Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland...
[linux-2.6] / drivers / infiniband / core / umem.c
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  * $Id: uverbs_mem.c 2743 2005-06-28 22:27:59Z roland $
35  */
36
37 #include <linux/mm.h>
38 #include <linux/dma-mapping.h>
39 #include <linux/sched.h>
40
41 #include "uverbs.h"
42
43 #define IB_UMEM_MAX_PAGE_CHUNK                                          \
44         ((PAGE_SIZE - offsetof(struct ib_umem_chunk, page_list)) /      \
45          ((void *) &((struct ib_umem_chunk *) 0)->page_list[1] -        \
46           (void *) &((struct ib_umem_chunk *) 0)->page_list[0]))
47
48 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
49 {
50         struct ib_umem_chunk *chunk, *tmp;
51         int i;
52
53         list_for_each_entry_safe(chunk, tmp, &umem->chunk_list, list) {
54                 ib_dma_unmap_sg(dev, chunk->page_list,
55                                 chunk->nents, DMA_BIDIRECTIONAL);
56                 for (i = 0; i < chunk->nents; ++i) {
57                         if (umem->writable && dirty)
58                                 set_page_dirty_lock(chunk->page_list[i].page);
59                         put_page(chunk->page_list[i].page);
60                 }
61
62                 kfree(chunk);
63         }
64 }
65
66 /**
67  * ib_umem_get - Pin and DMA map userspace memory.
68  * @context: userspace context to pin memory for
69  * @addr: userspace virtual address to start at
70  * @size: length of region to pin
71  * @access: IB_ACCESS_xxx flags for memory being pinned
72  */
73 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
74                             size_t size, int access)
75 {
76         struct ib_umem *umem;
77         struct page **page_list;
78         struct ib_umem_chunk *chunk;
79         unsigned long locked;
80         unsigned long lock_limit;
81         unsigned long cur_base;
82         unsigned long npages;
83         int ret;
84         int off;
85         int i;
86
87         if (!can_do_mlock())
88                 return ERR_PTR(-EPERM);
89
90         umem = kmalloc(sizeof *umem, GFP_KERNEL);
91         if (!umem)
92                 return ERR_PTR(-ENOMEM);
93
94         umem->context   = context;
95         umem->length    = size;
96         umem->offset    = addr & ~PAGE_MASK;
97         umem->page_size = PAGE_SIZE;
98         /*
99          * We ask for writable memory if any access flags other than
100          * "remote read" are set.  "Local write" and "remote write"
101          * obviously require write access.  "Remote atomic" can do
102          * things like fetch and add, which will modify memory, and
103          * "MW bind" can change permissions by binding a window.
104          */
105         umem->writable  = !!(access & ~IB_ACCESS_REMOTE_READ);
106
107         INIT_LIST_HEAD(&umem->chunk_list);
108
109         page_list = (struct page **) __get_free_page(GFP_KERNEL);
110         if (!page_list) {
111                 kfree(umem);
112                 return ERR_PTR(-ENOMEM);
113         }
114
115         npages = PAGE_ALIGN(size + umem->offset) >> PAGE_SHIFT;
116
117         down_write(&current->mm->mmap_sem);
118
119         locked     = npages + current->mm->locked_vm;
120         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
121
122         if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
123                 ret = -ENOMEM;
124                 goto out;
125         }
126
127         cur_base = addr & PAGE_MASK;
128
129         ret = 0;
130         while (npages) {
131                 ret = get_user_pages(current, current->mm, cur_base,
132                                      min_t(int, npages,
133                                            PAGE_SIZE / sizeof (struct page *)),
134                                      1, !umem->writable, page_list, NULL);
135
136                 if (ret < 0)
137                         goto out;
138
139                 cur_base += ret * PAGE_SIZE;
140                 npages   -= ret;
141
142                 off = 0;
143
144                 while (ret) {
145                         chunk = kmalloc(sizeof *chunk + sizeof (struct scatterlist) *
146                                         min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK),
147                                         GFP_KERNEL);
148                         if (!chunk) {
149                                 ret = -ENOMEM;
150                                 goto out;
151                         }
152
153                         chunk->nents = min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK);
154                         for (i = 0; i < chunk->nents; ++i) {
155                                 chunk->page_list[i].page   = page_list[i + off];
156                                 chunk->page_list[i].offset = 0;
157                                 chunk->page_list[i].length = PAGE_SIZE;
158                         }
159
160                         chunk->nmap = ib_dma_map_sg(context->device,
161                                                     &chunk->page_list[0],
162                                                     chunk->nents,
163                                                     DMA_BIDIRECTIONAL);
164                         if (chunk->nmap <= 0) {
165                                 for (i = 0; i < chunk->nents; ++i)
166                                         put_page(chunk->page_list[i].page);
167                                 kfree(chunk);
168
169                                 ret = -ENOMEM;
170                                 goto out;
171                         }
172
173                         ret -= chunk->nents;
174                         off += chunk->nents;
175                         list_add_tail(&chunk->list, &umem->chunk_list);
176                 }
177
178                 ret = 0;
179         }
180
181 out:
182         if (ret < 0) {
183                 __ib_umem_release(context->device, umem, 0);
184                 kfree(umem);
185         } else
186                 current->mm->locked_vm = locked;
187
188         up_write(&current->mm->mmap_sem);
189         free_page((unsigned long) page_list);
190
191         return ret < 0 ? ERR_PTR(ret) : umem;
192 }
193 EXPORT_SYMBOL(ib_umem_get);
194
195 static void ib_umem_account(struct work_struct *work)
196 {
197         struct ib_umem *umem = container_of(work, struct ib_umem, work);
198
199         down_write(&umem->mm->mmap_sem);
200         umem->mm->locked_vm -= umem->diff;
201         up_write(&umem->mm->mmap_sem);
202         mmput(umem->mm);
203         kfree(umem);
204 }
205
206 /**
207  * ib_umem_release - release memory pinned with ib_umem_get
208  * @umem: umem struct to release
209  */
210 void ib_umem_release(struct ib_umem *umem)
211 {
212         struct ib_ucontext *context = umem->context;
213         struct mm_struct *mm;
214         unsigned long diff;
215
216         __ib_umem_release(umem->context->device, umem, 1);
217
218         mm = get_task_mm(current);
219         if (!mm) {
220                 kfree(umem);
221                 return;
222         }
223
224         diff = PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT;
225
226         /*
227          * We may be called with the mm's mmap_sem already held.  This
228          * can happen when a userspace munmap() is the call that drops
229          * the last reference to our file and calls our release
230          * method.  If there are memory regions to destroy, we'll end
231          * up here and not be able to take the mmap_sem.  In that case
232          * we defer the vm_locked accounting to the system workqueue.
233          */
234         if (context->closing) {
235                 if (!down_write_trylock(&mm->mmap_sem)) {
236                         INIT_WORK(&umem->work, ib_umem_account);
237                         umem->mm   = mm;
238                         umem->diff = diff;
239
240                         schedule_work(&umem->work);
241                         return;
242                 }
243         } else
244                 down_write(&mm->mmap_sem);
245
246         current->mm->locked_vm -= diff;
247         up_write(&mm->mmap_sem);
248         mmput(mm);
249         kfree(umem);
250 }
251 EXPORT_SYMBOL(ib_umem_release);
252
253 int ib_umem_page_count(struct ib_umem *umem)
254 {
255         struct ib_umem_chunk *chunk;
256         int shift;
257         int i;
258         int n;
259
260         shift = ilog2(umem->page_size);
261
262         n = 0;
263         list_for_each_entry(chunk, &umem->chunk_list, list)
264                 for (i = 0; i < chunk->nmap; ++i)
265                         n += sg_dma_len(&chunk->page_list[i]) >> shift;
266
267         return n;
268 }
269 EXPORT_SYMBOL(ib_umem_page_count);