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 <linux/count_zeros.h>
43 #include <rdma/ib_umem_odp.h>
44
45 #include "uverbs.h"
46
__ib_umem_release(struct ib_device * dev,struct ib_umem * umem,int dirty)47 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
48 {
49 struct sg_page_iter sg_iter;
50 struct page *page;
51
52 if (umem->nmap > 0)
53 ib_dma_unmap_sg(dev, umem->sg_head.sgl, umem->sg_nents,
54 DMA_BIDIRECTIONAL);
55
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);
59 }
60
61 sg_free_table(&umem->sg_head);
62 }
63
64 /**
65 * ib_umem_find_best_pgsz - Find best HW page size to use for this MR
66 *
67 * @umem: umem struct
68 * @pgsz_bitmap: bitmap of HW supported page sizes
69 * @virt: IOVA
70 *
71 * This helper is intended for HW that support multiple page
72 * sizes but can do only a single page size in an MR.
73 *
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.
77 */
ib_umem_find_best_pgsz(struct ib_umem * umem,unsigned long pgsz_bitmap,unsigned long virt)78 unsigned long ib_umem_find_best_pgsz(struct ib_umem *umem,
79 unsigned long pgsz_bitmap,
80 unsigned long virt)
81 {
82 struct scatterlist *sg;
83 unsigned long va, pgoff;
84 dma_addr_t mask;
85 int i;
86
87 /* rdma_for_each_block() has a bug if the page size is smaller than the
88 * page size used to build the umem. For now prevent smaller page sizes
89 * from being returned.
90 */
91 pgsz_bitmap &= GENMASK(BITS_PER_LONG - 1, PAGE_SHIFT);
92
93 /* At minimum, drivers must support PAGE_SIZE or smaller */
94 if (WARN_ON(!(pgsz_bitmap & GENMASK(PAGE_SHIFT, 0))))
95 return 0;
96
97 umem->iova = va = virt;
98 /* The best result is the smallest page size that results in the minimum
99 * number of required pages. Compute the largest page size that could
100 * work based on VA address bits that don't change.
101 */
102 mask = pgsz_bitmap &
103 GENMASK(BITS_PER_LONG - 1,
104 bits_per((umem->length - 1 + virt) ^ virt));
105 /* offset into first SGL */
106 pgoff = umem->address & ~PAGE_MASK;
107
108 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) {
109 /* Walk SGL and reduce max page size if VA/PA bits differ
110 * for any address.
111 */
112 mask |= (sg_dma_address(sg) + pgoff) ^ va;
113 va += sg_dma_len(sg) - pgoff;
114 /* Except for the last entry, the ending iova alignment sets
115 * the maximum possible page size as the low bits of the iova
116 * must be zero when starting the next chunk.
117 */
118 if (i != (umem->nmap - 1))
119 mask |= va;
120 pgoff = 0;
121 }
122
123 /* The mask accumulates 1's in each position where the VA and physical
124 * address differ, thus the length of trailing 0 is the largest page
125 * size that can pass the VA through to the physical.
126 */
127 if (mask)
128 pgsz_bitmap &= GENMASK(count_trailing_zeros(mask), 0);
129 return pgsz_bitmap ? rounddown_pow_of_two(pgsz_bitmap) : 0;
130 }
131 EXPORT_SYMBOL(ib_umem_find_best_pgsz);
132
133 /**
134 * ib_umem_get - Pin and DMA map userspace memory.
135 *
136 * @device: IB device to connect UMEM
137 * @addr: userspace virtual address to start at
138 * @size: length of region to pin
139 * @access: IB_ACCESS_xxx flags for memory being pinned
140 */
ib_umem_get(struct ib_device * device,unsigned long addr,size_t size,int access)141 struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
142 size_t size, int access)
143 {
144 struct ib_umem *umem;
145 struct page **page_list;
146 unsigned long lock_limit;
147 unsigned long new_pinned;
148 unsigned long cur_base;
149 unsigned long dma_attr = 0;
150 struct mm_struct *mm;
151 unsigned long npages;
152 int ret;
153 struct scatterlist *sg = NULL;
154 unsigned int gup_flags = FOLL_WRITE;
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_FORCE;
210
211 while (npages) {
212 cond_resched();
213 ret = pin_user_pages_fast(cur_base,
214 min_t(unsigned long, npages,
215 PAGE_SIZE /
216 sizeof(struct page *)),
217 gup_flags | FOLL_LONGTERM, page_list);
218 if (ret < 0)
219 goto umem_release;
220
221 cur_base += ret * PAGE_SIZE;
222 npages -= ret;
223 sg = __sg_alloc_table_from_pages(&umem->sg_head, page_list, ret,
224 0, ret << PAGE_SHIFT,
225 ib_dma_max_seg_size(device), sg, npages,
226 GFP_KERNEL);
227 umem->sg_nents = umem->sg_head.nents;
228 if (IS_ERR(sg)) {
229 unpin_user_pages_dirty_lock(page_list, ret, 0);
230 ret = PTR_ERR(sg);
231 goto umem_release;
232 }
233 }
234
235 if (access & IB_ACCESS_RELAXED_ORDERING)
236 dma_attr |= DMA_ATTR_WEAK_ORDERING;
237
238 umem->nmap =
239 ib_dma_map_sg_attrs(device, umem->sg_head.sgl, umem->sg_nents,
240 DMA_BIDIRECTIONAL, dma_attr);
241
242 if (!umem->nmap) {
243 ret = -ENOMEM;
244 goto umem_release;
245 }
246
247 ret = 0;
248 goto out;
249
250 umem_release:
251 __ib_umem_release(device, umem, 0);
252 atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm);
253 out:
254 free_page((unsigned long) page_list);
255 umem_kfree:
256 if (ret) {
257 mmdrop(umem->owning_mm);
258 kfree(umem);
259 }
260 return ret ? ERR_PTR(ret) : umem;
261 }
262 EXPORT_SYMBOL(ib_umem_get);
263
264 /**
265 * ib_umem_release - release memory pinned with ib_umem_get
266 * @umem: umem struct to release
267 */
ib_umem_release(struct ib_umem * umem)268 void ib_umem_release(struct ib_umem *umem)
269 {
270 if (!umem)
271 return;
272 if (umem->is_odp)
273 return ib_umem_odp_release(to_ib_umem_odp(umem));
274
275 __ib_umem_release(umem->ibdev, umem, 1);
276
277 atomic64_sub(ib_umem_num_pages(umem), &umem->owning_mm->pinned_vm);
278 mmdrop(umem->owning_mm);
279 kfree(umem);
280 }
281 EXPORT_SYMBOL(ib_umem_release);
282
283 /*
284 * Copy from the given ib_umem's pages to the given buffer.
285 *
286 * umem - the umem to copy from
287 * offset - offset to start copying from
288 * dst - destination buffer
289 * length - buffer length
290 *
291 * Returns 0 on success, or an error code.
292 */
ib_umem_copy_from(void * dst,struct ib_umem * umem,size_t offset,size_t length)293 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
294 size_t length)
295 {
296 size_t end = offset + length;
297 int ret;
298
299 if (offset > umem->length || length > umem->length - offset) {
300 pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n",
301 offset, umem->length, end);
302 return -EINVAL;
303 }
304
305 ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->sg_nents, dst, length,
306 offset + ib_umem_offset(umem));
307
308 if (ret < 0)
309 return ret;
310 else if (ret != length)
311 return -EINVAL;
312 else
313 return 0;
314 }
315 EXPORT_SYMBOL(ib_umem_copy_from);
316