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 #include <linux/slab.h>
  39
  40 #include <asm/page.h>
  41
  42 #include "mthca_memfree.h"
  43 #include "mthca_dev.h"
  44 #include "mthca_cmd.h"
  45
  46 /*
  47  * We allocate in as big chunks as we can, up to a maximum of 256 KB
  48  * per chunk.
  49  */
  50 enum {
  51         MTHCA_ICM_ALLOC_SIZE   = 1 << 18,
  52         MTHCA_TABLE_CHUNK_SIZE = 1 << 18
  53 };
  54
  55 struct mthca_user_db_table {
  56         struct mutex mutex;
  57         struct {
  58                 u64                uvirt;
  59                 struct scatterlist mem;
  60                 int                refcount;
  61         }                page[0];
  62 };
  63
  64 static void mthca_free_icm_pages(struct mthca_dev *dev, struct mthca_icm_chunk *chunk)
  65 {
  66         int i;
  67
  68         if (chunk->nsg > 0)
  69                 pci_unmap_sg(dev->pdev, chunk->mem, chunk->npages,
  70                              PCI_DMA_BIDIRECTIONAL);
  71
  72         for (i = 0; i < chunk->npages; ++i)
  73                 __free_pages(sg_page(&chunk->mem[i]),
  74                              get_order(chunk->mem[i].length));
  75 }
  76
  77 static void mthca_free_icm_coherent(struct mthca_dev *dev, struct mthca_icm_chunk *chunk)
  78 {
  79         int i;
  80
  81         for (i = 0; i < chunk->npages; ++i) {
  82                 dma_free_coherent(&dev->pdev->dev, chunk->mem[i].length,
  83                                   lowmem_page_address(sg_page(&chunk->mem[i])),
  84                                   sg_dma_address(&chunk->mem[i]));
  85         }
  86 }
  87
  88 void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm, int coherent)
  89 {
  90         struct mthca_icm_chunk *chunk, *tmp;
  91
  92         if (!icm)
  93                 return;
  94
  95         list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) {
  96                 if (coherent)
  97                         mthca_free_icm_coherent(dev, chunk);
  98                 else
  99                         mthca_free_icm_pages(dev, chunk);
 100
 101                 kfree(chunk);
 102         }
 103
 104         kfree(icm);
 105 }
 106
 107 static int mthca_alloc_icm_pages(struct scatterlist *mem, int order, gfp_t gfp_mask)
 108 {
 109         struct page *page;
 110
 111         /*
 112          * Use __GFP_ZERO because buggy firmware assumes ICM pages are
 113          * cleared, and subtle failures are seen if they aren't.
 114          */
 115         page = alloc_pages(gfp_mask | __GFP_ZERO, order);
 116         if (!page)
 117                 return -ENOMEM;
 118
 119         sg_set_page(mem, page, PAGE_SIZE << order, 0);
 120         return 0;
 121 }
 122
 123 static int mthca_alloc_icm_coherent(struct device *dev, struct scatterlist *mem,
 124                                     int order, gfp_t gfp_mask)
 125 {
 126         void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order, &sg_dma_address(mem),
 127                                        gfp_mask);
 128         if (!buf)
 129                 return -ENOMEM;
 130
 131         sg_set_buf(mem, buf, PAGE_SIZE << order);
 132         BUG_ON(mem->offset);
 133         sg_dma_len(mem) = PAGE_SIZE << order;
 134         return 0;
 135 }
 136
 137 struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages,
 138                                   gfp_t gfp_mask, int coherent)
 139 {
 140         struct mthca_icm *icm;
 141         struct mthca_icm_chunk *chunk = NULL;
 142         int cur_order;
 143         int ret;
 144
 145         /* We use sg_set_buf for coherent allocs, which assumes low memory */
 146         BUG_ON(coherent && (gfp_mask & __GFP_HIGHMEM));
 147
 148         icm = kmalloc(sizeof *icm, gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
 149         if (!icm)
 150                 return icm;
 151
 152         icm->refcount = 0;
 153         INIT_LIST_HEAD(&icm->chunk_list);
 154
 155         cur_order = get_order(MTHCA_ICM_ALLOC_SIZE);
 156
 157         while (npages > 0) {
 158                 if (!chunk) {
 159                         chunk = kmalloc(sizeof *chunk,
 160                                         gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
 161                         if (!chunk)
 162                                 goto fail;
 163
 164                         sg_init_table(chunk->mem, MTHCA_ICM_CHUNK_LEN);
 165                         chunk->npages = 0;
 166                         chunk->nsg    = 0;
 167                         list_add_tail(&chunk->list, &icm->chunk_list);
 168                 }
 169
 170                 while (1 << cur_order > npages)
 171                         --cur_order;
 172
 173                 if (coherent)
 174                         ret = mthca_alloc_icm_coherent(&dev->pdev->dev,
 175                                                        &chunk->mem[chunk->npages],
 176                                                        cur_order, gfp_mask);
 177                 else
 178                         ret = mthca_alloc_icm_pages(&chunk->mem[chunk->npages],
 179                                                     cur_order, gfp_mask);
 180
 181                 if (!ret) {
 182                         ++chunk->npages;
 183
 184                         if (coherent)
 185                                 ++chunk->nsg;
 186                         else if (chunk->npages == MTHCA_ICM_CHUNK_LEN) {
 187                                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 188                                                         chunk->npages,
 189                                                         PCI_DMA_BIDIRECTIONAL);
 190
 191                                 if (chunk->nsg <= 0)
 192                                         goto fail;
 193                         }
 194
 195                         if (chunk->npages == MTHCA_ICM_CHUNK_LEN)
 196                                 chunk = NULL;
 197
 198                         npages -= 1 << cur_order;
 199                 } else {
 200                         --cur_order;
 201                         if (cur_order < 0)
 202                                 goto fail;
 203                 }
 204         }
 205
 206         if (!coherent && chunk) {
 207                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 208                                         chunk->npages,
 209                                         PCI_DMA_BIDIRECTIONAL);
 210
 211                 if (chunk->nsg <= 0)
 212                         goto fail;
 213         }
 214
 215         return icm;
 216
 217 fail:
 218         mthca_free_icm(dev, icm, coherent);
 219         return NULL;
 220 }
 221
 222 int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
 223 {
 224         int i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
 225         int ret = 0;
 226         u8 status;
 227
 228         mutex_lock(&table->mutex);
 229
 230         if (table->icm[i]) {
 231                 ++table->icm[i]->refcount;
 232                 goto out;
 233         }
 234
 235         table->icm[i] = mthca_alloc_icm(dev, MTHCA_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
 236                                         (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
 237                                         __GFP_NOWARN, table->coherent);
 238         if (!table->icm[i]) {
 239                 ret = -ENOMEM;
 240                 goto out;
 241         }
 242
 243         if (mthca_MAP_ICM(dev, table->icm[i], table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
 244                           &status) || status) {
 245                 mthca_free_icm(dev, table->icm[i], table->coherent);
 246                 table->icm[i] = NULL;
 247                 ret = -ENOMEM;
 248                 goto out;
 249         }
 250
 251         ++table->icm[i]->refcount;
 252
 253 out:
 254         mutex_unlock(&table->mutex);
 255         return ret;
 256 }
 257
 258 void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
 259 {
 260         int i;
 261         u8 status;
 262
 263         if (!mthca_is_memfree(dev))
 264                 return;
 265
 266         i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
 267
 268         mutex_lock(&table->mutex);
 269
 270         if (--table->icm[i]->refcount == 0) {
 271                 mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
 272                                 MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
 273                                 &status);
 274                 mthca_free_icm(dev, table->icm[i], table->coherent);
 275                 table->icm[i] = NULL;
 276         }
 277
 278         mutex_unlock(&table->mutex);
 279 }
 280
 281 void *mthca_table_find(struct mthca_icm_table *table, int obj, dma_addr_t *dma_handle)
 282 {
 283         int idx, offset, dma_offset, i;
 284         struct mthca_icm_chunk *chunk;
 285         struct mthca_icm *icm;
 286         struct page *page = NULL;
 287
 288         if (!table->lowmem)
 289                 return NULL;
 290
 291         mutex_lock(&table->mutex);
 292
 293         idx = (obj & (table->num_obj - 1)) * table->obj_size;
 294         icm = table->icm[idx / MTHCA_TABLE_CHUNK_SIZE];
 295         dma_offset = offset = idx % MTHCA_TABLE_CHUNK_SIZE;
 296
 297         if (!icm)
 298                 goto out;
 299
 300         list_for_each_entry(chunk, &icm->chunk_list, list) {
 301                 for (i = 0; i < chunk->npages; ++i) {
 302                         if (dma_handle && dma_offset >= 0) {
 303                                 if (sg_dma_len(&chunk->mem[i]) > dma_offset)
 304                                         *dma_handle = sg_dma_address(&chunk->mem[i]) +
 305                                                 dma_offset;
 306                                 dma_offset -= sg_dma_len(&chunk->mem[i]);
 307                         }
 308                         /* DMA mapping can merge pages but not split them,
 309                          * so if we found the page, dma_handle has already
 310                          * been assigned to. */
 311                         if (chunk->mem[i].length > offset) {
 312                                 page = sg_page(&chunk->mem[i]);
 313                                 goto out;
 314                         }
 315                         offset -= chunk->mem[i].length;
 316                 }
 317         }
 318
 319 out:
 320         mutex_unlock(&table->mutex);
 321         return page ? lowmem_page_address(page) + offset : NULL;
 322 }
 323
 324 int mthca_table_get_range(struct mthca_dev *dev, struct mthca_icm_table *table,
 325                           int start, int end)
 326 {
 327         int inc = MTHCA_TABLE_CHUNK_SIZE / table->obj_size;
 328         int i, err;
 329
 330         for (i = start; i <= end; i += inc) {
 331                 err = mthca_table_get(dev, table, i);
 332                 if (err)
 333                         goto fail;
 334         }
 335
 336         return 0;
 337
 338 fail:
 339         while (i > start) {
 340                 i -= inc;
 341                 mthca_table_put(dev, table, i);
 342         }
 343
 344         return err;
 345 }
 346
 347 void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table,
 348                            int start, int end)
 349 {
 350         int i;
 351
 352         if (!mthca_is_memfree(dev))
 353                 return;
 354
 355         for (i = start; i <= end; i += MTHCA_TABLE_CHUNK_SIZE / table->obj_size)
 356                 mthca_table_put(dev, table, i);
 357 }
 358
 359 struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
 360                                               u64 virt, int obj_size,
 361                                               int nobj, int reserved,
 362                                               int use_lowmem, int use_coherent)
 363 {
 364         struct mthca_icm_table *table;
 365         int obj_per_chunk;
 366         int num_icm;
 367         unsigned chunk_size;
 368         int i;
 369         u8 status;
 370
 371         obj_per_chunk = MTHCA_TABLE_CHUNK_SIZE / obj_size;
 372         num_icm = DIV_ROUND_UP(nobj, obj_per_chunk);
 373
 374         table = kmalloc(sizeof *table + num_icm * sizeof *table->icm, GFP_KERNEL);
 375         if (!table)
 376                 return NULL;
 377
 378         table->virt     = virt;
 379         table->num_icm  = num_icm;
 380         table->num_obj  = nobj;
 381         table->obj_size = obj_size;
 382         table->lowmem   = use_lowmem;
 383         table->coherent = use_coherent;
 384         mutex_init(&table->mutex);
 385
 386         for (i = 0; i < num_icm; ++i)
 387                 table->icm[i] = NULL;
 388
 389         for (i = 0; i * MTHCA_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
 390                 chunk_size = MTHCA_TABLE_CHUNK_SIZE;
 391                 if ((i + 1) * MTHCA_TABLE_CHUNK_SIZE > nobj * obj_size)
 392                         chunk_size = nobj * obj_size - i * MTHCA_TABLE_CHUNK_SIZE;
 393
 394                 table->icm[i] = mthca_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
 395                                                 (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
 396                                                 __GFP_NOWARN, use_coherent);
 397                 if (!table->icm[i])
 398                         goto err;
 399                 if (mthca_MAP_ICM(dev, table->icm[i], virt + i * MTHCA_TABLE_CHUNK_SIZE,
 400                                   &status) || status) {
 401                         mthca_free_icm(dev, table->icm[i], table->coherent);
 402                         table->icm[i] = NULL;
 403                         goto err;
 404                 }
 405
 406                 /*
 407                  * Add a reference to this ICM chunk so that it never
 408                  * gets freed (since it contains reserved firmware objects).
 409                  */
 410                 ++table->icm[i]->refcount;
 411         }
 412
 413         return table;
 414
 415 err:
 416         for (i = 0; i < num_icm; ++i)
 417                 if (table->icm[i]) {
 418                         mthca_UNMAP_ICM(dev, virt + i * MTHCA_TABLE_CHUNK_SIZE,
 419                                         MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
 420                                         &status);
 421                         mthca_free_icm(dev, table->icm[i], table->coherent);
 422                 }
 423
 424         kfree(table);
 425
 426         return NULL;
 427 }
 428
 429 void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table)
 430 {
 431         int i;
 432         u8 status;
 433
 434         for (i = 0; i < table->num_icm; ++i)
 435                 if (table->icm[i]) {
 436                         mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
 437                                         MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
 438                                         &status);
 439                         mthca_free_icm(dev, table->icm[i], table->coherent);
 440                 }
 441
 442         kfree(table);
 443 }
 444
 445 static u64 mthca_uarc_virt(struct mthca_dev *dev, struct mthca_uar *uar, int page)
 446 {
 447         return dev->uar_table.uarc_base +
 448                 uar->index * dev->uar_table.uarc_size +
 449                 page * MTHCA_ICM_PAGE_SIZE;
 450 }
 451
 452 int mthca_map_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
 453                       struct mthca_user_db_table *db_tab, int index, u64 uaddr)
 454 {
 455         struct page *pages[1];
 456         int ret = 0;
 457         u8 status;
 458         int i;
 459
 460         if (!mthca_is_memfree(dev))
 461                 return 0;
 462
 463         if (index < 0 || index > dev->uar_table.uarc_size / 8)
 464                 return -EINVAL;
 465
 466         mutex_lock(&db_tab->mutex);
 467
 468         i = index / MTHCA_DB_REC_PER_PAGE;
 469
 470         if ((db_tab->page[i].refcount >= MTHCA_DB_REC_PER_PAGE)       ||
 471             (db_tab->page[i].uvirt && db_tab->page[i].uvirt != uaddr) ||
 472             (uaddr & 4095)) {
 473                 ret = -EINVAL;
 474                 goto out;
 475         }
 476
 477         if (db_tab->page[i].refcount) {
 478                 ++db_tab->page[i].refcount;
 479                 goto out;
 480         }
 481
 482         ret = get_user_pages(current, current->mm, uaddr & PAGE_MASK, 1, 1, 0,
 483                              pages, NULL);
 484         if (ret < 0)
 485                 goto out;
 486
 487         sg_set_page(&db_tab->page[i].mem, pages[0], MTHCA_ICM_PAGE_SIZE,
 488                         uaddr & ~PAGE_MASK);
 489
 490         ret = pci_map_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
 491         if (ret < 0) {
 492                 put_page(pages[0]);
 493                 goto out;
 494         }
 495
 496         ret = mthca_MAP_ICM_page(dev, sg_dma_address(&db_tab->page[i].mem),
 497                                  mthca_uarc_virt(dev, uar, i), &status);
 498         if (!ret && status)
 499                 ret = -EINVAL;
 500         if (ret) {
 501                 pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
 502                 put_page(sg_page(&db_tab->page[i].mem));
 503                 goto out;
 504         }
 505
 506         db_tab->page[i].uvirt    = uaddr;
 507         db_tab->page[i].refcount = 1;
 508
 509 out:
 510         mutex_unlock(&db_tab->mutex);
 511         return ret;
 512 }
 513
 514 void mthca_unmap_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
 515                          struct mthca_user_db_table *db_tab, int index)
 516 {
 517         if (!mthca_is_memfree(dev))
 518                 return;
 519
 520         /*
 521          * To make our bookkeeping simpler, we don't unmap DB
 522          * pages until we clean up the whole db table.
 523          */
 524
 525         mutex_lock(&db_tab->mutex);
 526
 527         --db_tab->page[index / MTHCA_DB_REC_PER_PAGE].refcount;
 528
 529         mutex_unlock(&db_tab->mutex);
 530 }
 531
 532 struct mthca_user_db_table *mthca_init_user_db_tab(struct mthca_dev *dev)
 533 {
 534         struct mthca_user_db_table *db_tab;
 535         int npages;
 536         int i;
 537
 538         if (!mthca_is_memfree(dev))
 539                 return NULL;
 540
 541         npages = dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE;
 542         db_tab = kmalloc(sizeof *db_tab + npages * sizeof *db_tab->page, GFP_KERNEL);
 543         if (!db_tab)
 544                 return ERR_PTR(-ENOMEM);
 545
 546         mutex_init(&db_tab->mutex);
 547         for (i = 0; i < npages; ++i) {
 548                 db_tab->page[i].refcount = 0;
 549                 db_tab->page[i].uvirt    = 0;
 550                 sg_init_table(&db_tab->page[i].mem, 1);
 551         }
 552
 553         return db_tab;
 554 }
 555
 556 void mthca_cleanup_user_db_tab(struct mthca_dev *dev, struct mthca_uar *uar,
 557                                struct mthca_user_db_table *db_tab)
 558 {
 559         int i;
 560         u8 status;
 561
 562         if (!mthca_is_memfree(dev))
 563                 return;
 564
 565         for (i = 0; i < dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE; ++i) {
 566                 if (db_tab->page[i].uvirt) {
 567                         mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, uar, i), 1, &status);
 568                         pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
 569                         put_page(sg_page(&db_tab->page[i].mem));
 570                 }
 571         }
 572
 573         kfree(db_tab);
 574 }
 575
 576 int mthca_alloc_db(struct mthca_dev *dev, enum mthca_db_type type,
 577                    u32 qn, __be32 **db)
 578 {
 579         int group;
 580         int start, end, dir;
 581         int i, j;
 582         struct mthca_db_page *page;
 583         int ret = 0;
 584         u8 status;
 585
 586         mutex_lock(&dev->db_tab->mutex);
 587
 588         switch (type) {
 589         case MTHCA_DB_TYPE_CQ_ARM:
 590         case MTHCA_DB_TYPE_SQ:
 591                 group = 0;
 592                 start = 0;
 593                 end   = dev->db_tab->max_group1;
 594                 dir   = 1;
 595                 break;
 596
 597         case MTHCA_DB_TYPE_CQ_SET_CI:
 598         case MTHCA_DB_TYPE_RQ:
 599         case MTHCA_DB_TYPE_SRQ:
 600                 group = 1;
 601                 start = dev->db_tab->npages - 1;
 602                 end   = dev->db_tab->min_group2;
 603                 dir   = -1;
 604                 break;
 605
 606         default:
 607                 ret = -EINVAL;
 608                 goto out;
 609         }
 610
 611         for (i = start; i != end; i += dir)
 612                 if (dev->db_tab->page[i].db_rec &&
 613                     !bitmap_full(dev->db_tab->page[i].used,
 614                                  MTHCA_DB_REC_PER_PAGE)) {
 615                         page = dev->db_tab->page + i;
 616                         goto found;
 617                 }
 618
 619         for (i = start; i != end; i += dir)
 620                 if (!dev->db_tab->page[i].db_rec) {
 621                         page = dev->db_tab->page + i;
 622                         goto alloc;
 623                 }
 624
 625         if (dev->db_tab->max_group1 >= dev->db_tab->min_group2 - 1) {
 626                 ret = -ENOMEM;
 627                 goto out;
 628         }
 629
 630         if (group == 0)
 631                 ++dev->db_tab->max_group1;
 632         else
 633                 --dev->db_tab->min_group2;
 634
 635         page = dev->db_tab->page + end;
 636
 637 alloc:
 638         page->db_rec = dma_alloc_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
 639                                           &page->mapping, GFP_KERNEL);
 640         if (!page->db_rec) {
 641                 ret = -ENOMEM;
 642                 goto out;
 643         }
 644         memset(page->db_rec, 0, MTHCA_ICM_PAGE_SIZE);
 645
 646         ret = mthca_MAP_ICM_page(dev, page->mapping,
 647                                  mthca_uarc_virt(dev, &dev->driver_uar, i), &status);
 648         if (!ret && status)
 649                 ret = -EINVAL;
 650         if (ret) {
 651                 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
 652                                   page->db_rec, page->mapping);
 653                 goto out;
 654         }
 655
 656         bitmap_zero(page->used, MTHCA_DB_REC_PER_PAGE);
 657
 658 found:
 659         j = find_first_zero_bit(page->used, MTHCA_DB_REC_PER_PAGE);
 660         set_bit(j, page->used);
 661
 662         if (group == 1)
 663                 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
 664
 665         ret = i * MTHCA_DB_REC_PER_PAGE + j;
 666
 667         page->db_rec[j] = cpu_to_be64((qn << 8) | (type << 5));
 668
 669         *db = (__be32 *) &page->db_rec[j];
 670
 671 out:
 672         mutex_unlock(&dev->db_tab->mutex);
 673
 674         return ret;
 675 }
 676
 677 void mthca_free_db(struct mthca_dev *dev, int type, int db_index)
 678 {
 679         int i, j;
 680         struct mthca_db_page *page;
 681         u8 status;
 682
 683         i = db_index / MTHCA_DB_REC_PER_PAGE;
 684         j = db_index % MTHCA_DB_REC_PER_PAGE;
 685
 686         page = dev->db_tab->page + i;
 687
 688         mutex_lock(&dev->db_tab->mutex);
 689
 690         page->db_rec[j] = 0;
 691         if (i >= dev->db_tab->min_group2)
 692                 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
 693         clear_bit(j, page->used);
 694
 695         if (bitmap_empty(page->used, MTHCA_DB_REC_PER_PAGE) &&
 696             i >= dev->db_tab->max_group1 - 1) {
 697                 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
 698
 699                 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
 700                                   page->db_rec, page->mapping);
 701                 page->db_rec = NULL;
 702
 703                 if (i == dev->db_tab->max_group1) {
 704                         --dev->db_tab->max_group1;
 705                         /* XXX may be able to unmap more pages now */
 706                 }
 707                 if (i == dev->db_tab->min_group2)
 708                         ++dev->db_tab->min_group2;
 709         }
 710
 711         mutex_unlock(&dev->db_tab->mutex);
 712 }
 713
 714 int mthca_init_db_tab(struct mthca_dev *dev)
 715 {
 716         int i;
 717
 718         if (!mthca_is_memfree(dev))
 719                 return 0;
 720
 721         dev->db_tab = kmalloc(sizeof *dev->db_tab, GFP_KERNEL);
 722         if (!dev->db_tab)
 723                 return -ENOMEM;
 724
 725         mutex_init(&dev->db_tab->mutex);
 726
 727         dev->db_tab->npages     = dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE;
 728         dev->db_tab->max_group1 = 0;
 729         dev->db_tab->min_group2 = dev->db_tab->npages - 1;
 730
 731         dev->db_tab->page = kmalloc(dev->db_tab->npages *
 732                                     sizeof *dev->db_tab->page,
 733                                     GFP_KERNEL);
 734         if (!dev->db_tab->page) {
 735                 kfree(dev->db_tab);
 736                 return -ENOMEM;
 737         }
 738
 739         for (i = 0; i < dev->db_tab->npages; ++i)
 740                 dev->db_tab->page[i].db_rec = NULL;
 741
 742         return 0;
 743 }
 744
 745 void mthca_cleanup_db_tab(struct mthca_dev *dev)
 746 {
 747         int i;
 748         u8 status;
 749
 750         if (!mthca_is_memfree(dev))
 751                 return;
 752
 753         /*
 754          * Because we don't always free our UARC pages when they
 755          * become empty to make mthca_free_db() simpler we need to
 756          * make a sweep through the doorbell pages and free any
 757          * leftover pages now.
 758          */
 759         for (i = 0; i < dev->db_tab->npages; ++i) {
 760                 if (!dev->db_tab->page[i].db_rec)
 761                         continue;
 762
 763                 if (!bitmap_empty(dev->db_tab->page[i].used, MTHCA_DB_REC_PER_PAGE))
 764                         mthca_warn(dev, "Kernel UARC page %d not empty\n", i);
 765
 766                 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
 767
 768                 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
 769                                   dev->db_tab->page[i].db_rec,
 770                                   dev->db_tab->page[i].mapping);
 771         }
 772
 773         kfree(dev->db_tab->page);
 774         kfree(dev->db_tab);
 775 }