drivers/infiniband/hw/mlx4/mr.c

   1 /*
   2  * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
   3  * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <linux/slab.h>
  35 #include <rdma/ib_user_verbs.h>
  36
  37 #include "mlx4_ib.h"
  38
  39 static u32 convert_access(int acc)
  40 {
  41         return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX4_PERM_ATOMIC       : 0) |
  42                (acc & IB_ACCESS_REMOTE_WRITE  ? MLX4_PERM_REMOTE_WRITE : 0) |
  43                (acc & IB_ACCESS_REMOTE_READ   ? MLX4_PERM_REMOTE_READ  : 0) |
  44                (acc & IB_ACCESS_LOCAL_WRITE   ? MLX4_PERM_LOCAL_WRITE  : 0) |
  45                (acc & IB_ACCESS_MW_BIND       ? MLX4_PERM_BIND_MW      : 0) |
  46                MLX4_PERM_LOCAL_READ;
  47 }
  48
  49 static enum mlx4_mw_type to_mlx4_type(enum ib_mw_type type)
  50 {
  51         switch (type) {
  52         case IB_MW_TYPE_1:      return MLX4_MW_TYPE_1;
  53         case IB_MW_TYPE_2:      return MLX4_MW_TYPE_2;
  54         default:                return -1;
  55         }
  56 }
  57
  58 struct ib_mr *mlx4_ib_get_dma_mr(struct ib_pd *pd, int acc)
  59 {
  60         struct mlx4_ib_mr *mr;
  61         int err;
  62
  63         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
  64         if (!mr)
  65                 return ERR_PTR(-ENOMEM);
  66
  67         err = mlx4_mr_alloc(to_mdev(pd->device)->dev, to_mpd(pd)->pdn, 0,
  68                             ~0ull, convert_access(acc), 0, 0, &mr->mmr);
  69         if (err)
  70                 goto err_free;
  71
  72         err = mlx4_mr_enable(to_mdev(pd->device)->dev, &mr->mmr);
  73         if (err)
  74                 goto err_mr;
  75
  76         mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
  77         mr->umem = NULL;
  78
  79         return &mr->ibmr;
  80
  81 err_mr:
  82         (void) mlx4_mr_free(to_mdev(pd->device)->dev, &mr->mmr);
  83
  84 err_free:
  85         kfree(mr);
  86
  87         return ERR_PTR(err);
  88 }
  89
  90 int mlx4_ib_umem_write_mtt(struct mlx4_ib_dev *dev, struct mlx4_mtt *mtt,
  91                            struct ib_umem *umem)
  92 {
  93         u64 *pages;
  94         int i, k, entry;
  95         int n;
  96         int len;
  97         int err = 0;
  98         struct scatterlist *sg;
  99
 100         pages = (u64 *) __get_free_page(GFP_KERNEL);
 101         if (!pages)
 102                 return -ENOMEM;
 103
 104         i = n = 0;
 105
 106         for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
 107                 len = sg_dma_len(sg) >> mtt->page_shift;
 108                 for (k = 0; k < len; ++k) {
 109                         pages[i++] = sg_dma_address(sg) +
 110                                 umem->page_size * k;
 111                         /*
 112                          * Be friendly to mlx4_write_mtt() and
 113                          * pass it chunks of appropriate size.
 114                          */
 115                         if (i == PAGE_SIZE / sizeof (u64)) {
 116                                 err = mlx4_write_mtt(dev->dev, mtt, n,
 117                                                      i, pages);
 118                                 if (err)
 119                                         goto out;
 120                                 n += i;
 121                                 i = 0;
 122                         }
 123                 }
 124         }
 125
 126         if (i)
 127                 err = mlx4_write_mtt(dev->dev, mtt, n, i, pages);
 128
 129 out:
 130         free_page((unsigned long) pages);
 131         return err;
 132 }
 133
 134 struct ib_mr *mlx4_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 135                                   u64 virt_addr, int access_flags,
 136                                   struct ib_udata *udata)
 137 {
 138         struct mlx4_ib_dev *dev = to_mdev(pd->device);
 139         struct mlx4_ib_mr *mr;
 140         int shift;
 141         int err;
 142         int n;
 143
 144         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 145         if (!mr)
 146                 return ERR_PTR(-ENOMEM);
 147
 148         /* Force registering the memory as writable. */
 149         /* Used for memory re-registeration. HCA protects the access */
 150         mr->umem = ib_umem_get(pd->uobject->context, start, length,
 151                                access_flags | IB_ACCESS_LOCAL_WRITE, 0);
 152         if (IS_ERR(mr->umem)) {
 153                 err = PTR_ERR(mr->umem);
 154                 goto err_free;
 155         }
 156
 157         n = ib_umem_page_count(mr->umem);
 158         shift = ilog2(mr->umem->page_size);
 159
 160         err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, virt_addr, length,
 161                             convert_access(access_flags), n, shift, &mr->mmr);
 162         if (err)
 163                 goto err_umem;
 164
 165         err = mlx4_ib_umem_write_mtt(dev, &mr->mmr.mtt, mr->umem);
 166         if (err)
 167                 goto err_mr;
 168
 169         err = mlx4_mr_enable(dev->dev, &mr->mmr);
 170         if (err)
 171                 goto err_mr;
 172
 173         mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
 174
 175         return &mr->ibmr;
 176
 177 err_mr:
 178         (void) mlx4_mr_free(to_mdev(pd->device)->dev, &mr->mmr);
 179
 180 err_umem:
 181         ib_umem_release(mr->umem);
 182
 183 err_free:
 184         kfree(mr);
 185
 186         return ERR_PTR(err);
 187 }
 188
 189 int mlx4_ib_rereg_user_mr(struct ib_mr *mr, int flags,
 190                           u64 start, u64 length, u64 virt_addr,
 191                           int mr_access_flags, struct ib_pd *pd,
 192                           struct ib_udata *udata)
 193 {
 194         struct mlx4_ib_dev *dev = to_mdev(mr->device);
 195         struct mlx4_ib_mr *mmr = to_mmr(mr);
 196         struct mlx4_mpt_entry *mpt_entry;
 197         struct mlx4_mpt_entry **pmpt_entry = &mpt_entry;
 198         int err;
 199
 200         /* Since we synchronize this call and mlx4_ib_dereg_mr via uverbs,
 201          * we assume that the calls can't run concurrently. Otherwise, a
 202          * race exists.
 203          */
 204         err =  mlx4_mr_hw_get_mpt(dev->dev, &mmr->mmr, &pmpt_entry);
 205
 206         if (err)
 207                 return err;
 208
 209         if (flags & IB_MR_REREG_PD) {
 210                 err = mlx4_mr_hw_change_pd(dev->dev, *pmpt_entry,
 211                                            to_mpd(pd)->pdn);
 212
 213                 if (err)
 214                         goto release_mpt_entry;
 215         }
 216
 217         if (flags & IB_MR_REREG_ACCESS) {
 218                 err = mlx4_mr_hw_change_access(dev->dev, *pmpt_entry,
 219                                                convert_access(mr_access_flags));
 220
 221                 if (err)
 222                         goto release_mpt_entry;
 223         }
 224
 225         if (flags & IB_MR_REREG_TRANS) {
 226                 int shift;
 227                 int n;
 228
 229                 mlx4_mr_rereg_mem_cleanup(dev->dev, &mmr->mmr);
 230                 ib_umem_release(mmr->umem);
 231                 mmr->umem = ib_umem_get(mr->uobject->context, start, length,
 232                                         mr_access_flags |
 233                                         IB_ACCESS_LOCAL_WRITE,
 234                                         0);
 235                 if (IS_ERR(mmr->umem)) {
 236                         err = PTR_ERR(mmr->umem);
 237                         /* Prevent mlx4_ib_dereg_mr from free'ing invalid pointer */
 238                         mmr->umem = NULL;
 239                         goto release_mpt_entry;
 240                 }
 241                 n = ib_umem_page_count(mmr->umem);
 242                 shift = ilog2(mmr->umem->page_size);
 243
 244                 err = mlx4_mr_rereg_mem_write(dev->dev, &mmr->mmr,
 245                                               virt_addr, length, n, shift,
 246                                               *pmpt_entry);
 247                 if (err) {
 248                         ib_umem_release(mmr->umem);
 249                         goto release_mpt_entry;
 250                 }
 251                 mmr->mmr.iova       = virt_addr;
 252                 mmr->mmr.size       = length;
 253
 254                 err = mlx4_ib_umem_write_mtt(dev, &mmr->mmr.mtt, mmr->umem);
 255                 if (err) {
 256                         mlx4_mr_rereg_mem_cleanup(dev->dev, &mmr->mmr);
 257                         ib_umem_release(mmr->umem);
 258                         goto release_mpt_entry;
 259                 }
 260         }
 261
 262         /* If we couldn't transfer the MR to the HCA, just remember to
 263          * return a failure. But dereg_mr will free the resources.
 264          */
 265         err = mlx4_mr_hw_write_mpt(dev->dev, &mmr->mmr, pmpt_entry);
 266         if (!err && flags & IB_MR_REREG_ACCESS)
 267                 mmr->mmr.access = mr_access_flags;
 268
 269 release_mpt_entry:
 270         mlx4_mr_hw_put_mpt(dev->dev, pmpt_entry);
 271
 272         return err;
 273 }
 274
 275 static int
 276 mlx4_alloc_priv_pages(struct ib_device *device,
 277                       struct mlx4_ib_mr *mr,
 278                       int max_pages)
 279 {
 280         int ret;
 281
 282         /* Ensure that size is aligned to DMA cacheline
 283          * requirements.
 284          * max_pages is limited to MLX4_MAX_FAST_REG_PAGES
 285          * so page_map_size will never cross PAGE_SIZE.
 286          */
 287         mr->page_map_size = roundup(max_pages * sizeof(u64),
 288                                     MLX4_MR_PAGES_ALIGN);
 289
 290         /* Prevent cross page boundary allocation. */
 291         mr->pages = (__be64 *)get_zeroed_page(GFP_KERNEL);
 292         if (!mr->pages)
 293                 return -ENOMEM;
 294
 295         mr->page_map = dma_map_single(device->dma_device, mr->pages,
 296                                       mr->page_map_size, DMA_TO_DEVICE);
 297
 298         if (dma_mapping_error(device->dma_device, mr->page_map)) {
 299                 ret = -ENOMEM;
 300                 goto err;
 301         }
 302
 303         return 0;
 304
 305 err:
 306         free_page((unsigned long)mr->pages);
 307         return ret;
 308 }
 309
 310 static void
 311 mlx4_free_priv_pages(struct mlx4_ib_mr *mr)
 312 {
 313         if (mr->pages) {
 314                 struct ib_device *device = mr->ibmr.device;
 315
 316                 dma_unmap_single(device->dma_device, mr->page_map,
 317                                  mr->page_map_size, DMA_TO_DEVICE);
 318                 free_page((unsigned long)mr->pages);
 319                 mr->pages = NULL;
 320         }
 321 }
 322
 323 int mlx4_ib_dereg_mr(struct ib_mr *ibmr)
 324 {
 325         struct mlx4_ib_mr *mr = to_mmr(ibmr);
 326         int ret;
 327
 328         mlx4_free_priv_pages(mr);
 329
 330         ret = mlx4_mr_free(to_mdev(ibmr->device)->dev, &mr->mmr);
 331         if (ret)
 332                 return ret;
 333         if (mr->umem)
 334                 ib_umem_release(mr->umem);
 335         kfree(mr);
 336
 337         return 0;
 338 }
 339
 340 struct ib_mw *mlx4_ib_alloc_mw(struct ib_pd *pd, enum ib_mw_type type,
 341                                struct ib_udata *udata)
 342 {
 343         struct mlx4_ib_dev *dev = to_mdev(pd->device);
 344         struct mlx4_ib_mw *mw;
 345         int err;
 346
 347         mw = kmalloc(sizeof(*mw), GFP_KERNEL);
 348         if (!mw)
 349                 return ERR_PTR(-ENOMEM);
 350
 351         err = mlx4_mw_alloc(dev->dev, to_mpd(pd)->pdn,
 352                             to_mlx4_type(type), &mw->mmw);
 353         if (err)
 354                 goto err_free;
 355
 356         err = mlx4_mw_enable(dev->dev, &mw->mmw);
 357         if (err)
 358                 goto err_mw;
 359
 360         mw->ibmw.rkey = mw->mmw.key;
 361
 362         return &mw->ibmw;
 363
 364 err_mw:
 365         mlx4_mw_free(dev->dev, &mw->mmw);
 366
 367 err_free:
 368         kfree(mw);
 369
 370         return ERR_PTR(err);
 371 }
 372
 373 int mlx4_ib_dealloc_mw(struct ib_mw *ibmw)
 374 {
 375         struct mlx4_ib_mw *mw = to_mmw(ibmw);
 376
 377         mlx4_mw_free(to_mdev(ibmw->device)->dev, &mw->mmw);
 378         kfree(mw);
 379
 380         return 0;
 381 }
 382
 383 struct ib_mr *mlx4_ib_alloc_mr(struct ib_pd *pd,
 384                                enum ib_mr_type mr_type,
 385                                u32 max_num_sg)
 386 {
 387         struct mlx4_ib_dev *dev = to_mdev(pd->device);
 388         struct mlx4_ib_mr *mr;
 389         int err;
 390
 391         if (mr_type != IB_MR_TYPE_MEM_REG ||
 392             max_num_sg > MLX4_MAX_FAST_REG_PAGES)
 393                 return ERR_PTR(-EINVAL);
 394
 395         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 396         if (!mr)
 397                 return ERR_PTR(-ENOMEM);
 398
 399         err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, 0, 0, 0,
 400                             max_num_sg, 0, &mr->mmr);
 401         if (err)
 402                 goto err_free;
 403
 404         err = mlx4_alloc_priv_pages(pd->device, mr, max_num_sg);
 405         if (err)
 406                 goto err_free_mr;
 407
 408         mr->max_pages = max_num_sg;
 409
 410         err = mlx4_mr_enable(dev->dev, &mr->mmr);
 411         if (err)
 412                 goto err_free_pl;
 413
 414         mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
 415         mr->umem = NULL;
 416
 417         return &mr->ibmr;
 418
 419 err_free_pl:
 420         mlx4_free_priv_pages(mr);
 421 err_free_mr:
 422         (void) mlx4_mr_free(dev->dev, &mr->mmr);
 423 err_free:
 424         kfree(mr);
 425         return ERR_PTR(err);
 426 }
 427
 428 struct ib_fmr *mlx4_ib_fmr_alloc(struct ib_pd *pd, int acc,
 429                                  struct ib_fmr_attr *fmr_attr)
 430 {
 431         struct mlx4_ib_dev *dev = to_mdev(pd->device);
 432         struct mlx4_ib_fmr *fmr;
 433         int err = -ENOMEM;
 434
 435         fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
 436         if (!fmr)
 437                 return ERR_PTR(-ENOMEM);
 438
 439         err = mlx4_fmr_alloc(dev->dev, to_mpd(pd)->pdn, convert_access(acc),
 440                              fmr_attr->max_pages, fmr_attr->max_maps,
 441                              fmr_attr->page_shift, &fmr->mfmr);
 442         if (err)
 443                 goto err_free;
 444
 445         err = mlx4_fmr_enable(to_mdev(pd->device)->dev, &fmr->mfmr);
 446         if (err)
 447                 goto err_mr;
 448
 449         fmr->ibfmr.rkey = fmr->ibfmr.lkey = fmr->mfmr.mr.key;
 450
 451         return &fmr->ibfmr;
 452
 453 err_mr:
 454         (void) mlx4_mr_free(to_mdev(pd->device)->dev, &fmr->mfmr.mr);
 455
 456 err_free:
 457         kfree(fmr);
 458
 459         return ERR_PTR(err);
 460 }
 461
 462 int mlx4_ib_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list,
 463                       int npages, u64 iova)
 464 {
 465         struct mlx4_ib_fmr *ifmr = to_mfmr(ibfmr);
 466         struct mlx4_ib_dev *dev = to_mdev(ifmr->ibfmr.device);
 467
 468         return mlx4_map_phys_fmr(dev->dev, &ifmr->mfmr, page_list, npages, iova,
 469                                  &ifmr->ibfmr.lkey, &ifmr->ibfmr.rkey);
 470 }
 471
 472 int mlx4_ib_unmap_fmr(struct list_head *fmr_list)
 473 {
 474         struct ib_fmr *ibfmr;
 475         int err;
 476         struct mlx4_dev *mdev = NULL;
 477
 478         list_for_each_entry(ibfmr, fmr_list, list) {
 479                 if (mdev && to_mdev(ibfmr->device)->dev != mdev)
 480                         return -EINVAL;
 481                 mdev = to_mdev(ibfmr->device)->dev;
 482         }
 483
 484         if (!mdev)
 485                 return 0;
 486
 487         list_for_each_entry(ibfmr, fmr_list, list) {
 488                 struct mlx4_ib_fmr *ifmr = to_mfmr(ibfmr);
 489
 490                 mlx4_fmr_unmap(mdev, &ifmr->mfmr, &ifmr->ibfmr.lkey, &ifmr->ibfmr.rkey);
 491         }
 492
 493         /*
 494          * Make sure all MPT status updates are visible before issuing
 495          * SYNC_TPT firmware command.
 496          */
 497         wmb();
 498
 499         err = mlx4_SYNC_TPT(mdev);
 500         if (err)
 501                 pr_warn("SYNC_TPT error %d when "
 502                        "unmapping FMRs\n", err);
 503
 504         return 0;
 505 }
 506
 507 int mlx4_ib_fmr_dealloc(struct ib_fmr *ibfmr)
 508 {
 509         struct mlx4_ib_fmr *ifmr = to_mfmr(ibfmr);
 510         struct mlx4_ib_dev *dev = to_mdev(ibfmr->device);
 511         int err;
 512
 513         err = mlx4_fmr_free(dev->dev, &ifmr->mfmr);
 514
 515         if (!err)
 516                 kfree(ifmr);
 517
 518         return err;
 519 }
 520
 521 static int mlx4_set_page(struct ib_mr *ibmr, u64 addr)
 522 {
 523         struct mlx4_ib_mr *mr = to_mmr(ibmr);
 524
 525         if (unlikely(mr->npages == mr->max_pages))
 526                 return -ENOMEM;
 527
 528         mr->pages[mr->npages++] = cpu_to_be64(addr | MLX4_MTT_FLAG_PRESENT);
 529
 530         return 0;
 531 }
 532
 533 int mlx4_ib_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
 534                       unsigned int *sg_offset)
 535 {
 536         struct mlx4_ib_mr *mr = to_mmr(ibmr);
 537         int rc;
 538
 539         mr->npages = 0;
 540
 541         ib_dma_sync_single_for_cpu(ibmr->device, mr->page_map,
 542                                    mr->page_map_size, DMA_TO_DEVICE);
 543
 544         rc = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, mlx4_set_page);
 545
 546         ib_dma_sync_single_for_device(ibmr->device, mr->page_map,
 547                                       mr->page_map_size, DMA_TO_DEVICE);
 548
 549         return rc;
 550 }