219b10397b4d0fd64f19767adbdb3e24352f1476
[linux-2.6.git] / drivers / infiniband / hw / mlx4 / qp.c
1 /*
2  * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3  * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33
34 #include <linux/log2.h>
35
36 #include <rdma/ib_cache.h>
37 #include <rdma/ib_pack.h>
38
39 #include <linux/mlx4/qp.h>
40
41 #include "mlx4_ib.h"
42 #include "user.h"
43
44 enum {
45         MLX4_IB_ACK_REQ_FREQ    = 8,
46 };
47
48 enum {
49         MLX4_IB_DEFAULT_SCHED_QUEUE     = 0x83,
50         MLX4_IB_DEFAULT_QP0_SCHED_QUEUE = 0x3f
51 };
52
53 enum {
54         /*
55          * Largest possible UD header: send with GRH and immediate data.
56          */
57         MLX4_IB_UD_HEADER_SIZE          = 72
58 };
59
60 struct mlx4_ib_sqp {
61         struct mlx4_ib_qp       qp;
62         int                     pkey_index;
63         u32                     qkey;
64         u32                     send_psn;
65         struct ib_ud_header     ud_header;
66         u8                      header_buf[MLX4_IB_UD_HEADER_SIZE];
67 };
68
69 enum {
70         MLX4_IB_MIN_SQ_STRIDE = 6
71 };
72
73 static const __be32 mlx4_ib_opcode[] = {
74         [IB_WR_SEND]                    = cpu_to_be32(MLX4_OPCODE_SEND),
75         [IB_WR_LSO]                     = cpu_to_be32(MLX4_OPCODE_LSO),
76         [IB_WR_SEND_WITH_IMM]           = cpu_to_be32(MLX4_OPCODE_SEND_IMM),
77         [IB_WR_RDMA_WRITE]              = cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
78         [IB_WR_RDMA_WRITE_WITH_IMM]     = cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
79         [IB_WR_RDMA_READ]               = cpu_to_be32(MLX4_OPCODE_RDMA_READ),
80         [IB_WR_ATOMIC_CMP_AND_SWP]      = cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
81         [IB_WR_ATOMIC_FETCH_AND_ADD]    = cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
82         [IB_WR_SEND_WITH_INV]           = cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
83         [IB_WR_LOCAL_INV]               = cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
84         [IB_WR_FAST_REG_MR]             = cpu_to_be32(MLX4_OPCODE_FMR),
85 };
86
87 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
88 {
89         return container_of(mqp, struct mlx4_ib_sqp, qp);
90 }
91
92 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
93 {
94         return qp->mqp.qpn >= dev->dev->caps.sqp_start &&
95                 qp->mqp.qpn <= dev->dev->caps.sqp_start + 3;
96 }
97
98 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
99 {
100         return qp->mqp.qpn >= dev->dev->caps.sqp_start &&
101                 qp->mqp.qpn <= dev->dev->caps.sqp_start + 1;
102 }
103
104 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
105 {
106         return mlx4_buf_offset(&qp->buf, offset);
107 }
108
109 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
110 {
111         return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
112 }
113
114 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
115 {
116         return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
117 }
118
119 /*
120  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
121  * first four bytes of every 64 byte chunk with
122  *     0x7FFFFFF | (invalid_ownership_value << 31).
123  *
124  * When the max work request size is less than or equal to the WQE
125  * basic block size, as an optimization, we can stamp all WQEs with
126  * 0xffffffff, and skip the very first chunk of each WQE.
127  */
128 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
129 {
130         __be32 *wqe;
131         int i;
132         int s;
133         int ind;
134         void *buf;
135         __be32 stamp;
136         struct mlx4_wqe_ctrl_seg *ctrl;
137
138         if (qp->sq_max_wqes_per_wr > 1) {
139                 s = roundup(size, 1U << qp->sq.wqe_shift);
140                 for (i = 0; i < s; i += 64) {
141                         ind = (i >> qp->sq.wqe_shift) + n;
142                         stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
143                                                        cpu_to_be32(0xffffffff);
144                         buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
145                         wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
146                         *wqe = stamp;
147                 }
148         } else {
149                 ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
150                 s = (ctrl->fence_size & 0x3f) << 4;
151                 for (i = 64; i < s; i += 64) {
152                         wqe = buf + i;
153                         *wqe = cpu_to_be32(0xffffffff);
154                 }
155         }
156 }
157
158 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
159 {
160         struct mlx4_wqe_ctrl_seg *ctrl;
161         struct mlx4_wqe_inline_seg *inl;
162         void *wqe;
163         int s;
164
165         ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
166         s = sizeof(struct mlx4_wqe_ctrl_seg);
167
168         if (qp->ibqp.qp_type == IB_QPT_UD) {
169                 struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
170                 struct mlx4_av *av = (struct mlx4_av *)dgram->av;
171                 memset(dgram, 0, sizeof *dgram);
172                 av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
173                 s += sizeof(struct mlx4_wqe_datagram_seg);
174         }
175
176         /* Pad the remainder of the WQE with an inline data segment. */
177         if (size > s) {
178                 inl = wqe + s;
179                 inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
180         }
181         ctrl->srcrb_flags = 0;
182         ctrl->fence_size = size / 16;
183         /*
184          * Make sure descriptor is fully written before setting ownership bit
185          * (because HW can start executing as soon as we do).
186          */
187         wmb();
188
189         ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
190                 (n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
191
192         stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
193 }
194
195 /* Post NOP WQE to prevent wrap-around in the middle of WR */
196 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
197 {
198         unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
199         if (unlikely(s < qp->sq_max_wqes_per_wr)) {
200                 post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
201                 ind += s;
202         }
203         return ind;
204 }
205
206 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
207 {
208         struct ib_event event;
209         struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
210
211         if (type == MLX4_EVENT_TYPE_PATH_MIG)
212                 to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
213
214         if (ibqp->event_handler) {
215                 event.device     = ibqp->device;
216                 event.element.qp = ibqp;
217                 switch (type) {
218                 case MLX4_EVENT_TYPE_PATH_MIG:
219                         event.event = IB_EVENT_PATH_MIG;
220                         break;
221                 case MLX4_EVENT_TYPE_COMM_EST:
222                         event.event = IB_EVENT_COMM_EST;
223                         break;
224                 case MLX4_EVENT_TYPE_SQ_DRAINED:
225                         event.event = IB_EVENT_SQ_DRAINED;
226                         break;
227                 case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
228                         event.event = IB_EVENT_QP_LAST_WQE_REACHED;
229                         break;
230                 case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
231                         event.event = IB_EVENT_QP_FATAL;
232                         break;
233                 case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
234                         event.event = IB_EVENT_PATH_MIG_ERR;
235                         break;
236                 case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
237                         event.event = IB_EVENT_QP_REQ_ERR;
238                         break;
239                 case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
240                         event.event = IB_EVENT_QP_ACCESS_ERR;
241                         break;
242                 default:
243                         printk(KERN_WARNING "mlx4_ib: Unexpected event type %d "
244                                "on QP %06x\n", type, qp->qpn);
245                         return;
246                 }
247
248                 ibqp->event_handler(&event, ibqp->qp_context);
249         }
250 }
251
252 static int send_wqe_overhead(enum ib_qp_type type, u32 flags)
253 {
254         /*
255          * UD WQEs must have a datagram segment.
256          * RC and UC WQEs might have a remote address segment.
257          * MLX WQEs need two extra inline data segments (for the UD
258          * header and space for the ICRC).
259          */
260         switch (type) {
261         case IB_QPT_UD:
262                 return sizeof (struct mlx4_wqe_ctrl_seg) +
263                         sizeof (struct mlx4_wqe_datagram_seg) +
264                         ((flags & MLX4_IB_QP_LSO) ? 64 : 0);
265         case IB_QPT_UC:
266                 return sizeof (struct mlx4_wqe_ctrl_seg) +
267                         sizeof (struct mlx4_wqe_raddr_seg);
268         case IB_QPT_RC:
269                 return sizeof (struct mlx4_wqe_ctrl_seg) +
270                         sizeof (struct mlx4_wqe_atomic_seg) +
271                         sizeof (struct mlx4_wqe_raddr_seg);
272         case IB_QPT_SMI:
273         case IB_QPT_GSI:
274                 return sizeof (struct mlx4_wqe_ctrl_seg) +
275                         ALIGN(MLX4_IB_UD_HEADER_SIZE +
276                               DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
277                                            MLX4_INLINE_ALIGN) *
278                               sizeof (struct mlx4_wqe_inline_seg),
279                               sizeof (struct mlx4_wqe_data_seg)) +
280                         ALIGN(4 +
281                               sizeof (struct mlx4_wqe_inline_seg),
282                               sizeof (struct mlx4_wqe_data_seg));
283         default:
284                 return sizeof (struct mlx4_wqe_ctrl_seg);
285         }
286 }
287
288 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
289                        int is_user, int has_srq, struct mlx4_ib_qp *qp)
290 {
291         /* Sanity check RQ size before proceeding */
292         if (cap->max_recv_wr  > dev->dev->caps.max_wqes  ||
293             cap->max_recv_sge > dev->dev->caps.max_rq_sg)
294                 return -EINVAL;
295
296         if (has_srq) {
297                 /* QPs attached to an SRQ should have no RQ */
298                 if (cap->max_recv_wr)
299                         return -EINVAL;
300
301                 qp->rq.wqe_cnt = qp->rq.max_gs = 0;
302         } else {
303                 /* HW requires >= 1 RQ entry with >= 1 gather entry */
304                 if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
305                         return -EINVAL;
306
307                 qp->rq.wqe_cnt   = roundup_pow_of_two(max(1U, cap->max_recv_wr));
308                 qp->rq.max_gs    = roundup_pow_of_two(max(1U, cap->max_recv_sge));
309                 qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
310         }
311
312         cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
313         cap->max_recv_sge = qp->rq.max_gs;
314
315         return 0;
316 }
317
318 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
319                               enum ib_qp_type type, struct mlx4_ib_qp *qp)
320 {
321         int s;
322
323         /* Sanity check SQ size before proceeding */
324         if (cap->max_send_wr     > dev->dev->caps.max_wqes  ||
325             cap->max_send_sge    > dev->dev->caps.max_sq_sg ||
326             cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
327             sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
328                 return -EINVAL;
329
330         /*
331          * For MLX transport we need 2 extra S/G entries:
332          * one for the header and one for the checksum at the end
333          */
334         if ((type == IB_QPT_SMI || type == IB_QPT_GSI) &&
335             cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
336                 return -EINVAL;
337
338         s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
339                 cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
340                 send_wqe_overhead(type, qp->flags);
341
342         if (s > dev->dev->caps.max_sq_desc_sz)
343                 return -EINVAL;
344
345         /*
346          * Hermon supports shrinking WQEs, such that a single work
347          * request can include multiple units of 1 << wqe_shift.  This
348          * way, work requests can differ in size, and do not have to
349          * be a power of 2 in size, saving memory and speeding up send
350          * WR posting.  Unfortunately, if we do this then the
351          * wqe_index field in CQEs can't be used to look up the WR ID
352          * anymore, so we do this only if selective signaling is off.
353          *
354          * Further, on 32-bit platforms, we can't use vmap() to make
355          * the QP buffer virtually contigious.  Thus we have to use
356          * constant-sized WRs to make sure a WR is always fully within
357          * a single page-sized chunk.
358          *
359          * Finally, we use NOP work requests to pad the end of the
360          * work queue, to avoid wrap-around in the middle of WR.  We
361          * set NEC bit to avoid getting completions with error for
362          * these NOP WRs, but since NEC is only supported starting
363          * with firmware 2.2.232, we use constant-sized WRs for older
364          * firmware.
365          *
366          * And, since MLX QPs only support SEND, we use constant-sized
367          * WRs in this case.
368          *
369          * We look for the smallest value of wqe_shift such that the
370          * resulting number of wqes does not exceed device
371          * capabilities.
372          *
373          * We set WQE size to at least 64 bytes, this way stamping
374          * invalidates each WQE.
375          */
376         if (dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
377             qp->sq_signal_bits && BITS_PER_LONG == 64 &&
378             type != IB_QPT_SMI && type != IB_QPT_GSI)
379                 qp->sq.wqe_shift = ilog2(64);
380         else
381                 qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
382
383         for (;;) {
384                 qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
385
386                 /*
387                  * We need to leave 2 KB + 1 WR of headroom in the SQ to
388                  * allow HW to prefetch.
389                  */
390                 qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
391                 qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
392                                                     qp->sq_max_wqes_per_wr +
393                                                     qp->sq_spare_wqes);
394
395                 if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
396                         break;
397
398                 if (qp->sq_max_wqes_per_wr <= 1)
399                         return -EINVAL;
400
401                 ++qp->sq.wqe_shift;
402         }
403
404         qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
405                              (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
406                          send_wqe_overhead(type, qp->flags)) /
407                 sizeof (struct mlx4_wqe_data_seg);
408
409         qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
410                 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
411         if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
412                 qp->rq.offset = 0;
413                 qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
414         } else {
415                 qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
416                 qp->sq.offset = 0;
417         }
418
419         cap->max_send_wr  = qp->sq.max_post =
420                 (qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
421         cap->max_send_sge = min(qp->sq.max_gs,
422                                 min(dev->dev->caps.max_sq_sg,
423                                     dev->dev->caps.max_rq_sg));
424         /* We don't support inline sends for kernel QPs (yet) */
425         cap->max_inline_data = 0;
426
427         return 0;
428 }
429
430 static int set_user_sq_size(struct mlx4_ib_dev *dev,
431                             struct mlx4_ib_qp *qp,
432                             struct mlx4_ib_create_qp *ucmd)
433 {
434         /* Sanity check SQ size before proceeding */
435         if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes       ||
436             ucmd->log_sq_stride >
437                 ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
438             ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
439                 return -EINVAL;
440
441         qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
442         qp->sq.wqe_shift = ucmd->log_sq_stride;
443
444         qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
445                 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
446
447         return 0;
448 }
449
450 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
451                             struct ib_qp_init_attr *init_attr,
452                             struct ib_udata *udata, int sqpn, struct mlx4_ib_qp *qp)
453 {
454         int qpn;
455         int err;
456
457         mutex_init(&qp->mutex);
458         spin_lock_init(&qp->sq.lock);
459         spin_lock_init(&qp->rq.lock);
460
461         qp->state        = IB_QPS_RESET;
462         if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
463                 qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
464
465         err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, !!init_attr->srq, qp);
466         if (err)
467                 goto err;
468
469         if (pd->uobject) {
470                 struct mlx4_ib_create_qp ucmd;
471
472                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
473                         err = -EFAULT;
474                         goto err;
475                 }
476
477                 qp->sq_no_prefetch = ucmd.sq_no_prefetch;
478
479                 err = set_user_sq_size(dev, qp, &ucmd);
480                 if (err)
481                         goto err;
482
483                 qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
484                                        qp->buf_size, 0, 0);
485                 if (IS_ERR(qp->umem)) {
486                         err = PTR_ERR(qp->umem);
487                         goto err;
488                 }
489
490                 err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
491                                     ilog2(qp->umem->page_size), &qp->mtt);
492                 if (err)
493                         goto err_buf;
494
495                 err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
496                 if (err)
497                         goto err_mtt;
498
499                 if (!init_attr->srq) {
500                         err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
501                                                   ucmd.db_addr, &qp->db);
502                         if (err)
503                                 goto err_mtt;
504                 }
505         } else {
506                 qp->sq_no_prefetch = 0;
507
508                 if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
509                         qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
510
511                 if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
512                         qp->flags |= MLX4_IB_QP_LSO;
513
514                 err = set_kernel_sq_size(dev, &init_attr->cap, init_attr->qp_type, qp);
515                 if (err)
516                         goto err;
517
518                 if (!init_attr->srq) {
519                         err = mlx4_db_alloc(dev->dev, &qp->db, 0);
520                         if (err)
521                                 goto err;
522
523                         *qp->db.db = 0;
524                 }
525
526                 if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf)) {
527                         err = -ENOMEM;
528                         goto err_db;
529                 }
530
531                 err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
532                                     &qp->mtt);
533                 if (err)
534                         goto err_buf;
535
536                 err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
537                 if (err)
538                         goto err_mtt;
539
540                 qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof (u64), GFP_KERNEL);
541                 qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof (u64), GFP_KERNEL);
542
543                 if (!qp->sq.wrid || !qp->rq.wrid) {
544                         err = -ENOMEM;
545                         goto err_wrid;
546                 }
547         }
548
549         if (sqpn) {
550                 qpn = sqpn;
551         } else {
552                 err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn);
553                 if (err)
554                         goto err_wrid;
555         }
556
557         err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
558         if (err)
559                 goto err_qpn;
560
561         /*
562          * Hardware wants QPN written in big-endian order (after
563          * shifting) for send doorbell.  Precompute this value to save
564          * a little bit when posting sends.
565          */
566         qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
567
568         qp->mqp.event = mlx4_ib_qp_event;
569
570         return 0;
571
572 err_qpn:
573         if (!sqpn)
574                 mlx4_qp_release_range(dev->dev, qpn, 1);
575
576 err_wrid:
577         if (pd->uobject) {
578                 if (!init_attr->srq)
579                         mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context),
580                                               &qp->db);
581         } else {
582                 kfree(qp->sq.wrid);
583                 kfree(qp->rq.wrid);
584         }
585
586 err_mtt:
587         mlx4_mtt_cleanup(dev->dev, &qp->mtt);
588
589 err_buf:
590         if (pd->uobject)
591                 ib_umem_release(qp->umem);
592         else
593                 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
594
595 err_db:
596         if (!pd->uobject && !init_attr->srq)
597                 mlx4_db_free(dev->dev, &qp->db);
598
599 err:
600         return err;
601 }
602
603 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
604 {
605         switch (state) {
606         case IB_QPS_RESET:      return MLX4_QP_STATE_RST;
607         case IB_QPS_INIT:       return MLX4_QP_STATE_INIT;
608         case IB_QPS_RTR:        return MLX4_QP_STATE_RTR;
609         case IB_QPS_RTS:        return MLX4_QP_STATE_RTS;
610         case IB_QPS_SQD:        return MLX4_QP_STATE_SQD;
611         case IB_QPS_SQE:        return MLX4_QP_STATE_SQER;
612         case IB_QPS_ERR:        return MLX4_QP_STATE_ERR;
613         default:                return -1;
614         }
615 }
616
617 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
618         __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
619 {
620         if (send_cq == recv_cq) {
621                 spin_lock_irq(&send_cq->lock);
622                 __acquire(&recv_cq->lock);
623         } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
624                 spin_lock_irq(&send_cq->lock);
625                 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
626         } else {
627                 spin_lock_irq(&recv_cq->lock);
628                 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
629         }
630 }
631
632 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
633         __releases(&send_cq->lock) __releases(&recv_cq->lock)
634 {
635         if (send_cq == recv_cq) {
636                 __release(&recv_cq->lock);
637                 spin_unlock_irq(&send_cq->lock);
638         } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
639                 spin_unlock(&recv_cq->lock);
640                 spin_unlock_irq(&send_cq->lock);
641         } else {
642                 spin_unlock(&send_cq->lock);
643                 spin_unlock_irq(&recv_cq->lock);
644         }
645 }
646
647 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
648                               int is_user)
649 {
650         struct mlx4_ib_cq *send_cq, *recv_cq;
651
652         if (qp->state != IB_QPS_RESET)
653                 if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
654                                    MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
655                         printk(KERN_WARNING "mlx4_ib: modify QP %06x to RESET failed.\n",
656                                qp->mqp.qpn);
657
658         send_cq = to_mcq(qp->ibqp.send_cq);
659         recv_cq = to_mcq(qp->ibqp.recv_cq);
660
661         mlx4_ib_lock_cqs(send_cq, recv_cq);
662
663         if (!is_user) {
664                 __mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
665                                  qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
666                 if (send_cq != recv_cq)
667                         __mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
668         }
669
670         mlx4_qp_remove(dev->dev, &qp->mqp);
671
672         mlx4_ib_unlock_cqs(send_cq, recv_cq);
673
674         mlx4_qp_free(dev->dev, &qp->mqp);
675
676         if (!is_sqp(dev, qp))
677                 mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
678
679         mlx4_mtt_cleanup(dev->dev, &qp->mtt);
680
681         if (is_user) {
682                 if (!qp->ibqp.srq)
683                         mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
684                                               &qp->db);
685                 ib_umem_release(qp->umem);
686         } else {
687                 kfree(qp->sq.wrid);
688                 kfree(qp->rq.wrid);
689                 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
690                 if (!qp->ibqp.srq)
691                         mlx4_db_free(dev->dev, &qp->db);
692         }
693 }
694
695 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
696                                 struct ib_qp_init_attr *init_attr,
697                                 struct ib_udata *udata)
698 {
699         struct mlx4_ib_dev *dev = to_mdev(pd->device);
700         struct mlx4_ib_sqp *sqp;
701         struct mlx4_ib_qp *qp;
702         int err;
703
704         /*
705          * We only support LSO and multicast loopback blocking, and
706          * only for kernel UD QPs.
707          */
708         if (init_attr->create_flags & ~(IB_QP_CREATE_IPOIB_UD_LSO |
709                                         IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK))
710                 return ERR_PTR(-EINVAL);
711
712         if (init_attr->create_flags &&
713             (pd->uobject || init_attr->qp_type != IB_QPT_UD))
714                 return ERR_PTR(-EINVAL);
715
716         switch (init_attr->qp_type) {
717         case IB_QPT_RC:
718         case IB_QPT_UC:
719         case IB_QPT_UD:
720         {
721                 qp = kzalloc(sizeof *qp, GFP_KERNEL);
722                 if (!qp)
723                         return ERR_PTR(-ENOMEM);
724
725                 err = create_qp_common(dev, pd, init_attr, udata, 0, qp);
726                 if (err) {
727                         kfree(qp);
728                         return ERR_PTR(err);
729                 }
730
731                 qp->ibqp.qp_num = qp->mqp.qpn;
732
733                 break;
734         }
735         case IB_QPT_SMI:
736         case IB_QPT_GSI:
737         {
738                 /* Userspace is not allowed to create special QPs: */
739                 if (pd->uobject)
740                         return ERR_PTR(-EINVAL);
741
742                 sqp = kzalloc(sizeof *sqp, GFP_KERNEL);
743                 if (!sqp)
744                         return ERR_PTR(-ENOMEM);
745
746                 qp = &sqp->qp;
747
748                 err = create_qp_common(dev, pd, init_attr, udata,
749                                        dev->dev->caps.sqp_start +
750                                        (init_attr->qp_type == IB_QPT_SMI ? 0 : 2) +
751                                        init_attr->port_num - 1,
752                                        qp);
753                 if (err) {
754                         kfree(sqp);
755                         return ERR_PTR(err);
756                 }
757
758                 qp->port        = init_attr->port_num;
759                 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
760
761                 break;
762         }
763         default:
764                 /* Don't support raw QPs */
765                 return ERR_PTR(-EINVAL);
766         }
767
768         return &qp->ibqp;
769 }
770
771 int mlx4_ib_destroy_qp(struct ib_qp *qp)
772 {
773         struct mlx4_ib_dev *dev = to_mdev(qp->device);
774         struct mlx4_ib_qp *mqp = to_mqp(qp);
775
776         if (is_qp0(dev, mqp))
777                 mlx4_CLOSE_PORT(dev->dev, mqp->port);
778
779         destroy_qp_common(dev, mqp, !!qp->pd->uobject);
780
781         if (is_sqp(dev, mqp))
782                 kfree(to_msqp(mqp));
783         else
784                 kfree(mqp);
785
786         return 0;
787 }
788
789 static int to_mlx4_st(enum ib_qp_type type)
790 {
791         switch (type) {
792         case IB_QPT_RC:         return MLX4_QP_ST_RC;
793         case IB_QPT_UC:         return MLX4_QP_ST_UC;
794         case IB_QPT_UD:         return MLX4_QP_ST_UD;
795         case IB_QPT_SMI:
796         case IB_QPT_GSI:        return MLX4_QP_ST_MLX;
797         default:                return -1;
798         }
799 }
800
801 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
802                                    int attr_mask)
803 {
804         u8 dest_rd_atomic;
805         u32 access_flags;
806         u32 hw_access_flags = 0;
807
808         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
809                 dest_rd_atomic = attr->max_dest_rd_atomic;
810         else
811                 dest_rd_atomic = qp->resp_depth;
812
813         if (attr_mask & IB_QP_ACCESS_FLAGS)
814                 access_flags = attr->qp_access_flags;
815         else
816                 access_flags = qp->atomic_rd_en;
817
818         if (!dest_rd_atomic)
819                 access_flags &= IB_ACCESS_REMOTE_WRITE;
820
821         if (access_flags & IB_ACCESS_REMOTE_READ)
822                 hw_access_flags |= MLX4_QP_BIT_RRE;
823         if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
824                 hw_access_flags |= MLX4_QP_BIT_RAE;
825         if (access_flags & IB_ACCESS_REMOTE_WRITE)
826                 hw_access_flags |= MLX4_QP_BIT_RWE;
827
828         return cpu_to_be32(hw_access_flags);
829 }
830
831 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
832                             int attr_mask)
833 {
834         if (attr_mask & IB_QP_PKEY_INDEX)
835                 sqp->pkey_index = attr->pkey_index;
836         if (attr_mask & IB_QP_QKEY)
837                 sqp->qkey = attr->qkey;
838         if (attr_mask & IB_QP_SQ_PSN)
839                 sqp->send_psn = attr->sq_psn;
840 }
841
842 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
843 {
844         path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
845 }
846
847 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
848                          struct mlx4_qp_path *path, u8 port)
849 {
850         path->grh_mylmc     = ah->src_path_bits & 0x7f;
851         path->rlid          = cpu_to_be16(ah->dlid);
852         if (ah->static_rate) {
853                 path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
854                 while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
855                        !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
856                         --path->static_rate;
857         } else
858                 path->static_rate = 0;
859         path->counter_index = 0xff;
860
861         if (ah->ah_flags & IB_AH_GRH) {
862                 if (ah->grh.sgid_index >= dev->dev->caps.gid_table_len[port]) {
863                         printk(KERN_ERR "sgid_index (%u) too large. max is %d\n",
864                                ah->grh.sgid_index, dev->dev->caps.gid_table_len[port] - 1);
865                         return -1;
866                 }
867
868                 path->grh_mylmc |= 1 << 7;
869                 path->mgid_index = ah->grh.sgid_index;
870                 path->hop_limit  = ah->grh.hop_limit;
871                 path->tclass_flowlabel =
872                         cpu_to_be32((ah->grh.traffic_class << 20) |
873                                     (ah->grh.flow_label));
874                 memcpy(path->rgid, ah->grh.dgid.raw, 16);
875         }
876
877         path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
878                 ((port - 1) << 6) | ((ah->sl & 0xf) << 2);
879
880         return 0;
881 }
882
883 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
884                                const struct ib_qp_attr *attr, int attr_mask,
885                                enum ib_qp_state cur_state, enum ib_qp_state new_state)
886 {
887         struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
888         struct mlx4_ib_qp *qp = to_mqp(ibqp);
889         struct mlx4_qp_context *context;
890         enum mlx4_qp_optpar optpar = 0;
891         int sqd_event;
892         int err = -EINVAL;
893
894         context = kzalloc(sizeof *context, GFP_KERNEL);
895         if (!context)
896                 return -ENOMEM;
897
898         context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
899                                      (to_mlx4_st(ibqp->qp_type) << 16));
900         context->flags     |= cpu_to_be32(1 << 8); /* DE? */
901
902         if (!(attr_mask & IB_QP_PATH_MIG_STATE))
903                 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
904         else {
905                 optpar |= MLX4_QP_OPTPAR_PM_STATE;
906                 switch (attr->path_mig_state) {
907                 case IB_MIG_MIGRATED:
908                         context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
909                         break;
910                 case IB_MIG_REARM:
911                         context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
912                         break;
913                 case IB_MIG_ARMED:
914                         context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
915                         break;
916                 }
917         }
918
919         if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
920                 context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
921         else if (ibqp->qp_type == IB_QPT_UD) {
922                 if (qp->flags & MLX4_IB_QP_LSO)
923                         context->mtu_msgmax = (IB_MTU_4096 << 5) |
924                                               ilog2(dev->dev->caps.max_gso_sz);
925                 else
926                         context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
927         } else if (attr_mask & IB_QP_PATH_MTU) {
928                 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
929                         printk(KERN_ERR "path MTU (%u) is invalid\n",
930                                attr->path_mtu);
931                         goto out;
932                 }
933                 context->mtu_msgmax = (attr->path_mtu << 5) |
934                         ilog2(dev->dev->caps.max_msg_sz);
935         }
936
937         if (qp->rq.wqe_cnt)
938                 context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
939         context->rq_size_stride |= qp->rq.wqe_shift - 4;
940
941         if (qp->sq.wqe_cnt)
942                 context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
943         context->sq_size_stride |= qp->sq.wqe_shift - 4;
944
945         if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
946                 context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
947
948         if (qp->ibqp.uobject)
949                 context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
950         else
951                 context->usr_page = cpu_to_be32(dev->priv_uar.index);
952
953         if (attr_mask & IB_QP_DEST_QPN)
954                 context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
955
956         if (attr_mask & IB_QP_PORT) {
957                 if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
958                     !(attr_mask & IB_QP_AV)) {
959                         mlx4_set_sched(&context->pri_path, attr->port_num);
960                         optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
961                 }
962         }
963
964         if (attr_mask & IB_QP_PKEY_INDEX) {
965                 context->pri_path.pkey_index = attr->pkey_index;
966                 optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
967         }
968
969         if (attr_mask & IB_QP_AV) {
970                 if (mlx4_set_path(dev, &attr->ah_attr, &context->pri_path,
971                                   attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
972                         goto out;
973
974                 optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
975                            MLX4_QP_OPTPAR_SCHED_QUEUE);
976         }
977
978         if (attr_mask & IB_QP_TIMEOUT) {
979                 context->pri_path.ackto = attr->timeout << 3;
980                 optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
981         }
982
983         if (attr_mask & IB_QP_ALT_PATH) {
984                 if (attr->alt_port_num == 0 ||
985                     attr->alt_port_num > dev->dev->caps.num_ports)
986                         goto out;
987
988                 if (attr->alt_pkey_index >=
989                     dev->dev->caps.pkey_table_len[attr->alt_port_num])
990                         goto out;
991
992                 if (mlx4_set_path(dev, &attr->alt_ah_attr, &context->alt_path,
993                                   attr->alt_port_num))
994                         goto out;
995
996                 context->alt_path.pkey_index = attr->alt_pkey_index;
997                 context->alt_path.ackto = attr->alt_timeout << 3;
998                 optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
999         }
1000
1001         context->pd         = cpu_to_be32(to_mpd(ibqp->pd)->pdn);
1002         context->params1    = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1003
1004         /* Set "fast registration enabled" for all kernel QPs */
1005         if (!qp->ibqp.uobject)
1006                 context->params1 |= cpu_to_be32(1 << 11);
1007
1008         if (attr_mask & IB_QP_RNR_RETRY) {
1009                 context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1010                 optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1011         }
1012
1013         if (attr_mask & IB_QP_RETRY_CNT) {
1014                 context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1015                 optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1016         }
1017
1018         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1019                 if (attr->max_rd_atomic)
1020                         context->params1 |=
1021                                 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1022                 optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1023         }
1024
1025         if (attr_mask & IB_QP_SQ_PSN)
1026                 context->next_send_psn = cpu_to_be32(attr->sq_psn);
1027
1028         context->cqn_send = cpu_to_be32(to_mcq(ibqp->send_cq)->mcq.cqn);
1029
1030         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1031                 if (attr->max_dest_rd_atomic)
1032                         context->params2 |=
1033                                 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1034                 optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1035         }
1036
1037         if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1038                 context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1039                 optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1040         }
1041
1042         if (ibqp->srq)
1043                 context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1044
1045         if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1046                 context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1047                 optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1048         }
1049         if (attr_mask & IB_QP_RQ_PSN)
1050                 context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1051
1052         context->cqn_recv = cpu_to_be32(to_mcq(ibqp->recv_cq)->mcq.cqn);
1053
1054         if (attr_mask & IB_QP_QKEY) {
1055                 context->qkey = cpu_to_be32(attr->qkey);
1056                 optpar |= MLX4_QP_OPTPAR_Q_KEY;
1057         }
1058
1059         if (ibqp->srq)
1060                 context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1061
1062         if (!ibqp->srq && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1063                 context->db_rec_addr = cpu_to_be64(qp->db.dma);
1064
1065         if (cur_state == IB_QPS_INIT &&
1066             new_state == IB_QPS_RTR  &&
1067             (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1068              ibqp->qp_type == IB_QPT_UD)) {
1069                 context->pri_path.sched_queue = (qp->port - 1) << 6;
1070                 if (is_qp0(dev, qp))
1071                         context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1072                 else
1073                         context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1074         }
1075
1076         if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD  &&
1077             attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1078                 sqd_event = 1;
1079         else
1080                 sqd_event = 0;
1081
1082         if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1083                 context->rlkey |= (1 << 4);
1084
1085         /*
1086          * Before passing a kernel QP to the HW, make sure that the
1087          * ownership bits of the send queue are set and the SQ
1088          * headroom is stamped so that the hardware doesn't start
1089          * processing stale work requests.
1090          */
1091         if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1092                 struct mlx4_wqe_ctrl_seg *ctrl;
1093                 int i;
1094
1095                 for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1096                         ctrl = get_send_wqe(qp, i);
1097                         ctrl->owner_opcode = cpu_to_be32(1 << 31);
1098                         if (qp->sq_max_wqes_per_wr == 1)
1099                                 ctrl->fence_size = 1 << (qp->sq.wqe_shift - 4);
1100
1101                         stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1102                 }
1103         }
1104
1105         err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
1106                              to_mlx4_state(new_state), context, optpar,
1107                              sqd_event, &qp->mqp);
1108         if (err)
1109                 goto out;
1110
1111         qp->state = new_state;
1112
1113         if (attr_mask & IB_QP_ACCESS_FLAGS)
1114                 qp->atomic_rd_en = attr->qp_access_flags;
1115         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1116                 qp->resp_depth = attr->max_dest_rd_atomic;
1117         if (attr_mask & IB_QP_PORT)
1118                 qp->port = attr->port_num;
1119         if (attr_mask & IB_QP_ALT_PATH)
1120                 qp->alt_port = attr->alt_port_num;
1121
1122         if (is_sqp(dev, qp))
1123                 store_sqp_attrs(to_msqp(qp), attr, attr_mask);
1124
1125         /*
1126          * If we moved QP0 to RTR, bring the IB link up; if we moved
1127          * QP0 to RESET or ERROR, bring the link back down.
1128          */
1129         if (is_qp0(dev, qp)) {
1130                 if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
1131                         if (mlx4_INIT_PORT(dev->dev, qp->port))
1132                                 printk(KERN_WARNING "INIT_PORT failed for port %d\n",
1133                                        qp->port);
1134
1135                 if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1136                     (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1137                         mlx4_CLOSE_PORT(dev->dev, qp->port);
1138         }
1139
1140         /*
1141          * If we moved a kernel QP to RESET, clean up all old CQ
1142          * entries and reinitialize the QP.
1143          */
1144         if (new_state == IB_QPS_RESET && !ibqp->uobject) {
1145                 mlx4_ib_cq_clean(to_mcq(ibqp->recv_cq), qp->mqp.qpn,
1146                                  ibqp->srq ? to_msrq(ibqp->srq): NULL);
1147                 if (ibqp->send_cq != ibqp->recv_cq)
1148                         mlx4_ib_cq_clean(to_mcq(ibqp->send_cq), qp->mqp.qpn, NULL);
1149
1150                 qp->rq.head = 0;
1151                 qp->rq.tail = 0;
1152                 qp->sq.head = 0;
1153                 qp->sq.tail = 0;
1154                 qp->sq_next_wqe = 0;
1155                 if (!ibqp->srq)
1156                         *qp->db.db  = 0;
1157         }
1158
1159 out:
1160         kfree(context);
1161         return err;
1162 }
1163
1164 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1165                       int attr_mask, struct ib_udata *udata)
1166 {
1167         struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1168         struct mlx4_ib_qp *qp = to_mqp(ibqp);
1169         enum ib_qp_state cur_state, new_state;
1170         int err = -EINVAL;
1171
1172         mutex_lock(&qp->mutex);
1173
1174         cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1175         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1176
1177         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask))
1178                 goto out;
1179
1180         if ((attr_mask & IB_QP_PORT) &&
1181             (attr->port_num == 0 || attr->port_num > dev->dev->caps.num_ports)) {
1182                 goto out;
1183         }
1184
1185         if (attr_mask & IB_QP_PKEY_INDEX) {
1186                 int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1187                 if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p])
1188                         goto out;
1189         }
1190
1191         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1192             attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
1193                 goto out;
1194         }
1195
1196         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1197             attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
1198                 goto out;
1199         }
1200
1201         if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1202                 err = 0;
1203                 goto out;
1204         }
1205
1206         err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
1207
1208 out:
1209         mutex_unlock(&qp->mutex);
1210         return err;
1211 }
1212
1213 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
1214                             void *wqe, unsigned *mlx_seg_len)
1215 {
1216         struct ib_device *ib_dev = &to_mdev(sqp->qp.ibqp.device)->ib_dev;
1217         struct mlx4_wqe_mlx_seg *mlx = wqe;
1218         struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
1219         struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
1220         u16 pkey;
1221         int send_size;
1222         int header_size;
1223         int spc;
1224         int i;
1225
1226         send_size = 0;
1227         for (i = 0; i < wr->num_sge; ++i)
1228                 send_size += wr->sg_list[i].length;
1229
1230         ib_ud_header_init(send_size, mlx4_ib_ah_grh_present(ah), &sqp->ud_header);
1231
1232         sqp->ud_header.lrh.service_level   =
1233                 be32_to_cpu(ah->av.sl_tclass_flowlabel) >> 28;
1234         sqp->ud_header.lrh.destination_lid = ah->av.dlid;
1235         sqp->ud_header.lrh.source_lid      = cpu_to_be16(ah->av.g_slid & 0x7f);
1236         if (mlx4_ib_ah_grh_present(ah)) {
1237                 sqp->ud_header.grh.traffic_class =
1238                         (be32_to_cpu(ah->av.sl_tclass_flowlabel) >> 20) & 0xff;
1239                 sqp->ud_header.grh.flow_label    =
1240                         ah->av.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
1241                 sqp->ud_header.grh.hop_limit     = ah->av.hop_limit;
1242                 ib_get_cached_gid(ib_dev, be32_to_cpu(ah->av.port_pd) >> 24,
1243                                   ah->av.gid_index, &sqp->ud_header.grh.source_gid);
1244                 memcpy(sqp->ud_header.grh.destination_gid.raw,
1245                        ah->av.dgid, 16);
1246         }
1247
1248         mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1249         mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
1250                                   (sqp->ud_header.lrh.destination_lid ==
1251                                    IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
1252                                   (sqp->ud_header.lrh.service_level << 8));
1253         mlx->rlid   = sqp->ud_header.lrh.destination_lid;
1254
1255         switch (wr->opcode) {
1256         case IB_WR_SEND:
1257                 sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
1258                 sqp->ud_header.immediate_present = 0;
1259                 break;
1260         case IB_WR_SEND_WITH_IMM:
1261                 sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1262                 sqp->ud_header.immediate_present = 1;
1263                 sqp->ud_header.immediate_data    = wr->ex.imm_data;
1264                 break;
1265         default:
1266                 return -EINVAL;
1267         }
1268
1269         sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1270         if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1271                 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1272         sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1273         if (!sqp->qp.ibqp.qp_num)
1274                 ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
1275         else
1276                 ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey);
1277         sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1278         sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1279         sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1280         sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1281                                                sqp->qkey : wr->wr.ud.remote_qkey);
1282         sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1283
1284         header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
1285
1286         if (0) {
1287                 printk(KERN_ERR "built UD header of size %d:\n", header_size);
1288                 for (i = 0; i < header_size / 4; ++i) {
1289                         if (i % 8 == 0)
1290                                 printk("  [%02x] ", i * 4);
1291                         printk(" %08x",
1292                                be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
1293                         if ((i + 1) % 8 == 0)
1294                                 printk("\n");
1295                 }
1296                 printk("\n");
1297         }
1298
1299         /*
1300          * Inline data segments may not cross a 64 byte boundary.  If
1301          * our UD header is bigger than the space available up to the
1302          * next 64 byte boundary in the WQE, use two inline data
1303          * segments to hold the UD header.
1304          */
1305         spc = MLX4_INLINE_ALIGN -
1306                 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
1307         if (header_size <= spc) {
1308                 inl->byte_count = cpu_to_be32(1 << 31 | header_size);
1309                 memcpy(inl + 1, sqp->header_buf, header_size);
1310                 i = 1;
1311         } else {
1312                 inl->byte_count = cpu_to_be32(1 << 31 | spc);
1313                 memcpy(inl + 1, sqp->header_buf, spc);
1314
1315                 inl = (void *) (inl + 1) + spc;
1316                 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
1317                 /*
1318                  * Need a barrier here to make sure all the data is
1319                  * visible before the byte_count field is set.
1320                  * Otherwise the HCA prefetcher could grab the 64-byte
1321                  * chunk with this inline segment and get a valid (!=
1322                  * 0xffffffff) byte count but stale data, and end up
1323                  * generating a packet with bad headers.
1324                  *
1325                  * The first inline segment's byte_count field doesn't
1326                  * need a barrier, because it comes after a
1327                  * control/MLX segment and therefore is at an offset
1328                  * of 16 mod 64.
1329                  */
1330                 wmb();
1331                 inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
1332                 i = 2;
1333         }
1334
1335         *mlx_seg_len =
1336                 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
1337         return 0;
1338 }
1339
1340 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
1341 {
1342         unsigned cur;
1343         struct mlx4_ib_cq *cq;
1344
1345         cur = wq->head - wq->tail;
1346         if (likely(cur + nreq < wq->max_post))
1347                 return 0;
1348
1349         cq = to_mcq(ib_cq);
1350         spin_lock(&cq->lock);
1351         cur = wq->head - wq->tail;
1352         spin_unlock(&cq->lock);
1353
1354         return cur + nreq >= wq->max_post;
1355 }
1356
1357 static __be32 convert_access(int acc)
1358 {
1359         return (acc & IB_ACCESS_REMOTE_ATOMIC ? cpu_to_be32(MLX4_WQE_FMR_PERM_ATOMIC)       : 0) |
1360                (acc & IB_ACCESS_REMOTE_WRITE  ? cpu_to_be32(MLX4_WQE_FMR_PERM_REMOTE_WRITE) : 0) |
1361                (acc & IB_ACCESS_REMOTE_READ   ? cpu_to_be32(MLX4_WQE_FMR_PERM_REMOTE_READ)  : 0) |
1362                (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
1363                 cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
1364 }
1365
1366 static void set_fmr_seg(struct mlx4_wqe_fmr_seg *fseg, struct ib_send_wr *wr)
1367 {
1368         struct mlx4_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
1369         int i;
1370
1371         for (i = 0; i < wr->wr.fast_reg.page_list_len; ++i)
1372                 mfrpl->mapped_page_list[i] =
1373                         cpu_to_be64(wr->wr.fast_reg.page_list->page_list[i] |
1374                                     MLX4_MTT_FLAG_PRESENT);
1375
1376         fseg->flags             = convert_access(wr->wr.fast_reg.access_flags);
1377         fseg->mem_key           = cpu_to_be32(wr->wr.fast_reg.rkey);
1378         fseg->buf_list          = cpu_to_be64(mfrpl->map);
1379         fseg->start_addr        = cpu_to_be64(wr->wr.fast_reg.iova_start);
1380         fseg->reg_len           = cpu_to_be64(wr->wr.fast_reg.length);
1381         fseg->offset            = 0; /* XXX -- is this just for ZBVA? */
1382         fseg->page_size         = cpu_to_be32(wr->wr.fast_reg.page_shift);
1383         fseg->reserved[0]       = 0;
1384         fseg->reserved[1]       = 0;
1385 }
1386
1387 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
1388 {
1389         iseg->flags     = 0;
1390         iseg->mem_key   = cpu_to_be32(rkey);
1391         iseg->guest_id  = 0;
1392         iseg->pa        = 0;
1393 }
1394
1395 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
1396                                           u64 remote_addr, u32 rkey)
1397 {
1398         rseg->raddr    = cpu_to_be64(remote_addr);
1399         rseg->rkey     = cpu_to_be32(rkey);
1400         rseg->reserved = 0;
1401 }
1402
1403 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, struct ib_send_wr *wr)
1404 {
1405         if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1406                 aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
1407                 aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add);
1408         } else {
1409                 aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
1410                 aseg->compare  = 0;
1411         }
1412
1413 }
1414
1415 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
1416                              struct ib_send_wr *wr)
1417 {
1418         memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av));
1419         dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1420         dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1421 }
1422
1423 static void set_mlx_icrc_seg(void *dseg)
1424 {
1425         u32 *t = dseg;
1426         struct mlx4_wqe_inline_seg *iseg = dseg;
1427
1428         t[1] = 0;
1429
1430         /*
1431          * Need a barrier here before writing the byte_count field to
1432          * make sure that all the data is visible before the
1433          * byte_count field is set.  Otherwise, if the segment begins
1434          * a new cacheline, the HCA prefetcher could grab the 64-byte
1435          * chunk and get a valid (!= * 0xffffffff) byte count but
1436          * stale data, and end up sending the wrong data.
1437          */
1438         wmb();
1439
1440         iseg->byte_count = cpu_to_be32((1 << 31) | 4);
1441 }
1442
1443 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
1444 {
1445         dseg->lkey       = cpu_to_be32(sg->lkey);
1446         dseg->addr       = cpu_to_be64(sg->addr);
1447
1448         /*
1449          * Need a barrier here before writing the byte_count field to
1450          * make sure that all the data is visible before the
1451          * byte_count field is set.  Otherwise, if the segment begins
1452          * a new cacheline, the HCA prefetcher could grab the 64-byte
1453          * chunk and get a valid (!= * 0xffffffff) byte count but
1454          * stale data, and end up sending the wrong data.
1455          */
1456         wmb();
1457
1458         dseg->byte_count = cpu_to_be32(sg->length);
1459 }
1460
1461 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
1462 {
1463         dseg->byte_count = cpu_to_be32(sg->length);
1464         dseg->lkey       = cpu_to_be32(sg->lkey);
1465         dseg->addr       = cpu_to_be64(sg->addr);
1466 }
1467
1468 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_send_wr *wr,
1469                          struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
1470                          __be32 *lso_hdr_sz)
1471 {
1472         unsigned halign = ALIGN(sizeof *wqe + wr->wr.ud.hlen, 16);
1473
1474         /*
1475          * This is a temporary limitation and will be removed in
1476          * a forthcoming FW release:
1477          */
1478         if (unlikely(halign > 64))
1479                 return -EINVAL;
1480
1481         if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
1482                      wr->num_sge > qp->sq.max_gs - (halign >> 4)))
1483                 return -EINVAL;
1484
1485         memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen);
1486
1487         *lso_hdr_sz  = cpu_to_be32((wr->wr.ud.mss - wr->wr.ud.hlen) << 16 |
1488                                    wr->wr.ud.hlen);
1489         *lso_seg_len = halign;
1490         return 0;
1491 }
1492
1493 static __be32 send_ieth(struct ib_send_wr *wr)
1494 {
1495         switch (wr->opcode) {
1496         case IB_WR_SEND_WITH_IMM:
1497         case IB_WR_RDMA_WRITE_WITH_IMM:
1498                 return wr->ex.imm_data;
1499
1500         case IB_WR_SEND_WITH_INV:
1501                 return cpu_to_be32(wr->ex.invalidate_rkey);
1502
1503         default:
1504                 return 0;
1505         }
1506 }
1507
1508 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1509                       struct ib_send_wr **bad_wr)
1510 {
1511         struct mlx4_ib_qp *qp = to_mqp(ibqp);
1512         void *wqe;
1513         struct mlx4_wqe_ctrl_seg *ctrl;
1514         struct mlx4_wqe_data_seg *dseg;
1515         unsigned long flags;
1516         int nreq;
1517         int err = 0;
1518         unsigned ind;
1519         int uninitialized_var(stamp);
1520         int uninitialized_var(size);
1521         unsigned uninitialized_var(seglen);
1522         __be32 dummy;
1523         __be32 *lso_wqe;
1524         __be32 uninitialized_var(lso_hdr_sz);
1525         int i;
1526
1527         spin_lock_irqsave(&qp->sq.lock, flags);
1528
1529         ind = qp->sq_next_wqe;
1530
1531         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1532                 lso_wqe = &dummy;
1533
1534                 if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1535                         err = -ENOMEM;
1536                         *bad_wr = wr;
1537                         goto out;
1538                 }
1539
1540                 if (unlikely(wr->num_sge > qp->sq.max_gs)) {
1541                         err = -EINVAL;
1542                         *bad_wr = wr;
1543                         goto out;
1544                 }
1545
1546                 ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
1547                 qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
1548
1549                 ctrl->srcrb_flags =
1550                         (wr->send_flags & IB_SEND_SIGNALED ?
1551                          cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
1552                         (wr->send_flags & IB_SEND_SOLICITED ?
1553                          cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
1554                         ((wr->send_flags & IB_SEND_IP_CSUM) ?
1555                          cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
1556                                      MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
1557                         qp->sq_signal_bits;
1558
1559                 ctrl->imm = send_ieth(wr);
1560
1561                 wqe += sizeof *ctrl;
1562                 size = sizeof *ctrl / 16;
1563
1564                 switch (ibqp->qp_type) {
1565                 case IB_QPT_RC:
1566                 case IB_QPT_UC:
1567                         switch (wr->opcode) {
1568                         case IB_WR_ATOMIC_CMP_AND_SWP:
1569                         case IB_WR_ATOMIC_FETCH_AND_ADD:
1570                                 set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
1571                                               wr->wr.atomic.rkey);
1572                                 wqe  += sizeof (struct mlx4_wqe_raddr_seg);
1573
1574                                 set_atomic_seg(wqe, wr);
1575                                 wqe  += sizeof (struct mlx4_wqe_atomic_seg);
1576
1577                                 size += (sizeof (struct mlx4_wqe_raddr_seg) +
1578                                          sizeof (struct mlx4_wqe_atomic_seg)) / 16;
1579
1580                                 break;
1581
1582                         case IB_WR_RDMA_READ:
1583                         case IB_WR_RDMA_WRITE:
1584                         case IB_WR_RDMA_WRITE_WITH_IMM:
1585                                 set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
1586                                               wr->wr.rdma.rkey);
1587                                 wqe  += sizeof (struct mlx4_wqe_raddr_seg);
1588                                 size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
1589                                 break;
1590
1591                         case IB_WR_LOCAL_INV:
1592                                 ctrl->srcrb_flags |=
1593                                         cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
1594                                 set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
1595                                 wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
1596                                 size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
1597                                 break;
1598
1599                         case IB_WR_FAST_REG_MR:
1600                                 ctrl->srcrb_flags |=
1601                                         cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
1602                                 set_fmr_seg(wqe, wr);
1603                                 wqe  += sizeof (struct mlx4_wqe_fmr_seg);
1604                                 size += sizeof (struct mlx4_wqe_fmr_seg) / 16;
1605                                 break;
1606
1607                         default:
1608                                 /* No extra segments required for sends */
1609                                 break;
1610                         }
1611                         break;
1612
1613                 case IB_QPT_UD:
1614                         set_datagram_seg(wqe, wr);
1615                         wqe  += sizeof (struct mlx4_wqe_datagram_seg);
1616                         size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
1617
1618                         if (wr->opcode == IB_WR_LSO) {
1619                                 err = build_lso_seg(wqe, wr, qp, &seglen, &lso_hdr_sz);
1620                                 if (unlikely(err)) {
1621                                         *bad_wr = wr;
1622                                         goto out;
1623                                 }
1624                                 lso_wqe = (__be32 *) wqe;
1625                                 wqe  += seglen;
1626                                 size += seglen / 16;
1627                         }
1628                         break;
1629
1630                 case IB_QPT_SMI:
1631                 case IB_QPT_GSI:
1632                         err = build_mlx_header(to_msqp(qp), wr, ctrl, &seglen);
1633                         if (unlikely(err)) {
1634                                 *bad_wr = wr;
1635                                 goto out;
1636                         }
1637                         wqe  += seglen;
1638                         size += seglen / 16;
1639                         break;
1640
1641                 default:
1642                         break;
1643                 }
1644
1645                 /*
1646                  * Write data segments in reverse order, so as to
1647                  * overwrite cacheline stamp last within each
1648                  * cacheline.  This avoids issues with WQE
1649                  * prefetching.
1650                  */
1651
1652                 dseg = wqe;
1653                 dseg += wr->num_sge - 1;
1654                 size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
1655
1656                 /* Add one more inline data segment for ICRC for MLX sends */
1657                 if (unlikely(qp->ibqp.qp_type == IB_QPT_SMI ||
1658                              qp->ibqp.qp_type == IB_QPT_GSI)) {
1659                         set_mlx_icrc_seg(dseg + 1);
1660                         size += sizeof (struct mlx4_wqe_data_seg) / 16;
1661                 }
1662
1663                 for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
1664                         set_data_seg(dseg, wr->sg_list + i);
1665
1666                 /*
1667                  * Possibly overwrite stamping in cacheline with LSO
1668                  * segment only after making sure all data segments
1669                  * are written.
1670                  */
1671                 wmb();
1672                 *lso_wqe = lso_hdr_sz;
1673
1674                 ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
1675                                     MLX4_WQE_CTRL_FENCE : 0) | size;
1676
1677                 /*
1678                  * Make sure descriptor is fully written before
1679                  * setting ownership bit (because HW can start
1680                  * executing as soon as we do).
1681                  */
1682                 wmb();
1683
1684                 if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
1685                         err = -EINVAL;
1686                         goto out;
1687                 }
1688
1689                 ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
1690                         (ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
1691
1692                 stamp = ind + qp->sq_spare_wqes;
1693                 ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
1694
1695                 /*
1696                  * We can improve latency by not stamping the last
1697                  * send queue WQE until after ringing the doorbell, so
1698                  * only stamp here if there are still more WQEs to post.
1699                  *
1700                  * Same optimization applies to padding with NOP wqe
1701                  * in case of WQE shrinking (used to prevent wrap-around
1702                  * in the middle of WR).
1703                  */
1704                 if (wr->next) {
1705                         stamp_send_wqe(qp, stamp, size * 16);
1706                         ind = pad_wraparound(qp, ind);
1707                 }
1708         }
1709
1710 out:
1711         if (likely(nreq)) {
1712                 qp->sq.head += nreq;
1713
1714                 /*
1715                  * Make sure that descriptors are written before
1716                  * doorbell record.
1717                  */
1718                 wmb();
1719
1720                 writel(qp->doorbell_qpn,
1721                        to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
1722
1723                 /*
1724                  * Make sure doorbells don't leak out of SQ spinlock
1725                  * and reach the HCA out of order.
1726                  */
1727                 mmiowb();
1728
1729                 stamp_send_wqe(qp, stamp, size * 16);
1730
1731                 ind = pad_wraparound(qp, ind);
1732                 qp->sq_next_wqe = ind;
1733         }
1734
1735         spin_unlock_irqrestore(&qp->sq.lock, flags);
1736
1737         return err;
1738 }
1739
1740 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1741                       struct ib_recv_wr **bad_wr)
1742 {
1743         struct mlx4_ib_qp *qp = to_mqp(ibqp);
1744         struct mlx4_wqe_data_seg *scat;
1745         unsigned long flags;
1746         int err = 0;
1747         int nreq;
1748         int ind;
1749         int i;
1750
1751         spin_lock_irqsave(&qp->rq.lock, flags);
1752
1753         ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
1754
1755         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1756                 if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.send_cq)) {
1757                         err = -ENOMEM;
1758                         *bad_wr = wr;
1759                         goto out;
1760                 }
1761
1762                 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1763                         err = -EINVAL;
1764                         *bad_wr = wr;
1765                         goto out;
1766                 }
1767
1768                 scat = get_recv_wqe(qp, ind);
1769
1770                 for (i = 0; i < wr->num_sge; ++i)
1771                         __set_data_seg(scat + i, wr->sg_list + i);
1772
1773                 if (i < qp->rq.max_gs) {
1774                         scat[i].byte_count = 0;
1775                         scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
1776                         scat[i].addr       = 0;
1777                 }
1778
1779                 qp->rq.wrid[ind] = wr->wr_id;
1780
1781                 ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
1782         }
1783
1784 out:
1785         if (likely(nreq)) {
1786                 qp->rq.head += nreq;
1787
1788                 /*
1789                  * Make sure that descriptors are written before
1790                  * doorbell record.
1791                  */
1792                 wmb();
1793
1794                 *qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
1795         }
1796
1797         spin_unlock_irqrestore(&qp->rq.lock, flags);
1798
1799         return err;
1800 }
1801
1802 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
1803 {
1804         switch (mlx4_state) {
1805         case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
1806         case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
1807         case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
1808         case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
1809         case MLX4_QP_STATE_SQ_DRAINING:
1810         case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
1811         case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
1812         case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
1813         default:                     return -1;
1814         }
1815 }
1816
1817 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
1818 {
1819         switch (mlx4_mig_state) {
1820         case MLX4_QP_PM_ARMED:          return IB_MIG_ARMED;
1821         case MLX4_QP_PM_REARM:          return IB_MIG_REARM;
1822         case MLX4_QP_PM_MIGRATED:       return IB_MIG_MIGRATED;
1823         default: return -1;
1824         }
1825 }
1826
1827 static int to_ib_qp_access_flags(int mlx4_flags)
1828 {
1829         int ib_flags = 0;
1830
1831         if (mlx4_flags & MLX4_QP_BIT_RRE)
1832                 ib_flags |= IB_ACCESS_REMOTE_READ;
1833         if (mlx4_flags & MLX4_QP_BIT_RWE)
1834                 ib_flags |= IB_ACCESS_REMOTE_WRITE;
1835         if (mlx4_flags & MLX4_QP_BIT_RAE)
1836                 ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
1837
1838         return ib_flags;
1839 }
1840
1841 static void to_ib_ah_attr(struct mlx4_dev *dev, struct ib_ah_attr *ib_ah_attr,
1842                                 struct mlx4_qp_path *path)
1843 {
1844         memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
1845         ib_ah_attr->port_num      = path->sched_queue & 0x40 ? 2 : 1;
1846
1847         if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
1848                 return;
1849
1850         ib_ah_attr->dlid          = be16_to_cpu(path->rlid);
1851         ib_ah_attr->sl            = (path->sched_queue >> 2) & 0xf;
1852         ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
1853         ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
1854         ib_ah_attr->ah_flags      = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
1855         if (ib_ah_attr->ah_flags) {
1856                 ib_ah_attr->grh.sgid_index = path->mgid_index;
1857                 ib_ah_attr->grh.hop_limit  = path->hop_limit;
1858                 ib_ah_attr->grh.traffic_class =
1859                         (be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
1860                 ib_ah_attr->grh.flow_label =
1861                         be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
1862                 memcpy(ib_ah_attr->grh.dgid.raw,
1863                         path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
1864         }
1865 }
1866
1867 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
1868                      struct ib_qp_init_attr *qp_init_attr)
1869 {
1870         struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1871         struct mlx4_ib_qp *qp = to_mqp(ibqp);
1872         struct mlx4_qp_context context;
1873         int mlx4_state;
1874         int err = 0;
1875
1876         mutex_lock(&qp->mutex);
1877
1878         if (qp->state == IB_QPS_RESET) {
1879                 qp_attr->qp_state = IB_QPS_RESET;
1880                 goto done;
1881         }
1882
1883         err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
1884         if (err) {
1885                 err = -EINVAL;
1886                 goto out;
1887         }
1888
1889         mlx4_state = be32_to_cpu(context.flags) >> 28;
1890
1891         qp->state                    = to_ib_qp_state(mlx4_state);
1892         qp_attr->qp_state            = qp->state;
1893         qp_attr->path_mtu            = context.mtu_msgmax >> 5;
1894         qp_attr->path_mig_state      =
1895                 to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
1896         qp_attr->qkey                = be32_to_cpu(context.qkey);
1897         qp_attr->rq_psn              = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
1898         qp_attr->sq_psn              = be32_to_cpu(context.next_send_psn) & 0xffffff;
1899         qp_attr->dest_qp_num         = be32_to_cpu(context.remote_qpn) & 0xffffff;
1900         qp_attr->qp_access_flags     =
1901                 to_ib_qp_access_flags(be32_to_cpu(context.params2));
1902
1903         if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
1904                 to_ib_ah_attr(dev->dev, &qp_attr->ah_attr, &context.pri_path);
1905                 to_ib_ah_attr(dev->dev, &qp_attr->alt_ah_attr, &context.alt_path);
1906                 qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
1907                 qp_attr->alt_port_num   = qp_attr->alt_ah_attr.port_num;
1908         }
1909
1910         qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
1911         if (qp_attr->qp_state == IB_QPS_INIT)
1912                 qp_attr->port_num = qp->port;
1913         else
1914                 qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
1915
1916         /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
1917         qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
1918
1919         qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
1920
1921         qp_attr->max_dest_rd_atomic =
1922                 1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
1923         qp_attr->min_rnr_timer      =
1924                 (be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
1925         qp_attr->timeout            = context.pri_path.ackto >> 3;
1926         qp_attr->retry_cnt          = (be32_to_cpu(context.params1) >> 16) & 0x7;
1927         qp_attr->rnr_retry          = (be32_to_cpu(context.params1) >> 13) & 0x7;
1928         qp_attr->alt_timeout        = context.alt_path.ackto >> 3;
1929
1930 done:
1931         qp_attr->cur_qp_state        = qp_attr->qp_state;
1932         qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
1933         qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
1934
1935         if (!ibqp->uobject) {
1936                 qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
1937                 qp_attr->cap.max_send_sge = qp->sq.max_gs;
1938         } else {
1939                 qp_attr->cap.max_send_wr  = 0;
1940                 qp_attr->cap.max_send_sge = 0;
1941         }
1942
1943         /*
1944          * We don't support inline sends for kernel QPs (yet), and we
1945          * don't know what userspace's value should be.
1946          */
1947         qp_attr->cap.max_inline_data = 0;
1948
1949         qp_init_attr->cap            = qp_attr->cap;
1950
1951         qp_init_attr->create_flags = 0;
1952         if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
1953                 qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
1954
1955         if (qp->flags & MLX4_IB_QP_LSO)
1956                 qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
1957
1958 out:
1959         mutex_unlock(&qp->mutex);
1960         return err;
1961 }
1962