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