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/slab.h>
41 #include <linux/pagemap.h>
42 #include <rdma/ib_umem_odp.h>
46 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
48 struct sg_page_iter sg_iter;
52 ib_dma_unmap_sg(dev, umem->sg_head.sgl, umem->sg_nents,
55 for_each_sg_page(umem->sg_head.sgl, &sg_iter, umem->sg_nents, 0) {
56 page = sg_page_iter_page(&sg_iter);
57 put_user_pages_dirty_lock(&page, 1, umem->writable && dirty);
60 sg_free_table(&umem->sg_head);
63 /* ib_umem_add_sg_table - Add N contiguous pages to scatter table
65 * sg: current scatterlist entry
66 * page_list: array of npage struct page pointers
67 * npages: number of pages in page_list
68 * max_seg_sz: maximum segment size in bytes
69 * nents: [out] number of entries in the scatterlist
71 * Return new end of scatterlist
73 static struct scatterlist *ib_umem_add_sg_table(struct scatterlist *sg,
74 struct page **page_list,
76 unsigned int max_seg_sz,
79 unsigned long first_pfn;
81 bool update_cur_sg = false;
82 bool first = !sg_page(sg);
84 /* Check if new page_list is contiguous with end of previous page_list.
85 * sg->length here is a multiple of PAGE_SIZE and sg->offset is 0.
87 if (!first && (page_to_pfn(sg_page(sg)) + (sg->length >> PAGE_SHIFT) ==
88 page_to_pfn(page_list[0])))
93 struct page *first_page = page_list[i];
95 first_pfn = page_to_pfn(first_page);
97 /* Compute the number of contiguous pages we have starting
100 for (len = 0; i != npages &&
101 first_pfn + len == page_to_pfn(page_list[i]) &&
102 len < (max_seg_sz >> PAGE_SHIFT);
106 /* Squash N contiguous pages from page_list into current sge */
108 if ((max_seg_sz - sg->length) >= (len << PAGE_SHIFT)) {
109 sg_set_page(sg, sg_page(sg),
110 sg->length + (len << PAGE_SHIFT),
112 update_cur_sg = false;
115 update_cur_sg = false;
118 /* Squash N contiguous pages into next sge or first sge */
123 sg_set_page(sg, first_page, len << PAGE_SHIFT, 0);
131 * ib_umem_find_best_pgsz - Find best HW page size to use for this MR
134 * @pgsz_bitmap: bitmap of HW supported page sizes
137 * This helper is intended for HW that support multiple page
138 * sizes but can do only a single page size in an MR.
140 * Returns 0 if the umem requires page sizes not supported by
141 * the driver to be mapped. Drivers always supporting PAGE_SIZE
142 * or smaller will never see a 0 result.
144 unsigned long ib_umem_find_best_pgsz(struct ib_umem *umem,
145 unsigned long pgsz_bitmap,
148 struct scatterlist *sg;
149 unsigned int best_pg_bit;
150 unsigned long va, pgoff;
154 /* At minimum, drivers must support PAGE_SIZE or smaller */
155 if (WARN_ON(!(pgsz_bitmap & GENMASK(PAGE_SHIFT, 0))))
159 /* max page size not to exceed MR length */
160 mask = roundup_pow_of_two(umem->length);
161 /* offset into first SGL */
162 pgoff = umem->address & ~PAGE_MASK;
164 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) {
165 /* Walk SGL and reduce max page size if VA/PA bits differ
168 mask |= (sg_dma_address(sg) + pgoff) ^ va;
169 if (i && i != (umem->nmap - 1))
170 /* restrict by length as well for interior SGEs */
171 mask |= sg_dma_len(sg);
172 va += sg_dma_len(sg) - pgoff;
175 best_pg_bit = rdma_find_pg_bit(mask, pgsz_bitmap);
177 return BIT_ULL(best_pg_bit);
179 EXPORT_SYMBOL(ib_umem_find_best_pgsz);
182 * ib_umem_get - Pin and DMA map userspace memory.
184 * @udata: userspace context to pin memory for
185 * @addr: userspace virtual address to start at
186 * @size: length of region to pin
187 * @access: IB_ACCESS_xxx flags for memory being pinned
189 struct ib_umem *ib_umem_get(struct ib_udata *udata, unsigned long addr,
190 size_t size, int access)
192 struct ib_ucontext *context;
193 struct ib_umem *umem;
194 struct page **page_list;
195 unsigned long lock_limit;
196 unsigned long new_pinned;
197 unsigned long cur_base;
198 struct mm_struct *mm;
199 unsigned long npages;
201 struct scatterlist *sg;
202 unsigned int gup_flags = FOLL_WRITE;
205 return ERR_PTR(-EIO);
207 context = container_of(udata, struct uverbs_attr_bundle, driver_udata)
210 return ERR_PTR(-EIO);
213 * If the combination of the addr and size requested for this memory
214 * region causes an integer overflow, return error.
216 if (((addr + size) < addr) ||
217 PAGE_ALIGN(addr + size) < (addr + size))
218 return ERR_PTR(-EINVAL);
221 return ERR_PTR(-EPERM);
223 if (access & IB_ACCESS_ON_DEMAND)
224 return ERR_PTR(-EOPNOTSUPP);
226 umem = kzalloc(sizeof(*umem), GFP_KERNEL);
228 return ERR_PTR(-ENOMEM);
229 umem->ibdev = context->device;
231 umem->address = addr;
232 umem->writable = ib_access_writable(access);
233 umem->owning_mm = mm = current->mm;
236 page_list = (struct page **) __get_free_page(GFP_KERNEL);
242 npages = ib_umem_num_pages(umem);
243 if (npages == 0 || npages > UINT_MAX) {
248 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
250 new_pinned = atomic64_add_return(npages, &mm->pinned_vm);
251 if (new_pinned > lock_limit && !capable(CAP_IPC_LOCK)) {
252 atomic64_sub(npages, &mm->pinned_vm);
257 cur_base = addr & PAGE_MASK;
259 ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL);
264 gup_flags |= FOLL_FORCE;
266 sg = umem->sg_head.sgl;
269 down_read(&mm->mmap_sem);
270 ret = get_user_pages(cur_base,
271 min_t(unsigned long, npages,
272 PAGE_SIZE / sizeof (struct page *)),
273 gup_flags | FOLL_LONGTERM,
276 up_read(&mm->mmap_sem);
280 cur_base += ret * PAGE_SIZE;
283 sg = ib_umem_add_sg_table(sg, page_list, ret,
284 dma_get_max_seg_size(context->device->dma_device),
287 up_read(&mm->mmap_sem);
292 umem->nmap = ib_dma_map_sg(context->device,
306 __ib_umem_release(context->device, umem, 0);
308 atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm);
310 free_page((unsigned long) page_list);
313 mmdrop(umem->owning_mm);
316 return ret ? ERR_PTR(ret) : umem;
318 EXPORT_SYMBOL(ib_umem_get);
321 * ib_umem_release - release memory pinned with ib_umem_get
322 * @umem: umem struct to release
324 void ib_umem_release(struct ib_umem *umem)
329 return ib_umem_odp_release(to_ib_umem_odp(umem));
331 __ib_umem_release(umem->ibdev, umem, 1);
333 atomic64_sub(ib_umem_num_pages(umem), &umem->owning_mm->pinned_vm);
334 mmdrop(umem->owning_mm);
337 EXPORT_SYMBOL(ib_umem_release);
339 int ib_umem_page_count(struct ib_umem *umem)
342 struct scatterlist *sg;
344 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i)
345 n += sg_dma_len(sg) >> PAGE_SHIFT;
349 EXPORT_SYMBOL(ib_umem_page_count);
352 * Copy from the given ib_umem's pages to the given buffer.
354 * umem - the umem to copy from
355 * offset - offset to start copying from
356 * dst - destination buffer
357 * length - buffer length
359 * Returns 0 on success, or an error code.
361 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
364 size_t end = offset + length;
367 if (offset > umem->length || length > umem->length - offset) {
368 pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n",
369 offset, umem->length, end);
373 ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->sg_nents, dst, length,
374 offset + ib_umem_offset(umem));
378 else if (ret != length)
383 EXPORT_SYMBOL(ib_umem_copy_from);