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/hmm.h>
44 #include <linux/pagemap.h>
46 #include <rdma/ib_verbs.h>
47 #include <rdma/ib_umem.h>
48 #include <rdma/ib_umem_odp.h>
52 static inline int ib_init_umem_odp(struct ib_umem_odp *umem_odp,
53 const struct mmu_interval_notifier_ops *ops)
57 umem_odp->umem.is_odp = 1;
58 mutex_init(&umem_odp->umem_mutex);
60 if (!umem_odp->is_implicit_odp) {
61 size_t page_size = 1UL << umem_odp->page_shift;
66 start = ALIGN_DOWN(umem_odp->umem.address, page_size);
67 if (check_add_overflow(umem_odp->umem.address,
68 (unsigned long)umem_odp->umem.length,
71 end = ALIGN(end, page_size);
72 if (unlikely(end < page_size))
75 ndmas = (end - start) >> umem_odp->page_shift;
79 npfns = (end - start) >> PAGE_SHIFT;
80 umem_odp->pfn_list = kvcalloc(
81 npfns, sizeof(*umem_odp->pfn_list), GFP_KERNEL);
82 if (!umem_odp->pfn_list)
85 umem_odp->dma_list = kvcalloc(
86 ndmas, sizeof(*umem_odp->dma_list), GFP_KERNEL);
87 if (!umem_odp->dma_list) {
92 ret = mmu_interval_notifier_insert(&umem_odp->notifier,
93 umem_odp->umem.owning_mm,
94 start, end - start, ops);
102 kvfree(umem_odp->dma_list);
104 kvfree(umem_odp->pfn_list);
109 * ib_umem_odp_alloc_implicit - Allocate a parent implicit ODP umem
111 * Implicit ODP umems do not have a VA range and do not have any page lists.
112 * They exist only to hold the per_mm reference to help the driver create
115 * @device: IB device to create UMEM
116 * @access: ib_reg_mr access flags
118 struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_device *device,
121 struct ib_umem *umem;
122 struct ib_umem_odp *umem_odp;
125 if (access & IB_ACCESS_HUGETLB)
126 return ERR_PTR(-EINVAL);
128 umem_odp = kzalloc(sizeof(*umem_odp), GFP_KERNEL);
130 return ERR_PTR(-ENOMEM);
131 umem = &umem_odp->umem;
132 umem->ibdev = device;
133 umem->writable = ib_access_writable(access);
134 umem->owning_mm = current->mm;
135 umem_odp->is_implicit_odp = 1;
136 umem_odp->page_shift = PAGE_SHIFT;
138 umem_odp->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
139 ret = ib_init_umem_odp(umem_odp, NULL);
141 put_pid(umem_odp->tgid);
147 EXPORT_SYMBOL(ib_umem_odp_alloc_implicit);
150 * ib_umem_odp_alloc_child - Allocate a child ODP umem under an implicit
153 * @root: The parent umem enclosing the child. This must be allocated using
154 * ib_alloc_implicit_odp_umem()
155 * @addr: The starting userspace VA
156 * @size: The length of the userspace VA
157 * @ops: MMU interval ops, currently only @invalidate
160 ib_umem_odp_alloc_child(struct ib_umem_odp *root, unsigned long addr,
162 const struct mmu_interval_notifier_ops *ops)
165 * Caller must ensure that root cannot be freed during the call to
168 struct ib_umem_odp *odp_data;
169 struct ib_umem *umem;
172 if (WARN_ON(!root->is_implicit_odp))
173 return ERR_PTR(-EINVAL);
175 odp_data = kzalloc(sizeof(*odp_data), GFP_KERNEL);
177 return ERR_PTR(-ENOMEM);
178 umem = &odp_data->umem;
179 umem->ibdev = root->umem.ibdev;
181 umem->address = addr;
182 umem->writable = root->umem.writable;
183 umem->owning_mm = root->umem.owning_mm;
184 odp_data->page_shift = PAGE_SHIFT;
185 odp_data->notifier.ops = ops;
188 * A mmget must be held when registering a notifier, the owming_mm only
189 * has a mm_grab at this point.
191 if (!mmget_not_zero(umem->owning_mm)) {
196 odp_data->tgid = get_pid(root->tgid);
197 ret = ib_init_umem_odp(odp_data, ops);
200 mmput(umem->owning_mm);
204 put_pid(odp_data->tgid);
205 mmput(umem->owning_mm);
210 EXPORT_SYMBOL(ib_umem_odp_alloc_child);
213 * ib_umem_odp_get - Create a umem_odp for a userspace va
215 * @device: IB device struct to get UMEM
216 * @addr: userspace virtual address to start at
217 * @size: length of region to pin
218 * @access: IB_ACCESS_xxx flags for memory being pinned
219 * @ops: MMU interval ops, currently only @invalidate
221 * The driver should use when the access flags indicate ODP memory. It avoids
222 * pinning, instead, stores the mm for future page fault handling in
223 * conjunction with MMU notifiers.
225 struct ib_umem_odp *ib_umem_odp_get(struct ib_device *device,
226 unsigned long addr, size_t size, int access,
227 const struct mmu_interval_notifier_ops *ops)
229 struct ib_umem_odp *umem_odp;
230 struct mm_struct *mm;
233 if (WARN_ON_ONCE(!(access & IB_ACCESS_ON_DEMAND)))
234 return ERR_PTR(-EINVAL);
236 umem_odp = kzalloc(sizeof(struct ib_umem_odp), GFP_KERNEL);
238 return ERR_PTR(-ENOMEM);
240 umem_odp->umem.ibdev = device;
241 umem_odp->umem.length = size;
242 umem_odp->umem.address = addr;
243 umem_odp->umem.writable = ib_access_writable(access);
244 umem_odp->umem.owning_mm = mm = current->mm;
245 umem_odp->notifier.ops = ops;
247 umem_odp->page_shift = PAGE_SHIFT;
248 #ifdef CONFIG_HUGETLB_PAGE
249 if (access & IB_ACCESS_HUGETLB)
250 umem_odp->page_shift = HPAGE_SHIFT;
253 umem_odp->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
254 ret = ib_init_umem_odp(umem_odp, ops);
260 put_pid(umem_odp->tgid);
264 EXPORT_SYMBOL(ib_umem_odp_get);
266 void ib_umem_odp_release(struct ib_umem_odp *umem_odp)
269 * Ensure that no more pages are mapped in the umem.
271 * It is the driver's responsibility to ensure, before calling us,
272 * that the hardware will not attempt to access the MR any more.
274 if (!umem_odp->is_implicit_odp) {
275 mutex_lock(&umem_odp->umem_mutex);
276 ib_umem_odp_unmap_dma_pages(umem_odp, ib_umem_start(umem_odp),
277 ib_umem_end(umem_odp));
278 mutex_unlock(&umem_odp->umem_mutex);
279 mmu_interval_notifier_remove(&umem_odp->notifier);
280 kvfree(umem_odp->dma_list);
281 kvfree(umem_odp->pfn_list);
283 put_pid(umem_odp->tgid);
286 EXPORT_SYMBOL(ib_umem_odp_release);
289 * Map for DMA and insert a single page into the on-demand paging page tables.
291 * @umem: the umem to insert the page to.
292 * @dma_index: index in the umem to add the dma to.
293 * @page: the page struct to map and add.
294 * @access_mask: access permissions needed for this page.
295 * @current_seq: sequence number for synchronization with invalidations.
296 * the sequence number is taken from
297 * umem_odp->notifiers_seq.
299 * The function returns -EFAULT if the DMA mapping operation fails.
302 static int ib_umem_odp_map_dma_single_page(
303 struct ib_umem_odp *umem_odp,
304 unsigned int dma_index,
308 struct ib_device *dev = umem_odp->umem.ibdev;
309 dma_addr_t *dma_addr = &umem_odp->dma_list[dma_index];
313 * If the page is already dma mapped it means it went through
314 * a non-invalidating trasition, like read-only to writable.
317 *dma_addr = (*dma_addr & ODP_DMA_ADDR_MASK) | access_mask;
321 *dma_addr = ib_dma_map_page(dev, page, 0, 1 << umem_odp->page_shift,
323 if (ib_dma_mapping_error(dev, *dma_addr)) {
328 *dma_addr |= access_mask;
333 * ib_umem_odp_map_dma_and_lock - DMA map userspace memory in an ODP MR and lock it.
335 * Maps the range passed in the argument to DMA addresses.
336 * The DMA addresses of the mapped pages is updated in umem_odp->dma_list.
337 * Upon success the ODP MR will be locked to let caller complete its device
340 * Returns the number of pages mapped in success, negative error code
342 * @umem_odp: the umem to map and pin
343 * @user_virt: the address from which we need to map.
344 * @bcnt: the minimal number of bytes to pin and map. The mapping might be
345 * bigger due to alignment, and may also be smaller in case of an error
346 * pinning or mapping a page. The actual pages mapped is returned in
348 * @access_mask: bit mask of the requested access permissions for the given
350 * @fault: is faulting required for the given range
352 int ib_umem_odp_map_dma_and_lock(struct ib_umem_odp *umem_odp, u64 user_virt,
353 u64 bcnt, u64 access_mask, bool fault)
354 __acquires(&umem_odp->umem_mutex)
356 struct task_struct *owning_process = NULL;
357 struct mm_struct *owning_mm = umem_odp->umem.owning_mm;
358 int pfn_index, dma_index, ret = 0, start_idx;
359 unsigned int page_shift, hmm_order, pfn_start_idx;
360 unsigned long num_pfns, current_seq;
361 struct hmm_range range = {};
362 unsigned long timeout;
364 if (access_mask == 0)
367 if (user_virt < ib_umem_start(umem_odp) ||
368 user_virt + bcnt > ib_umem_end(umem_odp))
371 page_shift = umem_odp->page_shift;
374 * owning_process is allowed to be NULL, this means somehow the mm is
375 * existing beyond the lifetime of the originating process.. Presumably
376 * mmget_not_zero will fail in this case.
378 owning_process = get_pid_task(umem_odp->tgid, PIDTYPE_PID);
379 if (!owning_process || !mmget_not_zero(owning_mm)) {
384 range.notifier = &umem_odp->notifier;
385 range.start = ALIGN_DOWN(user_virt, 1UL << page_shift);
386 range.end = ALIGN(user_virt + bcnt, 1UL << page_shift);
387 pfn_start_idx = (range.start - ib_umem_start(umem_odp)) >> PAGE_SHIFT;
388 num_pfns = (range.end - range.start) >> PAGE_SHIFT;
390 range.default_flags = HMM_PFN_REQ_FAULT;
392 if (access_mask & ODP_WRITE_ALLOWED_BIT)
393 range.default_flags |= HMM_PFN_REQ_WRITE;
396 range.hmm_pfns = &(umem_odp->pfn_list[pfn_start_idx]);
397 timeout = jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
400 current_seq = range.notifier_seq =
401 mmu_interval_read_begin(&umem_odp->notifier);
403 mmap_read_lock(owning_mm);
404 ret = hmm_range_fault(&range);
405 mmap_read_unlock(owning_mm);
407 if (ret == -EBUSY && !time_after(jiffies, timeout))
412 start_idx = (range.start - ib_umem_start(umem_odp)) >> page_shift;
413 dma_index = start_idx;
415 mutex_lock(&umem_odp->umem_mutex);
416 if (mmu_interval_read_retry(&umem_odp->notifier, current_seq)) {
417 mutex_unlock(&umem_odp->umem_mutex);
421 for (pfn_index = 0; pfn_index < num_pfns;
422 pfn_index += 1 << (page_shift - PAGE_SHIFT), dma_index++) {
426 * Since we asked for hmm_range_fault() to populate
427 * pages it shouldn't return an error entry on success.
429 WARN_ON(range.hmm_pfns[pfn_index] & HMM_PFN_ERROR);
430 WARN_ON(!(range.hmm_pfns[pfn_index] & HMM_PFN_VALID));
432 if (!(range.hmm_pfns[pfn_index] & HMM_PFN_VALID)) {
433 WARN_ON(umem_odp->dma_list[dma_index]);
436 access_mask = ODP_READ_ALLOWED_BIT;
437 if (range.hmm_pfns[pfn_index] & HMM_PFN_WRITE)
438 access_mask |= ODP_WRITE_ALLOWED_BIT;
441 hmm_order = hmm_pfn_to_map_order(range.hmm_pfns[pfn_index]);
442 /* If a hugepage was detected and ODP wasn't set for, the umem
443 * page_shift will be used, the opposite case is an error.
445 if (hmm_order + PAGE_SHIFT < page_shift) {
447 ibdev_dbg(umem_odp->umem.ibdev,
448 "%s: un-expected hmm_order %d, page_shift %d\n",
449 __func__, hmm_order, page_shift);
453 ret = ib_umem_odp_map_dma_single_page(
454 umem_odp, dma_index, hmm_pfn_to_page(range.hmm_pfns[pfn_index]),
457 ibdev_dbg(umem_odp->umem.ibdev,
458 "ib_umem_odp_map_dma_single_page failed with error %d\n", ret);
462 /* upon sucesss lock should stay on hold for the callee */
464 ret = dma_index - start_idx;
466 mutex_unlock(&umem_odp->umem_mutex);
469 mmput_async(owning_mm);
472 put_task_struct(owning_process);
475 EXPORT_SYMBOL(ib_umem_odp_map_dma_and_lock);
477 void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, u64 virt,
484 struct ib_device *dev = umem_odp->umem.ibdev;
486 lockdep_assert_held(&umem_odp->umem_mutex);
488 virt = max_t(u64, virt, ib_umem_start(umem_odp));
489 bound = min_t(u64, bound, ib_umem_end(umem_odp));
490 for (addr = virt; addr < bound; addr += BIT(umem_odp->page_shift)) {
491 idx = (addr - ib_umem_start(umem_odp)) >> umem_odp->page_shift;
492 dma = umem_odp->dma_list[idx];
494 /* The access flags guaranteed a valid DMA address in case was NULL */
496 unsigned long pfn_idx = (addr - ib_umem_start(umem_odp)) >> PAGE_SHIFT;
497 struct page *page = hmm_pfn_to_page(umem_odp->pfn_list[pfn_idx]);
499 dma_addr = dma & ODP_DMA_ADDR_MASK;
500 ib_dma_unmap_page(dev, dma_addr,
501 BIT(umem_odp->page_shift),
503 if (dma & ODP_WRITE_ALLOWED_BIT) {
504 struct page *head_page = compound_head(page);
506 * set_page_dirty prefers being called with
507 * the page lock. However, MMU notifiers are
508 * called sometimes with and sometimes without
509 * the lock. We rely on the umem_mutex instead
510 * to prevent other mmu notifiers from
511 * continuing and allowing the page mapping to
514 set_page_dirty(head_page);
516 umem_odp->dma_list[idx] = 0;
521 EXPORT_SYMBOL(ib_umem_odp_unmap_dma_pages);