git.oblomov.eu Git - linux-2.6/blob - drivers/infiniband/hw/mthca/mthca_srq.c

   1 /*
   2  * Copyright (c) 2005 Cisco Systems. All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  * $Id: mthca_srq.c 3047 2005-08-10 03:59:35Z roland $
  33  */
  34
  35 #include <linux/slab.h>
  36 #include <linux/string.h>
  37
  38 #include "mthca_dev.h"
  39 #include "mthca_cmd.h"
  40 #include "mthca_memfree.h"
  41 #include "mthca_wqe.h"
  42
  43 enum {
  44         MTHCA_MAX_DIRECT_SRQ_SIZE = 4 * PAGE_SIZE
  45 };
  46
  47 struct mthca_tavor_srq_context {
  48         __be64 wqe_base_ds;     /* low 6 bits is descriptor size */
  49         __be32 state_pd;
  50         __be32 lkey;
  51         __be32 uar;
  52         __be16 limit_watermark;
  53         __be16 wqe_cnt;
  54         u32    reserved[2];
  55 };
  56
  57 struct mthca_arbel_srq_context {
  58         __be32 state_logsize_srqn;
  59         __be32 lkey;
  60         __be32 db_index;
  61         __be32 logstride_usrpage;
  62         __be64 wqe_base;
  63         __be32 eq_pd;
  64         __be16 limit_watermark;
  65         __be16 wqe_cnt;
  66         u16    reserved1;
  67         __be16 wqe_counter;
  68         u32    reserved2[3];
  69 };
  70
  71 static void *get_wqe(struct mthca_srq *srq, int n)
  72 {
  73         if (srq->is_direct)
  74                 return srq->queue.direct.buf + (n << srq->wqe_shift);
  75         else
  76                 return srq->queue.page_list[(n << srq->wqe_shift) >> PAGE_SHIFT].buf +
  77                         ((n << srq->wqe_shift) & (PAGE_SIZE - 1));
  78 }
  79
  80 /*
  81  * Return a pointer to the location within a WQE that we're using as a
  82  * link when the WQE is in the free list.  We use the imm field
  83  * because in the Tavor case, posting a WQE may overwrite the next
  84  * segment of the previous WQE, but a receive WQE will never touch the
  85  * imm field.  This avoids corrupting our free list if the previous
  86  * WQE has already completed and been put on the free list when we
  87  * post the next WQE.
  88  */
  89 static inline int *wqe_to_link(void *wqe)
  90 {
  91         return (int *) (wqe + offsetof(struct mthca_next_seg, imm));
  92 }
  93
  94 static void mthca_tavor_init_srq_context(struct mthca_dev *dev,
  95                                          struct mthca_pd *pd,
  96                                          struct mthca_srq *srq,
  97                                          struct mthca_tavor_srq_context *context)
  98 {
  99         memset(context, 0, sizeof *context);
 100
 101         context->wqe_base_ds = cpu_to_be64(1 << (srq->wqe_shift - 4));
 102         context->state_pd    = cpu_to_be32(pd->pd_num);
 103         context->lkey        = cpu_to_be32(srq->mr.ibmr.lkey);
 104
 105         if (pd->ibpd.uobject)
 106                 context->uar =
 107                         cpu_to_be32(to_mucontext(pd->ibpd.uobject->context)->uar.index);
 108         else
 109                 context->uar = cpu_to_be32(dev->driver_uar.index);
 110 }
 111
 112 static void mthca_arbel_init_srq_context(struct mthca_dev *dev,
 113                                          struct mthca_pd *pd,
 114                                          struct mthca_srq *srq,
 115                                          struct mthca_arbel_srq_context *context)
 116 {
 117         int logsize;
 118
 119         memset(context, 0, sizeof *context);
 120
 121         logsize = long_log2(srq->max) + srq->wqe_shift;
 122         context->state_logsize_srqn = cpu_to_be32(logsize << 24 | srq->srqn);
 123         context->lkey = cpu_to_be32(srq->mr.ibmr.lkey);
 124         context->db_index = cpu_to_be32(srq->db_index);
 125         context->logstride_usrpage = cpu_to_be32((srq->wqe_shift - 4) << 29);
 126         if (pd->ibpd.uobject)
 127                 context->logstride_usrpage |=
 128                         cpu_to_be32(to_mucontext(pd->ibpd.uobject->context)->uar.index);
 129         else
 130                 context->logstride_usrpage |= cpu_to_be32(dev->driver_uar.index);
 131         context->eq_pd = cpu_to_be32(MTHCA_EQ_ASYNC << 24 | pd->pd_num);
 132 }
 133
 134 static void mthca_free_srq_buf(struct mthca_dev *dev, struct mthca_srq *srq)
 135 {
 136         mthca_buf_free(dev, srq->max << srq->wqe_shift, &srq->queue,
 137                        srq->is_direct, &srq->mr);
 138         kfree(srq->wrid);
 139 }
 140
 141 static int mthca_alloc_srq_buf(struct mthca_dev *dev, struct mthca_pd *pd,
 142                                struct mthca_srq *srq)
 143 {
 144         struct mthca_data_seg *scatter;
 145         void *wqe;
 146         int err;
 147         int i;
 148
 149         if (pd->ibpd.uobject)
 150                 return 0;
 151
 152         srq->wrid = kmalloc(srq->max * sizeof (u64), GFP_KERNEL);
 153         if (!srq->wrid)
 154                 return -ENOMEM;
 155
 156         err = mthca_buf_alloc(dev, srq->max << srq->wqe_shift,
 157                               MTHCA_MAX_DIRECT_SRQ_SIZE,
 158                               &srq->queue, &srq->is_direct, pd, 1, &srq->mr);
 159         if (err) {
 160                 kfree(srq->wrid);
 161                 return err;
 162         }
 163
 164         /*
 165          * Now initialize the SRQ buffer so that all of the WQEs are
 166          * linked into the list of free WQEs.  In addition, set the
 167          * scatter list L_Keys to the sentry value of 0x100.
 168          */
 169         for (i = 0; i < srq->max; ++i) {
 170                 wqe = get_wqe(srq, i);
 171
 172                 *wqe_to_link(wqe) = i < srq->max - 1 ? i + 1 : -1;
 173
 174                 for (scatter = wqe + sizeof (struct mthca_next_seg);
 175                      (void *) scatter < wqe + (1 << srq->wqe_shift);
 176                      ++scatter)
 177                         scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
 178         }
 179
 180         srq->last = get_wqe(srq, srq->max - 1);
 181
 182         return 0;
 183 }
 184
 185 int mthca_alloc_srq(struct mthca_dev *dev, struct mthca_pd *pd,
 186                     struct ib_srq_attr *attr, struct mthca_srq *srq)
 187 {
 188         struct mthca_mailbox *mailbox;
 189         u8 status;
 190         int ds;
 191         int err;
 192
 193         /* Sanity check SRQ size before proceeding */
 194         if (attr->max_wr  > dev->limits.max_srq_wqes ||
 195             attr->max_sge > dev->limits.max_srq_sge)
 196                 return -EINVAL;
 197
 198         srq->max      = attr->max_wr;
 199         srq->max_gs   = attr->max_sge;
 200         srq->counter  = 0;
 201
 202         if (mthca_is_memfree(dev))
 203                 srq->max = roundup_pow_of_two(srq->max + 1);
 204
 205         ds = max(64UL,
 206                  roundup_pow_of_two(sizeof (struct mthca_next_seg) +
 207                                     srq->max_gs * sizeof (struct mthca_data_seg)));
 208
 209         if (!mthca_is_memfree(dev) && (ds > dev->limits.max_desc_sz))
 210                 return -EINVAL;
 211
 212         srq->wqe_shift = long_log2(ds);
 213
 214         srq->srqn = mthca_alloc(&dev->srq_table.alloc);
 215         if (srq->srqn == -1)
 216                 return -ENOMEM;
 217
 218         if (mthca_is_memfree(dev)) {
 219                 err = mthca_table_get(dev, dev->srq_table.table, srq->srqn);
 220                 if (err)
 221                         goto err_out;
 222
 223                 if (!pd->ibpd.uobject) {
 224                         srq->db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SRQ,
 225                                                        srq->srqn, &srq->db);
 226                         if (srq->db_index < 0) {
 227                                 err = -ENOMEM;
 228                                 goto err_out_icm;
 229                         }
 230                 }
 231         }
 232
 233         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
 234         if (IS_ERR(mailbox)) {
 235                 err = PTR_ERR(mailbox);
 236                 goto err_out_db;
 237         }
 238
 239         err = mthca_alloc_srq_buf(dev, pd, srq);
 240         if (err)
 241                 goto err_out_mailbox;
 242
 243         spin_lock_init(&srq->lock);
 244         srq->refcount = 1;
 245         init_waitqueue_head(&srq->wait);
 246         mutex_init(&srq->mutex);
 247
 248         if (mthca_is_memfree(dev))
 249                 mthca_arbel_init_srq_context(dev, pd, srq, mailbox->buf);
 250         else
 251                 mthca_tavor_init_srq_context(dev, pd, srq, mailbox->buf);
 252
 253         err = mthca_SW2HW_SRQ(dev, mailbox, srq->srqn, &status);
 254
 255         if (err) {
 256                 mthca_warn(dev, "SW2HW_SRQ failed (%d)\n", err);
 257                 goto err_out_free_buf;
 258         }
 259         if (status) {
 260                 mthca_warn(dev, "SW2HW_SRQ returned status 0x%02x\n",
 261                            status);
 262                 err = -EINVAL;
 263                 goto err_out_free_buf;
 264         }
 265
 266         spin_lock_irq(&dev->srq_table.lock);
 267         if (mthca_array_set(&dev->srq_table.srq,
 268                             srq->srqn & (dev->limits.num_srqs - 1),
 269                             srq)) {
 270                 spin_unlock_irq(&dev->srq_table.lock);
 271                 goto err_out_free_srq;
 272         }
 273         spin_unlock_irq(&dev->srq_table.lock);
 274
 275         mthca_free_mailbox(dev, mailbox);
 276
 277         srq->first_free = 0;
 278         srq->last_free  = srq->max - 1;
 279
 280         attr->max_wr    = (mthca_is_memfree(dev)) ? srq->max - 1 : srq->max;
 281         attr->max_sge   = srq->max_gs;
 282
 283         return 0;
 284
 285 err_out_free_srq:
 286         err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn, &status);
 287         if (err)
 288                 mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
 289         else if (status)
 290                 mthca_warn(dev, "HW2SW_SRQ returned status 0x%02x\n", status);
 291
 292 err_out_free_buf:
 293         if (!pd->ibpd.uobject)
 294                 mthca_free_srq_buf(dev, srq);
 295
 296 err_out_mailbox:
 297         mthca_free_mailbox(dev, mailbox);
 298
 299 err_out_db:
 300         if (!pd->ibpd.uobject && mthca_is_memfree(dev))
 301                 mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
 302
 303 err_out_icm:
 304         mthca_table_put(dev, dev->srq_table.table, srq->srqn);
 305
 306 err_out:
 307         mthca_free(&dev->srq_table.alloc, srq->srqn);
 308
 309         return err;
 310 }
 311
 312 static inline int get_srq_refcount(struct mthca_dev *dev, struct mthca_srq *srq)
 313 {
 314         int c;
 315
 316         spin_lock_irq(&dev->srq_table.lock);
 317         c = srq->refcount;
 318         spin_unlock_irq(&dev->srq_table.lock);
 319
 320         return c;
 321 }
 322
 323 void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq)
 324 {
 325         struct mthca_mailbox *mailbox;
 326         int err;
 327         u8 status;
 328
 329         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
 330         if (IS_ERR(mailbox)) {
 331                 mthca_warn(dev, "No memory for mailbox to free SRQ.\n");
 332                 return;
 333         }
 334
 335         err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn, &status);
 336         if (err)
 337                 mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
 338         else if (status)
 339                 mthca_warn(dev, "HW2SW_SRQ returned status 0x%02x\n", status);
 340
 341         spin_lock_irq(&dev->srq_table.lock);
 342         mthca_array_clear(&dev->srq_table.srq,
 343                           srq->srqn & (dev->limits.num_srqs - 1));
 344         --srq->refcount;
 345         spin_unlock_irq(&dev->srq_table.lock);
 346
 347         wait_event(srq->wait, !get_srq_refcount(dev, srq));
 348
 349         if (!srq->ibsrq.uobject) {
 350                 mthca_free_srq_buf(dev, srq);
 351                 if (mthca_is_memfree(dev))
 352                         mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
 353         }
 354
 355         mthca_table_put(dev, dev->srq_table.table, srq->srqn);
 356         mthca_free(&dev->srq_table.alloc, srq->srqn);
 357         mthca_free_mailbox(dev, mailbox);
 358 }
 359
 360 int mthca_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
 361                      enum ib_srq_attr_mask attr_mask)
 362 {
 363         struct mthca_dev *dev = to_mdev(ibsrq->device);
 364         struct mthca_srq *srq = to_msrq(ibsrq);
 365         int ret;
 366         u8 status;
 367
 368         /* We don't support resizing SRQs (yet?) */
 369         if (attr_mask & IB_SRQ_MAX_WR)
 370                 return -EINVAL;
 371
 372         if (attr_mask & IB_SRQ_LIMIT) {
 373                 if (attr->srq_limit > srq->max)
 374                         return -EINVAL;
 375
 376                 mutex_lock(&srq->mutex);
 377                 ret = mthca_ARM_SRQ(dev, srq->srqn, attr->srq_limit, &status);
 378                 mutex_unlock(&srq->mutex);
 379
 380                 if (ret)
 381                         return ret;
 382                 if (status)
 383                         return -EINVAL;
 384         }
 385
 386         return 0;
 387 }
 388
 389 int mthca_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
 390 {
 391         struct mthca_dev *dev = to_mdev(ibsrq->device);
 392         struct mthca_srq *srq = to_msrq(ibsrq);
 393         struct mthca_mailbox *mailbox;
 394         struct mthca_arbel_srq_context *arbel_ctx;
 395         struct mthca_tavor_srq_context *tavor_ctx;
 396         u8 status;
 397         int err;
 398
 399         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
 400         if (IS_ERR(mailbox))
 401                 return PTR_ERR(mailbox);
 402
 403         err = mthca_QUERY_SRQ(dev, srq->srqn, mailbox, &status);
 404         if (err)
 405                 goto out;
 406
 407         if (mthca_is_memfree(dev)) {
 408                 arbel_ctx = mailbox->buf;
 409                 srq_attr->srq_limit = be16_to_cpu(arbel_ctx->limit_watermark);
 410         } else {
 411                 tavor_ctx = mailbox->buf;
 412                 srq_attr->srq_limit = be16_to_cpu(tavor_ctx->limit_watermark);
 413         }
 414
 415         srq_attr->max_wr  = (mthca_is_memfree(dev)) ? srq->max - 1 : srq->max;
 416         srq_attr->max_sge = srq->max_gs;
 417
 418 out:
 419         mthca_free_mailbox(dev, mailbox);
 420
 421         return err;
 422 }
 423
 424 void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
 425                      enum ib_event_type event_type)
 426 {
 427         struct mthca_srq *srq;
 428         struct ib_event event;
 429
 430         spin_lock(&dev->srq_table.lock);
 431         srq = mthca_array_get(&dev->srq_table.srq, srqn & (dev->limits.num_srqs - 1));
 432         if (srq)
 433                 ++srq->refcount;
 434         spin_unlock(&dev->srq_table.lock);
 435
 436         if (!srq) {
 437                 mthca_warn(dev, "Async event for bogus SRQ %08x\n", srqn);
 438                 return;
 439         }
 440
 441         if (!srq->ibsrq.event_handler)
 442                 goto out;
 443
 444         event.device      = &dev->ib_dev;
 445         event.event       = event_type;
 446         event.element.srq = &srq->ibsrq;
 447         srq->ibsrq.event_handler(&event, srq->ibsrq.srq_context);
 448
 449 out:
 450         spin_lock(&dev->srq_table.lock);
 451         if (!--srq->refcount)
 452                 wake_up(&srq->wait);
 453         spin_unlock(&dev->srq_table.lock);
 454 }
 455
 456 /*
 457  * This function must be called with IRQs disabled.
 458  */
 459 void mthca_free_srq_wqe(struct mthca_srq *srq, u32 wqe_addr)
 460 {
 461         int ind;
 462
 463         ind = wqe_addr >> srq->wqe_shift;
 464
 465         spin_lock(&srq->lock);
 466
 467         if (likely(srq->first_free >= 0))
 468                 *wqe_to_link(get_wqe(srq, srq->last_free)) = ind;
 469         else
 470                 srq->first_free = ind;
 471
 472         *wqe_to_link(get_wqe(srq, ind)) = -1;
 473         srq->last_free = ind;
 474
 475         spin_unlock(&srq->lock);
 476 }
 477
 478 int mthca_tavor_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
 479                               struct ib_recv_wr **bad_wr)
 480 {
 481         struct mthca_dev *dev = to_mdev(ibsrq->device);
 482         struct mthca_srq *srq = to_msrq(ibsrq);
 483         __be32 doorbell[2];
 484         unsigned long flags;
 485         int err = 0;
 486         int first_ind;
 487         int ind;
 488         int next_ind;
 489         int nreq;
 490         int i;
 491         void *wqe;
 492         void *prev_wqe;
 493
 494         spin_lock_irqsave(&srq->lock, flags);
 495
 496         first_ind = srq->first_free;
 497
 498         for (nreq = 0; wr; wr = wr->next) {
 499                 ind = srq->first_free;
 500
 501                 if (ind < 0) {
 502                         mthca_err(dev, "SRQ %06x full\n", srq->srqn);
 503                         err = -ENOMEM;
 504                         *bad_wr = wr;
 505                         break;
 506                 }
 507
 508                 wqe       = get_wqe(srq, ind);
 509                 next_ind  = *wqe_to_link(wqe);
 510
 511                 if (next_ind < 0) {
 512                         mthca_err(dev, "SRQ %06x full\n", srq->srqn);
 513                         err = -ENOMEM;
 514                         *bad_wr = wr;
 515                         break;
 516                 }
 517
 518                 prev_wqe  = srq->last;
 519                 srq->last = wqe;
 520
 521                 ((struct mthca_next_seg *) wqe)->nda_op = 0;
 522                 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
 523                 /* flags field will always remain 0 */
 524
 525                 wqe += sizeof (struct mthca_next_seg);
 526
 527                 if (unlikely(wr->num_sge > srq->max_gs)) {
 528                         err = -EINVAL;
 529                         *bad_wr = wr;
 530                         srq->last = prev_wqe;
 531                         break;
 532                 }
 533
 534                 for (i = 0; i < wr->num_sge; ++i) {
 535                         ((struct mthca_data_seg *) wqe)->byte_count =
 536                                 cpu_to_be32(wr->sg_list[i].length);
 537                         ((struct mthca_data_seg *) wqe)->lkey =
 538                                 cpu_to_be32(wr->sg_list[i].lkey);
 539                         ((struct mthca_data_seg *) wqe)->addr =
 540                                 cpu_to_be64(wr->sg_list[i].addr);
 541                         wqe += sizeof (struct mthca_data_seg);
 542                 }
 543
 544                 if (i < srq->max_gs) {
 545                         ((struct mthca_data_seg *) wqe)->byte_count = 0;
 546                         ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
 547                         ((struct mthca_data_seg *) wqe)->addr = 0;
 548                 }
 549
 550                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
 551                         cpu_to_be32((ind << srq->wqe_shift) | 1);
 552                 wmb();
 553                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
 554                         cpu_to_be32(MTHCA_NEXT_DBD);
 555
 556                 srq->wrid[ind]  = wr->wr_id;
 557                 srq->first_free = next_ind;
 558
 559                 ++nreq;
 560                 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
 561                         nreq = 0;
 562
 563                         doorbell[0] = cpu_to_be32(first_ind << srq->wqe_shift);
 564                         doorbell[1] = cpu_to_be32(srq->srqn << 8);
 565
 566                         /*
 567                          * Make sure that descriptors are written
 568                          * before doorbell is rung.
 569                          */
 570                         wmb();
 571
 572                         mthca_write64(doorbell,
 573                                       dev->kar + MTHCA_RECEIVE_DOORBELL,
 574                                       MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
 575
 576                         first_ind = srq->first_free;
 577                 }
 578         }
 579
 580         if (likely(nreq)) {
 581                 doorbell[0] = cpu_to_be32(first_ind << srq->wqe_shift);
 582                 doorbell[1] = cpu_to_be32((srq->srqn << 8) | nreq);
 583
 584                 /*
 585                  * Make sure that descriptors are written before
 586                  * doorbell is rung.
 587                  */
 588                 wmb();
 589
 590                 mthca_write64(doorbell,
 591                               dev->kar + MTHCA_RECEIVE_DOORBELL,
 592                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
 593         }
 594
 595         spin_unlock_irqrestore(&srq->lock, flags);
 596         return err;
 597 }
 598
 599 int mthca_arbel_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
 600                               struct ib_recv_wr **bad_wr)
 601 {
 602         struct mthca_dev *dev = to_mdev(ibsrq->device);
 603         struct mthca_srq *srq = to_msrq(ibsrq);
 604         unsigned long flags;
 605         int err = 0;
 606         int ind;
 607         int next_ind;
 608         int nreq;
 609         int i;
 610         void *wqe;
 611
 612         spin_lock_irqsave(&srq->lock, flags);
 613
 614         for (nreq = 0; wr; ++nreq, wr = wr->next) {
 615                 ind = srq->first_free;
 616
 617                 if (ind < 0) {
 618                         mthca_err(dev, "SRQ %06x full\n", srq->srqn);
 619                         err = -ENOMEM;
 620                         *bad_wr = wr;
 621                         break;
 622                 }
 623
 624                 wqe       = get_wqe(srq, ind);
 625                 next_ind  = *wqe_to_link(wqe);
 626
 627                 if (next_ind < 0) {
 628                         mthca_err(dev, "SRQ %06x full\n", srq->srqn);
 629                         err = -ENOMEM;
 630                         *bad_wr = wr;
 631                         break;
 632                 }
 633
 634                 ((struct mthca_next_seg *) wqe)->nda_op =
 635                         cpu_to_be32((next_ind << srq->wqe_shift) | 1);
 636                 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
 637                 /* flags field will always remain 0 */
 638
 639                 wqe += sizeof (struct mthca_next_seg);
 640
 641                 if (unlikely(wr->num_sge > srq->max_gs)) {
 642                         err = -EINVAL;
 643                         *bad_wr = wr;
 644                         break;
 645                 }
 646
 647                 for (i = 0; i < wr->num_sge; ++i) {
 648                         ((struct mthca_data_seg *) wqe)->byte_count =
 649                                 cpu_to_be32(wr->sg_list[i].length);
 650                         ((struct mthca_data_seg *) wqe)->lkey =
 651                                 cpu_to_be32(wr->sg_list[i].lkey);
 652                         ((struct mthca_data_seg *) wqe)->addr =
 653                                 cpu_to_be64(wr->sg_list[i].addr);
 654                         wqe += sizeof (struct mthca_data_seg);
 655                 }
 656
 657                 if (i < srq->max_gs) {
 658                         ((struct mthca_data_seg *) wqe)->byte_count = 0;
 659                         ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
 660                         ((struct mthca_data_seg *) wqe)->addr = 0;
 661                 }
 662
 663                 srq->wrid[ind]  = wr->wr_id;
 664                 srq->first_free = next_ind;
 665         }
 666
 667         if (likely(nreq)) {
 668                 srq->counter += nreq;
 669
 670                 /*
 671                  * Make sure that descriptors are written before
 672                  * we write doorbell record.
 673                  */
 674                 wmb();
 675                 *srq->db = cpu_to_be32(srq->counter);
 676         }
 677
 678         spin_unlock_irqrestore(&srq->lock, flags);
 679         return err;
 680 }
 681
 682 int mthca_max_srq_sge(struct mthca_dev *dev)
 683 {
 684         if (mthca_is_memfree(dev))
 685                 return dev->limits.max_sg;
 686
 687         /*
 688          * SRQ allocations are based on powers of 2 for Tavor,
 689          * (although they only need to be multiples of 16 bytes).
 690          *
 691          * Therefore, we need to base the max number of sg entries on
 692          * the largest power of 2 descriptor size that is <= to the
 693          * actual max WQE descriptor size, rather than return the
 694          * max_sg value given by the firmware (which is based on WQE
 695          * sizes as multiples of 16, not powers of 2).
 696          *
 697          * If SRQ implementation is changed for Tavor to be based on
 698          * multiples of 16, the calculation below can be deleted and
 699          * the FW max_sg value returned.
 700          */
 701         return min_t(int, dev->limits.max_sg,
 702                      ((1 << (fls(dev->limits.max_desc_sz) - 1)) -
 703                       sizeof (struct mthca_next_seg)) /
 704                      sizeof (struct mthca_data_seg));
 705 }
 706
 707 int __devinit mthca_init_srq_table(struct mthca_dev *dev)
 708 {
 709         int err;
 710
 711         if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
 712                 return 0;
 713
 714         spin_lock_init(&dev->srq_table.lock);
 715
 716         err = mthca_alloc_init(&dev->srq_table.alloc,
 717                                dev->limits.num_srqs,
 718                                dev->limits.num_srqs - 1,
 719                                dev->limits.reserved_srqs);
 720         if (err)
 721                 return err;
 722
 723         err = mthca_array_init(&dev->srq_table.srq,
 724                                dev->limits.num_srqs);
 725         if (err)
 726                 mthca_alloc_cleanup(&dev->srq_table.alloc);
 727
 728         return err;
 729 }
 730
 731 void mthca_cleanup_srq_table(struct mthca_dev *dev)
 732 {
 733         if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
 734                 return;
 735
 736         mthca_array_cleanup(&dev->srq_table.srq, dev->limits.num_srqs);
 737         mthca_alloc_cleanup(&dev->srq_table.alloc);
 738 }