Loading...
1/*
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.
5 * Copyright (c) 2020 Intel Corporation. All rights reserved.
6 *
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
11 * OpenIB.org BSD license below:
12 *
13 * Redistribution and use in source and binary forms, with or
14 * without modification, are permitted provided that the following
15 * conditions are met:
16 *
17 * - Redistributions of source code must retain the above
18 * copyright notice, this list of conditions and the following
19 * disclaimer.
20 *
21 * - Redistributions in binary form must reproduce the above
22 * copyright notice, this list of conditions and the following
23 * disclaimer in the documentation and/or other materials
24 * provided with the distribution.
25 *
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 * SOFTWARE.
34 */
35
36#include <linux/mm.h>
37#include <linux/dma-mapping.h>
38#include <linux/sched/signal.h>
39#include <linux/sched/mm.h>
40#include <linux/export.h>
41#include <linux/slab.h>
42#include <linux/pagemap.h>
43#include <linux/count_zeros.h>
44#include <rdma/ib_umem_odp.h>
45
46#include "uverbs.h"
47
48static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
49{
50 bool make_dirty = umem->writable && dirty;
51 struct scatterlist *sg;
52 unsigned int i;
53
54 if (dirty)
55 ib_dma_unmap_sgtable_attrs(dev, &umem->sgt_append.sgt,
56 DMA_BIDIRECTIONAL, 0);
57
58 for_each_sgtable_sg(&umem->sgt_append.sgt, sg, i)
59 unpin_user_page_range_dirty_lock(sg_page(sg),
60 DIV_ROUND_UP(sg->length, PAGE_SIZE), make_dirty);
61
62 sg_free_append_table(&umem->sgt_append);
63}
64
65/**
66 * ib_umem_find_best_pgsz - Find best HW page size to use for this MR
67 *
68 * @umem: umem struct
69 * @pgsz_bitmap: bitmap of HW supported page sizes
70 * @virt: IOVA
71 *
72 * This helper is intended for HW that support multiple page
73 * sizes but can do only a single page size in an MR.
74 *
75 * Returns 0 if the umem requires page sizes not supported by
76 * the driver to be mapped. Drivers always supporting PAGE_SIZE
77 * or smaller will never see a 0 result.
78 */
79unsigned long ib_umem_find_best_pgsz(struct ib_umem *umem,
80 unsigned long pgsz_bitmap,
81 unsigned long virt)
82{
83 struct scatterlist *sg;
84 unsigned long va, pgoff;
85 dma_addr_t mask;
86 int i;
87
88 umem->iova = va = virt;
89
90 if (umem->is_odp) {
91 unsigned int page_size = BIT(to_ib_umem_odp(umem)->page_shift);
92
93 /* ODP must always be self consistent. */
94 if (!(pgsz_bitmap & page_size))
95 return 0;
96 return page_size;
97 }
98
99 /* The best result is the smallest page size that results in the minimum
100 * number of required pages. Compute the largest page size that could
101 * work based on VA address bits that don't change.
102 */
103 mask = pgsz_bitmap &
104 GENMASK(BITS_PER_LONG - 1,
105 bits_per((umem->length - 1 + virt) ^ virt));
106 /* offset into first SGL */
107 pgoff = umem->address & ~PAGE_MASK;
108
109 for_each_sgtable_dma_sg(&umem->sgt_append.sgt, sg, i) {
110 /* Walk SGL and reduce max page size if VA/PA bits differ
111 * for any address.
112 */
113 mask |= (sg_dma_address(sg) + pgoff) ^ va;
114 va += sg_dma_len(sg) - pgoff;
115 /* Except for the last entry, the ending iova alignment sets
116 * the maximum possible page size as the low bits of the iova
117 * must be zero when starting the next chunk.
118 */
119 if (i != (umem->sgt_append.sgt.nents - 1))
120 mask |= va;
121 pgoff = 0;
122 }
123
124 /* The mask accumulates 1's in each position where the VA and physical
125 * address differ, thus the length of trailing 0 is the largest page
126 * size that can pass the VA through to the physical.
127 */
128 if (mask)
129 pgsz_bitmap &= GENMASK(count_trailing_zeros(mask), 0);
130 return pgsz_bitmap ? rounddown_pow_of_two(pgsz_bitmap) : 0;
131}
132EXPORT_SYMBOL(ib_umem_find_best_pgsz);
133
134/**
135 * ib_umem_get - Pin and DMA map userspace memory.
136 *
137 * @device: IB device to connect UMEM
138 * @addr: userspace virtual address to start at
139 * @size: length of region to pin
140 * @access: IB_ACCESS_xxx flags for memory being pinned
141 */
142struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
143 size_t size, int access)
144{
145 struct ib_umem *umem;
146 struct page **page_list;
147 unsigned long lock_limit;
148 unsigned long new_pinned;
149 unsigned long cur_base;
150 unsigned long dma_attr = 0;
151 struct mm_struct *mm;
152 unsigned long npages;
153 int pinned, ret;
154 unsigned int gup_flags = FOLL_LONGTERM;
155
156 /*
157 * If the combination of the addr and size requested for this memory
158 * region causes an integer overflow, return error.
159 */
160 if (((addr + size) < addr) ||
161 PAGE_ALIGN(addr + size) < (addr + size))
162 return ERR_PTR(-EINVAL);
163
164 if (!can_do_mlock())
165 return ERR_PTR(-EPERM);
166
167 if (access & IB_ACCESS_ON_DEMAND)
168 return ERR_PTR(-EOPNOTSUPP);
169
170 umem = kzalloc(sizeof(*umem), GFP_KERNEL);
171 if (!umem)
172 return ERR_PTR(-ENOMEM);
173 umem->ibdev = device;
174 umem->length = size;
175 umem->address = addr;
176 /*
177 * Drivers should call ib_umem_find_best_pgsz() to set the iova
178 * correctly.
179 */
180 umem->iova = addr;
181 umem->writable = ib_access_writable(access);
182 umem->owning_mm = mm = current->mm;
183 mmgrab(mm);
184
185 page_list = (struct page **) __get_free_page(GFP_KERNEL);
186 if (!page_list) {
187 ret = -ENOMEM;
188 goto umem_kfree;
189 }
190
191 npages = ib_umem_num_pages(umem);
192 if (npages == 0 || npages > UINT_MAX) {
193 ret = -EINVAL;
194 goto out;
195 }
196
197 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
198
199 new_pinned = atomic64_add_return(npages, &mm->pinned_vm);
200 if (new_pinned > lock_limit && !capable(CAP_IPC_LOCK)) {
201 atomic64_sub(npages, &mm->pinned_vm);
202 ret = -ENOMEM;
203 goto out;
204 }
205
206 cur_base = addr & PAGE_MASK;
207
208 if (umem->writable)
209 gup_flags |= FOLL_WRITE;
210
211 while (npages) {
212 cond_resched();
213 pinned = pin_user_pages_fast(cur_base,
214 min_t(unsigned long, npages,
215 PAGE_SIZE /
216 sizeof(struct page *)),
217 gup_flags, page_list);
218 if (pinned < 0) {
219 ret = pinned;
220 goto umem_release;
221 }
222
223 cur_base += pinned * PAGE_SIZE;
224 npages -= pinned;
225 ret = sg_alloc_append_table_from_pages(
226 &umem->sgt_append, page_list, pinned, 0,
227 pinned << PAGE_SHIFT, ib_dma_max_seg_size(device),
228 npages, GFP_KERNEL);
229 if (ret) {
230 unpin_user_pages_dirty_lock(page_list, pinned, 0);
231 goto umem_release;
232 }
233 }
234
235 if (access & IB_ACCESS_RELAXED_ORDERING)
236 dma_attr |= DMA_ATTR_WEAK_ORDERING;
237
238 ret = ib_dma_map_sgtable_attrs(device, &umem->sgt_append.sgt,
239 DMA_BIDIRECTIONAL, dma_attr);
240 if (ret)
241 goto umem_release;
242 goto out;
243
244umem_release:
245 __ib_umem_release(device, umem, 0);
246 atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm);
247out:
248 free_page((unsigned long) page_list);
249umem_kfree:
250 if (ret) {
251 mmdrop(umem->owning_mm);
252 kfree(umem);
253 }
254 return ret ? ERR_PTR(ret) : umem;
255}
256EXPORT_SYMBOL(ib_umem_get);
257
258/**
259 * ib_umem_release - release memory pinned with ib_umem_get
260 * @umem: umem struct to release
261 */
262void ib_umem_release(struct ib_umem *umem)
263{
264 if (!umem)
265 return;
266 if (umem->is_dmabuf)
267 return ib_umem_dmabuf_release(to_ib_umem_dmabuf(umem));
268 if (umem->is_odp)
269 return ib_umem_odp_release(to_ib_umem_odp(umem));
270
271 __ib_umem_release(umem->ibdev, umem, 1);
272
273 atomic64_sub(ib_umem_num_pages(umem), &umem->owning_mm->pinned_vm);
274 mmdrop(umem->owning_mm);
275 kfree(umem);
276}
277EXPORT_SYMBOL(ib_umem_release);
278
279/*
280 * Copy from the given ib_umem's pages to the given buffer.
281 *
282 * umem - the umem to copy from
283 * offset - offset to start copying from
284 * dst - destination buffer
285 * length - buffer length
286 *
287 * Returns 0 on success, or an error code.
288 */
289int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
290 size_t length)
291{
292 size_t end = offset + length;
293 int ret;
294
295 if (offset > umem->length || length > umem->length - offset) {
296 pr_err("%s not in range. offset: %zd umem length: %zd end: %zd\n",
297 __func__, offset, umem->length, end);
298 return -EINVAL;
299 }
300
301 ret = sg_pcopy_to_buffer(umem->sgt_append.sgt.sgl,
302 umem->sgt_append.sgt.orig_nents, dst, length,
303 offset + ib_umem_offset(umem));
304
305 if (ret < 0)
306 return ret;
307 else if (ret != length)
308 return -EINVAL;
309 else
310 return 0;
311}
312EXPORT_SYMBOL(ib_umem_copy_from);
1/*
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.
5 *
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:
11 *
12 * Redistribution and use in source and binary forms, with or
13 * without modification, are permitted provided that the following
14 * conditions are met:
15 *
16 * - Redistributions of source code must retain the above
17 * copyright notice, this list of conditions and the following
18 * disclaimer.
19 *
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.
24 *
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
32 * SOFTWARE.
33 */
34
35#include <linux/mm.h>
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>
43
44#include "uverbs.h"
45
46static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
47{
48 struct sg_page_iter sg_iter;
49 struct page *page;
50
51 if (umem->nmap > 0)
52 ib_dma_unmap_sg(dev, umem->sg_head.sgl, umem->sg_nents,
53 DMA_BIDIRECTIONAL);
54
55 for_each_sg_page(umem->sg_head.sgl, &sg_iter, umem->sg_nents, 0) {
56 page = sg_page_iter_page(&sg_iter);
57 unpin_user_pages_dirty_lock(&page, 1, umem->writable && dirty);
58 }
59
60 sg_free_table(&umem->sg_head);
61}
62
63/* ib_umem_add_sg_table - Add N contiguous pages to scatter table
64 *
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
70 *
71 * Return new end of scatterlist
72 */
73static struct scatterlist *ib_umem_add_sg_table(struct scatterlist *sg,
74 struct page **page_list,
75 unsigned long npages,
76 unsigned int max_seg_sz,
77 int *nents)
78{
79 unsigned long first_pfn;
80 unsigned long i = 0;
81 bool update_cur_sg = false;
82 bool first = !sg_page(sg);
83
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.
86 */
87 if (!first && (page_to_pfn(sg_page(sg)) + (sg->length >> PAGE_SHIFT) ==
88 page_to_pfn(page_list[0])))
89 update_cur_sg = true;
90
91 while (i != npages) {
92 unsigned long len;
93 struct page *first_page = page_list[i];
94
95 first_pfn = page_to_pfn(first_page);
96
97 /* Compute the number of contiguous pages we have starting
98 * at i
99 */
100 for (len = 0; i != npages &&
101 first_pfn + len == page_to_pfn(page_list[i]) &&
102 len < (max_seg_sz >> PAGE_SHIFT);
103 len++)
104 i++;
105
106 /* Squash N contiguous pages from page_list into current sge */
107 if (update_cur_sg) {
108 if ((max_seg_sz - sg->length) >= (len << PAGE_SHIFT)) {
109 sg_set_page(sg, sg_page(sg),
110 sg->length + (len << PAGE_SHIFT),
111 0);
112 update_cur_sg = false;
113 continue;
114 }
115 update_cur_sg = false;
116 }
117
118 /* Squash N contiguous pages into next sge or first sge */
119 if (!first)
120 sg = sg_next(sg);
121
122 (*nents)++;
123 sg_set_page(sg, first_page, len << PAGE_SHIFT, 0);
124 first = false;
125 }
126
127 return sg;
128}
129
130/**
131 * ib_umem_find_best_pgsz - Find best HW page size to use for this MR
132 *
133 * @umem: umem struct
134 * @pgsz_bitmap: bitmap of HW supported page sizes
135 * @virt: IOVA
136 *
137 * This helper is intended for HW that support multiple page
138 * sizes but can do only a single page size in an MR.
139 *
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.
143 */
144unsigned long ib_umem_find_best_pgsz(struct ib_umem *umem,
145 unsigned long pgsz_bitmap,
146 unsigned long virt)
147{
148 struct scatterlist *sg;
149 unsigned int best_pg_bit;
150 unsigned long va, pgoff;
151 dma_addr_t mask;
152 int i;
153
154 /* At minimum, drivers must support PAGE_SIZE or smaller */
155 if (WARN_ON(!(pgsz_bitmap & GENMASK(PAGE_SHIFT, 0))))
156 return 0;
157
158 va = virt;
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;
163
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
166 * for any address.
167 */
168 mask |= (sg_dma_address(sg) + pgoff) ^ va;
169 va += sg_dma_len(sg) - pgoff;
170 /* Except for the last entry, the ending iova alignment sets
171 * the maximum possible page size as the low bits of the iova
172 * must be zero when starting the next chunk.
173 */
174 if (i != (umem->nmap - 1))
175 mask |= va;
176 pgoff = 0;
177 }
178 best_pg_bit = rdma_find_pg_bit(mask, pgsz_bitmap);
179
180 return BIT_ULL(best_pg_bit);
181}
182EXPORT_SYMBOL(ib_umem_find_best_pgsz);
183
184/**
185 * ib_umem_get - Pin and DMA map userspace memory.
186 *
187 * @device: IB device to connect UMEM
188 * @addr: userspace virtual address to start at
189 * @size: length of region to pin
190 * @access: IB_ACCESS_xxx flags for memory being pinned
191 */
192struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
193 size_t size, int access)
194{
195 struct ib_umem *umem;
196 struct page **page_list;
197 unsigned long lock_limit;
198 unsigned long new_pinned;
199 unsigned long cur_base;
200 unsigned long dma_attr = 0;
201 struct mm_struct *mm;
202 unsigned long npages;
203 int ret;
204 struct scatterlist *sg;
205 unsigned int gup_flags = FOLL_WRITE;
206
207 /*
208 * If the combination of the addr and size requested for this memory
209 * region causes an integer overflow, return error.
210 */
211 if (((addr + size) < addr) ||
212 PAGE_ALIGN(addr + size) < (addr + size))
213 return ERR_PTR(-EINVAL);
214
215 if (!can_do_mlock())
216 return ERR_PTR(-EPERM);
217
218 if (access & IB_ACCESS_ON_DEMAND)
219 return ERR_PTR(-EOPNOTSUPP);
220
221 umem = kzalloc(sizeof(*umem), GFP_KERNEL);
222 if (!umem)
223 return ERR_PTR(-ENOMEM);
224 umem->ibdev = device;
225 umem->length = size;
226 umem->address = addr;
227 umem->writable = ib_access_writable(access);
228 umem->owning_mm = mm = current->mm;
229 mmgrab(mm);
230
231 page_list = (struct page **) __get_free_page(GFP_KERNEL);
232 if (!page_list) {
233 ret = -ENOMEM;
234 goto umem_kfree;
235 }
236
237 npages = ib_umem_num_pages(umem);
238 if (npages == 0 || npages > UINT_MAX) {
239 ret = -EINVAL;
240 goto out;
241 }
242
243 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
244
245 new_pinned = atomic64_add_return(npages, &mm->pinned_vm);
246 if (new_pinned > lock_limit && !capable(CAP_IPC_LOCK)) {
247 atomic64_sub(npages, &mm->pinned_vm);
248 ret = -ENOMEM;
249 goto out;
250 }
251
252 cur_base = addr & PAGE_MASK;
253
254 ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL);
255 if (ret)
256 goto vma;
257
258 if (!umem->writable)
259 gup_flags |= FOLL_FORCE;
260
261 sg = umem->sg_head.sgl;
262
263 while (npages) {
264 cond_resched();
265 ret = pin_user_pages_fast(cur_base,
266 min_t(unsigned long, npages,
267 PAGE_SIZE /
268 sizeof(struct page *)),
269 gup_flags | FOLL_LONGTERM, page_list);
270 if (ret < 0)
271 goto umem_release;
272
273 cur_base += ret * PAGE_SIZE;
274 npages -= ret;
275
276 sg = ib_umem_add_sg_table(sg, page_list, ret,
277 dma_get_max_seg_size(device->dma_device),
278 &umem->sg_nents);
279 }
280
281 sg_mark_end(sg);
282
283 if (access & IB_ACCESS_RELAXED_ORDERING)
284 dma_attr |= DMA_ATTR_WEAK_ORDERING;
285
286 umem->nmap =
287 ib_dma_map_sg_attrs(device, umem->sg_head.sgl, umem->sg_nents,
288 DMA_BIDIRECTIONAL, dma_attr);
289
290 if (!umem->nmap) {
291 ret = -ENOMEM;
292 goto umem_release;
293 }
294
295 ret = 0;
296 goto out;
297
298umem_release:
299 __ib_umem_release(device, umem, 0);
300vma:
301 atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm);
302out:
303 free_page((unsigned long) page_list);
304umem_kfree:
305 if (ret) {
306 mmdrop(umem->owning_mm);
307 kfree(umem);
308 }
309 return ret ? ERR_PTR(ret) : umem;
310}
311EXPORT_SYMBOL(ib_umem_get);
312
313/**
314 * ib_umem_release - release memory pinned with ib_umem_get
315 * @umem: umem struct to release
316 */
317void ib_umem_release(struct ib_umem *umem)
318{
319 if (!umem)
320 return;
321 if (umem->is_odp)
322 return ib_umem_odp_release(to_ib_umem_odp(umem));
323
324 __ib_umem_release(umem->ibdev, umem, 1);
325
326 atomic64_sub(ib_umem_num_pages(umem), &umem->owning_mm->pinned_vm);
327 mmdrop(umem->owning_mm);
328 kfree(umem);
329}
330EXPORT_SYMBOL(ib_umem_release);
331
332int ib_umem_page_count(struct ib_umem *umem)
333{
334 int i, n = 0;
335 struct scatterlist *sg;
336
337 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i)
338 n += sg_dma_len(sg) >> PAGE_SHIFT;
339
340 return n;
341}
342EXPORT_SYMBOL(ib_umem_page_count);
343
344/*
345 * Copy from the given ib_umem's pages to the given buffer.
346 *
347 * umem - the umem to copy from
348 * offset - offset to start copying from
349 * dst - destination buffer
350 * length - buffer length
351 *
352 * Returns 0 on success, or an error code.
353 */
354int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
355 size_t length)
356{
357 size_t end = offset + length;
358 int ret;
359
360 if (offset > umem->length || length > umem->length - offset) {
361 pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n",
362 offset, umem->length, end);
363 return -EINVAL;
364 }
365
366 ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->sg_nents, dst, length,
367 offset + ib_umem_offset(umem));
368
369 if (ret < 0)
370 return ret;
371 else if (ret != length)
372 return -EINVAL;
373 else
374 return 0;
375}
376EXPORT_SYMBOL(ib_umem_copy_from);