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
 247         if (mthca_is_memfree(dev))
 248                 mthca_arbel_init_srq_context(dev, pd, srq, mailbox->buf);
 249         else
 250                 mthca_tavor_init_srq_context(dev, pd, srq, mailbox->buf);
 251
 252         err = mthca_SW2HW_SRQ(dev, mailbox, srq->srqn, &status);
 253
 254         if (err) {
 255                 mthca_warn(dev, "SW2HW_SRQ failed (%d)\n", err);
 256                 goto err_out_free_buf;
 257         }
 258         if (status) {
 259                 mthca_warn(dev, "SW2HW_SRQ returned status 0x%02x\n",
 260                            status);
 261                 err = -EINVAL;
 262                 goto err_out_free_buf;
 263         }
 264
 265         spin_lock_irq(&dev->srq_table.lock);
 266         if (mthca_array_set(&dev->srq_table.srq,
 267                             srq->srqn & (dev->limits.num_srqs - 1),
 268                             srq)) {
 269                 spin_unlock_irq(&dev->srq_table.lock);
 270                 goto err_out_free_srq;
 271         }
 272         spin_unlock_irq(&dev->srq_table.lock);
 273
 274         mthca_free_mailbox(dev, mailbox);
 275
 276         srq->first_free = 0;
 277         srq->last_free  = srq->max - 1;
 278
 279         attr->max_wr    = (mthca_is_memfree(dev)) ? srq->max - 1 : srq->max;
 280         attr->max_sge   = srq->max_gs;
 281
 282         return 0;
 283
 284 err_out_free_srq:
 285         err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn, &status);
 286         if (err)
 287                 mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
 288         else if (status)
 289                 mthca_warn(dev, "HW2SW_SRQ returned status 0x%02x\n", status);
 290
 291 err_out_free_buf:
 292         if (!pd->ibpd.uobject)
 293                 mthca_free_srq_buf(dev, srq);
 294
 295 err_out_mailbox:
 296         mthca_free_mailbox(dev, mailbox);
 297
 298 err_out_db:
 299         if (!pd->ibpd.uobject && mthca_is_memfree(dev))
 300                 mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
 301
 302 err_out_icm:
 303         mthca_table_put(dev, dev->srq_table.table, srq->srqn);
 304
 305 err_out:
 306         mthca_free(&dev->srq_table.alloc, srq->srqn);
 307
 308         return err;
 309 }
 310
 311 static inline int get_srq_refcount(struct mthca_dev *dev, struct mthca_srq *srq)
 312 {
 313         int c;
 314
 315         spin_lock_irq(&dev->srq_table.lock);
 316         c = srq->refcount;
 317         spin_unlock_irq(&dev->srq_table.lock);
 318
 319         return c;
 320 }
 321
 322 void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq)
 323 {
 324         struct mthca_mailbox *mailbox;
 325         int err;
 326         u8 status;
 327
 328         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
 329         if (IS_ERR(mailbox)) {
 330                 mthca_warn(dev, "No memory for mailbox to free SRQ.\n");
 331                 return;
 332         }
 333
 334         err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn, &status);
 335         if (err)
 336                 mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
 337         else if (status)
 338                 mthca_warn(dev, "HW2SW_SRQ returned status 0x%02x\n", status);
 339
 340         spin_lock_irq(&dev->srq_table.lock);
 341         mthca_array_clear(&dev->srq_table.srq,
 342                           srq->srqn & (dev->limits.num_srqs - 1));
 343         --srq->refcount;
 344         spin_unlock_irq(&dev->srq_table.lock);
 345
 346         wait_event(srq->wait, !get_srq_refcount(dev, srq));
 347
 348         if (!srq->ibsrq.uobject) {
 349                 mthca_free_srq_buf(dev, srq);
 350                 if (mthca_is_memfree(dev))
 351                         mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
 352         }
 353
 354         mthca_table_put(dev, dev->srq_table.table, srq->srqn);
 355         mthca_free(&dev->srq_table.alloc, srq->srqn);
 356         mthca_free_mailbox(dev, mailbox);
 357 }
 358
 359 int mthca_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
 360                      enum ib_srq_attr_mask attr_mask)
 361 {
 362         struct mthca_dev *dev = to_mdev(ibsrq->device);
 363         struct mthca_srq *srq = to_msrq(ibsrq);
 364         int ret;
 365         u8 status;
 366
 367         /* We don't support resizing SRQs (yet?) */
 368         if (attr_mask & IB_SRQ_MAX_WR)
 369                 return -EINVAL;
 370
 371         if (attr_mask & IB_SRQ_LIMIT) {
 372                 if (attr->srq_limit > srq->max)
 373                         return -EINVAL;
 374                 ret = mthca_ARM_SRQ(dev, srq->srqn, attr->srq_limit, &status);
 375                 if (ret)
 376                         return ret;
 377                 if (status)
 378                         return -EINVAL;
 379         }
 380
 381         return 0;
 382 }
 383
 384 int mthca_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
 385 {
 386         struct mthca_dev *dev = to_mdev(ibsrq->device);
 387         struct mthca_srq *srq = to_msrq(ibsrq);
 388         struct mthca_mailbox *mailbox;
 389         struct mthca_arbel_srq_context *arbel_ctx;
 390         struct mthca_tavor_srq_context *tavor_ctx;
 391         u8 status;
 392         int err;
 393
 394         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
 395         if (IS_ERR(mailbox))
 396                 return PTR_ERR(mailbox);
 397
 398         err = mthca_QUERY_SRQ(dev, srq->srqn, mailbox, &status);
 399         if (err)
 400                 goto out;
 401
 402         if (mthca_is_memfree(dev)) {
 403                 arbel_ctx = mailbox->buf;
 404                 srq_attr->srq_limit = be16_to_cpu(arbel_ctx->limit_watermark);
 405         } else {
 406                 tavor_ctx = mailbox->buf;
 407                 srq_attr->srq_limit = be16_to_cpu(tavor_ctx->limit_watermark);
 408         }
 409
 410         srq_attr->max_wr  = (mthca_is_memfree(dev)) ? srq->max - 1 : srq->max;
 411         srq_attr->max_sge = srq->max_gs;
 412
 413 out:
 414         mthca_free_mailbox(dev, mailbox);
 415
 416         return err;
 417 }
 418
 419 void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
 420                      enum ib_event_type event_type)
 421 {
 422         struct mthca_srq *srq;
 423         struct ib_event event;
 424
 425         spin_lock(&dev->srq_table.lock);
 426         srq = mthca_array_get(&dev->srq_table.srq, srqn & (dev->limits.num_srqs - 1));
 427         if (srq)
 428                 ++srq->refcount;
 429         spin_unlock(&dev->srq_table.lock);
 430
 431         if (!srq) {
 432                 mthca_warn(dev, "Async event for bogus SRQ %08x\n", srqn);
 433                 return;
 434         }
 435
 436         if (!srq->ibsrq.event_handler)
 437                 goto out;
 438
 439         event.device      = &dev->ib_dev;
 440         event.event       = event_type;
 441         event.element.srq = &srq->ibsrq;
 442         srq->ibsrq.event_handler(&event, srq->ibsrq.srq_context);
 443
 444 out:
 445         spin_lock(&dev->srq_table.lock);
 446         if (!--srq->refcount)
 447                 wake_up(&srq->wait);
 448         spin_unlock(&dev->srq_table.lock);
 449 }
 450
 451 /*
 452  * This function must be called with IRQs disabled.
 453  */
 454 void mthca_free_srq_wqe(struct mthca_srq *srq, u32 wqe_addr)
 455 {
 456         int ind;
 457
 458         ind = wqe_addr >> srq->wqe_shift;
 459
 460         spin_lock(&srq->lock);
 461
 462         if (likely(srq->first_free >= 0))
 463                 *wqe_to_link(get_wqe(srq, srq->last_free)) = ind;
 464         else
 465                 srq->first_free = ind;
 466
 467         *wqe_to_link(get_wqe(srq, ind)) = -1;
 468         srq->last_free = ind;
 469
 470         spin_unlock(&srq->lock);
 471 }
 472
 473 int mthca_tavor_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
 474                               struct ib_recv_wr **bad_wr)
 475 {
 476         struct mthca_dev *dev = to_mdev(ibsrq->device);
 477         struct mthca_srq *srq = to_msrq(ibsrq);
 478         __be32 doorbell[2];
 479         unsigned long flags;
 480         int err = 0;
 481         int first_ind;
 482         int ind;
 483         int next_ind;
 484         int nreq;
 485         int i;
 486         void *wqe;
 487         void *prev_wqe;
 488
 489         spin_lock_irqsave(&srq->lock, flags);
 490
 491         first_ind = srq->first_free;
 492
 493         for (nreq = 0; wr; wr = wr->next) {
 494                 ind = srq->first_free;
 495
 496                 if (ind < 0) {
 497                         mthca_err(dev, "SRQ %06x full\n", srq->srqn);
 498                         err = -ENOMEM;
 499                         *bad_wr = wr;
 500                         break;
 501                 }
 502
 503                 wqe       = get_wqe(srq, ind);
 504                 next_ind  = *wqe_to_link(wqe);
 505
 506                 if (next_ind < 0) {
 507                         mthca_err(dev, "SRQ %06x full\n", srq->srqn);
 508                         err = -ENOMEM;
 509                         *bad_wr = wr;
 510                         break;
 511                 }
 512
 513                 prev_wqe  = srq->last;
 514                 srq->last = wqe;
 515
 516                 ((struct mthca_next_seg *) wqe)->nda_op = 0;
 517                 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
 518                 /* flags field will always remain 0 */
 519
 520                 wqe += sizeof (struct mthca_next_seg);
 521
 522                 if (unlikely(wr->num_sge > srq->max_gs)) {
 523                         err = -EINVAL;
 524                         *bad_wr = wr;
 525                         srq->last = prev_wqe;
 526                         break;
 527                 }
 528
 529                 for (i = 0; i < wr->num_sge; ++i) {
 530                         ((struct mthca_data_seg *) wqe)->byte_count =
 531                                 cpu_to_be32(wr->sg_list[i].length);
 532                         ((struct mthca_data_seg *) wqe)->lkey =
 533                                 cpu_to_be32(wr->sg_list[i].lkey);
 534                         ((struct mthca_data_seg *) wqe)->addr =
 535                                 cpu_to_be64(wr->sg_list[i].addr);
 536                         wqe += sizeof (struct mthca_data_seg);
 537                 }
 538
 539                 if (i < srq->max_gs) {
 540                         ((struct mthca_data_seg *) wqe)->byte_count = 0;
 541                         ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
 542                         ((struct mthca_data_seg *) wqe)->addr = 0;
 543                 }
 544
 545                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
 546                         cpu_to_be32((ind << srq->wqe_shift) | 1);
 547                 wmb();
 548                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
 549                         cpu_to_be32(MTHCA_NEXT_DBD);
 550
 551                 srq->wrid[ind]  = wr->wr_id;
 552                 srq->first_free = next_ind;
 553
 554                 ++nreq;
 555                 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
 556                         nreq = 0;
 557
 558                         doorbell[0] = cpu_to_be32(first_ind << srq->wqe_shift);
 559                         doorbell[1] = cpu_to_be32(srq->srqn << 8);
 560
 561                         /*
 562                          * Make sure that descriptors are written
 563                          * before doorbell is rung.
 564                          */
 565                         wmb();
 566
 567                         mthca_write64(doorbell,
 568                                       dev->kar + MTHCA_RECEIVE_DOORBELL,
 569                                       MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
 570
 571                         first_ind = srq->first_free;
 572                 }
 573         }
 574
 575         if (likely(nreq)) {
 576                 doorbell[0] = cpu_to_be32(first_ind << srq->wqe_shift);
 577                 doorbell[1] = cpu_to_be32((srq->srqn << 8) | nreq);
 578
 579                 /*
 580                  * Make sure that descriptors are written before
 581                  * doorbell is rung.
 582                  */
 583                 wmb();
 584
 585                 mthca_write64(doorbell,
 586                               dev->kar + MTHCA_RECEIVE_DOORBELL,
 587                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
 588         }
 589
 590         spin_unlock_irqrestore(&srq->lock, flags);
 591         return err;
 592 }
 593
 594 int mthca_arbel_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
 595                               struct ib_recv_wr **bad_wr)
 596 {
 597         struct mthca_dev *dev = to_mdev(ibsrq->device);
 598         struct mthca_srq *srq = to_msrq(ibsrq);
 599         unsigned long flags;
 600         int err = 0;
 601         int ind;
 602         int next_ind;
 603         int nreq;
 604         int i;
 605         void *wqe;
 606
 607         spin_lock_irqsave(&srq->lock, flags);
 608
 609         for (nreq = 0; wr; ++nreq, wr = wr->next) {
 610                 ind = srq->first_free;
 611
 612                 if (ind < 0) {
 613                         mthca_err(dev, "SRQ %06x full\n", srq->srqn);
 614                         err = -ENOMEM;
 615                         *bad_wr = wr;
 616                         break;
 617                 }
 618
 619                 wqe       = get_wqe(srq, ind);
 620                 next_ind  = *wqe_to_link(wqe);
 621
 622                 if (next_ind < 0) {
 623                         mthca_err(dev, "SRQ %06x full\n", srq->srqn);
 624                         err = -ENOMEM;
 625                         *bad_wr = wr;
 626                         break;
 627                 }
 628
 629                 ((struct mthca_next_seg *) wqe)->nda_op =
 630                         cpu_to_be32((next_ind << srq->wqe_shift) | 1);
 631                 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
 632                 /* flags field will always remain 0 */
 633
 634                 wqe += sizeof (struct mthca_next_seg);
 635
 636                 if (unlikely(wr->num_sge > srq->max_gs)) {
 637                         err = -EINVAL;
 638                         *bad_wr = wr;
 639                         break;
 640                 }
 641
 642                 for (i = 0; i < wr->num_sge; ++i) {
 643                         ((struct mthca_data_seg *) wqe)->byte_count =
 644                                 cpu_to_be32(wr->sg_list[i].length);
 645                         ((struct mthca_data_seg *) wqe)->lkey =
 646                                 cpu_to_be32(wr->sg_list[i].lkey);
 647                         ((struct mthca_data_seg *) wqe)->addr =
 648                                 cpu_to_be64(wr->sg_list[i].addr);
 649                         wqe += sizeof (struct mthca_data_seg);
 650                 }
 651
 652                 if (i < srq->max_gs) {
 653                         ((struct mthca_data_seg *) wqe)->byte_count = 0;
 654                         ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
 655                         ((struct mthca_data_seg *) wqe)->addr = 0;
 656                 }
 657
 658                 srq->wrid[ind]  = wr->wr_id;
 659                 srq->first_free = next_ind;
 660         }
 661
 662         if (likely(nreq)) {
 663                 srq->counter += nreq;
 664
 665                 /*
 666                  * Make sure that descriptors are written before
 667                  * we write doorbell record.
 668                  */
 669                 wmb();
 670                 *srq->db = cpu_to_be32(srq->counter);
 671         }
 672
 673         spin_unlock_irqrestore(&srq->lock, flags);
 674         return err;
 675 }
 676
 677 int mthca_max_srq_sge(struct mthca_dev *dev)
 678 {
 679         if (mthca_is_memfree(dev))
 680                 return dev->limits.max_sg;
 681
 682         /*
 683          * SRQ allocations are based on powers of 2 for Tavor,
 684          * (although they only need to be multiples of 16 bytes).
 685          *
 686          * Therefore, we need to base the max number of sg entries on
 687          * the largest power of 2 descriptor size that is <= to the
 688          * actual max WQE descriptor size, rather than return the
 689          * max_sg value given by the firmware (which is based on WQE
 690          * sizes as multiples of 16, not powers of 2).
 691          *
 692          * If SRQ implementation is changed for Tavor to be based on
 693          * multiples of 16, the calculation below can be deleted and
 694          * the FW max_sg value returned.
 695          */
 696         return min_t(int, dev->limits.max_sg,
 697                      ((1 << (fls(dev->limits.max_desc_sz) - 1)) -
 698                       sizeof (struct mthca_next_seg)) /
 699                      sizeof (struct mthca_data_seg));
 700 }
 701
 702 int __devinit mthca_init_srq_table(struct mthca_dev *dev)
 703 {
 704         int err;
 705
 706         if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
 707                 return 0;
 708
 709         spin_lock_init(&dev->srq_table.lock);
 710
 711         err = mthca_alloc_init(&dev->srq_table.alloc,
 712                                dev->limits.num_srqs,
 713                                dev->limits.num_srqs - 1,
 714                                dev->limits.reserved_srqs);
 715         if (err)
 716                 return err;
 717
 718         err = mthca_array_init(&dev->srq_table.srq,
 719                                dev->limits.num_srqs);
 720         if (err)
 721                 mthca_alloc_cleanup(&dev->srq_table.alloc);
 722
 723         return err;
 724 }
 725
 726 void mthca_cleanup_srq_table(struct mthca_dev *dev)
 727 {
 728         if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
 729                 return;
 730
 731         mthca_array_cleanup(&dev->srq_table.srq, dev->limits.num_srqs);
 732         mthca_alloc_cleanup(&dev->srq_table.alloc);
 733 }