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
  35 #include <linux/mm.h>
  36 #include <linux/scatterlist.h>
  37 #include <linux/sched.h>
  38
  39 #include <asm/page.h>
  40
  41 #include "mthca_memfree.h"
  42 #include "mthca_dev.h"
  43 #include "mthca_cmd.h"
  44
  45 /*
  46  * We allocate in as big chunks as we can, up to a maximum of 256 KB
  47  * per chunk.
  48  */
  49 enum {
  50         MTHCA_ICM_ALLOC_SIZE   = 1 << 18,
  51         MTHCA_TABLE_CHUNK_SIZE = 1 << 18
  52 };
  53
  54 struct mthca_user_db_table {
  55         struct mutex mutex;
  56         struct {
  57                 u64                uvirt;
  58                 struct scatterlist mem;
  59                 int                refcount;
  60         }                page[0];
  61 };
  62
  63 static void mthca_free_icm_pages(struct mthca_dev *dev, struct mthca_icm_chunk *chunk)
  64 {
  65         int i;
  66
  67         if (chunk->nsg > 0)
  68                 pci_unmap_sg(dev->pdev, chunk->mem, chunk->npages,
  69                              PCI_DMA_BIDIRECTIONAL);
  70
  71         for (i = 0; i < chunk->npages; ++i)
  72                 __free_pages(sg_page(&chunk->mem[i]),
  73                              get_order(chunk->mem[i].length));
  74 }
  75
  76 static void mthca_free_icm_coherent(struct mthca_dev *dev, struct mthca_icm_chunk *chunk)
  77 {
  78         int i;
  79
  80         for (i = 0; i < chunk->npages; ++i) {
  81                 dma_free_coherent(&dev->pdev->dev, chunk->mem[i].length,
  82                                   lowmem_page_address(sg_page(&chunk->mem[i])),
  83                                   sg_dma_address(&chunk->mem[i]));
  84         }
  85 }
  86
  87 void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm, int coherent)
  88 {
  89         struct mthca_icm_chunk *chunk, *tmp;
  90
  91         if (!icm)
  92                 return;
  93
  94         list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) {
  95                 if (coherent)
  96                         mthca_free_icm_coherent(dev, chunk);
  97                 else
  98                         mthca_free_icm_pages(dev, chunk);
  99
 100                 kfree(chunk);
 101         }
 102
 103         kfree(icm);
 104 }
 105
 106 static int mthca_alloc_icm_pages(struct scatterlist *mem, int order, gfp_t gfp_mask)
 107 {
 108         struct page *page;
 109
 110         /*
 111          * Use __GFP_ZERO because buggy firmware assumes ICM pages are
 112          * cleared, and subtle failures are seen if they aren't.
 113          */
 114         page = alloc_pages(gfp_mask | __GFP_ZERO, order);
 115         if (!page)
 116                 return -ENOMEM;
 117
 118         sg_set_page(mem, page, PAGE_SIZE << order, 0);
 119         return 0;
 120 }
 121
 122 static int mthca_alloc_icm_coherent(struct device *dev, struct scatterlist *mem,
 123                                     int order, gfp_t gfp_mask)
 124 {
 125         void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order, &sg_dma_address(mem),
 126                                        gfp_mask);
 127         if (!buf)
 128                 return -ENOMEM;
 129
 130         sg_set_buf(mem, buf, PAGE_SIZE << order);
 131         BUG_ON(mem->offset);
 132         sg_dma_len(mem) = PAGE_SIZE << order;
 133         return 0;
 134 }
 135
 136 struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages,
 137                                   gfp_t gfp_mask, int coherent)
 138 {
 139         struct mthca_icm *icm;
 140         struct mthca_icm_chunk *chunk = NULL;
 141         int cur_order;
 142         int ret;
 143
 144         /* We use sg_set_buf for coherent allocs, which assumes low memory */
 145         BUG_ON(coherent && (gfp_mask & __GFP_HIGHMEM));
 146
 147         icm = kmalloc(sizeof *icm, gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
 148         if (!icm)
 149                 return icm;
 150
 151         icm->refcount = 0;
 152         INIT_LIST_HEAD(&icm->chunk_list);
 153
 154         cur_order = get_order(MTHCA_ICM_ALLOC_SIZE);
 155
 156         while (npages > 0) {
 157                 if (!chunk) {
 158                         chunk = kmalloc(sizeof *chunk,
 159                                         gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
 160                         if (!chunk)
 161                                 goto fail;
 162
 163                         sg_init_table(chunk->mem, MTHCA_ICM_CHUNK_LEN);
 164                         chunk->npages = 0;
 165                         chunk->nsg    = 0;
 166                         list_add_tail(&chunk->list, &icm->chunk_list);
 167                 }
 168
 169                 while (1 << cur_order > npages)
 170                         --cur_order;
 171
 172                 if (coherent)
 173                         ret = mthca_alloc_icm_coherent(&dev->pdev->dev,
 174                                                        &chunk->mem[chunk->npages],
 175                                                        cur_order, gfp_mask);
 176                 else
 177                         ret = mthca_alloc_icm_pages(&chunk->mem[chunk->npages],
 178                                                     cur_order, gfp_mask);
 179
 180                 if (!ret) {
 181                         ++chunk->npages;
 182
 183                         if (coherent)
 184                                 ++chunk->nsg;
 185                         else if (chunk->npages == MTHCA_ICM_CHUNK_LEN) {
 186                                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 187                                                         chunk->npages,
 188                                                         PCI_DMA_BIDIRECTIONAL);
 189
 190                                 if (chunk->nsg <= 0)
 191                                         goto fail;
 192                         }
 193
 194                         if (chunk->npages == MTHCA_ICM_CHUNK_LEN)
 195                                 chunk = NULL;
 196
 197                         npages -= 1 << cur_order;
 198                 } else {
 199                         --cur_order;
 200                         if (cur_order < 0)
 201                                 goto fail;
 202                 }
 203         }
 204
 205         if (!coherent && chunk) {
 206                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 207                                         chunk->npages,
 208                                         PCI_DMA_BIDIRECTIONAL);
 209
 210                 if (chunk->nsg <= 0)
 211                         goto fail;
 212         }
 213
 214         return icm;
 215
 216 fail:
 217         mthca_free_icm(dev, icm, coherent);
 218         return NULL;
 219 }
 220
 221 int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
 222 {
 223         int i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
 224         int ret = 0;
 225         u8 status;
 226
 227         mutex_lock(&table->mutex);
 228
 229         if (table->icm[i]) {
 230                 ++table->icm[i]->refcount;
 231                 goto out;
 232         }
 233
 234         table->icm[i] = mthca_alloc_icm(dev, MTHCA_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
 235                                         (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
 236                                         __GFP_NOWARN, table->coherent);
 237         if (!table->icm[i]) {
 238                 ret = -ENOMEM;
 239                 goto out;
 240         }
 241
 242         if (mthca_MAP_ICM(dev, table->icm[i], table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
 243                           &status) || status) {
 244                 mthca_free_icm(dev, table->icm[i], table->coherent);
 245                 table->icm[i] = NULL;
 246                 ret = -ENOMEM;
 247                 goto out;
 248         }
 249
 250         ++table->icm[i]->refcount;
 251
 252 out:
 253         mutex_unlock(&table->mutex);
 254         return ret;
 255 }
 256
 257 void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
 258 {
 259         int i;
 260         u8 status;
 261
 262         if (!mthca_is_memfree(dev))
 263                 return;
 264
 265         i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
 266
 267         mutex_lock(&table->mutex);
 268
 269         if (--table->icm[i]->refcount == 0) {
 270                 mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
 271                                 MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
 272                                 &status);
 273                 mthca_free_icm(dev, table->icm[i], table->coherent);
 274                 table->icm[i] = NULL;
 275         }
 276
 277         mutex_unlock(&table->mutex);
 278 }
 279
 280 void *mthca_table_find(struct mthca_icm_table *table, int obj, dma_addr_t *dma_handle)
 281 {
 282         int idx, offset, dma_offset, i;
 283         struct mthca_icm_chunk *chunk;
 284         struct mthca_icm *icm;
 285         struct page *page = NULL;
 286
 287         if (!table->lowmem)
 288                 return NULL;
 289
 290         mutex_lock(&table->mutex);
 291
 292         idx = (obj & (table->num_obj - 1)) * table->obj_size;
 293         icm = table->icm[idx / MTHCA_TABLE_CHUNK_SIZE];
 294         dma_offset = offset = idx % MTHCA_TABLE_CHUNK_SIZE;
 295
 296         if (!icm)
 297                 goto out;
 298
 299         list_for_each_entry(chunk, &icm->chunk_list, list) {
 300                 for (i = 0; i < chunk->npages; ++i) {
 301                         if (dma_handle && dma_offset >= 0) {
 302                                 if (sg_dma_len(&chunk->mem[i]) > dma_offset)
 303                                         *dma_handle = sg_dma_address(&chunk->mem[i]) +
 304                                                 dma_offset;
 305                                 dma_offset -= sg_dma_len(&chunk->mem[i]);
 306                         }
 307                         /* DMA mapping can merge pages but not split them,
 308                          * so if we found the page, dma_handle has already
 309                          * been assigned to. */
 310                         if (chunk->mem[i].length > offset) {
 311                                 page = sg_page(&chunk->mem[i]);
 312                                 goto out;
 313                         }
 314                         offset -= chunk->mem[i].length;
 315                 }
 316         }
 317
 318 out:
 319         mutex_unlock(&table->mutex);
 320         return page ? lowmem_page_address(page) + offset : NULL;
 321 }
 322
 323 int mthca_table_get_range(struct mthca_dev *dev, struct mthca_icm_table *table,
 324                           int start, int end)
 325 {
 326         int inc = MTHCA_TABLE_CHUNK_SIZE / table->obj_size;
 327         int i, err;
 328
 329         for (i = start; i <= end; i += inc) {
 330                 err = mthca_table_get(dev, table, i);
 331                 if (err)
 332                         goto fail;
 333         }
 334
 335         return 0;
 336
 337 fail:
 338         while (i > start) {
 339                 i -= inc;
 340                 mthca_table_put(dev, table, i);
 341         }
 342
 343         return err;
 344 }
 345
 346 void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table,
 347                            int start, int end)
 348 {
 349         int i;
 350
 351         if (!mthca_is_memfree(dev))
 352                 return;
 353
 354         for (i = start; i <= end; i += MTHCA_TABLE_CHUNK_SIZE / table->obj_size)
 355                 mthca_table_put(dev, table, i);
 356 }
 357
 358 struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
 359                                               u64 virt, int obj_size,
 360                                               int nobj, int reserved,
 361                                               int use_lowmem, int use_coherent)
 362 {
 363         struct mthca_icm_table *table;
 364         int obj_per_chunk;
 365         int num_icm;
 366         unsigned chunk_size;
 367         int i;
 368         u8 status;
 369
 370         obj_per_chunk = MTHCA_TABLE_CHUNK_SIZE / obj_size;
 371         num_icm = DIV_ROUND_UP(nobj, obj_per_chunk);
 372
 373         table = kmalloc(sizeof *table + num_icm * sizeof *table->icm, GFP_KERNEL);
 374         if (!table)
 375                 return NULL;
 376
 377         table->virt     = virt;
 378         table->num_icm  = num_icm;
 379         table->num_obj  = nobj;
 380         table->obj_size = obj_size;
 381         table->lowmem   = use_lowmem;
 382         table->coherent = use_coherent;
 383         mutex_init(&table->mutex);
 384
 385         for (i = 0; i < num_icm; ++i)
 386                 table->icm[i] = NULL;
 387
 388         for (i = 0; i * MTHCA_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
 389                 chunk_size = MTHCA_TABLE_CHUNK_SIZE;
 390                 if ((i + 1) * MTHCA_TABLE_CHUNK_SIZE > nobj * obj_size)
 391                         chunk_size = nobj * obj_size - i * MTHCA_TABLE_CHUNK_SIZE;
 392
 393                 table->icm[i] = mthca_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
 394                                                 (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
 395                                                 __GFP_NOWARN, use_coherent);
 396                 if (!table->icm[i])
 397                         goto err;
 398                 if (mthca_MAP_ICM(dev, table->icm[i], virt + i * MTHCA_TABLE_CHUNK_SIZE,
 399                                   &status) || status) {
 400                         mthca_free_icm(dev, table->icm[i], table->coherent);
 401                         table->icm[i] = NULL;
 402                         goto err;
 403                 }
 404
 405                 /*
 406                  * Add a reference to this ICM chunk so that it never
 407                  * gets freed (since it contains reserved firmware objects).
 408                  */
 409                 ++table->icm[i]->refcount;
 410         }
 411
 412         return table;
 413
 414 err:
 415         for (i = 0; i < num_icm; ++i)
 416                 if (table->icm[i]) {
 417                         mthca_UNMAP_ICM(dev, virt + i * MTHCA_TABLE_CHUNK_SIZE,
 418                                         MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
 419                                         &status);
 420                         mthca_free_icm(dev, table->icm[i], table->coherent);
 421                 }
 422
 423         kfree(table);
 424
 425         return NULL;
 426 }
 427
 428 void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table)
 429 {
 430         int i;
 431         u8 status;
 432
 433         for (i = 0; i < table->num_icm; ++i)
 434                 if (table->icm[i]) {
 435                         mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
 436                                         MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
 437                                         &status);
 438                         mthca_free_icm(dev, table->icm[i], table->coherent);
 439                 }
 440
 441         kfree(table);
 442 }
 443
 444 static u64 mthca_uarc_virt(struct mthca_dev *dev, struct mthca_uar *uar, int page)
 445 {
 446         return dev->uar_table.uarc_base +
 447                 uar->index * dev->uar_table.uarc_size +
 448                 page * MTHCA_ICM_PAGE_SIZE;
 449 }
 450
 451 int mthca_map_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
 452                       struct mthca_user_db_table *db_tab, int index, u64 uaddr)
 453 {
 454         struct page *pages[1];
 455         int ret = 0;
 456         u8 status;
 457         int i;
 458
 459         if (!mthca_is_memfree(dev))
 460                 return 0;
 461
 462         if (index < 0 || index > dev->uar_table.uarc_size / 8)
 463                 return -EINVAL;
 464
 465         mutex_lock(&db_tab->mutex);
 466
 467         i = index / MTHCA_DB_REC_PER_PAGE;
 468
 469         if ((db_tab->page[i].refcount >= MTHCA_DB_REC_PER_PAGE)       ||
 470             (db_tab->page[i].uvirt && db_tab->page[i].uvirt != uaddr) ||
 471             (uaddr & 4095)) {
 472                 ret = -EINVAL;
 473                 goto out;
 474         }
 475
 476         if (db_tab->page[i].refcount) {
 477                 ++db_tab->page[i].refcount;
 478                 goto out;
 479         }
 480
 481         ret = get_user_pages(current, current->mm, uaddr & PAGE_MASK, 1, 1, 0,
 482                              pages, NULL);
 483         if (ret < 0)
 484                 goto out;
 485
 486         sg_set_page(&db_tab->page[i].mem, pages[0], MTHCA_ICM_PAGE_SIZE,
 487                         uaddr & ~PAGE_MASK);
 488
 489         ret = pci_map_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
 490         if (ret < 0) {
 491                 put_page(pages[0]);
 492                 goto out;
 493         }
 494
 495         ret = mthca_MAP_ICM_page(dev, sg_dma_address(&db_tab->page[i].mem),
 496                                  mthca_uarc_virt(dev, uar, i), &status);
 497         if (!ret && status)
 498                 ret = -EINVAL;
 499         if (ret) {
 500                 pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
 501                 put_page(sg_page(&db_tab->page[i].mem));
 502                 goto out;
 503         }
 504
 505         db_tab->page[i].uvirt    = uaddr;
 506         db_tab->page[i].refcount = 1;
 507
 508 out:
 509         mutex_unlock(&db_tab->mutex);
 510         return ret;
 511 }
 512
 513 void mthca_unmap_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
 514                          struct mthca_user_db_table *db_tab, int index)
 515 {
 516         if (!mthca_is_memfree(dev))
 517                 return;
 518
 519         /*
 520          * To make our bookkeeping simpler, we don't unmap DB
 521          * pages until we clean up the whole db table.
 522          */
 523
 524         mutex_lock(&db_tab->mutex);
 525
 526         --db_tab->page[index / MTHCA_DB_REC_PER_PAGE].refcount;
 527
 528         mutex_unlock(&db_tab->mutex);
 529 }
 530
 531 struct mthca_user_db_table *mthca_init_user_db_tab(struct mthca_dev *dev)
 532 {
 533         struct mthca_user_db_table *db_tab;
 534         int npages;
 535         int i;
 536
 537         if (!mthca_is_memfree(dev))
 538                 return NULL;
 539
 540         npages = dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE;
 541         db_tab = kmalloc(sizeof *db_tab + npages * sizeof *db_tab->page, GFP_KERNEL);
 542         if (!db_tab)
 543                 return ERR_PTR(-ENOMEM);
 544
 545         mutex_init(&db_tab->mutex);
 546         for (i = 0; i < npages; ++i) {
 547                 db_tab->page[i].refcount = 0;
 548                 db_tab->page[i].uvirt    = 0;
 549                 sg_init_table(&db_tab->page[i].mem, 1);
 550         }
 551
 552         return db_tab;
 553 }
 554
 555 void mthca_cleanup_user_db_tab(struct mthca_dev *dev, struct mthca_uar *uar,
 556                                struct mthca_user_db_table *db_tab)
 557 {
 558         int i;
 559         u8 status;
 560
 561         if (!mthca_is_memfree(dev))
 562                 return;
 563
 564         for (i = 0; i < dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE; ++i) {
 565                 if (db_tab->page[i].uvirt) {
 566                         mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, uar, i), 1, &status);
 567                         pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
 568                         put_page(sg_page(&db_tab->page[i].mem));
 569                 }
 570         }
 571
 572         kfree(db_tab);
 573 }
 574
 575 int mthca_alloc_db(struct mthca_dev *dev, enum mthca_db_type type,
 576                    u32 qn, __be32 **db)
 577 {
 578         int group;
 579         int start, end, dir;
 580         int i, j;
 581         struct mthca_db_page *page;
 582         int ret = 0;
 583         u8 status;
 584
 585         mutex_lock(&dev->db_tab->mutex);
 586
 587         switch (type) {
 588         case MTHCA_DB_TYPE_CQ_ARM:
 589         case MTHCA_DB_TYPE_SQ:
 590                 group = 0;
 591                 start = 0;
 592                 end   = dev->db_tab->max_group1;
 593                 dir   = 1;
 594                 break;
 595
 596         case MTHCA_DB_TYPE_CQ_SET_CI:
 597         case MTHCA_DB_TYPE_RQ:
 598         case MTHCA_DB_TYPE_SRQ:
 599                 group = 1;
 600                 start = dev->db_tab->npages - 1;
 601                 end   = dev->db_tab->min_group2;
 602                 dir   = -1;
 603                 break;
 604
 605         default:
 606                 ret = -EINVAL;
 607                 goto out;
 608         }
 609
 610         for (i = start; i != end; i += dir)
 611                 if (dev->db_tab->page[i].db_rec &&
 612                     !bitmap_full(dev->db_tab->page[i].used,
 613                                  MTHCA_DB_REC_PER_PAGE)) {
 614                         page = dev->db_tab->page + i;
 615                         goto found;
 616                 }
 617
 618         for (i = start; i != end; i += dir)
 619                 if (!dev->db_tab->page[i].db_rec) {
 620                         page = dev->db_tab->page + i;
 621                         goto alloc;
 622                 }
 623
 624         if (dev->db_tab->max_group1 >= dev->db_tab->min_group2 - 1) {
 625                 ret = -ENOMEM;
 626                 goto out;
 627         }
 628
 629         if (group == 0)
 630                 ++dev->db_tab->max_group1;
 631         else
 632                 --dev->db_tab->min_group2;
 633
 634         page = dev->db_tab->page + end;
 635
 636 alloc:
 637         page->db_rec = dma_alloc_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
 638                                           &page->mapping, GFP_KERNEL);
 639         if (!page->db_rec) {
 640                 ret = -ENOMEM;
 641                 goto out;
 642         }
 643         memset(page->db_rec, 0, MTHCA_ICM_PAGE_SIZE);
 644
 645         ret = mthca_MAP_ICM_page(dev, page->mapping,
 646                                  mthca_uarc_virt(dev, &dev->driver_uar, i), &status);
 647         if (!ret && status)
 648                 ret = -EINVAL;
 649         if (ret) {
 650                 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
 651                                   page->db_rec, page->mapping);
 652                 goto out;
 653         }
 654
 655         bitmap_zero(page->used, MTHCA_DB_REC_PER_PAGE);
 656
 657 found:
 658         j = find_first_zero_bit(page->used, MTHCA_DB_REC_PER_PAGE);
 659         set_bit(j, page->used);
 660
 661         if (group == 1)
 662                 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
 663
 664         ret = i * MTHCA_DB_REC_PER_PAGE + j;
 665
 666         page->db_rec[j] = cpu_to_be64((qn << 8) | (type << 5));
 667
 668         *db = (__be32 *) &page->db_rec[j];
 669
 670 out:
 671         mutex_unlock(&dev->db_tab->mutex);
 672
 673         return ret;
 674 }
 675
 676 void mthca_free_db(struct mthca_dev *dev, int type, int db_index)
 677 {
 678         int i, j;
 679         struct mthca_db_page *page;
 680         u8 status;
 681
 682         i = db_index / MTHCA_DB_REC_PER_PAGE;
 683         j = db_index % MTHCA_DB_REC_PER_PAGE;
 684
 685         page = dev->db_tab->page + i;
 686
 687         mutex_lock(&dev->db_tab->mutex);
 688
 689         page->db_rec[j] = 0;
 690         if (i >= dev->db_tab->min_group2)
 691                 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
 692         clear_bit(j, page->used);
 693
 694         if (bitmap_empty(page->used, MTHCA_DB_REC_PER_PAGE) &&
 695             i >= dev->db_tab->max_group1 - 1) {
 696                 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
 697
 698                 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
 699                                   page->db_rec, page->mapping);
 700                 page->db_rec = NULL;
 701
 702                 if (i == dev->db_tab->max_group1) {
 703                         --dev->db_tab->max_group1;
 704                         /* XXX may be able to unmap more pages now */
 705                 }
 706                 if (i == dev->db_tab->min_group2)
 707                         ++dev->db_tab->min_group2;
 708         }
 709
 710         mutex_unlock(&dev->db_tab->mutex);
 711 }
 712
 713 int mthca_init_db_tab(struct mthca_dev *dev)
 714 {
 715         int i;
 716
 717         if (!mthca_is_memfree(dev))
 718                 return 0;
 719
 720         dev->db_tab = kmalloc(sizeof *dev->db_tab, GFP_KERNEL);
 721         if (!dev->db_tab)
 722                 return -ENOMEM;
 723
 724         mutex_init(&dev->db_tab->mutex);
 725
 726         dev->db_tab->npages     = dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE;
 727         dev->db_tab->max_group1 = 0;
 728         dev->db_tab->min_group2 = dev->db_tab->npages - 1;
 729
 730         dev->db_tab->page = kmalloc(dev->db_tab->npages *
 731                                     sizeof *dev->db_tab->page,
 732                                     GFP_KERNEL);
 733         if (!dev->db_tab->page) {
 734                 kfree(dev->db_tab);
 735                 return -ENOMEM;
 736         }
 737
 738         for (i = 0; i < dev->db_tab->npages; ++i)
 739                 dev->db_tab->page[i].db_rec = NULL;
 740
 741         return 0;
 742 }
 743
 744 void mthca_cleanup_db_tab(struct mthca_dev *dev)
 745 {
 746         int i;
 747         u8 status;
 748
 749         if (!mthca_is_memfree(dev))
 750                 return;
 751
 752         /*
 753          * Because we don't always free our UARC pages when they
 754          * become empty to make mthca_free_db() simpler we need to
 755          * make a sweep through the doorbell pages and free any
 756          * leftover pages now.
 757          */
 758         for (i = 0; i < dev->db_tab->npages; ++i) {
 759                 if (!dev->db_tab->page[i].db_rec)
 760                         continue;
 761
 762                 if (!bitmap_empty(dev->db_tab->page[i].used, MTHCA_DB_REC_PER_PAGE))
 763                         mthca_warn(dev, "Kernel UARC page %d not empty\n", i);
 764
 765                 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
 766
 767                 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
 768                                   dev->db_tab->page[i].db_rec,
 769                                   dev->db_tab->page[i].mapping);
 770         }
 771
 772         kfree(dev->db_tab->page);
 773         kfree(dev->db_tab);
 774 }