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. */ 46struct 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 97static u64 mlx5_imr_ksm_entries; 98 99static 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 = mr_to_mdev(imr)->mkeys.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 = mr_to_mdev(imr)->mkeys.null_mkey; 141 pklm->va = 0; 142 } 143 } 144} 145 146static 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 158static 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 174void 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 */ 190static 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 210static 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 226static 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); 310out: 311 mutex_unlock(&umem_odp->umem_mutex); 312 return true; 313} 314 315const struct mmu_interval_notifier_ops mlx5_mn_ops = { 316 .invalidate = mlx5_ib_invalidate_range, 317}; 318 319static 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 384static 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 405static 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, imr->access_flags, 421 MLX5_MKC_ACCESS_MODE_MTT, 422 MLX5_IMR_MTT_ENTRIES); 423 if (IS_ERR(mr)) { 424 ib_umem_odp_release(odp); 425 return mr; 426 } 427 428 mr->access_flags = imr->access_flags; 429 mr->ibmr.pd = imr->ibmr.pd; 430 mr->ibmr.device = &mr_to_mdev(imr)->ib_dev; 431 mr->umem = &odp->umem; 432 mr->ibmr.lkey = mr->mmkey.key; 433 mr->ibmr.rkey = mr->mmkey.key; 434 mr->ibmr.iova = idx * MLX5_IMR_MTT_SIZE; 435 mr->parent = imr; 436 odp->private = mr; 437 438 /* 439 * First refcount is owned by the xarray and second refconut 440 * is returned to the caller. 441 */ 442 refcount_set(&mr->mmkey.usecount, 2); 443 444 err = mlx5r_umr_update_xlt(mr, 0, 445 MLX5_IMR_MTT_ENTRIES, 446 PAGE_SHIFT, 447 MLX5_IB_UPD_XLT_ZAP | 448 MLX5_IB_UPD_XLT_ENABLE); 449 if (err) { 450 ret = ERR_PTR(err); 451 goto out_mr; 452 } 453 454 xa_lock(&imr->implicit_children); 455 ret = __xa_cmpxchg(&imr->implicit_children, idx, NULL, mr, 456 GFP_KERNEL); 457 if (unlikely(ret)) { 458 if (xa_is_err(ret)) { 459 ret = ERR_PTR(xa_err(ret)); 460 goto out_lock; 461 } 462 /* 463 * Another thread beat us to creating the child mr, use 464 * theirs. 465 */ 466 refcount_inc(&ret->mmkey.usecount); 467 goto out_lock; 468 } 469 xa_unlock(&imr->implicit_children); 470 471 mlx5_ib_dbg(mr_to_mdev(imr), "key %x mr %p\n", mr->mmkey.key, mr); 472 return mr; 473 474out_lock: 475 xa_unlock(&imr->implicit_children); 476out_mr: 477 mlx5_ib_dereg_mr(&mr->ibmr, NULL); 478 return ret; 479} 480 481struct mlx5_ib_mr *mlx5_ib_alloc_implicit_mr(struct mlx5_ib_pd *pd, 482 int access_flags) 483{ 484 struct mlx5_ib_dev *dev = to_mdev(pd->ibpd.device); 485 struct ib_umem_odp *umem_odp; 486 struct mlx5_ib_mr *imr; 487 int err; 488 489 if (!mlx5r_umr_can_load_pas(dev, MLX5_IMR_MTT_ENTRIES * PAGE_SIZE)) 490 return ERR_PTR(-EOPNOTSUPP); 491 492 umem_odp = ib_umem_odp_alloc_implicit(&dev->ib_dev, access_flags); 493 if (IS_ERR(umem_odp)) 494 return ERR_CAST(umem_odp); 495 496 imr = mlx5_mr_cache_alloc(dev, access_flags, MLX5_MKC_ACCESS_MODE_KSM, 497 mlx5_imr_ksm_entries); 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; 528out_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 534void 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) 552static 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 600out: 601 return ret; 602} 603 604static 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 */ 665out: 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 690static 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 */ 737static 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 764int 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 774int 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 784struct pf_frame { 785 struct pf_frame *next; 786 u32 key; 787 u64 io_virt; 788 size_t bcnt; 789 int depth; 790}; 791 792static 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 */ 813static 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 830next_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 948end: 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 */ 981static 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 __be32 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 = 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 && 1018 key == dev->mkeys.terminate_scatter_list_mkey && 1019 io_virt == 0) 1020 break; 1021 1022 if (!inline_segment && total_wqe_bytes) { 1023 *total_wqe_bytes += bcnt - min_t(size_t, bcnt, 1024 pfault->bytes_committed); 1025 } 1026 1027 /* A zero length data segment designates a length of 2GB. */ 1028 if (bcnt == 0) 1029 bcnt = 1U << 31; 1030 1031 if (inline_segment || bcnt <= pfault->bytes_committed) { 1032 pfault->bytes_committed -= 1033 min_t(size_t, bcnt, 1034 pfault->bytes_committed); 1035 continue; 1036 } 1037 1038 ret = pagefault_single_data_segment(dev, NULL, be32_to_cpu(key), 1039 io_virt, bcnt, 1040 &pfault->bytes_committed, 1041 bytes_mapped); 1042 if (ret < 0) 1043 break; 1044 npages += ret; 1045 } 1046 1047 return ret < 0 ? ret : npages; 1048} 1049 1050/* 1051 * Parse initiator WQE. Advances the wqe pointer to point at the 1052 * scatter-gather list, and set wqe_end to the end of the WQE. 1053 */ 1054static int mlx5_ib_mr_initiator_pfault_handler( 1055 struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault, 1056 struct mlx5_ib_qp *qp, void **wqe, void **wqe_end, int wqe_length) 1057{ 1058 struct mlx5_wqe_ctrl_seg *ctrl = *wqe; 1059 u16 wqe_index = pfault->wqe.wqe_index; 1060 struct mlx5_base_av *av; 1061 unsigned ds, opcode; 1062 u32 qpn = qp->trans_qp.base.mqp.qpn; 1063 1064 ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK; 1065 if (ds * MLX5_WQE_DS_UNITS > wqe_length) { 1066 mlx5_ib_err(dev, "Unable to read the complete WQE. ds = 0x%x, ret = 0x%x\n", 1067 ds, wqe_length); 1068 return -EFAULT; 1069 } 1070 1071 if (ds == 0) { 1072 mlx5_ib_err(dev, "Got WQE with zero DS. wqe_index=%x, qpn=%x\n", 1073 wqe_index, qpn); 1074 return -EFAULT; 1075 } 1076 1077 *wqe_end = *wqe + ds * MLX5_WQE_DS_UNITS; 1078 *wqe += sizeof(*ctrl); 1079 1080 opcode = be32_to_cpu(ctrl->opmod_idx_opcode) & 1081 MLX5_WQE_CTRL_OPCODE_MASK; 1082 1083 if (qp->type == IB_QPT_XRC_INI) 1084 *wqe += sizeof(struct mlx5_wqe_xrc_seg); 1085 1086 if (qp->type == IB_QPT_UD || qp->type == MLX5_IB_QPT_DCI) { 1087 av = *wqe; 1088 if (av->dqp_dct & cpu_to_be32(MLX5_EXTENDED_UD_AV)) 1089 *wqe += sizeof(struct mlx5_av); 1090 else 1091 *wqe += sizeof(struct mlx5_base_av); 1092 } 1093 1094 switch (opcode) { 1095 case MLX5_OPCODE_RDMA_WRITE: 1096 case MLX5_OPCODE_RDMA_WRITE_IMM: 1097 case MLX5_OPCODE_RDMA_READ: 1098 *wqe += sizeof(struct mlx5_wqe_raddr_seg); 1099 break; 1100 case MLX5_OPCODE_ATOMIC_CS: 1101 case MLX5_OPCODE_ATOMIC_FA: 1102 *wqe += sizeof(struct mlx5_wqe_raddr_seg); 1103 *wqe += sizeof(struct mlx5_wqe_atomic_seg); 1104 break; 1105 } 1106 1107 return 0; 1108} 1109 1110/* 1111 * Parse responder WQE and set wqe_end to the end of the WQE. 1112 */ 1113static int mlx5_ib_mr_responder_pfault_handler_srq(struct mlx5_ib_dev *dev, 1114 struct mlx5_ib_srq *srq, 1115 void **wqe, void **wqe_end, 1116 int wqe_length) 1117{ 1118 int wqe_size = 1 << srq->msrq.wqe_shift; 1119 1120 if (wqe_size > wqe_length) { 1121 mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n"); 1122 return -EFAULT; 1123 } 1124 1125 *wqe_end = *wqe + wqe_size; 1126 *wqe += sizeof(struct mlx5_wqe_srq_next_seg); 1127 1128 return 0; 1129} 1130 1131static int mlx5_ib_mr_responder_pfault_handler_rq(struct mlx5_ib_dev *dev, 1132 struct mlx5_ib_qp *qp, 1133 void *wqe, void **wqe_end, 1134 int wqe_length) 1135{ 1136 struct mlx5_ib_wq *wq = &qp->rq; 1137 int wqe_size = 1 << wq->wqe_shift; 1138 1139 if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE) { 1140 mlx5_ib_err(dev, "ODP fault with WQE signatures is not supported\n"); 1141 return -EFAULT; 1142 } 1143 1144 if (wqe_size > wqe_length) { 1145 mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n"); 1146 return -EFAULT; 1147 } 1148 1149 *wqe_end = wqe + wqe_size; 1150 1151 return 0; 1152} 1153 1154static inline struct mlx5_core_rsc_common *odp_get_rsc(struct mlx5_ib_dev *dev, 1155 u32 wq_num, int pf_type) 1156{ 1157 struct mlx5_core_rsc_common *common = NULL; 1158 struct mlx5_core_srq *srq; 1159 1160 switch (pf_type) { 1161 case MLX5_WQE_PF_TYPE_RMP: 1162 srq = mlx5_cmd_get_srq(dev, wq_num); 1163 if (srq) 1164 common = &srq->common; 1165 break; 1166 case MLX5_WQE_PF_TYPE_REQ_SEND_OR_WRITE: 1167 case MLX5_WQE_PF_TYPE_RESP: 1168 case MLX5_WQE_PF_TYPE_REQ_READ_OR_ATOMIC: 1169 common = mlx5_core_res_hold(dev, wq_num, MLX5_RES_QP); 1170 break; 1171 default: 1172 break; 1173 } 1174 1175 return common; 1176} 1177 1178static inline struct mlx5_ib_qp *res_to_qp(struct mlx5_core_rsc_common *res) 1179{ 1180 struct mlx5_core_qp *mqp = (struct mlx5_core_qp *)res; 1181 1182 return to_mibqp(mqp); 1183} 1184 1185static inline struct mlx5_ib_srq *res_to_srq(struct mlx5_core_rsc_common *res) 1186{ 1187 struct mlx5_core_srq *msrq = 1188 container_of(res, struct mlx5_core_srq, common); 1189 1190 return to_mibsrq(msrq); 1191} 1192 1193static void mlx5_ib_mr_wqe_pfault_handler(struct mlx5_ib_dev *dev, 1194 struct mlx5_pagefault *pfault) 1195{ 1196 bool sq = pfault->type & MLX5_PFAULT_REQUESTOR; 1197 u16 wqe_index = pfault->wqe.wqe_index; 1198 void *wqe, *wqe_start = NULL, *wqe_end = NULL; 1199 u32 bytes_mapped, total_wqe_bytes; 1200 struct mlx5_core_rsc_common *res; 1201 int resume_with_error = 1; 1202 struct mlx5_ib_qp *qp; 1203 size_t bytes_copied; 1204 int ret = 0; 1205 1206 res = odp_get_rsc(dev, pfault->wqe.wq_num, pfault->type); 1207 if (!res) { 1208 mlx5_ib_dbg(dev, "wqe page fault for missing resource %d\n", pfault->wqe.wq_num); 1209 return; 1210 } 1211 1212 if (res->res != MLX5_RES_QP && res->res != MLX5_RES_SRQ && 1213 res->res != MLX5_RES_XSRQ) { 1214 mlx5_ib_err(dev, "wqe page fault for unsupported type %d\n", 1215 pfault->type); 1216 goto resolve_page_fault; 1217 } 1218 1219 wqe_start = (void *)__get_free_page(GFP_KERNEL); 1220 if (!wqe_start) { 1221 mlx5_ib_err(dev, "Error allocating memory for IO page fault handling.\n"); 1222 goto resolve_page_fault; 1223 } 1224 1225 wqe = wqe_start; 1226 qp = (res->res == MLX5_RES_QP) ? res_to_qp(res) : NULL; 1227 if (qp && sq) { 1228 ret = mlx5_ib_read_wqe_sq(qp, wqe_index, wqe, PAGE_SIZE, 1229 &bytes_copied); 1230 if (ret) 1231 goto read_user; 1232 ret = mlx5_ib_mr_initiator_pfault_handler( 1233 dev, pfault, qp, &wqe, &wqe_end, bytes_copied); 1234 } else if (qp && !sq) { 1235 ret = mlx5_ib_read_wqe_rq(qp, wqe_index, wqe, PAGE_SIZE, 1236 &bytes_copied); 1237 if (ret) 1238 goto read_user; 1239 ret = mlx5_ib_mr_responder_pfault_handler_rq( 1240 dev, qp, wqe, &wqe_end, bytes_copied); 1241 } else if (!qp) { 1242 struct mlx5_ib_srq *srq = res_to_srq(res); 1243 1244 ret = mlx5_ib_read_wqe_srq(srq, wqe_index, wqe, PAGE_SIZE, 1245 &bytes_copied); 1246 if (ret) 1247 goto read_user; 1248 ret = mlx5_ib_mr_responder_pfault_handler_srq( 1249 dev, srq, &wqe, &wqe_end, bytes_copied); 1250 } 1251 1252 if (ret < 0 || wqe >= wqe_end) 1253 goto resolve_page_fault; 1254 1255 ret = pagefault_data_segments(dev, pfault, wqe, wqe_end, &bytes_mapped, 1256 &total_wqe_bytes, !sq); 1257 if (ret == -EAGAIN) 1258 goto out; 1259 1260 if (ret < 0 || total_wqe_bytes > bytes_mapped) 1261 goto resolve_page_fault; 1262 1263out: 1264 ret = 0; 1265 resume_with_error = 0; 1266 1267read_user: 1268 if (ret) 1269 mlx5_ib_err( 1270 dev, 1271 "Failed reading a WQE following page fault, error %d, wqe_index %x, qpn %x\n", 1272 ret, wqe_index, pfault->token); 1273 1274resolve_page_fault: 1275 mlx5_ib_page_fault_resume(dev, pfault, resume_with_error); 1276 mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x resume_with_error=%d, type: 0x%x\n", 1277 pfault->wqe.wq_num, resume_with_error, 1278 pfault->type); 1279 mlx5_core_res_put(res); 1280 free_page((unsigned long)wqe_start); 1281} 1282 1283static int pages_in_range(u64 address, u32 length) 1284{ 1285 return (ALIGN(address + length, PAGE_SIZE) - 1286 (address & PAGE_MASK)) >> PAGE_SHIFT; 1287} 1288 1289static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_dev *dev, 1290 struct mlx5_pagefault *pfault) 1291{ 1292 u64 address; 1293 u32 length; 1294 u32 prefetch_len = pfault->bytes_committed; 1295 int prefetch_activated = 0; 1296 u32 rkey = pfault->rdma.r_key; 1297 int ret; 1298 1299 /* The RDMA responder handler handles the page fault in two parts. 1300 * First it brings the necessary pages for the current packet 1301 * (and uses the pfault context), and then (after resuming the QP) 1302 * prefetches more pages. The second operation cannot use the pfault 1303 * context and therefore uses the dummy_pfault context allocated on 1304 * the stack */ 1305 pfault->rdma.rdma_va += pfault->bytes_committed; 1306 pfault->rdma.rdma_op_len -= min(pfault->bytes_committed, 1307 pfault->rdma.rdma_op_len); 1308 pfault->bytes_committed = 0; 1309 1310 address = pfault->rdma.rdma_va; 1311 length = pfault->rdma.rdma_op_len; 1312 1313 /* For some operations, the hardware cannot tell the exact message 1314 * length, and in those cases it reports zero. Use prefetch 1315 * logic. */ 1316 if (length == 0) { 1317 prefetch_activated = 1; 1318 length = pfault->rdma.packet_size; 1319 prefetch_len = min(MAX_PREFETCH_LEN, prefetch_len); 1320 } 1321 1322 ret = pagefault_single_data_segment(dev, NULL, rkey, address, length, 1323 &pfault->bytes_committed, NULL); 1324 if (ret == -EAGAIN) { 1325 /* We're racing with an invalidation, don't prefetch */ 1326 prefetch_activated = 0; 1327 } else if (ret < 0 || pages_in_range(address, length) > ret) { 1328 mlx5_ib_page_fault_resume(dev, pfault, 1); 1329 if (ret != -ENOENT) 1330 mlx5_ib_dbg(dev, "PAGE FAULT error %d. QP 0x%x, type: 0x%x\n", 1331 ret, pfault->token, pfault->type); 1332 return; 1333 } 1334 1335 mlx5_ib_page_fault_resume(dev, pfault, 0); 1336 mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x, type: 0x%x, prefetch_activated: %d\n", 1337 pfault->token, pfault->type, 1338 prefetch_activated); 1339 1340 /* At this point, there might be a new pagefault already arriving in 1341 * the eq, switch to the dummy pagefault for the rest of the 1342 * processing. We're still OK with the objects being alive as the 1343 * work-queue is being fenced. */ 1344 1345 if (prefetch_activated) { 1346 u32 bytes_committed = 0; 1347 1348 ret = pagefault_single_data_segment(dev, NULL, rkey, address, 1349 prefetch_len, 1350 &bytes_committed, NULL); 1351 if (ret < 0 && ret != -EAGAIN) { 1352 mlx5_ib_dbg(dev, "Prefetch failed. ret: %d, QP 0x%x, address: 0x%.16llx, length = 0x%.16x\n", 1353 ret, pfault->token, address, prefetch_len); 1354 } 1355 } 1356} 1357 1358static void mlx5_ib_pfault(struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault) 1359{ 1360 u8 event_subtype = pfault->event_subtype; 1361 1362 switch (event_subtype) { 1363 case MLX5_PFAULT_SUBTYPE_WQE: 1364 mlx5_ib_mr_wqe_pfault_handler(dev, pfault); 1365 break; 1366 case MLX5_PFAULT_SUBTYPE_RDMA: 1367 mlx5_ib_mr_rdma_pfault_handler(dev, pfault); 1368 break; 1369 default: 1370 mlx5_ib_err(dev, "Invalid page fault event subtype: 0x%x\n", 1371 event_subtype); 1372 mlx5_ib_page_fault_resume(dev, pfault, 1); 1373 } 1374} 1375 1376static void mlx5_ib_eqe_pf_action(struct work_struct *work) 1377{ 1378 struct mlx5_pagefault *pfault = container_of(work, 1379 struct mlx5_pagefault, 1380 work); 1381 struct mlx5_ib_pf_eq *eq = pfault->eq; 1382 1383 mlx5_ib_pfault(eq->dev, pfault); 1384 mempool_free(pfault, eq->pool); 1385} 1386 1387static void mlx5_ib_eq_pf_process(struct mlx5_ib_pf_eq *eq) 1388{ 1389 struct mlx5_eqe_page_fault *pf_eqe; 1390 struct mlx5_pagefault *pfault; 1391 struct mlx5_eqe *eqe; 1392 int cc = 0; 1393 1394 while ((eqe = mlx5_eq_get_eqe(eq->core, cc))) { 1395 pfault = mempool_alloc(eq->pool, GFP_ATOMIC); 1396 if (!pfault) { 1397 schedule_work(&eq->work); 1398 break; 1399 } 1400 1401 pf_eqe = &eqe->data.page_fault; 1402 pfault->event_subtype = eqe->sub_type; 1403 pfault->bytes_committed = be32_to_cpu(pf_eqe->bytes_committed); 1404 1405 mlx5_ib_dbg(eq->dev, 1406 "PAGE_FAULT: subtype: 0x%02x, bytes_committed: 0x%06x\n", 1407 eqe->sub_type, pfault->bytes_committed); 1408 1409 switch (eqe->sub_type) { 1410 case MLX5_PFAULT_SUBTYPE_RDMA: 1411 /* RDMA based event */ 1412 pfault->type = 1413 be32_to_cpu(pf_eqe->rdma.pftype_token) >> 24; 1414 pfault->token = 1415 be32_to_cpu(pf_eqe->rdma.pftype_token) & 1416 MLX5_24BIT_MASK; 1417 pfault->rdma.r_key = 1418 be32_to_cpu(pf_eqe->rdma.r_key); 1419 pfault->rdma.packet_size = 1420 be16_to_cpu(pf_eqe->rdma.packet_length); 1421 pfault->rdma.rdma_op_len = 1422 be32_to_cpu(pf_eqe->rdma.rdma_op_len); 1423 pfault->rdma.rdma_va = 1424 be64_to_cpu(pf_eqe->rdma.rdma_va); 1425 mlx5_ib_dbg(eq->dev, 1426 "PAGE_FAULT: type:0x%x, token: 0x%06x, r_key: 0x%08x\n", 1427 pfault->type, pfault->token, 1428 pfault->rdma.r_key); 1429 mlx5_ib_dbg(eq->dev, 1430 "PAGE_FAULT: rdma_op_len: 0x%08x, rdma_va: 0x%016llx\n", 1431 pfault->rdma.rdma_op_len, 1432 pfault->rdma.rdma_va); 1433 break; 1434 1435 case MLX5_PFAULT_SUBTYPE_WQE: 1436 /* WQE based event */ 1437 pfault->type = 1438 (be32_to_cpu(pf_eqe->wqe.pftype_wq) >> 24) & 0x7; 1439 pfault->token = 1440 be32_to_cpu(pf_eqe->wqe.token); 1441 pfault->wqe.wq_num = 1442 be32_to_cpu(pf_eqe->wqe.pftype_wq) & 1443 MLX5_24BIT_MASK; 1444 pfault->wqe.wqe_index = 1445 be16_to_cpu(pf_eqe->wqe.wqe_index); 1446 pfault->wqe.packet_size = 1447 be16_to_cpu(pf_eqe->wqe.packet_length); 1448 mlx5_ib_dbg(eq->dev, 1449 "PAGE_FAULT: type:0x%x, token: 0x%06x, wq_num: 0x%06x, wqe_index: 0x%04x\n", 1450 pfault->type, pfault->token, 1451 pfault->wqe.wq_num, 1452 pfault->wqe.wqe_index); 1453 break; 1454 1455 default: 1456 mlx5_ib_warn(eq->dev, 1457 "Unsupported page fault event sub-type: 0x%02hhx\n", 1458 eqe->sub_type); 1459 /* Unsupported page faults should still be 1460 * resolved by the page fault handler 1461 */ 1462 } 1463 1464 pfault->eq = eq; 1465 INIT_WORK(&pfault->work, mlx5_ib_eqe_pf_action); 1466 queue_work(eq->wq, &pfault->work); 1467 1468 cc = mlx5_eq_update_cc(eq->core, ++cc); 1469 } 1470 1471 mlx5_eq_update_ci(eq->core, cc, 1); 1472} 1473 1474static int mlx5_ib_eq_pf_int(struct notifier_block *nb, unsigned long type, 1475 void *data) 1476{ 1477 struct mlx5_ib_pf_eq *eq = 1478 container_of(nb, struct mlx5_ib_pf_eq, irq_nb); 1479 unsigned long flags; 1480 1481 if (spin_trylock_irqsave(&eq->lock, flags)) { 1482 mlx5_ib_eq_pf_process(eq); 1483 spin_unlock_irqrestore(&eq->lock, flags); 1484 } else { 1485 schedule_work(&eq->work); 1486 } 1487 1488 return IRQ_HANDLED; 1489} 1490 1491/* mempool_refill() was proposed but unfortunately wasn't accepted 1492 * http://lkml.iu.edu/hypermail/linux/kernel/1512.1/05073.html 1493 * Cheap workaround. 1494 */ 1495static void mempool_refill(mempool_t *pool) 1496{ 1497 while (pool->curr_nr < pool->min_nr) 1498 mempool_free(mempool_alloc(pool, GFP_KERNEL), pool); 1499} 1500 1501static void mlx5_ib_eq_pf_action(struct work_struct *work) 1502{ 1503 struct mlx5_ib_pf_eq *eq = 1504 container_of(work, struct mlx5_ib_pf_eq, work); 1505 1506 mempool_refill(eq->pool); 1507 1508 spin_lock_irq(&eq->lock); 1509 mlx5_ib_eq_pf_process(eq); 1510 spin_unlock_irq(&eq->lock); 1511} 1512 1513enum { 1514 MLX5_IB_NUM_PF_EQE = 0x1000, 1515 MLX5_IB_NUM_PF_DRAIN = 64, 1516}; 1517 1518int mlx5r_odp_create_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq) 1519{ 1520 struct mlx5_eq_param param = {}; 1521 int err = 0; 1522 1523 mutex_lock(&dev->odp_eq_mutex); 1524 if (eq->core) 1525 goto unlock; 1526 INIT_WORK(&eq->work, mlx5_ib_eq_pf_action); 1527 spin_lock_init(&eq->lock); 1528 eq->dev = dev; 1529 1530 eq->pool = mempool_create_kmalloc_pool(MLX5_IB_NUM_PF_DRAIN, 1531 sizeof(struct mlx5_pagefault)); 1532 if (!eq->pool) { 1533 err = -ENOMEM; 1534 goto unlock; 1535 } 1536 1537 eq->wq = alloc_workqueue("mlx5_ib_page_fault", 1538 WQ_HIGHPRI | WQ_UNBOUND | WQ_MEM_RECLAIM, 1539 MLX5_NUM_CMD_EQE); 1540 if (!eq->wq) { 1541 err = -ENOMEM; 1542 goto err_mempool; 1543 } 1544 1545 eq->irq_nb.notifier_call = mlx5_ib_eq_pf_int; 1546 param = (struct mlx5_eq_param) { 1547 .nent = MLX5_IB_NUM_PF_EQE, 1548 }; 1549 param.mask[0] = 1ull << MLX5_EVENT_TYPE_PAGE_FAULT; 1550 eq->core = mlx5_eq_create_generic(dev->mdev, ¶m); 1551 if (IS_ERR(eq->core)) { 1552 err = PTR_ERR(eq->core); 1553 goto err_wq; 1554 } 1555 err = mlx5_eq_enable(dev->mdev, eq->core, &eq->irq_nb); 1556 if (err) { 1557 mlx5_ib_err(dev, "failed to enable odp EQ %d\n", err); 1558 goto err_eq; 1559 } 1560 1561 mutex_unlock(&dev->odp_eq_mutex); 1562 return 0; 1563err_eq: 1564 mlx5_eq_destroy_generic(dev->mdev, eq->core); 1565err_wq: 1566 eq->core = NULL; 1567 destroy_workqueue(eq->wq); 1568err_mempool: 1569 mempool_destroy(eq->pool); 1570unlock: 1571 mutex_unlock(&dev->odp_eq_mutex); 1572 return err; 1573} 1574 1575static int 1576mlx5_ib_odp_destroy_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq) 1577{ 1578 int err; 1579 1580 if (!eq->core) 1581 return 0; 1582 mlx5_eq_disable(dev->mdev, eq->core, &eq->irq_nb); 1583 err = mlx5_eq_destroy_generic(dev->mdev, eq->core); 1584 cancel_work_sync(&eq->work); 1585 destroy_workqueue(eq->wq); 1586 mempool_destroy(eq->pool); 1587 1588 return err; 1589} 1590 1591int mlx5_odp_init_mkey_cache(struct mlx5_ib_dev *dev) 1592{ 1593 struct mlx5r_cache_rb_key rb_key = { 1594 .access_mode = MLX5_MKC_ACCESS_MODE_KSM, 1595 .ndescs = mlx5_imr_ksm_entries, 1596 }; 1597 struct mlx5_cache_ent *ent; 1598 1599 if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT)) 1600 return 0; 1601 1602 ent = mlx5r_cache_create_ent_locked(dev, rb_key, true); 1603 if (IS_ERR(ent)) 1604 return PTR_ERR(ent); 1605 1606 return 0; 1607} 1608 1609static const struct ib_device_ops mlx5_ib_dev_odp_ops = { 1610 .advise_mr = mlx5_ib_advise_mr, 1611}; 1612 1613int mlx5_ib_odp_init_one(struct mlx5_ib_dev *dev) 1614{ 1615 internal_fill_odp_caps(dev); 1616 1617 if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT)) 1618 return 0; 1619 1620 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_odp_ops); 1621 1622 mutex_init(&dev->odp_eq_mutex); 1623 return 0; 1624} 1625 1626void mlx5_ib_odp_cleanup_one(struct mlx5_ib_dev *dev) 1627{ 1628 if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT)) 1629 return; 1630 1631 mlx5_ib_odp_destroy_eq(dev, &dev->odp_pf_eq); 1632} 1633 1634int mlx5_ib_odp_init(void) 1635{ 1636 mlx5_imr_ksm_entries = BIT_ULL(get_order(TASK_SIZE) - 1637 MLX5_IMR_MTT_BITS); 1638 1639 return 0; 1640} 1641 1642struct prefetch_mr_work { 1643 struct work_struct work; 1644 u32 pf_flags; 1645 u32 num_sge; 1646 struct { 1647 u64 io_virt; 1648 struct mlx5_ib_mr *mr; 1649 size_t length; 1650 } frags[]; 1651}; 1652 1653static void destroy_prefetch_work(struct prefetch_mr_work *work) 1654{ 1655 u32 i; 1656 1657 for (i = 0; i < work->num_sge; ++i) 1658 mlx5r_deref_odp_mkey(&work->frags[i].mr->mmkey); 1659 1660 kvfree(work); 1661} 1662 1663static struct mlx5_ib_mr * 1664get_prefetchable_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice, 1665 u32 lkey) 1666{ 1667 struct mlx5_ib_dev *dev = to_mdev(pd->device); 1668 struct mlx5_ib_mr *mr = NULL; 1669 struct mlx5_ib_mkey *mmkey; 1670 1671 xa_lock(&dev->odp_mkeys); 1672 mmkey = xa_load(&dev->odp_mkeys, mlx5_base_mkey(lkey)); 1673 if (!mmkey || mmkey->key != lkey) { 1674 mr = ERR_PTR(-ENOENT); 1675 goto end; 1676 } 1677 if (mmkey->type != MLX5_MKEY_MR) { 1678 mr = ERR_PTR(-EINVAL); 1679 goto end; 1680 } 1681 1682 mr = container_of(mmkey, struct mlx5_ib_mr, mmkey); 1683 1684 if (mr->ibmr.pd != pd) { 1685 mr = ERR_PTR(-EPERM); 1686 goto end; 1687 } 1688 1689 /* prefetch with write-access must be supported by the MR */ 1690 if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && 1691 !mr->umem->writable) { 1692 mr = ERR_PTR(-EPERM); 1693 goto end; 1694 } 1695 1696 refcount_inc(&mmkey->usecount); 1697end: 1698 xa_unlock(&dev->odp_mkeys); 1699 return mr; 1700} 1701 1702static void mlx5_ib_prefetch_mr_work(struct work_struct *w) 1703{ 1704 struct prefetch_mr_work *work = 1705 container_of(w, struct prefetch_mr_work, work); 1706 u32 bytes_mapped = 0; 1707 int ret; 1708 u32 i; 1709 1710 /* We rely on IB/core that work is executed if we have num_sge != 0 only. */ 1711 WARN_ON(!work->num_sge); 1712 for (i = 0; i < work->num_sge; ++i) { 1713 ret = pagefault_mr(work->frags[i].mr, work->frags[i].io_virt, 1714 work->frags[i].length, &bytes_mapped, 1715 work->pf_flags); 1716 if (ret <= 0) 1717 continue; 1718 mlx5_update_odp_stats(work->frags[i].mr, prefetch, ret); 1719 } 1720 1721 destroy_prefetch_work(work); 1722} 1723 1724static int init_prefetch_work(struct ib_pd *pd, 1725 enum ib_uverbs_advise_mr_advice advice, 1726 u32 pf_flags, struct prefetch_mr_work *work, 1727 struct ib_sge *sg_list, u32 num_sge) 1728{ 1729 u32 i; 1730 1731 INIT_WORK(&work->work, mlx5_ib_prefetch_mr_work); 1732 work->pf_flags = pf_flags; 1733 1734 for (i = 0; i < num_sge; ++i) { 1735 struct mlx5_ib_mr *mr; 1736 1737 mr = get_prefetchable_mr(pd, advice, sg_list[i].lkey); 1738 if (IS_ERR(mr)) { 1739 work->num_sge = i; 1740 return PTR_ERR(mr); 1741 } 1742 work->frags[i].io_virt = sg_list[i].addr; 1743 work->frags[i].length = sg_list[i].length; 1744 work->frags[i].mr = mr; 1745 } 1746 work->num_sge = num_sge; 1747 return 0; 1748} 1749 1750static int mlx5_ib_prefetch_sg_list(struct ib_pd *pd, 1751 enum ib_uverbs_advise_mr_advice advice, 1752 u32 pf_flags, struct ib_sge *sg_list, 1753 u32 num_sge) 1754{ 1755 u32 bytes_mapped = 0; 1756 int ret = 0; 1757 u32 i; 1758 1759 for (i = 0; i < num_sge; ++i) { 1760 struct mlx5_ib_mr *mr; 1761 1762 mr = get_prefetchable_mr(pd, advice, sg_list[i].lkey); 1763 if (IS_ERR(mr)) 1764 return PTR_ERR(mr); 1765 ret = pagefault_mr(mr, sg_list[i].addr, sg_list[i].length, 1766 &bytes_mapped, pf_flags); 1767 if (ret < 0) { 1768 mlx5r_deref_odp_mkey(&mr->mmkey); 1769 return ret; 1770 } 1771 mlx5_update_odp_stats(mr, prefetch, ret); 1772 mlx5r_deref_odp_mkey(&mr->mmkey); 1773 } 1774 1775 return 0; 1776} 1777 1778int mlx5_ib_advise_mr_prefetch(struct ib_pd *pd, 1779 enum ib_uverbs_advise_mr_advice advice, 1780 u32 flags, struct ib_sge *sg_list, u32 num_sge) 1781{ 1782 u32 pf_flags = 0; 1783 struct prefetch_mr_work *work; 1784 int rc; 1785 1786 if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH) 1787 pf_flags |= MLX5_PF_FLAGS_DOWNGRADE; 1788 1789 if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT) 1790 pf_flags |= MLX5_PF_FLAGS_SNAPSHOT; 1791 1792 if (flags & IB_UVERBS_ADVISE_MR_FLAG_FLUSH) 1793 return mlx5_ib_prefetch_sg_list(pd, advice, pf_flags, sg_list, 1794 num_sge); 1795 1796 work = kvzalloc(struct_size(work, frags, num_sge), GFP_KERNEL); 1797 if (!work) 1798 return -ENOMEM; 1799 1800 rc = init_prefetch_work(pd, advice, pf_flags, work, sg_list, num_sge); 1801 if (rc) { 1802 destroy_prefetch_work(work); 1803 return rc; 1804 } 1805 queue_work(system_unbound_wq, &work->work); 1806 return 0; 1807} 1808