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

   1 /*
   2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Cisco Systems.  All rights reserved.
   4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  *
  34  * $Id$
  35  */
  36
  37 #include <linux/mm.h>
  38 #include <linux/scatterlist.h>
  39 #include <linux/sched.h>
  40
  41 #include <asm/page.h>
  42
  43 #include "mthca_memfree.h"
  44 #include "mthca_dev.h"
  45 #include "mthca_cmd.h"
  46
  47 /*
  48  * We allocate in as big chunks as we can, up to a maximum of 256 KB
  49  * per chunk.
  50  */
  51 enum {
  52         MTHCA_ICM_ALLOC_SIZE   = 1 << 18,
  53         MTHCA_TABLE_CHUNK_SIZE = 1 << 18
  54 };
  55
  56 struct mthca_user_db_table {
  57         struct mutex mutex;
  58         struct {
  59                 u64                uvirt;
  60                 struct scatterlist mem;
  61                 int                refcount;
  62         }                page[0];
  63 };
  64
  65 static void mthca_free_icm_pages(struct mthca_dev *dev, struct mthca_icm_chunk *chunk)
  66 {
  67         int i;
  68
  69         if (chunk->nsg > 0)
  70                 pci_unmap_sg(dev->pdev, chunk->mem, chunk->npages,
  71                              PCI_DMA_BIDIRECTIONAL);
  72
  73         for (i = 0; i < chunk->npages; ++i)
  74                 __free_pages(chunk->mem[i].page,
  75                              get_order(chunk->mem[i].length));
  76 }
  77
  78 static void mthca_free_icm_coherent(struct mthca_dev *dev, struct mthca_icm_chunk *chunk)
  79 {
  80         int i;
  81
  82         for (i = 0; i < chunk->npages; ++i) {
  83                 dma_free_coherent(&dev->pdev->dev, chunk->mem[i].length,
  84                                   lowmem_page_address(chunk->mem[i].page),
  85                                   sg_dma_address(&chunk->mem[i]));
  86         }
  87 }
  88
  89 void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm, int coherent)
  90 {
  91         struct mthca_icm_chunk *chunk, *tmp;
  92
  93         if (!icm)
  94                 return;
  95
  96         list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) {
  97                 if (coherent)
  98                         mthca_free_icm_coherent(dev, chunk);
  99                 else
 100                         mthca_free_icm_pages(dev, chunk);
 101
 102                 kfree(chunk);
 103         }
 104
 105         kfree(icm);
 106 }
 107
 108 static int mthca_alloc_icm_pages(struct scatterlist *mem, int order, gfp_t gfp_mask)
 109 {
 110         mem->page = alloc_pages(gfp_mask, order);
 111         if (!mem->page)
 112                 return -ENOMEM;
 113
 114         mem->length = PAGE_SIZE << order;
 115         mem->offset = 0;
 116         return 0;
 117 }
 118
 119 static int mthca_alloc_icm_coherent(struct device *dev, struct scatterlist *mem,
 120                                     int order, gfp_t gfp_mask)
 121 {
 122         void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order, &sg_dma_address(mem),
 123                                        gfp_mask);
 124         if (!buf)
 125                 return -ENOMEM;
 126
 127         sg_set_buf(mem, buf, PAGE_SIZE << order);
 128         BUG_ON(mem->offset);
 129         sg_dma_len(mem) = PAGE_SIZE << order;
 130         return 0;
 131 }
 132
 133 struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages,
 134                                   gfp_t gfp_mask, int coherent)
 135 {
 136         struct mthca_icm *icm;
 137         struct mthca_icm_chunk *chunk = NULL;
 138         int cur_order;
 139         int ret;
 140
 141         /* We use sg_set_buf for coherent allocs, which assumes low memory */
 142         BUG_ON(coherent && (gfp_mask & __GFP_HIGHMEM));
 143
 144         icm = kmalloc(sizeof *icm, gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
 145         if (!icm)
 146                 return icm;
 147
 148         icm->refcount = 0;
 149         INIT_LIST_HEAD(&icm->chunk_list);
 150
 151         cur_order = get_order(MTHCA_ICM_ALLOC_SIZE);
 152
 153         while (npages > 0) {
 154                 if (!chunk) {
 155                         chunk = kmalloc(sizeof *chunk,
 156                                         gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
 157                         if (!chunk)
 158                                 goto fail;
 159
 160                         chunk->npages = 0;
 161                         chunk->nsg    = 0;
 162                         list_add_tail(&chunk->list, &icm->chunk_list);
 163                 }
 164
 165                 while (1 << cur_order > npages)
 166                         --cur_order;
 167
 168                 if (coherent)
 169                         ret = mthca_alloc_icm_coherent(&dev->pdev->dev,
 170                                                        &chunk->mem[chunk->npages],
 171                                                        cur_order, gfp_mask);
 172                 else
 173                         ret = mthca_alloc_icm_pages(&chunk->mem[chunk->npages],
 174                                                     cur_order, gfp_mask);
 175
 176                 if (!ret) {
 177                         ++chunk->npages;
 178
 179                         if (coherent)
 180                                 ++chunk->nsg;
 181                         else if (chunk->npages == MTHCA_ICM_CHUNK_LEN) {
 182                                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 183                                                         chunk->npages,
 184                                                         PCI_DMA_BIDIRECTIONAL);
 185
 186                                 if (chunk->nsg <= 0)
 187                                         goto fail;
 188                         }
 189
 190                         if (chunk->npages == MTHCA_ICM_CHUNK_LEN)
 191                                 chunk = NULL;
 192
 193                         npages -= 1 << cur_order;
 194                 } else {
 195                         --cur_order;
 196                         if (cur_order < 0)
 197                                 goto fail;
 198                 }
 199         }
 200
 201         if (!coherent && chunk) {
 202                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 203                                         chunk->npages,
 204                                         PCI_DMA_BIDIRECTIONAL);
 205
 206                 if (chunk->nsg <= 0)
 207                         goto fail;
 208         }
 209
 210         return icm;
 211
 212 fail:
 213         mthca_free_icm(dev, icm, coherent);
 214         return NULL;
 215 }
 216
 217 int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
 218 {
 219         int i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
 220         int ret = 0;
 221         u8 status;
 222
 223         mutex_lock(&table->mutex);
 224
 225         if (table->icm[i]) {
 226                 ++table->icm[i]->refcount;
 227                 goto out;
 228         }
 229
 230         table->icm[i] = mthca_alloc_icm(dev, MTHCA_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
 231                                         (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
 232                                         __GFP_NOWARN, table->coherent);
 233         if (!table->icm[i]) {
 234                 ret = -ENOMEM;
 235                 goto out;
 236         }
 237
 238         if (mthca_MAP_ICM(dev, table->icm[i], table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
 239                           &status) || status) {
 240                 mthca_free_icm(dev, table->icm[i], table->coherent);
 241                 table->icm[i] = NULL;
 242                 ret = -ENOMEM;
 243                 goto out;
 244         }
 245
 246         ++table->icm[i]->refcount;
 247
 248 out:
 249         mutex_unlock(&table->mutex);
 250         return ret;
 251 }
 252
 253 void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
 254 {
 255         int i;
 256         u8 status;
 257
 258         if (!mthca_is_memfree(dev))
 259                 return;
 260
 261         i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
 262
 263         mutex_lock(&table->mutex);
 264
 265         if (--table->icm[i]->refcount == 0) {
 266                 mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
 267                                 MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
 268                                 &status);
 269                 mthca_free_icm(dev, table->icm[i], table->coherent);
 270                 table->icm[i] = NULL;
 271         }
 272
 273         mutex_unlock(&table->mutex);
 274 }
 275
 276 void *mthca_table_find(struct mthca_icm_table *table, int obj, dma_addr_t *dma_handle)
 277 {
 278         int idx, offset, dma_offset, i;
 279         struct mthca_icm_chunk *chunk;
 280         struct mthca_icm *icm;
 281         struct page *page = NULL;
 282
 283         if (!table->lowmem)
 284                 return NULL;
 285
 286         mutex_lock(&table->mutex);
 287
 288         idx = (obj & (table->num_obj - 1)) * table->obj_size;
 289         icm = table->icm[idx / MTHCA_TABLE_CHUNK_SIZE];
 290         dma_offset = offset = idx % MTHCA_TABLE_CHUNK_SIZE;
 291
 292         if (!icm)
 293                 goto out;
 294
 295         list_for_each_entry(chunk, &icm->chunk_list, list) {
 296                 for (i = 0; i < chunk->npages; ++i) {
 297                         if (dma_handle && dma_offset >= 0) {
 298                                 if (sg_dma_len(&chunk->mem[i]) > dma_offset)
 299                                         *dma_handle = sg_dma_address(&chunk->mem[i]) +
 300                                                 dma_offset;
 301                                 dma_offset -= sg_dma_len(&chunk->mem[i]);
 302                         }
 303                         /* DMA mapping can merge pages but not split them,
 304                          * so if we found the page, dma_handle has already
 305                          * been assigned to. */
 306                         if (chunk->mem[i].length > offset) {
 307                                 page = chunk->mem[i].page;
 308                                 goto out;
 309                         }
 310                         offset -= chunk->mem[i].length;
 311                 }
 312         }
 313
 314 out:
 315         mutex_unlock(&table->mutex);
 316         return page ? lowmem_page_address(page) + offset : NULL;
 317 }
 318
 319 int mthca_table_get_range(struct mthca_dev *dev, struct mthca_icm_table *table,
 320                           int start, int end)
 321 {
 322         int inc = MTHCA_TABLE_CHUNK_SIZE / table->obj_size;
 323         int i, err;
 324
 325         for (i = start; i <= end; i += inc) {
 326                 err = mthca_table_get(dev, table, i);
 327                 if (err)
 328                         goto fail;
 329         }
 330
 331         return 0;
 332
 333 fail:
 334         while (i > start) {
 335                 i -= inc;
 336                 mthca_table_put(dev, table, i);
 337         }
 338
 339         return err;
 340 }
 341
 342 void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table,
 343                            int start, int end)
 344 {
 345         int i;
 346
 347         if (!mthca_is_memfree(dev))
 348                 return;
 349
 350         for (i = start; i <= end; i += MTHCA_TABLE_CHUNK_SIZE / table->obj_size)
 351                 mthca_table_put(dev, table, i);
 352 }
 353
 354 struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
 355                                               u64 virt, int obj_size,
 356                                               int nobj, int reserved,
 357                                               int use_lowmem, int use_coherent)
 358 {
 359         struct mthca_icm_table *table;
 360         int num_icm;
 361         unsigned chunk_size;
 362         int i;
 363         u8 status;
 364
 365         num_icm = (obj_size * nobj + MTHCA_TABLE_CHUNK_SIZE - 1) / MTHCA_TABLE_CHUNK_SIZE;
 366
 367         table = kmalloc(sizeof *table + num_icm * sizeof *table->icm, GFP_KERNEL);
 368         if (!table)
 369                 return NULL;
 370
 371         table->virt     = virt;
 372         table->num_icm  = num_icm;
 373         table->num_obj  = nobj;
 374         table->obj_size = obj_size;
 375         table->lowmem   = use_lowmem;
 376         table->coherent = use_coherent;
 377         mutex_init(&table->mutex);
 378
 379         for (i = 0; i < num_icm; ++i)
 380                 table->icm[i] = NULL;
 381
 382         for (i = 0; i * MTHCA_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
 383                 chunk_size = MTHCA_TABLE_CHUNK_SIZE;
 384                 if ((i + 1) * MTHCA_TABLE_CHUNK_SIZE > nobj * obj_size)
 385                         chunk_size = nobj * obj_size - i * MTHCA_TABLE_CHUNK_SIZE;
 386
 387                 table->icm[i] = mthca_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
 388                                                 (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
 389                                                 __GFP_NOWARN, use_coherent);
 390                 if (!table->icm[i])
 391                         goto err;
 392                 if (mthca_MAP_ICM(dev, table->icm[i], virt + i * MTHCA_TABLE_CHUNK_SIZE,
 393                                   &status) || status) {
 394                         mthca_free_icm(dev, table->icm[i], table->coherent);
 395                         table->icm[i] = NULL;
 396                         goto err;
 397                 }
 398
 399                 /*
 400                  * Add a reference to this ICM chunk so that it never
 401                  * gets freed (since it contains reserved firmware objects).
 402                  */
 403                 ++table->icm[i]->refcount;
 404         }
 405
 406         return table;
 407
 408 err:
 409         for (i = 0; i < num_icm; ++i)
 410                 if (table->icm[i]) {
 411                         mthca_UNMAP_ICM(dev, virt + i * MTHCA_TABLE_CHUNK_SIZE,
 412                                         MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
 413                                         &status);
 414                         mthca_free_icm(dev, table->icm[i], table->coherent);
 415                 }
 416
 417         kfree(table);
 418
 419         return NULL;
 420 }
 421
 422 void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table)
 423 {
 424         int i;
 425         u8 status;
 426
 427         for (i = 0; i < table->num_icm; ++i)
 428                 if (table->icm[i]) {
 429                         mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
 430                                         MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
 431                                         &status);
 432                         mthca_free_icm(dev, table->icm[i], table->coherent);
 433                 }
 434
 435         kfree(table);
 436 }
 437
 438 static u64 mthca_uarc_virt(struct mthca_dev *dev, struct mthca_uar *uar, int page)
 439 {
 440         return dev->uar_table.uarc_base +
 441                 uar->index * dev->uar_table.uarc_size +
 442                 page * MTHCA_ICM_PAGE_SIZE;
 443 }
 444
 445 int mthca_map_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
 446                       struct mthca_user_db_table *db_tab, int index, u64 uaddr)
 447 {
 448         int ret = 0;
 449         u8 status;
 450         int i;
 451
 452         if (!mthca_is_memfree(dev))
 453                 return 0;
 454
 455         if (index < 0 || index > dev->uar_table.uarc_size / 8)
 456                 return -EINVAL;
 457
 458         mutex_lock(&db_tab->mutex);
 459
 460         i = index / MTHCA_DB_REC_PER_PAGE;
 461
 462         if ((db_tab->page[i].refcount >= MTHCA_DB_REC_PER_PAGE)       ||
 463             (db_tab->page[i].uvirt && db_tab->page[i].uvirt != uaddr) ||
 464             (uaddr & 4095)) {
 465                 ret = -EINVAL;
 466                 goto out;
 467         }
 468
 469         if (db_tab->page[i].refcount) {
 470                 ++db_tab->page[i].refcount;
 471                 goto out;
 472         }
 473
 474         ret = get_user_pages(current, current->mm, uaddr & PAGE_MASK, 1, 1, 0,
 475                              &db_tab->page[i].mem.page, NULL);
 476         if (ret < 0)
 477                 goto out;
 478
 479         db_tab->page[i].mem.length = MTHCA_ICM_PAGE_SIZE;
 480         db_tab->page[i].mem.offset = uaddr & ~PAGE_MASK;
 481
 482         ret = pci_map_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
 483         if (ret < 0) {
 484                 put_page(db_tab->page[i].mem.page);
 485                 goto out;
 486         }
 487
 488         ret = mthca_MAP_ICM_page(dev, sg_dma_address(&db_tab->page[i].mem),
 489                                  mthca_uarc_virt(dev, uar, i), &status);
 490         if (!ret && status)
 491                 ret = -EINVAL;
 492         if (ret) {
 493                 pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
 494                 put_page(db_tab->page[i].mem.page);
 495                 goto out;
 496         }
 497
 498         db_tab->page[i].uvirt    = uaddr;
 499         db_tab->page[i].refcount = 1;
 500
 501 out:
 502         mutex_unlock(&db_tab->mutex);
 503         return ret;
 504 }
 505
 506 void mthca_unmap_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
 507                          struct mthca_user_db_table *db_tab, int index)
 508 {
 509         if (!mthca_is_memfree(dev))
 510                 return;
 511
 512         /*
 513          * To make our bookkeeping simpler, we don't unmap DB
 514          * pages until we clean up the whole db table.
 515          */
 516
 517         mutex_lock(&db_tab->mutex);
 518
 519         --db_tab->page[index / MTHCA_DB_REC_PER_PAGE].refcount;
 520
 521         mutex_unlock(&db_tab->mutex);
 522 }
 523
 524 struct mthca_user_db_table *mthca_init_user_db_tab(struct mthca_dev *dev)
 525 {
 526         struct mthca_user_db_table *db_tab;
 527         int npages;
 528         int i;
 529
 530         if (!mthca_is_memfree(dev))
 531                 return NULL;
 532
 533         npages = dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE;
 534         db_tab = kmalloc(sizeof *db_tab + npages * sizeof *db_tab->page, GFP_KERNEL);
 535         if (!db_tab)
 536                 return ERR_PTR(-ENOMEM);
 537
 538         mutex_init(&db_tab->mutex);
 539         for (i = 0; i < npages; ++i) {
 540                 db_tab->page[i].refcount = 0;
 541                 db_tab->page[i].uvirt    = 0;
 542         }
 543
 544         return db_tab;
 545 }
 546
 547 void mthca_cleanup_user_db_tab(struct mthca_dev *dev, struct mthca_uar *uar,
 548                                struct mthca_user_db_table *db_tab)
 549 {
 550         int i;
 551         u8 status;
 552
 553         if (!mthca_is_memfree(dev))
 554                 return;
 555
 556         for (i = 0; i < dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE; ++i) {
 557                 if (db_tab->page[i].uvirt) {
 558                         mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, uar, i), 1, &status);
 559                         pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
 560                         put_page(db_tab->page[i].mem.page);
 561                 }
 562         }
 563
 564         kfree(db_tab);
 565 }
 566
 567 int mthca_alloc_db(struct mthca_dev *dev, enum mthca_db_type type,
 568                    u32 qn, __be32 **db)
 569 {
 570         int group;
 571         int start, end, dir;
 572         int i, j;
 573         struct mthca_db_page *page;
 574         int ret = 0;
 575         u8 status;
 576
 577         mutex_lock(&dev->db_tab->mutex);
 578
 579         switch (type) {
 580         case MTHCA_DB_TYPE_CQ_ARM:
 581         case MTHCA_DB_TYPE_SQ:
 582                 group = 0;
 583                 start = 0;
 584                 end   = dev->db_tab->max_group1;
 585                 dir   = 1;
 586                 break;
 587
 588         case MTHCA_DB_TYPE_CQ_SET_CI:
 589         case MTHCA_DB_TYPE_RQ:
 590         case MTHCA_DB_TYPE_SRQ:
 591                 group = 1;
 592                 start = dev->db_tab->npages - 1;
 593                 end   = dev->db_tab->min_group2;
 594                 dir   = -1;
 595                 break;
 596
 597         default:
 598                 ret = -EINVAL;
 599                 goto out;
 600         }
 601
 602         for (i = start; i != end; i += dir)
 603                 if (dev->db_tab->page[i].db_rec &&
 604                     !bitmap_full(dev->db_tab->page[i].used,
 605                                  MTHCA_DB_REC_PER_PAGE)) {
 606                         page = dev->db_tab->page + i;
 607                         goto found;
 608                 }
 609
 610         for (i = start; i != end; i += dir)
 611                 if (!dev->db_tab->page[i].db_rec) {
 612                         page = dev->db_tab->page + i;
 613                         goto alloc;
 614                 }
 615
 616         if (dev->db_tab->max_group1 >= dev->db_tab->min_group2 - 1) {
 617                 ret = -ENOMEM;
 618                 goto out;
 619         }
 620
 621         if (group == 0)
 622                 ++dev->db_tab->max_group1;
 623         else
 624                 --dev->db_tab->min_group2;
 625
 626         page = dev->db_tab->page + end;
 627
 628 alloc:
 629         page->db_rec = dma_alloc_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
 630                                           &page->mapping, GFP_KERNEL);
 631         if (!page->db_rec) {
 632                 ret = -ENOMEM;
 633                 goto out;
 634         }
 635         memset(page->db_rec, 0, MTHCA_ICM_PAGE_SIZE);
 636
 637         ret = mthca_MAP_ICM_page(dev, page->mapping,
 638                                  mthca_uarc_virt(dev, &dev->driver_uar, i), &status);
 639         if (!ret && status)
 640                 ret = -EINVAL;
 641         if (ret) {
 642                 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
 643                                   page->db_rec, page->mapping);
 644                 goto out;
 645         }
 646
 647         bitmap_zero(page->used, MTHCA_DB_REC_PER_PAGE);
 648
 649 found:
 650         j = find_first_zero_bit(page->used, MTHCA_DB_REC_PER_PAGE);
 651         set_bit(j, page->used);
 652
 653         if (group == 1)
 654                 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
 655
 656         ret = i * MTHCA_DB_REC_PER_PAGE + j;
 657
 658         page->db_rec[j] = cpu_to_be64((qn << 8) | (type << 5));
 659
 660         *db = (__be32 *) &page->db_rec[j];
 661
 662 out:
 663         mutex_unlock(&dev->db_tab->mutex);
 664
 665         return ret;
 666 }
 667
 668 void mthca_free_db(struct mthca_dev *dev, int type, int db_index)
 669 {
 670         int i, j;
 671         struct mthca_db_page *page;
 672         u8 status;
 673
 674         i = db_index / MTHCA_DB_REC_PER_PAGE;
 675         j = db_index % MTHCA_DB_REC_PER_PAGE;
 676
 677         page = dev->db_tab->page + i;
 678
 679         mutex_lock(&dev->db_tab->mutex);
 680
 681         page->db_rec[j] = 0;
 682         if (i >= dev->db_tab->min_group2)
 683                 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
 684         clear_bit(j, page->used);
 685
 686         if (bitmap_empty(page->used, MTHCA_DB_REC_PER_PAGE) &&
 687             i >= dev->db_tab->max_group1 - 1) {
 688                 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
 689
 690                 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
 691                                   page->db_rec, page->mapping);
 692                 page->db_rec = NULL;
 693
 694                 if (i == dev->db_tab->max_group1) {
 695                         --dev->db_tab->max_group1;
 696                         /* XXX may be able to unmap more pages now */
 697                 }
 698                 if (i == dev->db_tab->min_group2)
 699                         ++dev->db_tab->min_group2;
 700         }
 701
 702         mutex_unlock(&dev->db_tab->mutex);
 703 }
 704
 705 int mthca_init_db_tab(struct mthca_dev *dev)
 706 {
 707         int i;
 708
 709         if (!mthca_is_memfree(dev))
 710                 return 0;
 711
 712         dev->db_tab = kmalloc(sizeof *dev->db_tab, GFP_KERNEL);
 713         if (!dev->db_tab)
 714                 return -ENOMEM;
 715
 716         mutex_init(&dev->db_tab->mutex);
 717
 718         dev->db_tab->npages     = dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE;
 719         dev->db_tab->max_group1 = 0;
 720         dev->db_tab->min_group2 = dev->db_tab->npages - 1;
 721
 722         dev->db_tab->page = kmalloc(dev->db_tab->npages *
 723                                     sizeof *dev->db_tab->page,
 724                                     GFP_KERNEL);
 725         if (!dev->db_tab->page) {
 726                 kfree(dev->db_tab);
 727                 return -ENOMEM;
 728         }
 729
 730         for (i = 0; i < dev->db_tab->npages; ++i)
 731                 dev->db_tab->page[i].db_rec = NULL;
 732
 733         return 0;
 734 }
 735
 736 void mthca_cleanup_db_tab(struct mthca_dev *dev)
 737 {
 738         int i;
 739         u8 status;
 740
 741         if (!mthca_is_memfree(dev))
 742                 return;
 743
 744         /*
 745          * Because we don't always free our UARC pages when they
 746          * become empty to make mthca_free_db() simpler we need to
 747          * make a sweep through the doorbell pages and free any
 748          * leftover pages now.
 749          */
 750         for (i = 0; i < dev->db_tab->npages; ++i) {
 751                 if (!dev->db_tab->page[i].db_rec)
 752                         continue;
 753
 754                 if (!bitmap_empty(dev->db_tab->page[i].used, MTHCA_DB_REC_PER_PAGE))
 755                         mthca_warn(dev, "Kernel UARC page %d not empty\n", i);
 756
 757                 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
 758
 759                 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
 760                                   dev->db_tab->page[i].db_rec,
 761                                   dev->db_tab->page[i].mapping);
 762         }
 763
 764         kfree(dev->db_tab->page);
 765         kfree(dev->db_tab);
 766 }