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);
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) - 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)
111 * Increase the number of notifiers running, to
112 * prevent any further fault handling on this MR.
114 ib_umem_notifier_start_account(umem_odp);
116 /* Make sure that the fact the umem is dying is out before we release
117 * all pending page faults. */
119 complete_all(&umem_odp->notifier_completion);
120 umem_odp->umem.context->invalidate_range(
121 umem_odp, ib_umem_start(umem_odp), ib_umem_end(umem_odp));
125 static void ib_umem_notifier_release(struct mmu_notifier *mn,
126 struct mm_struct *mm)
128 struct ib_ucontext_per_mm *per_mm =
129 container_of(mn, struct ib_ucontext_per_mm, mn);
131 down_read(&per_mm->umem_rwsem);
133 rbt_ib_umem_for_each_in_range(
134 &per_mm->umem_tree, 0, ULLONG_MAX,
135 ib_umem_notifier_release_trampoline, true, NULL);
136 up_read(&per_mm->umem_rwsem);
139 static int invalidate_range_start_trampoline(struct ib_umem_odp *item,
140 u64 start, u64 end, void *cookie)
142 ib_umem_notifier_start_account(item);
143 item->umem.context->invalidate_range(item, start, end);
147 static int ib_umem_notifier_invalidate_range_start(struct mmu_notifier *mn,
148 const struct mmu_notifier_range *range)
150 struct ib_ucontext_per_mm *per_mm =
151 container_of(mn, struct ib_ucontext_per_mm, mn);
154 if (mmu_notifier_range_blockable(range))
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 rc = rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, range->start,
171 invalidate_range_start_trampoline,
172 mmu_notifier_range_blockable(range),
175 up_read(&per_mm->umem_rwsem);
179 static int invalidate_range_end_trampoline(struct ib_umem_odp *item, u64 start,
180 u64 end, void *cookie)
182 ib_umem_notifier_end_account(item);
186 static void ib_umem_notifier_invalidate_range_end(struct mmu_notifier *mn,
187 const struct mmu_notifier_range *range)
189 struct ib_ucontext_per_mm *per_mm =
190 container_of(mn, struct ib_ucontext_per_mm, mn);
192 if (unlikely(!per_mm->active))
195 rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, range->start,
197 invalidate_range_end_trampoline, true, NULL);
198 up_read(&per_mm->umem_rwsem);
201 static const struct mmu_notifier_ops ib_umem_notifiers = {
202 .release = ib_umem_notifier_release,
203 .invalidate_range_start = ib_umem_notifier_invalidate_range_start,
204 .invalidate_range_end = ib_umem_notifier_invalidate_range_end,
207 static void add_umem_to_per_mm(struct ib_umem_odp *umem_odp)
209 struct ib_ucontext_per_mm *per_mm = umem_odp->per_mm;
211 down_write(&per_mm->umem_rwsem);
212 if (likely(ib_umem_start(umem_odp) != ib_umem_end(umem_odp)))
213 rbt_ib_umem_insert(&umem_odp->interval_tree,
215 up_write(&per_mm->umem_rwsem);
218 static void remove_umem_from_per_mm(struct ib_umem_odp *umem_odp)
220 struct ib_ucontext_per_mm *per_mm = umem_odp->per_mm;
222 down_write(&per_mm->umem_rwsem);
223 if (likely(ib_umem_start(umem_odp) != ib_umem_end(umem_odp)))
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 odp_data->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 umem_odp->page_shift = PAGE_SHIFT;
409 if (access & IB_ACCESS_HUGETLB) {
410 struct vm_area_struct *vma;
413 down_read(&mm->mmap_sem);
414 vma = find_vma(mm, ib_umem_start(umem_odp));
415 if (!vma || !is_vm_hugetlb_page(vma)) {
416 up_read(&mm->mmap_sem);
420 umem_odp->page_shift = huge_page_shift(h);
421 up_read(&mm->mmap_sem);
424 mutex_init(&umem_odp->umem_mutex);
426 init_completion(&umem_odp->notifier_completion);
428 if (ib_umem_odp_num_pages(umem_odp)) {
429 umem_odp->page_list =
430 vzalloc(array_size(sizeof(*umem_odp->page_list),
431 ib_umem_odp_num_pages(umem_odp)));
432 if (!umem_odp->page_list)
436 vzalloc(array_size(sizeof(*umem_odp->dma_list),
437 ib_umem_odp_num_pages(umem_odp)));
438 if (!umem_odp->dma_list) {
444 ret_val = get_per_mm(umem_odp);
447 add_umem_to_per_mm(umem_odp);
452 vfree(umem_odp->dma_list);
454 vfree(umem_odp->page_list);
458 void ib_umem_odp_release(struct ib_umem_odp *umem_odp)
461 * Ensure that no more pages are mapped in the umem.
463 * It is the driver's responsibility to ensure, before calling us,
464 * that the hardware will not attempt to access the MR any more.
466 ib_umem_odp_unmap_dma_pages(umem_odp, ib_umem_start(umem_odp),
467 ib_umem_end(umem_odp));
469 remove_umem_from_per_mm(umem_odp);
470 put_per_mm(umem_odp);
471 vfree(umem_odp->dma_list);
472 vfree(umem_odp->page_list);
476 * Map for DMA and insert a single page into the on-demand paging page tables.
478 * @umem: the umem to insert the page to.
479 * @page_index: index in the umem to add the page to.
480 * @page: the page struct to map and add.
481 * @access_mask: access permissions needed for this page.
482 * @current_seq: sequence number for synchronization with invalidations.
483 * the sequence number is taken from
484 * umem_odp->notifiers_seq.
486 * The function returns -EFAULT if the DMA mapping operation fails. It returns
487 * -EAGAIN if a concurrent invalidation prevents us from updating the page.
489 * The page is released via put_user_page even if the operation failed. For
490 * on-demand pinning, the page is released whenever it isn't stored in the
493 static int ib_umem_odp_map_dma_single_page(
494 struct ib_umem_odp *umem_odp,
498 unsigned long current_seq)
500 struct ib_ucontext *context = umem_odp->umem.context;
501 struct ib_device *dev = context->device;
503 int remove_existing_mapping = 0;
507 * Note: we avoid writing if seq is different from the initial seq, to
508 * handle case of a racing notifier. This check also allows us to bail
509 * early if we have a notifier running in parallel with us.
511 if (ib_umem_mmu_notifier_retry(umem_odp, current_seq)) {
515 if (!(umem_odp->dma_list[page_index])) {
517 ib_dma_map_page(dev, page, 0, BIT(umem_odp->page_shift),
519 if (ib_dma_mapping_error(dev, dma_addr)) {
523 umem_odp->dma_list[page_index] = dma_addr | access_mask;
524 umem_odp->page_list[page_index] = page;
526 } else if (umem_odp->page_list[page_index] == page) {
527 umem_odp->dma_list[page_index] |= access_mask;
529 pr_err("error: got different pages in IB device and from get_user_pages. IB device page: %p, gup page: %p\n",
530 umem_odp->page_list[page_index], page);
531 /* Better remove the mapping now, to prevent any further
533 remove_existing_mapping = 1;
539 if (remove_existing_mapping) {
540 ib_umem_notifier_start_account(umem_odp);
541 context->invalidate_range(
543 ib_umem_start(umem_odp) +
544 (page_index << umem_odp->page_shift),
545 ib_umem_start(umem_odp) +
546 ((page_index + 1) << umem_odp->page_shift));
547 ib_umem_notifier_end_account(umem_odp);
555 * ib_umem_odp_map_dma_pages - Pin and DMA map userspace memory in an ODP MR.
557 * Pins the range of pages passed in the argument, and maps them to
558 * DMA addresses. The DMA addresses of the mapped pages is updated in
559 * umem_odp->dma_list.
561 * Returns the number of pages mapped in success, negative error code
563 * An -EAGAIN error code is returned when a concurrent mmu notifier prevents
564 * the function from completing its task.
565 * An -ENOENT error code indicates that userspace process is being terminated
566 * and mm was already destroyed.
567 * @umem_odp: the umem to map and pin
568 * @user_virt: the address from which we need to map.
569 * @bcnt: the minimal number of bytes to pin and map. The mapping might be
570 * bigger due to alignment, and may also be smaller in case of an error
571 * pinning or mapping a page. The actual pages mapped is returned in
573 * @access_mask: bit mask of the requested access permissions for the given
575 * @current_seq: the MMU notifiers sequance value for synchronization with
576 * invalidations. the sequance number is read from
577 * umem_odp->notifiers_seq before calling this function
579 int ib_umem_odp_map_dma_pages(struct ib_umem_odp *umem_odp, u64 user_virt,
580 u64 bcnt, u64 access_mask,
581 unsigned long current_seq)
583 struct task_struct *owning_process = NULL;
584 struct mm_struct *owning_mm = umem_odp->umem.owning_mm;
585 struct page **local_page_list = NULL;
587 int j, k, ret = 0, start_idx, npages = 0;
588 unsigned int flags = 0, page_shift;
591 if (access_mask == 0)
594 if (user_virt < ib_umem_start(umem_odp) ||
595 user_virt + bcnt > ib_umem_end(umem_odp))
598 local_page_list = (struct page **)__get_free_page(GFP_KERNEL);
599 if (!local_page_list)
602 page_shift = umem_odp->page_shift;
603 page_mask = ~(BIT(page_shift) - 1);
604 off = user_virt & (~page_mask);
605 user_virt = user_virt & page_mask;
606 bcnt += off; /* Charge for the first page offset as well. */
609 * owning_process is allowed to be NULL, this means somehow the mm is
610 * existing beyond the lifetime of the originating process.. Presumably
611 * mmget_not_zero will fail in this case.
613 owning_process = get_pid_task(umem_odp->per_mm->tgid, PIDTYPE_PID);
614 if (!owning_process || !mmget_not_zero(owning_mm)) {
619 if (access_mask & ODP_WRITE_ALLOWED_BIT)
622 start_idx = (user_virt - ib_umem_start(umem_odp)) >> page_shift;
626 const size_t gup_num_pages = min_t(size_t,
627 (bcnt + BIT(page_shift) - 1) >> page_shift,
628 PAGE_SIZE / sizeof(struct page *));
630 down_read(&owning_mm->mmap_sem);
632 * Note: this might result in redundent page getting. We can
633 * avoid this by checking dma_list to be 0 before calling
634 * get_user_pages. However, this make the code much more
635 * complex (and doesn't gain us much performance in most use
638 npages = get_user_pages_remote(owning_process, owning_mm,
639 user_virt, gup_num_pages,
640 flags, local_page_list, NULL, NULL);
641 up_read(&owning_mm->mmap_sem);
644 if (npages != -EAGAIN)
645 pr_warn("fail to get %zu user pages with error %d\n", gup_num_pages, npages);
647 pr_debug("fail to get %zu user pages with error %d\n", gup_num_pages, npages);
651 bcnt -= min_t(size_t, npages << PAGE_SHIFT, bcnt);
652 mutex_lock(&umem_odp->umem_mutex);
653 for (j = 0; j < npages; j++, user_virt += PAGE_SIZE) {
654 if (user_virt & ~page_mask) {
656 if (page_to_phys(local_page_list[j]) != p) {
660 put_user_page(local_page_list[j]);
664 ret = ib_umem_odp_map_dma_single_page(
665 umem_odp, k, local_page_list[j],
666 access_mask, current_seq);
669 pr_warn("ib_umem_odp_map_dma_single_page failed with error %d\n", ret);
671 pr_debug("ib_umem_odp_map_dma_single_page failed with error %d\n", ret);
675 p = page_to_phys(local_page_list[j]);
678 mutex_unlock(&umem_odp->umem_mutex);
682 * Release pages, remembering that the first page
683 * to hit an error was already released by
684 * ib_umem_odp_map_dma_single_page().
686 if (npages - (j + 1) > 0)
687 put_user_pages(&local_page_list[j+1],
694 if (npages < 0 && k == start_idx)
703 put_task_struct(owning_process);
704 free_page((unsigned long)local_page_list);
707 EXPORT_SYMBOL(ib_umem_odp_map_dma_pages);
709 void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, u64 virt,
714 struct ib_device *dev = umem_odp->umem.context->device;
716 virt = max_t(u64, virt, ib_umem_start(umem_odp));
717 bound = min_t(u64, bound, ib_umem_end(umem_odp));
718 /* Note that during the run of this function, the
719 * notifiers_count of the MR is > 0, preventing any racing
720 * faults from completion. We might be racing with other
721 * invalidations, so we must make sure we free each page only
723 mutex_lock(&umem_odp->umem_mutex);
724 for (addr = virt; addr < bound; addr += BIT(umem_odp->page_shift)) {
725 idx = (addr - ib_umem_start(umem_odp)) >> umem_odp->page_shift;
726 if (umem_odp->page_list[idx]) {
727 struct page *page = umem_odp->page_list[idx];
728 dma_addr_t dma = umem_odp->dma_list[idx];
729 dma_addr_t dma_addr = dma & ODP_DMA_ADDR_MASK;
733 ib_dma_unmap_page(dev, dma_addr,
734 BIT(umem_odp->page_shift),
736 if (dma & ODP_WRITE_ALLOWED_BIT) {
737 struct page *head_page = compound_head(page);
739 * set_page_dirty prefers being called with
740 * the page lock. However, MMU notifiers are
741 * called sometimes with and sometimes without
742 * the lock. We rely on the umem_mutex instead
743 * to prevent other mmu notifiers from
744 * continuing and allowing the page mapping to
747 set_page_dirty(head_page);
749 umem_odp->page_list[idx] = NULL;
750 umem_odp->dma_list[idx] = 0;
754 mutex_unlock(&umem_odp->umem_mutex);
756 EXPORT_SYMBOL(ib_umem_odp_unmap_dma_pages);
758 /* @last is not a part of the interval. See comment for function
761 int rbt_ib_umem_for_each_in_range(struct rb_root_cached *root,
768 struct umem_odp_node *node, *next;
769 struct ib_umem_odp *umem;
771 if (unlikely(start == last))
774 for (node = rbt_ib_umem_iter_first(root, start, last - 1);
776 /* TODO move the blockable decision up to the callback */
779 next = rbt_ib_umem_iter_next(node, start, last - 1);
780 umem = container_of(node, struct ib_umem_odp, interval_tree);
781 ret_val = cb(umem, start, last, cookie) || ret_val;
786 EXPORT_SYMBOL(rbt_ib_umem_for_each_in_range);
788 struct ib_umem_odp *rbt_ib_umem_lookup(struct rb_root_cached *root,
789 u64 addr, u64 length)
791 struct umem_odp_node *node;
793 node = rbt_ib_umem_iter_first(root, addr, addr + length - 1);
795 return container_of(node, struct ib_umem_odp, interval_tree);
799 EXPORT_SYMBOL(rbt_ib_umem_lookup);