err.no 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(sg_page(&chunk->mem[i]),
  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(sg_page(&chunk->mem[i])),
  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         struct page *page;
 111
 112         /*
 113          * Use __GFP_ZERO because buggy firmware assumes ICM pages are
 114          * cleared, and subtle failures are seen if they aren't.
 115          */
 116         page = alloc_pages(gfp_mask | __GFP_ZERO, order);
 117         if (!page)
 118                 return -ENOMEM;
 119
 120         sg_set_page(mem, page, PAGE_SIZE << order, 0);
 121         return 0;
 122 }
 123
 124 static int mthca_alloc_icm_coherent(struct device *dev, struct scatterlist *mem,
 125                                     int order, gfp_t gfp_mask)
 126 {
 127         void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order, &sg_dma_address(mem),
 128                                        gfp_mask);
 129         if (!buf)
 130                 return -ENOMEM;
 131
 132         sg_set_buf(mem, buf, PAGE_SIZE << order);
 133         BUG_ON(mem->offset);
 134         sg_dma_len(mem) = PAGE_SIZE << order;
 135         return 0;
 136 }
 137
 138 struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages,
 139                                   gfp_t gfp_mask, int coherent)
 140 {
 141         struct mthca_icm *icm;
 142         struct mthca_icm_chunk *chunk = NULL;
 143         int cur_order;
 144         int ret;
 145
 146         /* We use sg_set_buf for coherent allocs, which assumes low memory */
 147         BUG_ON(coherent && (gfp_mask & __GFP_HIGHMEM));
 148
 149         icm = kmalloc(sizeof *icm, gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
 150         if (!icm)
 151                 return icm;
 152
 153         icm->refcount = 0;
 154         INIT_LIST_HEAD(&icm->chunk_list);
 155
 156         cur_order = get_order(MTHCA_ICM_ALLOC_SIZE);
 157
 158         while (npages > 0) {
 159                 if (!chunk) {
 160                         chunk = kmalloc(sizeof *chunk,
 161                                         gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
 162                         if (!chunk)
 163                                 goto fail;
 164
 165                         sg_init_table(chunk->mem, MTHCA_ICM_CHUNK_LEN);
 166                         chunk->npages = 0;
 167                         chunk->nsg    = 0;
 168                         list_add_tail(&chunk->list, &icm->chunk_list);
 169                 }
 170
 171                 while (1 << cur_order > npages)
 172                         --cur_order;
 173
 174                 if (coherent)
 175                         ret = mthca_alloc_icm_coherent(&dev->pdev->dev,
 176                                                        &chunk->mem[chunk->npages],
 177                                                        cur_order, gfp_mask);
 178                 else
 179                         ret = mthca_alloc_icm_pages(&chunk->mem[chunk->npages],
 180                                                     cur_order, gfp_mask);
 181
 182                 if (!ret) {
 183                         ++chunk->npages;
 184
 185                         if (coherent)
 186                                 ++chunk->nsg;
 187                         else if (chunk->npages == MTHCA_ICM_CHUNK_LEN) {
 188                                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 189                                                         chunk->npages,
 190                                                         PCI_DMA_BIDIRECTIONAL);
 191
 192                                 if (chunk->nsg <= 0)
 193                                         goto fail;
 194                         }
 195
 196                         if (chunk->npages == MTHCA_ICM_CHUNK_LEN)
 197                                 chunk = NULL;
 198
 199                         npages -= 1 << cur_order;
 200                 } else {
 201                         --cur_order;
 202                         if (cur_order < 0)
 203                                 goto fail;
 204                 }
 205         }
 206
 207         if (!coherent && chunk) {
 208                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 209                                         chunk->npages,
 210                                         PCI_DMA_BIDIRECTIONAL);
 211
 212                 if (chunk->nsg <= 0)
 213                         goto fail;
 214         }
 215
 216         return icm;
 217
 218 fail:
 219         mthca_free_icm(dev, icm, coherent);
 220         return NULL;
 221 }
 222
 223 int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
 224 {
 225         int i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
 226         int ret = 0;
 227         u8 status;
 228
 229         mutex_lock(&table->mutex);
 230
 231         if (table->icm[i]) {
 232                 ++table->icm[i]->refcount;
 233                 goto out;
 234         }
 235
 236         table->icm[i] = mthca_alloc_icm(dev, MTHCA_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
 237                                         (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
 238                                         __GFP_NOWARN, table->coherent);
 239         if (!table->icm[i]) {
 240                 ret = -ENOMEM;
 241                 goto out;
 242         }
 243
 244         if (mthca_MAP_ICM(dev, table->icm[i], table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
 245                           &status) || status) {
 246                 mthca_free_icm(dev, table->icm[i], table->coherent);
 247                 table->icm[i] = NULL;
 248                 ret = -ENOMEM;
 249                 goto out;
 250         }
 251
 252         ++table->icm[i]->refcount;
 253
 254 out:
 255         mutex_unlock(&table->mutex);
 256         return ret;
 257 }
 258
 259 void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
 260 {
 261         int i;
 262         u8 status;
 263
 264         if (!mthca_is_memfree(dev))
 265                 return;
 266
 267         i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
 268
 269         mutex_lock(&table->mutex);
 270
 271         if (--table->icm[i]->refcount == 0) {
 272                 mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
 273                                 MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
 274                                 &status);
 275                 mthca_free_icm(dev, table->icm[i], table->coherent);
 276                 table->icm[i] = NULL;
 277         }
 278
 279         mutex_unlock(&table->mutex);
 280 }
 281
 282 void *mthca_table_find(struct mthca_icm_table *table, int obj, dma_addr_t *dma_handle)
 283 {
 284         int idx, offset, dma_offset, i;
 285         struct mthca_icm_chunk *chunk;
 286         struct mthca_icm *icm;
 287         struct page *page = NULL;
 288
 289         if (!table->lowmem)
 290                 return NULL;
 291
 292         mutex_lock(&table->mutex);
 293
 294         idx = (obj & (table->num_obj - 1)) * table->obj_size;
 295         icm = table->icm[idx / MTHCA_TABLE_CHUNK_SIZE];
 296         dma_offset = offset = idx % MTHCA_TABLE_CHUNK_SIZE;
 297
 298         if (!icm)
 299                 goto out;
 300
 301         list_for_each_entry(chunk, &icm->chunk_list, list) {
 302                 for (i = 0; i < chunk->npages; ++i) {
 303                         if (dma_handle && dma_offset >= 0) {
 304                                 if (sg_dma_len(&chunk->mem[i]) > dma_offset)
 305                                         *dma_handle = sg_dma_address(&chunk->mem[i]) +
 306                                                 dma_offset;
 307                                 dma_offset -= sg_dma_len(&chunk->mem[i]);
 308                         }
 309                         /* DMA mapping can merge pages but not split them,
 310                          * so if we found the page, dma_handle has already
 311                          * been assigned to. */
 312                         if (chunk->mem[i].length > offset) {
 313                                 page = sg_page(&chunk->mem[i]);
 314                                 goto out;
 315                         }
 316                         offset -= chunk->mem[i].length;
 317                 }
 318         }
 319
 320 out:
 321         mutex_unlock(&table->mutex);
 322         return page ? lowmem_page_address(page) + offset : NULL;
 323 }
 324
 325 int mthca_table_get_range(struct mthca_dev *dev, struct mthca_icm_table *table,
 326                           int start, int end)
 327 {
 328         int inc = MTHCA_TABLE_CHUNK_SIZE / table->obj_size;
 329         int i, err;
 330
 331         for (i = start; i <= end; i += inc) {
 332                 err = mthca_table_get(dev, table, i);
 333                 if (err)
 334                         goto fail;
 335         }
 336
 337         return 0;
 338
 339 fail:
 340         while (i > start) {
 341                 i -= inc;
 342                 mthca_table_put(dev, table, i);
 343         }
 344
 345         return err;
 346 }
 347
 348 void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table,
 349                            int start, int end)
 350 {
 351         int i;
 352
 353         if (!mthca_is_memfree(dev))
 354                 return;
 355
 356         for (i = start; i <= end; i += MTHCA_TABLE_CHUNK_SIZE / table->obj_size)
 357                 mthca_table_put(dev, table, i);
 358 }
 359
 360 struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
 361                                               u64 virt, int obj_size,
 362                                               int nobj, int reserved,
 363                                               int use_lowmem, int use_coherent)
 364 {
 365         struct mthca_icm_table *table;
 366         int obj_per_chunk;
 367         int num_icm;
 368         unsigned chunk_size;
 369         int i;
 370         u8 status;
 371
 372         obj_per_chunk = MTHCA_TABLE_CHUNK_SIZE / obj_size;
 373         num_icm = DIV_ROUND_UP(nobj, obj_per_chunk);
 374
 375         table = kmalloc(sizeof *table + num_icm * sizeof *table->icm, GFP_KERNEL);
 376         if (!table)
 377                 return NULL;
 378
 379         table->virt     = virt;
 380         table->num_icm  = num_icm;
 381         table->num_obj  = nobj;
 382         table->obj_size = obj_size;
 383         table->lowmem   = use_lowmem;
 384         table->coherent = use_coherent;
 385         mutex_init(&table->mutex);
 386
 387         for (i = 0; i < num_icm; ++i)
 388                 table->icm[i] = NULL;
 389
 390         for (i = 0; i * MTHCA_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
 391                 chunk_size = MTHCA_TABLE_CHUNK_SIZE;
 392                 if ((i + 1) * MTHCA_TABLE_CHUNK_SIZE > nobj * obj_size)
 393                         chunk_size = nobj * obj_size - i * MTHCA_TABLE_CHUNK_SIZE;
 394
 395                 table->icm[i] = mthca_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
 396                                                 (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
 397                                                 __GFP_NOWARN, use_coherent);
 398                 if (!table->icm[i])
 399                         goto err;
 400                 if (mthca_MAP_ICM(dev, table->icm[i], virt + i * MTHCA_TABLE_CHUNK_SIZE,
 401                                   &status) || status) {
 402                         mthca_free_icm(dev, table->icm[i], table->coherent);
 403                         table->icm[i] = NULL;
 404                         goto err;
 405                 }
 406
 407                 /*
 408                  * Add a reference to this ICM chunk so that it never
 409                  * gets freed (since it contains reserved firmware objects).
 410                  */
 411                 ++table->icm[i]->refcount;
 412         }
 413
 414         return table;
 415
 416 err:
 417         for (i = 0; i < num_icm; ++i)
 418                 if (table->icm[i]) {
 419                         mthca_UNMAP_ICM(dev, virt + i * MTHCA_TABLE_CHUNK_SIZE,
 420                                         MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
 421                                         &status);
 422                         mthca_free_icm(dev, table->icm[i], table->coherent);
 423                 }
 424
 425         kfree(table);
 426
 427         return NULL;
 428 }
 429
 430 void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table)
 431 {
 432         int i;
 433         u8 status;
 434
 435         for (i = 0; i < table->num_icm; ++i)
 436                 if (table->icm[i]) {
 437                         mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
 438                                         MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
 439                                         &status);
 440                         mthca_free_icm(dev, table->icm[i], table->coherent);
 441                 }
 442
 443         kfree(table);
 444 }
 445
 446 static u64 mthca_uarc_virt(struct mthca_dev *dev, struct mthca_uar *uar, int page)
 447 {
 448         return dev->uar_table.uarc_base +
 449                 uar->index * dev->uar_table.uarc_size +
 450                 page * MTHCA_ICM_PAGE_SIZE;
 451 }
 452
 453 int mthca_map_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
 454                       struct mthca_user_db_table *db_tab, int index, u64 uaddr)
 455 {
 456         struct page *pages[1];
 457         int ret = 0;
 458         u8 status;
 459         int i;
 460
 461         if (!mthca_is_memfree(dev))
 462                 return 0;
 463
 464         if (index < 0 || index > dev->uar_table.uarc_size / 8)
 465                 return -EINVAL;
 466
 467         mutex_lock(&db_tab->mutex);
 468
 469         i = index / MTHCA_DB_REC_PER_PAGE;
 470
 471         if ((db_tab->page[i].refcount >= MTHCA_DB_REC_PER_PAGE)       ||
 472             (db_tab->page[i].uvirt && db_tab->page[i].uvirt != uaddr) ||
 473             (uaddr & 4095)) {
 474                 ret = -EINVAL;
 475                 goto out;
 476         }
 477
 478         if (db_tab->page[i].refcount) {
 479                 ++db_tab->page[i].refcount;
 480                 goto out;
 481         }
 482
 483         ret = get_user_pages(current, current->mm, uaddr & PAGE_MASK, 1, 1, 0,
 484                              pages, NULL);
 485         if (ret < 0)
 486                 goto out;
 487
 488         sg_set_page(&db_tab->page[i].mem, pages[0], MTHCA_ICM_PAGE_SIZE,
 489                         uaddr & ~PAGE_MASK);
 490
 491         ret = pci_map_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
 492         if (ret < 0) {
 493                 put_page(pages[0]);
 494                 goto out;
 495         }
 496
 497         ret = mthca_MAP_ICM_page(dev, sg_dma_address(&db_tab->page[i].mem),
 498                                  mthca_uarc_virt(dev, uar, i), &status);
 499         if (!ret && status)
 500                 ret = -EINVAL;
 501         if (ret) {
 502                 pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
 503                 put_page(sg_page(&db_tab->page[i].mem));
 504                 goto out;
 505         }
 506
 507         db_tab->page[i].uvirt    = uaddr;
 508         db_tab->page[i].refcount = 1;
 509
 510 out:
 511         mutex_unlock(&db_tab->mutex);
 512         return ret;
 513 }
 514
 515 void mthca_unmap_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
 516                          struct mthca_user_db_table *db_tab, int index)
 517 {
 518         if (!mthca_is_memfree(dev))
 519                 return;
 520
 521         /*
 522          * To make our bookkeeping simpler, we don't unmap DB
 523          * pages until we clean up the whole db table.
 524          */
 525
 526         mutex_lock(&db_tab->mutex);
 527
 528         --db_tab->page[index / MTHCA_DB_REC_PER_PAGE].refcount;
 529
 530         mutex_unlock(&db_tab->mutex);
 531 }
 532
 533 struct mthca_user_db_table *mthca_init_user_db_tab(struct mthca_dev *dev)
 534 {
 535         struct mthca_user_db_table *db_tab;
 536         int npages;
 537         int i;
 538
 539         if (!mthca_is_memfree(dev))
 540                 return NULL;
 541
 542         npages = dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE;
 543         db_tab = kmalloc(sizeof *db_tab + npages * sizeof *db_tab->page, GFP_KERNEL);
 544         if (!db_tab)
 545                 return ERR_PTR(-ENOMEM);
 546
 547         mutex_init(&db_tab->mutex);
 548         for (i = 0; i < npages; ++i) {
 549                 db_tab->page[i].refcount = 0;
 550                 db_tab->page[i].uvirt    = 0;
 551                 sg_init_table(&db_tab->page[i].mem, 1);
 552         }
 553
 554         return db_tab;
 555 }
 556
 557 void mthca_cleanup_user_db_tab(struct mthca_dev *dev, struct mthca_uar *uar,
 558                                struct mthca_user_db_table *db_tab)
 559 {
 560         int i;
 561         u8 status;
 562
 563         if (!mthca_is_memfree(dev))
 564                 return;
 565
 566         for (i = 0; i < dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE; ++i) {
 567                 if (db_tab->page[i].uvirt) {
 568                         mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, uar, i), 1, &status);
 569                         pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
 570                         put_page(sg_page(&db_tab->page[i].mem));
 571                 }
 572         }
 573
 574         kfree(db_tab);
 575 }
 576
 577 int mthca_alloc_db(struct mthca_dev *dev, enum mthca_db_type type,
 578                    u32 qn, __be32 **db)
 579 {
 580         int group;
 581         int start, end, dir;
 582         int i, j;
 583         struct mthca_db_page *page;
 584         int ret = 0;
 585         u8 status;
 586
 587         mutex_lock(&dev->db_tab->mutex);
 588
 589         switch (type) {
 590         case MTHCA_DB_TYPE_CQ_ARM:
 591         case MTHCA_DB_TYPE_SQ:
 592                 group = 0;
 593                 start = 0;
 594                 end   = dev->db_tab->max_group1;
 595                 dir   = 1;
 596                 break;
 597
 598         case MTHCA_DB_TYPE_CQ_SET_CI:
 599         case MTHCA_DB_TYPE_RQ:
 600         case MTHCA_DB_TYPE_SRQ:
 601                 group = 1;
 602                 start = dev->db_tab->npages - 1;
 603                 end   = dev->db_tab->min_group2;
 604                 dir   = -1;
 605                 break;
 606
 607         default:
 608                 ret = -EINVAL;
 609                 goto out;
 610         }
 611
 612         for (i = start; i != end; i += dir)
 613                 if (dev->db_tab->page[i].db_rec &&
 614                     !bitmap_full(dev->db_tab->page[i].used,
 615                                  MTHCA_DB_REC_PER_PAGE)) {
 616                         page = dev->db_tab->page + i;
 617                         goto found;
 618                 }
 619
 620         for (i = start; i != end; i += dir)
 621                 if (!dev->db_tab->page[i].db_rec) {
 622                         page = dev->db_tab->page + i;
 623                         goto alloc;
 624                 }
 625
 626         if (dev->db_tab->max_group1 >= dev->db_tab->min_group2 - 1) {
 627                 ret = -ENOMEM;
 628                 goto out;
 629         }
 630
 631         if (group == 0)
 632                 ++dev->db_tab->max_group1;
 633         else
 634                 --dev->db_tab->min_group2;
 635
 636         page = dev->db_tab->page + end;
 637
 638 alloc:
 639         page->db_rec = dma_alloc_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
 640                                           &page->mapping, GFP_KERNEL);
 641         if (!page->db_rec) {
 642                 ret = -ENOMEM;
 643                 goto out;
 644         }
 645         memset(page->db_rec, 0, MTHCA_ICM_PAGE_SIZE);
 646
 647         ret = mthca_MAP_ICM_page(dev, page->mapping,
 648                                  mthca_uarc_virt(dev, &dev->driver_uar, i), &status);
 649         if (!ret && status)
 650                 ret = -EINVAL;
 651         if (ret) {
 652                 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
 653                                   page->db_rec, page->mapping);
 654                 goto out;
 655         }
 656
 657         bitmap_zero(page->used, MTHCA_DB_REC_PER_PAGE);
 658
 659 found:
 660         j = find_first_zero_bit(page->used, MTHCA_DB_REC_PER_PAGE);
 661         set_bit(j, page->used);
 662
 663         if (group == 1)
 664                 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
 665
 666         ret = i * MTHCA_DB_REC_PER_PAGE + j;
 667
 668         page->db_rec[j] = cpu_to_be64((qn << 8) | (type << 5));
 669
 670         *db = (__be32 *) &page->db_rec[j];
 671
 672 out:
 673         mutex_unlock(&dev->db_tab->mutex);
 674
 675         return ret;
 676 }
 677
 678 void mthca_free_db(struct mthca_dev *dev, int type, int db_index)
 679 {
 680         int i, j;
 681         struct mthca_db_page *page;
 682         u8 status;
 683
 684         i = db_index / MTHCA_DB_REC_PER_PAGE;
 685         j = db_index % MTHCA_DB_REC_PER_PAGE;
 686
 687         page = dev->db_tab->page + i;
 688
 689         mutex_lock(&dev->db_tab->mutex);
 690
 691         page->db_rec[j] = 0;
 692         if (i >= dev->db_tab->min_group2)
 693                 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
 694         clear_bit(j, page->used);
 695
 696         if (bitmap_empty(page->used, MTHCA_DB_REC_PER_PAGE) &&
 697             i >= dev->db_tab->max_group1 - 1) {
 698                 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
 699
 700                 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
 701                                   page->db_rec, page->mapping);
 702                 page->db_rec = NULL;
 703
 704                 if (i == dev->db_tab->max_group1) {
 705                         --dev->db_tab->max_group1;
 706                         /* XXX may be able to unmap more pages now */
 707                 }
 708                 if (i == dev->db_tab->min_group2)
 709                         ++dev->db_tab->min_group2;
 710         }
 711
 712         mutex_unlock(&dev->db_tab->mutex);
 713 }
 714
 715 int mthca_init_db_tab(struct mthca_dev *dev)
 716 {
 717         int i;
 718
 719         if (!mthca_is_memfree(dev))
 720                 return 0;
 721
 722         dev->db_tab = kmalloc(sizeof *dev->db_tab, GFP_KERNEL);
 723         if (!dev->db_tab)
 724                 return -ENOMEM;
 725
 726         mutex_init(&dev->db_tab->mutex);
 727
 728         dev->db_tab->npages     = dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE;
 729         dev->db_tab->max_group1 = 0;
 730         dev->db_tab->min_group2 = dev->db_tab->npages - 1;
 731
 732         dev->db_tab->page = kmalloc(dev->db_tab->npages *
 733                                     sizeof *dev->db_tab->page,
 734                                     GFP_KERNEL);
 735         if (!dev->db_tab->page) {
 736                 kfree(dev->db_tab);
 737                 return -ENOMEM;
 738         }
 739
 740         for (i = 0; i < dev->db_tab->npages; ++i)
 741                 dev->db_tab->page[i].db_rec = NULL;
 742
 743         return 0;
 744 }
 745
 746 void mthca_cleanup_db_tab(struct mthca_dev *dev)
 747 {
 748         int i;
 749         u8 status;
 750
 751         if (!mthca_is_memfree(dev))
 752                 return;
 753
 754         /*
 755          * Because we don't always free our UARC pages when they
 756          * become empty to make mthca_free_db() simpler we need to
 757          * make a sweep through the doorbell pages and free any
 758          * leftover pages now.
 759          */
 760         for (i = 0; i < dev->db_tab->npages; ++i) {
 761                 if (!dev->db_tab->page[i].db_rec)
 762                         continue;
 763
 764                 if (!bitmap_empty(dev->db_tab->page[i].used, MTHCA_DB_REC_PER_PAGE))
 765                         mthca_warn(dev, "Kernel UARC page %d not empty\n", i);
 766
 767                 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
 768
 769                 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
 770                                   dev->db_tab->page[i].db_rec,
 771                                   dev->db_tab->page[i].mapping);
 772         }
 773
 774         kfree(dev->db_tab->page);
 775         kfree(dev->db_tab);
 776 }