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/sched/mm.h>
36 #include <linux/sched/task.h>
37 #include <linux/pid.h>
38 #include <linux/slab.h>
39 #include <linux/export.h>
40 #include <linux/vmalloc.h>
41 #include <linux/hugetlb.h>
42 #include <linux/interval_tree_generic.h>
43 #include <linux/pagemap.h>
45 #include <rdma/ib_verbs.h>
46 #include <rdma/ib_umem.h>
47 #include <rdma/ib_umem_odp.h>
50 * The ib_umem list keeps track of memory regions for which the HW
51 * device request to receive notification when the related memory
54 * ib_umem_lock protects the list.
57 static u64 node_start(struct umem_odp_node *n)
59 struct ib_umem_odp *umem_odp =
60 container_of(n, struct ib_umem_odp, interval_tree);
62 return ib_umem_start(&umem_odp->umem);
65 /* Note that the representation of the intervals in the interval tree
66 * considers the ending point as contained in the interval, while the
67 * function ib_umem_end returns the first address which is not contained
70 static u64 node_last(struct umem_odp_node *n)
72 struct ib_umem_odp *umem_odp =
73 container_of(n, struct ib_umem_odp, interval_tree);
75 return ib_umem_end(&umem_odp->umem) - 1;
78 INTERVAL_TREE_DEFINE(struct umem_odp_node, rb, u64, __subtree_last,
79 node_start, node_last, static, rbt_ib_umem)
81 static void ib_umem_notifier_start_account(struct ib_umem_odp *umem_odp)
83 mutex_lock(&umem_odp->umem_mutex);
84 if (umem_odp->notifiers_count++ == 0)
86 * Initialize the completion object for waiting on
87 * notifiers. Since notifier_count is zero, no one should be
90 reinit_completion(&umem_odp->notifier_completion);
91 mutex_unlock(&umem_odp->umem_mutex);
94 static void ib_umem_notifier_end_account(struct ib_umem_odp *umem_odp)
96 mutex_lock(&umem_odp->umem_mutex);
98 * This sequence increase will notify the QP page fault that the page
99 * that is going to be mapped in the spte could have been freed.
101 ++umem_odp->notifiers_seq;
102 if (--umem_odp->notifiers_count == 0)
103 complete_all(&umem_odp->notifier_completion);
104 mutex_unlock(&umem_odp->umem_mutex);
107 static int ib_umem_notifier_release_trampoline(struct ib_umem_odp *umem_odp,
108 u64 start, u64 end, void *cookie)
110 struct ib_umem *umem = &umem_odp->umem;
113 * Increase the number of notifiers running, to
114 * prevent any further fault handling on this MR.
116 ib_umem_notifier_start_account(umem_odp);
118 /* Make sure that the fact the umem is dying is out before we release
119 * all pending page faults. */
121 complete_all(&umem_odp->notifier_completion);
122 umem->context->invalidate_range(umem_odp, ib_umem_start(umem),
127 static void ib_umem_notifier_release(struct mmu_notifier *mn,
128 struct mm_struct *mm)
130 struct ib_ucontext_per_mm *per_mm =
131 container_of(mn, struct ib_ucontext_per_mm, mn);
133 down_read(&per_mm->umem_rwsem);
135 rbt_ib_umem_for_each_in_range(
136 &per_mm->umem_tree, 0, ULLONG_MAX,
137 ib_umem_notifier_release_trampoline, true, NULL);
138 up_read(&per_mm->umem_rwsem);
141 static int invalidate_range_start_trampoline(struct ib_umem_odp *item,
142 u64 start, u64 end, void *cookie)
144 ib_umem_notifier_start_account(item);
145 item->umem.context->invalidate_range(item, start, end);
149 static int ib_umem_notifier_invalidate_range_start(struct mmu_notifier *mn,
150 const struct mmu_notifier_range *range)
152 struct ib_ucontext_per_mm *per_mm =
153 container_of(mn, struct ib_ucontext_per_mm, mn);
155 if (mmu_notifier_range_blockable(range))
156 down_read(&per_mm->umem_rwsem);
157 else if (!down_read_trylock(&per_mm->umem_rwsem))
160 if (!per_mm->active) {
161 up_read(&per_mm->umem_rwsem);
163 * At this point active is permanently set and visible to this
164 * CPU without a lock, that fact is relied on to skip the unlock
170 return rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, range->start,
172 invalidate_range_start_trampoline,
173 mmu_notifier_range_blockable(range),
177 static int invalidate_range_end_trampoline(struct ib_umem_odp *item, u64 start,
178 u64 end, void *cookie)
180 ib_umem_notifier_end_account(item);
184 static void ib_umem_notifier_invalidate_range_end(struct mmu_notifier *mn,
185 const struct mmu_notifier_range *range)
187 struct ib_ucontext_per_mm *per_mm =
188 container_of(mn, struct ib_ucontext_per_mm, mn);
190 if (unlikely(!per_mm->active))
193 rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, range->start,
195 invalidate_range_end_trampoline, true, NULL);
196 up_read(&per_mm->umem_rwsem);
199 static const struct mmu_notifier_ops ib_umem_notifiers = {
200 .release = ib_umem_notifier_release,
201 .invalidate_range_start = ib_umem_notifier_invalidate_range_start,
202 .invalidate_range_end = ib_umem_notifier_invalidate_range_end,
205 static void add_umem_to_per_mm(struct ib_umem_odp *umem_odp)
207 struct ib_ucontext_per_mm *per_mm = umem_odp->per_mm;
208 struct ib_umem *umem = &umem_odp->umem;
210 down_write(&per_mm->umem_rwsem);
211 if (likely(ib_umem_start(umem) != ib_umem_end(umem)))
212 rbt_ib_umem_insert(&umem_odp->interval_tree,
214 up_write(&per_mm->umem_rwsem);
217 static void remove_umem_from_per_mm(struct ib_umem_odp *umem_odp)
219 struct ib_ucontext_per_mm *per_mm = umem_odp->per_mm;
220 struct ib_umem *umem = &umem_odp->umem;
222 down_write(&per_mm->umem_rwsem);
223 if (likely(ib_umem_start(umem) != ib_umem_end(umem)))
224 rbt_ib_umem_remove(&umem_odp->interval_tree,
226 complete_all(&umem_odp->notifier_completion);
228 up_write(&per_mm->umem_rwsem);
231 static struct ib_ucontext_per_mm *alloc_per_mm(struct ib_ucontext *ctx,
232 struct mm_struct *mm)
234 struct ib_ucontext_per_mm *per_mm;
237 per_mm = kzalloc(sizeof(*per_mm), GFP_KERNEL);
239 return ERR_PTR(-ENOMEM);
241 per_mm->context = ctx;
243 per_mm->umem_tree = RB_ROOT_CACHED;
244 init_rwsem(&per_mm->umem_rwsem);
245 per_mm->active = true;
248 per_mm->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
251 WARN_ON(mm != current->mm);
253 per_mm->mn.ops = &ib_umem_notifiers;
254 ret = mmu_notifier_register(&per_mm->mn, per_mm->mm);
256 dev_err(&ctx->device->dev,
257 "Failed to register mmu_notifier %d\n", ret);
261 list_add(&per_mm->ucontext_list, &ctx->per_mm_list);
265 put_pid(per_mm->tgid);
270 static int get_per_mm(struct ib_umem_odp *umem_odp)
272 struct ib_ucontext *ctx = umem_odp->umem.context;
273 struct ib_ucontext_per_mm *per_mm;
276 * Generally speaking we expect only one or two per_mm in this list,
277 * so no reason to optimize this search today.
279 mutex_lock(&ctx->per_mm_list_lock);
280 list_for_each_entry(per_mm, &ctx->per_mm_list, ucontext_list) {
281 if (per_mm->mm == umem_odp->umem.owning_mm)
285 per_mm = alloc_per_mm(ctx, umem_odp->umem.owning_mm);
286 if (IS_ERR(per_mm)) {
287 mutex_unlock(&ctx->per_mm_list_lock);
288 return PTR_ERR(per_mm);
292 umem_odp->per_mm = per_mm;
293 per_mm->odp_mrs_count++;
294 mutex_unlock(&ctx->per_mm_list_lock);
299 static void free_per_mm(struct rcu_head *rcu)
301 kfree(container_of(rcu, struct ib_ucontext_per_mm, rcu));
304 static void put_per_mm(struct ib_umem_odp *umem_odp)
306 struct ib_ucontext_per_mm *per_mm = umem_odp->per_mm;
307 struct ib_ucontext *ctx = umem_odp->umem.context;
310 mutex_lock(&ctx->per_mm_list_lock);
311 umem_odp->per_mm = NULL;
312 per_mm->odp_mrs_count--;
313 need_free = per_mm->odp_mrs_count == 0;
315 list_del(&per_mm->ucontext_list);
316 mutex_unlock(&ctx->per_mm_list_lock);
322 * NOTE! mmu_notifier_unregister() can happen between a start/end
323 * callback, resulting in an start/end, and thus an unbalanced
324 * lock. This doesn't really matter to us since we are about to kfree
325 * the memory that holds the lock, however LOCKDEP doesn't like this.
327 down_write(&per_mm->umem_rwsem);
328 per_mm->active = false;
329 up_write(&per_mm->umem_rwsem);
331 WARN_ON(!RB_EMPTY_ROOT(&per_mm->umem_tree.rb_root));
332 mmu_notifier_unregister_no_release(&per_mm->mn, per_mm->mm);
333 put_pid(per_mm->tgid);
334 mmu_notifier_call_srcu(&per_mm->rcu, free_per_mm);
337 struct ib_umem_odp *ib_alloc_odp_umem(struct ib_umem_odp *root,
338 unsigned long addr, size_t size)
340 struct ib_ucontext_per_mm *per_mm = root->per_mm;
341 struct ib_ucontext *ctx = per_mm->context;
342 struct ib_umem_odp *odp_data;
343 struct ib_umem *umem;
344 int pages = size >> PAGE_SHIFT;
347 odp_data = kzalloc(sizeof(*odp_data), GFP_KERNEL);
349 return ERR_PTR(-ENOMEM);
350 umem = &odp_data->umem;
353 umem->address = addr;
354 umem->page_shift = PAGE_SHIFT;
355 umem->writable = root->umem.writable;
357 odp_data->per_mm = per_mm;
358 umem->owning_mm = per_mm->mm;
359 mmgrab(umem->owning_mm);
361 mutex_init(&odp_data->umem_mutex);
362 init_completion(&odp_data->notifier_completion);
364 odp_data->page_list =
365 vzalloc(array_size(pages, sizeof(*odp_data->page_list)));
366 if (!odp_data->page_list) {
372 vzalloc(array_size(pages, sizeof(*odp_data->dma_list)));
373 if (!odp_data->dma_list) {
379 * Caller must ensure that the umem_odp that the per_mm came from
380 * cannot be freed during the call to ib_alloc_odp_umem.
382 mutex_lock(&ctx->per_mm_list_lock);
383 per_mm->odp_mrs_count++;
384 mutex_unlock(&ctx->per_mm_list_lock);
385 add_umem_to_per_mm(odp_data);
390 vfree(odp_data->page_list);
392 mmdrop(umem->owning_mm);
396 EXPORT_SYMBOL(ib_alloc_odp_umem);
398 int ib_umem_odp_get(struct ib_umem_odp *umem_odp, int access)
400 struct ib_umem *umem = &umem_odp->umem;
402 * NOTE: This must called in a process context where umem->owning_mm
405 struct mm_struct *mm = umem->owning_mm;
408 if (access & IB_ACCESS_HUGETLB) {
409 struct vm_area_struct *vma;
412 down_read(&mm->mmap_sem);
413 vma = find_vma(mm, ib_umem_start(umem));
414 if (!vma || !is_vm_hugetlb_page(vma)) {
415 up_read(&mm->mmap_sem);
419 umem->page_shift = huge_page_shift(h);
420 up_read(&mm->mmap_sem);
423 mutex_init(&umem_odp->umem_mutex);
425 init_completion(&umem_odp->notifier_completion);
427 if (ib_umem_num_pages(umem)) {
428 umem_odp->page_list =
429 vzalloc(array_size(sizeof(*umem_odp->page_list),
430 ib_umem_num_pages(umem)));
431 if (!umem_odp->page_list)
435 vzalloc(array_size(sizeof(*umem_odp->dma_list),
436 ib_umem_num_pages(umem)));
437 if (!umem_odp->dma_list) {
443 ret_val = get_per_mm(umem_odp);
446 add_umem_to_per_mm(umem_odp);
451 vfree(umem_odp->dma_list);
453 vfree(umem_odp->page_list);
457 void ib_umem_odp_release(struct ib_umem_odp *umem_odp)
459 struct ib_umem *umem = &umem_odp->umem;
462 * Ensure that no more pages are mapped in the umem.
464 * It is the driver's responsibility to ensure, before calling us,
465 * that the hardware will not attempt to access the MR any more.
467 ib_umem_odp_unmap_dma_pages(umem_odp, ib_umem_start(umem),
470 remove_umem_from_per_mm(umem_odp);
471 put_per_mm(umem_odp);
472 vfree(umem_odp->dma_list);
473 vfree(umem_odp->page_list);
477 * Map for DMA and insert a single page into the on-demand paging page tables.
479 * @umem: the umem to insert the page to.
480 * @page_index: index in the umem to add the page to.
481 * @page: the page struct to map and add.
482 * @access_mask: access permissions needed for this page.
483 * @current_seq: sequence number for synchronization with invalidations.
484 * the sequence number is taken from
485 * umem_odp->notifiers_seq.
487 * The function returns -EFAULT if the DMA mapping operation fails. It returns
488 * -EAGAIN if a concurrent invalidation prevents us from updating the page.
490 * The page is released via put_page even if the operation failed. For
491 * on-demand pinning, the page is released whenever it isn't stored in the
494 static int ib_umem_odp_map_dma_single_page(
495 struct ib_umem_odp *umem_odp,
499 unsigned long current_seq)
501 struct ib_umem *umem = &umem_odp->umem;
502 struct ib_device *dev = umem->context->device;
504 int remove_existing_mapping = 0;
508 * Note: we avoid writing if seq is different from the initial seq, to
509 * handle case of a racing notifier. This check also allows us to bail
510 * early if we have a notifier running in parallel with us.
512 if (ib_umem_mmu_notifier_retry(umem_odp, current_seq)) {
516 if (!(umem_odp->dma_list[page_index])) {
517 dma_addr = ib_dma_map_page(dev,
519 0, BIT(umem->page_shift),
521 if (ib_dma_mapping_error(dev, dma_addr)) {
525 umem_odp->dma_list[page_index] = dma_addr | access_mask;
526 umem_odp->page_list[page_index] = page;
528 } else if (umem_odp->page_list[page_index] == page) {
529 umem_odp->dma_list[page_index] |= access_mask;
531 pr_err("error: got different pages in IB device and from get_user_pages. IB device page: %p, gup page: %p\n",
532 umem_odp->page_list[page_index], page);
533 /* Better remove the mapping now, to prevent any further
535 remove_existing_mapping = 1;
541 if (remove_existing_mapping) {
542 ib_umem_notifier_start_account(umem_odp);
543 umem->context->invalidate_range(
545 ib_umem_start(umem) + (page_index << umem->page_shift),
546 ib_umem_start(umem) +
547 ((page_index + 1) << umem->page_shift));
548 ib_umem_notifier_end_account(umem_odp);
556 * ib_umem_odp_map_dma_pages - Pin and DMA map userspace memory in an ODP MR.
558 * Pins the range of pages passed in the argument, and maps them to
559 * DMA addresses. The DMA addresses of the mapped pages is updated in
560 * umem_odp->dma_list.
562 * Returns the number of pages mapped in success, negative error code
564 * An -EAGAIN error code is returned when a concurrent mmu notifier prevents
565 * the function from completing its task.
566 * An -ENOENT error code indicates that userspace process is being terminated
567 * and mm was already destroyed.
568 * @umem_odp: the umem to map and pin
569 * @user_virt: the address from which we need to map.
570 * @bcnt: the minimal number of bytes to pin and map. The mapping might be
571 * bigger due to alignment, and may also be smaller in case of an error
572 * pinning or mapping a page. The actual pages mapped is returned in
574 * @access_mask: bit mask of the requested access permissions for the given
576 * @current_seq: the MMU notifiers sequance value for synchronization with
577 * invalidations. the sequance number is read from
578 * umem_odp->notifiers_seq before calling this function
580 int ib_umem_odp_map_dma_pages(struct ib_umem_odp *umem_odp, u64 user_virt,
581 u64 bcnt, u64 access_mask,
582 unsigned long current_seq)
584 struct ib_umem *umem = &umem_odp->umem;
585 struct task_struct *owning_process = NULL;
586 struct mm_struct *owning_mm = umem_odp->umem.owning_mm;
587 struct page **local_page_list = NULL;
589 int j, k, ret = 0, start_idx, npages = 0, page_shift;
590 unsigned int flags = 0;
593 if (access_mask == 0)
596 if (user_virt < ib_umem_start(umem) ||
597 user_virt + bcnt > ib_umem_end(umem))
600 local_page_list = (struct page **)__get_free_page(GFP_KERNEL);
601 if (!local_page_list)
604 page_shift = umem->page_shift;
605 page_mask = ~(BIT(page_shift) - 1);
606 off = user_virt & (~page_mask);
607 user_virt = user_virt & page_mask;
608 bcnt += off; /* Charge for the first page offset as well. */
611 * owning_process is allowed to be NULL, this means somehow the mm is
612 * existing beyond the lifetime of the originating process.. Presumably
613 * mmget_not_zero will fail in this case.
615 owning_process = get_pid_task(umem_odp->per_mm->tgid, PIDTYPE_PID);
616 if (!owning_process || !mmget_not_zero(owning_mm)) {
621 if (access_mask & ODP_WRITE_ALLOWED_BIT)
624 start_idx = (user_virt - ib_umem_start(umem)) >> page_shift;
628 const size_t gup_num_pages = min_t(size_t,
629 (bcnt + BIT(page_shift) - 1) >> page_shift,
630 PAGE_SIZE / sizeof(struct page *));
632 down_read(&owning_mm->mmap_sem);
634 * Note: this might result in redundent page getting. We can
635 * avoid this by checking dma_list to be 0 before calling
636 * get_user_pages. However, this make the code much more
637 * complex (and doesn't gain us much performance in most use
640 npages = get_user_pages_remote(owning_process, owning_mm,
641 user_virt, gup_num_pages,
642 flags, local_page_list, NULL, NULL);
643 up_read(&owning_mm->mmap_sem);
646 if (npages != -EAGAIN)
647 pr_warn("fail to get %zu user pages with error %d\n", gup_num_pages, npages);
649 pr_debug("fail to get %zu user pages with error %d\n", gup_num_pages, npages);
653 bcnt -= min_t(size_t, npages << PAGE_SHIFT, bcnt);
654 mutex_lock(&umem_odp->umem_mutex);
655 for (j = 0; j < npages; j++, user_virt += PAGE_SIZE) {
656 if (user_virt & ~page_mask) {
658 if (page_to_phys(local_page_list[j]) != p) {
662 put_page(local_page_list[j]);
666 ret = ib_umem_odp_map_dma_single_page(
667 umem_odp, k, local_page_list[j],
668 access_mask, current_seq);
671 pr_warn("ib_umem_odp_map_dma_single_page failed with error %d\n", ret);
673 pr_debug("ib_umem_odp_map_dma_single_page failed with error %d\n", ret);
677 p = page_to_phys(local_page_list[j]);
680 mutex_unlock(&umem_odp->umem_mutex);
684 * Release pages, remembering that the first page
685 * to hit an error was already released by
686 * ib_umem_odp_map_dma_single_page().
688 if (npages - (j + 1) > 0)
689 release_pages(&local_page_list[j+1],
696 if (npages < 0 && k == start_idx)
705 put_task_struct(owning_process);
706 free_page((unsigned long)local_page_list);
709 EXPORT_SYMBOL(ib_umem_odp_map_dma_pages);
711 void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, u64 virt,
714 struct ib_umem *umem = &umem_odp->umem;
717 struct ib_device *dev = umem->context->device;
719 virt = max_t(u64, virt, ib_umem_start(umem));
720 bound = min_t(u64, bound, ib_umem_end(umem));
721 /* Note that during the run of this function, the
722 * notifiers_count of the MR is > 0, preventing any racing
723 * faults from completion. We might be racing with other
724 * invalidations, so we must make sure we free each page only
726 mutex_lock(&umem_odp->umem_mutex);
727 for (addr = virt; addr < bound; addr += BIT(umem->page_shift)) {
728 idx = (addr - ib_umem_start(umem)) >> umem->page_shift;
729 if (umem_odp->page_list[idx]) {
730 struct page *page = umem_odp->page_list[idx];
731 dma_addr_t dma = umem_odp->dma_list[idx];
732 dma_addr_t dma_addr = dma & ODP_DMA_ADDR_MASK;
736 ib_dma_unmap_page(dev, dma_addr, PAGE_SIZE,
738 if (dma & ODP_WRITE_ALLOWED_BIT) {
739 struct page *head_page = compound_head(page);
741 * set_page_dirty prefers being called with
742 * the page lock. However, MMU notifiers are
743 * called sometimes with and sometimes without
744 * the lock. We rely on the umem_mutex instead
745 * to prevent other mmu notifiers from
746 * continuing and allowing the page mapping to
749 set_page_dirty(head_page);
751 umem_odp->page_list[idx] = NULL;
752 umem_odp->dma_list[idx] = 0;
756 mutex_unlock(&umem_odp->umem_mutex);
758 EXPORT_SYMBOL(ib_umem_odp_unmap_dma_pages);
760 /* @last is not a part of the interval. See comment for function
763 int rbt_ib_umem_for_each_in_range(struct rb_root_cached *root,
770 struct umem_odp_node *node, *next;
771 struct ib_umem_odp *umem;
773 if (unlikely(start == last))
776 for (node = rbt_ib_umem_iter_first(root, start, last - 1);
778 /* TODO move the blockable decision up to the callback */
781 next = rbt_ib_umem_iter_next(node, start, last - 1);
782 umem = container_of(node, struct ib_umem_odp, interval_tree);
783 ret_val = cb(umem, start, last, cookie) || ret_val;
788 EXPORT_SYMBOL(rbt_ib_umem_for_each_in_range);
790 struct ib_umem_odp *rbt_ib_umem_lookup(struct rb_root_cached *root,
791 u64 addr, u64 length)
793 struct umem_odp_node *node;
795 node = rbt_ib_umem_iter_first(root, addr, addr + length - 1);
797 return container_of(node, struct ib_umem_odp, interval_tree);
801 EXPORT_SYMBOL(rbt_ib_umem_lookup);