1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0 3 * 4 * Copyright (c) 2014 Mellanox Technologies. All rights reserved. 5 * 6 * This software is available to you under a choice of one of two 7 * licenses. You may choose to be licensed under the terms of the GNU 8 * General Public License (GPL) Version 2, available from the file 9 * COPYING in the main directory of this source tree, or the 10 * OpenIB.org BSD license below: 11 * 12 * Redistribution and use in source and binary forms, with or 13 * without modification, are permitted provided that the following 14 * conditions are met: 15 * 16 * - Redistributions of source code must retain the above 17 * copyright notice, this list of conditions and the following 18 * disclaimer. 19 * 20 * - Redistributions in binary form must reproduce the above 21 * copyright notice, this list of conditions and the following 22 * disclaimer in the documentation and/or other materials 23 * provided with the distribution. 24 * 25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 32 * SOFTWARE. 33 */ 34 35 #include <sys/cdefs.h> 36 #include <linux/types.h> 37 #include <linux/sched.h> 38 #include <linux/slab.h> 39 #include <linux/vmalloc.h> 40 41 #include <rdma/ib_verbs.h> 42 #include <rdma/ib_umem.h> 43 #include <rdma/ib_umem_odp.h> 44 45 static void ib_umem_notifier_start_account(struct ib_umem *item) 46 { 47 mutex_lock(&item->odp_data->umem_mutex); 48 49 /* Only update private counters for this umem if it has them. 50 * Otherwise skip it. All page faults will be delayed for this umem. */ 51 if (item->odp_data->mn_counters_active) { 52 int notifiers_count = item->odp_data->notifiers_count++; 53 54 if (notifiers_count == 0) 55 /* Initialize the completion object for waiting on 56 * notifiers. Since notifier_count is zero, no one 57 * should be waiting right now. */ 58 reinit_completion(&item->odp_data->notifier_completion); 59 } 60 mutex_unlock(&item->odp_data->umem_mutex); 61 } 62 63 static void ib_umem_notifier_end_account(struct ib_umem *item) 64 { 65 mutex_lock(&item->odp_data->umem_mutex); 66 67 /* Only update private counters for this umem if it has them. 68 * Otherwise skip it. All page faults will be delayed for this umem. */ 69 if (item->odp_data->mn_counters_active) { 70 /* 71 * This sequence increase will notify the QP page fault that 72 * the page that is going to be mapped in the spte could have 73 * been freed. 74 */ 75 ++item->odp_data->notifiers_seq; 76 if (--item->odp_data->notifiers_count == 0) 77 complete_all(&item->odp_data->notifier_completion); 78 } 79 mutex_unlock(&item->odp_data->umem_mutex); 80 } 81 82 /* Account for a new mmu notifier in an ib_ucontext. */ 83 static void ib_ucontext_notifier_start_account(struct ib_ucontext *context) 84 { 85 atomic_inc(&context->notifier_count); 86 } 87 88 /* Account for a terminating mmu notifier in an ib_ucontext. 89 * 90 * Must be called with the ib_ucontext->umem_rwsem semaphore unlocked, since 91 * the function takes the semaphore itself. */ 92 static void ib_ucontext_notifier_end_account(struct ib_ucontext *context) 93 { 94 int zero_notifiers = atomic_dec_and_test(&context->notifier_count); 95 96 if (zero_notifiers && 97 !list_empty(&context->no_private_counters)) { 98 /* No currently running mmu notifiers. Now is the chance to 99 * add private accounting to all previously added umems. */ 100 struct ib_umem_odp *odp_data, *next; 101 102 /* Prevent concurrent mmu notifiers from working on the 103 * no_private_counters list. */ 104 down_write(&context->umem_rwsem); 105 106 /* Read the notifier_count again, with the umem_rwsem 107 * semaphore taken for write. */ 108 if (!atomic_read(&context->notifier_count)) { 109 list_for_each_entry_safe(odp_data, next, 110 &context->no_private_counters, 111 no_private_counters) { 112 mutex_lock(&odp_data->umem_mutex); 113 odp_data->mn_counters_active = true; 114 list_del(&odp_data->no_private_counters); 115 complete_all(&odp_data->notifier_completion); 116 mutex_unlock(&odp_data->umem_mutex); 117 } 118 } 119 120 up_write(&context->umem_rwsem); 121 } 122 } 123 124 static int ib_umem_notifier_release_trampoline(struct ib_umem *item, u64 start, 125 u64 end, void *cookie) { 126 /* 127 * Increase the number of notifiers running, to 128 * prevent any further fault handling on this MR. 129 */ 130 ib_umem_notifier_start_account(item); 131 item->odp_data->dying = 1; 132 /* Make sure that the fact the umem is dying is out before we release 133 * all pending page faults. */ 134 smp_wmb(); 135 complete_all(&item->odp_data->notifier_completion); 136 item->context->invalidate_range(item, ib_umem_start(item), 137 ib_umem_end(item)); 138 return 0; 139 } 140 141 static void ib_umem_notifier_release(struct mmu_notifier *mn, 142 struct mm_struct *mm) 143 { 144 struct ib_ucontext *context = container_of(mn, struct ib_ucontext, mn); 145 146 if (!context->invalidate_range) 147 return; 148 149 ib_ucontext_notifier_start_account(context); 150 down_read(&context->umem_rwsem); 151 rbt_ib_umem_for_each_in_range(&context->umem_tree, 0, 152 ULLONG_MAX, 153 ib_umem_notifier_release_trampoline, 154 NULL); 155 up_read(&context->umem_rwsem); 156 } 157 158 static int invalidate_page_trampoline(struct ib_umem *item, u64 start, 159 u64 end, void *cookie) 160 { 161 ib_umem_notifier_start_account(item); 162 item->context->invalidate_range(item, start, start + PAGE_SIZE); 163 ib_umem_notifier_end_account(item); 164 return 0; 165 } 166 167 static void ib_umem_notifier_invalidate_page(struct mmu_notifier *mn, 168 struct mm_struct *mm, 169 unsigned long address) 170 { 171 struct ib_ucontext *context = container_of(mn, struct ib_ucontext, mn); 172 173 if (!context->invalidate_range) 174 return; 175 176 ib_ucontext_notifier_start_account(context); 177 down_read(&context->umem_rwsem); 178 rbt_ib_umem_for_each_in_range(&context->umem_tree, address, 179 address + PAGE_SIZE, 180 invalidate_page_trampoline, NULL); 181 up_read(&context->umem_rwsem); 182 ib_ucontext_notifier_end_account(context); 183 } 184 185 static int invalidate_range_start_trampoline(struct ib_umem *item, u64 start, 186 u64 end, void *cookie) 187 { 188 ib_umem_notifier_start_account(item); 189 item->context->invalidate_range(item, start, end); 190 return 0; 191 } 192 193 static void ib_umem_notifier_invalidate_range_start(struct mmu_notifier *mn, 194 struct mm_struct *mm, 195 unsigned long start, 196 unsigned long end) 197 { 198 struct ib_ucontext *context = container_of(mn, struct ib_ucontext, mn); 199 200 if (!context->invalidate_range) 201 return; 202 203 ib_ucontext_notifier_start_account(context); 204 down_read(&context->umem_rwsem); 205 rbt_ib_umem_for_each_in_range(&context->umem_tree, start, 206 end, 207 invalidate_range_start_trampoline, NULL); 208 up_read(&context->umem_rwsem); 209 } 210 211 static int invalidate_range_end_trampoline(struct ib_umem *item, u64 start, 212 u64 end, void *cookie) 213 { 214 ib_umem_notifier_end_account(item); 215 return 0; 216 } 217 218 static void ib_umem_notifier_invalidate_range_end(struct mmu_notifier *mn, 219 struct mm_struct *mm, 220 unsigned long start, 221 unsigned long end) 222 { 223 struct ib_ucontext *context = container_of(mn, struct ib_ucontext, mn); 224 225 if (!context->invalidate_range) 226 return; 227 228 down_read(&context->umem_rwsem); 229 rbt_ib_umem_for_each_in_range(&context->umem_tree, start, 230 end, 231 invalidate_range_end_trampoline, NULL); 232 up_read(&context->umem_rwsem); 233 ib_ucontext_notifier_end_account(context); 234 } 235 236 static const struct mmu_notifier_ops ib_umem_notifiers = { 237 .release = ib_umem_notifier_release, 238 .invalidate_page = ib_umem_notifier_invalidate_page, 239 .invalidate_range_start = ib_umem_notifier_invalidate_range_start, 240 .invalidate_range_end = ib_umem_notifier_invalidate_range_end, 241 }; 242 243 int ib_umem_odp_get(struct ib_ucontext *context, struct ib_umem *umem) 244 { 245 int ret_val; 246 pid_t our_pid; 247 struct mm_struct *mm = get_task_mm(current); 248 249 if (!mm) 250 return -EINVAL; 251 252 /* Prevent creating ODP MRs in child processes */ 253 rcu_read_lock(); 254 our_pid = get_pid(task_pid_group_leader(current)); 255 rcu_read_unlock(); 256 put_pid(our_pid); 257 if (context->tgid != our_pid) { 258 ret_val = -EINVAL; 259 goto out_mm; 260 } 261 262 umem->odp_data = kzalloc(sizeof(*umem->odp_data), GFP_KERNEL); 263 if (!umem->odp_data) { 264 ret_val = -ENOMEM; 265 goto out_mm; 266 } 267 umem->odp_data->umem = umem; 268 269 mutex_init(&umem->odp_data->umem_mutex); 270 271 init_completion(&umem->odp_data->notifier_completion); 272 273 umem->odp_data->page_list = vzalloc(ib_umem_num_pages(umem) * 274 sizeof(*umem->odp_data->page_list)); 275 if (!umem->odp_data->page_list) { 276 ret_val = -ENOMEM; 277 goto out_odp_data; 278 } 279 280 umem->odp_data->dma_list = vzalloc(ib_umem_num_pages(umem) * 281 sizeof(*umem->odp_data->dma_list)); 282 if (!umem->odp_data->dma_list) { 283 ret_val = -ENOMEM; 284 goto out_page_list; 285 } 286 287 /* 288 * When using MMU notifiers, we will get a 289 * notification before the "current" task (and MM) is 290 * destroyed. We use the umem_rwsem semaphore to synchronize. 291 */ 292 down_write(&context->umem_rwsem); 293 context->odp_mrs_count++; 294 if (likely(ib_umem_start(umem) != ib_umem_end(umem))) 295 rbt_ib_umem_insert(&umem->odp_data->interval_tree, 296 &context->umem_tree); 297 if (likely(!atomic_read(&context->notifier_count)) || 298 context->odp_mrs_count == 1) 299 umem->odp_data->mn_counters_active = true; 300 else 301 list_add(&umem->odp_data->no_private_counters, 302 &context->no_private_counters); 303 downgrade_write(&context->umem_rwsem); 304 305 if (context->odp_mrs_count == 1) { 306 /* 307 * Note that at this point, no MMU notifier is running 308 * for this context! 309 */ 310 atomic_set(&context->notifier_count, 0); 311 INIT_HLIST_NODE(&context->mn.hlist); 312 context->mn.ops = &ib_umem_notifiers; 313 /* 314 * Lock-dep detects a false positive for mmap_sem vs. 315 * umem_rwsem, due to not grasping downgrade_write correctly. 316 */ 317 ret_val = mmu_notifier_register(&context->mn, mm); 318 if (ret_val) { 319 pr_err("Failed to register mmu_notifier %d\n", ret_val); 320 ret_val = -EBUSY; 321 goto out_mutex; 322 } 323 } 324 325 up_read(&context->umem_rwsem); 326 327 /* 328 * Note that doing an mmput can cause a notifier for the relevant mm. 329 * If the notifier is called while we hold the umem_rwsem, this will 330 * cause a deadlock. Therefore, we release the reference only after we 331 * released the semaphore. 332 */ 333 mmput(mm); 334 return 0; 335 336 out_mutex: 337 up_read(&context->umem_rwsem); 338 vfree(umem->odp_data->dma_list); 339 out_page_list: 340 vfree(umem->odp_data->page_list); 341 out_odp_data: 342 kfree(umem->odp_data); 343 out_mm: 344 mmput(mm); 345 return ret_val; 346 } 347 348 void ib_umem_odp_release(struct ib_umem *umem) 349 { 350 struct ib_ucontext *context = umem->context; 351 352 /* 353 * Ensure that no more pages are mapped in the umem. 354 * 355 * It is the driver's responsibility to ensure, before calling us, 356 * that the hardware will not attempt to access the MR any more. 357 */ 358 ib_umem_odp_unmap_dma_pages(umem, ib_umem_start(umem), 359 ib_umem_end(umem)); 360 361 down_write(&context->umem_rwsem); 362 if (likely(ib_umem_start(umem) != ib_umem_end(umem))) 363 rbt_ib_umem_remove(&umem->odp_data->interval_tree, 364 &context->umem_tree); 365 context->odp_mrs_count--; 366 if (!umem->odp_data->mn_counters_active) { 367 list_del(&umem->odp_data->no_private_counters); 368 complete_all(&umem->odp_data->notifier_completion); 369 } 370 371 /* 372 * Downgrade the lock to a read lock. This ensures that the notifiers 373 * (who lock the mutex for reading) will be able to finish, and we 374 * will be able to enventually obtain the mmu notifiers SRCU. Note 375 * that since we are doing it atomically, no other user could register 376 * and unregister while we do the check. 377 */ 378 downgrade_write(&context->umem_rwsem); 379 if (!context->odp_mrs_count) { 380 struct task_struct *owning_process = NULL; 381 struct mm_struct *owning_mm = NULL; 382 383 owning_process = get_pid_task(context->tgid, 384 PIDTYPE_PID); 385 if (owning_process == NULL) 386 /* 387 * The process is already dead, notifier were removed 388 * already. 389 */ 390 goto out; 391 392 owning_mm = get_task_mm(owning_process); 393 if (owning_mm == NULL) 394 /* 395 * The process' mm is already dead, notifier were 396 * removed already. 397 */ 398 goto out_put_task; 399 mmu_notifier_unregister(&context->mn, owning_mm); 400 401 mmput(owning_mm); 402 403 out_put_task: 404 put_task_struct(owning_process); 405 } 406 out: 407 up_read(&context->umem_rwsem); 408 409 vfree(umem->odp_data->dma_list); 410 vfree(umem->odp_data->page_list); 411 kfree(umem->odp_data); 412 kfree(umem); 413 } 414 415 /* 416 * Map for DMA and insert a single page into the on-demand paging page tables. 417 * 418 * @umem: the umem to insert the page to. 419 * @page_index: index in the umem to add the page to. 420 * @page: the page struct to map and add. 421 * @access_mask: access permissions needed for this page. 422 * @current_seq: sequence number for synchronization with invalidations. 423 * the sequence number is taken from 424 * umem->odp_data->notifiers_seq. 425 * 426 * The function returns -EFAULT if the DMA mapping operation fails. It returns 427 * -EAGAIN if a concurrent invalidation prevents us from updating the page. 428 * 429 * The page is released via put_page even if the operation failed. For 430 * on-demand pinning, the page is released whenever it isn't stored in the 431 * umem. 432 */ 433 static int ib_umem_odp_map_dma_single_page( 434 struct ib_umem *umem, 435 int page_index, 436 u64 base_virt_addr, 437 struct page *page, 438 u64 access_mask, 439 unsigned long current_seq) 440 { 441 struct ib_device *dev = umem->context->device; 442 dma_addr_t dma_addr; 443 int stored_page = 0; 444 int remove_existing_mapping = 0; 445 int ret = 0; 446 447 /* 448 * Note: we avoid writing if seq is different from the initial seq, to 449 * handle case of a racing notifier. This check also allows us to bail 450 * early if we have a notifier running in parallel with us. 451 */ 452 if (ib_umem_mmu_notifier_retry(umem, current_seq)) { 453 ret = -EAGAIN; 454 goto out; 455 } 456 if (!(umem->odp_data->dma_list[page_index])) { 457 dma_addr = ib_dma_map_page(dev, 458 page, 459 0, PAGE_SIZE, 460 DMA_BIDIRECTIONAL); 461 if (ib_dma_mapping_error(dev, dma_addr)) { 462 ret = -EFAULT; 463 goto out; 464 } 465 umem->odp_data->dma_list[page_index] = dma_addr | access_mask; 466 umem->odp_data->page_list[page_index] = page; 467 stored_page = 1; 468 } else if (umem->odp_data->page_list[page_index] == page) { 469 umem->odp_data->dma_list[page_index] |= access_mask; 470 } else { 471 pr_err("error: got different pages in IB device and from get_user_pages. IB device page: %p, gup page: %p\n", 472 umem->odp_data->page_list[page_index], page); 473 /* Better remove the mapping now, to prevent any further 474 * damage. */ 475 remove_existing_mapping = 1; 476 } 477 478 out: 479 /* On Demand Paging - avoid pinning the page */ 480 if (umem->context->invalidate_range || !stored_page) 481 put_page(page); 482 483 if (remove_existing_mapping && umem->context->invalidate_range) { 484 invalidate_page_trampoline( 485 umem, 486 base_virt_addr + (page_index * PAGE_SIZE), 487 base_virt_addr + ((page_index+1)*PAGE_SIZE), 488 NULL); 489 ret = -EAGAIN; 490 } 491 492 return ret; 493 } 494 495 /** 496 * ib_umem_odp_map_dma_pages - Pin and DMA map userspace memory in an ODP MR. 497 * 498 * Pins the range of pages passed in the argument, and maps them to 499 * DMA addresses. The DMA addresses of the mapped pages is updated in 500 * umem->odp_data->dma_list. 501 * 502 * Returns the number of pages mapped in success, negative error code 503 * for failure. 504 * An -EAGAIN error code is returned when a concurrent mmu notifier prevents 505 * the function from completing its task. 506 * 507 * @umem: the umem to map and pin 508 * @user_virt: the address from which we need to map. 509 * @bcnt: the minimal number of bytes to pin and map. The mapping might be 510 * bigger due to alignment, and may also be smaller in case of an error 511 * pinning or mapping a page. The actual pages mapped is returned in 512 * the return value. 513 * @access_mask: bit mask of the requested access permissions for the given 514 * range. 515 * @current_seq: the MMU notifiers sequance value for synchronization with 516 * invalidations. the sequance number is read from 517 * umem->odp_data->notifiers_seq before calling this function 518 */ 519 int ib_umem_odp_map_dma_pages(struct ib_umem *umem, u64 user_virt, u64 bcnt, 520 u64 access_mask, unsigned long current_seq) 521 { 522 struct task_struct *owning_process = NULL; 523 struct mm_struct *owning_mm = NULL; 524 struct page **local_page_list = NULL; 525 u64 off; 526 int j, k, ret = 0, start_idx, npages = 0; 527 u64 base_virt_addr; 528 unsigned int flags = 0; 529 530 if (access_mask == 0) 531 return -EINVAL; 532 533 if (user_virt < ib_umem_start(umem) || 534 user_virt + bcnt > ib_umem_end(umem)) 535 return -EFAULT; 536 537 local_page_list = (struct page **)__get_free_page(GFP_KERNEL); 538 if (!local_page_list) 539 return -ENOMEM; 540 541 off = user_virt & (~PAGE_MASK); 542 user_virt = user_virt & PAGE_MASK; 543 base_virt_addr = user_virt; 544 bcnt += off; /* Charge for the first page offset as well. */ 545 546 owning_process = get_pid_task(umem->context->tgid, PIDTYPE_PID); 547 if (owning_process == NULL) { 548 ret = -EINVAL; 549 goto out_no_task; 550 } 551 552 owning_mm = get_task_mm(owning_process); 553 if (owning_mm == NULL) { 554 ret = -EINVAL; 555 goto out_put_task; 556 } 557 558 if (access_mask & ODP_WRITE_ALLOWED_BIT) 559 flags |= FOLL_WRITE; 560 561 start_idx = (user_virt - ib_umem_start(umem)) >> PAGE_SHIFT; 562 k = start_idx; 563 564 while (bcnt > 0) { 565 const size_t gup_num_pages = 566 min_t(size_t, ALIGN(bcnt, PAGE_SIZE) / PAGE_SIZE, 567 PAGE_SIZE / sizeof(struct page *)); 568 569 down_read(&owning_mm->mmap_sem); 570 /* 571 * Note: this might result in redundent page getting. We can 572 * avoid this by checking dma_list to be 0 before calling 573 * get_user_pages. However, this make the code much more 574 * complex (and doesn't gain us much performance in most use 575 * cases). 576 */ 577 npages = get_user_pages_remote(owning_process, owning_mm, 578 user_virt, gup_num_pages, 579 flags, local_page_list, NULL); 580 up_read(&owning_mm->mmap_sem); 581 582 if (npages < 0) 583 break; 584 585 bcnt -= min_t(size_t, npages << PAGE_SHIFT, bcnt); 586 user_virt += npages << PAGE_SHIFT; 587 mutex_lock(&umem->odp_data->umem_mutex); 588 for (j = 0; j < npages; ++j) { 589 ret = ib_umem_odp_map_dma_single_page( 590 umem, k, base_virt_addr, local_page_list[j], 591 access_mask, current_seq); 592 if (ret < 0) 593 break; 594 k++; 595 } 596 mutex_unlock(&umem->odp_data->umem_mutex); 597 598 if (ret < 0) { 599 /* Release left over pages when handling errors. */ 600 for (++j; j < npages; ++j) 601 put_page(local_page_list[j]); 602 break; 603 } 604 } 605 606 if (ret >= 0) { 607 if (npages < 0 && k == start_idx) 608 ret = npages; 609 else 610 ret = k - start_idx; 611 } 612 613 mmput(owning_mm); 614 out_put_task: 615 put_task_struct(owning_process); 616 out_no_task: 617 free_page((unsigned long)local_page_list); 618 return ret; 619 } 620 EXPORT_SYMBOL(ib_umem_odp_map_dma_pages); 621 622 void ib_umem_odp_unmap_dma_pages(struct ib_umem *umem, u64 virt, 623 u64 bound) 624 { 625 int idx; 626 u64 addr; 627 struct ib_device *dev = umem->context->device; 628 629 virt = max_t(u64, virt, ib_umem_start(umem)); 630 bound = min_t(u64, bound, ib_umem_end(umem)); 631 /* Note that during the run of this function, the 632 * notifiers_count of the MR is > 0, preventing any racing 633 * faults from completion. We might be racing with other 634 * invalidations, so we must make sure we free each page only 635 * once. */ 636 mutex_lock(&umem->odp_data->umem_mutex); 637 for (addr = virt; addr < bound; addr += (u64)umem->page_size) { 638 idx = (addr - ib_umem_start(umem)) / PAGE_SIZE; 639 if (umem->odp_data->page_list[idx]) { 640 struct page *page = umem->odp_data->page_list[idx]; 641 dma_addr_t dma = umem->odp_data->dma_list[idx]; 642 dma_addr_t dma_addr = dma & ODP_DMA_ADDR_MASK; 643 644 WARN_ON(!dma_addr); 645 646 ib_dma_unmap_page(dev, dma_addr, PAGE_SIZE, 647 DMA_BIDIRECTIONAL); 648 if (dma & ODP_WRITE_ALLOWED_BIT) { 649 struct page *head_page = compound_head(page); 650 /* 651 * set_page_dirty prefers being called with 652 * the page lock. However, MMU notifiers are 653 * called sometimes with and sometimes without 654 * the lock. We rely on the umem_mutex instead 655 * to prevent other mmu notifiers from 656 * continuing and allowing the page mapping to 657 * be removed. 658 */ 659 set_page_dirty(head_page); 660 } 661 /* on demand pinning support */ 662 if (!umem->context->invalidate_range) 663 put_page(page); 664 umem->odp_data->page_list[idx] = NULL; 665 umem->odp_data->dma_list[idx] = 0; 666 } 667 } 668 mutex_unlock(&umem->odp_data->umem_mutex); 669 } 670 EXPORT_SYMBOL(ib_umem_odp_unmap_dma_pages); 671