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 * 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 (umem->nmap > 0)
55 ib_dma_unmap_sg(dev, umem->sg_head.sgl, umem->sg_nents,
56 DMA_BIDIRECTIONAL);
57
58 for_each_sg(umem->sg_head.sgl, sg, umem->sg_nents, 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_table(&umem->sg_head);
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 if (umem->is_odp) {
89 unsigned int page_size = BIT(to_ib_umem_odp(umem)->page_shift);
90
91 /* ODP must always be self consistent. */
92 if (!(pgsz_bitmap & page_size))
93 return 0;
94 return page_size;
95 }
96
97 /* rdma_for_each_block() has a bug if the page size is smaller than the
98 * page size used to build the umem. For now prevent smaller page sizes
99 * from being returned.
100 */
101 pgsz_bitmap &= GENMASK(BITS_PER_LONG - 1, PAGE_SHIFT);
102
103 umem->iova = va = virt;
104 /* The best result is the smallest page size that results in the minimum
105 * number of required pages. Compute the largest page size that could
106 * work based on VA address bits that don't change.
107 */
108 mask = pgsz_bitmap &
109 GENMASK(BITS_PER_LONG - 1,
110 bits_per((umem->length - 1 + virt) ^ virt));
111 /* offset into first SGL */
112 pgoff = umem->address & ~PAGE_MASK;
113
114 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) {
115 /* Walk SGL and reduce max page size if VA/PA bits differ
116 * for any address.
117 */
118 mask |= (sg_dma_address(sg) + pgoff) ^ va;
119 va += sg_dma_len(sg) - pgoff;
120 /* Except for the last entry, the ending iova alignment sets
121 * the maximum possible page size as the low bits of the iova
122 * must be zero when starting the next chunk.
123 */
124 if (i != (umem->nmap - 1))
125 mask |= va;
126 pgoff = 0;
127 }
128
129 /* The mask accumulates 1's in each position where the VA and physical
130 * address differ, thus the length of trailing 0 is the largest page
131 * size that can pass the VA through to the physical.
132 */
133 if (mask)
134 pgsz_bitmap &= GENMASK(count_trailing_zeros(mask), 0);
135 return pgsz_bitmap ? rounddown_pow_of_two(pgsz_bitmap) : 0;
136}
137EXPORT_SYMBOL(ib_umem_find_best_pgsz);
138
139/**
140 * ib_umem_get - Pin and DMA map userspace memory.
141 *
142 * @device: IB device to connect UMEM
143 * @addr: userspace virtual address to start at
144 * @size: length of region to pin
145 * @access: IB_ACCESS_xxx flags for memory being pinned
146 */
147struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
148 size_t size, int access)
149{
150 struct ib_umem *umem;
151 struct page **page_list;
152 unsigned long lock_limit;
153 unsigned long new_pinned;
154 unsigned long cur_base;
155 unsigned long dma_attr = 0;
156 struct mm_struct *mm;
157 unsigned long npages;
158 int ret;
159 struct scatterlist *sg = NULL;
160 unsigned int gup_flags = FOLL_WRITE;
161
162 /*
163 * If the combination of the addr and size requested for this memory
164 * region causes an integer overflow, return error.
165 */
166 if (((addr + size) < addr) ||
167 PAGE_ALIGN(addr + size) < (addr + size))
168 return ERR_PTR(-EINVAL);
169
170 if (!can_do_mlock())
171 return ERR_PTR(-EPERM);
172
173 if (access & IB_ACCESS_ON_DEMAND)
174 return ERR_PTR(-EOPNOTSUPP);
175
176 umem = kzalloc(sizeof(*umem), GFP_KERNEL);
177 if (!umem)
178 return ERR_PTR(-ENOMEM);
179 umem->ibdev = device;
180 umem->length = size;
181 umem->address = addr;
182 /*
183 * Drivers should call ib_umem_find_best_pgsz() to set the iova
184 * correctly.
185 */
186 umem->iova = addr;
187 umem->writable = ib_access_writable(access);
188 umem->owning_mm = mm = current->mm;
189 mmgrab(mm);
190
191 page_list = (struct page **) __get_free_page(GFP_KERNEL);
192 if (!page_list) {
193 ret = -ENOMEM;
194 goto umem_kfree;
195 }
196
197 npages = ib_umem_num_pages(umem);
198 if (npages == 0 || npages > UINT_MAX) {
199 ret = -EINVAL;
200 goto out;
201 }
202
203 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
204
205 new_pinned = atomic64_add_return(npages, &mm->pinned_vm);
206 if (new_pinned > lock_limit && !capable(CAP_IPC_LOCK)) {
207 atomic64_sub(npages, &mm->pinned_vm);
208 ret = -ENOMEM;
209 goto out;
210 }
211
212 cur_base = addr & PAGE_MASK;
213
214 if (!umem->writable)
215 gup_flags |= FOLL_FORCE;
216
217 while (npages) {
218 cond_resched();
219 ret = pin_user_pages_fast(cur_base,
220 min_t(unsigned long, npages,
221 PAGE_SIZE /
222 sizeof(struct page *)),
223 gup_flags | FOLL_LONGTERM, page_list);
224 if (ret < 0)
225 goto umem_release;
226
227 cur_base += ret * PAGE_SIZE;
228 npages -= ret;
229 sg = __sg_alloc_table_from_pages(&umem->sg_head, page_list, ret,
230 0, ret << PAGE_SHIFT,
231 ib_dma_max_seg_size(device), sg, npages,
232 GFP_KERNEL);
233 umem->sg_nents = umem->sg_head.nents;
234 if (IS_ERR(sg)) {
235 unpin_user_pages_dirty_lock(page_list, ret, 0);
236 ret = PTR_ERR(sg);
237 goto umem_release;
238 }
239 }
240
241 if (access & IB_ACCESS_RELAXED_ORDERING)
242 dma_attr |= DMA_ATTR_WEAK_ORDERING;
243
244 umem->nmap =
245 ib_dma_map_sg_attrs(device, umem->sg_head.sgl, umem->sg_nents,
246 DMA_BIDIRECTIONAL, dma_attr);
247
248 if (!umem->nmap) {
249 ret = -ENOMEM;
250 goto umem_release;
251 }
252
253 ret = 0;
254 goto out;
255
256umem_release:
257 __ib_umem_release(device, umem, 0);
258 atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm);
259out:
260 free_page((unsigned long) page_list);
261umem_kfree:
262 if (ret) {
263 mmdrop(umem->owning_mm);
264 kfree(umem);
265 }
266 return ret ? ERR_PTR(ret) : umem;
267}
268EXPORT_SYMBOL(ib_umem_get);
269
270/**
271 * ib_umem_release - release memory pinned with ib_umem_get
272 * @umem: umem struct to release
273 */
274void ib_umem_release(struct ib_umem *umem)
275{
276 if (!umem)
277 return;
278 if (umem->is_dmabuf)
279 return ib_umem_dmabuf_release(to_ib_umem_dmabuf(umem));
280 if (umem->is_odp)
281 return ib_umem_odp_release(to_ib_umem_odp(umem));
282
283 __ib_umem_release(umem->ibdev, umem, 1);
284
285 atomic64_sub(ib_umem_num_pages(umem), &umem->owning_mm->pinned_vm);
286 mmdrop(umem->owning_mm);
287 kfree(umem);
288}
289EXPORT_SYMBOL(ib_umem_release);
290
291/*
292 * Copy from the given ib_umem's pages to the given buffer.
293 *
294 * umem - the umem to copy from
295 * offset - offset to start copying from
296 * dst - destination buffer
297 * length - buffer length
298 *
299 * Returns 0 on success, or an error code.
300 */
301int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
302 size_t length)
303{
304 size_t end = offset + length;
305 int ret;
306
307 if (offset > umem->length || length > umem->length - offset) {
308 pr_err("%s not in range. offset: %zd umem length: %zd end: %zd\n",
309 __func__, offset, umem->length, end);
310 return -EINVAL;
311 }
312
313 ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->sg_nents, dst, length,
314 offset + ib_umem_offset(umem));
315
316 if (ret < 0)
317 return ret;
318 else if (ret != length)
319 return -EINVAL;
320 else
321 return 0;
322}
323EXPORT_SYMBOL(ib_umem_copy_from);