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