2 * Copyright (c) 2005 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Cisco Systems. All rights reserved.
4 * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
6 * This software is available to you under a choice of one of two
7 * licenses. You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the
10 * OpenIB.org BSD license below:
12 * Redistribution and use in source and binary forms, with or
13 * without modification, are permitted provided that the following
16 * - Redistributions of source code must retain the above
17 * copyright notice, this list of conditions and the following
20 * - Redistributions in binary form must reproduce the above
21 * copyright notice, this list of conditions and the following
22 * disclaimer in the documentation and/or other materials
23 * provided with the distribution.
25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36 #include <linux/dma-mapping.h>
37 #include <linux/sched/signal.h>
38 #include <linux/sched/mm.h>
39 #include <linux/export.h>
40 #include <linux/hugetlb.h>
41 #include <linux/slab.h>
42 #include <rdma/ib_umem_odp.h>
47 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
49 struct scatterlist *sg;
54 ib_dma_unmap_sg(dev, umem->sg_head.sgl,
58 for_each_sg(umem->sg_head.sgl, sg, umem->npages, i) {
61 if (!PageDirty(page) && umem->writable && dirty)
62 set_page_dirty_lock(page);
66 sg_free_table(&umem->sg_head);
70 * ib_umem_get - Pin and DMA map userspace memory.
72 * If access flags indicate ODP memory, avoid pinning. Instead, stores
73 * the mm for future page fault handling in conjunction with MMU notifiers.
75 * @context: userspace context to pin memory for
76 * @addr: userspace virtual address to start at
77 * @size: length of region to pin
78 * @access: IB_ACCESS_xxx flags for memory being pinned
79 * @dmasync: flush in-flight DMA when the memory region is written
81 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
82 size_t size, int access, int dmasync)
85 struct page **page_list;
86 struct vm_area_struct **vma_list;
87 unsigned long lock_limit;
88 unsigned long new_pinned;
89 unsigned long cur_base;
94 unsigned long dma_attrs = 0;
95 struct scatterlist *sg, *sg_list_start;
96 unsigned int gup_flags = FOLL_WRITE;
99 dma_attrs |= DMA_ATTR_WRITE_BARRIER;
102 * If the combination of the addr and size requested for this memory
103 * region causes an integer overflow, return error.
105 if (((addr + size) < addr) ||
106 PAGE_ALIGN(addr + size) < (addr + size))
107 return ERR_PTR(-EINVAL);
110 return ERR_PTR(-EPERM);
112 if (access & IB_ACCESS_ON_DEMAND) {
113 umem = kzalloc(sizeof(struct ib_umem_odp), GFP_KERNEL);
115 return ERR_PTR(-ENOMEM);
118 umem = kzalloc(sizeof(*umem), GFP_KERNEL);
120 return ERR_PTR(-ENOMEM);
123 umem->context = context;
125 umem->address = addr;
126 umem->page_shift = PAGE_SHIFT;
127 umem->writable = ib_access_writable(access);
128 umem->owning_mm = mm = current->mm;
131 if (access & IB_ACCESS_ON_DEMAND) {
132 ret = ib_umem_odp_get(to_ib_umem_odp(umem), access);
138 /* We assume the memory is from hugetlb until proved otherwise */
141 page_list = (struct page **) __get_free_page(GFP_KERNEL);
148 * if we can't alloc the vma_list, it's not so bad;
149 * just assume the memory is not hugetlb memory
151 vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL);
155 npages = ib_umem_num_pages(umem);
156 if (npages == 0 || npages > UINT_MAX) {
161 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
163 down_write(&mm->mmap_sem);
164 if (check_add_overflow(mm->pinned_vm, npages, &new_pinned) ||
165 (new_pinned > lock_limit && !capable(CAP_IPC_LOCK))) {
166 up_write(&mm->mmap_sem);
170 mm->pinned_vm = new_pinned;
171 up_write(&mm->mmap_sem);
173 cur_base = addr & PAGE_MASK;
175 ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL);
180 gup_flags |= FOLL_FORCE;
182 sg_list_start = umem->sg_head.sgl;
185 down_read(&mm->mmap_sem);
186 ret = get_user_pages_longterm(cur_base,
187 min_t(unsigned long, npages,
188 PAGE_SIZE / sizeof (struct page *)),
189 gup_flags, page_list, vma_list);
191 up_read(&mm->mmap_sem);
196 cur_base += ret * PAGE_SIZE;
199 /* Continue to hold the mmap_sem as vma_list access
200 * needs to be protected.
202 for_each_sg(sg_list_start, sg, ret, i) {
203 if (vma_list && !is_vm_hugetlb_page(vma_list[i]))
206 sg_set_page(sg, page_list[i], PAGE_SIZE, 0);
208 up_read(&mm->mmap_sem);
210 /* preparing for next loop */
214 umem->nmap = ib_dma_map_sg_attrs(context->device,
229 __ib_umem_release(context->device, umem, 0);
231 down_write(&mm->mmap_sem);
232 mm->pinned_vm -= ib_umem_num_pages(umem);
233 up_write(&mm->mmap_sem);
236 free_page((unsigned long) vma_list);
237 free_page((unsigned long) page_list);
240 mmdrop(umem->owning_mm);
243 return ret ? ERR_PTR(ret) : umem;
245 EXPORT_SYMBOL(ib_umem_get);
247 static void __ib_umem_release_tail(struct ib_umem *umem)
249 mmdrop(umem->owning_mm);
251 kfree(to_ib_umem_odp(umem));
256 static void ib_umem_release_defer(struct work_struct *work)
258 struct ib_umem *umem = container_of(work, struct ib_umem, work);
260 down_write(&umem->owning_mm->mmap_sem);
261 umem->owning_mm->pinned_vm -= ib_umem_num_pages(umem);
262 up_write(&umem->owning_mm->mmap_sem);
264 __ib_umem_release_tail(umem);
268 * ib_umem_release - release memory pinned with ib_umem_get
269 * @umem: umem struct to release
271 void ib_umem_release(struct ib_umem *umem)
273 struct ib_ucontext *context = umem->context;
276 ib_umem_odp_release(to_ib_umem_odp(umem));
277 __ib_umem_release_tail(umem);
281 __ib_umem_release(umem->context->device, umem, 1);
284 * We may be called with the mm's mmap_sem already held. This
285 * can happen when a userspace munmap() is the call that drops
286 * the last reference to our file and calls our release
287 * method. If there are memory regions to destroy, we'll end
288 * up here and not be able to take the mmap_sem. In that case
289 * we defer the vm_locked accounting a workqueue.
291 if (context->closing) {
292 if (!down_write_trylock(&umem->owning_mm->mmap_sem)) {
293 INIT_WORK(&umem->work, ib_umem_release_defer);
294 queue_work(ib_wq, &umem->work);
298 down_write(&umem->owning_mm->mmap_sem);
300 umem->owning_mm->pinned_vm -= ib_umem_num_pages(umem);
301 up_write(&umem->owning_mm->mmap_sem);
303 __ib_umem_release_tail(umem);
305 EXPORT_SYMBOL(ib_umem_release);
307 int ib_umem_page_count(struct ib_umem *umem)
311 struct scatterlist *sg;
314 return ib_umem_num_pages(umem);
317 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i)
318 n += sg_dma_len(sg) >> umem->page_shift;
322 EXPORT_SYMBOL(ib_umem_page_count);
325 * Copy from the given ib_umem's pages to the given buffer.
327 * umem - the umem to copy from
328 * offset - offset to start copying from
329 * dst - destination buffer
330 * length - buffer length
332 * Returns 0 on success, or an error code.
334 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
337 size_t end = offset + length;
340 if (offset > umem->length || length > umem->length - offset) {
341 pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n",
342 offset, umem->length, end);
346 ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->npages, dst, length,
347 offset + ib_umem_offset(umem));
351 else if (ret != length)
356 EXPORT_SYMBOL(ib_umem_copy_from);