1 /* 2 * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved. 3 * 4 * This software is available to you under a choice of one of two 5 * licenses. You may choose to be licensed under the terms of the GNU 6 * General Public License (GPL) Version 2, available from the file 7 * COPYING in the main directory of this source tree, or the 8 * OpenIB.org BSD license below: 9 * 10 * Redistribution and use in source and binary forms, with or 11 * without modification, are permitted provided that the following 12 * conditions are met: 13 * 14 * - Redistributions of source code must retain the above 15 * copyright notice, this list of conditions and the following 16 * disclaimer. 17 * 18 * - Redistributions in binary form must reproduce the above 19 * copyright notice, this list of conditions and the following 20 * disclaimer in the documentation and/or other materials 21 * provided with the distribution. 22 * 23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 30 * SOFTWARE. 31 */ 32 33 #include <rdma/ib_umem_odp.h> 34 #include <linux/kernel.h> 35 #include <linux/dma-buf.h> 36 #include <linux/dma-resv.h> 37 38 #include "mlx5_ib.h" 39 #include "cmd.h" 40 #include "umr.h" 41 #include "qp.h" 42 43 #include <linux/mlx5/eq.h> 44 45 /* Contains the details of a pagefault. */ 46 struct mlx5_pagefault { 47 u32 bytes_committed; 48 u32 token; 49 u8 event_subtype; 50 u8 type; 51 union { 52 /* Initiator or send message responder pagefault details. */ 53 struct { 54 /* Received packet size, only valid for responders. */ 55 u32 packet_size; 56 /* 57 * Number of resource holding WQE, depends on type. 58 */ 59 u32 wq_num; 60 /* 61 * WQE index. Refers to either the send queue or 62 * receive queue, according to event_subtype. 63 */ 64 u16 wqe_index; 65 } wqe; 66 /* RDMA responder pagefault details */ 67 struct { 68 u32 r_key; 69 /* 70 * Received packet size, minimal size page fault 71 * resolution required for forward progress. 72 */ 73 u32 packet_size; 74 u32 rdma_op_len; 75 u64 rdma_va; 76 } rdma; 77 }; 78 79 struct mlx5_ib_pf_eq *eq; 80 struct work_struct work; 81 }; 82 83 #define MAX_PREFETCH_LEN (4*1024*1024U) 84 85 /* Timeout in ms to wait for an active mmu notifier to complete when handling 86 * a pagefault. */ 87 #define MMU_NOTIFIER_TIMEOUT 1000 88 89 #define MLX5_IMR_MTT_BITS (30 - PAGE_SHIFT) 90 #define MLX5_IMR_MTT_SHIFT (MLX5_IMR_MTT_BITS + PAGE_SHIFT) 91 #define MLX5_IMR_MTT_ENTRIES BIT_ULL(MLX5_IMR_MTT_BITS) 92 #define MLX5_IMR_MTT_SIZE BIT_ULL(MLX5_IMR_MTT_SHIFT) 93 #define MLX5_IMR_MTT_MASK (~(MLX5_IMR_MTT_SIZE - 1)) 94 95 #define MLX5_KSM_PAGE_SHIFT MLX5_IMR_MTT_SHIFT 96 97 static u64 mlx5_imr_ksm_entries; 98 99 static void populate_klm(struct mlx5_klm *pklm, size_t idx, size_t nentries, 100 struct mlx5_ib_mr *imr, int flags) 101 { 102 struct mlx5_klm *end = pklm + nentries; 103 104 if (flags & MLX5_IB_UPD_XLT_ZAP) { 105 for (; pklm != end; pklm++, idx++) { 106 pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE); 107 pklm->key = cpu_to_be32(mr_to_mdev(imr)->null_mkey); 108 pklm->va = 0; 109 } 110 return; 111 } 112 113 /* 114 * The locking here is pretty subtle. Ideally the implicit_children 115 * xarray would be protected by the umem_mutex, however that is not 116 * possible. Instead this uses a weaker update-then-lock pattern: 117 * 118 * xa_store() 119 * mutex_lock(umem_mutex) 120 * mlx5r_umr_update_xlt() 121 * mutex_unlock(umem_mutex) 122 * destroy lkey 123 * 124 * ie any change the xarray must be followed by the locked update_xlt 125 * before destroying. 126 * 127 * The umem_mutex provides the acquire/release semantic needed to make 128 * the xa_store() visible to a racing thread. 129 */ 130 lockdep_assert_held(&to_ib_umem_odp(imr->umem)->umem_mutex); 131 132 for (; pklm != end; pklm++, idx++) { 133 struct mlx5_ib_mr *mtt = xa_load(&imr->implicit_children, idx); 134 135 pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE); 136 if (mtt) { 137 pklm->key = cpu_to_be32(mtt->ibmr.lkey); 138 pklm->va = cpu_to_be64(idx * MLX5_IMR_MTT_SIZE); 139 } else { 140 pklm->key = cpu_to_be32(mr_to_mdev(imr)->null_mkey); 141 pklm->va = 0; 142 } 143 } 144 } 145 146 static u64 umem_dma_to_mtt(dma_addr_t umem_dma) 147 { 148 u64 mtt_entry = umem_dma & ODP_DMA_ADDR_MASK; 149 150 if (umem_dma & ODP_READ_ALLOWED_BIT) 151 mtt_entry |= MLX5_IB_MTT_READ; 152 if (umem_dma & ODP_WRITE_ALLOWED_BIT) 153 mtt_entry |= MLX5_IB_MTT_WRITE; 154 155 return mtt_entry; 156 } 157 158 static void populate_mtt(__be64 *pas, size_t idx, size_t nentries, 159 struct mlx5_ib_mr *mr, int flags) 160 { 161 struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem); 162 dma_addr_t pa; 163 size_t i; 164 165 if (flags & MLX5_IB_UPD_XLT_ZAP) 166 return; 167 168 for (i = 0; i < nentries; i++) { 169 pa = odp->dma_list[idx + i]; 170 pas[i] = cpu_to_be64(umem_dma_to_mtt(pa)); 171 } 172 } 173 174 void mlx5_odp_populate_xlt(void *xlt, size_t idx, size_t nentries, 175 struct mlx5_ib_mr *mr, int flags) 176 { 177 if (flags & MLX5_IB_UPD_XLT_INDIRECT) { 178 populate_klm(xlt, idx, nentries, mr, flags); 179 } else { 180 populate_mtt(xlt, idx, nentries, mr, flags); 181 } 182 } 183 184 /* 185 * This must be called after the mr has been removed from implicit_children. 186 * NOTE: The MR does not necessarily have to be 187 * empty here, parallel page faults could have raced with the free process and 188 * added pages to it. 189 */ 190 static void free_implicit_child_mr_work(struct work_struct *work) 191 { 192 struct mlx5_ib_mr *mr = 193 container_of(work, struct mlx5_ib_mr, odp_destroy.work); 194 struct mlx5_ib_mr *imr = mr->parent; 195 struct ib_umem_odp *odp_imr = to_ib_umem_odp(imr->umem); 196 struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem); 197 198 mlx5r_deref_wait_odp_mkey(&mr->mmkey); 199 200 mutex_lock(&odp_imr->umem_mutex); 201 mlx5r_umr_update_xlt(mr->parent, 202 ib_umem_start(odp) >> MLX5_IMR_MTT_SHIFT, 1, 0, 203 MLX5_IB_UPD_XLT_INDIRECT | MLX5_IB_UPD_XLT_ATOMIC); 204 mutex_unlock(&odp_imr->umem_mutex); 205 mlx5_ib_dereg_mr(&mr->ibmr, NULL); 206 207 mlx5r_deref_odp_mkey(&imr->mmkey); 208 } 209 210 static void destroy_unused_implicit_child_mr(struct mlx5_ib_mr *mr) 211 { 212 struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem); 213 unsigned long idx = ib_umem_start(odp) >> MLX5_IMR_MTT_SHIFT; 214 struct mlx5_ib_mr *imr = mr->parent; 215 216 if (!refcount_inc_not_zero(&imr->mmkey.usecount)) 217 return; 218 219 xa_erase(&imr->implicit_children, idx); 220 221 /* Freeing a MR is a sleeping operation, so bounce to a work queue */ 222 INIT_WORK(&mr->odp_destroy.work, free_implicit_child_mr_work); 223 queue_work(system_unbound_wq, &mr->odp_destroy.work); 224 } 225 226 static bool mlx5_ib_invalidate_range(struct mmu_interval_notifier *mni, 227 const struct mmu_notifier_range *range, 228 unsigned long cur_seq) 229 { 230 struct ib_umem_odp *umem_odp = 231 container_of(mni, struct ib_umem_odp, notifier); 232 struct mlx5_ib_mr *mr; 233 const u64 umr_block_mask = MLX5_UMR_MTT_NUM_ENTRIES_ALIGNMENT - 1; 234 u64 idx = 0, blk_start_idx = 0; 235 u64 invalidations = 0; 236 unsigned long start; 237 unsigned long end; 238 int in_block = 0; 239 u64 addr; 240 241 if (!mmu_notifier_range_blockable(range)) 242 return false; 243 244 mutex_lock(&umem_odp->umem_mutex); 245 mmu_interval_set_seq(mni, cur_seq); 246 /* 247 * If npages is zero then umem_odp->private may not be setup yet. This 248 * does not complete until after the first page is mapped for DMA. 249 */ 250 if (!umem_odp->npages) 251 goto out; 252 mr = umem_odp->private; 253 254 start = max_t(u64, ib_umem_start(umem_odp), range->start); 255 end = min_t(u64, ib_umem_end(umem_odp), range->end); 256 257 /* 258 * Iteration one - zap the HW's MTTs. The notifiers_count ensures that 259 * while we are doing the invalidation, no page fault will attempt to 260 * overwrite the same MTTs. Concurent invalidations might race us, 261 * but they will write 0s as well, so no difference in the end result. 262 */ 263 for (addr = start; addr < end; addr += BIT(umem_odp->page_shift)) { 264 idx = (addr - ib_umem_start(umem_odp)) >> umem_odp->page_shift; 265 /* 266 * Strive to write the MTTs in chunks, but avoid overwriting 267 * non-existing MTTs. The huristic here can be improved to 268 * estimate the cost of another UMR vs. the cost of bigger 269 * UMR. 270 */ 271 if (umem_odp->dma_list[idx] & 272 (ODP_READ_ALLOWED_BIT | ODP_WRITE_ALLOWED_BIT)) { 273 if (!in_block) { 274 blk_start_idx = idx; 275 in_block = 1; 276 } 277 278 /* Count page invalidations */ 279 invalidations += idx - blk_start_idx + 1; 280 } else { 281 u64 umr_offset = idx & umr_block_mask; 282 283 if (in_block && umr_offset == 0) { 284 mlx5r_umr_update_xlt(mr, blk_start_idx, 285 idx - blk_start_idx, 0, 286 MLX5_IB_UPD_XLT_ZAP | 287 MLX5_IB_UPD_XLT_ATOMIC); 288 in_block = 0; 289 } 290 } 291 } 292 if (in_block) 293 mlx5r_umr_update_xlt(mr, blk_start_idx, 294 idx - blk_start_idx + 1, 0, 295 MLX5_IB_UPD_XLT_ZAP | 296 MLX5_IB_UPD_XLT_ATOMIC); 297 298 mlx5_update_odp_stats(mr, invalidations, invalidations); 299 300 /* 301 * We are now sure that the device will not access the 302 * memory. We can safely unmap it, and mark it as dirty if 303 * needed. 304 */ 305 306 ib_umem_odp_unmap_dma_pages(umem_odp, start, end); 307 308 if (unlikely(!umem_odp->npages && mr->parent)) 309 destroy_unused_implicit_child_mr(mr); 310 out: 311 mutex_unlock(&umem_odp->umem_mutex); 312 return true; 313 } 314 315 const struct mmu_interval_notifier_ops mlx5_mn_ops = { 316 .invalidate = mlx5_ib_invalidate_range, 317 }; 318 319 static void internal_fill_odp_caps(struct mlx5_ib_dev *dev) 320 { 321 struct ib_odp_caps *caps = &dev->odp_caps; 322 323 memset(caps, 0, sizeof(*caps)); 324 325 if (!MLX5_CAP_GEN(dev->mdev, pg) || !mlx5r_umr_can_load_pas(dev, 0)) 326 return; 327 328 caps->general_caps = IB_ODP_SUPPORT; 329 330 if (MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset)) 331 dev->odp_max_size = U64_MAX; 332 else 333 dev->odp_max_size = BIT_ULL(MLX5_MAX_UMR_SHIFT + PAGE_SHIFT); 334 335 if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.send)) 336 caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SEND; 337 338 if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.srq_receive)) 339 caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV; 340 341 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.send)) 342 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SEND; 343 344 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.receive)) 345 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_RECV; 346 347 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.write)) 348 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_WRITE; 349 350 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.read)) 351 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_READ; 352 353 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.atomic)) 354 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC; 355 356 if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.srq_receive)) 357 caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV; 358 359 if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.send)) 360 caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_SEND; 361 362 if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.receive)) 363 caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_RECV; 364 365 if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.write)) 366 caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_WRITE; 367 368 if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.read)) 369 caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_READ; 370 371 if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.atomic)) 372 caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_ATOMIC; 373 374 if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.srq_receive)) 375 caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV; 376 377 if (MLX5_CAP_GEN(dev->mdev, fixed_buffer_size) && 378 MLX5_CAP_GEN(dev->mdev, null_mkey) && 379 MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset) && 380 !MLX5_CAP_GEN(dev->mdev, umr_indirect_mkey_disabled)) 381 caps->general_caps |= IB_ODP_SUPPORT_IMPLICIT; 382 } 383 384 static void mlx5_ib_page_fault_resume(struct mlx5_ib_dev *dev, 385 struct mlx5_pagefault *pfault, 386 int error) 387 { 388 int wq_num = pfault->event_subtype == MLX5_PFAULT_SUBTYPE_WQE ? 389 pfault->wqe.wq_num : pfault->token; 390 u32 in[MLX5_ST_SZ_DW(page_fault_resume_in)] = {}; 391 int err; 392 393 MLX5_SET(page_fault_resume_in, in, opcode, MLX5_CMD_OP_PAGE_FAULT_RESUME); 394 MLX5_SET(page_fault_resume_in, in, page_fault_type, pfault->type); 395 MLX5_SET(page_fault_resume_in, in, token, pfault->token); 396 MLX5_SET(page_fault_resume_in, in, wq_number, wq_num); 397 MLX5_SET(page_fault_resume_in, in, error, !!error); 398 399 err = mlx5_cmd_exec_in(dev->mdev, page_fault_resume, in); 400 if (err) 401 mlx5_ib_err(dev, "Failed to resolve the page fault on WQ 0x%x err %d\n", 402 wq_num, err); 403 } 404 405 static struct mlx5_ib_mr *implicit_get_child_mr(struct mlx5_ib_mr *imr, 406 unsigned long idx) 407 { 408 struct mlx5_ib_dev *dev = mr_to_mdev(imr); 409 struct ib_umem_odp *odp; 410 struct mlx5_ib_mr *mr; 411 struct mlx5_ib_mr *ret; 412 int err; 413 414 odp = ib_umem_odp_alloc_child(to_ib_umem_odp(imr->umem), 415 idx * MLX5_IMR_MTT_SIZE, 416 MLX5_IMR_MTT_SIZE, &mlx5_mn_ops); 417 if (IS_ERR(odp)) 418 return ERR_CAST(odp); 419 420 mr = mlx5_mr_cache_alloc(dev, &dev->cache.ent[MLX5_IMR_MTT_CACHE_ENTRY], 421 imr->access_flags); 422 if (IS_ERR(mr)) { 423 ib_umem_odp_release(odp); 424 return mr; 425 } 426 427 mr->access_flags = imr->access_flags; 428 mr->ibmr.pd = imr->ibmr.pd; 429 mr->ibmr.device = &mr_to_mdev(imr)->ib_dev; 430 mr->umem = &odp->umem; 431 mr->ibmr.lkey = mr->mmkey.key; 432 mr->ibmr.rkey = mr->mmkey.key; 433 mr->ibmr.iova = idx * MLX5_IMR_MTT_SIZE; 434 mr->parent = imr; 435 odp->private = mr; 436 437 /* 438 * First refcount is owned by the xarray and second refconut 439 * is returned to the caller. 440 */ 441 refcount_set(&mr->mmkey.usecount, 2); 442 443 err = mlx5r_umr_update_xlt(mr, 0, 444 MLX5_IMR_MTT_ENTRIES, 445 PAGE_SHIFT, 446 MLX5_IB_UPD_XLT_ZAP | 447 MLX5_IB_UPD_XLT_ENABLE); 448 if (err) { 449 ret = ERR_PTR(err); 450 goto out_mr; 451 } 452 453 xa_lock(&imr->implicit_children); 454 ret = __xa_cmpxchg(&imr->implicit_children, idx, NULL, mr, 455 GFP_KERNEL); 456 if (unlikely(ret)) { 457 if (xa_is_err(ret)) { 458 ret = ERR_PTR(xa_err(ret)); 459 goto out_lock; 460 } 461 /* 462 * Another thread beat us to creating the child mr, use 463 * theirs. 464 */ 465 refcount_inc(&ret->mmkey.usecount); 466 goto out_lock; 467 } 468 xa_unlock(&imr->implicit_children); 469 470 mlx5_ib_dbg(mr_to_mdev(imr), "key %x mr %p\n", mr->mmkey.key, mr); 471 return mr; 472 473 out_lock: 474 xa_unlock(&imr->implicit_children); 475 out_mr: 476 mlx5_ib_dereg_mr(&mr->ibmr, NULL); 477 return ret; 478 } 479 480 struct mlx5_ib_mr *mlx5_ib_alloc_implicit_mr(struct mlx5_ib_pd *pd, 481 int access_flags) 482 { 483 struct mlx5_ib_dev *dev = to_mdev(pd->ibpd.device); 484 struct ib_umem_odp *umem_odp; 485 struct mlx5_ib_mr *imr; 486 int err; 487 488 if (!mlx5r_umr_can_load_pas(dev, MLX5_IMR_MTT_ENTRIES * PAGE_SIZE)) 489 return ERR_PTR(-EOPNOTSUPP); 490 491 umem_odp = ib_umem_odp_alloc_implicit(&dev->ib_dev, access_flags); 492 if (IS_ERR(umem_odp)) 493 return ERR_CAST(umem_odp); 494 495 imr = mlx5_mr_cache_alloc(dev, 496 &dev->cache.ent[MLX5_IMR_KSM_CACHE_ENTRY], 497 access_flags); 498 if (IS_ERR(imr)) { 499 ib_umem_odp_release(umem_odp); 500 return imr; 501 } 502 503 imr->access_flags = access_flags; 504 imr->ibmr.pd = &pd->ibpd; 505 imr->ibmr.iova = 0; 506 imr->umem = &umem_odp->umem; 507 imr->ibmr.lkey = imr->mmkey.key; 508 imr->ibmr.rkey = imr->mmkey.key; 509 imr->ibmr.device = &dev->ib_dev; 510 imr->is_odp_implicit = true; 511 xa_init(&imr->implicit_children); 512 513 err = mlx5r_umr_update_xlt(imr, 0, 514 mlx5_imr_ksm_entries, 515 MLX5_KSM_PAGE_SHIFT, 516 MLX5_IB_UPD_XLT_INDIRECT | 517 MLX5_IB_UPD_XLT_ZAP | 518 MLX5_IB_UPD_XLT_ENABLE); 519 if (err) 520 goto out_mr; 521 522 err = mlx5r_store_odp_mkey(dev, &imr->mmkey); 523 if (err) 524 goto out_mr; 525 526 mlx5_ib_dbg(dev, "key %x mr %p\n", imr->mmkey.key, imr); 527 return imr; 528 out_mr: 529 mlx5_ib_err(dev, "Failed to register MKEY %d\n", err); 530 mlx5_ib_dereg_mr(&imr->ibmr, NULL); 531 return ERR_PTR(err); 532 } 533 534 void mlx5_ib_free_odp_mr(struct mlx5_ib_mr *mr) 535 { 536 struct mlx5_ib_mr *mtt; 537 unsigned long idx; 538 539 /* 540 * If this is an implicit MR it is already invalidated so we can just 541 * delete the children mkeys. 542 */ 543 xa_for_each(&mr->implicit_children, idx, mtt) { 544 xa_erase(&mr->implicit_children, idx); 545 mlx5_ib_dereg_mr(&mtt->ibmr, NULL); 546 } 547 } 548 549 #define MLX5_PF_FLAGS_DOWNGRADE BIT(1) 550 #define MLX5_PF_FLAGS_SNAPSHOT BIT(2) 551 #define MLX5_PF_FLAGS_ENABLE BIT(3) 552 static int pagefault_real_mr(struct mlx5_ib_mr *mr, struct ib_umem_odp *odp, 553 u64 user_va, size_t bcnt, u32 *bytes_mapped, 554 u32 flags) 555 { 556 int page_shift, ret, np; 557 bool downgrade = flags & MLX5_PF_FLAGS_DOWNGRADE; 558 u64 access_mask; 559 u64 start_idx; 560 bool fault = !(flags & MLX5_PF_FLAGS_SNAPSHOT); 561 u32 xlt_flags = MLX5_IB_UPD_XLT_ATOMIC; 562 563 if (flags & MLX5_PF_FLAGS_ENABLE) 564 xlt_flags |= MLX5_IB_UPD_XLT_ENABLE; 565 566 page_shift = odp->page_shift; 567 start_idx = (user_va - ib_umem_start(odp)) >> page_shift; 568 access_mask = ODP_READ_ALLOWED_BIT; 569 570 if (odp->umem.writable && !downgrade) 571 access_mask |= ODP_WRITE_ALLOWED_BIT; 572 573 np = ib_umem_odp_map_dma_and_lock(odp, user_va, bcnt, access_mask, fault); 574 if (np < 0) 575 return np; 576 577 /* 578 * No need to check whether the MTTs really belong to this MR, since 579 * ib_umem_odp_map_dma_and_lock already checks this. 580 */ 581 ret = mlx5r_umr_update_xlt(mr, start_idx, np, page_shift, xlt_flags); 582 mutex_unlock(&odp->umem_mutex); 583 584 if (ret < 0) { 585 if (ret != -EAGAIN) 586 mlx5_ib_err(mr_to_mdev(mr), 587 "Failed to update mkey page tables\n"); 588 goto out; 589 } 590 591 if (bytes_mapped) { 592 u32 new_mappings = (np << page_shift) - 593 (user_va - round_down(user_va, 1 << page_shift)); 594 595 *bytes_mapped += min_t(u32, new_mappings, bcnt); 596 } 597 598 return np << (page_shift - PAGE_SHIFT); 599 600 out: 601 return ret; 602 } 603 604 static int pagefault_implicit_mr(struct mlx5_ib_mr *imr, 605 struct ib_umem_odp *odp_imr, u64 user_va, 606 size_t bcnt, u32 *bytes_mapped, u32 flags) 607 { 608 unsigned long end_idx = (user_va + bcnt - 1) >> MLX5_IMR_MTT_SHIFT; 609 unsigned long upd_start_idx = end_idx + 1; 610 unsigned long upd_len = 0; 611 unsigned long npages = 0; 612 int err; 613 int ret; 614 615 if (unlikely(user_va >= mlx5_imr_ksm_entries * MLX5_IMR_MTT_SIZE || 616 mlx5_imr_ksm_entries * MLX5_IMR_MTT_SIZE - user_va < bcnt)) 617 return -EFAULT; 618 619 /* Fault each child mr that intersects with our interval. */ 620 while (bcnt) { 621 unsigned long idx = user_va >> MLX5_IMR_MTT_SHIFT; 622 struct ib_umem_odp *umem_odp; 623 struct mlx5_ib_mr *mtt; 624 u64 len; 625 626 xa_lock(&imr->implicit_children); 627 mtt = xa_load(&imr->implicit_children, idx); 628 if (unlikely(!mtt)) { 629 xa_unlock(&imr->implicit_children); 630 mtt = implicit_get_child_mr(imr, idx); 631 if (IS_ERR(mtt)) { 632 ret = PTR_ERR(mtt); 633 goto out; 634 } 635 upd_start_idx = min(upd_start_idx, idx); 636 upd_len = idx - upd_start_idx + 1; 637 } else { 638 refcount_inc(&mtt->mmkey.usecount); 639 xa_unlock(&imr->implicit_children); 640 } 641 642 umem_odp = to_ib_umem_odp(mtt->umem); 643 len = min_t(u64, user_va + bcnt, ib_umem_end(umem_odp)) - 644 user_va; 645 646 ret = pagefault_real_mr(mtt, umem_odp, user_va, len, 647 bytes_mapped, flags); 648 649 mlx5r_deref_odp_mkey(&mtt->mmkey); 650 651 if (ret < 0) 652 goto out; 653 user_va += len; 654 bcnt -= len; 655 npages += ret; 656 } 657 658 ret = npages; 659 660 /* 661 * Any time the implicit_children are changed we must perform an 662 * update of the xlt before exiting to ensure the HW and the 663 * implicit_children remains synchronized. 664 */ 665 out: 666 if (likely(!upd_len)) 667 return ret; 668 669 /* 670 * Notice this is not strictly ordered right, the KSM is updated after 671 * the implicit_children is updated, so a parallel page fault could 672 * see a MR that is not yet visible in the KSM. This is similar to a 673 * parallel page fault seeing a MR that is being concurrently removed 674 * from the KSM. Both of these improbable situations are resolved 675 * safely by resuming the HW and then taking another page fault. The 676 * next pagefault handler will see the new information. 677 */ 678 mutex_lock(&odp_imr->umem_mutex); 679 err = mlx5r_umr_update_xlt(imr, upd_start_idx, upd_len, 0, 680 MLX5_IB_UPD_XLT_INDIRECT | 681 MLX5_IB_UPD_XLT_ATOMIC); 682 mutex_unlock(&odp_imr->umem_mutex); 683 if (err) { 684 mlx5_ib_err(mr_to_mdev(imr), "Failed to update PAS\n"); 685 return err; 686 } 687 return ret; 688 } 689 690 static int pagefault_dmabuf_mr(struct mlx5_ib_mr *mr, size_t bcnt, 691 u32 *bytes_mapped, u32 flags) 692 { 693 struct ib_umem_dmabuf *umem_dmabuf = to_ib_umem_dmabuf(mr->umem); 694 u32 xlt_flags = 0; 695 int err; 696 unsigned int page_size; 697 698 if (flags & MLX5_PF_FLAGS_ENABLE) 699 xlt_flags |= MLX5_IB_UPD_XLT_ENABLE; 700 701 dma_resv_lock(umem_dmabuf->attach->dmabuf->resv, NULL); 702 err = ib_umem_dmabuf_map_pages(umem_dmabuf); 703 if (err) { 704 dma_resv_unlock(umem_dmabuf->attach->dmabuf->resv); 705 return err; 706 } 707 708 page_size = mlx5_umem_find_best_pgsz(&umem_dmabuf->umem, mkc, 709 log_page_size, 0, 710 umem_dmabuf->umem.iova); 711 if (unlikely(page_size < PAGE_SIZE)) { 712 ib_umem_dmabuf_unmap_pages(umem_dmabuf); 713 err = -EINVAL; 714 } else { 715 err = mlx5r_umr_update_mr_pas(mr, xlt_flags); 716 } 717 dma_resv_unlock(umem_dmabuf->attach->dmabuf->resv); 718 719 if (err) 720 return err; 721 722 if (bytes_mapped) 723 *bytes_mapped += bcnt; 724 725 return ib_umem_num_pages(mr->umem); 726 } 727 728 /* 729 * Returns: 730 * -EFAULT: The io_virt->bcnt is not within the MR, it covers pages that are 731 * not accessible, or the MR is no longer valid. 732 * -EAGAIN/-ENOMEM: The operation should be retried 733 * 734 * -EINVAL/others: General internal malfunction 735 * >0: Number of pages mapped 736 */ 737 static int pagefault_mr(struct mlx5_ib_mr *mr, u64 io_virt, size_t bcnt, 738 u32 *bytes_mapped, u32 flags) 739 { 740 struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem); 741 742 if (unlikely(io_virt < mr->ibmr.iova)) 743 return -EFAULT; 744 745 if (mr->umem->is_dmabuf) 746 return pagefault_dmabuf_mr(mr, bcnt, bytes_mapped, flags); 747 748 if (!odp->is_implicit_odp) { 749 u64 user_va; 750 751 if (check_add_overflow(io_virt - mr->ibmr.iova, 752 (u64)odp->umem.address, &user_va)) 753 return -EFAULT; 754 if (unlikely(user_va >= ib_umem_end(odp) || 755 ib_umem_end(odp) - user_va < bcnt)) 756 return -EFAULT; 757 return pagefault_real_mr(mr, odp, user_va, bcnt, bytes_mapped, 758 flags); 759 } 760 return pagefault_implicit_mr(mr, odp, io_virt, bcnt, bytes_mapped, 761 flags); 762 } 763 764 int mlx5_ib_init_odp_mr(struct mlx5_ib_mr *mr) 765 { 766 int ret; 767 768 ret = pagefault_real_mr(mr, to_ib_umem_odp(mr->umem), mr->umem->address, 769 mr->umem->length, NULL, 770 MLX5_PF_FLAGS_SNAPSHOT | MLX5_PF_FLAGS_ENABLE); 771 return ret >= 0 ? 0 : ret; 772 } 773 774 int mlx5_ib_init_dmabuf_mr(struct mlx5_ib_mr *mr) 775 { 776 int ret; 777 778 ret = pagefault_dmabuf_mr(mr, mr->umem->length, NULL, 779 MLX5_PF_FLAGS_ENABLE); 780 781 return ret >= 0 ? 0 : ret; 782 } 783 784 struct pf_frame { 785 struct pf_frame *next; 786 u32 key; 787 u64 io_virt; 788 size_t bcnt; 789 int depth; 790 }; 791 792 static bool mkey_is_eq(struct mlx5_ib_mkey *mmkey, u32 key) 793 { 794 if (!mmkey) 795 return false; 796 if (mmkey->type == MLX5_MKEY_MW || 797 mmkey->type == MLX5_MKEY_INDIRECT_DEVX) 798 return mlx5_base_mkey(mmkey->key) == mlx5_base_mkey(key); 799 return mmkey->key == key; 800 } 801 802 /* 803 * Handle a single data segment in a page-fault WQE or RDMA region. 804 * 805 * Returns number of OS pages retrieved on success. The caller may continue to 806 * the next data segment. 807 * Can return the following error codes: 808 * -EAGAIN to designate a temporary error. The caller will abort handling the 809 * page fault and resolve it. 810 * -EFAULT when there's an error mapping the requested pages. The caller will 811 * abort the page fault handling. 812 */ 813 static int pagefault_single_data_segment(struct mlx5_ib_dev *dev, 814 struct ib_pd *pd, u32 key, 815 u64 io_virt, size_t bcnt, 816 u32 *bytes_committed, 817 u32 *bytes_mapped) 818 { 819 int npages = 0, ret, i, outlen, cur_outlen = 0, depth = 0; 820 struct pf_frame *head = NULL, *frame; 821 struct mlx5_ib_mkey *mmkey; 822 struct mlx5_ib_mr *mr; 823 struct mlx5_klm *pklm; 824 u32 *out = NULL; 825 size_t offset; 826 827 io_virt += *bytes_committed; 828 bcnt -= *bytes_committed; 829 830 next_mr: 831 xa_lock(&dev->odp_mkeys); 832 mmkey = xa_load(&dev->odp_mkeys, mlx5_base_mkey(key)); 833 if (!mmkey) { 834 xa_unlock(&dev->odp_mkeys); 835 mlx5_ib_dbg( 836 dev, 837 "skipping non ODP MR (lkey=0x%06x) in page fault handler.\n", 838 key); 839 if (bytes_mapped) 840 *bytes_mapped += bcnt; 841 /* 842 * The user could specify a SGL with multiple lkeys and only 843 * some of them are ODP. Treat the non-ODP ones as fully 844 * faulted. 845 */ 846 ret = 0; 847 goto end; 848 } 849 refcount_inc(&mmkey->usecount); 850 xa_unlock(&dev->odp_mkeys); 851 852 if (!mkey_is_eq(mmkey, key)) { 853 mlx5_ib_dbg(dev, "failed to find mkey %x\n", key); 854 ret = -EFAULT; 855 goto end; 856 } 857 858 switch (mmkey->type) { 859 case MLX5_MKEY_MR: 860 mr = container_of(mmkey, struct mlx5_ib_mr, mmkey); 861 862 ret = pagefault_mr(mr, io_virt, bcnt, bytes_mapped, 0); 863 if (ret < 0) 864 goto end; 865 866 mlx5_update_odp_stats(mr, faults, ret); 867 868 npages += ret; 869 ret = 0; 870 break; 871 872 case MLX5_MKEY_MW: 873 case MLX5_MKEY_INDIRECT_DEVX: 874 if (depth >= MLX5_CAP_GEN(dev->mdev, max_indirection)) { 875 mlx5_ib_dbg(dev, "indirection level exceeded\n"); 876 ret = -EFAULT; 877 goto end; 878 } 879 880 outlen = MLX5_ST_SZ_BYTES(query_mkey_out) + 881 sizeof(*pklm) * (mmkey->ndescs - 2); 882 883 if (outlen > cur_outlen) { 884 kfree(out); 885 out = kzalloc(outlen, GFP_KERNEL); 886 if (!out) { 887 ret = -ENOMEM; 888 goto end; 889 } 890 cur_outlen = outlen; 891 } 892 893 pklm = (struct mlx5_klm *)MLX5_ADDR_OF(query_mkey_out, out, 894 bsf0_klm0_pas_mtt0_1); 895 896 ret = mlx5_core_query_mkey(dev->mdev, mmkey->key, out, outlen); 897 if (ret) 898 goto end; 899 900 offset = io_virt - MLX5_GET64(query_mkey_out, out, 901 memory_key_mkey_entry.start_addr); 902 903 for (i = 0; bcnt && i < mmkey->ndescs; i++, pklm++) { 904 if (offset >= be32_to_cpu(pklm->bcount)) { 905 offset -= be32_to_cpu(pklm->bcount); 906 continue; 907 } 908 909 frame = kzalloc(sizeof(*frame), GFP_KERNEL); 910 if (!frame) { 911 ret = -ENOMEM; 912 goto end; 913 } 914 915 frame->key = be32_to_cpu(pklm->key); 916 frame->io_virt = be64_to_cpu(pklm->va) + offset; 917 frame->bcnt = min_t(size_t, bcnt, 918 be32_to_cpu(pklm->bcount) - offset); 919 frame->depth = depth + 1; 920 frame->next = head; 921 head = frame; 922 923 bcnt -= frame->bcnt; 924 offset = 0; 925 } 926 break; 927 928 default: 929 mlx5_ib_dbg(dev, "wrong mkey type %d\n", mmkey->type); 930 ret = -EFAULT; 931 goto end; 932 } 933 934 if (head) { 935 frame = head; 936 head = frame->next; 937 938 key = frame->key; 939 io_virt = frame->io_virt; 940 bcnt = frame->bcnt; 941 depth = frame->depth; 942 kfree(frame); 943 944 mlx5r_deref_odp_mkey(mmkey); 945 goto next_mr; 946 } 947 948 end: 949 if (mmkey) 950 mlx5r_deref_odp_mkey(mmkey); 951 while (head) { 952 frame = head; 953 head = frame->next; 954 kfree(frame); 955 } 956 kfree(out); 957 958 *bytes_committed = 0; 959 return ret ? ret : npages; 960 } 961 962 /* 963 * Parse a series of data segments for page fault handling. 964 * 965 * @dev: Pointer to mlx5 IB device 966 * @pfault: contains page fault information. 967 * @wqe: points at the first data segment in the WQE. 968 * @wqe_end: points after the end of the WQE. 969 * @bytes_mapped: receives the number of bytes that the function was able to 970 * map. This allows the caller to decide intelligently whether 971 * enough memory was mapped to resolve the page fault 972 * successfully (e.g. enough for the next MTU, or the entire 973 * WQE). 974 * @total_wqe_bytes: receives the total data size of this WQE in bytes (minus 975 * the committed bytes). 976 * @receive_queue: receive WQE end of sg list 977 * 978 * Returns the number of pages loaded if positive, zero for an empty WQE, or a 979 * negative error code. 980 */ 981 static int pagefault_data_segments(struct mlx5_ib_dev *dev, 982 struct mlx5_pagefault *pfault, 983 void *wqe, 984 void *wqe_end, u32 *bytes_mapped, 985 u32 *total_wqe_bytes, bool receive_queue) 986 { 987 int ret = 0, npages = 0; 988 u64 io_virt; 989 u32 key; 990 u32 byte_count; 991 size_t bcnt; 992 int inline_segment; 993 994 if (bytes_mapped) 995 *bytes_mapped = 0; 996 if (total_wqe_bytes) 997 *total_wqe_bytes = 0; 998 999 while (wqe < wqe_end) { 1000 struct mlx5_wqe_data_seg *dseg = wqe; 1001 1002 io_virt = be64_to_cpu(dseg->addr); 1003 key = be32_to_cpu(dseg->lkey); 1004 byte_count = be32_to_cpu(dseg->byte_count); 1005 inline_segment = !!(byte_count & MLX5_INLINE_SEG); 1006 bcnt = byte_count & ~MLX5_INLINE_SEG; 1007 1008 if (inline_segment) { 1009 bcnt = bcnt & MLX5_WQE_INLINE_SEG_BYTE_COUNT_MASK; 1010 wqe += ALIGN(sizeof(struct mlx5_wqe_inline_seg) + bcnt, 1011 16); 1012 } else { 1013 wqe += sizeof(*dseg); 1014 } 1015 1016 /* receive WQE end of sg list. */ 1017 if (receive_queue && bcnt == 0 && key == MLX5_INVALID_LKEY && 1018 io_virt == 0) 1019 break; 1020 1021 if (!inline_segment && total_wqe_bytes) { 1022 *total_wqe_bytes += bcnt - min_t(size_t, bcnt, 1023 pfault->bytes_committed); 1024 } 1025 1026 /* A zero length data segment designates a length of 2GB. */ 1027 if (bcnt == 0) 1028 bcnt = 1U << 31; 1029 1030 if (inline_segment || bcnt <= pfault->bytes_committed) { 1031 pfault->bytes_committed -= 1032 min_t(size_t, bcnt, 1033 pfault->bytes_committed); 1034 continue; 1035 } 1036 1037 ret = pagefault_single_data_segment(dev, NULL, key, 1038 io_virt, bcnt, 1039 &pfault->bytes_committed, 1040 bytes_mapped); 1041 if (ret < 0) 1042 break; 1043 npages += ret; 1044 } 1045 1046 return ret < 0 ? ret : npages; 1047 } 1048 1049 /* 1050 * Parse initiator WQE. Advances the wqe pointer to point at the 1051 * scatter-gather list, and set wqe_end to the end of the WQE. 1052 */ 1053 static int mlx5_ib_mr_initiator_pfault_handler( 1054 struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault, 1055 struct mlx5_ib_qp *qp, void **wqe, void **wqe_end, int wqe_length) 1056 { 1057 struct mlx5_wqe_ctrl_seg *ctrl = *wqe; 1058 u16 wqe_index = pfault->wqe.wqe_index; 1059 struct mlx5_base_av *av; 1060 unsigned ds, opcode; 1061 u32 qpn = qp->trans_qp.base.mqp.qpn; 1062 1063 ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK; 1064 if (ds * MLX5_WQE_DS_UNITS > wqe_length) { 1065 mlx5_ib_err(dev, "Unable to read the complete WQE. ds = 0x%x, ret = 0x%x\n", 1066 ds, wqe_length); 1067 return -EFAULT; 1068 } 1069 1070 if (ds == 0) { 1071 mlx5_ib_err(dev, "Got WQE with zero DS. wqe_index=%x, qpn=%x\n", 1072 wqe_index, qpn); 1073 return -EFAULT; 1074 } 1075 1076 *wqe_end = *wqe + ds * MLX5_WQE_DS_UNITS; 1077 *wqe += sizeof(*ctrl); 1078 1079 opcode = be32_to_cpu(ctrl->opmod_idx_opcode) & 1080 MLX5_WQE_CTRL_OPCODE_MASK; 1081 1082 if (qp->type == IB_QPT_XRC_INI) 1083 *wqe += sizeof(struct mlx5_wqe_xrc_seg); 1084 1085 if (qp->type == IB_QPT_UD || qp->type == MLX5_IB_QPT_DCI) { 1086 av = *wqe; 1087 if (av->dqp_dct & cpu_to_be32(MLX5_EXTENDED_UD_AV)) 1088 *wqe += sizeof(struct mlx5_av); 1089 else 1090 *wqe += sizeof(struct mlx5_base_av); 1091 } 1092 1093 switch (opcode) { 1094 case MLX5_OPCODE_RDMA_WRITE: 1095 case MLX5_OPCODE_RDMA_WRITE_IMM: 1096 case MLX5_OPCODE_RDMA_READ: 1097 *wqe += sizeof(struct mlx5_wqe_raddr_seg); 1098 break; 1099 case MLX5_OPCODE_ATOMIC_CS: 1100 case MLX5_OPCODE_ATOMIC_FA: 1101 *wqe += sizeof(struct mlx5_wqe_raddr_seg); 1102 *wqe += sizeof(struct mlx5_wqe_atomic_seg); 1103 break; 1104 } 1105 1106 return 0; 1107 } 1108 1109 /* 1110 * Parse responder WQE and set wqe_end to the end of the WQE. 1111 */ 1112 static int mlx5_ib_mr_responder_pfault_handler_srq(struct mlx5_ib_dev *dev, 1113 struct mlx5_ib_srq *srq, 1114 void **wqe, void **wqe_end, 1115 int wqe_length) 1116 { 1117 int wqe_size = 1 << srq->msrq.wqe_shift; 1118 1119 if (wqe_size > wqe_length) { 1120 mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n"); 1121 return -EFAULT; 1122 } 1123 1124 *wqe_end = *wqe + wqe_size; 1125 *wqe += sizeof(struct mlx5_wqe_srq_next_seg); 1126 1127 return 0; 1128 } 1129 1130 static int mlx5_ib_mr_responder_pfault_handler_rq(struct mlx5_ib_dev *dev, 1131 struct mlx5_ib_qp *qp, 1132 void *wqe, void **wqe_end, 1133 int wqe_length) 1134 { 1135 struct mlx5_ib_wq *wq = &qp->rq; 1136 int wqe_size = 1 << wq->wqe_shift; 1137 1138 if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE) { 1139 mlx5_ib_err(dev, "ODP fault with WQE signatures is not supported\n"); 1140 return -EFAULT; 1141 } 1142 1143 if (wqe_size > wqe_length) { 1144 mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n"); 1145 return -EFAULT; 1146 } 1147 1148 *wqe_end = wqe + wqe_size; 1149 1150 return 0; 1151 } 1152 1153 static inline struct mlx5_core_rsc_common *odp_get_rsc(struct mlx5_ib_dev *dev, 1154 u32 wq_num, int pf_type) 1155 { 1156 struct mlx5_core_rsc_common *common = NULL; 1157 struct mlx5_core_srq *srq; 1158 1159 switch (pf_type) { 1160 case MLX5_WQE_PF_TYPE_RMP: 1161 srq = mlx5_cmd_get_srq(dev, wq_num); 1162 if (srq) 1163 common = &srq->common; 1164 break; 1165 case MLX5_WQE_PF_TYPE_REQ_SEND_OR_WRITE: 1166 case MLX5_WQE_PF_TYPE_RESP: 1167 case MLX5_WQE_PF_TYPE_REQ_READ_OR_ATOMIC: 1168 common = mlx5_core_res_hold(dev, wq_num, MLX5_RES_QP); 1169 break; 1170 default: 1171 break; 1172 } 1173 1174 return common; 1175 } 1176 1177 static inline struct mlx5_ib_qp *res_to_qp(struct mlx5_core_rsc_common *res) 1178 { 1179 struct mlx5_core_qp *mqp = (struct mlx5_core_qp *)res; 1180 1181 return to_mibqp(mqp); 1182 } 1183 1184 static inline struct mlx5_ib_srq *res_to_srq(struct mlx5_core_rsc_common *res) 1185 { 1186 struct mlx5_core_srq *msrq = 1187 container_of(res, struct mlx5_core_srq, common); 1188 1189 return to_mibsrq(msrq); 1190 } 1191 1192 static void mlx5_ib_mr_wqe_pfault_handler(struct mlx5_ib_dev *dev, 1193 struct mlx5_pagefault *pfault) 1194 { 1195 bool sq = pfault->type & MLX5_PFAULT_REQUESTOR; 1196 u16 wqe_index = pfault->wqe.wqe_index; 1197 void *wqe, *wqe_start = NULL, *wqe_end = NULL; 1198 u32 bytes_mapped, total_wqe_bytes; 1199 struct mlx5_core_rsc_common *res; 1200 int resume_with_error = 1; 1201 struct mlx5_ib_qp *qp; 1202 size_t bytes_copied; 1203 int ret = 0; 1204 1205 res = odp_get_rsc(dev, pfault->wqe.wq_num, pfault->type); 1206 if (!res) { 1207 mlx5_ib_dbg(dev, "wqe page fault for missing resource %d\n", pfault->wqe.wq_num); 1208 return; 1209 } 1210 1211 if (res->res != MLX5_RES_QP && res->res != MLX5_RES_SRQ && 1212 res->res != MLX5_RES_XSRQ) { 1213 mlx5_ib_err(dev, "wqe page fault for unsupported type %d\n", 1214 pfault->type); 1215 goto resolve_page_fault; 1216 } 1217 1218 wqe_start = (void *)__get_free_page(GFP_KERNEL); 1219 if (!wqe_start) { 1220 mlx5_ib_err(dev, "Error allocating memory for IO page fault handling.\n"); 1221 goto resolve_page_fault; 1222 } 1223 1224 wqe = wqe_start; 1225 qp = (res->res == MLX5_RES_QP) ? res_to_qp(res) : NULL; 1226 if (qp && sq) { 1227 ret = mlx5_ib_read_wqe_sq(qp, wqe_index, wqe, PAGE_SIZE, 1228 &bytes_copied); 1229 if (ret) 1230 goto read_user; 1231 ret = mlx5_ib_mr_initiator_pfault_handler( 1232 dev, pfault, qp, &wqe, &wqe_end, bytes_copied); 1233 } else if (qp && !sq) { 1234 ret = mlx5_ib_read_wqe_rq(qp, wqe_index, wqe, PAGE_SIZE, 1235 &bytes_copied); 1236 if (ret) 1237 goto read_user; 1238 ret = mlx5_ib_mr_responder_pfault_handler_rq( 1239 dev, qp, wqe, &wqe_end, bytes_copied); 1240 } else if (!qp) { 1241 struct mlx5_ib_srq *srq = res_to_srq(res); 1242 1243 ret = mlx5_ib_read_wqe_srq(srq, wqe_index, wqe, PAGE_SIZE, 1244 &bytes_copied); 1245 if (ret) 1246 goto read_user; 1247 ret = mlx5_ib_mr_responder_pfault_handler_srq( 1248 dev, srq, &wqe, &wqe_end, bytes_copied); 1249 } 1250 1251 if (ret < 0 || wqe >= wqe_end) 1252 goto resolve_page_fault; 1253 1254 ret = pagefault_data_segments(dev, pfault, wqe, wqe_end, &bytes_mapped, 1255 &total_wqe_bytes, !sq); 1256 if (ret == -EAGAIN) 1257 goto out; 1258 1259 if (ret < 0 || total_wqe_bytes > bytes_mapped) 1260 goto resolve_page_fault; 1261 1262 out: 1263 ret = 0; 1264 resume_with_error = 0; 1265 1266 read_user: 1267 if (ret) 1268 mlx5_ib_err( 1269 dev, 1270 "Failed reading a WQE following page fault, error %d, wqe_index %x, qpn %x\n", 1271 ret, wqe_index, pfault->token); 1272 1273 resolve_page_fault: 1274 mlx5_ib_page_fault_resume(dev, pfault, resume_with_error); 1275 mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x resume_with_error=%d, type: 0x%x\n", 1276 pfault->wqe.wq_num, resume_with_error, 1277 pfault->type); 1278 mlx5_core_res_put(res); 1279 free_page((unsigned long)wqe_start); 1280 } 1281 1282 static int pages_in_range(u64 address, u32 length) 1283 { 1284 return (ALIGN(address + length, PAGE_SIZE) - 1285 (address & PAGE_MASK)) >> PAGE_SHIFT; 1286 } 1287 1288 static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_dev *dev, 1289 struct mlx5_pagefault *pfault) 1290 { 1291 u64 address; 1292 u32 length; 1293 u32 prefetch_len = pfault->bytes_committed; 1294 int prefetch_activated = 0; 1295 u32 rkey = pfault->rdma.r_key; 1296 int ret; 1297 1298 /* The RDMA responder handler handles the page fault in two parts. 1299 * First it brings the necessary pages for the current packet 1300 * (and uses the pfault context), and then (after resuming the QP) 1301 * prefetches more pages. The second operation cannot use the pfault 1302 * context and therefore uses the dummy_pfault context allocated on 1303 * the stack */ 1304 pfault->rdma.rdma_va += pfault->bytes_committed; 1305 pfault->rdma.rdma_op_len -= min(pfault->bytes_committed, 1306 pfault->rdma.rdma_op_len); 1307 pfault->bytes_committed = 0; 1308 1309 address = pfault->rdma.rdma_va; 1310 length = pfault->rdma.rdma_op_len; 1311 1312 /* For some operations, the hardware cannot tell the exact message 1313 * length, and in those cases it reports zero. Use prefetch 1314 * logic. */ 1315 if (length == 0) { 1316 prefetch_activated = 1; 1317 length = pfault->rdma.packet_size; 1318 prefetch_len = min(MAX_PREFETCH_LEN, prefetch_len); 1319 } 1320 1321 ret = pagefault_single_data_segment(dev, NULL, rkey, address, length, 1322 &pfault->bytes_committed, NULL); 1323 if (ret == -EAGAIN) { 1324 /* We're racing with an invalidation, don't prefetch */ 1325 prefetch_activated = 0; 1326 } else if (ret < 0 || pages_in_range(address, length) > ret) { 1327 mlx5_ib_page_fault_resume(dev, pfault, 1); 1328 if (ret != -ENOENT) 1329 mlx5_ib_dbg(dev, "PAGE FAULT error %d. QP 0x%x, type: 0x%x\n", 1330 ret, pfault->token, pfault->type); 1331 return; 1332 } 1333 1334 mlx5_ib_page_fault_resume(dev, pfault, 0); 1335 mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x, type: 0x%x, prefetch_activated: %d\n", 1336 pfault->token, pfault->type, 1337 prefetch_activated); 1338 1339 /* At this point, there might be a new pagefault already arriving in 1340 * the eq, switch to the dummy pagefault for the rest of the 1341 * processing. We're still OK with the objects being alive as the 1342 * work-queue is being fenced. */ 1343 1344 if (prefetch_activated) { 1345 u32 bytes_committed = 0; 1346 1347 ret = pagefault_single_data_segment(dev, NULL, rkey, address, 1348 prefetch_len, 1349 &bytes_committed, NULL); 1350 if (ret < 0 && ret != -EAGAIN) { 1351 mlx5_ib_dbg(dev, "Prefetch failed. ret: %d, QP 0x%x, address: 0x%.16llx, length = 0x%.16x\n", 1352 ret, pfault->token, address, prefetch_len); 1353 } 1354 } 1355 } 1356 1357 static void mlx5_ib_pfault(struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault) 1358 { 1359 u8 event_subtype = pfault->event_subtype; 1360 1361 switch (event_subtype) { 1362 case MLX5_PFAULT_SUBTYPE_WQE: 1363 mlx5_ib_mr_wqe_pfault_handler(dev, pfault); 1364 break; 1365 case MLX5_PFAULT_SUBTYPE_RDMA: 1366 mlx5_ib_mr_rdma_pfault_handler(dev, pfault); 1367 break; 1368 default: 1369 mlx5_ib_err(dev, "Invalid page fault event subtype: 0x%x\n", 1370 event_subtype); 1371 mlx5_ib_page_fault_resume(dev, pfault, 1); 1372 } 1373 } 1374 1375 static void mlx5_ib_eqe_pf_action(struct work_struct *work) 1376 { 1377 struct mlx5_pagefault *pfault = container_of(work, 1378 struct mlx5_pagefault, 1379 work); 1380 struct mlx5_ib_pf_eq *eq = pfault->eq; 1381 1382 mlx5_ib_pfault(eq->dev, pfault); 1383 mempool_free(pfault, eq->pool); 1384 } 1385 1386 static void mlx5_ib_eq_pf_process(struct mlx5_ib_pf_eq *eq) 1387 { 1388 struct mlx5_eqe_page_fault *pf_eqe; 1389 struct mlx5_pagefault *pfault; 1390 struct mlx5_eqe *eqe; 1391 int cc = 0; 1392 1393 while ((eqe = mlx5_eq_get_eqe(eq->core, cc))) { 1394 pfault = mempool_alloc(eq->pool, GFP_ATOMIC); 1395 if (!pfault) { 1396 schedule_work(&eq->work); 1397 break; 1398 } 1399 1400 pf_eqe = &eqe->data.page_fault; 1401 pfault->event_subtype = eqe->sub_type; 1402 pfault->bytes_committed = be32_to_cpu(pf_eqe->bytes_committed); 1403 1404 mlx5_ib_dbg(eq->dev, 1405 "PAGE_FAULT: subtype: 0x%02x, bytes_committed: 0x%06x\n", 1406 eqe->sub_type, pfault->bytes_committed); 1407 1408 switch (eqe->sub_type) { 1409 case MLX5_PFAULT_SUBTYPE_RDMA: 1410 /* RDMA based event */ 1411 pfault->type = 1412 be32_to_cpu(pf_eqe->rdma.pftype_token) >> 24; 1413 pfault->token = 1414 be32_to_cpu(pf_eqe->rdma.pftype_token) & 1415 MLX5_24BIT_MASK; 1416 pfault->rdma.r_key = 1417 be32_to_cpu(pf_eqe->rdma.r_key); 1418 pfault->rdma.packet_size = 1419 be16_to_cpu(pf_eqe->rdma.packet_length); 1420 pfault->rdma.rdma_op_len = 1421 be32_to_cpu(pf_eqe->rdma.rdma_op_len); 1422 pfault->rdma.rdma_va = 1423 be64_to_cpu(pf_eqe->rdma.rdma_va); 1424 mlx5_ib_dbg(eq->dev, 1425 "PAGE_FAULT: type:0x%x, token: 0x%06x, r_key: 0x%08x\n", 1426 pfault->type, pfault->token, 1427 pfault->rdma.r_key); 1428 mlx5_ib_dbg(eq->dev, 1429 "PAGE_FAULT: rdma_op_len: 0x%08x, rdma_va: 0x%016llx\n", 1430 pfault->rdma.rdma_op_len, 1431 pfault->rdma.rdma_va); 1432 break; 1433 1434 case MLX5_PFAULT_SUBTYPE_WQE: 1435 /* WQE based event */ 1436 pfault->type = 1437 (be32_to_cpu(pf_eqe->wqe.pftype_wq) >> 24) & 0x7; 1438 pfault->token = 1439 be32_to_cpu(pf_eqe->wqe.token); 1440 pfault->wqe.wq_num = 1441 be32_to_cpu(pf_eqe->wqe.pftype_wq) & 1442 MLX5_24BIT_MASK; 1443 pfault->wqe.wqe_index = 1444 be16_to_cpu(pf_eqe->wqe.wqe_index); 1445 pfault->wqe.packet_size = 1446 be16_to_cpu(pf_eqe->wqe.packet_length); 1447 mlx5_ib_dbg(eq->dev, 1448 "PAGE_FAULT: type:0x%x, token: 0x%06x, wq_num: 0x%06x, wqe_index: 0x%04x\n", 1449 pfault->type, pfault->token, 1450 pfault->wqe.wq_num, 1451 pfault->wqe.wqe_index); 1452 break; 1453 1454 default: 1455 mlx5_ib_warn(eq->dev, 1456 "Unsupported page fault event sub-type: 0x%02hhx\n", 1457 eqe->sub_type); 1458 /* Unsupported page faults should still be 1459 * resolved by the page fault handler 1460 */ 1461 } 1462 1463 pfault->eq = eq; 1464 INIT_WORK(&pfault->work, mlx5_ib_eqe_pf_action); 1465 queue_work(eq->wq, &pfault->work); 1466 1467 cc = mlx5_eq_update_cc(eq->core, ++cc); 1468 } 1469 1470 mlx5_eq_update_ci(eq->core, cc, 1); 1471 } 1472 1473 static int mlx5_ib_eq_pf_int(struct notifier_block *nb, unsigned long type, 1474 void *data) 1475 { 1476 struct mlx5_ib_pf_eq *eq = 1477 container_of(nb, struct mlx5_ib_pf_eq, irq_nb); 1478 unsigned long flags; 1479 1480 if (spin_trylock_irqsave(&eq->lock, flags)) { 1481 mlx5_ib_eq_pf_process(eq); 1482 spin_unlock_irqrestore(&eq->lock, flags); 1483 } else { 1484 schedule_work(&eq->work); 1485 } 1486 1487 return IRQ_HANDLED; 1488 } 1489 1490 /* mempool_refill() was proposed but unfortunately wasn't accepted 1491 * http://lkml.iu.edu/hypermail/linux/kernel/1512.1/05073.html 1492 * Cheap workaround. 1493 */ 1494 static void mempool_refill(mempool_t *pool) 1495 { 1496 while (pool->curr_nr < pool->min_nr) 1497 mempool_free(mempool_alloc(pool, GFP_KERNEL), pool); 1498 } 1499 1500 static void mlx5_ib_eq_pf_action(struct work_struct *work) 1501 { 1502 struct mlx5_ib_pf_eq *eq = 1503 container_of(work, struct mlx5_ib_pf_eq, work); 1504 1505 mempool_refill(eq->pool); 1506 1507 spin_lock_irq(&eq->lock); 1508 mlx5_ib_eq_pf_process(eq); 1509 spin_unlock_irq(&eq->lock); 1510 } 1511 1512 enum { 1513 MLX5_IB_NUM_PF_EQE = 0x1000, 1514 MLX5_IB_NUM_PF_DRAIN = 64, 1515 }; 1516 1517 int mlx5r_odp_create_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq) 1518 { 1519 struct mlx5_eq_param param = {}; 1520 int err = 0; 1521 1522 mutex_lock(&dev->odp_eq_mutex); 1523 if (eq->core) 1524 goto unlock; 1525 INIT_WORK(&eq->work, mlx5_ib_eq_pf_action); 1526 spin_lock_init(&eq->lock); 1527 eq->dev = dev; 1528 1529 eq->pool = mempool_create_kmalloc_pool(MLX5_IB_NUM_PF_DRAIN, 1530 sizeof(struct mlx5_pagefault)); 1531 if (!eq->pool) { 1532 err = -ENOMEM; 1533 goto unlock; 1534 } 1535 1536 eq->wq = alloc_workqueue("mlx5_ib_page_fault", 1537 WQ_HIGHPRI | WQ_UNBOUND | WQ_MEM_RECLAIM, 1538 MLX5_NUM_CMD_EQE); 1539 if (!eq->wq) { 1540 err = -ENOMEM; 1541 goto err_mempool; 1542 } 1543 1544 eq->irq_nb.notifier_call = mlx5_ib_eq_pf_int; 1545 param = (struct mlx5_eq_param) { 1546 .nent = MLX5_IB_NUM_PF_EQE, 1547 }; 1548 param.mask[0] = 1ull << MLX5_EVENT_TYPE_PAGE_FAULT; 1549 eq->core = mlx5_eq_create_generic(dev->mdev, ¶m); 1550 if (IS_ERR(eq->core)) { 1551 err = PTR_ERR(eq->core); 1552 goto err_wq; 1553 } 1554 err = mlx5_eq_enable(dev->mdev, eq->core, &eq->irq_nb); 1555 if (err) { 1556 mlx5_ib_err(dev, "failed to enable odp EQ %d\n", err); 1557 goto err_eq; 1558 } 1559 1560 mutex_unlock(&dev->odp_eq_mutex); 1561 return 0; 1562 err_eq: 1563 mlx5_eq_destroy_generic(dev->mdev, eq->core); 1564 err_wq: 1565 eq->core = NULL; 1566 destroy_workqueue(eq->wq); 1567 err_mempool: 1568 mempool_destroy(eq->pool); 1569 unlock: 1570 mutex_unlock(&dev->odp_eq_mutex); 1571 return err; 1572 } 1573 1574 static int 1575 mlx5_ib_odp_destroy_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq) 1576 { 1577 int err; 1578 1579 if (!eq->core) 1580 return 0; 1581 mlx5_eq_disable(dev->mdev, eq->core, &eq->irq_nb); 1582 err = mlx5_eq_destroy_generic(dev->mdev, eq->core); 1583 cancel_work_sync(&eq->work); 1584 destroy_workqueue(eq->wq); 1585 mempool_destroy(eq->pool); 1586 1587 return err; 1588 } 1589 1590 void mlx5_odp_init_mkey_cache_entry(struct mlx5_cache_ent *ent) 1591 { 1592 if (!(ent->dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT)) 1593 return; 1594 1595 switch (ent->order - 2) { 1596 case MLX5_IMR_MTT_CACHE_ENTRY: 1597 ent->page = PAGE_SHIFT; 1598 ent->ndescs = MLX5_IMR_MTT_ENTRIES; 1599 ent->access_mode = MLX5_MKC_ACCESS_MODE_MTT; 1600 ent->limit = 0; 1601 break; 1602 1603 case MLX5_IMR_KSM_CACHE_ENTRY: 1604 ent->page = MLX5_KSM_PAGE_SHIFT; 1605 ent->ndescs = mlx5_imr_ksm_entries; 1606 ent->access_mode = MLX5_MKC_ACCESS_MODE_KSM; 1607 ent->limit = 0; 1608 break; 1609 } 1610 } 1611 1612 static const struct ib_device_ops mlx5_ib_dev_odp_ops = { 1613 .advise_mr = mlx5_ib_advise_mr, 1614 }; 1615 1616 int mlx5_ib_odp_init_one(struct mlx5_ib_dev *dev) 1617 { 1618 int ret = 0; 1619 1620 internal_fill_odp_caps(dev); 1621 1622 if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT)) 1623 return ret; 1624 1625 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_odp_ops); 1626 1627 if (dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT) { 1628 ret = mlx5_cmd_null_mkey(dev->mdev, &dev->null_mkey); 1629 if (ret) { 1630 mlx5_ib_err(dev, "Error getting null_mkey %d\n", ret); 1631 return ret; 1632 } 1633 } 1634 1635 mutex_init(&dev->odp_eq_mutex); 1636 return ret; 1637 } 1638 1639 void mlx5_ib_odp_cleanup_one(struct mlx5_ib_dev *dev) 1640 { 1641 if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT)) 1642 return; 1643 1644 mlx5_ib_odp_destroy_eq(dev, &dev->odp_pf_eq); 1645 } 1646 1647 int mlx5_ib_odp_init(void) 1648 { 1649 mlx5_imr_ksm_entries = BIT_ULL(get_order(TASK_SIZE) - 1650 MLX5_IMR_MTT_BITS); 1651 1652 return 0; 1653 } 1654 1655 struct prefetch_mr_work { 1656 struct work_struct work; 1657 u32 pf_flags; 1658 u32 num_sge; 1659 struct { 1660 u64 io_virt; 1661 struct mlx5_ib_mr *mr; 1662 size_t length; 1663 } frags[]; 1664 }; 1665 1666 static void destroy_prefetch_work(struct prefetch_mr_work *work) 1667 { 1668 u32 i; 1669 1670 for (i = 0; i < work->num_sge; ++i) 1671 mlx5r_deref_odp_mkey(&work->frags[i].mr->mmkey); 1672 1673 kvfree(work); 1674 } 1675 1676 static struct mlx5_ib_mr * 1677 get_prefetchable_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice, 1678 u32 lkey) 1679 { 1680 struct mlx5_ib_dev *dev = to_mdev(pd->device); 1681 struct mlx5_ib_mr *mr = NULL; 1682 struct mlx5_ib_mkey *mmkey; 1683 1684 xa_lock(&dev->odp_mkeys); 1685 mmkey = xa_load(&dev->odp_mkeys, mlx5_base_mkey(lkey)); 1686 if (!mmkey || mmkey->key != lkey) { 1687 mr = ERR_PTR(-ENOENT); 1688 goto end; 1689 } 1690 if (mmkey->type != MLX5_MKEY_MR) { 1691 mr = ERR_PTR(-EINVAL); 1692 goto end; 1693 } 1694 1695 mr = container_of(mmkey, struct mlx5_ib_mr, mmkey); 1696 1697 if (mr->ibmr.pd != pd) { 1698 mr = ERR_PTR(-EPERM); 1699 goto end; 1700 } 1701 1702 /* prefetch with write-access must be supported by the MR */ 1703 if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && 1704 !mr->umem->writable) { 1705 mr = ERR_PTR(-EPERM); 1706 goto end; 1707 } 1708 1709 refcount_inc(&mmkey->usecount); 1710 end: 1711 xa_unlock(&dev->odp_mkeys); 1712 return mr; 1713 } 1714 1715 static void mlx5_ib_prefetch_mr_work(struct work_struct *w) 1716 { 1717 struct prefetch_mr_work *work = 1718 container_of(w, struct prefetch_mr_work, work); 1719 u32 bytes_mapped = 0; 1720 int ret; 1721 u32 i; 1722 1723 /* We rely on IB/core that work is executed if we have num_sge != 0 only. */ 1724 WARN_ON(!work->num_sge); 1725 for (i = 0; i < work->num_sge; ++i) { 1726 ret = pagefault_mr(work->frags[i].mr, work->frags[i].io_virt, 1727 work->frags[i].length, &bytes_mapped, 1728 work->pf_flags); 1729 if (ret <= 0) 1730 continue; 1731 mlx5_update_odp_stats(work->frags[i].mr, prefetch, ret); 1732 } 1733 1734 destroy_prefetch_work(work); 1735 } 1736 1737 static int init_prefetch_work(struct ib_pd *pd, 1738 enum ib_uverbs_advise_mr_advice advice, 1739 u32 pf_flags, struct prefetch_mr_work *work, 1740 struct ib_sge *sg_list, u32 num_sge) 1741 { 1742 u32 i; 1743 1744 INIT_WORK(&work->work, mlx5_ib_prefetch_mr_work); 1745 work->pf_flags = pf_flags; 1746 1747 for (i = 0; i < num_sge; ++i) { 1748 struct mlx5_ib_mr *mr; 1749 1750 mr = get_prefetchable_mr(pd, advice, sg_list[i].lkey); 1751 if (IS_ERR(mr)) { 1752 work->num_sge = i; 1753 return PTR_ERR(mr); 1754 } 1755 work->frags[i].io_virt = sg_list[i].addr; 1756 work->frags[i].length = sg_list[i].length; 1757 work->frags[i].mr = mr; 1758 } 1759 work->num_sge = num_sge; 1760 return 0; 1761 } 1762 1763 static int mlx5_ib_prefetch_sg_list(struct ib_pd *pd, 1764 enum ib_uverbs_advise_mr_advice advice, 1765 u32 pf_flags, struct ib_sge *sg_list, 1766 u32 num_sge) 1767 { 1768 u32 bytes_mapped = 0; 1769 int ret = 0; 1770 u32 i; 1771 1772 for (i = 0; i < num_sge; ++i) { 1773 struct mlx5_ib_mr *mr; 1774 1775 mr = get_prefetchable_mr(pd, advice, sg_list[i].lkey); 1776 if (IS_ERR(mr)) 1777 return PTR_ERR(mr); 1778 ret = pagefault_mr(mr, sg_list[i].addr, sg_list[i].length, 1779 &bytes_mapped, pf_flags); 1780 if (ret < 0) { 1781 mlx5r_deref_odp_mkey(&mr->mmkey); 1782 return ret; 1783 } 1784 mlx5_update_odp_stats(mr, prefetch, ret); 1785 mlx5r_deref_odp_mkey(&mr->mmkey); 1786 } 1787 1788 return 0; 1789 } 1790 1791 int mlx5_ib_advise_mr_prefetch(struct ib_pd *pd, 1792 enum ib_uverbs_advise_mr_advice advice, 1793 u32 flags, struct ib_sge *sg_list, u32 num_sge) 1794 { 1795 u32 pf_flags = 0; 1796 struct prefetch_mr_work *work; 1797 int rc; 1798 1799 if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH) 1800 pf_flags |= MLX5_PF_FLAGS_DOWNGRADE; 1801 1802 if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT) 1803 pf_flags |= MLX5_PF_FLAGS_SNAPSHOT; 1804 1805 if (flags & IB_UVERBS_ADVISE_MR_FLAG_FLUSH) 1806 return mlx5_ib_prefetch_sg_list(pd, advice, pf_flags, sg_list, 1807 num_sge); 1808 1809 work = kvzalloc(struct_size(work, frags, num_sge), GFP_KERNEL); 1810 if (!work) 1811 return -ENOMEM; 1812 1813 rc = init_prefetch_work(pd, advice, pf_flags, work, sg_list, num_sge); 1814 if (rc) { 1815 destroy_prefetch_work(work); 1816 return rc; 1817 } 1818 queue_work(system_unbound_wq, &work->work); 1819 return 0; 1820 } 1821