include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit...
[linux-3.10.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 #include <linux/slab.h>
36
37 #include <rdma/ib_cache.h>
38 #include <rdma/ib_pack.h>
39
40 #include <linux/mlx4/qp.h>
41
42 #include "mlx4_ib.h"
43 #include "user.h"
44
45 enum {
46         MLX4_IB_ACK_REQ_FREQ    = 8,
47 };
48
49 enum {
50         MLX4_IB_DEFAULT_SCHED_QUEUE     = 0x83,
51         MLX4_IB_DEFAULT_QP0_SCHED_QUEUE = 0x3f
52 };
53
54 enum {
55         /*
56          * Largest possible UD header: send with GRH and immediate data.
57          */
58         MLX4_IB_UD_HEADER_SIZE          = 72,
59         MLX4_IB_LSO_HEADER_SPARE        = 128,
60 };
61
62 struct mlx4_ib_sqp {
63         struct mlx4_ib_qp       qp;
64         int                     pkey_index;
65         u32                     qkey;
66         u32                     send_psn;
67         struct ib_ud_header     ud_header;
68         u8                      header_buf[MLX4_IB_UD_HEADER_SIZE];
69 };
70
71 enum {
72         MLX4_IB_MIN_SQ_STRIDE   = 6,
73         MLX4_IB_CACHE_LINE_SIZE = 64,
74 };
75
76 static const __be32 mlx4_ib_opcode[] = {
77         [IB_WR_SEND]                    = cpu_to_be32(MLX4_OPCODE_SEND),
78         [IB_WR_LSO]                     = cpu_to_be32(MLX4_OPCODE_LSO),
79         [IB_WR_SEND_WITH_IMM]           = cpu_to_be32(MLX4_OPCODE_SEND_IMM),
80         [IB_WR_RDMA_WRITE]              = cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
81         [IB_WR_RDMA_WRITE_WITH_IMM]     = cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
82         [IB_WR_RDMA_READ]               = cpu_to_be32(MLX4_OPCODE_RDMA_READ),
83         [IB_WR_ATOMIC_CMP_AND_SWP]      = cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
84         [IB_WR_ATOMIC_FETCH_AND_ADD]    = cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
85         [IB_WR_SEND_WITH_INV]           = cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
86         [IB_WR_LOCAL_INV]               = cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
87         [IB_WR_FAST_REG_MR]             = cpu_to_be32(MLX4_OPCODE_FMR),
88 };
89
90 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
91 {
92         return container_of(mqp, struct mlx4_ib_sqp, qp);
93 }
94
95 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
96 {
97         return qp->mqp.qpn >= dev->dev->caps.sqp_start &&
98                 qp->mqp.qpn <= dev->dev->caps.sqp_start + 3;
99 }
100
101 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
102 {
103         return qp->mqp.qpn >= dev->dev->caps.sqp_start &&
104                 qp->mqp.qpn <= dev->dev->caps.sqp_start + 1;
105 }
106
107 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
108 {
109         return mlx4_buf_offset(&qp->buf, offset);
110 }
111
112 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
113 {
114         return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
115 }
116
117 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
118 {
119         return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
120 }
121
122 /*
123  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
124  * first four bytes of every 64 byte chunk with
125  *     0x7FFFFFF | (invalid_ownership_value << 31).
126  *
127  * When the max work request size is less than or equal to the WQE
128  * basic block size, as an optimization, we can stamp all WQEs with
129  * 0xffffffff, and skip the very first chunk of each WQE.
130  */
131 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
132 {
133         __be32 *wqe;
134         int i;
135         int s;
136         int ind;
137         void *buf;
138         __be32 stamp;
139         struct mlx4_wqe_ctrl_seg *ctrl;
140
141         if (qp->sq_max_wqes_per_wr > 1) {
142                 s = roundup(size, 1U << qp->sq.wqe_shift);
143                 for (i = 0; i < s; i += 64) {
144                         ind = (i >> qp->sq.wqe_shift) + n;
145                         stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
146                                                        cpu_to_be32(0xffffffff);
147                         buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
148                         wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
149                         *wqe = stamp;
150                 }
151         } else {
152                 ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
153                 s = (ctrl->fence_size & 0x3f) << 4;
154                 for (i = 64; i < s; i += 64) {
155                         wqe = buf + i;
156                         *wqe = cpu_to_be32(0xffffffff);
157                 }
158         }
159 }
160
161 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
162 {
163         struct mlx4_wqe_ctrl_seg *ctrl;
164         struct mlx4_wqe_inline_seg *inl;
165         void *wqe;
166         int s;
167
168         ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
169         s = sizeof(struct mlx4_wqe_ctrl_seg);
170
171         if (qp->ibqp.qp_type == IB_QPT_UD) {
172                 struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
173                 struct mlx4_av *av = (struct mlx4_av *)dgram->av;
174                 memset(dgram, 0, sizeof *dgram);
175                 av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
176                 s += sizeof(struct mlx4_wqe_datagram_seg);
177         }
178
179         /* Pad the remainder of the WQE with an inline data segment. */
180         if (size > s) {
181                 inl = wqe + s;
182                 inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
183         }
184         ctrl->srcrb_flags = 0;
185         ctrl->fence_size = size / 16;
186         /*
187          * Make sure descriptor is fully written before setting ownership bit
188          * (because HW can start executing as soon as we do).
189          */
190         wmb();
191
192         ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
193                 (n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
194
195         stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
196 }
197
198 /* Post NOP WQE to prevent wrap-around in the middle of WR */
199 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
200 {
201         unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
202         if (unlikely(s < qp->sq_max_wqes_per_wr)) {
203                 post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
204                 ind += s;
205         }
206         return ind;
207 }
208
209 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
210 {
211         struct ib_event event;
212         struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
213
214         if (type == MLX4_EVENT_TYPE_PATH_MIG)
215                 to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
216
217         if (ibqp->event_handler) {
218                 event.device     = ibqp->device;
219                 event.element.qp = ibqp;
220                 switch (type) {
221                 case MLX4_EVENT_TYPE_PATH_MIG:
222                         event.event = IB_EVENT_PATH_MIG;
223                         break;
224                 case MLX4_EVENT_TYPE_COMM_EST:
225                         event.event = IB_EVENT_COMM_EST;
226                         break;
227                 case MLX4_EVENT_TYPE_SQ_DRAINED:
228                         event.event = IB_EVENT_SQ_DRAINED;
229                         break;
230                 case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
231                         event.event = IB_EVENT_QP_LAST_WQE_REACHED;
232                         break;
233                 case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
234                         event.event = IB_EVENT_QP_FATAL;
235                         break;
236                 case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
237                         event.event = IB_EVENT_PATH_MIG_ERR;
238                         break;
239                 case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
240                         event.event = IB_EVENT_QP_REQ_ERR;
241                         break;
242                 case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
243                         event.event = IB_EVENT_QP_ACCESS_ERR;
244                         break;
245                 default:
246                         printk(KERN_WARNING "mlx4_ib: Unexpected event type %d "
247                                "on QP %06x\n", type, qp->qpn);
248                         return;
249                 }
250
251                 ibqp->event_handler(&event, ibqp->qp_context);
252         }
253 }
254
255 static int send_wqe_overhead(enum ib_qp_type type, u32 flags)
256 {
257         /*
258          * UD WQEs must have a datagram segment.
259          * RC and UC WQEs might have a remote address segment.
260          * MLX WQEs need two extra inline data segments (for the UD
261          * header and space for the ICRC).
262          */
263         switch (type) {
264         case IB_QPT_UD:
265                 return sizeof (struct mlx4_wqe_ctrl_seg) +
266                         sizeof (struct mlx4_wqe_datagram_seg) +
267                         ((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
268         case IB_QPT_UC:
269                 return sizeof (struct mlx4_wqe_ctrl_seg) +
270                         sizeof (struct mlx4_wqe_raddr_seg);
271         case IB_QPT_RC:
272                 return sizeof (struct mlx4_wqe_ctrl_seg) +
273                         sizeof (struct mlx4_wqe_atomic_seg) +
274                         sizeof (struct mlx4_wqe_raddr_seg);
275         case IB_QPT_SMI:
276         case IB_QPT_GSI:
277                 return sizeof (struct mlx4_wqe_ctrl_seg) +
278                         ALIGN(MLX4_IB_UD_HEADER_SIZE +
279                               DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
280                                            MLX4_INLINE_ALIGN) *
281                               sizeof (struct mlx4_wqe_inline_seg),
282                               sizeof (struct mlx4_wqe_data_seg)) +
283                         ALIGN(4 +
284                               sizeof (struct mlx4_wqe_inline_seg),
285                               sizeof (struct mlx4_wqe_data_seg));
286         default:
287                 return sizeof (struct mlx4_wqe_ctrl_seg);
288         }
289 }
290
291 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
292                        int is_user, int has_srq, struct mlx4_ib_qp *qp)
293 {
294         /* Sanity check RQ size before proceeding */
295         if (cap->max_recv_wr  > dev->dev->caps.max_wqes  ||
296             cap->max_recv_sge > dev->dev->caps.max_rq_sg)
297                 return -EINVAL;
298
299         if (has_srq) {
300                 /* QPs attached to an SRQ should have no RQ */
301                 if (cap->max_recv_wr)
302                         return -EINVAL;
303
304                 qp->rq.wqe_cnt = qp->rq.max_gs = 0;
305         } else {
306                 /* HW requires >= 1 RQ entry with >= 1 gather entry */
307                 if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
308                         return -EINVAL;
309
310                 qp->rq.wqe_cnt   = roundup_pow_of_two(max(1U, cap->max_recv_wr));
311                 qp->rq.max_gs    = roundup_pow_of_two(max(1U, cap->max_recv_sge));
312                 qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
313         }
314
315         cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
316         cap->max_recv_sge = qp->rq.max_gs;
317
318         return 0;
319 }
320
321 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
322                               enum ib_qp_type type, struct mlx4_ib_qp *qp)
323 {
324         int s;
325
326         /* Sanity check SQ size before proceeding */
327         if (cap->max_send_wr     > dev->dev->caps.max_wqes  ||
328             cap->max_send_sge    > dev->dev->caps.max_sq_sg ||
329             cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
330             sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
331                 return -EINVAL;
332
333         /*
334          * For MLX transport we need 2 extra S/G entries:
335          * one for the header and one for the checksum at the end
336          */
337         if ((type == IB_QPT_SMI || type == IB_QPT_GSI) &&
338             cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
339                 return -EINVAL;
340
341         s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
342                 cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
343                 send_wqe_overhead(type, qp->flags);
344
345         if (s > dev->dev->caps.max_sq_desc_sz)
346                 return -EINVAL;
347
348         /*
349          * Hermon supports shrinking WQEs, such that a single work
350          * request can include multiple units of 1 << wqe_shift.  This
351          * way, work requests can differ in size, and do not have to
352          * be a power of 2 in size, saving memory and speeding up send
353          * WR posting.  Unfortunately, if we do this then the
354          * wqe_index field in CQEs can't be used to look up the WR ID
355          * anymore, so we do this only if selective signaling is off.
356          *
357          * Further, on 32-bit platforms, we can't use vmap() to make
358          * the QP buffer virtually contiguous.  Thus we have to use
359          * constant-sized WRs to make sure a WR is always fully within
360          * a single page-sized chunk.
361          *
362          * Finally, we use NOP work requests to pad the end of the
363          * work queue, to avoid wrap-around in the middle of WR.  We
364          * set NEC bit to avoid getting completions with error for
365          * these NOP WRs, but since NEC is only supported starting
366          * with firmware 2.2.232, we use constant-sized WRs for older
367          * firmware.
368          *
369          * And, since MLX QPs only support SEND, we use constant-sized
370          * WRs in this case.
371          *
372          * We look for the smallest value of wqe_shift such that the
373          * resulting number of wqes does not exceed device
374          * capabilities.
375          *
376          * We set WQE size to at least 64 bytes, this way stamping
377          * invalidates each WQE.
378          */
379         if (dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
380             qp->sq_signal_bits && BITS_PER_LONG == 64 &&
381             type != IB_QPT_SMI && type != IB_QPT_GSI)
382                 qp->sq.wqe_shift = ilog2(64);
383         else
384                 qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
385
386         for (;;) {
387                 qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
388
389                 /*
390                  * We need to leave 2 KB + 1 WR of headroom in the SQ to
391                  * allow HW to prefetch.
392                  */
393                 qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
394                 qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
395                                                     qp->sq_max_wqes_per_wr +
396                                                     qp->sq_spare_wqes);
397
398                 if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
399                         break;
400
401                 if (qp->sq_max_wqes_per_wr <= 1)
402                         return -EINVAL;
403
404                 ++qp->sq.wqe_shift;
405         }
406
407         qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
408                              (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
409                          send_wqe_overhead(type, qp->flags)) /
410                 sizeof (struct mlx4_wqe_data_seg);
411
412         qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
413                 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
414         if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
415                 qp->rq.offset = 0;
416                 qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
417         } else {
418                 qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
419                 qp->sq.offset = 0;
420         }
421
422         cap->max_send_wr  = qp->sq.max_post =
423                 (qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
424         cap->max_send_sge = min(qp->sq.max_gs,
425                                 min(dev->dev->caps.max_sq_sg,
426                                     dev->dev->caps.max_rq_sg));
427         /* We don't support inline sends for kernel QPs (yet) */
428         cap->max_inline_data = 0;
429
430         return 0;
431 }
432
433 static int set_user_sq_size(struct mlx4_ib_dev *dev,
434                             struct mlx4_ib_qp *qp,
435                             struct mlx4_ib_create_qp *ucmd)
436 {
437         /* Sanity check SQ size before proceeding */
438         if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes       ||
439             ucmd->log_sq_stride >
440                 ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
441             ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
442                 return -EINVAL;
443
444         qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
445         qp->sq.wqe_shift = ucmd->log_sq_stride;
446
447         qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
448                 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
449
450         return 0;
451 }
452
453 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
454                             struct ib_qp_init_attr *init_attr,
455                             struct ib_udata *udata, int sqpn, struct mlx4_ib_qp *qp)
456 {
457         int qpn;
458         int err;
459
460         mutex_init(&qp->mutex);
461         spin_lock_init(&qp->sq.lock);
462         spin_lock_init(&qp->rq.lock);
463
464         qp->state        = IB_QPS_RESET;
465         if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
466                 qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
467
468         err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, !!init_attr->srq, qp);
469         if (err)
470                 goto err;
471
472         if (pd->uobject) {
473                 struct mlx4_ib_create_qp ucmd;
474
475                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
476                         err = -EFAULT;
477                         goto err;
478                 }
479
480                 qp->sq_no_prefetch = ucmd.sq_no_prefetch;
481
482                 err = set_user_sq_size(dev, qp, &ucmd);
483                 if (err)
484                         goto err;
485
486                 qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
487                                        qp->buf_size, 0, 0);
488                 if (IS_ERR(qp->umem)) {
489                         err = PTR_ERR(qp->umem);
490                         goto err;
491                 }
492
493                 err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
494                                     ilog2(qp->umem->page_size), &qp->mtt);
495                 if (err)
496                         goto err_buf;
497
498                 err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
499                 if (err)
500                         goto err_mtt;
501
502                 if (!init_attr->srq) {
503                         err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
504                                                   ucmd.db_addr, &qp->db);
505                         if (err)
506                                 goto err_mtt;
507                 }
508         } else {
509                 qp->sq_no_prefetch = 0;
510
511                 if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
512                         qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
513
514                 if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
515                         qp->flags |= MLX4_IB_QP_LSO;
516
517                 err = set_kernel_sq_size(dev, &init_attr->cap, init_attr->qp_type, qp);
518                 if (err)
519                         goto err;
520
521                 if (!init_attr->srq) {
522                         err = mlx4_db_alloc(dev->dev, &qp->db, 0);
523                         if (err)
524                                 goto err;
525
526                         *qp->db.db = 0;
527                 }
528
529                 if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf)) {
530                         err = -ENOMEM;
531                         goto err_db;
532                 }
533
534                 err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
535                                     &qp->mtt);
536                 if (err)
537                         goto err_buf;
538
539                 err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
540                 if (err)
541                         goto err_mtt;
542
543                 qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof (u64), GFP_KERNEL);
544                 qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof (u64), GFP_KERNEL);
545
546                 if (!qp->sq.wrid || !qp->rq.wrid) {
547                         err = -ENOMEM;
548                         goto err_wrid;
549                 }
550         }
551
552         if (sqpn) {
553                 qpn = sqpn;
554         } else {
555                 err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn);
556                 if (err)
557                         goto err_wrid;
558         }
559
560         err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
561         if (err)
562                 goto err_qpn;
563
564         /*
565          * Hardware wants QPN written in big-endian order (after
566          * shifting) for send doorbell.  Precompute this value to save
567          * a little bit when posting sends.
568          */
569         qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
570
571         qp->mqp.event = mlx4_ib_qp_event;
572
573         return 0;
574
575 err_qpn:
576         if (!sqpn)
577                 mlx4_qp_release_range(dev->dev, qpn, 1);
578
579 err_wrid:
580         if (pd->uobject) {
581                 if (!init_attr->srq)
582                         mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context),
583                                               &qp->db);
584         } else {
585                 kfree(qp->sq.wrid);
586                 kfree(qp->rq.wrid);
587         }
588
589 err_mtt:
590         mlx4_mtt_cleanup(dev->dev, &qp->mtt);
591
592 err_buf:
593         if (pd->uobject)
594                 ib_umem_release(qp->umem);
595         else
596                 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
597
598 err_db:
599         if (!pd->uobject && !init_attr->srq)
600                 mlx4_db_free(dev->dev, &qp->db);
601
602 err:
603         return err;
604 }
605
606 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
607 {
608         switch (state) {
609         case IB_QPS_RESET:      return MLX4_QP_STATE_RST;
610         case IB_QPS_INIT:       return MLX4_QP_STATE_INIT;
611         case IB_QPS_RTR:        return MLX4_QP_STATE_RTR;
612         case IB_QPS_RTS:        return MLX4_QP_STATE_RTS;
613         case IB_QPS_SQD:        return MLX4_QP_STATE_SQD;
614         case IB_QPS_SQE:        return MLX4_QP_STATE_SQER;
615         case IB_QPS_ERR:        return MLX4_QP_STATE_ERR;
616         default:                return -1;
617         }
618 }
619
620 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
621         __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
622 {
623         if (send_cq == recv_cq) {
624                 spin_lock_irq(&send_cq->lock);
625                 __acquire(&recv_cq->lock);
626         } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
627                 spin_lock_irq(&send_cq->lock);
628                 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
629         } else {
630                 spin_lock_irq(&recv_cq->lock);
631                 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
632         }
633 }
634
635 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
636         __releases(&send_cq->lock) __releases(&recv_cq->lock)
637 {
638         if (send_cq == recv_cq) {
639                 __release(&recv_cq->lock);
640                 spin_unlock_irq(&send_cq->lock);
641         } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
642                 spin_unlock(&recv_cq->lock);
643                 spin_unlock_irq(&send_cq->lock);
644         } else {
645                 spin_unlock(&send_cq->lock);
646                 spin_unlock_irq(&recv_cq->lock);
647         }
648 }
649
650 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
651                               int is_user)
652 {
653         struct mlx4_ib_cq *send_cq, *recv_cq;
654
655         if (qp->state != IB_QPS_RESET)
656                 if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
657                                    MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
658                         printk(KERN_WARNING "mlx4_ib: modify QP %06x to RESET failed.\n",
659                                qp->mqp.qpn);
660
661         send_cq = to_mcq(qp->ibqp.send_cq);
662         recv_cq = to_mcq(qp->ibqp.recv_cq);
663
664         mlx4_ib_lock_cqs(send_cq, recv_cq);
665
666         if (!is_user) {
667                 __mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
668                                  qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
669                 if (send_cq != recv_cq)
670                         __mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
671         }
672
673         mlx4_qp_remove(dev->dev, &qp->mqp);
674
675         mlx4_ib_unlock_cqs(send_cq, recv_cq);
676
677         mlx4_qp_free(dev->dev, &qp->mqp);
678
679         if (!is_sqp(dev, qp))
680                 mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
681
682         mlx4_mtt_cleanup(dev->dev, &qp->mtt);
683
684         if (is_user) {
685                 if (!qp->ibqp.srq)
686                         mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
687                                               &qp->db);
688                 ib_umem_release(qp->umem);
689         } else {
690                 kfree(qp->sq.wrid);
691                 kfree(qp->rq.wrid);
692                 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
693                 if (!qp->ibqp.srq)
694                         mlx4_db_free(dev->dev, &qp->db);
695         }
696 }
697
698 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
699                                 struct ib_qp_init_attr *init_attr,
700                                 struct ib_udata *udata)
701 {
702         struct mlx4_ib_dev *dev = to_mdev(pd->device);
703         struct mlx4_ib_sqp *sqp;
704         struct mlx4_ib_qp *qp;
705         int err;
706
707         /*
708          * We only support LSO and multicast loopback blocking, and
709          * only for kernel UD QPs.
710          */
711         if (init_attr->create_flags & ~(IB_QP_CREATE_IPOIB_UD_LSO |
712                                         IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK))
713                 return ERR_PTR(-EINVAL);
714
715         if (init_attr->create_flags &&
716             (pd->uobject || init_attr->qp_type != IB_QPT_UD))
717                 return ERR_PTR(-EINVAL);
718
719         switch (init_attr->qp_type) {
720         case IB_QPT_RC:
721         case IB_QPT_UC:
722         case IB_QPT_UD:
723         {
724                 qp = kzalloc(sizeof *qp, GFP_KERNEL);
725                 if (!qp)
726                         return ERR_PTR(-ENOMEM);
727
728                 err = create_qp_common(dev, pd, init_attr, udata, 0, qp);
729                 if (err) {
730                         kfree(qp);
731                         return ERR_PTR(err);
732                 }
733
734                 qp->ibqp.qp_num = qp->mqp.qpn;
735
736                 break;
737         }
738         case IB_QPT_SMI:
739         case IB_QPT_GSI:
740         {
741                 /* Userspace is not allowed to create special QPs: */
742                 if (pd->uobject)
743                         return ERR_PTR(-EINVAL);
744
745                 sqp = kzalloc(sizeof *sqp, GFP_KERNEL);
746                 if (!sqp)
747                         return ERR_PTR(-ENOMEM);
748
749                 qp = &sqp->qp;
750
751                 err = create_qp_common(dev, pd, init_attr, udata,
752                                        dev->dev->caps.sqp_start +
753                                        (init_attr->qp_type == IB_QPT_SMI ? 0 : 2) +
754                                        init_attr->port_num - 1,
755                                        qp);
756                 if (err) {
757                         kfree(sqp);
758                         return ERR_PTR(err);
759                 }
760
761                 qp->port        = init_attr->port_num;
762                 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
763
764                 break;
765         }
766         default:
767                 /* Don't support raw QPs */
768                 return ERR_PTR(-EINVAL);
769         }
770
771         return &qp->ibqp;
772 }
773
774 int mlx4_ib_destroy_qp(struct ib_qp *qp)
775 {
776         struct mlx4_ib_dev *dev = to_mdev(qp->device);
777         struct mlx4_ib_qp *mqp = to_mqp(qp);
778
779         if (is_qp0(dev, mqp))
780                 mlx4_CLOSE_PORT(dev->dev, mqp->port);
781
782         destroy_qp_common(dev, mqp, !!qp->pd->uobject);
783
784         if (is_sqp(dev, mqp))
785                 kfree(to_msqp(mqp));
786         else
787                 kfree(mqp);
788
789         return 0;
790 }
791
792 static int to_mlx4_st(enum ib_qp_type type)
793 {
794         switch (type) {
795         case IB_QPT_RC:         return MLX4_QP_ST_RC;
796         case IB_QPT_UC:         return MLX4_QP_ST_UC;
797         case IB_QPT_UD:         return MLX4_QP_ST_UD;
798         case IB_QPT_SMI:
799         case IB_QPT_GSI:        return MLX4_QP_ST_MLX;
800         default:                return -1;
801         }
802 }
803
804 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
805                                    int attr_mask)
806 {
807         u8 dest_rd_atomic;
808         u32 access_flags;
809         u32 hw_access_flags = 0;
810
811         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
812                 dest_rd_atomic = attr->max_dest_rd_atomic;
813         else
814                 dest_rd_atomic = qp->resp_depth;
815
816         if (attr_mask & IB_QP_ACCESS_FLAGS)
817                 access_flags = attr->qp_access_flags;
818         else
819                 access_flags = qp->atomic_rd_en;
820
821         if (!dest_rd_atomic)
822                 access_flags &= IB_ACCESS_REMOTE_WRITE;
823
824         if (access_flags & IB_ACCESS_REMOTE_READ)
825                 hw_access_flags |= MLX4_QP_BIT_RRE;
826         if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
827                 hw_access_flags |= MLX4_QP_BIT_RAE;
828         if (access_flags & IB_ACCESS_REMOTE_WRITE)
829                 hw_access_flags |= MLX4_QP_BIT_RWE;
830
831         return cpu_to_be32(hw_access_flags);
832 }
833
834 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
835                             int attr_mask)
836 {
837         if (attr_mask & IB_QP_PKEY_INDEX)
838                 sqp->pkey_index = attr->pkey_index;
839         if (attr_mask & IB_QP_QKEY)
840                 sqp->qkey = attr->qkey;
841         if (attr_mask & IB_QP_SQ_PSN)
842                 sqp->send_psn = attr->sq_psn;
843 }
844
845 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
846 {
847         path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
848 }
849
850 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
851                          struct mlx4_qp_path *path, u8 port)
852 {
853         path->grh_mylmc     = ah->src_path_bits & 0x7f;
854         path->rlid          = cpu_to_be16(ah->dlid);
855         if (ah->static_rate) {
856                 path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
857                 while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
858                        !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
859                         --path->static_rate;
860         } else
861                 path->static_rate = 0;
862         path->counter_index = 0xff;
863
864         if (ah->ah_flags & IB_AH_GRH) {
865                 if (ah->grh.sgid_index >= dev->dev->caps.gid_table_len[port]) {
866                         printk(KERN_ERR "sgid_index (%u) too large. max is %d\n",
867                                ah->grh.sgid_index, dev->dev->caps.gid_table_len[port] - 1);
868                         return -1;
869                 }
870
871                 path->grh_mylmc |= 1 << 7;
872                 path->mgid_index = ah->grh.sgid_index;
873                 path->hop_limit  = ah->grh.hop_limit;
874                 path->tclass_flowlabel =
875                         cpu_to_be32((ah->grh.traffic_class << 20) |
876                                     (ah->grh.flow_label));
877                 memcpy(path->rgid, ah->grh.dgid.raw, 16);
878         }
879
880         path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
881                 ((port - 1) << 6) | ((ah->sl & 0xf) << 2);
882
883         return 0;
884 }
885
886 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
887                                const struct ib_qp_attr *attr, int attr_mask,
888                                enum ib_qp_state cur_state, enum ib_qp_state new_state)
889 {
890         struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
891         struct mlx4_ib_qp *qp = to_mqp(ibqp);
892         struct mlx4_qp_context *context;
893         enum mlx4_qp_optpar optpar = 0;
894         int sqd_event;
895         int err = -EINVAL;
896
897         context = kzalloc(sizeof *context, GFP_KERNEL);
898         if (!context)
899                 return -ENOMEM;
900
901         context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
902                                      (to_mlx4_st(ibqp->qp_type) << 16));
903
904         if (!(attr_mask & IB_QP_PATH_MIG_STATE))
905                 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
906         else {
907                 optpar |= MLX4_QP_OPTPAR_PM_STATE;
908                 switch (attr->path_mig_state) {
909                 case IB_MIG_MIGRATED:
910                         context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
911                         break;
912                 case IB_MIG_REARM:
913                         context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
914                         break;
915                 case IB_MIG_ARMED:
916                         context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
917                         break;
918                 }
919         }
920
921         if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
922                 context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
923         else if (ibqp->qp_type == IB_QPT_UD) {
924                 if (qp->flags & MLX4_IB_QP_LSO)
925                         context->mtu_msgmax = (IB_MTU_4096 << 5) |
926                                               ilog2(dev->dev->caps.max_gso_sz);
927                 else
928                         context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
929         } else if (attr_mask & IB_QP_PATH_MTU) {
930                 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
931                         printk(KERN_ERR "path MTU (%u) is invalid\n",
932                                attr->path_mtu);
933                         goto out;
934                 }
935                 context->mtu_msgmax = (attr->path_mtu << 5) |
936                         ilog2(dev->dev->caps.max_msg_sz);
937         }
938
939         if (qp->rq.wqe_cnt)
940                 context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
941         context->rq_size_stride |= qp->rq.wqe_shift - 4;
942
943         if (qp->sq.wqe_cnt)
944                 context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
945         context->sq_size_stride |= qp->sq.wqe_shift - 4;
946
947         if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
948                 context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
949
950         if (qp->ibqp.uobject)
951                 context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
952         else
953                 context->usr_page = cpu_to_be32(dev->priv_uar.index);
954
955         if (attr_mask & IB_QP_DEST_QPN)
956                 context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
957
958         if (attr_mask & IB_QP_PORT) {
959                 if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
960                     !(attr_mask & IB_QP_AV)) {
961                         mlx4_set_sched(&context->pri_path, attr->port_num);
962                         optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
963                 }
964         }
965
966         if (attr_mask & IB_QP_PKEY_INDEX) {
967                 context->pri_path.pkey_index = attr->pkey_index;
968                 optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
969         }
970
971         if (attr_mask & IB_QP_AV) {
972                 if (mlx4_set_path(dev, &attr->ah_attr, &context->pri_path,
973                                   attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
974                         goto out;
975
976                 optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
977                            MLX4_QP_OPTPAR_SCHED_QUEUE);
978         }
979
980         if (attr_mask & IB_QP_TIMEOUT) {
981                 context->pri_path.ackto = attr->timeout << 3;
982                 optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
983         }
984
985         if (attr_mask & IB_QP_ALT_PATH) {
986                 if (attr->alt_port_num == 0 ||
987                     attr->alt_port_num > dev->dev->caps.num_ports)
988                         goto out;
989
990                 if (attr->alt_pkey_index >=
991                     dev->dev->caps.pkey_table_len[attr->alt_port_num])
992                         goto out;
993
994                 if (mlx4_set_path(dev, &attr->alt_ah_attr, &context->alt_path,
995                                   attr->alt_port_num))
996                         goto out;
997
998                 context->alt_path.pkey_index = attr->alt_pkey_index;
999                 context->alt_path.ackto = attr->alt_timeout << 3;
1000                 optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
1001         }
1002
1003         context->pd         = cpu_to_be32(to_mpd(ibqp->pd)->pdn);
1004         context->params1    = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1005
1006         /* Set "fast registration enabled" for all kernel QPs */
1007         if (!qp->ibqp.uobject)
1008                 context->params1 |= cpu_to_be32(1 << 11);
1009
1010         if (attr_mask & IB_QP_RNR_RETRY) {
1011                 context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1012                 optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1013         }
1014
1015         if (attr_mask & IB_QP_RETRY_CNT) {
1016                 context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1017                 optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1018         }
1019
1020         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1021                 if (attr->max_rd_atomic)
1022                         context->params1 |=
1023                                 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1024                 optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1025         }
1026
1027         if (attr_mask & IB_QP_SQ_PSN)
1028                 context->next_send_psn = cpu_to_be32(attr->sq_psn);
1029
1030         context->cqn_send = cpu_to_be32(to_mcq(ibqp->send_cq)->mcq.cqn);
1031
1032         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1033                 if (attr->max_dest_rd_atomic)
1034                         context->params2 |=
1035                                 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1036                 optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1037         }
1038
1039         if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1040                 context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1041                 optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1042         }
1043
1044         if (ibqp->srq)
1045                 context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1046
1047         if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1048                 context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1049                 optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1050         }
1051         if (attr_mask & IB_QP_RQ_PSN)
1052                 context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1053
1054         context->cqn_recv = cpu_to_be32(to_mcq(ibqp->recv_cq)->mcq.cqn);
1055
1056         if (attr_mask & IB_QP_QKEY) {
1057                 context->qkey = cpu_to_be32(attr->qkey);
1058                 optpar |= MLX4_QP_OPTPAR_Q_KEY;
1059         }
1060
1061         if (ibqp->srq)
1062                 context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1063
1064         if (!ibqp->srq && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1065                 context->db_rec_addr = cpu_to_be64(qp->db.dma);
1066
1067         if (cur_state == IB_QPS_INIT &&
1068             new_state == IB_QPS_RTR  &&
1069             (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1070              ibqp->qp_type == IB_QPT_UD)) {
1071                 context->pri_path.sched_queue = (qp->port - 1) << 6;
1072                 if (is_qp0(dev, qp))
1073                         context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1074                 else
1075                         context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1076         }
1077
1078         if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD  &&
1079             attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1080                 sqd_event = 1;
1081         else
1082                 sqd_event = 0;
1083
1084         if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1085                 context->rlkey |= (1 << 4);
1086
1087         /*
1088          * Before passing a kernel QP to the HW, make sure that the
1089          * ownership bits of the send queue are set and the SQ
1090          * headroom is stamped so that the hardware doesn't start
1091          * processing stale work requests.
1092          */
1093         if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1094                 struct mlx4_wqe_ctrl_seg *ctrl;
1095                 int i;
1096
1097                 for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1098                         ctrl = get_send_wqe(qp, i);
1099                         ctrl->owner_opcode = cpu_to_be32(1 << 31);
1100                         if (qp->sq_max_wqes_per_wr == 1)
1101                                 ctrl->fence_size = 1 << (qp->sq.wqe_shift - 4);
1102
1103                         stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1104                 }
1105         }
1106
1107         err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
1108                              to_mlx4_state(new_state), context, optpar,
1109                              sqd_event, &qp->mqp);
1110         if (err)
1111                 goto out;
1112
1113         qp->state = new_state;
1114
1115         if (attr_mask & IB_QP_ACCESS_FLAGS)
1116                 qp->atomic_rd_en = attr->qp_access_flags;
1117         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1118                 qp->resp_depth = attr->max_dest_rd_atomic;
1119         if (attr_mask & IB_QP_PORT)
1120                 qp->port = attr->port_num;
1121         if (attr_mask & IB_QP_ALT_PATH)
1122                 qp->alt_port = attr->alt_port_num;
1123
1124         if (is_sqp(dev, qp))
1125                 store_sqp_attrs(to_msqp(qp), attr, attr_mask);
1126
1127         /*
1128          * If we moved QP0 to RTR, bring the IB link up; if we moved
1129          * QP0 to RESET or ERROR, bring the link back down.
1130          */
1131         if (is_qp0(dev, qp)) {
1132                 if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
1133                         if (mlx4_INIT_PORT(dev->dev, qp->port))
1134                                 printk(KERN_WARNING "INIT_PORT failed for port %d\n",
1135                                        qp->port);
1136
1137                 if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1138                     (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1139                         mlx4_CLOSE_PORT(dev->dev, qp->port);
1140         }
1141
1142         /*
1143          * If we moved a kernel QP to RESET, clean up all old CQ
1144          * entries and reinitialize the QP.
1145          */
1146         if (new_state == IB_QPS_RESET && !ibqp->uobject) {
1147                 mlx4_ib_cq_clean(to_mcq(ibqp->recv_cq), qp->mqp.qpn,
1148                                  ibqp->srq ? to_msrq(ibqp->srq): NULL);
1149                 if (ibqp->send_cq != ibqp->recv_cq)
1150                         mlx4_ib_cq_clean(to_mcq(ibqp->send_cq), qp->mqp.qpn, NULL);
1151
1152                 qp->rq.head = 0;
1153                 qp->rq.tail = 0;
1154                 qp->sq.head = 0;
1155                 qp->sq.tail = 0;
1156                 qp->sq_next_wqe = 0;
1157                 if (!ibqp->srq)
1158                         *qp->db.db  = 0;
1159         }
1160
1161 out:
1162         kfree(context);
1163         return err;
1164 }
1165
1166 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1167                       int attr_mask, struct ib_udata *udata)
1168 {
1169         struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1170         struct mlx4_ib_qp *qp = to_mqp(ibqp);
1171         enum ib_qp_state cur_state, new_state;
1172         int err = -EINVAL;
1173
1174         mutex_lock(&qp->mutex);
1175
1176         cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1177         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1178
1179         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask))
1180                 goto out;
1181
1182         if ((attr_mask & IB_QP_PORT) &&
1183             (attr->port_num == 0 || attr->port_num > dev->dev->caps.num_ports)) {
1184                 goto out;
1185         }
1186
1187         if (attr_mask & IB_QP_PKEY_INDEX) {
1188                 int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1189                 if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p])
1190                         goto out;
1191         }
1192
1193         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1194             attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
1195                 goto out;
1196         }
1197
1198         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1199             attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
1200                 goto out;
1201         }
1202
1203         if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1204                 err = 0;
1205                 goto out;
1206         }
1207
1208         err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
1209
1210 out:
1211         mutex_unlock(&qp->mutex);
1212         return err;
1213 }
1214
1215 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
1216                             void *wqe, unsigned *mlx_seg_len)
1217 {
1218         struct ib_device *ib_dev = sqp->qp.ibqp.device;
1219         struct mlx4_wqe_mlx_seg *mlx = wqe;
1220         struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
1221         struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
1222         u16 pkey;
1223         int send_size;
1224         int header_size;
1225         int spc;
1226         int i;
1227
1228         send_size = 0;
1229         for (i = 0; i < wr->num_sge; ++i)
1230                 send_size += wr->sg_list[i].length;
1231
1232         ib_ud_header_init(send_size, mlx4_ib_ah_grh_present(ah), 0, &sqp->ud_header);
1233
1234         sqp->ud_header.lrh.service_level   =
1235                 be32_to_cpu(ah->av.sl_tclass_flowlabel) >> 28;
1236         sqp->ud_header.lrh.destination_lid = ah->av.dlid;
1237         sqp->ud_header.lrh.source_lid      = cpu_to_be16(ah->av.g_slid & 0x7f);
1238         if (mlx4_ib_ah_grh_present(ah)) {
1239                 sqp->ud_header.grh.traffic_class =
1240                         (be32_to_cpu(ah->av.sl_tclass_flowlabel) >> 20) & 0xff;
1241                 sqp->ud_header.grh.flow_label    =
1242                         ah->av.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
1243                 sqp->ud_header.grh.hop_limit     = ah->av.hop_limit;
1244                 ib_get_cached_gid(ib_dev, be32_to_cpu(ah->av.port_pd) >> 24,
1245                                   ah->av.gid_index, &sqp->ud_header.grh.source_gid);
1246                 memcpy(sqp->ud_header.grh.destination_gid.raw,
1247                        ah->av.dgid, 16);
1248         }
1249
1250         mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1251         mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
1252                                   (sqp->ud_header.lrh.destination_lid ==
1253                                    IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
1254                                   (sqp->ud_header.lrh.service_level << 8));
1255         mlx->rlid   = sqp->ud_header.lrh.destination_lid;
1256
1257         switch (wr->opcode) {
1258         case IB_WR_SEND:
1259                 sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
1260                 sqp->ud_header.immediate_present = 0;
1261                 break;
1262         case IB_WR_SEND_WITH_IMM:
1263                 sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1264                 sqp->ud_header.immediate_present = 1;
1265                 sqp->ud_header.immediate_data    = wr->ex.imm_data;
1266                 break;
1267         default:
1268                 return -EINVAL;
1269         }
1270
1271         sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1272         if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1273                 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1274         sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1275         if (!sqp->qp.ibqp.qp_num)
1276                 ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
1277         else
1278                 ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey);
1279         sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1280         sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1281         sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1282         sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1283                                                sqp->qkey : wr->wr.ud.remote_qkey);
1284         sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1285
1286         header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
1287
1288         if (0) {
1289                 printk(KERN_ERR "built UD header of size %d:\n", header_size);
1290                 for (i = 0; i < header_size / 4; ++i) {
1291                         if (i % 8 == 0)
1292                                 printk("  [%02x] ", i * 4);
1293                         printk(" %08x",
1294                                be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
1295                         if ((i + 1) % 8 == 0)
1296                                 printk("\n");
1297                 }
1298                 printk("\n");
1299         }
1300
1301         /*
1302          * Inline data segments may not cross a 64 byte boundary.  If
1303          * our UD header is bigger than the space available up to the
1304          * next 64 byte boundary in the WQE, use two inline data
1305          * segments to hold the UD header.
1306          */
1307         spc = MLX4_INLINE_ALIGN -
1308                 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
1309         if (header_size <= spc) {
1310                 inl->byte_count = cpu_to_be32(1 << 31 | header_size);
1311                 memcpy(inl + 1, sqp->header_buf, header_size);
1312                 i = 1;
1313         } else {
1314                 inl->byte_count = cpu_to_be32(1 << 31 | spc);
1315                 memcpy(inl + 1, sqp->header_buf, spc);
1316
1317                 inl = (void *) (inl + 1) + spc;
1318                 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
1319                 /*
1320                  * Need a barrier here to make sure all the data is
1321                  * visible before the byte_count field is set.
1322                  * Otherwise the HCA prefetcher could grab the 64-byte
1323                  * chunk with this inline segment and get a valid (!=
1324                  * 0xffffffff) byte count but stale data, and end up
1325                  * generating a packet with bad headers.
1326                  *
1327                  * The first inline segment's byte_count field doesn't
1328                  * need a barrier, because it comes after a
1329                  * control/MLX segment and therefore is at an offset
1330                  * of 16 mod 64.
1331                  */
1332                 wmb();
1333                 inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
1334                 i = 2;
1335         }
1336
1337         *mlx_seg_len =
1338                 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
1339         return 0;
1340 }
1341
1342 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
1343 {
1344         unsigned cur;
1345         struct mlx4_ib_cq *cq;
1346
1347         cur = wq->head - wq->tail;
1348         if (likely(cur + nreq < wq->max_post))
1349                 return 0;
1350
1351         cq = to_mcq(ib_cq);
1352         spin_lock(&cq->lock);
1353         cur = wq->head - wq->tail;
1354         spin_unlock(&cq->lock);
1355
1356         return cur + nreq >= wq->max_post;
1357 }
1358
1359 static __be32 convert_access(int acc)
1360 {
1361         return (acc & IB_ACCESS_REMOTE_ATOMIC ? cpu_to_be32(MLX4_WQE_FMR_PERM_ATOMIC)       : 0) |
1362                (acc & IB_ACCESS_REMOTE_WRITE  ? cpu_to_be32(MLX4_WQE_FMR_PERM_REMOTE_WRITE) : 0) |
1363                (acc & IB_ACCESS_REMOTE_READ   ? cpu_to_be32(MLX4_WQE_FMR_PERM_REMOTE_READ)  : 0) |
1364                (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
1365                 cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
1366 }
1367
1368 static void set_fmr_seg(struct mlx4_wqe_fmr_seg *fseg, struct ib_send_wr *wr)
1369 {
1370         struct mlx4_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
1371         int i;
1372
1373         for (i = 0; i < wr->wr.fast_reg.page_list_len; ++i)
1374                 mfrpl->mapped_page_list[i] =
1375                         cpu_to_be64(wr->wr.fast_reg.page_list->page_list[i] |
1376                                     MLX4_MTT_FLAG_PRESENT);
1377
1378         fseg->flags             = convert_access(wr->wr.fast_reg.access_flags);
1379         fseg->mem_key           = cpu_to_be32(wr->wr.fast_reg.rkey);
1380         fseg->buf_list          = cpu_to_be64(mfrpl->map);
1381         fseg->start_addr        = cpu_to_be64(wr->wr.fast_reg.iova_start);
1382         fseg->reg_len           = cpu_to_be64(wr->wr.fast_reg.length);
1383         fseg->offset            = 0; /* XXX -- is this just for ZBVA? */
1384         fseg->page_size         = cpu_to_be32(wr->wr.fast_reg.page_shift);
1385         fseg->reserved[0]       = 0;
1386         fseg->reserved[1]       = 0;
1387 }
1388
1389 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
1390 {
1391         iseg->flags     = 0;
1392         iseg->mem_key   = cpu_to_be32(rkey);
1393         iseg->guest_id  = 0;
1394         iseg->pa        = 0;
1395 }
1396
1397 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
1398                                           u64 remote_addr, u32 rkey)
1399 {
1400         rseg->raddr    = cpu_to_be64(remote_addr);
1401         rseg->rkey     = cpu_to_be32(rkey);
1402         rseg->reserved = 0;
1403 }
1404
1405 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, struct ib_send_wr *wr)
1406 {
1407         if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1408                 aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
1409                 aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add);
1410         } else {
1411                 aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
1412                 aseg->compare  = 0;
1413         }
1414
1415 }
1416
1417 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
1418                              struct ib_send_wr *wr)
1419 {
1420         memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av));
1421         dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1422         dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1423 }
1424
1425 static void set_mlx_icrc_seg(void *dseg)
1426 {
1427         u32 *t = dseg;
1428         struct mlx4_wqe_inline_seg *iseg = dseg;
1429
1430         t[1] = 0;
1431
1432         /*
1433          * Need a barrier here before writing the byte_count field to
1434          * make sure that all the data is visible before the
1435          * byte_count field is set.  Otherwise, if the segment begins
1436          * a new cacheline, the HCA prefetcher could grab the 64-byte
1437          * chunk and get a valid (!= * 0xffffffff) byte count but
1438          * stale data, and end up sending the wrong data.
1439          */
1440         wmb();
1441
1442         iseg->byte_count = cpu_to_be32((1 << 31) | 4);
1443 }
1444
1445 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
1446 {
1447         dseg->lkey       = cpu_to_be32(sg->lkey);
1448         dseg->addr       = cpu_to_be64(sg->addr);
1449
1450         /*
1451          * Need a barrier here before writing the byte_count field to
1452          * make sure that all the data is visible before the
1453          * byte_count field is set.  Otherwise, if the segment begins
1454          * a new cacheline, the HCA prefetcher could grab the 64-byte
1455          * chunk and get a valid (!= * 0xffffffff) byte count but
1456          * stale data, and end up sending the wrong data.
1457          */
1458         wmb();
1459
1460         dseg->byte_count = cpu_to_be32(sg->length);
1461 }
1462
1463 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
1464 {
1465         dseg->byte_count = cpu_to_be32(sg->length);
1466         dseg->lkey       = cpu_to_be32(sg->lkey);
1467         dseg->addr       = cpu_to_be64(sg->addr);
1468 }
1469
1470 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_send_wr *wr,
1471                          struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
1472                          __be32 *lso_hdr_sz, __be32 *blh)
1473 {
1474         unsigned halign = ALIGN(sizeof *wqe + wr->wr.ud.hlen, 16);
1475
1476         if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
1477                 *blh = cpu_to_be32(1 << 6);
1478
1479         if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
1480                      wr->num_sge > qp->sq.max_gs - (halign >> 4)))
1481                 return -EINVAL;
1482
1483         memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen);
1484
1485         *lso_hdr_sz  = cpu_to_be32((wr->wr.ud.mss - wr->wr.ud.hlen) << 16 |
1486                                    wr->wr.ud.hlen);
1487         *lso_seg_len = halign;
1488         return 0;
1489 }
1490
1491 static __be32 send_ieth(struct ib_send_wr *wr)
1492 {
1493         switch (wr->opcode) {
1494         case IB_WR_SEND_WITH_IMM:
1495         case IB_WR_RDMA_WRITE_WITH_IMM:
1496                 return wr->ex.imm_data;
1497
1498         case IB_WR_SEND_WITH_INV:
1499                 return cpu_to_be32(wr->ex.invalidate_rkey);
1500
1501         default:
1502                 return 0;
1503         }
1504 }
1505
1506 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1507                       struct ib_send_wr **bad_wr)
1508 {
1509         struct mlx4_ib_qp *qp = to_mqp(ibqp);
1510         void *wqe;
1511         struct mlx4_wqe_ctrl_seg *ctrl;
1512         struct mlx4_wqe_data_seg *dseg;
1513         unsigned long flags;
1514         int nreq;
1515         int err = 0;
1516         unsigned ind;
1517         int uninitialized_var(stamp);
1518         int uninitialized_var(size);
1519         unsigned uninitialized_var(seglen);
1520         __be32 dummy;
1521         __be32 *lso_wqe;
1522         __be32 uninitialized_var(lso_hdr_sz);
1523         __be32 blh;
1524         int i;
1525
1526         spin_lock_irqsave(&qp->sq.lock, flags);
1527
1528         ind = qp->sq_next_wqe;
1529
1530         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1531                 lso_wqe = &dummy;
1532                 blh = 0;
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, &blh);
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) | blh;
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.recv_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