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/hugetlb.h>
41 #include <linux/slab.h>
42 #include <rdma/ib_umem_odp.h>
43
44 #include "uverbs.h"
45
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 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->npages,
56 DMA_BIDIRECTIONAL);
57
58 for_each_sg(umem->sg_head.sgl, sg, umem->npages, i) {
59
60 page = sg_page(sg);
61 if (!PageDirty(page) && 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 */
ib_umem_get(struct ib_ucontext * context,unsigned long addr,size_t size,int access,int dmasync)83 struct 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 unsigned long dma_attrs = 0;
96 struct scatterlist *sg, *sg_list_start;
97 int need_release = 0;
98 unsigned int gup_flags = FOLL_WRITE;
99
100 if (dmasync)
101 dma_attrs |= DMA_ATTR_WRITE_BARRIER;
102
103 /*
104 * If the combination of the addr and size requested for this memory
105 * region causes an integer overflow, return error.
106 */
107 if (((addr + size) < addr) ||
108 PAGE_ALIGN(addr + size) < (addr + size))
109 return ERR_PTR(-EINVAL);
110
111 if (!can_do_mlock())
112 return ERR_PTR(-EPERM);
113
114 umem = kzalloc(sizeof *umem, GFP_KERNEL);
115 if (!umem)
116 return ERR_PTR(-ENOMEM);
117
118 umem->context = context;
119 umem->length = size;
120 umem->address = addr;
121 umem->page_shift = PAGE_SHIFT;
122 umem->writable = ib_access_writable(access);
123
124 if (access & IB_ACCESS_ON_DEMAND) {
125 ret = ib_umem_odp_get(context, umem, access);
126 if (ret) {
127 kfree(umem);
128 return ERR_PTR(ret);
129 }
130 return umem;
131 }
132
133 umem->odp_data = NULL;
134
135 /* We assume the memory is from hugetlb until proved otherwise */
136 umem->hugetlb = 1;
137
138 page_list = (struct page **) __get_free_page(GFP_KERNEL);
139 if (!page_list) {
140 kfree(umem);
141 return ERR_PTR(-ENOMEM);
142 }
143
144 /*
145 * if we can't alloc the vma_list, it's not so bad;
146 * just assume the memory is not hugetlb memory
147 */
148 vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL);
149 if (!vma_list)
150 umem->hugetlb = 0;
151
152 npages = ib_umem_num_pages(umem);
153
154 down_write(¤t->mm->mmap_sem);
155
156 locked = npages + current->mm->pinned_vm;
157 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
158
159 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
160 ret = -ENOMEM;
161 goto out;
162 }
163
164 cur_base = addr & PAGE_MASK;
165
166 if (npages == 0 || npages > UINT_MAX) {
167 ret = -EINVAL;
168 goto out;
169 }
170
171 ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL);
172 if (ret)
173 goto out;
174
175 if (!umem->writable)
176 gup_flags |= FOLL_FORCE;
177
178 need_release = 1;
179 sg_list_start = umem->sg_head.sgl;
180
181 while (npages) {
182 ret = get_user_pages_longterm(cur_base,
183 min_t(unsigned long, npages,
184 PAGE_SIZE / sizeof (struct page *)),
185 gup_flags, page_list, vma_list);
186
187 if (ret < 0)
188 goto out;
189
190 umem->npages += ret;
191 cur_base += ret * PAGE_SIZE;
192 npages -= ret;
193
194 for_each_sg(sg_list_start, sg, ret, i) {
195 if (vma_list && !is_vm_hugetlb_page(vma_list[i]))
196 umem->hugetlb = 0;
197
198 sg_set_page(sg, page_list[i], PAGE_SIZE, 0);
199 }
200
201 /* preparing for next loop */
202 sg_list_start = sg;
203 }
204
205 umem->nmap = ib_dma_map_sg_attrs(context->device,
206 umem->sg_head.sgl,
207 umem->npages,
208 DMA_BIDIRECTIONAL,
209 dma_attrs);
210
211 if (umem->nmap <= 0) {
212 ret = -ENOMEM;
213 goto out;
214 }
215
216 ret = 0;
217
218 out:
219 if (ret < 0) {
220 if (need_release)
221 __ib_umem_release(context->device, umem, 0);
222 kfree(umem);
223 } else
224 current->mm->pinned_vm = locked;
225
226 up_write(¤t->mm->mmap_sem);
227 if (vma_list)
228 free_page((unsigned long) vma_list);
229 free_page((unsigned long) page_list);
230
231 return ret < 0 ? ERR_PTR(ret) : umem;
232 }
233 EXPORT_SYMBOL(ib_umem_get);
234
ib_umem_account(struct work_struct * work)235 static void ib_umem_account(struct work_struct *work)
236 {
237 struct ib_umem *umem = container_of(work, struct ib_umem, work);
238
239 down_write(&umem->mm->mmap_sem);
240 umem->mm->pinned_vm -= umem->diff;
241 up_write(&umem->mm->mmap_sem);
242 mmput(umem->mm);
243 kfree(umem);
244 }
245
246 /**
247 * ib_umem_release - release memory pinned with ib_umem_get
248 * @umem: umem struct to release
249 */
ib_umem_release(struct ib_umem * umem)250 void ib_umem_release(struct ib_umem *umem)
251 {
252 struct ib_ucontext *context = umem->context;
253 struct mm_struct *mm;
254 struct task_struct *task;
255 unsigned long diff;
256
257 if (umem->odp_data) {
258 ib_umem_odp_release(umem);
259 return;
260 }
261
262 __ib_umem_release(umem->context->device, umem, 1);
263
264 task = get_pid_task(umem->context->tgid, PIDTYPE_PID);
265 if (!task)
266 goto out;
267 mm = get_task_mm(task);
268 put_task_struct(task);
269 if (!mm)
270 goto out;
271
272 diff = ib_umem_num_pages(umem);
273
274 /*
275 * We may be called with the mm's mmap_sem already held. This
276 * can happen when a userspace munmap() is the call that drops
277 * the last reference to our file and calls our release
278 * method. If there are memory regions to destroy, we'll end
279 * up here and not be able to take the mmap_sem. In that case
280 * we defer the vm_locked accounting to the system workqueue.
281 */
282 if (context->closing) {
283 if (!down_write_trylock(&mm->mmap_sem)) {
284 INIT_WORK(&umem->work, ib_umem_account);
285 umem->mm = mm;
286 umem->diff = diff;
287
288 queue_work(ib_wq, &umem->work);
289 return;
290 }
291 } else
292 down_write(&mm->mmap_sem);
293
294 mm->pinned_vm -= diff;
295 up_write(&mm->mmap_sem);
296 mmput(mm);
297 out:
298 kfree(umem);
299 }
300 EXPORT_SYMBOL(ib_umem_release);
301
ib_umem_page_count(struct ib_umem * umem)302 int ib_umem_page_count(struct ib_umem *umem)
303 {
304 int i;
305 int n;
306 struct scatterlist *sg;
307
308 if (umem->odp_data)
309 return ib_umem_num_pages(umem);
310
311 n = 0;
312 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i)
313 n += sg_dma_len(sg) >> umem->page_shift;
314
315 return n;
316 }
317 EXPORT_SYMBOL(ib_umem_page_count);
318
319 /*
320 * Copy from the given ib_umem's pages to the given buffer.
321 *
322 * umem - the umem to copy from
323 * offset - offset to start copying from
324 * dst - destination buffer
325 * length - buffer length
326 *
327 * Returns 0 on success, or an error code.
328 */
ib_umem_copy_from(void * dst,struct ib_umem * umem,size_t offset,size_t length)329 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
330 size_t length)
331 {
332 size_t end = offset + length;
333 int ret;
334
335 if (offset > umem->length || length > umem->length - offset) {
336 pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n",
337 offset, umem->length, end);
338 return -EINVAL;
339 }
340
341 ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->npages, dst, length,
342 offset + ib_umem_offset(umem));
343
344 if (ret < 0)
345 return ret;
346 else if (ret != length)
347 return -EINVAL;
348 else
349 return 0;
350 }
351 EXPORT_SYMBOL(ib_umem_copy_from);
352