]> bbs.cooldavid.org Git - net-next-2.6.git/blob - drivers/infiniband/hw/mthca/mthca_qp.c
c2d3300dace9798984e1d44a95d6ee324aefc035
[net-next-2.6.git] / drivers / infiniband / hw / mthca / mthca_qp.c
1 /*
2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Cisco Systems. All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
6  *
7  * This software is available to you under a choice of one of two
8  * licenses.  You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the
11  * OpenIB.org BSD license below:
12  *
13  *     Redistribution and use in source and binary forms, with or
14  *     without modification, are permitted provided that the following
15  *     conditions are met:
16  *
17  *      - Redistributions of source code must retain the above
18  *        copyright notice, this list of conditions and the following
19  *        disclaimer.
20  *
21  *      - Redistributions in binary form must reproduce the above
22  *        copyright notice, this list of conditions and the following
23  *        disclaimer in the documentation and/or other materials
24  *        provided with the distribution.
25  *
26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33  * SOFTWARE.
34  *
35  * $Id: mthca_qp.c 1355 2004-12-17 15:23:43Z roland $
36  */
37
38 #include <linux/init.h>
39 #include <linux/string.h>
40 #include <linux/slab.h>
41
42 #include <rdma/ib_verbs.h>
43 #include <rdma/ib_cache.h>
44 #include <rdma/ib_pack.h>
45
46 #include "mthca_dev.h"
47 #include "mthca_cmd.h"
48 #include "mthca_memfree.h"
49 #include "mthca_wqe.h"
50
51 enum {
52         MTHCA_MAX_DIRECT_QP_SIZE = 4 * PAGE_SIZE,
53         MTHCA_ACK_REQ_FREQ       = 10,
54         MTHCA_FLIGHT_LIMIT       = 9,
55         MTHCA_UD_HEADER_SIZE     = 72, /* largest UD header possible */
56         MTHCA_INLINE_HEADER_SIZE = 4,  /* data segment overhead for inline */
57         MTHCA_INLINE_CHUNK_SIZE  = 16  /* inline data segment chunk */
58 };
59
60 enum {
61         MTHCA_QP_STATE_RST  = 0,
62         MTHCA_QP_STATE_INIT = 1,
63         MTHCA_QP_STATE_RTR  = 2,
64         MTHCA_QP_STATE_RTS  = 3,
65         MTHCA_QP_STATE_SQE  = 4,
66         MTHCA_QP_STATE_SQD  = 5,
67         MTHCA_QP_STATE_ERR  = 6,
68         MTHCA_QP_STATE_DRAINING = 7
69 };
70
71 enum {
72         MTHCA_QP_ST_RC  = 0x0,
73         MTHCA_QP_ST_UC  = 0x1,
74         MTHCA_QP_ST_RD  = 0x2,
75         MTHCA_QP_ST_UD  = 0x3,
76         MTHCA_QP_ST_MLX = 0x7
77 };
78
79 enum {
80         MTHCA_QP_PM_MIGRATED = 0x3,
81         MTHCA_QP_PM_ARMED    = 0x0,
82         MTHCA_QP_PM_REARM    = 0x1
83 };
84
85 enum {
86         /* qp_context flags */
87         MTHCA_QP_BIT_DE  = 1 <<  8,
88         /* params1 */
89         MTHCA_QP_BIT_SRE = 1 << 15,
90         MTHCA_QP_BIT_SWE = 1 << 14,
91         MTHCA_QP_BIT_SAE = 1 << 13,
92         MTHCA_QP_BIT_SIC = 1 <<  4,
93         MTHCA_QP_BIT_SSC = 1 <<  3,
94         /* params2 */
95         MTHCA_QP_BIT_RRE = 1 << 15,
96         MTHCA_QP_BIT_RWE = 1 << 14,
97         MTHCA_QP_BIT_RAE = 1 << 13,
98         MTHCA_QP_BIT_RIC = 1 <<  4,
99         MTHCA_QP_BIT_RSC = 1 <<  3
100 };
101
102 struct mthca_qp_path {
103         __be32 port_pkey;
104         u8     rnr_retry;
105         u8     g_mylmc;
106         __be16 rlid;
107         u8     ackto;
108         u8     mgid_index;
109         u8     static_rate;
110         u8     hop_limit;
111         __be32 sl_tclass_flowlabel;
112         u8     rgid[16];
113 } __attribute__((packed));
114
115 struct mthca_qp_context {
116         __be32 flags;
117         __be32 tavor_sched_queue; /* Reserved on Arbel */
118         u8     mtu_msgmax;
119         u8     rq_size_stride;  /* Reserved on Tavor */
120         u8     sq_size_stride;  /* Reserved on Tavor */
121         u8     rlkey_arbel_sched_queue; /* Reserved on Tavor */
122         __be32 usr_page;
123         __be32 local_qpn;
124         __be32 remote_qpn;
125         u32    reserved1[2];
126         struct mthca_qp_path pri_path;
127         struct mthca_qp_path alt_path;
128         __be32 rdd;
129         __be32 pd;
130         __be32 wqe_base;
131         __be32 wqe_lkey;
132         __be32 params1;
133         __be32 reserved2;
134         __be32 next_send_psn;
135         __be32 cqn_snd;
136         __be32 snd_wqe_base_l;  /* Next send WQE on Tavor */
137         __be32 snd_db_index;    /* (debugging only entries) */
138         __be32 last_acked_psn;
139         __be32 ssn;
140         __be32 params2;
141         __be32 rnr_nextrecvpsn;
142         __be32 ra_buff_indx;
143         __be32 cqn_rcv;
144         __be32 rcv_wqe_base_l;  /* Next recv WQE on Tavor */
145         __be32 rcv_db_index;    /* (debugging only entries) */
146         __be32 qkey;
147         __be32 srqn;
148         __be32 rmsn;
149         __be16 rq_wqe_counter;  /* reserved on Tavor */
150         __be16 sq_wqe_counter;  /* reserved on Tavor */
151         u32    reserved3[18];
152 } __attribute__((packed));
153
154 struct mthca_qp_param {
155         __be32 opt_param_mask;
156         u32    reserved1;
157         struct mthca_qp_context context;
158         u32    reserved2[62];
159 } __attribute__((packed));
160
161 enum {
162         MTHCA_QP_OPTPAR_ALT_ADDR_PATH     = 1 << 0,
163         MTHCA_QP_OPTPAR_RRE               = 1 << 1,
164         MTHCA_QP_OPTPAR_RAE               = 1 << 2,
165         MTHCA_QP_OPTPAR_RWE               = 1 << 3,
166         MTHCA_QP_OPTPAR_PKEY_INDEX        = 1 << 4,
167         MTHCA_QP_OPTPAR_Q_KEY             = 1 << 5,
168         MTHCA_QP_OPTPAR_RNR_TIMEOUT       = 1 << 6,
169         MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7,
170         MTHCA_QP_OPTPAR_SRA_MAX           = 1 << 8,
171         MTHCA_QP_OPTPAR_RRA_MAX           = 1 << 9,
172         MTHCA_QP_OPTPAR_PM_STATE          = 1 << 10,
173         MTHCA_QP_OPTPAR_PORT_NUM          = 1 << 11,
174         MTHCA_QP_OPTPAR_RETRY_COUNT       = 1 << 12,
175         MTHCA_QP_OPTPAR_ALT_RNR_RETRY     = 1 << 13,
176         MTHCA_QP_OPTPAR_ACK_TIMEOUT       = 1 << 14,
177         MTHCA_QP_OPTPAR_RNR_RETRY         = 1 << 15,
178         MTHCA_QP_OPTPAR_SCHED_QUEUE       = 1 << 16
179 };
180
181 static const u8 mthca_opcode[] = {
182         [IB_WR_SEND]                 = MTHCA_OPCODE_SEND,
183         [IB_WR_SEND_WITH_IMM]        = MTHCA_OPCODE_SEND_IMM,
184         [IB_WR_RDMA_WRITE]           = MTHCA_OPCODE_RDMA_WRITE,
185         [IB_WR_RDMA_WRITE_WITH_IMM]  = MTHCA_OPCODE_RDMA_WRITE_IMM,
186         [IB_WR_RDMA_READ]            = MTHCA_OPCODE_RDMA_READ,
187         [IB_WR_ATOMIC_CMP_AND_SWP]   = MTHCA_OPCODE_ATOMIC_CS,
188         [IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
189 };
190
191 static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
192 {
193         return qp->qpn >= dev->qp_table.sqp_start &&
194                 qp->qpn <= dev->qp_table.sqp_start + 3;
195 }
196
197 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
198 {
199         return qp->qpn >= dev->qp_table.sqp_start &&
200                 qp->qpn <= dev->qp_table.sqp_start + 1;
201 }
202
203 static void *get_recv_wqe(struct mthca_qp *qp, int n)
204 {
205         if (qp->is_direct)
206                 return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
207         else
208                 return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
209                         ((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
210 }
211
212 static void *get_send_wqe(struct mthca_qp *qp, int n)
213 {
214         if (qp->is_direct)
215                 return qp->queue.direct.buf + qp->send_wqe_offset +
216                         (n << qp->sq.wqe_shift);
217         else
218                 return qp->queue.page_list[(qp->send_wqe_offset +
219                                             (n << qp->sq.wqe_shift)) >>
220                                            PAGE_SHIFT].buf +
221                         ((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
222                          (PAGE_SIZE - 1));
223 }
224
225 static void mthca_wq_init(struct mthca_wq *wq)
226 {
227         spin_lock_init(&wq->lock);
228         wq->next_ind  = 0;
229         wq->last_comp = wq->max - 1;
230         wq->head      = 0;
231         wq->tail      = 0;
232 }
233
234 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
235                     enum ib_event_type event_type)
236 {
237         struct mthca_qp *qp;
238         struct ib_event event;
239
240         spin_lock(&dev->qp_table.lock);
241         qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
242         if (qp)
243                 atomic_inc(&qp->refcount);
244         spin_unlock(&dev->qp_table.lock);
245
246         if (!qp) {
247                 mthca_warn(dev, "Async event for bogus QP %08x\n", qpn);
248                 return;
249         }
250
251         event.device      = &dev->ib_dev;
252         event.event       = event_type;
253         event.element.qp  = &qp->ibqp;
254         if (qp->ibqp.event_handler)
255                 qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
256
257         if (atomic_dec_and_test(&qp->refcount))
258                 wake_up(&qp->wait);
259 }
260
261 static int to_mthca_state(enum ib_qp_state ib_state)
262 {
263         switch (ib_state) {
264         case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
265         case IB_QPS_INIT:  return MTHCA_QP_STATE_INIT;
266         case IB_QPS_RTR:   return MTHCA_QP_STATE_RTR;
267         case IB_QPS_RTS:   return MTHCA_QP_STATE_RTS;
268         case IB_QPS_SQD:   return MTHCA_QP_STATE_SQD;
269         case IB_QPS_SQE:   return MTHCA_QP_STATE_SQE;
270         case IB_QPS_ERR:   return MTHCA_QP_STATE_ERR;
271         default:                return -1;
272         }
273 }
274
275 enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
276
277 static int to_mthca_st(int transport)
278 {
279         switch (transport) {
280         case RC:  return MTHCA_QP_ST_RC;
281         case UC:  return MTHCA_QP_ST_UC;
282         case UD:  return MTHCA_QP_ST_UD;
283         case RD:  return MTHCA_QP_ST_RD;
284         case MLX: return MTHCA_QP_ST_MLX;
285         default:  return -1;
286         }
287 }
288
289 static void store_attrs(struct mthca_sqp *sqp, struct ib_qp_attr *attr,
290                         int attr_mask)
291 {
292         if (attr_mask & IB_QP_PKEY_INDEX)
293                 sqp->pkey_index = attr->pkey_index;
294         if (attr_mask & IB_QP_QKEY)
295                 sqp->qkey = attr->qkey;
296         if (attr_mask & IB_QP_SQ_PSN)
297                 sqp->send_psn = attr->sq_psn;
298 }
299
300 static void init_port(struct mthca_dev *dev, int port)
301 {
302         int err;
303         u8 status;
304         struct mthca_init_ib_param param;
305
306         memset(&param, 0, sizeof param);
307
308         param.port_width = dev->limits.port_width_cap;
309         param.vl_cap     = dev->limits.vl_cap;
310         param.mtu_cap    = dev->limits.mtu_cap;
311         param.gid_cap    = dev->limits.gid_table_len;
312         param.pkey_cap   = dev->limits.pkey_table_len;
313
314         err = mthca_INIT_IB(dev, &param, port, &status);
315         if (err)
316                 mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
317         if (status)
318                 mthca_warn(dev, "INIT_IB returned status %02x.\n", status);
319 }
320
321 static __be32 get_hw_access_flags(struct mthca_qp *qp, struct ib_qp_attr *attr,
322                                   int attr_mask)
323 {
324         u8 dest_rd_atomic;
325         u32 access_flags;
326         u32 hw_access_flags = 0;
327
328         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
329                 dest_rd_atomic = attr->max_dest_rd_atomic;
330         else
331                 dest_rd_atomic = qp->resp_depth;
332
333         if (attr_mask & IB_QP_ACCESS_FLAGS)
334                 access_flags = attr->qp_access_flags;
335         else
336                 access_flags = qp->atomic_rd_en;
337
338         if (!dest_rd_atomic)
339                 access_flags &= IB_ACCESS_REMOTE_WRITE;
340
341         if (access_flags & IB_ACCESS_REMOTE_READ)
342                 hw_access_flags |= MTHCA_QP_BIT_RRE;
343         if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
344                 hw_access_flags |= MTHCA_QP_BIT_RAE;
345         if (access_flags & IB_ACCESS_REMOTE_WRITE)
346                 hw_access_flags |= MTHCA_QP_BIT_RWE;
347
348         return cpu_to_be32(hw_access_flags);
349 }
350
351 static void mthca_path_set(struct ib_ah_attr *ah, struct mthca_qp_path *path)
352 {
353         path->g_mylmc     = ah->src_path_bits & 0x7f;
354         path->rlid        = cpu_to_be16(ah->dlid);
355         path->static_rate = !!ah->static_rate;
356
357         if (ah->ah_flags & IB_AH_GRH) {
358                 path->g_mylmc   |= 1 << 7;
359                 path->mgid_index = ah->grh.sgid_index;
360                 path->hop_limit  = ah->grh.hop_limit;
361                 path->sl_tclass_flowlabel =
362                         cpu_to_be32((ah->sl << 28)                |
363                                     (ah->grh.traffic_class << 20) |
364                                     (ah->grh.flow_label));
365                 memcpy(path->rgid, ah->grh.dgid.raw, 16);
366         } else
367                 path->sl_tclass_flowlabel = cpu_to_be32(ah->sl << 28);
368 }
369
370 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
371 {
372         struct mthca_dev *dev = to_mdev(ibqp->device);
373         struct mthca_qp *qp = to_mqp(ibqp);
374         enum ib_qp_state cur_state, new_state;
375         struct mthca_mailbox *mailbox;
376         struct mthca_qp_param *qp_param;
377         struct mthca_qp_context *qp_context;
378         u32 sqd_event = 0;
379         u8 status;
380         int err;
381
382         if (attr_mask & IB_QP_CUR_STATE) {
383                 cur_state = attr->cur_qp_state;
384         } else {
385                 spin_lock_irq(&qp->sq.lock);
386                 spin_lock(&qp->rq.lock);
387                 cur_state = qp->state;
388                 spin_unlock(&qp->rq.lock);
389                 spin_unlock_irq(&qp->sq.lock);
390         }
391
392         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
393
394         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask)) {
395                 mthca_dbg(dev, "Bad QP transition (transport %d) "
396                           "%d->%d with attr 0x%08x\n",
397                           qp->transport, cur_state, new_state,
398                           attr_mask);
399                 return -EINVAL;
400         }
401
402         if ((attr_mask & IB_QP_PKEY_INDEX) &&
403              attr->pkey_index >= dev->limits.pkey_table_len) {
404                 mthca_dbg(dev, "PKey index (%u) too large. max is %d\n",
405                           attr->pkey_index,dev->limits.pkey_table_len-1);
406                 return -EINVAL;
407         }
408
409         if ((attr_mask & IB_QP_PORT) &&
410             (attr->port_num == 0 || attr->port_num > dev->limits.num_ports)) {
411                 mthca_dbg(dev, "Port number (%u) is invalid\n", attr->port_num);
412                 return -EINVAL;
413         }
414
415         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
416             attr->max_rd_atomic > dev->limits.max_qp_init_rdma) {
417                 mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n",
418                           attr->max_rd_atomic, dev->limits.max_qp_init_rdma);
419                 return -EINVAL;
420         }
421
422         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
423             attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) {
424                 mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n",
425                           attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift);
426                 return -EINVAL;
427         }
428
429         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
430         if (IS_ERR(mailbox))
431                 return PTR_ERR(mailbox);
432         qp_param = mailbox->buf;
433         qp_context = &qp_param->context;
434         memset(qp_param, 0, sizeof *qp_param);
435
436         qp_context->flags      = cpu_to_be32((to_mthca_state(new_state) << 28) |
437                                              (to_mthca_st(qp->transport) << 16));
438         qp_context->flags     |= cpu_to_be32(MTHCA_QP_BIT_DE);
439         if (!(attr_mask & IB_QP_PATH_MIG_STATE))
440                 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
441         else {
442                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
443                 switch (attr->path_mig_state) {
444                 case IB_MIG_MIGRATED:
445                         qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
446                         break;
447                 case IB_MIG_REARM:
448                         qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
449                         break;
450                 case IB_MIG_ARMED:
451                         qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
452                         break;
453                 }
454         }
455
456         /* leave tavor_sched_queue as 0 */
457
458         if (qp->transport == MLX || qp->transport == UD)
459                 qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
460         else if (attr_mask & IB_QP_PATH_MTU)
461                 qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
462
463         if (mthca_is_memfree(dev)) {
464                 if (qp->rq.max)
465                         qp_context->rq_size_stride = long_log2(qp->rq.max) << 3;
466                 qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
467
468                 if (qp->sq.max)
469                         qp_context->sq_size_stride = long_log2(qp->sq.max) << 3;
470                 qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
471         }
472
473         /* leave arbel_sched_queue as 0 */
474
475         if (qp->ibqp.uobject)
476                 qp_context->usr_page =
477                         cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index);
478         else
479                 qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
480         qp_context->local_qpn  = cpu_to_be32(qp->qpn);
481         if (attr_mask & IB_QP_DEST_QPN) {
482                 qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
483         }
484
485         if (qp->transport == MLX)
486                 qp_context->pri_path.port_pkey |=
487                         cpu_to_be32(to_msqp(qp)->port << 24);
488         else {
489                 if (attr_mask & IB_QP_PORT) {
490                         qp_context->pri_path.port_pkey |=
491                                 cpu_to_be32(attr->port_num << 24);
492                         qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
493                 }
494         }
495
496         if (attr_mask & IB_QP_PKEY_INDEX) {
497                 qp_context->pri_path.port_pkey |=
498                         cpu_to_be32(attr->pkey_index);
499                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
500         }
501
502         if (attr_mask & IB_QP_RNR_RETRY) {
503                 qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry =
504                         attr->rnr_retry << 5;
505                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY |
506                                                         MTHCA_QP_OPTPAR_ALT_RNR_RETRY);
507         }
508
509         if (attr_mask & IB_QP_AV) {
510                 mthca_path_set(&attr->ah_attr, &qp_context->pri_path);
511                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
512         }
513
514         if (attr_mask & IB_QP_TIMEOUT) {
515                 qp_context->pri_path.ackto = attr->timeout << 3;
516                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
517         }
518
519         if (attr_mask & IB_QP_ALT_PATH) {
520                 if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) {
521                         mthca_dbg(dev, "Alternate port number (%u) is invalid\n",
522                                 attr->alt_port_num);
523                         return -EINVAL;
524                 }
525
526                 mthca_path_set(&attr->alt_ah_attr, &qp_context->alt_path);
527                 qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index |
528                                                               attr->alt_port_num << 24);
529                 qp_context->alt_path.ackto = attr->alt_timeout << 3;
530                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH);
531         }
532
533         /* leave rdd as 0 */
534         qp_context->pd         = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
535         /* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
536         qp_context->wqe_lkey   = cpu_to_be32(qp->mr.ibmr.lkey);
537         qp_context->params1    = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
538                                              (MTHCA_FLIGHT_LIMIT << 24) |
539                                              MTHCA_QP_BIT_SWE);
540         if (qp->sq_policy == IB_SIGNAL_ALL_WR)
541                 qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
542         if (attr_mask & IB_QP_RETRY_CNT) {
543                 qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
544                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
545         }
546
547         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
548                 if (attr->max_rd_atomic) {
549                         qp_context->params1 |=
550                                 cpu_to_be32(MTHCA_QP_BIT_SRE |
551                                             MTHCA_QP_BIT_SAE);
552                         qp_context->params1 |=
553                                 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
554                 }
555                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
556         }
557
558         if (attr_mask & IB_QP_SQ_PSN)
559                 qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
560         qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
561
562         if (mthca_is_memfree(dev)) {
563                 qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
564                 qp_context->snd_db_index   = cpu_to_be32(qp->sq.db_index);
565         }
566
567         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
568                 if (attr->max_dest_rd_atomic)
569                         qp_context->params2 |=
570                                 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
571
572                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
573         }
574
575         if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
576                 qp_context->params2      |= get_hw_access_flags(qp, attr, attr_mask);
577                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
578                                                         MTHCA_QP_OPTPAR_RRE |
579                                                         MTHCA_QP_OPTPAR_RAE);
580         }
581
582         qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
583
584         if (ibqp->srq)
585                 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
586
587         if (attr_mask & IB_QP_MIN_RNR_TIMER) {
588                 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
589                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
590         }
591         if (attr_mask & IB_QP_RQ_PSN)
592                 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
593
594         qp_context->ra_buff_indx =
595                 cpu_to_be32(dev->qp_table.rdb_base +
596                             ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
597                              dev->qp_table.rdb_shift));
598
599         qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
600
601         if (mthca_is_memfree(dev))
602                 qp_context->rcv_db_index   = cpu_to_be32(qp->rq.db_index);
603
604         if (attr_mask & IB_QP_QKEY) {
605                 qp_context->qkey = cpu_to_be32(attr->qkey);
606                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
607         }
608
609         if (ibqp->srq)
610                 qp_context->srqn = cpu_to_be32(1 << 24 |
611                                                to_msrq(ibqp->srq)->srqn);
612
613         if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD  &&
614             attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY               &&
615             attr->en_sqd_async_notify)
616                 sqd_event = 1 << 31;
617
618         err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0,
619                               mailbox, sqd_event, &status);
620         if (status) {
621                 mthca_warn(dev, "modify QP %d->%d returned status %02x.\n",
622                            cur_state, new_state, status);
623                 err = -EINVAL;
624         }
625
626         if (!err) {
627                 qp->state = new_state;
628                 if (attr_mask & IB_QP_ACCESS_FLAGS)
629                         qp->atomic_rd_en = attr->qp_access_flags;
630                 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
631                         qp->resp_depth = attr->max_dest_rd_atomic;
632         }
633
634         mthca_free_mailbox(dev, mailbox);
635
636         if (is_sqp(dev, qp))
637                 store_attrs(to_msqp(qp), attr, attr_mask);
638
639         /*
640          * If we moved QP0 to RTR, bring the IB link up; if we moved
641          * QP0 to RESET or ERROR, bring the link back down.
642          */
643         if (is_qp0(dev, qp)) {
644                 if (cur_state != IB_QPS_RTR &&
645                     new_state == IB_QPS_RTR)
646                         init_port(dev, to_msqp(qp)->port);
647
648                 if (cur_state != IB_QPS_RESET &&
649                     cur_state != IB_QPS_ERR &&
650                     (new_state == IB_QPS_RESET ||
651                      new_state == IB_QPS_ERR))
652                         mthca_CLOSE_IB(dev, to_msqp(qp)->port, &status);
653         }
654
655         /*
656          * If we moved a kernel QP to RESET, clean up all old CQ
657          * entries and reinitialize the QP.
658          */
659         if (!err && new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
660                 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
661                                qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
662                 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
663                         mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
664                                        qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
665
666                 mthca_wq_init(&qp->sq);
667                 qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
668
669                 mthca_wq_init(&qp->rq);
670                 qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
671
672                 if (mthca_is_memfree(dev)) {
673                         *qp->sq.db = 0;
674                         *qp->rq.db = 0;
675                 }
676         }
677
678         return err;
679 }
680
681 static int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz)
682 {
683         /*
684          * Calculate the maximum size of WQE s/g segments, excluding
685          * the next segment and other non-data segments.
686          */
687         int max_data_size = desc_sz - sizeof (struct mthca_next_seg);
688
689         switch (qp->transport) {
690         case MLX:
691                 max_data_size -= 2 * sizeof (struct mthca_data_seg);
692                 break;
693
694         case UD:
695                 if (mthca_is_memfree(dev))
696                         max_data_size -= sizeof (struct mthca_arbel_ud_seg);
697                 else
698                         max_data_size -= sizeof (struct mthca_tavor_ud_seg);
699                 break;
700
701         default:
702                 max_data_size -= sizeof (struct mthca_raddr_seg);
703                 break;
704         }
705
706         return max_data_size;
707 }
708
709 static inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size)
710 {
711         /* We don't support inline data for kernel QPs (yet). */
712         return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0;
713 }
714
715 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
716                                  struct mthca_pd *pd,
717                                  struct mthca_qp *qp)
718 {
719         int max_data_size = mthca_max_data_size(dev, qp,
720                                                 min(dev->limits.max_desc_sz,
721                                                     1 << qp->sq.wqe_shift));
722
723         qp->max_inline_data = mthca_max_inline_data(pd, max_data_size);
724
725         qp->sq.max_gs = min_t(int, dev->limits.max_sg,
726                               max_data_size / sizeof (struct mthca_data_seg));
727         qp->rq.max_gs = min_t(int, dev->limits.max_sg,
728                                (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
729                                 sizeof (struct mthca_next_seg)) /
730                                sizeof (struct mthca_data_seg));
731 }
732
733 /*
734  * Allocate and register buffer for WQEs.  qp->rq.max, sq.max,
735  * rq.max_gs and sq.max_gs must all be assigned.
736  * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
737  * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
738  * queue)
739  */
740 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
741                                struct mthca_pd *pd,
742                                struct mthca_qp *qp)
743 {
744         int size;
745         int err = -ENOMEM;
746
747         size = sizeof (struct mthca_next_seg) +
748                 qp->rq.max_gs * sizeof (struct mthca_data_seg);
749
750         if (size > dev->limits.max_desc_sz)
751                 return -EINVAL;
752
753         for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
754              qp->rq.wqe_shift++)
755                 ; /* nothing */
756
757         size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
758         switch (qp->transport) {
759         case MLX:
760                 size += 2 * sizeof (struct mthca_data_seg);
761                 break;
762
763         case UD:
764                 size += mthca_is_memfree(dev) ?
765                         sizeof (struct mthca_arbel_ud_seg) :
766                         sizeof (struct mthca_tavor_ud_seg);
767                 break;
768
769         case UC:
770                 size += sizeof (struct mthca_raddr_seg);
771                 break;
772
773         case RC:
774                 size += sizeof (struct mthca_raddr_seg);
775                 /*
776                  * An atomic op will require an atomic segment, a
777                  * remote address segment and one scatter entry.
778                  */
779                 size = max_t(int, size,
780                              sizeof (struct mthca_atomic_seg) +
781                              sizeof (struct mthca_raddr_seg) +
782                              sizeof (struct mthca_data_seg));
783                 break;
784
785         default:
786                 break;
787         }
788
789         /* Make sure that we have enough space for a bind request */
790         size = max_t(int, size, sizeof (struct mthca_bind_seg));
791
792         size += sizeof (struct mthca_next_seg);
793
794         if (size > dev->limits.max_desc_sz)
795                 return -EINVAL;
796
797         for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
798              qp->sq.wqe_shift++)
799                 ; /* nothing */
800
801         qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
802                                     1 << qp->sq.wqe_shift);
803
804         /*
805          * If this is a userspace QP, we don't actually have to
806          * allocate anything.  All we need is to calculate the WQE
807          * sizes and the send_wqe_offset, so we're done now.
808          */
809         if (pd->ibpd.uobject)
810                 return 0;
811
812         size = PAGE_ALIGN(qp->send_wqe_offset +
813                           (qp->sq.max << qp->sq.wqe_shift));
814
815         qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
816                            GFP_KERNEL);
817         if (!qp->wrid)
818                 goto err_out;
819
820         err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
821                               &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
822         if (err)
823                 goto err_out;
824
825         return 0;
826
827 err_out:
828         kfree(qp->wrid);
829         return err;
830 }
831
832 static void mthca_free_wqe_buf(struct mthca_dev *dev,
833                                struct mthca_qp *qp)
834 {
835         mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
836                                        (qp->sq.max << qp->sq.wqe_shift)),
837                        &qp->queue, qp->is_direct, &qp->mr);
838         kfree(qp->wrid);
839 }
840
841 static int mthca_map_memfree(struct mthca_dev *dev,
842                              struct mthca_qp *qp)
843 {
844         int ret;
845
846         if (mthca_is_memfree(dev)) {
847                 ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
848                 if (ret)
849                         return ret;
850
851                 ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
852                 if (ret)
853                         goto err_qpc;
854
855                 ret = mthca_table_get(dev, dev->qp_table.rdb_table,
856                                       qp->qpn << dev->qp_table.rdb_shift);
857                 if (ret)
858                         goto err_eqpc;
859
860         }
861
862         return 0;
863
864 err_eqpc:
865         mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
866
867 err_qpc:
868         mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
869
870         return ret;
871 }
872
873 static void mthca_unmap_memfree(struct mthca_dev *dev,
874                                 struct mthca_qp *qp)
875 {
876         mthca_table_put(dev, dev->qp_table.rdb_table,
877                         qp->qpn << dev->qp_table.rdb_shift);
878         mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
879         mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
880 }
881
882 static int mthca_alloc_memfree(struct mthca_dev *dev,
883                                struct mthca_qp *qp)
884 {
885         int ret = 0;
886
887         if (mthca_is_memfree(dev)) {
888                 qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
889                                                  qp->qpn, &qp->rq.db);
890                 if (qp->rq.db_index < 0)
891                         return ret;
892
893                 qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
894                                                  qp->qpn, &qp->sq.db);
895                 if (qp->sq.db_index < 0)
896                         mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
897         }
898
899         return ret;
900 }
901
902 static void mthca_free_memfree(struct mthca_dev *dev,
903                                struct mthca_qp *qp)
904 {
905         if (mthca_is_memfree(dev)) {
906                 mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
907                 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
908         }
909 }
910
911 static int mthca_alloc_qp_common(struct mthca_dev *dev,
912                                  struct mthca_pd *pd,
913                                  struct mthca_cq *send_cq,
914                                  struct mthca_cq *recv_cq,
915                                  enum ib_sig_type send_policy,
916                                  struct mthca_qp *qp)
917 {
918         int ret;
919         int i;
920
921         atomic_set(&qp->refcount, 1);
922         init_waitqueue_head(&qp->wait);
923         qp->state        = IB_QPS_RESET;
924         qp->atomic_rd_en = 0;
925         qp->resp_depth   = 0;
926         qp->sq_policy    = send_policy;
927         mthca_wq_init(&qp->sq);
928         mthca_wq_init(&qp->rq);
929
930         ret = mthca_map_memfree(dev, qp);
931         if (ret)
932                 return ret;
933
934         ret = mthca_alloc_wqe_buf(dev, pd, qp);
935         if (ret) {
936                 mthca_unmap_memfree(dev, qp);
937                 return ret;
938         }
939
940         mthca_adjust_qp_caps(dev, pd, qp);
941
942         /*
943          * If this is a userspace QP, we're done now.  The doorbells
944          * will be allocated and buffers will be initialized in
945          * userspace.
946          */
947         if (pd->ibpd.uobject)
948                 return 0;
949
950         ret = mthca_alloc_memfree(dev, qp);
951         if (ret) {
952                 mthca_free_wqe_buf(dev, qp);
953                 mthca_unmap_memfree(dev, qp);
954                 return ret;
955         }
956
957         if (mthca_is_memfree(dev)) {
958                 struct mthca_next_seg *next;
959                 struct mthca_data_seg *scatter;
960                 int size = (sizeof (struct mthca_next_seg) +
961                             qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
962
963                 for (i = 0; i < qp->rq.max; ++i) {
964                         next = get_recv_wqe(qp, i);
965                         next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
966                                                    qp->rq.wqe_shift);
967                         next->ee_nds = cpu_to_be32(size);
968
969                         for (scatter = (void *) (next + 1);
970                              (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
971                              ++scatter)
972                                 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
973                 }
974
975                 for (i = 0; i < qp->sq.max; ++i) {
976                         next = get_send_wqe(qp, i);
977                         next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
978                                                     qp->sq.wqe_shift) +
979                                                    qp->send_wqe_offset);
980                 }
981         }
982
983         qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
984         qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
985
986         return 0;
987 }
988
989 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
990                              struct mthca_pd *pd, struct mthca_qp *qp)
991 {
992         int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz);
993
994         /* Sanity check QP size before proceeding */
995         if (cap->max_send_wr     > dev->limits.max_wqes ||
996             cap->max_recv_wr     > dev->limits.max_wqes ||
997             cap->max_send_sge    > dev->limits.max_sg   ||
998             cap->max_recv_sge    > dev->limits.max_sg   ||
999             cap->max_inline_data > mthca_max_inline_data(pd, max_data_size))
1000                 return -EINVAL;
1001
1002         /*
1003          * For MLX transport we need 2 extra S/G entries:
1004          * one for the header and one for the checksum at the end
1005          */
1006         if (qp->transport == MLX && cap->max_recv_sge + 2 > dev->limits.max_sg)
1007                 return -EINVAL;
1008
1009         if (mthca_is_memfree(dev)) {
1010                 qp->rq.max = cap->max_recv_wr ?
1011                         roundup_pow_of_two(cap->max_recv_wr) : 0;
1012                 qp->sq.max = cap->max_send_wr ?
1013                         roundup_pow_of_two(cap->max_send_wr) : 0;
1014         } else {
1015                 qp->rq.max = cap->max_recv_wr;
1016                 qp->sq.max = cap->max_send_wr;
1017         }
1018
1019         qp->rq.max_gs = cap->max_recv_sge;
1020         qp->sq.max_gs = max_t(int, cap->max_send_sge,
1021                               ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1022                                     MTHCA_INLINE_CHUNK_SIZE) /
1023                               sizeof (struct mthca_data_seg));
1024
1025         return 0;
1026 }
1027
1028 int mthca_alloc_qp(struct mthca_dev *dev,
1029                    struct mthca_pd *pd,
1030                    struct mthca_cq *send_cq,
1031                    struct mthca_cq *recv_cq,
1032                    enum ib_qp_type type,
1033                    enum ib_sig_type send_policy,
1034                    struct ib_qp_cap *cap,
1035                    struct mthca_qp *qp)
1036 {
1037         int err;
1038
1039         err = mthca_set_qp_size(dev, cap, pd, qp);
1040         if (err)
1041                 return err;
1042
1043         switch (type) {
1044         case IB_QPT_RC: qp->transport = RC; break;
1045         case IB_QPT_UC: qp->transport = UC; break;
1046         case IB_QPT_UD: qp->transport = UD; break;
1047         default: return -EINVAL;
1048         }
1049
1050         qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1051         if (qp->qpn == -1)
1052                 return -ENOMEM;
1053
1054         err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1055                                     send_policy, qp);
1056         if (err) {
1057                 mthca_free(&dev->qp_table.alloc, qp->qpn);
1058                 return err;
1059         }
1060
1061         spin_lock_irq(&dev->qp_table.lock);
1062         mthca_array_set(&dev->qp_table.qp,
1063                         qp->qpn & (dev->limits.num_qps - 1), qp);
1064         spin_unlock_irq(&dev->qp_table.lock);
1065
1066         return 0;
1067 }
1068
1069 int mthca_alloc_sqp(struct mthca_dev *dev,
1070                     struct mthca_pd *pd,
1071                     struct mthca_cq *send_cq,
1072                     struct mthca_cq *recv_cq,
1073                     enum ib_sig_type send_policy,
1074                     struct ib_qp_cap *cap,
1075                     int qpn,
1076                     int port,
1077                     struct mthca_sqp *sqp)
1078 {
1079         u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1080         int err;
1081
1082         err = mthca_set_qp_size(dev, cap, pd, &sqp->qp);
1083         if (err)
1084                 return err;
1085
1086         sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1087         sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1088                                              &sqp->header_dma, GFP_KERNEL);
1089         if (!sqp->header_buf)
1090                 return -ENOMEM;
1091
1092         spin_lock_irq(&dev->qp_table.lock);
1093         if (mthca_array_get(&dev->qp_table.qp, mqpn))
1094                 err = -EBUSY;
1095         else
1096                 mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1097         spin_unlock_irq(&dev->qp_table.lock);
1098
1099         if (err)
1100                 goto err_out;
1101
1102         sqp->port = port;
1103         sqp->qp.qpn       = mqpn;
1104         sqp->qp.transport = MLX;
1105
1106         err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1107                                     send_policy, &sqp->qp);
1108         if (err)
1109                 goto err_out_free;
1110
1111         atomic_inc(&pd->sqp_count);
1112
1113         return 0;
1114
1115  err_out_free:
1116         /*
1117          * Lock CQs here, so that CQ polling code can do QP lookup
1118          * without taking a lock.
1119          */
1120         spin_lock_irq(&send_cq->lock);
1121         if (send_cq != recv_cq)
1122                 spin_lock(&recv_cq->lock);
1123
1124         spin_lock(&dev->qp_table.lock);
1125         mthca_array_clear(&dev->qp_table.qp, mqpn);
1126         spin_unlock(&dev->qp_table.lock);
1127
1128         if (send_cq != recv_cq)
1129                 spin_unlock(&recv_cq->lock);
1130         spin_unlock_irq(&send_cq->lock);
1131
1132  err_out:
1133         dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1134                           sqp->header_buf, sqp->header_dma);
1135
1136         return err;
1137 }
1138
1139 void mthca_free_qp(struct mthca_dev *dev,
1140                    struct mthca_qp *qp)
1141 {
1142         u8 status;
1143         struct mthca_cq *send_cq;
1144         struct mthca_cq *recv_cq;
1145
1146         send_cq = to_mcq(qp->ibqp.send_cq);
1147         recv_cq = to_mcq(qp->ibqp.recv_cq);
1148
1149         /*
1150          * Lock CQs here, so that CQ polling code can do QP lookup
1151          * without taking a lock.
1152          */
1153         spin_lock_irq(&send_cq->lock);
1154         if (send_cq != recv_cq)
1155                 spin_lock(&recv_cq->lock);
1156
1157         spin_lock(&dev->qp_table.lock);
1158         mthca_array_clear(&dev->qp_table.qp,
1159                           qp->qpn & (dev->limits.num_qps - 1));
1160         spin_unlock(&dev->qp_table.lock);
1161
1162         if (send_cq != recv_cq)
1163                 spin_unlock(&recv_cq->lock);
1164         spin_unlock_irq(&send_cq->lock);
1165
1166         atomic_dec(&qp->refcount);
1167         wait_event(qp->wait, !atomic_read(&qp->refcount));
1168
1169         if (qp->state != IB_QPS_RESET)
1170                 mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0,
1171                                 NULL, 0, &status);
1172
1173         /*
1174          * If this is a userspace QP, the buffers, MR, CQs and so on
1175          * will be cleaned up in userspace, so all we have to do is
1176          * unref the mem-free tables and free the QPN in our table.
1177          */
1178         if (!qp->ibqp.uobject) {
1179                 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
1180                                qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1181                 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
1182                         mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
1183                                        qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1184
1185                 mthca_free_memfree(dev, qp);
1186                 mthca_free_wqe_buf(dev, qp);
1187         }
1188
1189         mthca_unmap_memfree(dev, qp);
1190
1191         if (is_sqp(dev, qp)) {
1192                 atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1193                 dma_free_coherent(&dev->pdev->dev,
1194                                   to_msqp(qp)->header_buf_size,
1195                                   to_msqp(qp)->header_buf,
1196                                   to_msqp(qp)->header_dma);
1197         } else
1198                 mthca_free(&dev->qp_table.alloc, qp->qpn);
1199 }
1200
1201 /* Create UD header for an MLX send and build a data segment for it */
1202 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1203                             int ind, struct ib_send_wr *wr,
1204                             struct mthca_mlx_seg *mlx,
1205                             struct mthca_data_seg *data)
1206 {
1207         int header_size;
1208         int err;
1209         u16 pkey;
1210
1211         ib_ud_header_init(256, /* assume a MAD */
1212                           mthca_ah_grh_present(to_mah(wr->wr.ud.ah)),
1213                           &sqp->ud_header);
1214
1215         err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1216         if (err)
1217                 return err;
1218         mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1219         mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1220                                   (sqp->ud_header.lrh.destination_lid ==
1221                                    IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1222                                   (sqp->ud_header.lrh.service_level << 8));
1223         mlx->rlid = sqp->ud_header.lrh.destination_lid;
1224         mlx->vcrc = 0;
1225
1226         switch (wr->opcode) {
1227         case IB_WR_SEND:
1228                 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1229                 sqp->ud_header.immediate_present = 0;
1230                 break;
1231         case IB_WR_SEND_WITH_IMM:
1232                 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1233                 sqp->ud_header.immediate_present = 1;
1234                 sqp->ud_header.immediate_data = wr->imm_data;
1235                 break;
1236         default:
1237                 return -EINVAL;
1238         }
1239
1240         sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1241         if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1242                 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1243         sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1244         if (!sqp->qp.ibqp.qp_num)
1245                 ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1246                                    sqp->pkey_index, &pkey);
1247         else
1248                 ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1249                                    wr->wr.ud.pkey_index, &pkey);
1250         sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1251         sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1252         sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1253         sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1254                                                sqp->qkey : wr->wr.ud.remote_qkey);
1255         sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1256
1257         header_size = ib_ud_header_pack(&sqp->ud_header,
1258                                         sqp->header_buf +
1259                                         ind * MTHCA_UD_HEADER_SIZE);
1260
1261         data->byte_count = cpu_to_be32(header_size);
1262         data->lkey       = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1263         data->addr       = cpu_to_be64(sqp->header_dma +
1264                                        ind * MTHCA_UD_HEADER_SIZE);
1265
1266         return 0;
1267 }
1268
1269 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1270                                     struct ib_cq *ib_cq)
1271 {
1272         unsigned cur;
1273         struct mthca_cq *cq;
1274
1275         cur = wq->head - wq->tail;
1276         if (likely(cur + nreq < wq->max))
1277                 return 0;
1278
1279         cq = to_mcq(ib_cq);
1280         spin_lock(&cq->lock);
1281         cur = wq->head - wq->tail;
1282         spin_unlock(&cq->lock);
1283
1284         return cur + nreq >= wq->max;
1285 }
1286
1287 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1288                           struct ib_send_wr **bad_wr)
1289 {
1290         struct mthca_dev *dev = to_mdev(ibqp->device);
1291         struct mthca_qp *qp = to_mqp(ibqp);
1292         void *wqe;
1293         void *prev_wqe;
1294         unsigned long flags;
1295         int err = 0;
1296         int nreq;
1297         int i;
1298         int size;
1299         int size0 = 0;
1300         u32 f0 = 0;
1301         int ind;
1302         u8 op0 = 0;
1303
1304         spin_lock_irqsave(&qp->sq.lock, flags);
1305
1306         /* XXX check that state is OK to post send */
1307
1308         ind = qp->sq.next_ind;
1309
1310         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1311                 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1312                         mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1313                                         " %d max, %d nreq)\n", qp->qpn,
1314                                         qp->sq.head, qp->sq.tail,
1315                                         qp->sq.max, nreq);
1316                         err = -ENOMEM;
1317                         *bad_wr = wr;
1318                         goto out;
1319                 }
1320
1321                 wqe = get_send_wqe(qp, ind);
1322                 prev_wqe = qp->sq.last;
1323                 qp->sq.last = wqe;
1324
1325                 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1326                 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
1327                 ((struct mthca_next_seg *) wqe)->flags =
1328                         ((wr->send_flags & IB_SEND_SIGNALED) ?
1329                          cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1330                         ((wr->send_flags & IB_SEND_SOLICITED) ?
1331                          cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1332                         cpu_to_be32(1);
1333                 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1334                     wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1335                         ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1336
1337                 wqe += sizeof (struct mthca_next_seg);
1338                 size = sizeof (struct mthca_next_seg) / 16;
1339
1340                 switch (qp->transport) {
1341                 case RC:
1342                         switch (wr->opcode) {
1343                         case IB_WR_ATOMIC_CMP_AND_SWP:
1344                         case IB_WR_ATOMIC_FETCH_AND_ADD:
1345                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1346                                         cpu_to_be64(wr->wr.atomic.remote_addr);
1347                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1348                                         cpu_to_be32(wr->wr.atomic.rkey);
1349                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1350
1351                                 wqe += sizeof (struct mthca_raddr_seg);
1352
1353                                 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1354                                         ((struct mthca_atomic_seg *) wqe)->swap_add =
1355                                                 cpu_to_be64(wr->wr.atomic.swap);
1356                                         ((struct mthca_atomic_seg *) wqe)->compare =
1357                                                 cpu_to_be64(wr->wr.atomic.compare_add);
1358                                 } else {
1359                                         ((struct mthca_atomic_seg *) wqe)->swap_add =
1360                                                 cpu_to_be64(wr->wr.atomic.compare_add);
1361                                         ((struct mthca_atomic_seg *) wqe)->compare = 0;
1362                                 }
1363
1364                                 wqe += sizeof (struct mthca_atomic_seg);
1365                                 size += (sizeof (struct mthca_raddr_seg) +
1366                                          sizeof (struct mthca_atomic_seg)) / 16;
1367                                 break;
1368
1369                         case IB_WR_RDMA_WRITE:
1370                         case IB_WR_RDMA_WRITE_WITH_IMM:
1371                         case IB_WR_RDMA_READ:
1372                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1373                                         cpu_to_be64(wr->wr.rdma.remote_addr);
1374                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1375                                         cpu_to_be32(wr->wr.rdma.rkey);
1376                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1377                                 wqe += sizeof (struct mthca_raddr_seg);
1378                                 size += sizeof (struct mthca_raddr_seg) / 16;
1379                                 break;
1380
1381                         default:
1382                                 /* No extra segments required for sends */
1383                                 break;
1384                         }
1385
1386                         break;
1387
1388                 case UC:
1389                         switch (wr->opcode) {
1390                         case IB_WR_RDMA_WRITE:
1391                         case IB_WR_RDMA_WRITE_WITH_IMM:
1392                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1393                                         cpu_to_be64(wr->wr.rdma.remote_addr);
1394                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1395                                         cpu_to_be32(wr->wr.rdma.rkey);
1396                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1397                                 wqe += sizeof (struct mthca_raddr_seg);
1398                                 size += sizeof (struct mthca_raddr_seg) / 16;
1399                                 break;
1400
1401                         default:
1402                                 /* No extra segments required for sends */
1403                                 break;
1404                         }
1405
1406                         break;
1407
1408                 case UD:
1409                         ((struct mthca_tavor_ud_seg *) wqe)->lkey =
1410                                 cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1411                         ((struct mthca_tavor_ud_seg *) wqe)->av_addr =
1412                                 cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1413                         ((struct mthca_tavor_ud_seg *) wqe)->dqpn =
1414                                 cpu_to_be32(wr->wr.ud.remote_qpn);
1415                         ((struct mthca_tavor_ud_seg *) wqe)->qkey =
1416                                 cpu_to_be32(wr->wr.ud.remote_qkey);
1417
1418                         wqe += sizeof (struct mthca_tavor_ud_seg);
1419                         size += sizeof (struct mthca_tavor_ud_seg) / 16;
1420                         break;
1421
1422                 case MLX:
1423                         err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1424                                                wqe - sizeof (struct mthca_next_seg),
1425                                                wqe);
1426                         if (err) {
1427                                 *bad_wr = wr;
1428                                 goto out;
1429                         }
1430                         wqe += sizeof (struct mthca_data_seg);
1431                         size += sizeof (struct mthca_data_seg) / 16;
1432                         break;
1433                 }
1434
1435                 if (wr->num_sge > qp->sq.max_gs) {
1436                         mthca_err(dev, "too many gathers\n");
1437                         err = -EINVAL;
1438                         *bad_wr = wr;
1439                         goto out;
1440                 }
1441
1442                 for (i = 0; i < wr->num_sge; ++i) {
1443                         ((struct mthca_data_seg *) wqe)->byte_count =
1444                                 cpu_to_be32(wr->sg_list[i].length);
1445                         ((struct mthca_data_seg *) wqe)->lkey =
1446                                 cpu_to_be32(wr->sg_list[i].lkey);
1447                         ((struct mthca_data_seg *) wqe)->addr =
1448                                 cpu_to_be64(wr->sg_list[i].addr);
1449                         wqe += sizeof (struct mthca_data_seg);
1450                         size += sizeof (struct mthca_data_seg) / 16;
1451                 }
1452
1453                 /* Add one more inline data segment for ICRC */
1454                 if (qp->transport == MLX) {
1455                         ((struct mthca_data_seg *) wqe)->byte_count =
1456                                 cpu_to_be32((1 << 31) | 4);
1457                         ((u32 *) wqe)[1] = 0;
1458                         wqe += sizeof (struct mthca_data_seg);
1459                         size += sizeof (struct mthca_data_seg) / 16;
1460                 }
1461
1462                 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1463
1464                 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1465                         mthca_err(dev, "opcode invalid\n");
1466                         err = -EINVAL;
1467                         *bad_wr = wr;
1468                         goto out;
1469                 }
1470
1471                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1472                         cpu_to_be32(((ind << qp->sq.wqe_shift) +
1473                                      qp->send_wqe_offset) |
1474                                     mthca_opcode[wr->opcode]);
1475                 wmb();
1476                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1477                         cpu_to_be32((size0 ? 0 : MTHCA_NEXT_DBD) | size);
1478
1479                 if (!size0) {
1480                         size0 = size;
1481                         op0   = mthca_opcode[wr->opcode];
1482                 }
1483
1484                 ++ind;
1485                 if (unlikely(ind >= qp->sq.max))
1486                         ind -= qp->sq.max;
1487         }
1488
1489 out:
1490         if (likely(nreq)) {
1491                 __be32 doorbell[2];
1492
1493                 doorbell[0] = cpu_to_be32(((qp->sq.next_ind << qp->sq.wqe_shift) +
1494                                            qp->send_wqe_offset) | f0 | op0);
1495                 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1496
1497                 wmb();
1498
1499                 mthca_write64(doorbell,
1500                               dev->kar + MTHCA_SEND_DOORBELL,
1501                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1502         }
1503
1504         qp->sq.next_ind = ind;
1505         qp->sq.head    += nreq;
1506
1507         spin_unlock_irqrestore(&qp->sq.lock, flags);
1508         return err;
1509 }
1510
1511 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1512                              struct ib_recv_wr **bad_wr)
1513 {
1514         struct mthca_dev *dev = to_mdev(ibqp->device);
1515         struct mthca_qp *qp = to_mqp(ibqp);
1516         __be32 doorbell[2];
1517         unsigned long flags;
1518         int err = 0;
1519         int nreq;
1520         int i;
1521         int size;
1522         int size0 = 0;
1523         int ind;
1524         void *wqe;
1525         void *prev_wqe;
1526
1527         spin_lock_irqsave(&qp->rq.lock, flags);
1528
1529         /* XXX check that state is OK to post receive */
1530
1531         ind = qp->rq.next_ind;
1532
1533         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1534                 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1535                         nreq = 0;
1536
1537                         doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1538                         doorbell[1] = cpu_to_be32(qp->qpn << 8);
1539
1540                         wmb();
1541
1542                         mthca_write64(doorbell,
1543                                       dev->kar + MTHCA_RECEIVE_DOORBELL,
1544                                       MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1545
1546                         qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1547                         size0 = 0;
1548                 }
1549
1550                 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1551                         mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1552                                         " %d max, %d nreq)\n", qp->qpn,
1553                                         qp->rq.head, qp->rq.tail,
1554                                         qp->rq.max, nreq);
1555                         err = -ENOMEM;
1556                         *bad_wr = wr;
1557                         goto out;
1558                 }
1559
1560                 wqe = get_recv_wqe(qp, ind);
1561                 prev_wqe = qp->rq.last;
1562                 qp->rq.last = wqe;
1563
1564                 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1565                 ((struct mthca_next_seg *) wqe)->ee_nds =
1566                         cpu_to_be32(MTHCA_NEXT_DBD);
1567                 ((struct mthca_next_seg *) wqe)->flags = 0;
1568
1569                 wqe += sizeof (struct mthca_next_seg);
1570                 size = sizeof (struct mthca_next_seg) / 16;
1571
1572                 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1573                         err = -EINVAL;
1574                         *bad_wr = wr;
1575                         goto out;
1576                 }
1577
1578                 for (i = 0; i < wr->num_sge; ++i) {
1579                         ((struct mthca_data_seg *) wqe)->byte_count =
1580                                 cpu_to_be32(wr->sg_list[i].length);
1581                         ((struct mthca_data_seg *) wqe)->lkey =
1582                                 cpu_to_be32(wr->sg_list[i].lkey);
1583                         ((struct mthca_data_seg *) wqe)->addr =
1584                                 cpu_to_be64(wr->sg_list[i].addr);
1585                         wqe += sizeof (struct mthca_data_seg);
1586                         size += sizeof (struct mthca_data_seg) / 16;
1587                 }
1588
1589                 qp->wrid[ind] = wr->wr_id;
1590
1591                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1592                         cpu_to_be32((ind << qp->rq.wqe_shift) | 1);
1593                 wmb();
1594                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1595                         cpu_to_be32(MTHCA_NEXT_DBD | size);
1596
1597                 if (!size0)
1598                         size0 = size;
1599
1600                 ++ind;
1601                 if (unlikely(ind >= qp->rq.max))
1602                         ind -= qp->rq.max;
1603         }
1604
1605 out:
1606         if (likely(nreq)) {
1607                 doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1608                 doorbell[1] = cpu_to_be32((qp->qpn << 8) | nreq);
1609
1610                 wmb();
1611
1612                 mthca_write64(doorbell,
1613                               dev->kar + MTHCA_RECEIVE_DOORBELL,
1614                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1615         }
1616
1617         qp->rq.next_ind = ind;
1618         qp->rq.head    += nreq;
1619
1620         spin_unlock_irqrestore(&qp->rq.lock, flags);
1621         return err;
1622 }
1623
1624 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1625                           struct ib_send_wr **bad_wr)
1626 {
1627         struct mthca_dev *dev = to_mdev(ibqp->device);
1628         struct mthca_qp *qp = to_mqp(ibqp);
1629         __be32 doorbell[2];
1630         void *wqe;
1631         void *prev_wqe;
1632         unsigned long flags;
1633         int err = 0;
1634         int nreq;
1635         int i;
1636         int size;
1637         int size0 = 0;
1638         u32 f0 = 0;
1639         int ind;
1640         u8 op0 = 0;
1641
1642         spin_lock_irqsave(&qp->sq.lock, flags);
1643
1644         /* XXX check that state is OK to post send */
1645
1646         ind = qp->sq.head & (qp->sq.max - 1);
1647
1648         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1649                 if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
1650                         nreq = 0;
1651
1652                         doorbell[0] = cpu_to_be32((MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
1653                                                   ((qp->sq.head & 0xffff) << 8) |
1654                                                   f0 | op0);
1655                         doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1656
1657                         qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;
1658                         size0 = 0;
1659
1660                         /*
1661                          * Make sure that descriptors are written before
1662                          * doorbell record.
1663                          */
1664                         wmb();
1665                         *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1666
1667                         /*
1668                          * Make sure doorbell record is written before we
1669                          * write MMIO send doorbell.
1670                          */
1671                         wmb();
1672                         mthca_write64(doorbell,
1673                                       dev->kar + MTHCA_SEND_DOORBELL,
1674                                       MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1675                 }
1676
1677                 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1678                         mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1679                                         " %d max, %d nreq)\n", qp->qpn,
1680                                         qp->sq.head, qp->sq.tail,
1681                                         qp->sq.max, nreq);
1682                         err = -ENOMEM;
1683                         *bad_wr = wr;
1684                         goto out;
1685                 }
1686
1687                 wqe = get_send_wqe(qp, ind);
1688                 prev_wqe = qp->sq.last;
1689                 qp->sq.last = wqe;
1690
1691                 ((struct mthca_next_seg *) wqe)->flags =
1692                         ((wr->send_flags & IB_SEND_SIGNALED) ?
1693                          cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1694                         ((wr->send_flags & IB_SEND_SOLICITED) ?
1695                          cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1696                         cpu_to_be32(1);
1697                 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1698                     wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1699                         ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1700
1701                 wqe += sizeof (struct mthca_next_seg);
1702                 size = sizeof (struct mthca_next_seg) / 16;
1703
1704                 switch (qp->transport) {
1705                 case RC:
1706                         switch (wr->opcode) {
1707                         case IB_WR_ATOMIC_CMP_AND_SWP:
1708                         case IB_WR_ATOMIC_FETCH_AND_ADD:
1709                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1710                                         cpu_to_be64(wr->wr.atomic.remote_addr);
1711                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1712                                         cpu_to_be32(wr->wr.atomic.rkey);
1713                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1714
1715                                 wqe += sizeof (struct mthca_raddr_seg);
1716
1717                                 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1718                                         ((struct mthca_atomic_seg *) wqe)->swap_add =
1719                                                 cpu_to_be64(wr->wr.atomic.swap);
1720                                         ((struct mthca_atomic_seg *) wqe)->compare =
1721                                                 cpu_to_be64(wr->wr.atomic.compare_add);
1722                                 } else {
1723                                         ((struct mthca_atomic_seg *) wqe)->swap_add =
1724                                                 cpu_to_be64(wr->wr.atomic.compare_add);
1725                                         ((struct mthca_atomic_seg *) wqe)->compare = 0;
1726                                 }
1727
1728                                 wqe += sizeof (struct mthca_atomic_seg);
1729                                 size += (sizeof (struct mthca_raddr_seg) +
1730                                          sizeof (struct mthca_atomic_seg)) / 16;
1731                                 break;
1732
1733                         case IB_WR_RDMA_READ:
1734                         case IB_WR_RDMA_WRITE:
1735                         case IB_WR_RDMA_WRITE_WITH_IMM:
1736                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1737                                         cpu_to_be64(wr->wr.rdma.remote_addr);
1738                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1739                                         cpu_to_be32(wr->wr.rdma.rkey);
1740                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1741                                 wqe += sizeof (struct mthca_raddr_seg);
1742                                 size += sizeof (struct mthca_raddr_seg) / 16;
1743                                 break;
1744
1745                         default:
1746                                 /* No extra segments required for sends */
1747                                 break;
1748                         }
1749
1750                         break;
1751
1752                 case UC:
1753                         switch (wr->opcode) {
1754                         case IB_WR_RDMA_WRITE:
1755                         case IB_WR_RDMA_WRITE_WITH_IMM:
1756                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1757                                         cpu_to_be64(wr->wr.rdma.remote_addr);
1758                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1759                                         cpu_to_be32(wr->wr.rdma.rkey);
1760                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1761                                 wqe += sizeof (struct mthca_raddr_seg);
1762                                 size += sizeof (struct mthca_raddr_seg) / 16;
1763                                 break;
1764
1765                         default:
1766                                 /* No extra segments required for sends */
1767                                 break;
1768                         }
1769
1770                         break;
1771
1772                 case UD:
1773                         memcpy(((struct mthca_arbel_ud_seg *) wqe)->av,
1774                                to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
1775                         ((struct mthca_arbel_ud_seg *) wqe)->dqpn =
1776                                 cpu_to_be32(wr->wr.ud.remote_qpn);
1777                         ((struct mthca_arbel_ud_seg *) wqe)->qkey =
1778                                 cpu_to_be32(wr->wr.ud.remote_qkey);
1779
1780                         wqe += sizeof (struct mthca_arbel_ud_seg);
1781                         size += sizeof (struct mthca_arbel_ud_seg) / 16;
1782                         break;
1783
1784                 case MLX:
1785                         err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1786                                                wqe - sizeof (struct mthca_next_seg),
1787                                                wqe);
1788                         if (err) {
1789                                 *bad_wr = wr;
1790                                 goto out;
1791                         }
1792                         wqe += sizeof (struct mthca_data_seg);
1793                         size += sizeof (struct mthca_data_seg) / 16;
1794                         break;
1795                 }
1796
1797                 if (wr->num_sge > qp->sq.max_gs) {
1798                         mthca_err(dev, "too many gathers\n");
1799                         err = -EINVAL;
1800                         *bad_wr = wr;
1801                         goto out;
1802                 }
1803
1804                 for (i = 0; i < wr->num_sge; ++i) {
1805                         ((struct mthca_data_seg *) wqe)->byte_count =
1806                                 cpu_to_be32(wr->sg_list[i].length);
1807                         ((struct mthca_data_seg *) wqe)->lkey =
1808                                 cpu_to_be32(wr->sg_list[i].lkey);
1809                         ((struct mthca_data_seg *) wqe)->addr =
1810                                 cpu_to_be64(wr->sg_list[i].addr);
1811                         wqe += sizeof (struct mthca_data_seg);
1812                         size += sizeof (struct mthca_data_seg) / 16;
1813                 }
1814
1815                 /* Add one more inline data segment for ICRC */
1816                 if (qp->transport == MLX) {
1817                         ((struct mthca_data_seg *) wqe)->byte_count =
1818                                 cpu_to_be32((1 << 31) | 4);
1819                         ((u32 *) wqe)[1] = 0;
1820                         wqe += sizeof (struct mthca_data_seg);
1821                         size += sizeof (struct mthca_data_seg) / 16;
1822                 }
1823
1824                 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1825
1826                 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1827                         mthca_err(dev, "opcode invalid\n");
1828                         err = -EINVAL;
1829                         *bad_wr = wr;
1830                         goto out;
1831                 }
1832
1833                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1834                         cpu_to_be32(((ind << qp->sq.wqe_shift) +
1835                                      qp->send_wqe_offset) |
1836                                     mthca_opcode[wr->opcode]);
1837                 wmb();
1838                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1839                         cpu_to_be32(MTHCA_NEXT_DBD | size);
1840
1841                 if (!size0) {
1842                         size0 = size;
1843                         op0   = mthca_opcode[wr->opcode];
1844                 }
1845
1846                 ++ind;
1847                 if (unlikely(ind >= qp->sq.max))
1848                         ind -= qp->sq.max;
1849         }
1850
1851 out:
1852         if (likely(nreq)) {
1853                 doorbell[0] = cpu_to_be32((nreq << 24)                  |
1854                                           ((qp->sq.head & 0xffff) << 8) |
1855                                           f0 | op0);
1856                 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1857
1858                 qp->sq.head += nreq;
1859
1860                 /*
1861                  * Make sure that descriptors are written before
1862                  * doorbell record.
1863                  */
1864                 wmb();
1865                 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1866
1867                 /*
1868                  * Make sure doorbell record is written before we
1869                  * write MMIO send doorbell.
1870                  */
1871                 wmb();
1872                 mthca_write64(doorbell,
1873                               dev->kar + MTHCA_SEND_DOORBELL,
1874                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1875         }
1876
1877         spin_unlock_irqrestore(&qp->sq.lock, flags);
1878         return err;
1879 }
1880
1881 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1882                              struct ib_recv_wr **bad_wr)
1883 {
1884         struct mthca_dev *dev = to_mdev(ibqp->device);
1885         struct mthca_qp *qp = to_mqp(ibqp);
1886         unsigned long flags;
1887         int err = 0;
1888         int nreq;
1889         int ind;
1890         int i;
1891         void *wqe;
1892
1893         spin_lock_irqsave(&qp->rq.lock, flags);
1894
1895         /* XXX check that state is OK to post receive */
1896
1897         ind = qp->rq.head & (qp->rq.max - 1);
1898
1899         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1900                 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1901                         mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1902                                         " %d max, %d nreq)\n", qp->qpn,
1903                                         qp->rq.head, qp->rq.tail,
1904                                         qp->rq.max, nreq);
1905                         err = -ENOMEM;
1906                         *bad_wr = wr;
1907                         goto out;
1908                 }
1909
1910                 wqe = get_recv_wqe(qp, ind);
1911
1912                 ((struct mthca_next_seg *) wqe)->flags = 0;
1913
1914                 wqe += sizeof (struct mthca_next_seg);
1915
1916                 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1917                         err = -EINVAL;
1918                         *bad_wr = wr;
1919                         goto out;
1920                 }
1921
1922                 for (i = 0; i < wr->num_sge; ++i) {
1923                         ((struct mthca_data_seg *) wqe)->byte_count =
1924                                 cpu_to_be32(wr->sg_list[i].length);
1925                         ((struct mthca_data_seg *) wqe)->lkey =
1926                                 cpu_to_be32(wr->sg_list[i].lkey);
1927                         ((struct mthca_data_seg *) wqe)->addr =
1928                                 cpu_to_be64(wr->sg_list[i].addr);
1929                         wqe += sizeof (struct mthca_data_seg);
1930                 }
1931
1932                 if (i < qp->rq.max_gs) {
1933                         ((struct mthca_data_seg *) wqe)->byte_count = 0;
1934                         ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1935                         ((struct mthca_data_seg *) wqe)->addr = 0;
1936                 }
1937
1938                 qp->wrid[ind] = wr->wr_id;
1939
1940                 ++ind;
1941                 if (unlikely(ind >= qp->rq.max))
1942                         ind -= qp->rq.max;
1943         }
1944 out:
1945         if (likely(nreq)) {
1946                 qp->rq.head += nreq;
1947
1948                 /*
1949                  * Make sure that descriptors are written before
1950                  * doorbell record.
1951                  */
1952                 wmb();
1953                 *qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
1954         }
1955
1956         spin_unlock_irqrestore(&qp->rq.lock, flags);
1957         return err;
1958 }
1959
1960 void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
1961                         int index, int *dbd, __be32 *new_wqe)
1962 {
1963         struct mthca_next_seg *next;
1964
1965         /*
1966          * For SRQs, all WQEs generate a CQE, so we're always at the
1967          * end of the doorbell chain.
1968          */
1969         if (qp->ibqp.srq) {
1970                 *new_wqe = 0;
1971                 return;
1972         }
1973
1974         if (is_send)
1975                 next = get_send_wqe(qp, index);
1976         else
1977                 next = get_recv_wqe(qp, index);
1978
1979         *dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
1980         if (next->ee_nds & cpu_to_be32(0x3f))
1981                 *new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
1982                         (next->ee_nds & cpu_to_be32(0x3f));
1983         else
1984                 *new_wqe = 0;
1985 }
1986
1987 int __devinit mthca_init_qp_table(struct mthca_dev *dev)
1988 {
1989         int err;
1990         u8 status;
1991         int i;
1992
1993         spin_lock_init(&dev->qp_table.lock);
1994
1995         /*
1996          * We reserve 2 extra QPs per port for the special QPs.  The
1997          * special QP for port 1 has to be even, so round up.
1998          */
1999         dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2000         err = mthca_alloc_init(&dev->qp_table.alloc,
2001                                dev->limits.num_qps,
2002                                (1 << 24) - 1,
2003                                dev->qp_table.sqp_start +
2004                                MTHCA_MAX_PORTS * 2);
2005         if (err)
2006                 return err;
2007
2008         err = mthca_array_init(&dev->qp_table.qp,
2009                                dev->limits.num_qps);
2010         if (err) {
2011                 mthca_alloc_cleanup(&dev->qp_table.alloc);
2012                 return err;
2013         }
2014
2015         for (i = 0; i < 2; ++i) {
2016                 err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2017                                             dev->qp_table.sqp_start + i * 2,
2018                                             &status);
2019                 if (err)
2020                         goto err_out;
2021                 if (status) {
2022                         mthca_warn(dev, "CONF_SPECIAL_QP returned "
2023                                    "status %02x, aborting.\n",
2024                                    status);
2025                         err = -EINVAL;
2026                         goto err_out;
2027                 }
2028         }
2029         return 0;
2030
2031  err_out:
2032         for (i = 0; i < 2; ++i)
2033                 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2034
2035         mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2036         mthca_alloc_cleanup(&dev->qp_table.alloc);
2037
2038         return err;
2039 }
2040
2041 void __devexit mthca_cleanup_qp_table(struct mthca_dev *dev)
2042 {
2043         int i;
2044         u8 status;
2045
2046         for (i = 0; i < 2; ++i)
2047                 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2048
2049         mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2050         mthca_alloc_cleanup(&dev->qp_table.alloc);
2051 }