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 <linux/count_zeros.h>
43 #include <rdma/ib_umem_odp.h>
47 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
49 struct sg_page_iter sg_iter;
53 ib_dma_unmap_sg(dev, umem->sg_head.sgl, umem->sg_nents,
56 for_each_sg_page(umem->sg_head.sgl, &sg_iter, umem->sg_nents, 0) {
57 page = sg_page_iter_page(&sg_iter);
58 unpin_user_pages_dirty_lock(&page, 1, umem->writable && dirty);
61 sg_free_table(&umem->sg_head);
65 * ib_umem_find_best_pgsz - Find best HW page size to use for this MR
68 * @pgsz_bitmap: bitmap of HW supported page sizes
71 * This helper is intended for HW that support multiple page
72 * sizes but can do only a single page size in an MR.
74 * Returns 0 if the umem requires page sizes not supported by
75 * the driver to be mapped. Drivers always supporting PAGE_SIZE
76 * or smaller will never see a 0 result.
78 unsigned long ib_umem_find_best_pgsz(struct ib_umem *umem,
79 unsigned long pgsz_bitmap,
82 struct scatterlist *sg;
83 unsigned long va, pgoff;
88 unsigned int page_size = BIT(to_ib_umem_odp(umem)->page_shift);
90 /* ODP must always be self consistent. */
91 if (!(pgsz_bitmap & page_size))
96 /* rdma_for_each_block() has a bug if the page size is smaller than the
97 * page size used to build the umem. For now prevent smaller page sizes
98 * from being returned.
100 pgsz_bitmap &= GENMASK(BITS_PER_LONG - 1, PAGE_SHIFT);
102 /* At minimum, drivers must support PAGE_SIZE or smaller */
103 if (WARN_ON(!(pgsz_bitmap & GENMASK(PAGE_SHIFT, 0))))
106 umem->iova = va = virt;
107 /* The best result is the smallest page size that results in the minimum
108 * number of required pages. Compute the largest page size that could
109 * work based on VA address bits that don't change.
112 GENMASK(BITS_PER_LONG - 1,
113 bits_per((umem->length - 1 + virt) ^ virt));
114 /* offset into first SGL */
115 pgoff = umem->address & ~PAGE_MASK;
117 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) {
118 /* Walk SGL and reduce max page size if VA/PA bits differ
121 mask |= (sg_dma_address(sg) + pgoff) ^ va;
122 va += sg_dma_len(sg) - pgoff;
123 /* Except for the last entry, the ending iova alignment sets
124 * the maximum possible page size as the low bits of the iova
125 * must be zero when starting the next chunk.
127 if (i != (umem->nmap - 1))
132 /* The mask accumulates 1's in each position where the VA and physical
133 * address differ, thus the length of trailing 0 is the largest page
134 * size that can pass the VA through to the physical.
137 pgsz_bitmap &= GENMASK(count_trailing_zeros(mask), 0);
138 return rounddown_pow_of_two(pgsz_bitmap);
140 EXPORT_SYMBOL(ib_umem_find_best_pgsz);
143 * ib_umem_get - Pin and DMA map userspace memory.
145 * @device: IB device to connect UMEM
146 * @addr: userspace virtual address to start at
147 * @size: length of region to pin
148 * @access: IB_ACCESS_xxx flags for memory being pinned
150 struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
151 size_t size, int access)
153 struct ib_umem *umem;
154 struct page **page_list;
155 unsigned long lock_limit;
156 unsigned long new_pinned;
157 unsigned long cur_base;
158 unsigned long dma_attr = 0;
159 struct mm_struct *mm;
160 unsigned long npages;
162 struct scatterlist *sg = NULL;
163 unsigned int gup_flags = FOLL_WRITE;
166 * If the combination of the addr and size requested for this memory
167 * region causes an integer overflow, return error.
169 if (((addr + size) < addr) ||
170 PAGE_ALIGN(addr + size) < (addr + size))
171 return ERR_PTR(-EINVAL);
174 return ERR_PTR(-EPERM);
176 if (access & IB_ACCESS_ON_DEMAND)
177 return ERR_PTR(-EOPNOTSUPP);
179 umem = kzalloc(sizeof(*umem), GFP_KERNEL);
181 return ERR_PTR(-ENOMEM);
182 umem->ibdev = device;
184 umem->address = addr;
186 * Drivers should call ib_umem_find_best_pgsz() to set the iova
190 umem->writable = ib_access_writable(access);
191 umem->owning_mm = mm = current->mm;
194 page_list = (struct page **) __get_free_page(GFP_KERNEL);
200 npages = ib_umem_num_pages(umem);
201 if (npages == 0 || npages > UINT_MAX) {
206 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
208 new_pinned = atomic64_add_return(npages, &mm->pinned_vm);
209 if (new_pinned > lock_limit && !capable(CAP_IPC_LOCK)) {
210 atomic64_sub(npages, &mm->pinned_vm);
215 cur_base = addr & PAGE_MASK;
218 gup_flags |= FOLL_FORCE;
222 ret = pin_user_pages_fast(cur_base,
223 min_t(unsigned long, npages,
225 sizeof(struct page *)),
226 gup_flags | FOLL_LONGTERM, page_list);
230 cur_base += ret * PAGE_SIZE;
232 sg = __sg_alloc_table_from_pages(&umem->sg_head, page_list, ret,
233 0, ret << PAGE_SHIFT,
234 ib_dma_max_seg_size(device), sg, npages,
236 umem->sg_nents = umem->sg_head.nents;
238 unpin_user_pages_dirty_lock(page_list, ret, 0);
244 if (access & IB_ACCESS_RELAXED_ORDERING)
245 dma_attr |= DMA_ATTR_WEAK_ORDERING;
248 ib_dma_map_sg_attrs(device, umem->sg_head.sgl, umem->sg_nents,
249 DMA_BIDIRECTIONAL, dma_attr);
260 __ib_umem_release(device, umem, 0);
261 atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm);
263 free_page((unsigned long) page_list);
266 mmdrop(umem->owning_mm);
269 return ret ? ERR_PTR(ret) : umem;
271 EXPORT_SYMBOL(ib_umem_get);
274 * ib_umem_release - release memory pinned with ib_umem_get
275 * @umem: umem struct to release
277 void ib_umem_release(struct ib_umem *umem)
282 return ib_umem_odp_release(to_ib_umem_odp(umem));
284 __ib_umem_release(umem->ibdev, umem, 1);
286 atomic64_sub(ib_umem_num_pages(umem), &umem->owning_mm->pinned_vm);
287 mmdrop(umem->owning_mm);
290 EXPORT_SYMBOL(ib_umem_release);
293 * Copy from the given ib_umem's pages to the given buffer.
295 * umem - the umem to copy from
296 * offset - offset to start copying from
297 * dst - destination buffer
298 * length - buffer length
300 * Returns 0 on success, or an error code.
302 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
305 size_t end = offset + length;
308 if (offset > umem->length || length > umem->length - offset) {
309 pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n",
310 offset, umem->length, end);
314 ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->sg_nents, dst, length,
315 offset + ib_umem_offset(umem));
319 else if (ret != length)
324 EXPORT_SYMBOL(ib_umem_copy_from);