Merge branch 'linus' into x86/i8259
[linux-2.6] / drivers / infiniband / hw / mlx4 / qp.c
1 /*
2  * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32
33 #include <linux/log2.h>
34
35 #include <rdma/ib_cache.h>
36 #include <rdma/ib_pack.h>
37
38 #include <linux/mlx4/qp.h>
39
40 #include "mlx4_ib.h"
41 #include "user.h"
42
43 enum {
44         MLX4_IB_ACK_REQ_FREQ    = 8,
45 };
46
47 enum {
48         MLX4_IB_DEFAULT_SCHED_QUEUE     = 0x83,
49         MLX4_IB_DEFAULT_QP0_SCHED_QUEUE = 0x3f
50 };
51
52 enum {
53         /*
54          * Largest possible UD header: send with GRH and immediate data.
55          */
56         MLX4_IB_UD_HEADER_SIZE          = 72
57 };
58
59 struct mlx4_ib_sqp {
60         struct mlx4_ib_qp       qp;
61         int                     pkey_index;
62         u32                     qkey;
63         u32                     send_psn;
64         struct ib_ud_header     ud_header;
65         u8                      header_buf[MLX4_IB_UD_HEADER_SIZE];
66 };
67
68 enum {
69         MLX4_IB_MIN_SQ_STRIDE = 6
70 };
71
72 static const __be32 mlx4_ib_opcode[] = {
73         [IB_WR_SEND]                    = __constant_cpu_to_be32(MLX4_OPCODE_SEND),
74         [IB_WR_LSO]                     = __constant_cpu_to_be32(MLX4_OPCODE_LSO),
75         [IB_WR_SEND_WITH_IMM]           = __constant_cpu_to_be32(MLX4_OPCODE_SEND_IMM),
76         [IB_WR_RDMA_WRITE]              = __constant_cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
77         [IB_WR_RDMA_WRITE_WITH_IMM]     = __constant_cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
78         [IB_WR_RDMA_READ]               = __constant_cpu_to_be32(MLX4_OPCODE_RDMA_READ),
79         [IB_WR_ATOMIC_CMP_AND_SWP]      = __constant_cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
80         [IB_WR_ATOMIC_FETCH_AND_ADD]    = __constant_cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
81 };
82
83 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
84 {
85         return container_of(mqp, struct mlx4_ib_sqp, qp);
86 }
87
88 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
89 {
90         return qp->mqp.qpn >= dev->dev->caps.sqp_start &&
91                 qp->mqp.qpn <= dev->dev->caps.sqp_start + 3;
92 }
93
94 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
95 {
96         return qp->mqp.qpn >= dev->dev->caps.sqp_start &&
97                 qp->mqp.qpn <= dev->dev->caps.sqp_start + 1;
98 }
99
100 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
101 {
102         return mlx4_buf_offset(&qp->buf, offset);
103 }
104
105 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
106 {
107         return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
108 }
109
110 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
111 {
112         return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
113 }
114
115 /*
116  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
117  * first four bytes of every 64 byte chunk with
118  *     0x7FFFFFF | (invalid_ownership_value << 31).
119  *
120  * When the max work request size is less than or equal to the WQE
121  * basic block size, as an optimization, we can stamp all WQEs with
122  * 0xffffffff, and skip the very first chunk of each WQE.
123  */
124 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
125 {
126         __be32 *wqe;
127         int i;
128         int s;
129         int ind;
130         void *buf;
131         __be32 stamp;
132
133         s = roundup(size, 1U << qp->sq.wqe_shift);
134         if (qp->sq_max_wqes_per_wr > 1) {
135                 for (i = 0; i < s; i += 64) {
136                         ind = (i >> qp->sq.wqe_shift) + n;
137                         stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
138                                                        cpu_to_be32(0xffffffff);
139                         buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
140                         wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
141                         *wqe = stamp;
142                 }
143         } else {
144                 buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
145                 for (i = 64; i < s; i += 64) {
146                         wqe = buf + i;
147                         *wqe = cpu_to_be32(0xffffffff);
148                 }
149         }
150 }
151
152 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
153 {
154         struct mlx4_wqe_ctrl_seg *ctrl;
155         struct mlx4_wqe_inline_seg *inl;
156         void *wqe;
157         int s;
158
159         ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
160         s = sizeof(struct mlx4_wqe_ctrl_seg);
161
162         if (qp->ibqp.qp_type == IB_QPT_UD) {
163                 struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
164                 struct mlx4_av *av = (struct mlx4_av *)dgram->av;
165                 memset(dgram, 0, sizeof *dgram);
166                 av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
167                 s += sizeof(struct mlx4_wqe_datagram_seg);
168         }
169
170         /* Pad the remainder of the WQE with an inline data segment. */
171         if (size > s) {
172                 inl = wqe + s;
173                 inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
174         }
175         ctrl->srcrb_flags = 0;
176         ctrl->fence_size = size / 16;
177         /*
178          * Make sure descriptor is fully written before setting ownership bit
179          * (because HW can start executing as soon as we do).
180          */
181         wmb();
182
183         ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
184                 (n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
185
186         stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
187 }
188
189 /* Post NOP WQE to prevent wrap-around in the middle of WR */
190 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
191 {
192         unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
193         if (unlikely(s < qp->sq_max_wqes_per_wr)) {
194                 post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
195                 ind += s;
196         }
197         return ind;
198 }
199
200 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
201 {
202         struct ib_event event;
203         struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
204
205         if (type == MLX4_EVENT_TYPE_PATH_MIG)
206                 to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
207
208         if (ibqp->event_handler) {
209                 event.device     = ibqp->device;
210                 event.element.qp = ibqp;
211                 switch (type) {
212                 case MLX4_EVENT_TYPE_PATH_MIG:
213                         event.event = IB_EVENT_PATH_MIG;
214                         break;
215                 case MLX4_EVENT_TYPE_COMM_EST:
216                         event.event = IB_EVENT_COMM_EST;
217                         break;
218                 case MLX4_EVENT_TYPE_SQ_DRAINED:
219                         event.event = IB_EVENT_SQ_DRAINED;
220                         break;
221                 case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
222                         event.event = IB_EVENT_QP_LAST_WQE_REACHED;
223                         break;
224                 case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
225                         event.event = IB_EVENT_QP_FATAL;
226                         break;
227                 case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
228                         event.event = IB_EVENT_PATH_MIG_ERR;
229                         break;
230                 case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
231                         event.event = IB_EVENT_QP_REQ_ERR;
232                         break;
233                 case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
234                         event.event = IB_EVENT_QP_ACCESS_ERR;
235                         break;
236                 default:
237                         printk(KERN_WARNING "mlx4_ib: Unexpected event type %d "
238                                "on QP %06x\n", type, qp->qpn);
239                         return;
240                 }
241
242                 ibqp->event_handler(&event, ibqp->qp_context);
243         }
244 }
245
246 static int send_wqe_overhead(enum ib_qp_type type, u32 flags)
247 {
248         /*
249          * UD WQEs must have a datagram segment.
250          * RC and UC WQEs might have a remote address segment.
251          * MLX WQEs need two extra inline data segments (for the UD
252          * header and space for the ICRC).
253          */
254         switch (type) {
255         case IB_QPT_UD:
256                 return sizeof (struct mlx4_wqe_ctrl_seg) +
257                         sizeof (struct mlx4_wqe_datagram_seg) +
258                         ((flags & MLX4_IB_QP_LSO) ? 64 : 0);
259         case IB_QPT_UC:
260                 return sizeof (struct mlx4_wqe_ctrl_seg) +
261                         sizeof (struct mlx4_wqe_raddr_seg);
262         case IB_QPT_RC:
263                 return sizeof (struct mlx4_wqe_ctrl_seg) +
264                         sizeof (struct mlx4_wqe_atomic_seg) +
265                         sizeof (struct mlx4_wqe_raddr_seg);
266         case IB_QPT_SMI:
267         case IB_QPT_GSI:
268                 return sizeof (struct mlx4_wqe_ctrl_seg) +
269                         ALIGN(MLX4_IB_UD_HEADER_SIZE +
270                               DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
271                                            MLX4_INLINE_ALIGN) *
272                               sizeof (struct mlx4_wqe_inline_seg),
273                               sizeof (struct mlx4_wqe_data_seg)) +
274                         ALIGN(4 +
275                               sizeof (struct mlx4_wqe_inline_seg),
276                               sizeof (struct mlx4_wqe_data_seg));
277         default:
278                 return sizeof (struct mlx4_wqe_ctrl_seg);
279         }
280 }
281
282 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
283                        int is_user, int has_srq, struct mlx4_ib_qp *qp)
284 {
285         /* Sanity check RQ size before proceeding */
286         if (cap->max_recv_wr  > dev->dev->caps.max_wqes  ||
287             cap->max_recv_sge > dev->dev->caps.max_rq_sg)
288                 return -EINVAL;
289
290         if (has_srq) {
291                 /* QPs attached to an SRQ should have no RQ */
292                 if (cap->max_recv_wr)
293                         return -EINVAL;
294
295                 qp->rq.wqe_cnt = qp->rq.max_gs = 0;
296         } else {
297                 /* HW requires >= 1 RQ entry with >= 1 gather entry */
298                 if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
299                         return -EINVAL;
300
301                 qp->rq.wqe_cnt   = roundup_pow_of_two(max(1U, cap->max_recv_wr));
302                 qp->rq.max_gs    = roundup_pow_of_two(max(1U, cap->max_recv_sge));
303                 qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
304         }
305
306         cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
307         cap->max_recv_sge = qp->rq.max_gs;
308
309         return 0;
310 }
311
312 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
313                               enum ib_qp_type type, struct mlx4_ib_qp *qp)
314 {
315         int s;
316
317         /* Sanity check SQ size before proceeding */
318         if (cap->max_send_wr     > dev->dev->caps.max_wqes  ||
319             cap->max_send_sge    > dev->dev->caps.max_sq_sg ||
320             cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
321             sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
322                 return -EINVAL;
323
324         /*
325          * For MLX transport we need 2 extra S/G entries:
326          * one for the header and one for the checksum at the end
327          */
328         if ((type == IB_QPT_SMI || type == IB_QPT_GSI) &&
329             cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
330                 return -EINVAL;
331
332         s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
333                 cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
334                 send_wqe_overhead(type, qp->flags);
335
336         if (s > dev->dev->caps.max_sq_desc_sz)
337                 return -EINVAL;
338
339         /*
340          * Hermon supports shrinking WQEs, such that a single work
341          * request can include multiple units of 1 << wqe_shift.  This
342          * way, work requests can differ in size, and do not have to
343          * be a power of 2 in size, saving memory and speeding up send
344          * WR posting.  Unfortunately, if we do this then the
345          * wqe_index field in CQEs can't be used to look up the WR ID
346          * anymore, so we do this only if selective signaling is off.
347          *
348          * Further, on 32-bit platforms, we can't use vmap() to make
349          * the QP buffer virtually contigious.  Thus we have to use
350          * constant-sized WRs to make sure a WR is always fully within
351          * a single page-sized chunk.
352          *
353          * Finally, we use NOP work requests to pad the end of the
354          * work queue, to avoid wrap-around in the middle of WR.  We
355          * set NEC bit to avoid getting completions with error for
356          * these NOP WRs, but since NEC is only supported starting
357          * with firmware 2.2.232, we use constant-sized WRs for older
358          * firmware.
359          *
360          * And, since MLX QPs only support SEND, we use constant-sized
361          * WRs in this case.
362          *
363          * We look for the smallest value of wqe_shift such that the
364          * resulting number of wqes does not exceed device
365          * capabilities.
366          *
367          * We set WQE size to at least 64 bytes, this way stamping
368          * invalidates each WQE.
369          */
370         if (dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
371             qp->sq_signal_bits && BITS_PER_LONG == 64 &&
372             type != IB_QPT_SMI && type != IB_QPT_GSI)
373                 qp->sq.wqe_shift = ilog2(64);
374         else
375                 qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
376
377         for (;;) {
378                 qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
379
380                 /*
381                  * We need to leave 2 KB + 1 WR of headroom in the SQ to
382                  * allow HW to prefetch.
383                  */
384                 qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
385                 qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
386                                                     qp->sq_max_wqes_per_wr +
387                                                     qp->sq_spare_wqes);
388
389                 if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
390                         break;
391
392                 if (qp->sq_max_wqes_per_wr <= 1)
393                         return -EINVAL;
394
395                 ++qp->sq.wqe_shift;
396         }
397
398         qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
399                              (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
400                          send_wqe_overhead(type, qp->flags)) /
401                 sizeof (struct mlx4_wqe_data_seg);
402
403         qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
404                 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
405         if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
406                 qp->rq.offset = 0;
407                 qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
408         } else {
409                 qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
410                 qp->sq.offset = 0;
411         }
412
413         cap->max_send_wr  = qp->sq.max_post =
414                 (qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
415         cap->max_send_sge = min(qp->sq.max_gs,
416                                 min(dev->dev->caps.max_sq_sg,
417                                     dev->dev->caps.max_rq_sg));
418         /* We don't support inline sends for kernel QPs (yet) */
419         cap->max_inline_data = 0;
420
421         return 0;
422 }
423
424 static int set_user_sq_size(struct mlx4_ib_dev *dev,
425                             struct mlx4_ib_qp *qp,
426                             struct mlx4_ib_create_qp *ucmd)
427 {
428         /* Sanity check SQ size before proceeding */
429         if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes       ||
430             ucmd->log_sq_stride >
431                 ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
432             ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
433                 return -EINVAL;
434
435         qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
436         qp->sq.wqe_shift = ucmd->log_sq_stride;
437
438         qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
439                 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
440
441         return 0;
442 }
443
444 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
445                             struct ib_qp_init_attr *init_attr,
446                             struct ib_udata *udata, int sqpn, struct mlx4_ib_qp *qp)
447 {
448         int err;
449
450         mutex_init(&qp->mutex);
451         spin_lock_init(&qp->sq.lock);
452         spin_lock_init(&qp->rq.lock);
453
454         qp->state        = IB_QPS_RESET;
455         qp->atomic_rd_en = 0;
456         qp->resp_depth   = 0;
457
458         qp->rq.head         = 0;
459         qp->rq.tail         = 0;
460         qp->sq.head         = 0;
461         qp->sq.tail         = 0;
462         qp->sq_next_wqe     = 0;
463
464         if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
465                 qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
466         else
467                 qp->sq_signal_bits = 0;
468
469         err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, !!init_attr->srq, qp);
470         if (err)
471                 goto err;
472
473         if (pd->uobject) {
474                 struct mlx4_ib_create_qp ucmd;
475
476                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
477                         err = -EFAULT;
478                         goto err;
479                 }
480
481                 qp->sq_no_prefetch = ucmd.sq_no_prefetch;
482
483                 err = set_user_sq_size(dev, qp, &ucmd);
484                 if (err)
485                         goto err;
486
487                 qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
488                                        qp->buf_size, 0, 0);
489                 if (IS_ERR(qp->umem)) {
490                         err = PTR_ERR(qp->umem);
491                         goto err;
492                 }
493
494                 err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
495                                     ilog2(qp->umem->page_size), &qp->mtt);
496                 if (err)
497                         goto err_buf;
498
499                 err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
500                 if (err)
501                         goto err_mtt;
502
503                 if (!init_attr->srq) {
504                         err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
505                                                   ucmd.db_addr, &qp->db);
506                         if (err)
507                                 goto err_mtt;
508                 }
509         } else {
510                 qp->sq_no_prefetch = 0;
511
512                 if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
513                         qp->flags |= MLX4_IB_QP_LSO;
514
515                 err = set_kernel_sq_size(dev, &init_attr->cap, init_attr->qp_type, qp);
516                 if (err)
517                         goto err;
518
519                 if (!init_attr->srq) {
520                         err = mlx4_db_alloc(dev->dev, &qp->db, 0);
521                         if (err)
522                                 goto err;
523
524                         *qp->db.db = 0;
525                 }
526
527                 if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf)) {
528                         err = -ENOMEM;
529                         goto err_db;
530                 }
531
532                 err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
533                                     &qp->mtt);
534                 if (err)
535                         goto err_buf;
536
537                 err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
538                 if (err)
539                         goto err_mtt;
540
541                 qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof (u64), GFP_KERNEL);
542                 qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof (u64), GFP_KERNEL);
543
544                 if (!qp->sq.wrid || !qp->rq.wrid) {
545                         err = -ENOMEM;
546                         goto err_wrid;
547                 }
548         }
549
550         err = mlx4_qp_alloc(dev->dev, sqpn, &qp->mqp);
551         if (err)
552                 goto err_wrid;
553
554         /*
555          * Hardware wants QPN written in big-endian order (after
556          * shifting) for send doorbell.  Precompute this value to save
557          * a little bit when posting sends.
558          */
559         qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
560
561         qp->mqp.event = mlx4_ib_qp_event;
562
563         return 0;
564
565 err_wrid:
566         if (pd->uobject) {
567                 if (!init_attr->srq)
568                         mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context),
569                                               &qp->db);
570         } else {
571                 kfree(qp->sq.wrid);
572                 kfree(qp->rq.wrid);
573         }
574
575 err_mtt:
576         mlx4_mtt_cleanup(dev->dev, &qp->mtt);
577
578 err_buf:
579         if (pd->uobject)
580                 ib_umem_release(qp->umem);
581         else
582                 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
583
584 err_db:
585         if (!pd->uobject && !init_attr->srq)
586                 mlx4_db_free(dev->dev, &qp->db);
587
588 err:
589         return err;
590 }
591
592 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
593 {
594         switch (state) {
595         case IB_QPS_RESET:      return MLX4_QP_STATE_RST;
596         case IB_QPS_INIT:       return MLX4_QP_STATE_INIT;
597         case IB_QPS_RTR:        return MLX4_QP_STATE_RTR;
598         case IB_QPS_RTS:        return MLX4_QP_STATE_RTS;
599         case IB_QPS_SQD:        return MLX4_QP_STATE_SQD;
600         case IB_QPS_SQE:        return MLX4_QP_STATE_SQER;
601         case IB_QPS_ERR:        return MLX4_QP_STATE_ERR;
602         default:                return -1;
603         }
604 }
605
606 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
607 {
608         if (send_cq == recv_cq)
609                 spin_lock_irq(&send_cq->lock);
610         else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
611                 spin_lock_irq(&send_cq->lock);
612                 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
613         } else {
614                 spin_lock_irq(&recv_cq->lock);
615                 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
616         }
617 }
618
619 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
620 {
621         if (send_cq == recv_cq)
622                 spin_unlock_irq(&send_cq->lock);
623         else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
624                 spin_unlock(&recv_cq->lock);
625                 spin_unlock_irq(&send_cq->lock);
626         } else {
627                 spin_unlock(&send_cq->lock);
628                 spin_unlock_irq(&recv_cq->lock);
629         }
630 }
631
632 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
633                               int is_user)
634 {
635         struct mlx4_ib_cq *send_cq, *recv_cq;
636
637         if (qp->state != IB_QPS_RESET)
638                 if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
639                                    MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
640                         printk(KERN_WARNING "mlx4_ib: modify QP %06x to RESET failed.\n",
641                                qp->mqp.qpn);
642
643         send_cq = to_mcq(qp->ibqp.send_cq);
644         recv_cq = to_mcq(qp->ibqp.recv_cq);
645
646         mlx4_ib_lock_cqs(send_cq, recv_cq);
647
648         if (!is_user) {
649                 __mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
650                                  qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
651                 if (send_cq != recv_cq)
652                         __mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
653         }
654
655         mlx4_qp_remove(dev->dev, &qp->mqp);
656
657         mlx4_ib_unlock_cqs(send_cq, recv_cq);
658
659         mlx4_qp_free(dev->dev, &qp->mqp);
660         mlx4_mtt_cleanup(dev->dev, &qp->mtt);
661
662         if (is_user) {
663                 if (!qp->ibqp.srq)
664                         mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
665                                               &qp->db);
666                 ib_umem_release(qp->umem);
667         } else {
668                 kfree(qp->sq.wrid);
669                 kfree(qp->rq.wrid);
670                 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
671                 if (!qp->ibqp.srq)
672                         mlx4_db_free(dev->dev, &qp->db);
673         }
674 }
675
676 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
677                                 struct ib_qp_init_attr *init_attr,
678                                 struct ib_udata *udata)
679 {
680         struct mlx4_ib_dev *dev = to_mdev(pd->device);
681         struct mlx4_ib_sqp *sqp;
682         struct mlx4_ib_qp *qp;
683         int err;
684
685         /* We only support LSO, and only for kernel UD QPs. */
686         if (init_attr->create_flags & ~IB_QP_CREATE_IPOIB_UD_LSO)
687                 return ERR_PTR(-EINVAL);
688         if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO &&
689             (pd->uobject || init_attr->qp_type != IB_QPT_UD))
690                 return ERR_PTR(-EINVAL);
691
692         switch (init_attr->qp_type) {
693         case IB_QPT_RC:
694         case IB_QPT_UC:
695         case IB_QPT_UD:
696         {
697                 qp = kmalloc(sizeof *qp, GFP_KERNEL);
698                 if (!qp)
699                         return ERR_PTR(-ENOMEM);
700
701                 err = create_qp_common(dev, pd, init_attr, udata, 0, qp);
702                 if (err) {
703                         kfree(qp);
704                         return ERR_PTR(err);
705                 }
706
707                 qp->ibqp.qp_num = qp->mqp.qpn;
708
709                 break;
710         }
711         case IB_QPT_SMI:
712         case IB_QPT_GSI:
713         {
714                 /* Userspace is not allowed to create special QPs: */
715                 if (pd->uobject)
716                         return ERR_PTR(-EINVAL);
717
718                 sqp = kmalloc(sizeof *sqp, GFP_KERNEL);
719                 if (!sqp)
720                         return ERR_PTR(-ENOMEM);
721
722                 qp = &sqp->qp;
723
724                 err = create_qp_common(dev, pd, init_attr, udata,
725                                        dev->dev->caps.sqp_start +
726                                        (init_attr->qp_type == IB_QPT_SMI ? 0 : 2) +
727                                        init_attr->port_num - 1,
728                                        qp);
729                 if (err) {
730                         kfree(sqp);
731                         return ERR_PTR(err);
732                 }
733
734                 qp->port        = init_attr->port_num;
735                 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
736
737                 break;
738         }
739         default:
740                 /* Don't support raw QPs */
741                 return ERR_PTR(-EINVAL);
742         }
743
744         return &qp->ibqp;
745 }
746
747 int mlx4_ib_destroy_qp(struct ib_qp *qp)
748 {
749         struct mlx4_ib_dev *dev = to_mdev(qp->device);
750         struct mlx4_ib_qp *mqp = to_mqp(qp);
751
752         if (is_qp0(dev, mqp))
753                 mlx4_CLOSE_PORT(dev->dev, mqp->port);
754
755         destroy_qp_common(dev, mqp, !!qp->pd->uobject);
756
757         if (is_sqp(dev, mqp))
758                 kfree(to_msqp(mqp));
759         else
760                 kfree(mqp);
761
762         return 0;
763 }
764
765 static int to_mlx4_st(enum ib_qp_type type)
766 {
767         switch (type) {
768         case IB_QPT_RC:         return MLX4_QP_ST_RC;
769         case IB_QPT_UC:         return MLX4_QP_ST_UC;
770         case IB_QPT_UD:         return MLX4_QP_ST_UD;
771         case IB_QPT_SMI:
772         case IB_QPT_GSI:        return MLX4_QP_ST_MLX;
773         default:                return -1;
774         }
775 }
776
777 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
778                                    int attr_mask)
779 {
780         u8 dest_rd_atomic;
781         u32 access_flags;
782         u32 hw_access_flags = 0;
783
784         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
785                 dest_rd_atomic = attr->max_dest_rd_atomic;
786         else
787                 dest_rd_atomic = qp->resp_depth;
788
789         if (attr_mask & IB_QP_ACCESS_FLAGS)
790                 access_flags = attr->qp_access_flags;
791         else
792                 access_flags = qp->atomic_rd_en;
793
794         if (!dest_rd_atomic)
795                 access_flags &= IB_ACCESS_REMOTE_WRITE;
796
797         if (access_flags & IB_ACCESS_REMOTE_READ)
798                 hw_access_flags |= MLX4_QP_BIT_RRE;
799         if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
800                 hw_access_flags |= MLX4_QP_BIT_RAE;
801         if (access_flags & IB_ACCESS_REMOTE_WRITE)
802                 hw_access_flags |= MLX4_QP_BIT_RWE;
803
804         return cpu_to_be32(hw_access_flags);
805 }
806
807 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
808                             int attr_mask)
809 {
810         if (attr_mask & IB_QP_PKEY_INDEX)
811                 sqp->pkey_index = attr->pkey_index;
812         if (attr_mask & IB_QP_QKEY)
813                 sqp->qkey = attr->qkey;
814         if (attr_mask & IB_QP_SQ_PSN)
815                 sqp->send_psn = attr->sq_psn;
816 }
817
818 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
819 {
820         path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
821 }
822
823 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
824                          struct mlx4_qp_path *path, u8 port)
825 {
826         path->grh_mylmc     = ah->src_path_bits & 0x7f;
827         path->rlid          = cpu_to_be16(ah->dlid);
828         if (ah->static_rate) {
829                 path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
830                 while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
831                        !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
832                         --path->static_rate;
833         } else
834                 path->static_rate = 0;
835         path->counter_index = 0xff;
836
837         if (ah->ah_flags & IB_AH_GRH) {
838                 if (ah->grh.sgid_index >= dev->dev->caps.gid_table_len[port]) {
839                         printk(KERN_ERR "sgid_index (%u) too large. max is %d\n",
840                                ah->grh.sgid_index, dev->dev->caps.gid_table_len[port] - 1);
841                         return -1;
842                 }
843
844                 path->grh_mylmc |= 1 << 7;
845                 path->mgid_index = ah->grh.sgid_index;
846                 path->hop_limit  = ah->grh.hop_limit;
847                 path->tclass_flowlabel =
848                         cpu_to_be32((ah->grh.traffic_class << 20) |
849                                     (ah->grh.flow_label));
850                 memcpy(path->rgid, ah->grh.dgid.raw, 16);
851         }
852
853         path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
854                 ((port - 1) << 6) | ((ah->sl & 0xf) << 2);
855
856         return 0;
857 }
858
859 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
860                                const struct ib_qp_attr *attr, int attr_mask,
861                                enum ib_qp_state cur_state, enum ib_qp_state new_state)
862 {
863         struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
864         struct mlx4_ib_qp *qp = to_mqp(ibqp);
865         struct mlx4_qp_context *context;
866         enum mlx4_qp_optpar optpar = 0;
867         int sqd_event;
868         int err = -EINVAL;
869
870         context = kzalloc(sizeof *context, GFP_KERNEL);
871         if (!context)
872                 return -ENOMEM;
873
874         context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
875                                      (to_mlx4_st(ibqp->qp_type) << 16));
876         context->flags     |= cpu_to_be32(1 << 8); /* DE? */
877
878         if (!(attr_mask & IB_QP_PATH_MIG_STATE))
879                 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
880         else {
881                 optpar |= MLX4_QP_OPTPAR_PM_STATE;
882                 switch (attr->path_mig_state) {
883                 case IB_MIG_MIGRATED:
884                         context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
885                         break;
886                 case IB_MIG_REARM:
887                         context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
888                         break;
889                 case IB_MIG_ARMED:
890                         context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
891                         break;
892                 }
893         }
894
895         if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
896                 context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
897         else if (ibqp->qp_type == IB_QPT_UD) {
898                 if (qp->flags & MLX4_IB_QP_LSO)
899                         context->mtu_msgmax = (IB_MTU_4096 << 5) |
900                                               ilog2(dev->dev->caps.max_gso_sz);
901                 else
902                         context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
903         } else if (attr_mask & IB_QP_PATH_MTU) {
904                 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
905                         printk(KERN_ERR "path MTU (%u) is invalid\n",
906                                attr->path_mtu);
907                         goto out;
908                 }
909                 context->mtu_msgmax = (attr->path_mtu << 5) | 31;
910         }
911
912         if (qp->rq.wqe_cnt)
913                 context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
914         context->rq_size_stride |= qp->rq.wqe_shift - 4;
915
916         if (qp->sq.wqe_cnt)
917                 context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
918         context->sq_size_stride |= qp->sq.wqe_shift - 4;
919
920         if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
921                 context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
922
923         if (qp->ibqp.uobject)
924                 context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
925         else
926                 context->usr_page = cpu_to_be32(dev->priv_uar.index);
927
928         if (attr_mask & IB_QP_DEST_QPN)
929                 context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
930
931         if (attr_mask & IB_QP_PORT) {
932                 if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
933                     !(attr_mask & IB_QP_AV)) {
934                         mlx4_set_sched(&context->pri_path, attr->port_num);
935                         optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
936                 }
937         }
938
939         if (attr_mask & IB_QP_PKEY_INDEX) {
940                 context->pri_path.pkey_index = attr->pkey_index;
941                 optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
942         }
943
944         if (attr_mask & IB_QP_AV) {
945                 if (mlx4_set_path(dev, &attr->ah_attr, &context->pri_path,
946                                   attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
947                         goto out;
948
949                 optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
950                            MLX4_QP_OPTPAR_SCHED_QUEUE);
951         }
952
953         if (attr_mask & IB_QP_TIMEOUT) {
954                 context->pri_path.ackto = attr->timeout << 3;
955                 optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
956         }
957
958         if (attr_mask & IB_QP_ALT_PATH) {
959                 if (attr->alt_port_num == 0 ||
960                     attr->alt_port_num > dev->dev->caps.num_ports)
961                         goto out;
962
963                 if (attr->alt_pkey_index >=
964                     dev->dev->caps.pkey_table_len[attr->alt_port_num])
965                         goto out;
966
967                 if (mlx4_set_path(dev, &attr->alt_ah_attr, &context->alt_path,
968                                   attr->alt_port_num))
969                         goto out;
970
971                 context->alt_path.pkey_index = attr->alt_pkey_index;
972                 context->alt_path.ackto = attr->alt_timeout << 3;
973                 optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
974         }
975
976         context->pd         = cpu_to_be32(to_mpd(ibqp->pd)->pdn);
977         context->params1    = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
978
979         if (attr_mask & IB_QP_RNR_RETRY) {
980                 context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
981                 optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
982         }
983
984         if (attr_mask & IB_QP_RETRY_CNT) {
985                 context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
986                 optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
987         }
988
989         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
990                 if (attr->max_rd_atomic)
991                         context->params1 |=
992                                 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
993                 optpar |= MLX4_QP_OPTPAR_SRA_MAX;
994         }
995
996         if (attr_mask & IB_QP_SQ_PSN)
997                 context->next_send_psn = cpu_to_be32(attr->sq_psn);
998
999         context->cqn_send = cpu_to_be32(to_mcq(ibqp->send_cq)->mcq.cqn);
1000
1001         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1002                 if (attr->max_dest_rd_atomic)
1003                         context->params2 |=
1004                                 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1005                 optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1006         }
1007
1008         if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1009                 context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1010                 optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1011         }
1012
1013         if (ibqp->srq)
1014                 context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1015
1016         if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1017                 context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1018                 optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1019         }
1020         if (attr_mask & IB_QP_RQ_PSN)
1021                 context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1022
1023         context->cqn_recv = cpu_to_be32(to_mcq(ibqp->recv_cq)->mcq.cqn);
1024
1025         if (attr_mask & IB_QP_QKEY) {
1026                 context->qkey = cpu_to_be32(attr->qkey);
1027                 optpar |= MLX4_QP_OPTPAR_Q_KEY;
1028         }
1029
1030         if (ibqp->srq)
1031                 context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1032
1033         if (!ibqp->srq && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1034                 context->db_rec_addr = cpu_to_be64(qp->db.dma);
1035
1036         if (cur_state == IB_QPS_INIT &&
1037             new_state == IB_QPS_RTR  &&
1038             (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1039              ibqp->qp_type == IB_QPT_UD)) {
1040                 context->pri_path.sched_queue = (qp->port - 1) << 6;
1041                 if (is_qp0(dev, qp))
1042                         context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1043                 else
1044                         context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1045         }
1046
1047         if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD  &&
1048             attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1049                 sqd_event = 1;
1050         else
1051                 sqd_event = 0;
1052
1053         /*
1054          * Before passing a kernel QP to the HW, make sure that the
1055          * ownership bits of the send queue are set and the SQ
1056          * headroom is stamped so that the hardware doesn't start
1057          * processing stale work requests.
1058          */
1059         if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1060                 struct mlx4_wqe_ctrl_seg *ctrl;
1061                 int i;
1062
1063                 for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1064                         ctrl = get_send_wqe(qp, i);
1065                         ctrl->owner_opcode = cpu_to_be32(1 << 31);
1066
1067                         stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1068                 }
1069         }
1070
1071         err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
1072                              to_mlx4_state(new_state), context, optpar,
1073                              sqd_event, &qp->mqp);
1074         if (err)
1075                 goto out;
1076
1077         qp->state = new_state;
1078
1079         if (attr_mask & IB_QP_ACCESS_FLAGS)
1080                 qp->atomic_rd_en = attr->qp_access_flags;
1081         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1082                 qp->resp_depth = attr->max_dest_rd_atomic;
1083         if (attr_mask & IB_QP_PORT)
1084                 qp->port = attr->port_num;
1085         if (attr_mask & IB_QP_ALT_PATH)
1086                 qp->alt_port = attr->alt_port_num;
1087
1088         if (is_sqp(dev, qp))
1089                 store_sqp_attrs(to_msqp(qp), attr, attr_mask);
1090
1091         /*
1092          * If we moved QP0 to RTR, bring the IB link up; if we moved
1093          * QP0 to RESET or ERROR, bring the link back down.
1094          */
1095         if (is_qp0(dev, qp)) {
1096                 if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
1097                         if (mlx4_INIT_PORT(dev->dev, qp->port))
1098                                 printk(KERN_WARNING "INIT_PORT failed for port %d\n",
1099                                        qp->port);
1100
1101                 if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1102                     (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1103                         mlx4_CLOSE_PORT(dev->dev, qp->port);
1104         }
1105
1106         /*
1107          * If we moved a kernel QP to RESET, clean up all old CQ
1108          * entries and reinitialize the QP.
1109          */
1110         if (new_state == IB_QPS_RESET && !ibqp->uobject) {
1111                 mlx4_ib_cq_clean(to_mcq(ibqp->recv_cq), qp->mqp.qpn,
1112                                  ibqp->srq ? to_msrq(ibqp->srq): NULL);
1113                 if (ibqp->send_cq != ibqp->recv_cq)
1114                         mlx4_ib_cq_clean(to_mcq(ibqp->send_cq), qp->mqp.qpn, NULL);
1115
1116                 qp->rq.head = 0;
1117                 qp->rq.tail = 0;
1118                 qp->sq.head = 0;
1119                 qp->sq.tail = 0;
1120                 qp->sq_next_wqe = 0;
1121                 if (!ibqp->srq)
1122                         *qp->db.db  = 0;
1123         }
1124
1125 out:
1126         kfree(context);
1127         return err;
1128 }
1129
1130 static const struct ib_qp_attr mlx4_ib_qp_attr = { .port_num = 1 };
1131 static const int mlx4_ib_qp_attr_mask_table[IB_QPT_UD + 1] = {
1132                 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX                |
1133                                 IB_QP_PORT                      |
1134                                 IB_QP_QKEY),
1135                 [IB_QPT_UC]  = (IB_QP_PKEY_INDEX                |
1136                                 IB_QP_PORT                      |
1137                                 IB_QP_ACCESS_FLAGS),
1138                 [IB_QPT_RC]  = (IB_QP_PKEY_INDEX                |
1139                                 IB_QP_PORT                      |
1140                                 IB_QP_ACCESS_FLAGS),
1141                 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX                |
1142                                 IB_QP_QKEY),
1143                 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX                |
1144                                 IB_QP_QKEY),
1145 };
1146
1147 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1148                       int attr_mask, struct ib_udata *udata)
1149 {
1150         struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1151         struct mlx4_ib_qp *qp = to_mqp(ibqp);
1152         enum ib_qp_state cur_state, new_state;
1153         int err = -EINVAL;
1154
1155         mutex_lock(&qp->mutex);
1156
1157         cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1158         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1159
1160         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask))
1161                 goto out;
1162
1163         if ((attr_mask & IB_QP_PORT) &&
1164             (attr->port_num == 0 || attr->port_num > dev->dev->caps.num_ports)) {
1165                 goto out;
1166         }
1167
1168         if (attr_mask & IB_QP_PKEY_INDEX) {
1169                 int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1170                 if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p])
1171                         goto out;
1172         }
1173
1174         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1175             attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
1176                 goto out;
1177         }
1178
1179         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1180             attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
1181                 goto out;
1182         }
1183
1184         if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1185                 err = 0;
1186                 goto out;
1187         }
1188
1189         if (cur_state == IB_QPS_RESET && new_state == IB_QPS_ERR) {
1190                 err = __mlx4_ib_modify_qp(ibqp, &mlx4_ib_qp_attr,
1191                                           mlx4_ib_qp_attr_mask_table[ibqp->qp_type],
1192                                           IB_QPS_RESET, IB_QPS_INIT);
1193                 if (err)
1194                         goto out;
1195                 cur_state = IB_QPS_INIT;
1196         }
1197
1198         err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
1199
1200 out:
1201         mutex_unlock(&qp->mutex);
1202         return err;
1203 }
1204
1205 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
1206                             void *wqe, unsigned *mlx_seg_len)
1207 {
1208         struct ib_device *ib_dev = &to_mdev(sqp->qp.ibqp.device)->ib_dev;
1209         struct mlx4_wqe_mlx_seg *mlx = wqe;
1210         struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
1211         struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
1212         u16 pkey;
1213         int send_size;
1214         int header_size;
1215         int spc;
1216         int i;
1217
1218         send_size = 0;
1219         for (i = 0; i < wr->num_sge; ++i)
1220                 send_size += wr->sg_list[i].length;
1221
1222         ib_ud_header_init(send_size, mlx4_ib_ah_grh_present(ah), &sqp->ud_header);
1223
1224         sqp->ud_header.lrh.service_level   =
1225                 be32_to_cpu(ah->av.sl_tclass_flowlabel) >> 28;
1226         sqp->ud_header.lrh.destination_lid = ah->av.dlid;
1227         sqp->ud_header.lrh.source_lid      = cpu_to_be16(ah->av.g_slid & 0x7f);
1228         if (mlx4_ib_ah_grh_present(ah)) {
1229                 sqp->ud_header.grh.traffic_class =
1230                         (be32_to_cpu(ah->av.sl_tclass_flowlabel) >> 20) & 0xff;
1231                 sqp->ud_header.grh.flow_label    =
1232                         ah->av.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
1233                 sqp->ud_header.grh.hop_limit     = ah->av.hop_limit;
1234                 ib_get_cached_gid(ib_dev, be32_to_cpu(ah->av.port_pd) >> 24,
1235                                   ah->av.gid_index, &sqp->ud_header.grh.source_gid);
1236                 memcpy(sqp->ud_header.grh.destination_gid.raw,
1237                        ah->av.dgid, 16);
1238         }
1239
1240         mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1241         mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
1242                                   (sqp->ud_header.lrh.destination_lid ==
1243                                    IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
1244                                   (sqp->ud_header.lrh.service_level << 8));
1245         mlx->rlid   = sqp->ud_header.lrh.destination_lid;
1246
1247         switch (wr->opcode) {
1248         case IB_WR_SEND:
1249                 sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
1250                 sqp->ud_header.immediate_present = 0;
1251                 break;
1252         case IB_WR_SEND_WITH_IMM:
1253                 sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1254                 sqp->ud_header.immediate_present = 1;
1255                 sqp->ud_header.immediate_data    = wr->ex.imm_data;
1256                 break;
1257         default:
1258                 return -EINVAL;
1259         }
1260
1261         sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1262         if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1263                 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1264         sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1265         if (!sqp->qp.ibqp.qp_num)
1266                 ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
1267         else
1268                 ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey);
1269         sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1270         sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1271         sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1272         sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1273                                                sqp->qkey : wr->wr.ud.remote_qkey);
1274         sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1275
1276         header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
1277
1278         if (0) {
1279                 printk(KERN_ERR "built UD header of size %d:\n", header_size);
1280                 for (i = 0; i < header_size / 4; ++i) {
1281                         if (i % 8 == 0)
1282                                 printk("  [%02x] ", i * 4);
1283                         printk(" %08x",
1284                                be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
1285                         if ((i + 1) % 8 == 0)
1286                                 printk("\n");
1287                 }
1288                 printk("\n");
1289         }
1290
1291         /*
1292          * Inline data segments may not cross a 64 byte boundary.  If
1293          * our UD header is bigger than the space available up to the
1294          * next 64 byte boundary in the WQE, use two inline data
1295          * segments to hold the UD header.
1296          */
1297         spc = MLX4_INLINE_ALIGN -
1298                 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
1299         if (header_size <= spc) {
1300                 inl->byte_count = cpu_to_be32(1 << 31 | header_size);
1301                 memcpy(inl + 1, sqp->header_buf, header_size);
1302                 i = 1;
1303         } else {
1304                 inl->byte_count = cpu_to_be32(1 << 31 | spc);
1305                 memcpy(inl + 1, sqp->header_buf, spc);
1306
1307                 inl = (void *) (inl + 1) + spc;
1308                 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
1309                 /*
1310                  * Need a barrier here to make sure all the data is
1311                  * visible before the byte_count field is set.
1312                  * Otherwise the HCA prefetcher could grab the 64-byte
1313                  * chunk with this inline segment and get a valid (!=
1314                  * 0xffffffff) byte count but stale data, and end up
1315                  * generating a packet with bad headers.
1316                  *
1317                  * The first inline segment's byte_count field doesn't
1318                  * need a barrier, because it comes after a
1319                  * control/MLX segment and therefore is at an offset
1320                  * of 16 mod 64.
1321                  */
1322                 wmb();
1323                 inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
1324                 i = 2;
1325         }
1326
1327         *mlx_seg_len =
1328                 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
1329         return 0;
1330 }
1331
1332 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
1333 {
1334         unsigned cur;
1335         struct mlx4_ib_cq *cq;
1336
1337         cur = wq->head - wq->tail;
1338         if (likely(cur + nreq < wq->max_post))
1339                 return 0;
1340
1341         cq = to_mcq(ib_cq);
1342         spin_lock(&cq->lock);
1343         cur = wq->head - wq->tail;
1344         spin_unlock(&cq->lock);
1345
1346         return cur + nreq >= wq->max_post;
1347 }
1348
1349 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
1350                                           u64 remote_addr, u32 rkey)
1351 {
1352         rseg->raddr    = cpu_to_be64(remote_addr);
1353         rseg->rkey     = cpu_to_be32(rkey);
1354         rseg->reserved = 0;
1355 }
1356
1357 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, struct ib_send_wr *wr)
1358 {
1359         if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1360                 aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
1361                 aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add);
1362         } else {
1363                 aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
1364                 aseg->compare  = 0;
1365         }
1366
1367 }
1368
1369 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
1370                              struct ib_send_wr *wr)
1371 {
1372         memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av));
1373         dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1374         dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1375 }
1376
1377 static void set_mlx_icrc_seg(void *dseg)
1378 {
1379         u32 *t = dseg;
1380         struct mlx4_wqe_inline_seg *iseg = dseg;
1381
1382         t[1] = 0;
1383
1384         /*
1385          * Need a barrier here before writing the byte_count field to
1386          * make sure that all the data is visible before the
1387          * byte_count field is set.  Otherwise, if the segment begins
1388          * a new cacheline, the HCA prefetcher could grab the 64-byte
1389          * chunk and get a valid (!= * 0xffffffff) byte count but
1390          * stale data, and end up sending the wrong data.
1391          */
1392         wmb();
1393
1394         iseg->byte_count = cpu_to_be32((1 << 31) | 4);
1395 }
1396
1397 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
1398 {
1399         dseg->lkey       = cpu_to_be32(sg->lkey);
1400         dseg->addr       = cpu_to_be64(sg->addr);
1401
1402         /*
1403          * Need a barrier here before writing the byte_count field to
1404          * make sure that all the data is visible before the
1405          * byte_count field is set.  Otherwise, if the segment begins
1406          * a new cacheline, the HCA prefetcher could grab the 64-byte
1407          * chunk and get a valid (!= * 0xffffffff) byte count but
1408          * stale data, and end up sending the wrong data.
1409          */
1410         wmb();
1411
1412         dseg->byte_count = cpu_to_be32(sg->length);
1413 }
1414
1415 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
1416 {
1417         dseg->byte_count = cpu_to_be32(sg->length);
1418         dseg->lkey       = cpu_to_be32(sg->lkey);
1419         dseg->addr       = cpu_to_be64(sg->addr);
1420 }
1421
1422 static int build_lso_seg(struct mlx4_lso_seg *wqe, struct ib_send_wr *wr,
1423                          struct mlx4_ib_qp *qp, unsigned *lso_seg_len)
1424 {
1425         unsigned halign = ALIGN(sizeof *wqe + wr->wr.ud.hlen, 16);
1426
1427         /*
1428          * This is a temporary limitation and will be removed in
1429          * a forthcoming FW release:
1430          */
1431         if (unlikely(halign > 64))
1432                 return -EINVAL;
1433
1434         if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
1435                      wr->num_sge > qp->sq.max_gs - (halign >> 4)))
1436                 return -EINVAL;
1437
1438         memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen);
1439
1440         /* make sure LSO header is written before overwriting stamping */
1441         wmb();
1442
1443         wqe->mss_hdr_size = cpu_to_be32((wr->wr.ud.mss - wr->wr.ud.hlen) << 16 |
1444                                         wr->wr.ud.hlen);
1445
1446         *lso_seg_len = halign;
1447         return 0;
1448 }
1449
1450 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1451                       struct ib_send_wr **bad_wr)
1452 {
1453         struct mlx4_ib_qp *qp = to_mqp(ibqp);
1454         void *wqe;
1455         struct mlx4_wqe_ctrl_seg *ctrl;
1456         struct mlx4_wqe_data_seg *dseg;
1457         unsigned long flags;
1458         int nreq;
1459         int err = 0;
1460         unsigned ind;
1461         int uninitialized_var(stamp);
1462         int uninitialized_var(size);
1463         unsigned uninitialized_var(seglen);
1464         int i;
1465
1466         spin_lock_irqsave(&qp->sq.lock, flags);
1467
1468         ind = qp->sq_next_wqe;
1469
1470         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1471                 if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1472                         err = -ENOMEM;
1473                         *bad_wr = wr;
1474                         goto out;
1475                 }
1476
1477                 if (unlikely(wr->num_sge > qp->sq.max_gs)) {
1478                         err = -EINVAL;
1479                         *bad_wr = wr;
1480                         goto out;
1481                 }
1482
1483                 ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
1484                 qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
1485
1486                 ctrl->srcrb_flags =
1487                         (wr->send_flags & IB_SEND_SIGNALED ?
1488                          cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
1489                         (wr->send_flags & IB_SEND_SOLICITED ?
1490                          cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
1491                         ((wr->send_flags & IB_SEND_IP_CSUM) ?
1492                          cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
1493                                      MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
1494                         qp->sq_signal_bits;
1495
1496                 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1497                     wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1498                         ctrl->imm = wr->ex.imm_data;
1499                 else
1500                         ctrl->imm = 0;
1501
1502                 wqe += sizeof *ctrl;
1503                 size = sizeof *ctrl / 16;
1504
1505                 switch (ibqp->qp_type) {
1506                 case IB_QPT_RC:
1507                 case IB_QPT_UC:
1508                         switch (wr->opcode) {
1509                         case IB_WR_ATOMIC_CMP_AND_SWP:
1510                         case IB_WR_ATOMIC_FETCH_AND_ADD:
1511                                 set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
1512                                               wr->wr.atomic.rkey);
1513                                 wqe  += sizeof (struct mlx4_wqe_raddr_seg);
1514
1515                                 set_atomic_seg(wqe, wr);
1516                                 wqe  += sizeof (struct mlx4_wqe_atomic_seg);
1517
1518                                 size += (sizeof (struct mlx4_wqe_raddr_seg) +
1519                                          sizeof (struct mlx4_wqe_atomic_seg)) / 16;
1520
1521                                 break;
1522
1523                         case IB_WR_RDMA_READ:
1524                         case IB_WR_RDMA_WRITE:
1525                         case IB_WR_RDMA_WRITE_WITH_IMM:
1526                                 set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
1527                                               wr->wr.rdma.rkey);
1528                                 wqe  += sizeof (struct mlx4_wqe_raddr_seg);
1529                                 size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
1530                                 break;
1531
1532                         default:
1533                                 /* No extra segments required for sends */
1534                                 break;
1535                         }
1536                         break;
1537
1538                 case IB_QPT_UD:
1539                         set_datagram_seg(wqe, wr);
1540                         wqe  += sizeof (struct mlx4_wqe_datagram_seg);
1541                         size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
1542
1543                         if (wr->opcode == IB_WR_LSO) {
1544                                 err = build_lso_seg(wqe, wr, qp, &seglen);
1545                                 if (unlikely(err)) {
1546                                         *bad_wr = wr;
1547                                         goto out;
1548                                 }
1549                                 wqe  += seglen;
1550                                 size += seglen / 16;
1551                         }
1552                         break;
1553
1554                 case IB_QPT_SMI:
1555                 case IB_QPT_GSI:
1556                         err = build_mlx_header(to_msqp(qp), wr, ctrl, &seglen);
1557                         if (unlikely(err)) {
1558                                 *bad_wr = wr;
1559                                 goto out;
1560                         }
1561                         wqe  += seglen;
1562                         size += seglen / 16;
1563                         break;
1564
1565                 default:
1566                         break;
1567                 }
1568
1569                 /*
1570                  * Write data segments in reverse order, so as to
1571                  * overwrite cacheline stamp last within each
1572                  * cacheline.  This avoids issues with WQE
1573                  * prefetching.
1574                  */
1575
1576                 dseg = wqe;
1577                 dseg += wr->num_sge - 1;
1578                 size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
1579
1580                 /* Add one more inline data segment for ICRC for MLX sends */
1581                 if (unlikely(qp->ibqp.qp_type == IB_QPT_SMI ||
1582                              qp->ibqp.qp_type == IB_QPT_GSI)) {
1583                         set_mlx_icrc_seg(dseg + 1);
1584                         size += sizeof (struct mlx4_wqe_data_seg) / 16;
1585                 }
1586
1587                 for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
1588                         set_data_seg(dseg, wr->sg_list + i);
1589
1590                 ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
1591                                     MLX4_WQE_CTRL_FENCE : 0) | size;
1592
1593                 /*
1594                  * Make sure descriptor is fully written before
1595                  * setting ownership bit (because HW can start
1596                  * executing as soon as we do).
1597                  */
1598                 wmb();
1599
1600                 if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
1601                         err = -EINVAL;
1602                         goto out;
1603                 }
1604
1605                 ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
1606                         (ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
1607
1608                 stamp = ind + qp->sq_spare_wqes;
1609                 ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
1610
1611                 /*
1612                  * We can improve latency by not stamping the last
1613                  * send queue WQE until after ringing the doorbell, so
1614                  * only stamp here if there are still more WQEs to post.
1615                  *
1616                  * Same optimization applies to padding with NOP wqe
1617                  * in case of WQE shrinking (used to prevent wrap-around
1618                  * in the middle of WR).
1619                  */
1620                 if (wr->next) {
1621                         stamp_send_wqe(qp, stamp, size * 16);
1622                         ind = pad_wraparound(qp, ind);
1623                 }
1624
1625         }
1626
1627 out:
1628         if (likely(nreq)) {
1629                 qp->sq.head += nreq;
1630
1631                 /*
1632                  * Make sure that descriptors are written before
1633                  * doorbell record.
1634                  */
1635                 wmb();
1636
1637                 writel(qp->doorbell_qpn,
1638                        to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
1639
1640                 /*
1641                  * Make sure doorbells don't leak out of SQ spinlock
1642                  * and reach the HCA out of order.
1643                  */
1644                 mmiowb();
1645
1646                 stamp_send_wqe(qp, stamp, size * 16);
1647
1648                 ind = pad_wraparound(qp, ind);
1649                 qp->sq_next_wqe = ind;
1650         }
1651
1652         spin_unlock_irqrestore(&qp->sq.lock, flags);
1653
1654         return err;
1655 }
1656
1657 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1658                       struct ib_recv_wr **bad_wr)
1659 {
1660         struct mlx4_ib_qp *qp = to_mqp(ibqp);
1661         struct mlx4_wqe_data_seg *scat;
1662         unsigned long flags;
1663         int err = 0;
1664         int nreq;
1665         int ind;
1666         int i;
1667
1668         spin_lock_irqsave(&qp->rq.lock, flags);
1669
1670         ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
1671
1672         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1673                 if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.send_cq)) {
1674                         err = -ENOMEM;
1675                         *bad_wr = wr;
1676                         goto out;
1677                 }
1678
1679                 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1680                         err = -EINVAL;
1681                         *bad_wr = wr;
1682                         goto out;
1683                 }
1684
1685                 scat = get_recv_wqe(qp, ind);
1686
1687                 for (i = 0; i < wr->num_sge; ++i)
1688                         __set_data_seg(scat + i, wr->sg_list + i);
1689
1690                 if (i < qp->rq.max_gs) {
1691                         scat[i].byte_count = 0;
1692                         scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
1693                         scat[i].addr       = 0;
1694                 }
1695
1696                 qp->rq.wrid[ind] = wr->wr_id;
1697
1698                 ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
1699         }
1700
1701 out:
1702         if (likely(nreq)) {
1703                 qp->rq.head += nreq;
1704
1705                 /*
1706                  * Make sure that descriptors are written before
1707                  * doorbell record.
1708                  */
1709                 wmb();
1710
1711                 *qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
1712         }
1713
1714         spin_unlock_irqrestore(&qp->rq.lock, flags);
1715
1716         return err;
1717 }
1718
1719 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
1720 {
1721         switch (mlx4_state) {
1722         case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
1723         case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
1724         case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
1725         case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
1726         case MLX4_QP_STATE_SQ_DRAINING:
1727         case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
1728         case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
1729         case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
1730         default:                     return -1;
1731         }
1732 }
1733
1734 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
1735 {
1736         switch (mlx4_mig_state) {
1737         case MLX4_QP_PM_ARMED:          return IB_MIG_ARMED;
1738         case MLX4_QP_PM_REARM:          return IB_MIG_REARM;
1739         case MLX4_QP_PM_MIGRATED:       return IB_MIG_MIGRATED;
1740         default: return -1;
1741         }
1742 }
1743
1744 static int to_ib_qp_access_flags(int mlx4_flags)
1745 {
1746         int ib_flags = 0;
1747
1748         if (mlx4_flags & MLX4_QP_BIT_RRE)
1749                 ib_flags |= IB_ACCESS_REMOTE_READ;
1750         if (mlx4_flags & MLX4_QP_BIT_RWE)
1751                 ib_flags |= IB_ACCESS_REMOTE_WRITE;
1752         if (mlx4_flags & MLX4_QP_BIT_RAE)
1753                 ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
1754
1755         return ib_flags;
1756 }
1757
1758 static void to_ib_ah_attr(struct mlx4_dev *dev, struct ib_ah_attr *ib_ah_attr,
1759                                 struct mlx4_qp_path *path)
1760 {
1761         memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
1762         ib_ah_attr->port_num      = path->sched_queue & 0x40 ? 2 : 1;
1763
1764         if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
1765                 return;
1766
1767         ib_ah_attr->dlid          = be16_to_cpu(path->rlid);
1768         ib_ah_attr->sl            = (path->sched_queue >> 2) & 0xf;
1769         ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
1770         ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
1771         ib_ah_attr->ah_flags      = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
1772         if (ib_ah_attr->ah_flags) {
1773                 ib_ah_attr->grh.sgid_index = path->mgid_index;
1774                 ib_ah_attr->grh.hop_limit  = path->hop_limit;
1775                 ib_ah_attr->grh.traffic_class =
1776                         (be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
1777                 ib_ah_attr->grh.flow_label =
1778                         be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
1779                 memcpy(ib_ah_attr->grh.dgid.raw,
1780                         path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
1781         }
1782 }
1783
1784 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
1785                      struct ib_qp_init_attr *qp_init_attr)
1786 {
1787         struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1788         struct mlx4_ib_qp *qp = to_mqp(ibqp);
1789         struct mlx4_qp_context context;
1790         int mlx4_state;
1791         int err = 0;
1792
1793         mutex_lock(&qp->mutex);
1794
1795         if (qp->state == IB_QPS_RESET) {
1796                 qp_attr->qp_state = IB_QPS_RESET;
1797                 goto done;
1798         }
1799
1800         err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
1801         if (err) {
1802                 err = -EINVAL;
1803                 goto out;
1804         }
1805
1806         mlx4_state = be32_to_cpu(context.flags) >> 28;
1807
1808         qp->state                    = to_ib_qp_state(mlx4_state);
1809         qp_attr->qp_state            = qp->state;
1810         qp_attr->path_mtu            = context.mtu_msgmax >> 5;
1811         qp_attr->path_mig_state      =
1812                 to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
1813         qp_attr->qkey                = be32_to_cpu(context.qkey);
1814         qp_attr->rq_psn              = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
1815         qp_attr->sq_psn              = be32_to_cpu(context.next_send_psn) & 0xffffff;
1816         qp_attr->dest_qp_num         = be32_to_cpu(context.remote_qpn) & 0xffffff;
1817         qp_attr->qp_access_flags     =
1818                 to_ib_qp_access_flags(be32_to_cpu(context.params2));
1819
1820         if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
1821                 to_ib_ah_attr(dev->dev, &qp_attr->ah_attr, &context.pri_path);
1822                 to_ib_ah_attr(dev->dev, &qp_attr->alt_ah_attr, &context.alt_path);
1823                 qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
1824                 qp_attr->alt_port_num   = qp_attr->alt_ah_attr.port_num;
1825         }
1826
1827         qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
1828         if (qp_attr->qp_state == IB_QPS_INIT)
1829                 qp_attr->port_num = qp->port;
1830         else
1831                 qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
1832
1833         /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
1834         qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
1835
1836         qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
1837
1838         qp_attr->max_dest_rd_atomic =
1839                 1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
1840         qp_attr->min_rnr_timer      =
1841                 (be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
1842         qp_attr->timeout            = context.pri_path.ackto >> 3;
1843         qp_attr->retry_cnt          = (be32_to_cpu(context.params1) >> 16) & 0x7;
1844         qp_attr->rnr_retry          = (be32_to_cpu(context.params1) >> 13) & 0x7;
1845         qp_attr->alt_timeout        = context.alt_path.ackto >> 3;
1846
1847 done:
1848         qp_attr->cur_qp_state        = qp_attr->qp_state;
1849         qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
1850         qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
1851
1852         if (!ibqp->uobject) {
1853                 qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
1854                 qp_attr->cap.max_send_sge = qp->sq.max_gs;
1855         } else {
1856                 qp_attr->cap.max_send_wr  = 0;
1857                 qp_attr->cap.max_send_sge = 0;
1858         }
1859
1860         /*
1861          * We don't support inline sends for kernel QPs (yet), and we
1862          * don't know what userspace's value should be.
1863          */
1864         qp_attr->cap.max_inline_data = 0;
1865
1866         qp_init_attr->cap            = qp_attr->cap;
1867
1868 out:
1869         mutex_unlock(&qp->mutex);
1870         return err;
1871 }
1872