2 * Copyright (c) 2014 Mellanox Technologies. All rights reserved.
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:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
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.
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
33 #include <linux/types.h>
34 #include <linux/sched.h>
35 #include <linux/pid.h>
36 #include <linux/slab.h>
37 #include <linux/export.h>
38 #include <linux/vmalloc.h>
40 #include <rdma/ib_verbs.h>
41 #include <rdma/ib_umem.h>
42 #include <rdma/ib_umem_odp.h>
44 static void ib_umem_notifier_start_account(struct ib_umem *item)
46 mutex_lock(&item->odp_data->umem_mutex);
48 /* Only update private counters for this umem if it has them.
49 * Otherwise skip it. All page faults will be delayed for this umem. */
50 if (item->odp_data->mn_counters_active) {
51 int notifiers_count = item->odp_data->notifiers_count++;
53 if (notifiers_count == 0)
54 /* Initialize the completion object for waiting on
55 * notifiers. Since notifier_count is zero, no one
56 * should be waiting right now. */
57 reinit_completion(&item->odp_data->notifier_completion);
59 mutex_unlock(&item->odp_data->umem_mutex);
62 static void ib_umem_notifier_end_account(struct ib_umem *item)
64 mutex_lock(&item->odp_data->umem_mutex);
66 /* Only update private counters for this umem if it has them.
67 * Otherwise skip it. All page faults will be delayed for this umem. */
68 if (item->odp_data->mn_counters_active) {
70 * This sequence increase will notify the QP page fault that
71 * the page that is going to be mapped in the spte could have
74 ++item->odp_data->notifiers_seq;
75 if (--item->odp_data->notifiers_count == 0)
76 complete_all(&item->odp_data->notifier_completion);
78 mutex_unlock(&item->odp_data->umem_mutex);
81 /* Account for a new mmu notifier in an ib_ucontext. */
82 static void ib_ucontext_notifier_start_account(struct ib_ucontext *context)
84 atomic_inc(&context->notifier_count);
87 /* Account for a terminating mmu notifier in an ib_ucontext.
89 * Must be called with the ib_ucontext->umem_rwsem semaphore unlocked, since
90 * the function takes the semaphore itself. */
91 static void ib_ucontext_notifier_end_account(struct ib_ucontext *context)
93 int zero_notifiers = atomic_dec_and_test(&context->notifier_count);
96 !list_empty(&context->no_private_counters)) {
97 /* No currently running mmu notifiers. Now is the chance to
98 * add private accounting to all previously added umems. */
99 struct ib_umem_odp *odp_data, *next;
101 /* Prevent concurrent mmu notifiers from working on the
102 * no_private_counters list. */
103 down_write(&context->umem_rwsem);
105 /* Read the notifier_count again, with the umem_rwsem
106 * semaphore taken for write. */
107 if (!atomic_read(&context->notifier_count)) {
108 list_for_each_entry_safe(odp_data, next,
109 &context->no_private_counters,
110 no_private_counters) {
111 mutex_lock(&odp_data->umem_mutex);
112 odp_data->mn_counters_active = true;
113 list_del(&odp_data->no_private_counters);
114 complete_all(&odp_data->notifier_completion);
115 mutex_unlock(&odp_data->umem_mutex);
119 up_write(&context->umem_rwsem);
123 static int ib_umem_notifier_release_trampoline(struct ib_umem *item, u64 start,
124 u64 end, void *cookie) {
126 * Increase the number of notifiers running, to
127 * prevent any further fault handling on this MR.
129 ib_umem_notifier_start_account(item);
130 item->odp_data->dying = 1;
131 /* Make sure that the fact the umem is dying is out before we release
132 * all pending page faults. */
134 complete_all(&item->odp_data->notifier_completion);
135 item->context->invalidate_range(item, ib_umem_start(item),
140 static void ib_umem_notifier_release(struct mmu_notifier *mn,
141 struct mm_struct *mm)
143 struct ib_ucontext *context = container_of(mn, struct ib_ucontext, mn);
145 if (!context->invalidate_range)
148 ib_ucontext_notifier_start_account(context);
149 down_read(&context->umem_rwsem);
150 rbt_ib_umem_for_each_in_range(&context->umem_tree, 0,
152 ib_umem_notifier_release_trampoline,
154 up_read(&context->umem_rwsem);
157 static int invalidate_page_trampoline(struct ib_umem *item, u64 start,
158 u64 end, void *cookie)
160 ib_umem_notifier_start_account(item);
161 item->context->invalidate_range(item, start, start + PAGE_SIZE);
162 ib_umem_notifier_end_account(item);
166 static void ib_umem_notifier_invalidate_page(struct mmu_notifier *mn,
167 struct mm_struct *mm,
168 unsigned long address)
170 struct ib_ucontext *context = container_of(mn, struct ib_ucontext, mn);
172 if (!context->invalidate_range)
175 ib_ucontext_notifier_start_account(context);
176 down_read(&context->umem_rwsem);
177 rbt_ib_umem_for_each_in_range(&context->umem_tree, address,
179 invalidate_page_trampoline, NULL);
180 up_read(&context->umem_rwsem);
181 ib_ucontext_notifier_end_account(context);
184 static int invalidate_range_start_trampoline(struct ib_umem *item, u64 start,
185 u64 end, void *cookie)
187 ib_umem_notifier_start_account(item);
188 item->context->invalidate_range(item, start, end);
192 static void ib_umem_notifier_invalidate_range_start(struct mmu_notifier *mn,
193 struct mm_struct *mm,
197 struct ib_ucontext *context = container_of(mn, struct ib_ucontext, mn);
199 if (!context->invalidate_range)
202 ib_ucontext_notifier_start_account(context);
203 down_read(&context->umem_rwsem);
204 rbt_ib_umem_for_each_in_range(&context->umem_tree, start,
206 invalidate_range_start_trampoline, NULL);
207 up_read(&context->umem_rwsem);
210 static int invalidate_range_end_trampoline(struct ib_umem *item, u64 start,
211 u64 end, void *cookie)
213 ib_umem_notifier_end_account(item);
217 static void ib_umem_notifier_invalidate_range_end(struct mmu_notifier *mn,
218 struct mm_struct *mm,
222 struct ib_ucontext *context = container_of(mn, struct ib_ucontext, mn);
224 if (!context->invalidate_range)
227 down_read(&context->umem_rwsem);
228 rbt_ib_umem_for_each_in_range(&context->umem_tree, start,
230 invalidate_range_end_trampoline, NULL);
231 up_read(&context->umem_rwsem);
232 ib_ucontext_notifier_end_account(context);
235 static const struct mmu_notifier_ops ib_umem_notifiers = {
236 .release = ib_umem_notifier_release,
237 .invalidate_page = ib_umem_notifier_invalidate_page,
238 .invalidate_range_start = ib_umem_notifier_invalidate_range_start,
239 .invalidate_range_end = ib_umem_notifier_invalidate_range_end,
242 struct ib_umem *ib_alloc_odp_umem(struct ib_ucontext *context,
246 struct ib_umem *umem;
247 struct ib_umem_odp *odp_data;
248 int pages = size >> PAGE_SHIFT;
251 umem = kzalloc(sizeof(*umem), GFP_KERNEL);
253 return ERR_PTR(-ENOMEM);
255 umem->context = context;
257 umem->address = addr;
258 umem->page_size = PAGE_SIZE;
261 odp_data = kzalloc(sizeof(*odp_data), GFP_KERNEL);
266 odp_data->umem = umem;
268 mutex_init(&odp_data->umem_mutex);
269 init_completion(&odp_data->notifier_completion);
271 odp_data->page_list = vzalloc(pages * sizeof(*odp_data->page_list));
272 if (!odp_data->page_list) {
277 odp_data->dma_list = vzalloc(pages * sizeof(*odp_data->dma_list));
278 if (!odp_data->dma_list) {
283 down_write(&context->umem_rwsem);
284 context->odp_mrs_count++;
285 rbt_ib_umem_insert(&odp_data->interval_tree, &context->umem_tree);
286 if (likely(!atomic_read(&context->notifier_count)))
287 odp_data->mn_counters_active = true;
289 list_add(&odp_data->no_private_counters,
290 &context->no_private_counters);
291 up_write(&context->umem_rwsem);
293 umem->odp_data = odp_data;
298 vfree(odp_data->page_list);
305 EXPORT_SYMBOL(ib_alloc_odp_umem);
307 int ib_umem_odp_get(struct ib_ucontext *context, struct ib_umem *umem)
311 struct mm_struct *mm = get_task_mm(current);
316 /* Prevent creating ODP MRs in child processes */
318 our_pid = get_task_pid(current->group_leader, PIDTYPE_PID);
321 if (context->tgid != our_pid) {
327 umem->odp_data = kzalloc(sizeof(*umem->odp_data), GFP_KERNEL);
328 if (!umem->odp_data) {
332 umem->odp_data->umem = umem;
334 mutex_init(&umem->odp_data->umem_mutex);
336 init_completion(&umem->odp_data->notifier_completion);
338 if (ib_umem_num_pages(umem)) {
339 umem->odp_data->page_list = vzalloc(ib_umem_num_pages(umem) *
340 sizeof(*umem->odp_data->page_list));
341 if (!umem->odp_data->page_list) {
346 umem->odp_data->dma_list = vzalloc(ib_umem_num_pages(umem) *
347 sizeof(*umem->odp_data->dma_list));
348 if (!umem->odp_data->dma_list) {
355 * When using MMU notifiers, we will get a
356 * notification before the "current" task (and MM) is
357 * destroyed. We use the umem_rwsem semaphore to synchronize.
359 down_write(&context->umem_rwsem);
360 context->odp_mrs_count++;
361 if (likely(ib_umem_start(umem) != ib_umem_end(umem)))
362 rbt_ib_umem_insert(&umem->odp_data->interval_tree,
363 &context->umem_tree);
364 if (likely(!atomic_read(&context->notifier_count)) ||
365 context->odp_mrs_count == 1)
366 umem->odp_data->mn_counters_active = true;
368 list_add(&umem->odp_data->no_private_counters,
369 &context->no_private_counters);
370 downgrade_write(&context->umem_rwsem);
372 if (context->odp_mrs_count == 1) {
374 * Note that at this point, no MMU notifier is running
377 atomic_set(&context->notifier_count, 0);
378 INIT_HLIST_NODE(&context->mn.hlist);
379 context->mn.ops = &ib_umem_notifiers;
381 * Lock-dep detects a false positive for mmap_sem vs.
382 * umem_rwsem, due to not grasping downgrade_write correctly.
385 ret_val = mmu_notifier_register(&context->mn, mm);
388 pr_err("Failed to register mmu_notifier %d\n", ret_val);
394 up_read(&context->umem_rwsem);
397 * Note that doing an mmput can cause a notifier for the relevant mm.
398 * If the notifier is called while we hold the umem_rwsem, this will
399 * cause a deadlock. Therefore, we release the reference only after we
400 * released the semaphore.
406 up_read(&context->umem_rwsem);
407 vfree(umem->odp_data->dma_list);
409 vfree(umem->odp_data->page_list);
411 kfree(umem->odp_data);
417 void ib_umem_odp_release(struct ib_umem *umem)
419 struct ib_ucontext *context = umem->context;
422 * Ensure that no more pages are mapped in the umem.
424 * It is the driver's responsibility to ensure, before calling us,
425 * that the hardware will not attempt to access the MR any more.
427 ib_umem_odp_unmap_dma_pages(umem, ib_umem_start(umem),
430 down_write(&context->umem_rwsem);
431 if (likely(ib_umem_start(umem) != ib_umem_end(umem)))
432 rbt_ib_umem_remove(&umem->odp_data->interval_tree,
433 &context->umem_tree);
434 context->odp_mrs_count--;
435 if (!umem->odp_data->mn_counters_active) {
436 list_del(&umem->odp_data->no_private_counters);
437 complete_all(&umem->odp_data->notifier_completion);
441 * Downgrade the lock to a read lock. This ensures that the notifiers
442 * (who lock the mutex for reading) will be able to finish, and we
443 * will be able to enventually obtain the mmu notifiers SRCU. Note
444 * that since we are doing it atomically, no other user could register
445 * and unregister while we do the check.
447 downgrade_write(&context->umem_rwsem);
448 if (!context->odp_mrs_count) {
449 struct task_struct *owning_process = NULL;
450 struct mm_struct *owning_mm = NULL;
452 owning_process = get_pid_task(context->tgid,
454 if (owning_process == NULL)
456 * The process is already dead, notifier were removed
461 owning_mm = get_task_mm(owning_process);
462 if (owning_mm == NULL)
464 * The process' mm is already dead, notifier were
468 mmu_notifier_unregister(&context->mn, owning_mm);
473 put_task_struct(owning_process);
476 up_read(&context->umem_rwsem);
478 vfree(umem->odp_data->dma_list);
479 vfree(umem->odp_data->page_list);
480 kfree(umem->odp_data);
485 * Map for DMA and insert a single page into the on-demand paging page tables.
487 * @umem: the umem to insert the page to.
488 * @page_index: index in the umem to add the page to.
489 * @page: the page struct to map and add.
490 * @access_mask: access permissions needed for this page.
491 * @current_seq: sequence number for synchronization with invalidations.
492 * the sequence number is taken from
493 * umem->odp_data->notifiers_seq.
495 * The function returns -EFAULT if the DMA mapping operation fails. It returns
496 * -EAGAIN if a concurrent invalidation prevents us from updating the page.
498 * The page is released via put_page even if the operation failed. For
499 * on-demand pinning, the page is released whenever it isn't stored in the
502 static int ib_umem_odp_map_dma_single_page(
503 struct ib_umem *umem,
508 unsigned long current_seq)
510 struct ib_device *dev = umem->context->device;
513 int remove_existing_mapping = 0;
517 * Note: we avoid writing if seq is different from the initial seq, to
518 * handle case of a racing notifier. This check also allows us to bail
519 * early if we have a notifier running in parallel with us.
521 if (ib_umem_mmu_notifier_retry(umem, current_seq)) {
525 if (!(umem->odp_data->dma_list[page_index])) {
526 dma_addr = ib_dma_map_page(dev,
530 if (ib_dma_mapping_error(dev, dma_addr)) {
534 umem->odp_data->dma_list[page_index] = dma_addr | access_mask;
535 umem->odp_data->page_list[page_index] = page;
538 } else if (umem->odp_data->page_list[page_index] == page) {
539 umem->odp_data->dma_list[page_index] |= access_mask;
541 pr_err("error: got different pages in IB device and from get_user_pages. IB device page: %p, gup page: %p\n",
542 umem->odp_data->page_list[page_index], page);
543 /* Better remove the mapping now, to prevent any further
545 remove_existing_mapping = 1;
549 /* On Demand Paging - avoid pinning the page */
550 if (umem->context->invalidate_range || !stored_page)
553 if (remove_existing_mapping && umem->context->invalidate_range) {
554 invalidate_page_trampoline(
556 base_virt_addr + (page_index * PAGE_SIZE),
557 base_virt_addr + ((page_index+1)*PAGE_SIZE),
566 * ib_umem_odp_map_dma_pages - Pin and DMA map userspace memory in an ODP MR.
568 * Pins the range of pages passed in the argument, and maps them to
569 * DMA addresses. The DMA addresses of the mapped pages is updated in
570 * umem->odp_data->dma_list.
572 * Returns the number of pages mapped in success, negative error code
574 * An -EAGAIN error code is returned when a concurrent mmu notifier prevents
575 * the function from completing its task.
577 * @umem: the umem to map and pin
578 * @user_virt: the address from which we need to map.
579 * @bcnt: the minimal number of bytes to pin and map. The mapping might be
580 * bigger due to alignment, and may also be smaller in case of an error
581 * pinning or mapping a page. The actual pages mapped is returned in
583 * @access_mask: bit mask of the requested access permissions for the given
585 * @current_seq: the MMU notifiers sequance value for synchronization with
586 * invalidations. the sequance number is read from
587 * umem->odp_data->notifiers_seq before calling this function
589 int ib_umem_odp_map_dma_pages(struct ib_umem *umem, u64 user_virt, u64 bcnt,
590 u64 access_mask, unsigned long current_seq)
592 struct task_struct *owning_process = NULL;
593 struct mm_struct *owning_mm = NULL;
594 struct page **local_page_list = NULL;
596 int j, k, ret = 0, start_idx, npages = 0;
598 unsigned int flags = 0;
600 if (access_mask == 0)
603 if (user_virt < ib_umem_start(umem) ||
604 user_virt + bcnt > ib_umem_end(umem))
607 local_page_list = (struct page **)__get_free_page(GFP_KERNEL);
608 if (!local_page_list)
611 off = user_virt & (~PAGE_MASK);
612 user_virt = user_virt & PAGE_MASK;
613 base_virt_addr = user_virt;
614 bcnt += off; /* Charge for the first page offset as well. */
616 owning_process = get_pid_task(umem->context->tgid, PIDTYPE_PID);
617 if (owning_process == NULL) {
622 owning_mm = get_task_mm(owning_process);
623 if (owning_mm == NULL) {
628 if (access_mask & ODP_WRITE_ALLOWED_BIT)
631 start_idx = (user_virt - ib_umem_start(umem)) >> PAGE_SHIFT;
635 const size_t gup_num_pages =
636 min_t(size_t, ALIGN(bcnt, PAGE_SIZE) / PAGE_SIZE,
637 PAGE_SIZE / sizeof(struct page *));
639 down_read(&owning_mm->mmap_sem);
641 * Note: this might result in redundent page getting. We can
642 * avoid this by checking dma_list to be 0 before calling
643 * get_user_pages. However, this make the code much more
644 * complex (and doesn't gain us much performance in most use
647 npages = get_user_pages_remote(owning_process, owning_mm,
648 user_virt, gup_num_pages,
649 flags, local_page_list, NULL, NULL);
650 up_read(&owning_mm->mmap_sem);
655 bcnt -= min_t(size_t, npages << PAGE_SHIFT, bcnt);
656 user_virt += npages << PAGE_SHIFT;
657 mutex_lock(&umem->odp_data->umem_mutex);
658 for (j = 0; j < npages; ++j) {
659 ret = ib_umem_odp_map_dma_single_page(
660 umem, k, base_virt_addr, local_page_list[j],
661 access_mask, current_seq);
666 mutex_unlock(&umem->odp_data->umem_mutex);
669 /* Release left over pages when handling errors. */
670 for (++j; j < npages; ++j)
671 put_page(local_page_list[j]);
677 if (npages < 0 && k == start_idx)
685 put_task_struct(owning_process);
687 free_page((unsigned long)local_page_list);
690 EXPORT_SYMBOL(ib_umem_odp_map_dma_pages);
692 void ib_umem_odp_unmap_dma_pages(struct ib_umem *umem, u64 virt,
697 struct ib_device *dev = umem->context->device;
699 virt = max_t(u64, virt, ib_umem_start(umem));
700 bound = min_t(u64, bound, ib_umem_end(umem));
701 /* Note that during the run of this function, the
702 * notifiers_count of the MR is > 0, preventing any racing
703 * faults from completion. We might be racing with other
704 * invalidations, so we must make sure we free each page only
706 mutex_lock(&umem->odp_data->umem_mutex);
707 for (addr = virt; addr < bound; addr += (u64)umem->page_size) {
708 idx = (addr - ib_umem_start(umem)) / PAGE_SIZE;
709 if (umem->odp_data->page_list[idx]) {
710 struct page *page = umem->odp_data->page_list[idx];
711 dma_addr_t dma = umem->odp_data->dma_list[idx];
712 dma_addr_t dma_addr = dma & ODP_DMA_ADDR_MASK;
716 ib_dma_unmap_page(dev, dma_addr, PAGE_SIZE,
718 if (dma & ODP_WRITE_ALLOWED_BIT) {
719 struct page *head_page = compound_head(page);
721 * set_page_dirty prefers being called with
722 * the page lock. However, MMU notifiers are
723 * called sometimes with and sometimes without
724 * the lock. We rely on the umem_mutex instead
725 * to prevent other mmu notifiers from
726 * continuing and allowing the page mapping to
729 set_page_dirty(head_page);
731 /* on demand pinning support */
732 if (!umem->context->invalidate_range)
734 umem->odp_data->page_list[idx] = NULL;
735 umem->odp_data->dma_list[idx] = 0;
739 mutex_unlock(&umem->odp_data->umem_mutex);
741 EXPORT_SYMBOL(ib_umem_odp_unmap_dma_pages);