Linux Kernel  3.7.1
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
mthca_qp.c
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2004 Topspin Communications. All rights reserved.
3  * Copyright (c) 2005 Cisco Systems. All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
6  *
7  * This software is available to you under a choice of one of two
8  * licenses. You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the
11  * OpenIB.org BSD license below:
12  *
13  * Redistribution and use in source and binary forms, with or
14  * without modification, are permitted provided that the following
15  * conditions are met:
16  *
17  * - Redistributions of source code must retain the above
18  * copyright notice, this list of conditions and the following
19  * disclaimer.
20  *
21  * - Redistributions in binary form must reproduce the above
22  * copyright notice, this list of conditions and the following
23  * disclaimer in the documentation and/or other materials
24  * provided with the distribution.
25  *
26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33  * SOFTWARE.
34  */
35 
36 #include <linux/string.h>
37 #include <linux/slab.h>
38 #include <linux/sched.h>
39 
40 #include <asm/io.h>
41 
42 #include <rdma/ib_verbs.h>
43 #include <rdma/ib_cache.h>
44 #include <rdma/ib_pack.h>
45 
46 #include "mthca_dev.h"
47 #include "mthca_cmd.h"
48 #include "mthca_memfree.h"
49 #include "mthca_wqe.h"
50 
51 enum {
55  MTHCA_UD_HEADER_SIZE = 72, /* largest UD header possible */
56  MTHCA_INLINE_HEADER_SIZE = 4, /* data segment overhead for inline */
57  MTHCA_INLINE_CHUNK_SIZE = 16 /* inline data segment chunk */
58 };
59 
60 enum {
69 };
70 
71 enum {
77 };
78 
79 enum {
83 };
84 
85 enum {
86  /* qp_context flags */
87  MTHCA_QP_BIT_DE = 1 << 8,
88  /* params1 */
89  MTHCA_QP_BIT_SRE = 1 << 15,
90  MTHCA_QP_BIT_SWE = 1 << 14,
91  MTHCA_QP_BIT_SAE = 1 << 13,
92  MTHCA_QP_BIT_SIC = 1 << 4,
93  MTHCA_QP_BIT_SSC = 1 << 3,
94  /* params2 */
95  MTHCA_QP_BIT_RRE = 1 << 15,
96  MTHCA_QP_BIT_RWE = 1 << 14,
97  MTHCA_QP_BIT_RAE = 1 << 13,
98  MTHCA_QP_BIT_RIC = 1 << 4,
100 };
101 
102 enum {
104 };
105 
116  u8 rgid[16];
117 } __attribute__((packed));
121  __be32 tavor_sched_queue; /* Reserved on Arbel */
123  u8 rq_size_stride; /* Reserved on Tavor */
124  u8 sq_size_stride; /* Reserved on Tavor */
125  u8 rlkey_arbel_sched_queue; /* Reserved on Tavor */
140  __be32 snd_wqe_base_l; /* Next send WQE on Tavor */
141  __be32 snd_db_index; /* (debugging only entries) */
148  __be32 rcv_wqe_base_l; /* Next recv WQE on Tavor */
149  __be32 rcv_db_index; /* (debugging only entries) */
153  __be16 rq_wqe_counter; /* reserved on Tavor */
154  __be16 sq_wqe_counter; /* reserved on Tavor */
156 } __attribute__((packed));
163 } __attribute__((packed));
165 enum {
183 };
185 static const u8 mthca_opcode[] = {
193 };
194 
195 static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
196 {
197  return qp->qpn >= dev->qp_table.sqp_start &&
198  qp->qpn <= dev->qp_table.sqp_start + 3;
199 }
200 
201 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
202 {
203  return qp->qpn >= dev->qp_table.sqp_start &&
204  qp->qpn <= dev->qp_table.sqp_start + 1;
205 }
206 
207 static void *get_recv_wqe(struct mthca_qp *qp, int n)
208 {
209  if (qp->is_direct)
210  return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
211  else
212  return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
213  ((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
214 }
215 
216 static void *get_send_wqe(struct mthca_qp *qp, int n)
217 {
218  if (qp->is_direct)
219  return qp->queue.direct.buf + qp->send_wqe_offset +
220  (n << qp->sq.wqe_shift);
221  else
222  return qp->queue.page_list[(qp->send_wqe_offset +
223  (n << qp->sq.wqe_shift)) >>
224  PAGE_SHIFT].buf +
225  ((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
226  (PAGE_SIZE - 1));
227 }
228 
229 static void mthca_wq_reset(struct mthca_wq *wq)
230 {
231  wq->next_ind = 0;
232  wq->last_comp = wq->max - 1;
233  wq->head = 0;
234  wq->tail = 0;
235 }
236 
237 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
239 {
240  struct mthca_qp *qp;
241  struct ib_event event;
242 
243  spin_lock(&dev->qp_table.lock);
244  qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
245  if (qp)
246  ++qp->refcount;
247  spin_unlock(&dev->qp_table.lock);
248 
249  if (!qp) {
250  mthca_warn(dev, "Async event %d for bogus QP %08x\n",
251  event_type, qpn);
252  return;
253  }
254 
255  if (event_type == IB_EVENT_PATH_MIG)
256  qp->port = qp->alt_port;
257 
258  event.device = &dev->ib_dev;
259  event.event = event_type;
260  event.element.qp = &qp->ibqp;
261  if (qp->ibqp.event_handler)
262  qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
263 
264  spin_lock(&dev->qp_table.lock);
265  if (!--qp->refcount)
266  wake_up(&qp->wait);
267  spin_unlock(&dev->qp_table.lock);
268 }
269 
270 static int to_mthca_state(enum ib_qp_state ib_state)
271 {
272  switch (ib_state) {
273  case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
274  case IB_QPS_INIT: return MTHCA_QP_STATE_INIT;
275  case IB_QPS_RTR: return MTHCA_QP_STATE_RTR;
276  case IB_QPS_RTS: return MTHCA_QP_STATE_RTS;
277  case IB_QPS_SQD: return MTHCA_QP_STATE_SQD;
278  case IB_QPS_SQE: return MTHCA_QP_STATE_SQE;
279  case IB_QPS_ERR: return MTHCA_QP_STATE_ERR;
280  default: return -1;
281  }
282 }
283 
284 enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
285 
286 static int to_mthca_st(int transport)
287 {
288  switch (transport) {
289  case RC: return MTHCA_QP_ST_RC;
290  case UC: return MTHCA_QP_ST_UC;
291  case UD: return MTHCA_QP_ST_UD;
292  case RD: return MTHCA_QP_ST_RD;
293  case MLX: return MTHCA_QP_ST_MLX;
294  default: return -1;
295  }
296 }
297 
298 static void store_attrs(struct mthca_sqp *sqp, const struct ib_qp_attr *attr,
299  int attr_mask)
300 {
301  if (attr_mask & IB_QP_PKEY_INDEX)
302  sqp->pkey_index = attr->pkey_index;
303  if (attr_mask & IB_QP_QKEY)
304  sqp->qkey = attr->qkey;
305  if (attr_mask & IB_QP_SQ_PSN)
306  sqp->send_psn = attr->sq_psn;
307 }
308 
309 static void init_port(struct mthca_dev *dev, int port)
310 {
311  int err;
312  struct mthca_init_ib_param param;
313 
314  memset(&param, 0, sizeof param);
315 
316  param.port_width = dev->limits.port_width_cap;
317  param.vl_cap = dev->limits.vl_cap;
318  param.mtu_cap = dev->limits.mtu_cap;
319  param.gid_cap = dev->limits.gid_table_len;
320  param.pkey_cap = dev->limits.pkey_table_len;
321 
322  err = mthca_INIT_IB(dev, &param, port);
323  if (err)
324  mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
325 }
326 
327 static __be32 get_hw_access_flags(struct mthca_qp *qp, const struct ib_qp_attr *attr,
328  int attr_mask)
329 {
330  u8 dest_rd_atomic;
331  u32 access_flags;
332  u32 hw_access_flags = 0;
333 
334  if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
335  dest_rd_atomic = attr->max_dest_rd_atomic;
336  else
337  dest_rd_atomic = qp->resp_depth;
338 
339  if (attr_mask & IB_QP_ACCESS_FLAGS)
340  access_flags = attr->qp_access_flags;
341  else
342  access_flags = qp->atomic_rd_en;
343 
344  if (!dest_rd_atomic)
345  access_flags &= IB_ACCESS_REMOTE_WRITE;
346 
347  if (access_flags & IB_ACCESS_REMOTE_READ)
348  hw_access_flags |= MTHCA_QP_BIT_RRE;
349  if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
350  hw_access_flags |= MTHCA_QP_BIT_RAE;
351  if (access_flags & IB_ACCESS_REMOTE_WRITE)
352  hw_access_flags |= MTHCA_QP_BIT_RWE;
353 
354  return cpu_to_be32(hw_access_flags);
355 }
356 
357 static inline enum ib_qp_state to_ib_qp_state(int mthca_state)
358 {
359  switch (mthca_state) {
360  case MTHCA_QP_STATE_RST: return IB_QPS_RESET;
361  case MTHCA_QP_STATE_INIT: return IB_QPS_INIT;
362  case MTHCA_QP_STATE_RTR: return IB_QPS_RTR;
363  case MTHCA_QP_STATE_RTS: return IB_QPS_RTS;
365  case MTHCA_QP_STATE_SQD: return IB_QPS_SQD;
366  case MTHCA_QP_STATE_SQE: return IB_QPS_SQE;
367  case MTHCA_QP_STATE_ERR: return IB_QPS_ERR;
368  default: return -1;
369  }
370 }
371 
372 static inline enum ib_mig_state to_ib_mig_state(int mthca_mig_state)
373 {
374  switch (mthca_mig_state) {
375  case 0: return IB_MIG_ARMED;
376  case 1: return IB_MIG_REARM;
377  case 3: return IB_MIG_MIGRATED;
378  default: return -1;
379  }
380 }
381 
382 static int to_ib_qp_access_flags(int mthca_flags)
383 {
384  int ib_flags = 0;
385 
386  if (mthca_flags & MTHCA_QP_BIT_RRE)
387  ib_flags |= IB_ACCESS_REMOTE_READ;
388  if (mthca_flags & MTHCA_QP_BIT_RWE)
389  ib_flags |= IB_ACCESS_REMOTE_WRITE;
390  if (mthca_flags & MTHCA_QP_BIT_RAE)
391  ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
392 
393  return ib_flags;
394 }
395 
396 static void to_ib_ah_attr(struct mthca_dev *dev, struct ib_ah_attr *ib_ah_attr,
397  struct mthca_qp_path *path)
398 {
399  memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
400  ib_ah_attr->port_num = (be32_to_cpu(path->port_pkey) >> 24) & 0x3;
401 
402  if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->limits.num_ports)
403  return;
404 
405  ib_ah_attr->dlid = be16_to_cpu(path->rlid);
406  ib_ah_attr->sl = be32_to_cpu(path->sl_tclass_flowlabel) >> 28;
407  ib_ah_attr->src_path_bits = path->g_mylmc & 0x7f;
408  ib_ah_attr->static_rate = mthca_rate_to_ib(dev,
409  path->static_rate & 0xf,
410  ib_ah_attr->port_num);
411  ib_ah_attr->ah_flags = (path->g_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
412  if (ib_ah_attr->ah_flags) {
413  ib_ah_attr->grh.sgid_index = path->mgid_index & (dev->limits.gid_table_len - 1);
414  ib_ah_attr->grh.hop_limit = path->hop_limit;
415  ib_ah_attr->grh.traffic_class =
416  (be32_to_cpu(path->sl_tclass_flowlabel) >> 20) & 0xff;
417  ib_ah_attr->grh.flow_label =
418  be32_to_cpu(path->sl_tclass_flowlabel) & 0xfffff;
419  memcpy(ib_ah_attr->grh.dgid.raw,
420  path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
421  }
422 }
423 
424 int mthca_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
425  struct ib_qp_init_attr *qp_init_attr)
426 {
427  struct mthca_dev *dev = to_mdev(ibqp->device);
428  struct mthca_qp *qp = to_mqp(ibqp);
429  int err = 0;
430  struct mthca_mailbox *mailbox = NULL;
431  struct mthca_qp_param *qp_param;
432  struct mthca_qp_context *context;
433  int mthca_state;
434 
435  mutex_lock(&qp->mutex);
436 
437  if (qp->state == IB_QPS_RESET) {
438  qp_attr->qp_state = IB_QPS_RESET;
439  goto done;
440  }
441 
442  mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
443  if (IS_ERR(mailbox)) {
444  err = PTR_ERR(mailbox);
445  goto out;
446  }
447 
448  err = mthca_QUERY_QP(dev, qp->qpn, 0, mailbox);
449  if (err) {
450  mthca_warn(dev, "QUERY_QP failed (%d)\n", err);
451  goto out_mailbox;
452  }
453 
454  qp_param = mailbox->buf;
455  context = &qp_param->context;
456  mthca_state = be32_to_cpu(context->flags) >> 28;
457 
458  qp->state = to_ib_qp_state(mthca_state);
459  qp_attr->qp_state = qp->state;
460  qp_attr->path_mtu = context->mtu_msgmax >> 5;
461  qp_attr->path_mig_state =
462  to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
463  qp_attr->qkey = be32_to_cpu(context->qkey);
464  qp_attr->rq_psn = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
465  qp_attr->sq_psn = be32_to_cpu(context->next_send_psn) & 0xffffff;
466  qp_attr->dest_qp_num = be32_to_cpu(context->remote_qpn) & 0xffffff;
467  qp_attr->qp_access_flags =
468  to_ib_qp_access_flags(be32_to_cpu(context->params2));
469 
470  if (qp->transport == RC || qp->transport == UC) {
471  to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
472  to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
473  qp_attr->alt_pkey_index =
474  be32_to_cpu(context->alt_path.port_pkey) & 0x7f;
475  qp_attr->alt_port_num = qp_attr->alt_ah_attr.port_num;
476  }
477 
478  qp_attr->pkey_index = be32_to_cpu(context->pri_path.port_pkey) & 0x7f;
479  qp_attr->port_num =
480  (be32_to_cpu(context->pri_path.port_pkey) >> 24) & 0x3;
481 
482  /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
483  qp_attr->sq_draining = mthca_state == MTHCA_QP_STATE_DRAINING;
484 
485  qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
486 
487  qp_attr->max_dest_rd_atomic =
488  1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
489  qp_attr->min_rnr_timer =
490  (be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
491  qp_attr->timeout = context->pri_path.ackto >> 3;
492  qp_attr->retry_cnt = (be32_to_cpu(context->params1) >> 16) & 0x7;
493  qp_attr->rnr_retry = context->pri_path.rnr_retry >> 5;
494  qp_attr->alt_timeout = context->alt_path.ackto >> 3;
495 
496 done:
497  qp_attr->cur_qp_state = qp_attr->qp_state;
498  qp_attr->cap.max_send_wr = qp->sq.max;
499  qp_attr->cap.max_recv_wr = qp->rq.max;
500  qp_attr->cap.max_send_sge = qp->sq.max_gs;
501  qp_attr->cap.max_recv_sge = qp->rq.max_gs;
502  qp_attr->cap.max_inline_data = qp->max_inline_data;
503 
504  qp_init_attr->cap = qp_attr->cap;
505  qp_init_attr->sq_sig_type = qp->sq_policy;
506 
507 out_mailbox:
508  mthca_free_mailbox(dev, mailbox);
509 
510 out:
511  mutex_unlock(&qp->mutex);
512  return err;
513 }
514 
515 static int mthca_path_set(struct mthca_dev *dev, const struct ib_ah_attr *ah,
516  struct mthca_qp_path *path, u8 port)
517 {
518  path->g_mylmc = ah->src_path_bits & 0x7f;
519  path->rlid = cpu_to_be16(ah->dlid);
520  path->static_rate = mthca_get_rate(dev, ah->static_rate, port);
521 
522  if (ah->ah_flags & IB_AH_GRH) {
523  if (ah->grh.sgid_index >= dev->limits.gid_table_len) {
524  mthca_dbg(dev, "sgid_index (%u) too large. max is %d\n",
525  ah->grh.sgid_index, dev->limits.gid_table_len-1);
526  return -1;
527  }
528 
529  path->g_mylmc |= 1 << 7;
530  path->mgid_index = ah->grh.sgid_index;
531  path->hop_limit = ah->grh.hop_limit;
532  path->sl_tclass_flowlabel =
533  cpu_to_be32((ah->sl << 28) |
534  (ah->grh.traffic_class << 20) |
535  (ah->grh.flow_label));
536  memcpy(path->rgid, ah->grh.dgid.raw, 16);
537  } else
538  path->sl_tclass_flowlabel = cpu_to_be32(ah->sl << 28);
539 
540  return 0;
541 }
542 
543 static int __mthca_modify_qp(struct ib_qp *ibqp,
544  const struct ib_qp_attr *attr, int attr_mask,
545  enum ib_qp_state cur_state, enum ib_qp_state new_state)
546 {
547  struct mthca_dev *dev = to_mdev(ibqp->device);
548  struct mthca_qp *qp = to_mqp(ibqp);
549  struct mthca_mailbox *mailbox;
550  struct mthca_qp_param *qp_param;
551  struct mthca_qp_context *qp_context;
552  u32 sqd_event = 0;
553  int err = -EINVAL;
554 
555  mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
556  if (IS_ERR(mailbox)) {
557  err = PTR_ERR(mailbox);
558  goto out;
559  }
560  qp_param = mailbox->buf;
561  qp_context = &qp_param->context;
562  memset(qp_param, 0, sizeof *qp_param);
563 
564  qp_context->flags = cpu_to_be32((to_mthca_state(new_state) << 28) |
565  (to_mthca_st(qp->transport) << 16));
566  qp_context->flags |= cpu_to_be32(MTHCA_QP_BIT_DE);
567  if (!(attr_mask & IB_QP_PATH_MIG_STATE))
568  qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
569  else {
571  switch (attr->path_mig_state) {
572  case IB_MIG_MIGRATED:
573  qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
574  break;
575  case IB_MIG_REARM:
576  qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
577  break;
578  case IB_MIG_ARMED:
579  qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
580  break;
581  }
582  }
583 
584  /* leave tavor_sched_queue as 0 */
585 
586  if (qp->transport == MLX || qp->transport == UD)
587  qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
588  else if (attr_mask & IB_QP_PATH_MTU) {
589  if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_2048) {
590  mthca_dbg(dev, "path MTU (%u) is invalid\n",
591  attr->path_mtu);
592  goto out_mailbox;
593  }
594  qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
595  }
596 
597  if (mthca_is_memfree(dev)) {
598  if (qp->rq.max)
599  qp_context->rq_size_stride = ilog2(qp->rq.max) << 3;
600  qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
601 
602  if (qp->sq.max)
603  qp_context->sq_size_stride = ilog2(qp->sq.max) << 3;
604  qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
605  }
606 
607  /* leave arbel_sched_queue as 0 */
608 
609  if (qp->ibqp.uobject)
610  qp_context->usr_page =
611  cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index);
612  else
613  qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
614  qp_context->local_qpn = cpu_to_be32(qp->qpn);
615  if (attr_mask & IB_QP_DEST_QPN) {
616  qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
617  }
618 
619  if (qp->transport == MLX)
620  qp_context->pri_path.port_pkey |=
621  cpu_to_be32(qp->port << 24);
622  else {
623  if (attr_mask & IB_QP_PORT) {
624  qp_context->pri_path.port_pkey |=
625  cpu_to_be32(attr->port_num << 24);
627  }
628  }
629 
630  if (attr_mask & IB_QP_PKEY_INDEX) {
631  qp_context->pri_path.port_pkey |=
632  cpu_to_be32(attr->pkey_index);
634  }
635 
636  if (attr_mask & IB_QP_RNR_RETRY) {
637  qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry =
638  attr->rnr_retry << 5;
641  }
642 
643  if (attr_mask & IB_QP_AV) {
644  if (mthca_path_set(dev, &attr->ah_attr, &qp_context->pri_path,
645  attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
646  goto out_mailbox;
647 
649  }
650 
651  if (ibqp->qp_type == IB_QPT_RC &&
652  cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
653  u8 sched_queue = ibqp->uobject ? 0x2 : 0x1;
654 
655  if (mthca_is_memfree(dev))
656  qp_context->rlkey_arbel_sched_queue |= sched_queue;
657  else
658  qp_context->tavor_sched_queue |= cpu_to_be32(sched_queue);
659 
660  qp_param->opt_param_mask |=
662  }
663 
664  if (attr_mask & IB_QP_TIMEOUT) {
665  qp_context->pri_path.ackto = attr->timeout << 3;
667  }
668 
669  if (attr_mask & IB_QP_ALT_PATH) {
670  if (attr->alt_pkey_index >= dev->limits.pkey_table_len) {
671  mthca_dbg(dev, "Alternate P_Key index (%u) too large. max is %d\n",
672  attr->alt_pkey_index, dev->limits.pkey_table_len-1);
673  goto out_mailbox;
674  }
675 
676  if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) {
677  mthca_dbg(dev, "Alternate port number (%u) is invalid\n",
678  attr->alt_port_num);
679  goto out_mailbox;
680  }
681 
682  if (mthca_path_set(dev, &attr->alt_ah_attr, &qp_context->alt_path,
683  attr->alt_ah_attr.port_num))
684  goto out_mailbox;
685 
686  qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index |
687  attr->alt_port_num << 24);
688  qp_context->alt_path.ackto = attr->alt_timeout << 3;
690  }
691 
692  /* leave rdd as 0 */
693  qp_context->pd = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
694  /* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
695  qp_context->wqe_lkey = cpu_to_be32(qp->mr.ibmr.lkey);
696  qp_context->params1 = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
697  (MTHCA_FLIGHT_LIMIT << 24) |
699  if (qp->sq_policy == IB_SIGNAL_ALL_WR)
700  qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
701  if (attr_mask & IB_QP_RETRY_CNT) {
702  qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
704  }
705 
706  if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
707  if (attr->max_rd_atomic) {
708  qp_context->params1 |=
711  qp_context->params1 |=
712  cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
713  }
715  }
716 
717  if (attr_mask & IB_QP_SQ_PSN)
718  qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
719  qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
720 
721  if (mthca_is_memfree(dev)) {
722  qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
723  qp_context->snd_db_index = cpu_to_be32(qp->sq.db_index);
724  }
725 
726  if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
727  if (attr->max_dest_rd_atomic)
728  qp_context->params2 |=
729  cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
730 
732  }
733 
734  if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
735  qp_context->params2 |= get_hw_access_flags(qp, attr, attr_mask);
739  }
740 
741  qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
742 
743  if (ibqp->srq)
744  qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
745 
746  if (attr_mask & IB_QP_MIN_RNR_TIMER) {
747  qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
749  }
750  if (attr_mask & IB_QP_RQ_PSN)
751  qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
752 
753  qp_context->ra_buff_indx =
754  cpu_to_be32(dev->qp_table.rdb_base +
755  ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
756  dev->qp_table.rdb_shift));
757 
758  qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
759 
760  if (mthca_is_memfree(dev))
761  qp_context->rcv_db_index = cpu_to_be32(qp->rq.db_index);
762 
763  if (attr_mask & IB_QP_QKEY) {
764  qp_context->qkey = cpu_to_be32(attr->qkey);
766  }
767 
768  if (ibqp->srq)
769  qp_context->srqn = cpu_to_be32(1 << 24 |
770  to_msrq(ibqp->srq)->srqn);
771 
772  if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD &&
773  attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY &&
774  attr->en_sqd_async_notify)
775  sqd_event = 1 << 31;
776 
777  err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0,
778  mailbox, sqd_event);
779  if (err) {
780  mthca_warn(dev, "modify QP %d->%d returned %d.\n",
781  cur_state, new_state, err);
782  goto out_mailbox;
783  }
784 
785  qp->state = new_state;
786  if (attr_mask & IB_QP_ACCESS_FLAGS)
787  qp->atomic_rd_en = attr->qp_access_flags;
788  if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
789  qp->resp_depth = attr->max_dest_rd_atomic;
790  if (attr_mask & IB_QP_PORT)
791  qp->port = attr->port_num;
792  if (attr_mask & IB_QP_ALT_PATH)
793  qp->alt_port = attr->alt_port_num;
794 
795  if (is_sqp(dev, qp))
796  store_attrs(to_msqp(qp), attr, attr_mask);
797 
798  /*
799  * If we moved QP0 to RTR, bring the IB link up; if we moved
800  * QP0 to RESET or ERROR, bring the link back down.
801  */
802  if (is_qp0(dev, qp)) {
803  if (cur_state != IB_QPS_RTR &&
804  new_state == IB_QPS_RTR)
805  init_port(dev, qp->port);
806 
807  if (cur_state != IB_QPS_RESET &&
808  cur_state != IB_QPS_ERR &&
809  (new_state == IB_QPS_RESET ||
810  new_state == IB_QPS_ERR))
811  mthca_CLOSE_IB(dev, qp->port);
812  }
813 
814  /*
815  * If we moved a kernel QP to RESET, clean up all old CQ
816  * entries and reinitialize the QP.
817  */
818  if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
819  mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn,
820  qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
821  if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
822  mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn, NULL);
823 
824  mthca_wq_reset(&qp->sq);
825  qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
826 
827  mthca_wq_reset(&qp->rq);
828  qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
829 
830  if (mthca_is_memfree(dev)) {
831  *qp->sq.db = 0;
832  *qp->rq.db = 0;
833  }
834  }
835 
836 out_mailbox:
837  mthca_free_mailbox(dev, mailbox);
838 out:
839  return err;
840 }
841 
842 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
843  struct ib_udata *udata)
844 {
845  struct mthca_dev *dev = to_mdev(ibqp->device);
846  struct mthca_qp *qp = to_mqp(ibqp);
847  enum ib_qp_state cur_state, new_state;
848  int err = -EINVAL;
849 
850  mutex_lock(&qp->mutex);
851  if (attr_mask & IB_QP_CUR_STATE) {
852  cur_state = attr->cur_qp_state;
853  } else {
854  spin_lock_irq(&qp->sq.lock);
855  spin_lock(&qp->rq.lock);
856  cur_state = qp->state;
857  spin_unlock(&qp->rq.lock);
858  spin_unlock_irq(&qp->sq.lock);
859  }
860 
861  new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
862 
863  if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask)) {
864  mthca_dbg(dev, "Bad QP transition (transport %d) "
865  "%d->%d with attr 0x%08x\n",
866  qp->transport, cur_state, new_state,
867  attr_mask);
868  goto out;
869  }
870 
871  if ((attr_mask & IB_QP_PKEY_INDEX) &&
872  attr->pkey_index >= dev->limits.pkey_table_len) {
873  mthca_dbg(dev, "P_Key index (%u) too large. max is %d\n",
874  attr->pkey_index, dev->limits.pkey_table_len-1);
875  goto out;
876  }
877 
878  if ((attr_mask & IB_QP_PORT) &&
879  (attr->port_num == 0 || attr->port_num > dev->limits.num_ports)) {
880  mthca_dbg(dev, "Port number (%u) is invalid\n", attr->port_num);
881  goto out;
882  }
883 
884  if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
885  attr->max_rd_atomic > dev->limits.max_qp_init_rdma) {
886  mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n",
887  attr->max_rd_atomic, dev->limits.max_qp_init_rdma);
888  goto out;
889  }
890 
891  if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
892  attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) {
893  mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n",
894  attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift);
895  goto out;
896  }
897 
898  if (cur_state == new_state && cur_state == IB_QPS_RESET) {
899  err = 0;
900  goto out;
901  }
902 
903  err = __mthca_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
904 
905 out:
906  mutex_unlock(&qp->mutex);
907  return err;
908 }
909 
910 static int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz)
911 {
912  /*
913  * Calculate the maximum size of WQE s/g segments, excluding
914  * the next segment and other non-data segments.
915  */
916  int max_data_size = desc_sz - sizeof (struct mthca_next_seg);
917 
918  switch (qp->transport) {
919  case MLX:
920  max_data_size -= 2 * sizeof (struct mthca_data_seg);
921  break;
922 
923  case UD:
924  if (mthca_is_memfree(dev))
925  max_data_size -= sizeof (struct mthca_arbel_ud_seg);
926  else
927  max_data_size -= sizeof (struct mthca_tavor_ud_seg);
928  break;
929 
930  default:
931  max_data_size -= sizeof (struct mthca_raddr_seg);
932  break;
933  }
934 
935  return max_data_size;
936 }
937 
938 static inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size)
939 {
940  /* We don't support inline data for kernel QPs (yet). */
941  return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0;
942 }
943 
944 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
945  struct mthca_pd *pd,
946  struct mthca_qp *qp)
947 {
948  int max_data_size = mthca_max_data_size(dev, qp,
949  min(dev->limits.max_desc_sz,
950  1 << qp->sq.wqe_shift));
951 
952  qp->max_inline_data = mthca_max_inline_data(pd, max_data_size);
953 
954  qp->sq.max_gs = min_t(int, dev->limits.max_sg,
955  max_data_size / sizeof (struct mthca_data_seg));
956  qp->rq.max_gs = min_t(int, dev->limits.max_sg,
957  (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
958  sizeof (struct mthca_next_seg)) /
959  sizeof (struct mthca_data_seg));
960 }
961 
962 /*
963  * Allocate and register buffer for WQEs. qp->rq.max, sq.max,
964  * rq.max_gs and sq.max_gs must all be assigned.
965  * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
966  * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
967  * queue)
968  */
969 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
970  struct mthca_pd *pd,
971  struct mthca_qp *qp)
972 {
973  int size;
974  int err = -ENOMEM;
975 
976  size = sizeof (struct mthca_next_seg) +
977  qp->rq.max_gs * sizeof (struct mthca_data_seg);
978 
979  if (size > dev->limits.max_desc_sz)
980  return -EINVAL;
981 
982  for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
983  qp->rq.wqe_shift++)
984  ; /* nothing */
985 
986  size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
987  switch (qp->transport) {
988  case MLX:
989  size += 2 * sizeof (struct mthca_data_seg);
990  break;
991 
992  case UD:
993  size += mthca_is_memfree(dev) ?
994  sizeof (struct mthca_arbel_ud_seg) :
995  sizeof (struct mthca_tavor_ud_seg);
996  break;
997 
998  case UC:
999  size += sizeof (struct mthca_raddr_seg);
1000  break;
1001 
1002  case RC:
1003  size += sizeof (struct mthca_raddr_seg);
1004  /*
1005  * An atomic op will require an atomic segment, a
1006  * remote address segment and one scatter entry.
1007  */
1008  size = max_t(int, size,
1009  sizeof (struct mthca_atomic_seg) +
1010  sizeof (struct mthca_raddr_seg) +
1011  sizeof (struct mthca_data_seg));
1012  break;
1013 
1014  default:
1015  break;
1016  }
1017 
1018  /* Make sure that we have enough space for a bind request */
1019  size = max_t(int, size, sizeof (struct mthca_bind_seg));
1020 
1021  size += sizeof (struct mthca_next_seg);
1022 
1023  if (size > dev->limits.max_desc_sz)
1024  return -EINVAL;
1025 
1026  for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
1027  qp->sq.wqe_shift++)
1028  ; /* nothing */
1029 
1030  qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
1031  1 << qp->sq.wqe_shift);
1032 
1033  /*
1034  * If this is a userspace QP, we don't actually have to
1035  * allocate anything. All we need is to calculate the WQE
1036  * sizes and the send_wqe_offset, so we're done now.
1037  */
1038  if (pd->ibpd.uobject)
1039  return 0;
1040 
1041  size = PAGE_ALIGN(qp->send_wqe_offset +
1042  (qp->sq.max << qp->sq.wqe_shift));
1043 
1044  qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
1045  GFP_KERNEL);
1046  if (!qp->wrid)
1047  goto err_out;
1048 
1049  err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
1050  &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1051  if (err)
1052  goto err_out;
1053 
1054  return 0;
1055 
1056 err_out:
1057  kfree(qp->wrid);
1058  return err;
1059 }
1060 
1061 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1062  struct mthca_qp *qp)
1063 {
1065  (qp->sq.max << qp->sq.wqe_shift)),
1066  &qp->queue, qp->is_direct, &qp->mr);
1067  kfree(qp->wrid);
1068 }
1069 
1070 static int mthca_map_memfree(struct mthca_dev *dev,
1071  struct mthca_qp *qp)
1072 {
1073  int ret;
1074 
1075  if (mthca_is_memfree(dev)) {
1076  ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1077  if (ret)
1078  return ret;
1079 
1080  ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1081  if (ret)
1082  goto err_qpc;
1083 
1084  ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1085  qp->qpn << dev->qp_table.rdb_shift);
1086  if (ret)
1087  goto err_eqpc;
1088 
1089  }
1090 
1091  return 0;
1092 
1093 err_eqpc:
1094  mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1095 
1096 err_qpc:
1097  mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1098 
1099  return ret;
1100 }
1101 
1102 static void mthca_unmap_memfree(struct mthca_dev *dev,
1103  struct mthca_qp *qp)
1104 {
1105  mthca_table_put(dev, dev->qp_table.rdb_table,
1106  qp->qpn << dev->qp_table.rdb_shift);
1107  mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1108  mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1109 }
1110 
1111 static int mthca_alloc_memfree(struct mthca_dev *dev,
1112  struct mthca_qp *qp)
1113 {
1114  if (mthca_is_memfree(dev)) {
1115  qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1116  qp->qpn, &qp->rq.db);
1117  if (qp->rq.db_index < 0)
1118  return -ENOMEM;
1119 
1120  qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1121  qp->qpn, &qp->sq.db);
1122  if (qp->sq.db_index < 0) {
1123  mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1124  return -ENOMEM;
1125  }
1126  }
1127 
1128  return 0;
1129 }
1130 
1131 static void mthca_free_memfree(struct mthca_dev *dev,
1132  struct mthca_qp *qp)
1133 {
1134  if (mthca_is_memfree(dev)) {
1135  mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1136  mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1137  }
1138 }
1139 
1140 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1141  struct mthca_pd *pd,
1142  struct mthca_cq *send_cq,
1143  struct mthca_cq *recv_cq,
1144  enum ib_sig_type send_policy,
1145  struct mthca_qp *qp)
1146 {
1147  int ret;
1148  int i;
1149  struct mthca_next_seg *next;
1150 
1151  qp->refcount = 1;
1152  init_waitqueue_head(&qp->wait);
1153  mutex_init(&qp->mutex);
1154  qp->state = IB_QPS_RESET;
1155  qp->atomic_rd_en = 0;
1156  qp->resp_depth = 0;
1157  qp->sq_policy = send_policy;
1158  mthca_wq_reset(&qp->sq);
1159  mthca_wq_reset(&qp->rq);
1160 
1161  spin_lock_init(&qp->sq.lock);
1162  spin_lock_init(&qp->rq.lock);
1163 
1164  ret = mthca_map_memfree(dev, qp);
1165  if (ret)
1166  return ret;
1167 
1168  ret = mthca_alloc_wqe_buf(dev, pd, qp);
1169  if (ret) {
1170  mthca_unmap_memfree(dev, qp);
1171  return ret;
1172  }
1173 
1174  mthca_adjust_qp_caps(dev, pd, qp);
1175 
1176  /*
1177  * If this is a userspace QP, we're done now. The doorbells
1178  * will be allocated and buffers will be initialized in
1179  * userspace.
1180  */
1181  if (pd->ibpd.uobject)
1182  return 0;
1183 
1184  ret = mthca_alloc_memfree(dev, qp);
1185  if (ret) {
1186  mthca_free_wqe_buf(dev, qp);
1187  mthca_unmap_memfree(dev, qp);
1188  return ret;
1189  }
1190 
1191  if (mthca_is_memfree(dev)) {
1192  struct mthca_data_seg *scatter;
1193  int size = (sizeof (struct mthca_next_seg) +
1194  qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1195 
1196  for (i = 0; i < qp->rq.max; ++i) {
1197  next = get_recv_wqe(qp, i);
1198  next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1199  qp->rq.wqe_shift);
1200  next->ee_nds = cpu_to_be32(size);
1201 
1202  for (scatter = (void *) (next + 1);
1203  (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1204  ++scatter)
1205  scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1206  }
1207 
1208  for (i = 0; i < qp->sq.max; ++i) {
1209  next = get_send_wqe(qp, i);
1210  next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1211  qp->sq.wqe_shift) +
1212  qp->send_wqe_offset);
1213  }
1214  } else {
1215  for (i = 0; i < qp->rq.max; ++i) {
1216  next = get_recv_wqe(qp, i);
1217  next->nda_op = htonl((((i + 1) % qp->rq.max) <<
1218  qp->rq.wqe_shift) | 1);
1219  }
1220 
1221  }
1222 
1223  qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1224  qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1225 
1226  return 0;
1227 }
1228 
1229 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1230  struct mthca_pd *pd, struct mthca_qp *qp)
1231 {
1232  int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz);
1233 
1234  /* Sanity check QP size before proceeding */
1235  if (cap->max_send_wr > dev->limits.max_wqes ||
1236  cap->max_recv_wr > dev->limits.max_wqes ||
1237  cap->max_send_sge > dev->limits.max_sg ||
1238  cap->max_recv_sge > dev->limits.max_sg ||
1239  cap->max_inline_data > mthca_max_inline_data(pd, max_data_size))
1240  return -EINVAL;
1241 
1242  /*
1243  * For MLX transport we need 2 extra send gather entries:
1244  * one for the header and one for the checksum at the end
1245  */
1246  if (qp->transport == MLX && cap->max_send_sge + 2 > dev->limits.max_sg)
1247  return -EINVAL;
1248 
1249  if (mthca_is_memfree(dev)) {
1250  qp->rq.max = cap->max_recv_wr ?
1251  roundup_pow_of_two(cap->max_recv_wr) : 0;
1252  qp->sq.max = cap->max_send_wr ?
1253  roundup_pow_of_two(cap->max_send_wr) : 0;
1254  } else {
1255  qp->rq.max = cap->max_recv_wr;
1256  qp->sq.max = cap->max_send_wr;
1257  }
1258 
1259  qp->rq.max_gs = cap->max_recv_sge;
1260  qp->sq.max_gs = max_t(int, cap->max_send_sge,
1263  sizeof (struct mthca_data_seg));
1264 
1265  return 0;
1266 }
1267 
1268 int mthca_alloc_qp(struct mthca_dev *dev,
1269  struct mthca_pd *pd,
1270  struct mthca_cq *send_cq,
1271  struct mthca_cq *recv_cq,
1272  enum ib_qp_type type,
1273  enum ib_sig_type send_policy,
1274  struct ib_qp_cap *cap,
1275  struct mthca_qp *qp)
1276 {
1277  int err;
1278 
1279  switch (type) {
1280  case IB_QPT_RC: qp->transport = RC; break;
1281  case IB_QPT_UC: qp->transport = UC; break;
1282  case IB_QPT_UD: qp->transport = UD; break;
1283  default: return -EINVAL;
1284  }
1285 
1286  err = mthca_set_qp_size(dev, cap, pd, qp);
1287  if (err)
1288  return err;
1289 
1290  qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1291  if (qp->qpn == -1)
1292  return -ENOMEM;
1293 
1294  /* initialize port to zero for error-catching. */
1295  qp->port = 0;
1296 
1297  err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1298  send_policy, qp);
1299  if (err) {
1300  mthca_free(&dev->qp_table.alloc, qp->qpn);
1301  return err;
1302  }
1303 
1304  spin_lock_irq(&dev->qp_table.lock);
1305  mthca_array_set(&dev->qp_table.qp,
1306  qp->qpn & (dev->limits.num_qps - 1), qp);
1307  spin_unlock_irq(&dev->qp_table.lock);
1308 
1309  return 0;
1310 }
1311 
1312 static void mthca_lock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1313  __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1314 {
1315  if (send_cq == recv_cq) {
1316  spin_lock_irq(&send_cq->lock);
1317  __acquire(&recv_cq->lock);
1318  } else if (send_cq->cqn < recv_cq->cqn) {
1319  spin_lock_irq(&send_cq->lock);
1321  } else {
1322  spin_lock_irq(&recv_cq->lock);
1324  }
1325 }
1326 
1327 static void mthca_unlock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1328  __releases(&send_cq->lock) __releases(&recv_cq->lock)
1329 {
1330  if (send_cq == recv_cq) {
1331  __release(&recv_cq->lock);
1332  spin_unlock_irq(&send_cq->lock);
1333  } else if (send_cq->cqn < recv_cq->cqn) {
1334  spin_unlock(&recv_cq->lock);
1335  spin_unlock_irq(&send_cq->lock);
1336  } else {
1337  spin_unlock(&send_cq->lock);
1338  spin_unlock_irq(&recv_cq->lock);
1339  }
1340 }
1341 
1342 int mthca_alloc_sqp(struct mthca_dev *dev,
1343  struct mthca_pd *pd,
1344  struct mthca_cq *send_cq,
1345  struct mthca_cq *recv_cq,
1346  enum ib_sig_type send_policy,
1347  struct ib_qp_cap *cap,
1348  int qpn,
1349  int port,
1350  struct mthca_sqp *sqp)
1351 {
1352  u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1353  int err;
1354 
1355  sqp->qp.transport = MLX;
1356  err = mthca_set_qp_size(dev, cap, pd, &sqp->qp);
1357  if (err)
1358  return err;
1359 
1360  sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1361  sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1362  &sqp->header_dma, GFP_KERNEL);
1363  if (!sqp->header_buf)
1364  return -ENOMEM;
1365 
1366  spin_lock_irq(&dev->qp_table.lock);
1367  if (mthca_array_get(&dev->qp_table.qp, mqpn))
1368  err = -EBUSY;
1369  else
1370  mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1371  spin_unlock_irq(&dev->qp_table.lock);
1372 
1373  if (err)
1374  goto err_out;
1375 
1376  sqp->qp.port = port;
1377  sqp->qp.qpn = mqpn;
1378  sqp->qp.transport = MLX;
1379 
1380  err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1381  send_policy, &sqp->qp);
1382  if (err)
1383  goto err_out_free;
1384 
1385  atomic_inc(&pd->sqp_count);
1386 
1387  return 0;
1388 
1389  err_out_free:
1390  /*
1391  * Lock CQs here, so that CQ polling code can do QP lookup
1392  * without taking a lock.
1393  */
1394  mthca_lock_cqs(send_cq, recv_cq);
1395 
1396  spin_lock(&dev->qp_table.lock);
1397  mthca_array_clear(&dev->qp_table.qp, mqpn);
1398  spin_unlock(&dev->qp_table.lock);
1399 
1400  mthca_unlock_cqs(send_cq, recv_cq);
1401 
1402  err_out:
1403  dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1404  sqp->header_buf, sqp->header_dma);
1405 
1406  return err;
1407 }
1408 
1409 static inline int get_qp_refcount(struct mthca_dev *dev, struct mthca_qp *qp)
1410 {
1411  int c;
1412 
1413  spin_lock_irq(&dev->qp_table.lock);
1414  c = qp->refcount;
1415  spin_unlock_irq(&dev->qp_table.lock);
1416 
1417  return c;
1418 }
1419 
1420 void mthca_free_qp(struct mthca_dev *dev,
1421  struct mthca_qp *qp)
1422 {
1423  struct mthca_cq *send_cq;
1424  struct mthca_cq *recv_cq;
1425 
1426  send_cq = to_mcq(qp->ibqp.send_cq);
1427  recv_cq = to_mcq(qp->ibqp.recv_cq);
1428 
1429  /*
1430  * Lock CQs here, so that CQ polling code can do QP lookup
1431  * without taking a lock.
1432  */
1433  mthca_lock_cqs(send_cq, recv_cq);
1434 
1435  spin_lock(&dev->qp_table.lock);
1436  mthca_array_clear(&dev->qp_table.qp,
1437  qp->qpn & (dev->limits.num_qps - 1));
1438  --qp->refcount;
1439  spin_unlock(&dev->qp_table.lock);
1440 
1441  mthca_unlock_cqs(send_cq, recv_cq);
1442 
1443  wait_event(qp->wait, !get_qp_refcount(dev, qp));
1444 
1445  if (qp->state != IB_QPS_RESET)
1446  mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0,
1447  NULL, 0);
1448 
1449  /*
1450  * If this is a userspace QP, the buffers, MR, CQs and so on
1451  * will be cleaned up in userspace, so all we have to do is
1452  * unref the mem-free tables and free the QPN in our table.
1453  */
1454  if (!qp->ibqp.uobject) {
1455  mthca_cq_clean(dev, recv_cq, qp->qpn,
1456  qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1457  if (send_cq != recv_cq)
1458  mthca_cq_clean(dev, send_cq, qp->qpn, NULL);
1459 
1460  mthca_free_memfree(dev, qp);
1461  mthca_free_wqe_buf(dev, qp);
1462  }
1463 
1464  mthca_unmap_memfree(dev, qp);
1465 
1466  if (is_sqp(dev, qp)) {
1467  atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1468  dma_free_coherent(&dev->pdev->dev,
1469  to_msqp(qp)->header_buf_size,
1470  to_msqp(qp)->header_buf,
1471  to_msqp(qp)->header_dma);
1472  } else
1473  mthca_free(&dev->qp_table.alloc, qp->qpn);
1474 }
1475 
1476 /* Create UD header for an MLX send and build a data segment for it */
1477 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1478  int ind, struct ib_send_wr *wr,
1479  struct mthca_mlx_seg *mlx,
1480  struct mthca_data_seg *data)
1481 {
1482  int header_size;
1483  int err;
1484  u16 pkey;
1485 
1486  ib_ud_header_init(256, /* assume a MAD */ 1, 0, 0,
1487  mthca_ah_grh_present(to_mah(wr->wr.ud.ah)), 0,
1488  &sqp->ud_header);
1489 
1490  err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1491  if (err)
1492  return err;
1493  mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1494  mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1495  (sqp->ud_header.lrh.destination_lid ==
1497  (sqp->ud_header.lrh.service_level << 8));
1498  mlx->rlid = sqp->ud_header.lrh.destination_lid;
1499  mlx->vcrc = 0;
1500 
1501  switch (wr->opcode) {
1502  case IB_WR_SEND:
1503  sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1504  sqp->ud_header.immediate_present = 0;
1505  break;
1506  case IB_WR_SEND_WITH_IMM:
1507  sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1508  sqp->ud_header.immediate_present = 1;
1509  sqp->ud_header.immediate_data = wr->ex.imm_data;
1510  break;
1511  default:
1512  return -EINVAL;
1513  }
1514 
1515  sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 : 0;
1516  if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1517  sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1518  sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1519  if (!sqp->qp.ibqp.qp_num)
1520  ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1521  sqp->pkey_index, &pkey);
1522  else
1523  ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1524  wr->wr.ud.pkey_index, &pkey);
1525  sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1526  sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1527  sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1528  sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1529  sqp->qkey : wr->wr.ud.remote_qkey);
1530  sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1531 
1532  header_size = ib_ud_header_pack(&sqp->ud_header,
1533  sqp->header_buf +
1534  ind * MTHCA_UD_HEADER_SIZE);
1535 
1536  data->byte_count = cpu_to_be32(header_size);
1537  data->lkey = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1538  data->addr = cpu_to_be64(sqp->header_dma +
1539  ind * MTHCA_UD_HEADER_SIZE);
1540 
1541  return 0;
1542 }
1543 
1544 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1545  struct ib_cq *ib_cq)
1546 {
1547  unsigned cur;
1548  struct mthca_cq *cq;
1549 
1550  cur = wq->head - wq->tail;
1551  if (likely(cur + nreq < wq->max))
1552  return 0;
1553 
1554  cq = to_mcq(ib_cq);
1555  spin_lock(&cq->lock);
1556  cur = wq->head - wq->tail;
1557  spin_unlock(&cq->lock);
1558 
1559  return cur + nreq >= wq->max;
1560 }
1561 
1562 static __always_inline void set_raddr_seg(struct mthca_raddr_seg *rseg,
1564 {
1565  rseg->raddr = cpu_to_be64(remote_addr);
1566  rseg->rkey = cpu_to_be32(rkey);
1567  rseg->reserved = 0;
1568 }
1569 
1570 static __always_inline void set_atomic_seg(struct mthca_atomic_seg *aseg,
1571  struct ib_send_wr *wr)
1572 {
1573  if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1574  aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
1575  aseg->compare = cpu_to_be64(wr->wr.atomic.compare_add);
1576  } else {
1577  aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
1578  aseg->compare = 0;
1579  }
1580 
1581 }
1582 
1583 static void set_tavor_ud_seg(struct mthca_tavor_ud_seg *useg,
1584  struct ib_send_wr *wr)
1585 {
1586  useg->lkey = cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1587  useg->av_addr = cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1588  useg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1589  useg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1590 
1591 }
1592 
1593 static void set_arbel_ud_seg(struct mthca_arbel_ud_seg *useg,
1594  struct ib_send_wr *wr)
1595 {
1596  memcpy(useg->av, to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
1597  useg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1598  useg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1599 }
1600 
1601 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1602  struct ib_send_wr **bad_wr)
1603 {
1604  struct mthca_dev *dev = to_mdev(ibqp->device);
1605  struct mthca_qp *qp = to_mqp(ibqp);
1606  void *wqe;
1607  void *prev_wqe;
1608  unsigned long flags;
1609  int err = 0;
1610  int nreq;
1611  int i;
1612  int size;
1613  /*
1614  * f0 and size0 are only used if nreq != 0, and they will
1615  * always be initialized the first time through the main loop
1616  * before nreq is incremented. So nreq cannot become non-zero
1617  * without initializing f0 and size0, and they are in fact
1618  * never used uninitialized.
1619  */
1620  int uninitialized_var(size0);
1622  int ind;
1623  u8 op0 = 0;
1624 
1625  spin_lock_irqsave(&qp->sq.lock, flags);
1626 
1627  /* XXX check that state is OK to post send */
1628 
1629  ind = qp->sq.next_ind;
1630 
1631  for (nreq = 0; wr; ++nreq, wr = wr->next) {
1632  if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1633  mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1634  " %d max, %d nreq)\n", qp->qpn,
1635  qp->sq.head, qp->sq.tail,
1636  qp->sq.max, nreq);
1637  err = -ENOMEM;
1638  *bad_wr = wr;
1639  goto out;
1640  }
1641 
1642  wqe = get_send_wqe(qp, ind);
1643  prev_wqe = qp->sq.last;
1644  qp->sq.last = wqe;
1645 
1646  ((struct mthca_next_seg *) wqe)->nda_op = 0;
1647  ((struct mthca_next_seg *) wqe)->ee_nds = 0;
1648  ((struct mthca_next_seg *) wqe)->flags =
1649  ((wr->send_flags & IB_SEND_SIGNALED) ?
1651  ((wr->send_flags & IB_SEND_SOLICITED) ?
1653  cpu_to_be32(1);
1654  if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1655  wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1656  ((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
1657 
1658  wqe += sizeof (struct mthca_next_seg);
1659  size = sizeof (struct mthca_next_seg) / 16;
1660 
1661  switch (qp->transport) {
1662  case RC:
1663  switch (wr->opcode) {
1666  set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
1667  wr->wr.atomic.rkey);
1668  wqe += sizeof (struct mthca_raddr_seg);
1669 
1670  set_atomic_seg(wqe, wr);
1671  wqe += sizeof (struct mthca_atomic_seg);
1672  size += (sizeof (struct mthca_raddr_seg) +
1673  sizeof (struct mthca_atomic_seg)) / 16;
1674  break;
1675 
1676  case IB_WR_RDMA_WRITE:
1678  case IB_WR_RDMA_READ:
1679  set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
1680  wr->wr.rdma.rkey);
1681  wqe += sizeof (struct mthca_raddr_seg);
1682  size += sizeof (struct mthca_raddr_seg) / 16;
1683  break;
1684 
1685  default:
1686  /* No extra segments required for sends */
1687  break;
1688  }
1689 
1690  break;
1691 
1692  case UC:
1693  switch (wr->opcode) {
1694  case IB_WR_RDMA_WRITE:
1696  set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
1697  wr->wr.rdma.rkey);
1698  wqe += sizeof (struct mthca_raddr_seg);
1699  size += sizeof (struct mthca_raddr_seg) / 16;
1700  break;
1701 
1702  default:
1703  /* No extra segments required for sends */
1704  break;
1705  }
1706 
1707  break;
1708 
1709  case UD:
1710  set_tavor_ud_seg(wqe, wr);
1711  wqe += sizeof (struct mthca_tavor_ud_seg);
1712  size += sizeof (struct mthca_tavor_ud_seg) / 16;
1713  break;
1714 
1715  case MLX:
1716  err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1717  wqe - sizeof (struct mthca_next_seg),
1718  wqe);
1719  if (err) {
1720  *bad_wr = wr;
1721  goto out;
1722  }
1723  wqe += sizeof (struct mthca_data_seg);
1724  size += sizeof (struct mthca_data_seg) / 16;
1725  break;
1726  }
1727 
1728  if (wr->num_sge > qp->sq.max_gs) {
1729  mthca_err(dev, "too many gathers\n");
1730  err = -EINVAL;
1731  *bad_wr = wr;
1732  goto out;
1733  }
1734 
1735  for (i = 0; i < wr->num_sge; ++i) {
1736  mthca_set_data_seg(wqe, wr->sg_list + i);
1737  wqe += sizeof (struct mthca_data_seg);
1738  size += sizeof (struct mthca_data_seg) / 16;
1739  }
1740 
1741  /* Add one more inline data segment for ICRC */
1742  if (qp->transport == MLX) {
1743  ((struct mthca_data_seg *) wqe)->byte_count =
1744  cpu_to_be32((1 << 31) | 4);
1745  ((u32 *) wqe)[1] = 0;
1746  wqe += sizeof (struct mthca_data_seg);
1747  size += sizeof (struct mthca_data_seg) / 16;
1748  }
1749 
1750  qp->wrid[ind + qp->rq.max] = wr->wr_id;
1751 
1752  if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1753  mthca_err(dev, "opcode invalid\n");
1754  err = -EINVAL;
1755  *bad_wr = wr;
1756  goto out;
1757  }
1758 
1759  ((struct mthca_next_seg *) prev_wqe)->nda_op =
1760  cpu_to_be32(((ind << qp->sq.wqe_shift) +
1761  qp->send_wqe_offset) |
1762  mthca_opcode[wr->opcode]);
1763  wmb();
1764  ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1765  cpu_to_be32((nreq ? 0 : MTHCA_NEXT_DBD) | size |
1766  ((wr->send_flags & IB_SEND_FENCE) ?
1767  MTHCA_NEXT_FENCE : 0));
1768 
1769  if (!nreq) {
1770  size0 = size;
1771  op0 = mthca_opcode[wr->opcode];
1772  f0 = wr->send_flags & IB_SEND_FENCE ?
1774  }
1775 
1776  ++ind;
1777  if (unlikely(ind >= qp->sq.max))
1778  ind -= qp->sq.max;
1779  }
1780 
1781 out:
1782  if (likely(nreq)) {
1783  wmb();
1784 
1785  mthca_write64(((qp->sq.next_ind << qp->sq.wqe_shift) +
1786  qp->send_wqe_offset) | f0 | op0,
1787  (qp->qpn << 8) | size0,
1788  dev->kar + MTHCA_SEND_DOORBELL,
1789  MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1790  /*
1791  * Make sure doorbells don't leak out of SQ spinlock
1792  * and reach the HCA out of order:
1793  */
1794  mmiowb();
1795  }
1796 
1797  qp->sq.next_ind = ind;
1798  qp->sq.head += nreq;
1799 
1800  spin_unlock_irqrestore(&qp->sq.lock, flags);
1801  return err;
1802 }
1803 
1804 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1805  struct ib_recv_wr **bad_wr)
1806 {
1807  struct mthca_dev *dev = to_mdev(ibqp->device);
1808  struct mthca_qp *qp = to_mqp(ibqp);
1809  unsigned long flags;
1810  int err = 0;
1811  int nreq;
1812  int i;
1813  int size;
1814  /*
1815  * size0 is only used if nreq != 0, and it will always be
1816  * initialized the first time through the main loop before
1817  * nreq is incremented. So nreq cannot become non-zero
1818  * without initializing size0, and it is in fact never used
1819  * uninitialized.
1820  */
1821  int uninitialized_var(size0);
1822  int ind;
1823  void *wqe;
1824  void *prev_wqe;
1825 
1826  spin_lock_irqsave(&qp->rq.lock, flags);
1827 
1828  /* XXX check that state is OK to post receive */
1829 
1830  ind = qp->rq.next_ind;
1831 
1832  for (nreq = 0; wr; wr = wr->next) {
1833  if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1834  mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1835  " %d max, %d nreq)\n", qp->qpn,
1836  qp->rq.head, qp->rq.tail,
1837  qp->rq.max, nreq);
1838  err = -ENOMEM;
1839  *bad_wr = wr;
1840  goto out;
1841  }
1842 
1843  wqe = get_recv_wqe(qp, ind);
1844  prev_wqe = qp->rq.last;
1845  qp->rq.last = wqe;
1846 
1847  ((struct mthca_next_seg *) wqe)->ee_nds =
1849  ((struct mthca_next_seg *) wqe)->flags = 0;
1850 
1851  wqe += sizeof (struct mthca_next_seg);
1852  size = sizeof (struct mthca_next_seg) / 16;
1853 
1854  if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1855  err = -EINVAL;
1856  *bad_wr = wr;
1857  goto out;
1858  }
1859 
1860  for (i = 0; i < wr->num_sge; ++i) {
1861  mthca_set_data_seg(wqe, wr->sg_list + i);
1862  wqe += sizeof (struct mthca_data_seg);
1863  size += sizeof (struct mthca_data_seg) / 16;
1864  }
1865 
1866  qp->wrid[ind] = wr->wr_id;
1867 
1868  ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1869  cpu_to_be32(MTHCA_NEXT_DBD | size);
1870 
1871  if (!nreq)
1872  size0 = size;
1873 
1874  ++ind;
1875  if (unlikely(ind >= qp->rq.max))
1876  ind -= qp->rq.max;
1877 
1878  ++nreq;
1880  nreq = 0;
1881 
1882  wmb();
1883 
1884  mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1885  qp->qpn << 8, dev->kar + MTHCA_RECEIVE_DOORBELL,
1886  MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1887 
1888  qp->rq.next_ind = ind;
1890  }
1891  }
1892 
1893 out:
1894  if (likely(nreq)) {
1895  wmb();
1896 
1897  mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1898  qp->qpn << 8 | nreq, dev->kar + MTHCA_RECEIVE_DOORBELL,
1899  MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1900  }
1901 
1902  qp->rq.next_ind = ind;
1903  qp->rq.head += nreq;
1904 
1905  /*
1906  * Make sure doorbells don't leak out of RQ spinlock and reach
1907  * the HCA out of order:
1908  */
1909  mmiowb();
1910 
1911  spin_unlock_irqrestore(&qp->rq.lock, flags);
1912  return err;
1913 }
1914 
1915 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1916  struct ib_send_wr **bad_wr)
1917 {
1918  struct mthca_dev *dev = to_mdev(ibqp->device);
1919  struct mthca_qp *qp = to_mqp(ibqp);
1920  u32 dbhi;
1921  void *wqe;
1922  void *prev_wqe;
1923  unsigned long flags;
1924  int err = 0;
1925  int nreq;
1926  int i;
1927  int size;
1928  /*
1929  * f0 and size0 are only used if nreq != 0, and they will
1930  * always be initialized the first time through the main loop
1931  * before nreq is incremented. So nreq cannot become non-zero
1932  * without initializing f0 and size0, and they are in fact
1933  * never used uninitialized.
1934  */
1935  int uninitialized_var(size0);
1937  int ind;
1938  u8 op0 = 0;
1939 
1940  spin_lock_irqsave(&qp->sq.lock, flags);
1941 
1942  /* XXX check that state is OK to post send */
1943 
1944  ind = qp->sq.head & (qp->sq.max - 1);
1945 
1946  for (nreq = 0; wr; ++nreq, wr = wr->next) {
1948  nreq = 0;
1949 
1950  dbhi = (MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
1951  ((qp->sq.head & 0xffff) << 8) | f0 | op0;
1952 
1954 
1955  /*
1956  * Make sure that descriptors are written before
1957  * doorbell record.
1958  */
1959  wmb();
1960  *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1961 
1962  /*
1963  * Make sure doorbell record is written before we
1964  * write MMIO send doorbell.
1965  */
1966  wmb();
1967 
1968  mthca_write64(dbhi, (qp->qpn << 8) | size0,
1969  dev->kar + MTHCA_SEND_DOORBELL,
1970  MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1971  }
1972 
1973  if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1974  mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1975  " %d max, %d nreq)\n", qp->qpn,
1976  qp->sq.head, qp->sq.tail,
1977  qp->sq.max, nreq);
1978  err = -ENOMEM;
1979  *bad_wr = wr;
1980  goto out;
1981  }
1982 
1983  wqe = get_send_wqe(qp, ind);
1984  prev_wqe = qp->sq.last;
1985  qp->sq.last = wqe;
1986 
1987  ((struct mthca_next_seg *) wqe)->flags =
1988  ((wr->send_flags & IB_SEND_SIGNALED) ?
1990  ((wr->send_flags & IB_SEND_SOLICITED) ?
1992  ((wr->send_flags & IB_SEND_IP_CSUM) ?
1994  cpu_to_be32(1);
1995  if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1996  wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1997  ((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
1998 
1999  wqe += sizeof (struct mthca_next_seg);
2000  size = sizeof (struct mthca_next_seg) / 16;
2001 
2002  switch (qp->transport) {
2003  case RC:
2004  switch (wr->opcode) {
2007  set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2008  wr->wr.atomic.rkey);
2009  wqe += sizeof (struct mthca_raddr_seg);
2010 
2011  set_atomic_seg(wqe, wr);
2012  wqe += sizeof (struct mthca_atomic_seg);
2013  size += (sizeof (struct mthca_raddr_seg) +
2014  sizeof (struct mthca_atomic_seg)) / 16;
2015  break;
2016 
2017  case IB_WR_RDMA_READ:
2018  case IB_WR_RDMA_WRITE:
2020  set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
2021  wr->wr.rdma.rkey);
2022  wqe += sizeof (struct mthca_raddr_seg);
2023  size += sizeof (struct mthca_raddr_seg) / 16;
2024  break;
2025 
2026  default:
2027  /* No extra segments required for sends */
2028  break;
2029  }
2030 
2031  break;
2032 
2033  case UC:
2034  switch (wr->opcode) {
2035  case IB_WR_RDMA_WRITE:
2037  set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
2038  wr->wr.rdma.rkey);
2039  wqe += sizeof (struct mthca_raddr_seg);
2040  size += sizeof (struct mthca_raddr_seg) / 16;
2041  break;
2042 
2043  default:
2044  /* No extra segments required for sends */
2045  break;
2046  }
2047 
2048  break;
2049 
2050  case UD:
2051  set_arbel_ud_seg(wqe, wr);
2052  wqe += sizeof (struct mthca_arbel_ud_seg);
2053  size += sizeof (struct mthca_arbel_ud_seg) / 16;
2054  break;
2055 
2056  case MLX:
2057  err = build_mlx_header(dev, to_msqp(qp), ind, wr,
2058  wqe - sizeof (struct mthca_next_seg),
2059  wqe);
2060  if (err) {
2061  *bad_wr = wr;
2062  goto out;
2063  }
2064  wqe += sizeof (struct mthca_data_seg);
2065  size += sizeof (struct mthca_data_seg) / 16;
2066  break;
2067  }
2068 
2069  if (wr->num_sge > qp->sq.max_gs) {
2070  mthca_err(dev, "too many gathers\n");
2071  err = -EINVAL;
2072  *bad_wr = wr;
2073  goto out;
2074  }
2075 
2076  for (i = 0; i < wr->num_sge; ++i) {
2077  mthca_set_data_seg(wqe, wr->sg_list + i);
2078  wqe += sizeof (struct mthca_data_seg);
2079  size += sizeof (struct mthca_data_seg) / 16;
2080  }
2081 
2082  /* Add one more inline data segment for ICRC */
2083  if (qp->transport == MLX) {
2084  ((struct mthca_data_seg *) wqe)->byte_count =
2085  cpu_to_be32((1 << 31) | 4);
2086  ((u32 *) wqe)[1] = 0;
2087  wqe += sizeof (struct mthca_data_seg);
2088  size += sizeof (struct mthca_data_seg) / 16;
2089  }
2090 
2091  qp->wrid[ind + qp->rq.max] = wr->wr_id;
2092 
2093  if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
2094  mthca_err(dev, "opcode invalid\n");
2095  err = -EINVAL;
2096  *bad_wr = wr;
2097  goto out;
2098  }
2099 
2100  ((struct mthca_next_seg *) prev_wqe)->nda_op =
2101  cpu_to_be32(((ind << qp->sq.wqe_shift) +
2102  qp->send_wqe_offset) |
2103  mthca_opcode[wr->opcode]);
2104  wmb();
2105  ((struct mthca_next_seg *) prev_wqe)->ee_nds =
2106  cpu_to_be32(MTHCA_NEXT_DBD | size |
2107  ((wr->send_flags & IB_SEND_FENCE) ?
2108  MTHCA_NEXT_FENCE : 0));
2109 
2110  if (!nreq) {
2111  size0 = size;
2112  op0 = mthca_opcode[wr->opcode];
2113  f0 = wr->send_flags & IB_SEND_FENCE ?
2115  }
2116 
2117  ++ind;
2118  if (unlikely(ind >= qp->sq.max))
2119  ind -= qp->sq.max;
2120  }
2121 
2122 out:
2123  if (likely(nreq)) {
2124  dbhi = (nreq << 24) | ((qp->sq.head & 0xffff) << 8) | f0 | op0;
2125 
2126  qp->sq.head += nreq;
2127 
2128  /*
2129  * Make sure that descriptors are written before
2130  * doorbell record.
2131  */
2132  wmb();
2133  *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2134 
2135  /*
2136  * Make sure doorbell record is written before we
2137  * write MMIO send doorbell.
2138  */
2139  wmb();
2140 
2141  mthca_write64(dbhi, (qp->qpn << 8) | size0, dev->kar + MTHCA_SEND_DOORBELL,
2142  MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2143  }
2144 
2145  /*
2146  * Make sure doorbells don't leak out of SQ spinlock and reach
2147  * the HCA out of order:
2148  */
2149  mmiowb();
2150 
2151  spin_unlock_irqrestore(&qp->sq.lock, flags);
2152  return err;
2153 }
2154 
2155 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2156  struct ib_recv_wr **bad_wr)
2157 {
2158  struct mthca_dev *dev = to_mdev(ibqp->device);
2159  struct mthca_qp *qp = to_mqp(ibqp);
2160  unsigned long flags;
2161  int err = 0;
2162  int nreq;
2163  int ind;
2164  int i;
2165  void *wqe;
2166 
2167  spin_lock_irqsave(&qp->rq.lock, flags);
2168 
2169  /* XXX check that state is OK to post receive */
2170 
2171  ind = qp->rq.head & (qp->rq.max - 1);
2172 
2173  for (nreq = 0; wr; ++nreq, wr = wr->next) {
2174  if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2175  mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2176  " %d max, %d nreq)\n", qp->qpn,
2177  qp->rq.head, qp->rq.tail,
2178  qp->rq.max, nreq);
2179  err = -ENOMEM;
2180  *bad_wr = wr;
2181  goto out;
2182  }
2183 
2184  wqe = get_recv_wqe(qp, ind);
2185 
2186  ((struct mthca_next_seg *) wqe)->flags = 0;
2187 
2188  wqe += sizeof (struct mthca_next_seg);
2189 
2190  if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2191  err = -EINVAL;
2192  *bad_wr = wr;
2193  goto out;
2194  }
2195 
2196  for (i = 0; i < wr->num_sge; ++i) {
2197  mthca_set_data_seg(wqe, wr->sg_list + i);
2198  wqe += sizeof (struct mthca_data_seg);
2199  }
2200 
2201  if (i < qp->rq.max_gs)
2202  mthca_set_data_seg_inval(wqe);
2203 
2204  qp->wrid[ind] = wr->wr_id;
2205 
2206  ++ind;
2207  if (unlikely(ind >= qp->rq.max))
2208  ind -= qp->rq.max;
2209  }
2210 out:
2211  if (likely(nreq)) {
2212  qp->rq.head += nreq;
2213 
2214  /*
2215  * Make sure that descriptors are written before
2216  * doorbell record.
2217  */
2218  wmb();
2219  *qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2220  }
2221 
2222  spin_unlock_irqrestore(&qp->rq.lock, flags);
2223  return err;
2224 }
2225 
2226 void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2227  int index, int *dbd, __be32 *new_wqe)
2228 {
2229  struct mthca_next_seg *next;
2230 
2231  /*
2232  * For SRQs, all receive WQEs generate a CQE, so we're always
2233  * at the end of the doorbell chain.
2234  */
2235  if (qp->ibqp.srq && !is_send) {
2236  *new_wqe = 0;
2237  return;
2238  }
2239 
2240  if (is_send)
2241  next = get_send_wqe(qp, index);
2242  else
2243  next = get_recv_wqe(qp, index);
2244 
2245  *dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2246  if (next->ee_nds & cpu_to_be32(0x3f))
2247  *new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2248  (next->ee_nds & cpu_to_be32(0x3f));
2249  else
2250  *new_wqe = 0;
2251 }
2252 
2254 {
2255  int err;
2256  int i;
2257 
2258  spin_lock_init(&dev->qp_table.lock);
2259 
2260  /*
2261  * We reserve 2 extra QPs per port for the special QPs. The
2262  * special QP for port 1 has to be even, so round up.
2263  */
2264  dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2265  err = mthca_alloc_init(&dev->qp_table.alloc,
2266  dev->limits.num_qps,
2267  (1 << 24) - 1,
2268  dev->qp_table.sqp_start +
2269  MTHCA_MAX_PORTS * 2);
2270  if (err)
2271  return err;
2272 
2273  err = mthca_array_init(&dev->qp_table.qp,
2274  dev->limits.num_qps);
2275  if (err) {
2276  mthca_alloc_cleanup(&dev->qp_table.alloc);
2277  return err;
2278  }
2279 
2280  for (i = 0; i < 2; ++i) {
2281  err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2282  dev->qp_table.sqp_start + i * 2);
2283  if (err) {
2284  mthca_warn(dev, "CONF_SPECIAL_QP returned "
2285  "%d, aborting.\n", err);
2286  goto err_out;
2287  }
2288  }
2289  return 0;
2290 
2291  err_out:
2292  for (i = 0; i < 2; ++i)
2293  mthca_CONF_SPECIAL_QP(dev, i, 0);
2294 
2295  mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2296  mthca_alloc_cleanup(&dev->qp_table.alloc);
2297 
2298  return err;
2299 }
2300 
2302 {
2303  int i;
2304 
2305  for (i = 0; i < 2; ++i)
2306  mthca_CONF_SPECIAL_QP(dev, i, 0);
2307 
2308  mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2309  mthca_alloc_cleanup(&dev->qp_table.alloc);
2310 }