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>
44 #include <rdma/ib_verbs.h>
45 #include <rdma/ib_umem.h>
46 #include <rdma/ib_umem_odp.h>
49 * The ib_umem list keeps track of memory regions for which the HW
50 * device request to receive notification when the related memory
53 * ib_umem_lock protects the list.
56 static u64 node_start(struct umem_odp_node *n)
58 struct ib_umem_odp *umem_odp =
59 container_of(n, struct ib_umem_odp, interval_tree);
61 return ib_umem_start(&umem_odp->umem);
64 /* Note that the representation of the intervals in the interval tree
65 * considers the ending point as contained in the interval, while the
66 * function ib_umem_end returns the first address which is not contained
69 static u64 node_last(struct umem_odp_node *n)
71 struct ib_umem_odp *umem_odp =
72 container_of(n, struct ib_umem_odp, interval_tree);
74 return ib_umem_end(&umem_odp->umem) - 1;
77 INTERVAL_TREE_DEFINE(struct umem_odp_node, rb, u64, __subtree_last,
78 node_start, node_last, static, rbt_ib_umem)
80 static void ib_umem_notifier_start_account(struct ib_umem_odp *umem_odp)
82 mutex_lock(&umem_odp->umem_mutex);
83 if (umem_odp->notifiers_count++ == 0)
85 * Initialize the completion object for waiting on
86 * notifiers. Since notifier_count is zero, no one should be
89 reinit_completion(&umem_odp->notifier_completion);
90 mutex_unlock(&umem_odp->umem_mutex);
93 static void ib_umem_notifier_end_account(struct ib_umem_odp *umem_odp)
95 mutex_lock(&umem_odp->umem_mutex);
97 * This sequence increase will notify the QP page fault that the page
98 * that is going to be mapped in the spte could have been freed.
100 ++umem_odp->notifiers_seq;
101 if (--umem_odp->notifiers_count == 0)
102 complete_all(&umem_odp->notifier_completion);
103 mutex_unlock(&umem_odp->umem_mutex);
106 static int ib_umem_notifier_release_trampoline(struct ib_umem_odp *umem_odp,
107 u64 start, u64 end, void *cookie)
109 struct ib_umem *umem = &umem_odp->umem;
112 * Increase the number of notifiers running, to
113 * prevent any further fault handling on this MR.
115 ib_umem_notifier_start_account(umem_odp);
117 /* Make sure that the fact the umem is dying is out before we release
118 * all pending page faults. */
120 complete_all(&umem_odp->notifier_completion);
121 umem->context->invalidate_range(umem_odp, ib_umem_start(umem),
126 static void ib_umem_notifier_release(struct mmu_notifier *mn,
127 struct mm_struct *mm)
129 struct ib_ucontext_per_mm *per_mm =
130 container_of(mn, struct ib_ucontext_per_mm, mn);
132 down_read(&per_mm->umem_rwsem);
134 rbt_ib_umem_for_each_in_range(
135 &per_mm->umem_tree, 0, ULLONG_MAX,
136 ib_umem_notifier_release_trampoline, true, NULL);
137 up_read(&per_mm->umem_rwsem);
140 static int invalidate_range_start_trampoline(struct ib_umem_odp *item,
141 u64 start, u64 end, void *cookie)
143 ib_umem_notifier_start_account(item);
144 item->umem.context->invalidate_range(item, start, end);
148 static int ib_umem_notifier_invalidate_range_start(struct mmu_notifier *mn,
149 const struct mmu_notifier_range *range)
151 struct ib_ucontext_per_mm *per_mm =
152 container_of(mn, struct ib_ucontext_per_mm, mn);
154 if (range->blockable)
155 down_read(&per_mm->umem_rwsem);
156 else if (!down_read_trylock(&per_mm->umem_rwsem))
159 if (!per_mm->active) {
160 up_read(&per_mm->umem_rwsem);
162 * At this point active is permanently set and visible to this
163 * CPU without a lock, that fact is relied on to skip the unlock
169 return rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, range->start,
171 invalidate_range_start_trampoline,
172 range->blockable, NULL);
175 static int invalidate_range_end_trampoline(struct ib_umem_odp *item, u64 start,
176 u64 end, void *cookie)
178 ib_umem_notifier_end_account(item);
182 static void ib_umem_notifier_invalidate_range_end(struct mmu_notifier *mn,
183 const struct mmu_notifier_range *range)
185 struct ib_ucontext_per_mm *per_mm =
186 container_of(mn, struct ib_ucontext_per_mm, mn);
188 if (unlikely(!per_mm->active))
191 rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, range->start,
193 invalidate_range_end_trampoline, true, NULL);
194 up_read(&per_mm->umem_rwsem);
197 static const struct mmu_notifier_ops ib_umem_notifiers = {
198 .release = ib_umem_notifier_release,
199 .invalidate_range_start = ib_umem_notifier_invalidate_range_start,
200 .invalidate_range_end = ib_umem_notifier_invalidate_range_end,
203 static void add_umem_to_per_mm(struct ib_umem_odp *umem_odp)
205 struct ib_ucontext_per_mm *per_mm = umem_odp->per_mm;
206 struct ib_umem *umem = &umem_odp->umem;
208 down_write(&per_mm->umem_rwsem);
209 if (likely(ib_umem_start(umem) != ib_umem_end(umem)))
210 rbt_ib_umem_insert(&umem_odp->interval_tree,
212 up_write(&per_mm->umem_rwsem);
215 static void remove_umem_from_per_mm(struct ib_umem_odp *umem_odp)
217 struct ib_ucontext_per_mm *per_mm = umem_odp->per_mm;
218 struct ib_umem *umem = &umem_odp->umem;
220 down_write(&per_mm->umem_rwsem);
221 if (likely(ib_umem_start(umem) != ib_umem_end(umem)))
222 rbt_ib_umem_remove(&umem_odp->interval_tree,
224 complete_all(&umem_odp->notifier_completion);
226 up_write(&per_mm->umem_rwsem);
229 static struct ib_ucontext_per_mm *alloc_per_mm(struct ib_ucontext *ctx,
230 struct mm_struct *mm)
232 struct ib_ucontext_per_mm *per_mm;
235 per_mm = kzalloc(sizeof(*per_mm), GFP_KERNEL);
237 return ERR_PTR(-ENOMEM);
239 per_mm->context = ctx;
241 per_mm->umem_tree = RB_ROOT_CACHED;
242 init_rwsem(&per_mm->umem_rwsem);
243 per_mm->active = ctx->invalidate_range;
246 per_mm->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
249 WARN_ON(mm != current->mm);
251 per_mm->mn.ops = &ib_umem_notifiers;
252 ret = mmu_notifier_register(&per_mm->mn, per_mm->mm);
254 dev_err(&ctx->device->dev,
255 "Failed to register mmu_notifier %d\n", ret);
259 list_add(&per_mm->ucontext_list, &ctx->per_mm_list);
263 put_pid(per_mm->tgid);
268 static int get_per_mm(struct ib_umem_odp *umem_odp)
270 struct ib_ucontext *ctx = umem_odp->umem.context;
271 struct ib_ucontext_per_mm *per_mm;
274 * Generally speaking we expect only one or two per_mm in this list,
275 * so no reason to optimize this search today.
277 mutex_lock(&ctx->per_mm_list_lock);
278 list_for_each_entry(per_mm, &ctx->per_mm_list, ucontext_list) {
279 if (per_mm->mm == umem_odp->umem.owning_mm)
283 per_mm = alloc_per_mm(ctx, umem_odp->umem.owning_mm);
284 if (IS_ERR(per_mm)) {
285 mutex_unlock(&ctx->per_mm_list_lock);
286 return PTR_ERR(per_mm);
290 umem_odp->per_mm = per_mm;
291 per_mm->odp_mrs_count++;
292 mutex_unlock(&ctx->per_mm_list_lock);
297 static void free_per_mm(struct rcu_head *rcu)
299 kfree(container_of(rcu, struct ib_ucontext_per_mm, rcu));
302 void put_per_mm(struct ib_umem_odp *umem_odp)
304 struct ib_ucontext_per_mm *per_mm = umem_odp->per_mm;
305 struct ib_ucontext *ctx = umem_odp->umem.context;
308 mutex_lock(&ctx->per_mm_list_lock);
309 umem_odp->per_mm = NULL;
310 per_mm->odp_mrs_count--;
311 need_free = per_mm->odp_mrs_count == 0;
313 list_del(&per_mm->ucontext_list);
314 mutex_unlock(&ctx->per_mm_list_lock);
320 * NOTE! mmu_notifier_unregister() can happen between a start/end
321 * callback, resulting in an start/end, and thus an unbalanced
322 * lock. This doesn't really matter to us since we are about to kfree
323 * the memory that holds the lock, however LOCKDEP doesn't like this.
325 down_write(&per_mm->umem_rwsem);
326 per_mm->active = false;
327 up_write(&per_mm->umem_rwsem);
329 WARN_ON(!RB_EMPTY_ROOT(&per_mm->umem_tree.rb_root));
330 mmu_notifier_unregister_no_release(&per_mm->mn, per_mm->mm);
331 put_pid(per_mm->tgid);
332 mmu_notifier_call_srcu(&per_mm->rcu, free_per_mm);
335 struct ib_umem_odp *ib_alloc_odp_umem(struct ib_ucontext_per_mm *per_mm,
336 unsigned long addr, size_t size)
338 struct ib_ucontext *ctx = per_mm->context;
339 struct ib_umem_odp *odp_data;
340 struct ib_umem *umem;
341 int pages = size >> PAGE_SHIFT;
344 odp_data = kzalloc(sizeof(*odp_data), GFP_KERNEL);
346 return ERR_PTR(-ENOMEM);
347 umem = &odp_data->umem;
350 umem->address = addr;
351 umem->page_shift = PAGE_SHIFT;
354 odp_data->per_mm = per_mm;
355 umem->owning_mm = per_mm->mm;
356 mmgrab(umem->owning_mm);
358 mutex_init(&odp_data->umem_mutex);
359 init_completion(&odp_data->notifier_completion);
361 odp_data->page_list =
362 vzalloc(array_size(pages, sizeof(*odp_data->page_list)));
363 if (!odp_data->page_list) {
369 vzalloc(array_size(pages, sizeof(*odp_data->dma_list)));
370 if (!odp_data->dma_list) {
376 * Caller must ensure that the umem_odp that the per_mm came from
377 * cannot be freed during the call to ib_alloc_odp_umem.
379 mutex_lock(&ctx->per_mm_list_lock);
380 per_mm->odp_mrs_count++;
381 mutex_unlock(&ctx->per_mm_list_lock);
382 add_umem_to_per_mm(odp_data);
387 vfree(odp_data->page_list);
389 mmdrop(umem->owning_mm);
393 EXPORT_SYMBOL(ib_alloc_odp_umem);
395 int ib_umem_odp_get(struct ib_umem_odp *umem_odp, int access)
397 struct ib_umem *umem = &umem_odp->umem;
399 * NOTE: This must called in a process context where umem->owning_mm
402 struct mm_struct *mm = umem->owning_mm;
405 if (access & IB_ACCESS_HUGETLB) {
406 struct vm_area_struct *vma;
409 down_read(&mm->mmap_sem);
410 vma = find_vma(mm, ib_umem_start(umem));
411 if (!vma || !is_vm_hugetlb_page(vma)) {
412 up_read(&mm->mmap_sem);
416 umem->page_shift = huge_page_shift(h);
417 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;
505 int remove_existing_mapping = 0;
509 * Note: we avoid writing if seq is different from the initial seq, to
510 * handle case of a racing notifier. This check also allows us to bail
511 * early if we have a notifier running in parallel with us.
513 if (ib_umem_mmu_notifier_retry(umem_odp, current_seq)) {
517 if (!(umem_odp->dma_list[page_index])) {
518 dma_addr = ib_dma_map_page(dev,
520 0, BIT(umem->page_shift),
522 if (ib_dma_mapping_error(dev, dma_addr)) {
526 umem_odp->dma_list[page_index] = dma_addr | access_mask;
527 umem_odp->page_list[page_index] = page;
530 } else if (umem_odp->page_list[page_index] == page) {
531 umem_odp->dma_list[page_index] |= access_mask;
533 pr_err("error: got different pages in IB device and from get_user_pages. IB device page: %p, gup page: %p\n",
534 umem_odp->page_list[page_index], page);
535 /* Better remove the mapping now, to prevent any further
537 remove_existing_mapping = 1;
541 /* On Demand Paging - avoid pinning the page */
542 if (umem->context->invalidate_range || !stored_page)
545 if (remove_existing_mapping && umem->context->invalidate_range) {
546 ib_umem_notifier_start_account(umem_odp);
547 umem->context->invalidate_range(
549 ib_umem_start(umem) + (page_index << umem->page_shift),
550 ib_umem_start(umem) +
551 ((page_index + 1) << umem->page_shift));
552 ib_umem_notifier_end_account(umem_odp);
560 * ib_umem_odp_map_dma_pages - Pin and DMA map userspace memory in an ODP MR.
562 * Pins the range of pages passed in the argument, and maps them to
563 * DMA addresses. The DMA addresses of the mapped pages is updated in
564 * umem_odp->dma_list.
566 * Returns the number of pages mapped in success, negative error code
568 * An -EAGAIN error code is returned when a concurrent mmu notifier prevents
569 * the function from completing its task.
570 * An -ENOENT error code indicates that userspace process is being terminated
571 * and mm was already destroyed.
572 * @umem_odp: the umem to map and pin
573 * @user_virt: the address from which we need to map.
574 * @bcnt: the minimal number of bytes to pin and map. The mapping might be
575 * bigger due to alignment, and may also be smaller in case of an error
576 * pinning or mapping a page. The actual pages mapped is returned in
578 * @access_mask: bit mask of the requested access permissions for the given
580 * @current_seq: the MMU notifiers sequance value for synchronization with
581 * invalidations. the sequance number is read from
582 * umem_odp->notifiers_seq before calling this function
584 int ib_umem_odp_map_dma_pages(struct ib_umem_odp *umem_odp, u64 user_virt,
585 u64 bcnt, u64 access_mask,
586 unsigned long current_seq)
588 struct ib_umem *umem = &umem_odp->umem;
589 struct task_struct *owning_process = NULL;
590 struct mm_struct *owning_mm = umem_odp->umem.owning_mm;
591 struct page **local_page_list = NULL;
593 int j, k, ret = 0, start_idx, npages = 0, page_shift;
594 unsigned int flags = 0;
597 if (access_mask == 0)
600 if (user_virt < ib_umem_start(umem) ||
601 user_virt + bcnt > ib_umem_end(umem))
604 local_page_list = (struct page **)__get_free_page(GFP_KERNEL);
605 if (!local_page_list)
608 page_shift = umem->page_shift;
609 page_mask = ~(BIT(page_shift) - 1);
610 off = user_virt & (~page_mask);
611 user_virt = user_virt & page_mask;
612 bcnt += off; /* Charge for the first page offset as well. */
615 * owning_process is allowed to be NULL, this means somehow the mm is
616 * existing beyond the lifetime of the originating process.. Presumably
617 * mmget_not_zero will fail in this case.
619 owning_process = get_pid_task(umem_odp->per_mm->tgid, PIDTYPE_PID);
620 if (WARN_ON(!mmget_not_zero(umem_odp->umem.owning_mm))) {
625 if (access_mask & ODP_WRITE_ALLOWED_BIT)
628 start_idx = (user_virt - ib_umem_start(umem)) >> page_shift;
632 const size_t gup_num_pages = min_t(size_t,
633 (bcnt + BIT(page_shift) - 1) >> page_shift,
634 PAGE_SIZE / sizeof(struct page *));
636 down_read(&owning_mm->mmap_sem);
638 * Note: this might result in redundent page getting. We can
639 * avoid this by checking dma_list to be 0 before calling
640 * get_user_pages. However, this make the code much more
641 * complex (and doesn't gain us much performance in most use
644 npages = get_user_pages_remote(owning_process, owning_mm,
645 user_virt, gup_num_pages,
646 flags, local_page_list, NULL, NULL);
647 up_read(&owning_mm->mmap_sem);
650 if (npages != -EAGAIN)
651 pr_warn("fail to get %zu user pages with error %d\n", gup_num_pages, npages);
653 pr_debug("fail to get %zu user pages with error %d\n", gup_num_pages, npages);
657 bcnt -= min_t(size_t, npages << PAGE_SHIFT, bcnt);
658 mutex_lock(&umem_odp->umem_mutex);
659 for (j = 0; j < npages; j++, user_virt += PAGE_SIZE) {
660 if (user_virt & ~page_mask) {
662 if (page_to_phys(local_page_list[j]) != p) {
666 put_page(local_page_list[j]);
670 ret = ib_umem_odp_map_dma_single_page(
671 umem_odp, k, local_page_list[j],
672 access_mask, current_seq);
675 pr_warn("ib_umem_odp_map_dma_single_page failed with error %d\n", ret);
677 pr_debug("ib_umem_odp_map_dma_single_page failed with error %d\n", ret);
681 p = page_to_phys(local_page_list[j]);
684 mutex_unlock(&umem_odp->umem_mutex);
687 /* Release left over pages when handling errors. */
688 for (++j; j < npages; ++j)
689 put_page(local_page_list[j]);
695 if (npages < 0 && k == start_idx)
704 put_task_struct(owning_process);
705 free_page((unsigned long)local_page_list);
708 EXPORT_SYMBOL(ib_umem_odp_map_dma_pages);
710 void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, u64 virt,
713 struct ib_umem *umem = &umem_odp->umem;
716 struct ib_device *dev = umem->context->device;
718 virt = max_t(u64, virt, ib_umem_start(umem));
719 bound = min_t(u64, bound, ib_umem_end(umem));
720 /* Note that during the run of this function, the
721 * notifiers_count of the MR is > 0, preventing any racing
722 * faults from completion. We might be racing with other
723 * invalidations, so we must make sure we free each page only
725 mutex_lock(&umem_odp->umem_mutex);
726 for (addr = virt; addr < bound; addr += BIT(umem->page_shift)) {
727 idx = (addr - ib_umem_start(umem)) >> umem->page_shift;
728 if (umem_odp->page_list[idx]) {
729 struct page *page = umem_odp->page_list[idx];
730 dma_addr_t dma = umem_odp->dma_list[idx];
731 dma_addr_t dma_addr = dma & ODP_DMA_ADDR_MASK;
735 ib_dma_unmap_page(dev, dma_addr, PAGE_SIZE,
737 if (dma & ODP_WRITE_ALLOWED_BIT) {
738 struct page *head_page = compound_head(page);
740 * set_page_dirty prefers being called with
741 * the page lock. However, MMU notifiers are
742 * called sometimes with and sometimes without
743 * the lock. We rely on the umem_mutex instead
744 * to prevent other mmu notifiers from
745 * continuing and allowing the page mapping to
748 set_page_dirty(head_page);
750 /* on demand pinning support */
751 if (!umem->context->invalidate_range)
753 umem_odp->page_list[idx] = NULL;
754 umem_odp->dma_list[idx] = 0;
758 mutex_unlock(&umem_odp->umem_mutex);
760 EXPORT_SYMBOL(ib_umem_odp_unmap_dma_pages);
762 /* @last is not a part of the interval. See comment for function
765 int rbt_ib_umem_for_each_in_range(struct rb_root_cached *root,
772 struct umem_odp_node *node, *next;
773 struct ib_umem_odp *umem;
775 if (unlikely(start == last))
778 for (node = rbt_ib_umem_iter_first(root, start, last - 1);
780 /* TODO move the blockable decision up to the callback */
783 next = rbt_ib_umem_iter_next(node, start, last - 1);
784 umem = container_of(node, struct ib_umem_odp, interval_tree);
785 ret_val = cb(umem, start, last, cookie) || ret_val;
790 EXPORT_SYMBOL(rbt_ib_umem_for_each_in_range);
792 struct ib_umem_odp *rbt_ib_umem_lookup(struct rb_root_cached *root,
793 u64 addr, u64 length)
795 struct umem_odp_node *node;
797 node = rbt_ib_umem_iter_first(root, addr, addr + length - 1);
799 return container_of(node, struct ib_umem_odp, interval_tree);
803 EXPORT_SYMBOL(rbt_ib_umem_lookup);