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.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>
51 static void ib_umem_notifier_start_account(struct ib_umem_odp *umem_odp)
53 mutex_lock(&umem_odp->umem_mutex);
54 if (umem_odp->notifiers_count++ == 0)
56 * Initialize the completion object for waiting on
57 * notifiers. Since notifier_count is zero, no one should be
60 reinit_completion(&umem_odp->notifier_completion);
61 mutex_unlock(&umem_odp->umem_mutex);
64 static void ib_umem_notifier_end_account(struct ib_umem_odp *umem_odp)
66 mutex_lock(&umem_odp->umem_mutex);
68 * This sequence increase will notify the QP page fault that the page
69 * that is going to be mapped in the spte could have been freed.
71 ++umem_odp->notifiers_seq;
72 if (--umem_odp->notifiers_count == 0)
73 complete_all(&umem_odp->notifier_completion);
74 mutex_unlock(&umem_odp->umem_mutex);
77 static void ib_umem_notifier_release(struct mmu_notifier *mn,
80 struct ib_ucontext_per_mm *per_mm =
81 container_of(mn, struct ib_ucontext_per_mm, mn);
84 down_read(&per_mm->umem_rwsem);
85 if (!per_mm->mn.users)
88 for (node = rb_first_cached(&per_mm->umem_tree); node;
89 node = rb_next(node)) {
90 struct ib_umem_odp *umem_odp =
91 rb_entry(node, struct ib_umem_odp, interval_tree.rb);
94 * Increase the number of notifiers running, to prevent any
95 * further fault handling on this MR.
97 ib_umem_notifier_start_account(umem_odp);
98 complete_all(&umem_odp->notifier_completion);
99 umem_odp->umem.ibdev->ops.invalidate_range(
100 umem_odp, ib_umem_start(umem_odp),
101 ib_umem_end(umem_odp));
105 up_read(&per_mm->umem_rwsem);
108 static int invalidate_range_start_trampoline(struct ib_umem_odp *item,
109 u64 start, u64 end, void *cookie)
111 ib_umem_notifier_start_account(item);
112 item->umem.ibdev->ops.invalidate_range(item, start, end);
116 static int ib_umem_notifier_invalidate_range_start(struct mmu_notifier *mn,
117 const struct mmu_notifier_range *range)
119 struct ib_ucontext_per_mm *per_mm =
120 container_of(mn, struct ib_ucontext_per_mm, mn);
123 if (mmu_notifier_range_blockable(range))
124 down_read(&per_mm->umem_rwsem);
125 else if (!down_read_trylock(&per_mm->umem_rwsem))
128 if (!per_mm->mn.users) {
129 up_read(&per_mm->umem_rwsem);
131 * At this point users is permanently zero and visible to this
132 * CPU without a lock, that fact is relied on to skip the unlock
138 rc = rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, range->start,
140 invalidate_range_start_trampoline,
141 mmu_notifier_range_blockable(range),
144 up_read(&per_mm->umem_rwsem);
148 static int invalidate_range_end_trampoline(struct ib_umem_odp *item, u64 start,
149 u64 end, void *cookie)
151 ib_umem_notifier_end_account(item);
155 static void ib_umem_notifier_invalidate_range_end(struct mmu_notifier *mn,
156 const struct mmu_notifier_range *range)
158 struct ib_ucontext_per_mm *per_mm =
159 container_of(mn, struct ib_ucontext_per_mm, mn);
161 if (unlikely(!per_mm->mn.users))
164 rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, range->start,
166 invalidate_range_end_trampoline, true, NULL);
167 up_read(&per_mm->umem_rwsem);
170 static struct mmu_notifier *ib_umem_alloc_notifier(struct mm_struct *mm)
172 struct ib_ucontext_per_mm *per_mm;
174 per_mm = kzalloc(sizeof(*per_mm), GFP_KERNEL);
176 return ERR_PTR(-ENOMEM);
178 per_mm->umem_tree = RB_ROOT_CACHED;
179 init_rwsem(&per_mm->umem_rwsem);
181 WARN_ON(mm != current->mm);
183 per_mm->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
188 static void ib_umem_free_notifier(struct mmu_notifier *mn)
190 struct ib_ucontext_per_mm *per_mm =
191 container_of(mn, struct ib_ucontext_per_mm, mn);
193 WARN_ON(!RB_EMPTY_ROOT(&per_mm->umem_tree.rb_root));
195 put_pid(per_mm->tgid);
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,
203 .alloc_notifier = ib_umem_alloc_notifier,
204 .free_notifier = ib_umem_free_notifier,
207 static inline int ib_init_umem_odp(struct ib_umem_odp *umem_odp)
209 struct ib_ucontext_per_mm *per_mm;
210 struct mmu_notifier *mn;
213 umem_odp->umem.is_odp = 1;
214 if (!umem_odp->is_implicit_odp) {
215 size_t page_size = 1UL << umem_odp->page_shift;
218 umem_odp->interval_tree.start =
219 ALIGN_DOWN(umem_odp->umem.address, page_size);
220 if (check_add_overflow(umem_odp->umem.address,
221 (unsigned long)umem_odp->umem.length,
222 &umem_odp->interval_tree.last))
224 umem_odp->interval_tree.last =
225 ALIGN(umem_odp->interval_tree.last, page_size);
226 if (unlikely(umem_odp->interval_tree.last < page_size))
229 pages = (umem_odp->interval_tree.last -
230 umem_odp->interval_tree.start) >>
231 umem_odp->page_shift;
236 * Note that the representation of the intervals in the
237 * interval tree considers the ending point as contained in
240 umem_odp->interval_tree.last--;
242 umem_odp->page_list = kvcalloc(
243 pages, sizeof(*umem_odp->page_list), GFP_KERNEL);
244 if (!umem_odp->page_list)
247 umem_odp->dma_list = kvcalloc(
248 pages, sizeof(*umem_odp->dma_list), GFP_KERNEL);
249 if (!umem_odp->dma_list) {
255 mn = mmu_notifier_get(&ib_umem_notifiers, umem_odp->umem.owning_mm);
260 umem_odp->per_mm = per_mm =
261 container_of(mn, struct ib_ucontext_per_mm, mn);
263 mutex_init(&umem_odp->umem_mutex);
264 init_completion(&umem_odp->notifier_completion);
266 if (!umem_odp->is_implicit_odp) {
267 down_write(&per_mm->umem_rwsem);
268 interval_tree_insert(&umem_odp->interval_tree,
270 up_write(&per_mm->umem_rwsem);
272 mmgrab(umem_odp->umem.owning_mm);
277 kvfree(umem_odp->dma_list);
279 kvfree(umem_odp->page_list);
284 * ib_umem_odp_alloc_implicit - Allocate a parent implicit ODP umem
286 * Implicit ODP umems do not have a VA range and do not have any page lists.
287 * They exist only to hold the per_mm reference to help the driver create
290 * @udata: udata from the syscall being used to create the umem
291 * @access: ib_reg_mr access flags
293 struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_udata *udata,
296 struct ib_ucontext *context =
297 container_of(udata, struct uverbs_attr_bundle, driver_udata)
299 struct ib_umem *umem;
300 struct ib_umem_odp *umem_odp;
303 if (access & IB_ACCESS_HUGETLB)
304 return ERR_PTR(-EINVAL);
307 return ERR_PTR(-EIO);
308 if (WARN_ON_ONCE(!context->device->ops.invalidate_range))
309 return ERR_PTR(-EINVAL);
311 umem_odp = kzalloc(sizeof(*umem_odp), GFP_KERNEL);
313 return ERR_PTR(-ENOMEM);
314 umem = &umem_odp->umem;
315 umem->ibdev = context->device;
316 umem->writable = ib_access_writable(access);
317 umem->owning_mm = current->mm;
318 umem_odp->is_implicit_odp = 1;
319 umem_odp->page_shift = PAGE_SHIFT;
321 ret = ib_init_umem_odp(umem_odp);
328 EXPORT_SYMBOL(ib_umem_odp_alloc_implicit);
331 * ib_umem_odp_alloc_child - Allocate a child ODP umem under an implicit
334 * @root: The parent umem enclosing the child. This must be allocated using
335 * ib_alloc_implicit_odp_umem()
336 * @addr: The starting userspace VA
337 * @size: The length of the userspace VA
339 struct ib_umem_odp *ib_umem_odp_alloc_child(struct ib_umem_odp *root,
340 unsigned long addr, size_t size)
343 * Caller must ensure that root cannot be freed during the call to
346 struct ib_umem_odp *odp_data;
347 struct ib_umem *umem;
350 if (WARN_ON(!root->is_implicit_odp))
351 return ERR_PTR(-EINVAL);
353 odp_data = kzalloc(sizeof(*odp_data), GFP_KERNEL);
355 return ERR_PTR(-ENOMEM);
356 umem = &odp_data->umem;
357 umem->ibdev = root->umem.ibdev;
359 umem->address = addr;
360 umem->writable = root->umem.writable;
361 umem->owning_mm = root->umem.owning_mm;
362 odp_data->page_shift = PAGE_SHIFT;
364 ret = ib_init_umem_odp(odp_data);
371 EXPORT_SYMBOL(ib_umem_odp_alloc_child);
374 * ib_umem_odp_get - Create a umem_odp for a userspace va
376 * @udata: userspace context to pin memory for
377 * @addr: userspace virtual address to start at
378 * @size: length of region to pin
379 * @access: IB_ACCESS_xxx flags for memory being pinned
381 * The driver should use when the access flags indicate ODP memory. It avoids
382 * pinning, instead, stores the mm for future page fault handling in
383 * conjunction with MMU notifiers.
385 struct ib_umem_odp *ib_umem_odp_get(struct ib_udata *udata, unsigned long addr,
386 size_t size, int access)
388 struct ib_umem_odp *umem_odp;
389 struct ib_ucontext *context;
390 struct mm_struct *mm;
394 return ERR_PTR(-EIO);
396 context = container_of(udata, struct uverbs_attr_bundle, driver_udata)
399 return ERR_PTR(-EIO);
401 if (WARN_ON_ONCE(!(access & IB_ACCESS_ON_DEMAND)) ||
402 WARN_ON_ONCE(!context->device->ops.invalidate_range))
403 return ERR_PTR(-EINVAL);
405 umem_odp = kzalloc(sizeof(struct ib_umem_odp), GFP_KERNEL);
407 return ERR_PTR(-ENOMEM);
409 umem_odp->umem.ibdev = context->device;
410 umem_odp->umem.length = size;
411 umem_odp->umem.address = addr;
412 umem_odp->umem.writable = ib_access_writable(access);
413 umem_odp->umem.owning_mm = mm = current->mm;
415 umem_odp->page_shift = PAGE_SHIFT;
416 if (access & IB_ACCESS_HUGETLB) {
417 struct vm_area_struct *vma;
420 down_read(&mm->mmap_sem);
421 vma = find_vma(mm, ib_umem_start(umem_odp));
422 if (!vma || !is_vm_hugetlb_page(vma)) {
423 up_read(&mm->mmap_sem);
428 umem_odp->page_shift = huge_page_shift(h);
429 up_read(&mm->mmap_sem);
432 ret = ib_init_umem_odp(umem_odp);
441 EXPORT_SYMBOL(ib_umem_odp_get);
443 void ib_umem_odp_release(struct ib_umem_odp *umem_odp)
445 struct ib_ucontext_per_mm *per_mm = umem_odp->per_mm;
448 * Ensure that no more pages are mapped in the umem.
450 * It is the driver's responsibility to ensure, before calling us,
451 * that the hardware will not attempt to access the MR any more.
453 if (!umem_odp->is_implicit_odp) {
454 mutex_lock(&umem_odp->umem_mutex);
455 ib_umem_odp_unmap_dma_pages(umem_odp, ib_umem_start(umem_odp),
456 ib_umem_end(umem_odp));
457 mutex_unlock(&umem_odp->umem_mutex);
458 kvfree(umem_odp->dma_list);
459 kvfree(umem_odp->page_list);
462 down_write(&per_mm->umem_rwsem);
463 if (!umem_odp->is_implicit_odp) {
464 interval_tree_remove(&umem_odp->interval_tree,
466 complete_all(&umem_odp->notifier_completion);
469 * NOTE! mmu_notifier_unregister() can happen between a start/end
470 * callback, resulting in a missing end, and thus an unbalanced
471 * lock. This doesn't really matter to us since we are about to kfree
472 * the memory that holds the lock, however LOCKDEP doesn't like this.
473 * Thus we call the mmu_notifier_put under the rwsem and test the
474 * internal users count to reliably see if we are past this point.
476 mmu_notifier_put(&per_mm->mn);
477 up_write(&per_mm->umem_rwsem);
479 mmdrop(umem_odp->umem.owning_mm);
482 EXPORT_SYMBOL(ib_umem_odp_release);
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->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_user_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_odp *umem_odp,
507 unsigned long current_seq)
509 struct ib_device *dev = umem_odp->umem.ibdev;
511 int remove_existing_mapping = 0;
515 * Note: we avoid writing if seq is different from the initial seq, to
516 * handle case of a racing notifier. This check also allows us to bail
517 * early if we have a notifier running in parallel with us.
519 if (ib_umem_mmu_notifier_retry(umem_odp, current_seq)) {
523 if (!(umem_odp->dma_list[page_index])) {
525 ib_dma_map_page(dev, page, 0, BIT(umem_odp->page_shift),
527 if (ib_dma_mapping_error(dev, dma_addr)) {
531 umem_odp->dma_list[page_index] = dma_addr | access_mask;
532 umem_odp->page_list[page_index] = page;
534 } else if (umem_odp->page_list[page_index] == page) {
535 umem_odp->dma_list[page_index] |= access_mask;
537 pr_err("error: got different pages in IB device and from get_user_pages. IB device page: %p, gup page: %p\n",
538 umem_odp->page_list[page_index], page);
539 /* Better remove the mapping now, to prevent any further
541 remove_existing_mapping = 1;
547 if (remove_existing_mapping) {
548 ib_umem_notifier_start_account(umem_odp);
549 dev->ops.invalidate_range(
551 ib_umem_start(umem_odp) +
552 (page_index << umem_odp->page_shift),
553 ib_umem_start(umem_odp) +
554 ((page_index + 1) << umem_odp->page_shift));
555 ib_umem_notifier_end_account(umem_odp);
563 * ib_umem_odp_map_dma_pages - Pin and DMA map userspace memory in an ODP MR.
565 * Pins the range of pages passed in the argument, and maps them to
566 * DMA addresses. The DMA addresses of the mapped pages is updated in
567 * umem_odp->dma_list.
569 * Returns the number of pages mapped in success, negative error code
571 * An -EAGAIN error code is returned when a concurrent mmu notifier prevents
572 * the function from completing its task.
573 * An -ENOENT error code indicates that userspace process is being terminated
574 * and mm was already destroyed.
575 * @umem_odp: the umem to map and pin
576 * @user_virt: the address from which we need to map.
577 * @bcnt: the minimal number of bytes to pin and map. The mapping might be
578 * bigger due to alignment, and may also be smaller in case of an error
579 * pinning or mapping a page. The actual pages mapped is returned in
581 * @access_mask: bit mask of the requested access permissions for the given
583 * @current_seq: the MMU notifiers sequance value for synchronization with
584 * invalidations. the sequance number is read from
585 * umem_odp->notifiers_seq before calling this function
587 int ib_umem_odp_map_dma_pages(struct ib_umem_odp *umem_odp, u64 user_virt,
588 u64 bcnt, u64 access_mask,
589 unsigned long current_seq)
591 struct task_struct *owning_process = NULL;
592 struct mm_struct *owning_mm = umem_odp->umem.owning_mm;
593 struct page **local_page_list = NULL;
595 int j, k, ret = 0, start_idx, npages = 0;
596 unsigned int flags = 0, page_shift;
599 if (access_mask == 0)
602 if (user_virt < ib_umem_start(umem_odp) ||
603 user_virt + bcnt > ib_umem_end(umem_odp))
606 local_page_list = (struct page **)__get_free_page(GFP_KERNEL);
607 if (!local_page_list)
610 page_shift = umem_odp->page_shift;
611 page_mask = ~(BIT(page_shift) - 1);
612 off = user_virt & (~page_mask);
613 user_virt = user_virt & page_mask;
614 bcnt += off; /* Charge for the first page offset as well. */
617 * owning_process is allowed to be NULL, this means somehow the mm is
618 * existing beyond the lifetime of the originating process.. Presumably
619 * mmget_not_zero will fail in this case.
621 owning_process = get_pid_task(umem_odp->per_mm->tgid, PIDTYPE_PID);
622 if (!owning_process || !mmget_not_zero(owning_mm)) {
627 if (access_mask & ODP_WRITE_ALLOWED_BIT)
630 start_idx = (user_virt - ib_umem_start(umem_odp)) >> page_shift;
634 const size_t gup_num_pages = min_t(size_t,
635 (bcnt + BIT(page_shift) - 1) >> page_shift,
636 PAGE_SIZE / sizeof(struct page *));
638 down_read(&owning_mm->mmap_sem);
640 * Note: this might result in redundent page getting. We can
641 * avoid this by checking dma_list to be 0 before calling
642 * get_user_pages. However, this make the code much more
643 * complex (and doesn't gain us much performance in most use
646 npages = get_user_pages_remote(owning_process, owning_mm,
647 user_virt, gup_num_pages,
648 flags, local_page_list, NULL, NULL);
649 up_read(&owning_mm->mmap_sem);
652 if (npages != -EAGAIN)
653 pr_warn("fail to get %zu user pages with error %d\n", gup_num_pages, npages);
655 pr_debug("fail to get %zu user pages with error %d\n", gup_num_pages, npages);
659 bcnt -= min_t(size_t, npages << PAGE_SHIFT, bcnt);
660 mutex_lock(&umem_odp->umem_mutex);
661 for (j = 0; j < npages; j++, user_virt += PAGE_SIZE) {
662 if (user_virt & ~page_mask) {
664 if (page_to_phys(local_page_list[j]) != p) {
668 put_user_page(local_page_list[j]);
672 ret = ib_umem_odp_map_dma_single_page(
673 umem_odp, k, local_page_list[j],
674 access_mask, current_seq);
677 pr_warn("ib_umem_odp_map_dma_single_page failed with error %d\n", ret);
679 pr_debug("ib_umem_odp_map_dma_single_page failed with error %d\n", ret);
683 p = page_to_phys(local_page_list[j]);
686 mutex_unlock(&umem_odp->umem_mutex);
690 * Release pages, remembering that the first page
691 * to hit an error was already released by
692 * ib_umem_odp_map_dma_single_page().
694 if (npages - (j + 1) > 0)
695 put_user_pages(&local_page_list[j+1],
702 if (npages < 0 && k == start_idx)
711 put_task_struct(owning_process);
712 free_page((unsigned long)local_page_list);
715 EXPORT_SYMBOL(ib_umem_odp_map_dma_pages);
717 void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, u64 virt,
722 struct ib_device *dev = umem_odp->umem.ibdev;
724 lockdep_assert_held(&umem_odp->umem_mutex);
726 virt = max_t(u64, virt, ib_umem_start(umem_odp));
727 bound = min_t(u64, bound, ib_umem_end(umem_odp));
728 /* Note that during the run of this function, the
729 * notifiers_count of the MR is > 0, preventing any racing
730 * faults from completion. We might be racing with other
731 * invalidations, so we must make sure we free each page only
733 for (addr = virt; addr < bound; addr += BIT(umem_odp->page_shift)) {
734 idx = (addr - ib_umem_start(umem_odp)) >> umem_odp->page_shift;
735 if (umem_odp->page_list[idx]) {
736 struct page *page = umem_odp->page_list[idx];
737 dma_addr_t dma = umem_odp->dma_list[idx];
738 dma_addr_t dma_addr = dma & ODP_DMA_ADDR_MASK;
742 ib_dma_unmap_page(dev, dma_addr,
743 BIT(umem_odp->page_shift),
745 if (dma & ODP_WRITE_ALLOWED_BIT) {
746 struct page *head_page = compound_head(page);
748 * set_page_dirty prefers being called with
749 * the page lock. However, MMU notifiers are
750 * called sometimes with and sometimes without
751 * the lock. We rely on the umem_mutex instead
752 * to prevent other mmu notifiers from
753 * continuing and allowing the page mapping to
756 set_page_dirty(head_page);
758 umem_odp->page_list[idx] = NULL;
759 umem_odp->dma_list[idx] = 0;
764 EXPORT_SYMBOL(ib_umem_odp_unmap_dma_pages);
766 /* @last is not a part of the interval. See comment for function
769 int rbt_ib_umem_for_each_in_range(struct rb_root_cached *root,
776 struct interval_tree_node *node, *next;
777 struct ib_umem_odp *umem;
779 if (unlikely(start == last))
782 for (node = interval_tree_iter_first(root, start, last - 1);
784 /* TODO move the blockable decision up to the callback */
787 next = interval_tree_iter_next(node, start, last - 1);
788 umem = container_of(node, struct ib_umem_odp, interval_tree);
789 ret_val = cb(umem, start, last, cookie) || ret_val;