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.h>
38#include <linux/export.h>
39#include <linux/hugetlb.h>
40#include <linux/dma-attrs.h>
41#include <linux/slab.h>
42#include <rdma/ib_umem_odp.h>
43
44#include "uverbs.h"
45
46
47static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
48{
49 struct scatterlist *sg;
50 struct page *page;
51 int i;
52
53 if (umem->nmap > 0)
54 ib_dma_unmap_sg(dev, umem->sg_head.sgl,
55 umem->nmap,
56 DMA_BIDIRECTIONAL);
57
58 for_each_sg(umem->sg_head.sgl, sg, umem->npages, i) {
59
60 page = sg_page(sg);
61 if (umem->writable && dirty)
62 set_page_dirty_lock(page);
63 put_page(page);
64 }
65
66 sg_free_table(&umem->sg_head);
67 return;
68
69}
70
71/**
72 * ib_umem_get - Pin and DMA map userspace memory.
73 *
74 * If access flags indicate ODP memory, avoid pinning. Instead, stores
75 * the mm for future page fault handling in conjunction with MMU notifiers.
76 *
77 * @context: userspace context to pin memory for
78 * @addr: userspace virtual address to start at
79 * @size: length of region to pin
80 * @access: IB_ACCESS_xxx flags for memory being pinned
81 * @dmasync: flush in-flight DMA when the memory region is written
82 */
83struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
84 size_t size, int access, int dmasync)
85{
86 struct ib_umem *umem;
87 struct page **page_list;
88 struct vm_area_struct **vma_list;
89 unsigned long locked;
90 unsigned long lock_limit;
91 unsigned long cur_base;
92 unsigned long npages;
93 int ret;
94 int i;
95 DEFINE_DMA_ATTRS(attrs);
96 struct scatterlist *sg, *sg_list_start;
97 int need_release = 0;
98
99 if (dmasync)
100 dma_set_attr(DMA_ATTR_WRITE_BARRIER, &attrs);
101
102 if (!size)
103 return ERR_PTR(-EINVAL);
104
105 /*
106 * If the combination of the addr and size requested for this memory
107 * region causes an integer overflow, return error.
108 */
109 if (((addr + size) < addr) ||
110 PAGE_ALIGN(addr + size) < (addr + size))
111 return ERR_PTR(-EINVAL);
112
113 if (!can_do_mlock())
114 return ERR_PTR(-EPERM);
115
116 umem = kzalloc(sizeof *umem, GFP_KERNEL);
117 if (!umem)
118 return ERR_PTR(-ENOMEM);
119
120 umem->context = context;
121 umem->length = size;
122 umem->address = addr;
123 umem->page_size = PAGE_SIZE;
124 umem->pid = get_task_pid(current, PIDTYPE_PID);
125 /*
126 * We ask for writable memory if any of the following
127 * access flags are set. "Local write" and "remote write"
128 * obviously require write access. "Remote atomic" can do
129 * things like fetch and add, which will modify memory, and
130 * "MW bind" can change permissions by binding a window.
131 */
132 umem->writable = !!(access &
133 (IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE |
134 IB_ACCESS_REMOTE_ATOMIC | IB_ACCESS_MW_BIND));
135
136 if (access & IB_ACCESS_ON_DEMAND) {
137 ret = ib_umem_odp_get(context, umem);
138 if (ret) {
139 kfree(umem);
140 return ERR_PTR(ret);
141 }
142 return umem;
143 }
144
145 umem->odp_data = NULL;
146
147 /* We assume the memory is from hugetlb until proved otherwise */
148 umem->hugetlb = 1;
149
150 page_list = (struct page **) __get_free_page(GFP_KERNEL);
151 if (!page_list) {
152 kfree(umem);
153 return ERR_PTR(-ENOMEM);
154 }
155
156 /*
157 * if we can't alloc the vma_list, it's not so bad;
158 * just assume the memory is not hugetlb memory
159 */
160 vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL);
161 if (!vma_list)
162 umem->hugetlb = 0;
163
164 npages = ib_umem_num_pages(umem);
165
166 down_write(¤t->mm->mmap_sem);
167
168 locked = npages + current->mm->pinned_vm;
169 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
170
171 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
172 ret = -ENOMEM;
173 goto out;
174 }
175
176 cur_base = addr & PAGE_MASK;
177
178 if (npages == 0) {
179 ret = -EINVAL;
180 goto out;
181 }
182
183 ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL);
184 if (ret)
185 goto out;
186
187 need_release = 1;
188 sg_list_start = umem->sg_head.sgl;
189
190 while (npages) {
191 ret = get_user_pages(cur_base,
192 min_t(unsigned long, npages,
193 PAGE_SIZE / sizeof (struct page *)),
194 1, !umem->writable, page_list, vma_list);
195
196 if (ret < 0)
197 goto out;
198
199 umem->npages += ret;
200 cur_base += ret * PAGE_SIZE;
201 npages -= ret;
202
203 for_each_sg(sg_list_start, sg, ret, i) {
204 if (vma_list && !is_vm_hugetlb_page(vma_list[i]))
205 umem->hugetlb = 0;
206
207 sg_set_page(sg, page_list[i], PAGE_SIZE, 0);
208 }
209
210 /* preparing for next loop */
211 sg_list_start = sg;
212 }
213
214 umem->nmap = ib_dma_map_sg_attrs(context->device,
215 umem->sg_head.sgl,
216 umem->npages,
217 DMA_BIDIRECTIONAL,
218 &attrs);
219
220 if (umem->nmap <= 0) {
221 ret = -ENOMEM;
222 goto out;
223 }
224
225 ret = 0;
226
227out:
228 if (ret < 0) {
229 if (need_release)
230 __ib_umem_release(context->device, umem, 0);
231 put_pid(umem->pid);
232 kfree(umem);
233 } else
234 current->mm->pinned_vm = locked;
235
236 up_write(¤t->mm->mmap_sem);
237 if (vma_list)
238 free_page((unsigned long) vma_list);
239 free_page((unsigned long) page_list);
240
241 return ret < 0 ? ERR_PTR(ret) : umem;
242}
243EXPORT_SYMBOL(ib_umem_get);
244
245static void ib_umem_account(struct work_struct *work)
246{
247 struct ib_umem *umem = container_of(work, struct ib_umem, work);
248
249 down_write(&umem->mm->mmap_sem);
250 umem->mm->pinned_vm -= umem->diff;
251 up_write(&umem->mm->mmap_sem);
252 mmput(umem->mm);
253 kfree(umem);
254}
255
256/**
257 * ib_umem_release - release memory pinned with ib_umem_get
258 * @umem: umem struct to release
259 */
260void ib_umem_release(struct ib_umem *umem)
261{
262 struct ib_ucontext *context = umem->context;
263 struct mm_struct *mm;
264 struct task_struct *task;
265 unsigned long diff;
266
267 if (umem->odp_data) {
268 ib_umem_odp_release(umem);
269 return;
270 }
271
272 __ib_umem_release(umem->context->device, umem, 1);
273
274 task = get_pid_task(umem->pid, PIDTYPE_PID);
275 put_pid(umem->pid);
276 if (!task)
277 goto out;
278 mm = get_task_mm(task);
279 put_task_struct(task);
280 if (!mm)
281 goto out;
282
283 diff = ib_umem_num_pages(umem);
284
285 /*
286 * We may be called with the mm's mmap_sem already held. This
287 * can happen when a userspace munmap() is the call that drops
288 * the last reference to our file and calls our release
289 * method. If there are memory regions to destroy, we'll end
290 * up here and not be able to take the mmap_sem. In that case
291 * we defer the vm_locked accounting to the system workqueue.
292 */
293 if (context->closing) {
294 if (!down_write_trylock(&mm->mmap_sem)) {
295 INIT_WORK(&umem->work, ib_umem_account);
296 umem->mm = mm;
297 umem->diff = diff;
298
299 queue_work(ib_wq, &umem->work);
300 return;
301 }
302 } else
303 down_write(&mm->mmap_sem);
304
305 mm->pinned_vm -= diff;
306 up_write(&mm->mmap_sem);
307 mmput(mm);
308out:
309 kfree(umem);
310}
311EXPORT_SYMBOL(ib_umem_release);
312
313int ib_umem_page_count(struct ib_umem *umem)
314{
315 int shift;
316 int i;
317 int n;
318 struct scatterlist *sg;
319
320 if (umem->odp_data)
321 return ib_umem_num_pages(umem);
322
323 shift = ilog2(umem->page_size);
324
325 n = 0;
326 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i)
327 n += sg_dma_len(sg) >> shift;
328
329 return n;
330}
331EXPORT_SYMBOL(ib_umem_page_count);
332
333/*
334 * Copy from the given ib_umem's pages to the given buffer.
335 *
336 * umem - the umem to copy from
337 * offset - offset to start copying from
338 * dst - destination buffer
339 * length - buffer length
340 *
341 * Returns 0 on success, or an error code.
342 */
343int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
344 size_t length)
345{
346 size_t end = offset + length;
347 int ret;
348
349 if (offset > umem->length || length > umem->length - offset) {
350 pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n",
351 offset, umem->length, end);
352 return -EINVAL;
353 }
354
355 ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->nmap, dst, length,
356 offset + ib_umem_offset(umem));
357
358 if (ret < 0)
359 return ret;
360 else if (ret != length)
361 return -EINVAL;
362 else
363 return 0;
364}
365EXPORT_SYMBOL(ib_umem_copy_from);