARE.
35  *
36  * $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     vendor_err;
132         __be16 db_cnt;
133         u32    reserved2;
134         __be32 wqe;
135         u8     opcode;
136         u8     reserved3[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_from_buf(struct mthca_cq_buf *buf,
154                                                  int entry)
155 {
156         if (buf->is_direct)
157                 return buf->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
158         else
159                 return buf->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
160                         + (entry * MTHCA_CQ_ENTRY_SIZE) % PAGE_SIZE;
161 }
162
163 static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
164 {
165         return get_cqe_from_buf(&cq->buf, entry);
166 }
167
168 static inline struct mthca_cqe *cqe_sw(struct mthca_cqe *cqe)
169 {
170         return MTHCA_CQ_ENTRY_OWNER_HW & cqe->owner ? NULL : cqe;
171 }
172
173 static inline struct mthca_cqe *next_cqe_sw(struct mthca_cq *cq)
174 {
175         return cqe_sw(get_cqe(cq, cq->cons_index & cq->ibcq.cqe));
176 }
177
178 static inline void set_cqe_hw(struct mthca_cqe *cqe)
179 {
180         cqe->owner = MTHCA_CQ_ENTRY_OWNER_HW;
181 }
182
183 static void dump_cqe(struct mthca_dev *dev, void *cqe_ptr)
184 {
185         __be32 *cqe = cqe_ptr;
186
187         (void) cqe;     /* avoid warning if mthca_dbg compiled away... */
188         mthca_dbg(dev, "CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
189                   be32_to_cpu(cqe[0]), be32_to_cpu(cqe[1]), be32_to_cpu(cqe[2]),
190                   be32_to_cpu(cqe[3]), be32_to_cpu(cqe[4]), be32_to_cpu(cqe[5]),
191                   be32_to_cpu(cqe[6]), be32_to_cpu(cqe[7]));
192 }
193
194 /*
195  * incr is ignored in native Arbel (mem-free) mode, so cq->cons_index
196  * should be correct before calling update_cons_index().
197  */
198 static inline void update_cons_index(struct mthca_dev *dev, struct mthca_cq *cq,
199                                      int incr)
200 {
201         __be32 doorbell[2];
202
203         if (mthca_is_memfree(dev)) {
204                 *cq->set_ci_db = cpu_to_be32(cq->cons_index);
205                 wmb();
206         } else {
207                 doorbell[0] = cpu_to_be32(MTHCA_TAVOR_CQ_DB_INC_CI | cq->cqn);
208                 doorbell[1] = cpu_to_be32(incr - 1);
209
210                 mthca_write64(doorbell,
211                               dev->kar + MTHCA_CQ_DOORBELL,
212                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
213         }
214 }
215
216 void mthca_cq_completion(struct mthca_dev *dev, u32 cqn)
217 {
218         struct mthca_cq *cq;
219
220         cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
221
222         if (!cq) {
223                 mthca_warn(dev, "Completion event for bogus CQ %08x\n", cqn);
224                 return;
225         }
226
227         ++cq->arm_sn;
228
229         cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
230 }
231
232 void mthca_cq_event(struct mthca_dev *dev, u32 cqn,
233                     enum ib_event_type event_type)
234 {
235         struct mthca_cq *cq;
236         struct ib_event event;
237
238         spin_lock(&dev->cq_table.lock);
239
240         cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
241         if (cq)
242                 ++cq->refcount;
243
244         spin_unlock(&dev->cq_table.lock);
245
246         if (!cq) {
247                 mthca_warn(dev, "Async event for bogus CQ %08x\n", cqn);
248                 return;
249         }
250
251         event.device      = &dev->ib_dev;
252         event.event       = event_type;
253         event.element.cq  = &cq->ibcq;
254         if (cq->ibcq.event_handler)
255                 cq->ibcq.event_handler(&event, cq->ibcq.cq_context);
256
257         spin_lock(&dev->cq_table.lock);
258         if (!--cq->refcount)
259                 wake_up(&cq->wait);
260         spin_unlock(&dev->cq_table.lock);
261 }
262
263 static inline int is_recv_cqe(struct mthca_cqe *cqe)
264 {
265         if ((cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
266             MTHCA_ERROR_CQE_OPCODE_MASK)
267                 return !(cqe->opcode & 0x01);
268         else
269                 return !(cqe->is_send & 0x80);
270 }
271
272 void mthca_cq_clean(struct mthca_dev *dev, struct mthca_cq *cq, u32 qpn,
273                     struct mthca_srq *srq)
274 {
275         struct mthca_cqe *cqe;
276         u32 prod_index;
277         int nfreed = 0;
278
279         spin_lock_irq(&cq->lock);
280
281         /*
282          * First we need to find the current producer index, so we
283          * know where to start cleaning from.  It doesn't matter if HW
284          * adds new entries after this loop -- the QP we're worried
285          * about is already in RESET, so the new entries won't come
286          * from our QP and therefore don't need to be checked.
287          */
288         for (prod_index = cq->cons_index;
289              cqe_sw(get_cqe(cq, prod_index & cq->ibcq.cqe));
290              ++prod_index)
291                 if (prod_index == cq->cons_index + cq->ibcq.cqe)
292                         break;
293
294         if (0)
295                 mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
296                           qpn, cq->cqn, cq->cons_index, prod_index);
297
298         /*
299          * Now sweep backwards through the CQ, removing CQ entries
300          * that match our QP by copying older entries on top of them.
301          */
302         while ((int) --prod_index - (int) cq->cons_index >= 0) {
303                 cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
304                 if (cqe->my_qpn == cpu_to_be32(qpn)) {
305                         if (srq && is_recv_cqe(cqe))
306                                 mthca_free_srq_wqe(srq, be32_to_cpu(cqe->wqe));
307                         ++nfreed;
308                 } else if (nfreed)
309                         memcpy(get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe),
310                                cqe, MTHCA_CQ_ENTRY_SIZE);
311         }
312
313         if (nfreed) {
314                 wmb();
315                 cq->cons_index += nfreed;
316                 update_cons_index(dev, cq, nfreed);
317         }
318
319         spin_unlock_irq(&cq->lock);
320 }
321
322 void mthca_cq_resize_copy_cqes(struct mthca_cq *cq)
323 {
324         int i;
325
326         /*
327          * In Tavor mode, the hardware keeps the consumer and producer
328          * indices mod the CQ size.  Since we might be making the CQ
329          * bigger, we need to deal with the case where the producer
330          * index wrapped around before the CQ was resized.
331          */
332         if (!mthca_is_memfree(to_mdev(cq->ibcq.device)) &&
333             cq->ibcq.cqe < cq->resize_buf->cqe) {
334                 cq->cons_index &= cq->ibcq.cqe;
335                 if (cqe_sw(get_cqe(cq, cq->ibcq.cqe)))
336                         cq->cons_index -= cq->ibcq.cqe + 1;
337         }
338
339         for (i = cq->cons_index; cqe_sw(get_cqe(cq, i & cq->ibcq.cqe)); ++i)
340                 memcpy(get_cqe_from_buf(&cq->resize_buf->buf,
341                                         i & cq->resize_buf->cqe),
342                        get_cqe(cq, i & cq->ibcq.cqe), MTHCA_CQ_ENTRY_SIZE);
343 }
344
345 int mthca_alloc_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int nent)
346 {
347         int ret;
348         int i;
349
350         ret = mthca_buf_alloc(dev, nent * MTHCA_CQ_ENTRY_SIZE,
351                               MTHCA_MAX_DIRECT_CQ_SIZE,
352                               &buf->queue, &buf->is_direct,
353                               &dev->driver_pd, 1, &buf->mr);
354         if (ret)
355                 return ret;
356
357         for (i = 0; i < nent; ++i)
358                 set_cqe_hw(get_cqe_from_buf(buf, i));
359
360         return 0;
361 }
362
363 void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int cqe)
364 {
365         mthca_buf_free(dev, (cqe + 1) * MTHCA_CQ_ENTRY_SIZE, &buf->queue,
366                        buf->is_direct, &buf->mr);
367 }
368
369 static void handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
370                              struct mthca_qp *qp, int wqe_index, int is_send,
371                              struct mthca_err_cqe *cqe,
372                              struct ib_wc *entry, int *free_cqe)
373 {
374         int dbd;
375         __be32 new_wqe;
376
377         if (cqe->syndrome == SYNDROME_LOCAL_QP_OP_ERR) {
378                 mthca_dbg(dev, "local QP operation err "
379                           "(QPN %06x, WQE @ %08x, CQN %06x, index %d)\n",
380                           be32_to_cpu(cqe->my_qpn), be32_to_cpu(cqe->wqe),
381                           cq->cqn, cq->cons_index);
382                 dump_cqe(dev, cqe);
383         }
384
385         /*
386          * For completions in error, only work request ID, status, vendor error
387          * (and freed resource count for RD) have to be set.
388          */
389         switch (cqe->syndrome) {
390         case SYNDROME_LOCAL_LENGTH_ERR:
391                 entry->status = IB_WC_LOC_LEN_ERR;
392                 break;
393         case SYNDROME_LOCAL_QP_OP_ERR:
394                 entry->status = IB_WC_LOC_QP_OP_ERR;
395                 break;
396         case SYNDROME_LOCAL_EEC_OP_ERR:
397                 entry->status = IB_WC_LOC_EEC_OP_ERR;
398                 break;
399         case SYNDROME_LOCAL_PROT_ERR:
400                 entry->status = IB_WC_LOC_PROT_ERR;
401                 break;
402         case SYNDROME_WR_FLUSH_ERR:
403                 entry->status = IB_WC_WR_FLUSH_ERR;
404                 break;
405         case SYNDROME_MW_BIND_ERR:
406                 entry->status = IB_WC_MW_BIND_ERR;
407                 break;
408         case SYNDROME_BAD_RESP_ERR:
409                 entry->status = IB_WC_BAD_RESP_ERR;
410                 break;
411         case SYNDROME_LOCAL_ACCESS_ERR:
412                 entry->status = IB_WC_LOC_ACCESS_ERR;
413                 break;
414         case SYNDROME_REMOTE_INVAL_REQ_ERR:
415                 entry->status = IB_WC_REM_INV_REQ_ERR;
416                 break;
417         case SYNDROME_REMOTE_ACCESS_ERR:
418                 entry->status = IB_WC_REM_ACCESS_ERR;
419                 break;
420         case SYNDROME_REMOTE_OP_ERR:
421                 entry->status = IB_WC_REM_OP_ERR;
422                 break;
423         case SYNDROME_RETRY_EXC_ERR:
424                 entry->status = IB_WC_RETRY_EXC_ERR;
425                 break;
426         case SYNDROME_RNR_RETRY_EXC_ERR:
427                 entry->status = IB_WC_RNR_RETRY_EXC_ERR;
428                 break;
429         case SYNDROME_LOCAL_RDD_VIOL_ERR:
430                 entry->status = IB_WC_LOC_RDD_VIOL_ERR;
431                 break;
432         case SYNDROME_REMOTE_INVAL_RD_REQ_ERR:
433                 entry->status = IB_WC_REM_INV_RD_REQ_ERR;
434                 break;
435         case SYNDROME_REMOTE_ABORTED_ERR:
436                 entry->status = IB_WC_REM_ABORT_ERR;
437                 break;
438         case SYNDROME_INVAL_EECN_ERR:
439                 entry->status = IB_WC_INV_EECN_ERR;
440                 break;
441         case SYNDROME_INVAL_EEC_STATE_ERR:
442                 entry->status = IB_WC_INV_EEC_STATE_ERR;
443                 break;
444         default:
445                 entry->status = IB_WC_GENERAL_ERR;
446                 break;
447         }
448
449         entry->vendor_err = cqe->vendor_err;
450
451         /*
452          * Mem-free HCAs always generate one CQE per WQE, even in the
453          * error case, so we don't have to check the doorbell count, etc.
454          */
455         if (mthca_is_memfree(dev))
456                 return;
457
458         mthca_free_err_wqe(dev, qp, is_send, wqe_index, &dbd, &new_wqe);
459
460         /*
461          * If we're at the end of the WQE chain, or we've used up our
462          * doorbell count, free the CQE.  Otherwise just update it for
463          * the next poll operation.
464          */
465         if (!(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
466                 return;
467
468         cqe->db_cnt   = cpu_to_be16(be16_to_cpu(cqe->db_cnt) - dbd);
469         cqe->wqe      = new_wqe;
470         cqe->syndrome = SYNDROME_WR_FLUSH_ERR;
471
472         *free_cqe = 0;
473 }
474
475 static inline int mthca_poll_one(struct mthca_dev *dev,
476                                  struct mthca_cq *cq,
477                                  struct mthca_qp **cur_qp,
478                                  int *freed,
479                                  struct ib_wc *entry)
480 {
481         struct mthca_wq *wq;
482         struct mthca_cqe *cqe;
483         int wqe_index;
484         int is_error;
485         int is_send;
486         int free_cqe = 1;
487         int err = 0;
488
489         cqe = next_cqe_sw(cq);
490         if (!cqe)
491                 return -EAGAIN;
492
493         /*
494          * Make sure we read CQ entry contents after we've checked the
495          * ownership bit.
496          */
497         rmb();
498
499         if (0) {
500                 mthca_dbg(dev, "%x/%d: CQE -> QPN %06x, WQE @ %08x\n",
501                           cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
502                           be32_to_cpu(cqe->wqe));
503                 dump_cqe(dev, cqe);
504         }
505
506         is_error = (cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
507                 MTHCA_ERROR_CQE_OPCODE_MASK;
508         is_send  = is_error ? cqe->opcode & 0x01 : cqe->is_send & 0x80;
509
510         if (!*cur_qp || be32_to_cpu(cqe->my_qpn) != (*cur_qp)->qpn) {
511                 /*
512                  * We do not have to take the QP table lock here,
513                  * because CQs will be locked while QPs are removed
514                  * from the table.
515                  */
516                 *cur_qp = mthca_array_get(&dev->qp_table.qp,
517                                           be32_to_cpu(cqe->my_qpn) &
518                                           (dev->limits.num_qps - 1));
519                 if (!*cur_qp) {
520                         mthca_warn(dev, "CQ entry for unknown QP %06x\n",
521                                    be32_to_cpu(cqe->my_qpn) & 0xffffff);
522                         err = -EINVAL;
523                         goto out;
524                 }
525         }
526
527         entry->qp_num = (*cur_qp)->qpn;
528
529         if (is_send) {
530                 wq = &(*cur_qp)->sq;
531                 wqe_index = ((be32_to_cpu(cqe->wqe) - (*cur_qp)->send_wqe_offset)
532                              >> wq->wqe_shift);
533                 entry->wr_id = (*cur_qp)->wrid[wqe_index +
534                                                (*cur_qp)->rq.max];
535         } else if ((*cur_qp)->ibqp.srq) {
536                 struct mthca_srq *srq = to_msrq((*cur_qp)->ibqp.srq);
537                 u32 wqe = be32_to_cpu(cqe->wqe);
538                 wq = NULL;
539                 wqe_index = wqe >> srq->wqe_shift;
540                 entry->wr_id = srq->wrid[wqe_index];
541                 mthca_free_srq_wqe(srq, wqe);
542         } else {
543                 s32 wqe;
544                 wq = &(*cur_qp)->rq;
545                 wqe = be32_to_cpu(cqe->wqe);
546                 wqe_index = wqe >> wq->wqe_shift;
547                 /*
548                  * WQE addr == base - 1 might be reported in receive completion
549                  * with error instead of (rq size - 1) by Sinai FW 1.0.800 and
550                  * Arbel FW 5.1.400.  This bug should be fixed in later FW revs.
551                  */
552                 if (unlikely(wqe_index < 0))
553                         wqe_index = wq->max - 1;
554                 entry->wr_id = (*cur_qp)->wrid[wqe_index];
555         }
556
557         if (wq) {
558                 if (wq->last_comp < wqe_index)
559                         wq->tail += wqe_index - wq->last_comp;
560                 else
561                         wq->tail += wqe_index + wq->max - wq->last_comp;
562
563                 wq->last_comp = wqe_index;
564         }
565
566         if (is_error) {
567                 handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,
568                                  (struct mthca_err_cqe *) cqe,
569                                  entry, &free_cqe);
570                 goto out;
571         }
572
573         if (is_send) {
574                 entry->wc_flags = 0;
575                 switch (cqe->opcode) {
576                 case MTHCA_OPCODE_RDMA_WRITE:
577                         entry->opcode    = IB_WC_RDMA_WRITE;
578                         break;
579                 case MTHCA_OPCODE_RDMA_WRITE_IMM:
580                         entry->opcode    = IB_WC_RDMA_WRITE;
581                         entry->wc_flags |= IB_WC_WITH_IMM;
582                         break;
583                 case MTHCA_OPCODE_SEND:
584                         entry->opcode    = IB_WC_SEND;
585                         break;
586                 case MTHCA_OPCODE_SEND_IMM:
587                         entry->opcode    = IB_WC_SEND;
588                         entry->wc_flags |= IB_WC_WITH_IMM;
589                         break;
590                 case MTHCA_OPCODE_RDMA_READ:
591                         entry->opcode    = IB_WC_RDMA_READ;
592                         entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
593                         break;
594                 case MTHCA_OPCODE_ATOMIC_CS:
595                         entry->opcode    = IB_WC_COMP_SWAP;
596                         entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
597                         break;
598                 case MTHCA_OPCODE_ATOMIC_FA:
599                         entry->opcode    = IB_WC_FETCH_ADD;
600                         entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
601                         break;
602                 case MTHCA_OPCODE_BIND_MW:
603                         entry->opcode    = IB_WC_BIND_MW;
604                         break;
605                 default:
606                         entry->opcode    = MTHCA_OPCODE_INVALID;
607                         break;
608                 }
609         } else {
610                 entry->byte_len = be32_to_cpu(cqe->byte_cnt);
611                 switch (cqe->opcode & 0x1f) {
612                 case IB_OPCODE_SEND_LAST_WITH_IMMEDIATE:
613                 case IB_OPCODE_SEND_ONLY_WITH_IMMEDIATE:
614                         entry->wc_flags = IB_WC_WITH_IMM;
615                         entry->imm_data = cqe->imm_etype_pkey_eec;
616                         entry->opcode = IB_WC_RECV;
617                         break;
618                 case IB_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
619                 case IB_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
620                         entry->wc_flags = IB_WC_WITH_IMM;
621                         entry->imm_data = cqe->imm_etype_pkey_eec;
622                         entry->opcode = IB_WC_RECV_RDMA_WITH_IMM;
623                         break;
624                 default:
625                         entry->wc_flags = 0;
626                         entry->opcode = IB_WC_RECV;
627                         break;
628                 }
629                 entry->slid        = be16_to_cpu(cqe->rlid);
630                 entry->sl          = be16_to_cpu(cqe->sl_g_mlpath) >> 12;
631                 entry->src_qp      = be32_to_cpu(cqe->rqpn) & 0xffffff;
632                 entry->dlid_path_bits = be16_to_cpu(cqe->sl_g_mlpath) & 0x7f;
633                 entry->pkey_index  = be32_to_cpu(cqe->imm_etype_pkey_eec) >> 16;
634                 entry->wc_flags   |= be16_to_cpu(cqe->sl_g_mlpath) & 0x80 ?
635                                         IB_WC_GRH : 0;
636         }
637
638         entry->status = IB_WC_SUCCESS;
639
640  out:
641         if (likely(free_cqe)) {
642                 set_cqe_hw(cqe);
643                 ++(*freed);
644                 ++cq->cons_index;
645         }
646
647         return err;
648 }
649
650 int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
651                   struct ib_wc *entry)
652 {
653         struct mthca_dev *dev = to_mdev(ibcq->device);
654         struct mthca_cq *cq = to_mcq(ibcq);
655         struct mthca_qp *qp = NULL;
656         unsigned long flags;
657         int err = 0;
658         int freed = 0;
659         int npolled;
660
661         spin_lock_irqsave(&cq->lock, flags);
662
663         npolled = 0;
664 repoll:
665         while (npolled < num_entries) {
666                 err = mthca_poll_one(dev, cq, &qp,
667                                      &freed, entry + npolled);
668                 if (err)
669                         break;
670                 ++npolled;
671         }
672
673         if (freed) {
674                 wmb();
675                 update_cons_index(dev, cq, freed);
676         }
677
678         /*
679          * If a CQ resize is in progress and we discovered that the
680          * old buffer is empty, then peek in the new buffer, and if
681          * it's not empty, switch to the new buffer and continue
682          * polling there.
683          */
684         if (unlikely(err == -EAGAIN && cq->resize_buf &&
685                      cq->resize_buf->state == CQ_RESIZE_READY)) {
686                 /*
687                  * In Tavor mode, the hardware keeps the producer
688                  * index modulo the CQ size.  Since we might be making
689                  * the CQ bigger, we need to mask our consumer index
690                  * using the size of the old CQ buffer before looking
691                  * in the new CQ buffer.
692                  */
693                 if (!mthca_is_memfree(dev))
694                         cq->cons_index &= cq->ibcq.cqe;
695
696                 if (cqe_sw(get_cqe_from_buf(&cq->resize_buf->buf,
697                                             cq->cons_index & cq->resize_buf->cqe))) {
698                         struct mthca_cq_buf tbuf;
699                         int tcqe;
700
701                         tbuf         = cq->buf;
702                         tcqe         = cq->ibcq.cqe;
703                         cq->buf      = cq->resize_buf->buf;
704                         cq->ibcq.cqe = cq->resize_buf->cqe;
705
706                         cq->resize_buf->buf   = tbuf;
707                         cq->resize_buf->cqe   = tcqe;
708                         cq->resize_buf->state = CQ_RESIZE_SWAPPED;
709
710                         goto repoll;
711                 }
712         }
713
714         spin_unlock_irqrestore(&cq->lock, flags);
715
716         return err == 0 || err == -EAGAIN ? npolled : err;
717 }
718
719 int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify notify)
720 {
721         __be32 doorbell[2];
722
723         doorbell[0] = cpu_to_be32((notify == IB_CQ_SOLICITED ?
724                                    MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL :
725                                    MTHCA_TAVOR_CQ_DB_REQ_NOT)      |
726                                   to_mcq(cq)->cqn);
727         doorbell[1] = (__force __be32) 0xffffffff;
728
729         mthca_write64(doorbell,
730                       to_mdev(cq->device)->kar + MTHCA_CQ_DOORBELL,
731                       MTHCA_GET_DOORBELL_LOCK(&to_mdev(cq->device)->doorbell_lock));
732
733         return 0;
734 }
735
736 int mthca_arbel_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify notify)
737 {
738         struct mthca_cq *cq = to_mcq(ibcq);
739         __be32 doorbell[2];
740         u32 sn;
741         __be32 ci;
742
743         sn = cq->arm_sn & 3;
744         ci = cpu_to_be32(cq->cons_index);
745
746         doorbell[0] = ci;
747         doorbell[1] = cpu_to_be32((cq->cqn << 8) | (2 << 5) | (sn << 3) |
748                                   (notify == IB_CQ_SOLICITED ? 1 : 2));
749
750         mthca_write_db_rec(doorbell, cq->arm_db);
751
752         /*
753          * Make sure that the doorbell record in host memory is
754          * written before ringing the doorbell via PCI MMIO.
755          */
756         wmb();
757
758         doorbell[0] = cpu_to_be32((sn << 28)                       |
759                                   (notify == IB_CQ_SOLICITED ?
760                                    MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL :
761                                    MTHCA_ARBEL_CQ_DB_REQ_NOT)      |
762                                   cq->cqn);
763         doorbell[1] = ci;
764
765         mthca_write64(doorbell,
766                       to_mdev(ibcq->device)->kar + MTHCA_CQ_DOORBELL,
767                       MTHCA_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->doorbell_lock));
768
769         return 0;
770 }
771
772 int mthca_init_cq(struct mthca_dev *dev, int nent,
773                   struct mthca_ucontext *ctx, u32 pdn,
774                   struct mthca_cq *cq)
775 {
776         struct mthca_mailbox *mailbox;
777         struct mthca_cq_context *cq_context;
778         int err = -ENOMEM;
779         u8 status;
780
781         cq->ibcq.cqe  = nent - 1;
782         cq->is_kernel = !ctx;
783
784         cq->cqn = mthca_alloc(&dev->cq_table.alloc);
785         if (cq->cqn == -1)
786                 return -ENOMEM;
787
788         if (mthca_is_memfree(dev)) {
789                 err = mthca_table_get(dev, dev->cq_table.table, cq->cqn);
790                 if (err)
791                         goto err_out;
792
793                 if (cq->is_kernel) {
794                         cq->arm_sn = 1;
795
796                         err = -ENOMEM;
797
798                         cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
799                                                              cq->cqn, &cq->set_ci_db);
800                         if (cq->set_ci_db_index < 0)
801                                 goto err_out_icm;
802
803                         cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
804                                                           cq->cqn, &cq->arm_db);
805                         if (cq->arm_db_index < 0)
806                                 goto err_out_ci;
807                 }
808         }
809
810         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
811         if (IS_ERR(mailbox))
812                 goto err_out_arm;
813
814         cq_context = mailbox->buf;
815
816         if (cq->is_kernel) {
817                 err = mthca_alloc_cq_buf(dev, &cq->buf, nent);
818                 if (err)
819                         goto err_out_mailbox;
820         }
821
822         spin_lock_init(&cq->lock);
823         cq->refcount = 1;
824         init_waitqueue_head(&cq->wait);
825         mutex_init(&cq->mutex);
826
827         memset(cq_context, 0, sizeof *cq_context);
828         cq_context->flags           = cpu_to_be32(MTHCA_CQ_STATUS_OK      |
829                                                   MTHCA_CQ_STATE_DISARMED |
830                                                   MTHCA_CQ_FLAG_TR);
831         cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24);
832         if (ctx)
833                 cq_context->logsize_usrpage |= cpu_to_be32(ctx->uar.index);
834         else
835                 cq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
836         cq_context->error_eqn       = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
837         cq_context->comp_eqn        = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
838         cq_context->pd              = cpu_to_be32(pdn);
839         cq_context->lkey            = cpu_to_be32(cq->buf.mr.ibmr.lkey);
840         cq_context->cqn             = cpu_to_be32(cq->cqn);
841
842         if (mthca_is_memfree(dev)) {
843                 cq_context->ci_db    = cpu_to_be32(cq->set_ci_db_index);
844                 cq_context->state_db = cpu_to_be32(cq->arm_db_index);
845         }
846
847         err = mthca_SW2HW_CQ(dev, mailbox, cq->cqn, &status);
848         if (err) {
849                 mthca_warn(dev, "SW2HW_CQ failed (%d)\n", err);
850                 goto err_out_free_mr;
851         }
852
853         if (status) {
854                 mthca_warn(dev, "SW2HW_CQ returned status 0x%02x\n",
855                            status);
856                 err = -EINVAL;
857                 goto err_out_free_mr;
858         }
859
860         spin_lock_irq(&dev->cq_table.lock);
861         if (mthca_array_set(&dev->cq_table.cq,
862                             cq->cqn & (dev->limits.num_cqs - 1),
863                             cq)) {
864                 spin_unlock_irq(&dev->cq_table.lock);
865                 goto err_out_free_mr;
866         }
867         spin_unlock_irq(&dev->cq_table.lock);
868
869         cq->cons_index = 0;
870
871         mthca_free_mailbox(dev, mailbox);
872
873         return 0;
874
875 err_out_free_mr:
876         if (cq->is_kernel)
877                 mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
878
879 err_out_mailbox:
880         mthca_free_mailbox(dev, mailbox);
881
882 err_out_arm:
883         if (cq->is_kernel && mthca_is_memfree(dev))
884                 mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
885
886 err_out_ci:
887         if (cq->is_kernel && mthca_is_memfree(dev))
888                 mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
889
890 err_out_icm:
891         mthca_table_put(dev, dev->cq_table.table, cq->cqn);
892
893 err_out:
894         mthca_free(&dev->cq_table.alloc, cq->cqn);
895
896         return err;
897 }
898
899 static inline int get_cq_refcount(struct mthca_dev *dev, struct mthca_cq *cq)
900 {
901         int c;
902
903         spin_lock_irq(&dev->cq_table.lock);
904         c = cq->refcount;
905         spin_unlock_irq(&dev->cq_table.lock);
906
907         return c;
908 }
909
910 void mthca_free_cq(struct mthca_dev *dev,
911                    struct mthca_cq *cq)
912 {
913         struct mthca_mailbox *mailbox;
914         int err;
915         u8 status;
916
917         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
918         if (IS_ERR(mailbox)) {
919                 mthca_warn(dev, "No memory for mailbox to free CQ.\n");
920                 return;
921         }
922
923         err = mthca_HW2SW_CQ(dev, mailbox, cq->cqn, &status);
924         if (err)
925                 mthca_warn(dev, "HW2SW_CQ failed (%d)\n", err);
926         else if (status)
927                 mthca_warn(dev, "HW2SW_CQ returned status 0x%02x\n", status);
928
929         if (0) {
930                 __be32 *ctx = mailbox->buf;
931                 int j;
932
933                 printk(KERN_ERR "context for CQN %x (cons index %x, next sw %d)\n",
934                        cq->cqn, cq->cons_index,
935                        cq->is_kernel ? !!next_cqe_sw(cq) : 0);
936                 for (j = 0; j < 16; ++j)
937                         printk(KERN_ERR "[%2x] %08x\n", j * 4, be32_to_cpu(ctx[j]));
938         }
939
940         spin_lock_irq(&dev->cq_table.lock);
941         mthca_array_clear(&dev->cq_table.cq,
942                           cq->cqn & (dev->limits.num_cqs - 1));
943         --cq->refcount;
944         spin_unlock_irq(&dev->cq_table.lock);
945
946         if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
947                 synchronize_irq(dev->eq_table.eq[MTHCA_EQ_COMP].msi_x_vector);
948         else
949                 synchronize_irq(dev->pdev->irq);
950
951         wait_event(cq->wait, !get_cq_refcount(dev, cq));
952
953         if (cq->is_kernel) {
954                 mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
955                 if (mthca_is_memfree(dev)) {
956                         mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM,    cq->arm_db_index);
957                         mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
958                 }
959         }
960
961         mthca_table_put(dev, dev->cq_table.table, cq->cqn);
962         mthca_free(&dev->cq_table.alloc, cq->cqn);
963         mthca_free_mailbox(dev, mailbox);
964 }
965
966 int __devinit mthca_init_cq_table(struct mthca_dev *dev)
967 {
968         int err;
969
970         spin_lock_init(&dev->cq_table.lock);
971
972         err = mthca_alloc_init(&dev->cq_table.alloc,
973                                dev->limits.num_cqs,
974                                (1 << 24) - 1,
975                                dev->limits.reserved_cqs);
976         if (err)
977                 return err;
978
979         err = mthca_array_init(&dev->cq_table.cq,
980                                dev->limits.num_cqs);
981         if (err)
982                 mthca_alloc_cleanup(&dev->cq_table.alloc);
983
984         return err;
985 }
986
987 void mthca_cleanup_cq_table(struct mthca_dev *dev)
988 {
989         mthca_array_cleanup(&dev->cq_table.cq, dev->limits.num_cqs);
990         mthca_alloc_cleanup(&dev->cq_table.alloc);
991 }