Merge branches 'omap1-upstream' and 'omap2-upstream' into devel
[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 #include <linux/hugetlb.h>
41
42 #include "uverbs.h"
43
44 #define IB_UMEM_MAX_PAGE_CHUNK                                          \
45         ((PAGE_SIZE - offsetof(struct ib_umem_chunk, page_list)) /      \
46          ((void *) &((struct ib_umem_chunk *) 0)->page_list[1] -        \
47           (void *) &((struct ib_umem_chunk *) 0)->page_list[0]))
48
49 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
50 {
51         struct ib_umem_chunk *chunk, *tmp;
52         int i;
53
54         list_for_each_entry_safe(chunk, tmp, &umem->chunk_list, list) {
55                 ib_dma_unmap_sg(dev, chunk->page_list,
56                                 chunk->nents, DMA_BIDIRECTIONAL);
57                 for (i = 0; i < chunk->nents; ++i) {
58                         if (umem->writable && dirty)
59                                 set_page_dirty_lock(chunk->page_list[i].page);
60                         put_page(chunk->page_list[i].page);
61                 }
62
63                 kfree(chunk);
64         }
65 }
66
67 /**
68  * ib_umem_get - Pin and DMA map userspace memory.
69  * @context: userspace context to pin memory for
70  * @addr: userspace virtual address to start at
71  * @size: length of region to pin
72  * @access: IB_ACCESS_xxx flags for memory being pinned
73  */
74 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
75                             size_t size, int access)
76 {
77         struct ib_umem *umem;
78         struct page **page_list;
79         struct vm_area_struct **vma_list;
80         struct ib_umem_chunk *chunk;
81         unsigned long locked;
82         unsigned long lock_limit;
83         unsigned long cur_base;
84         unsigned long npages;
85         int ret;
86         int off;
87         int i;
88
89         if (!can_do_mlock())
90                 return ERR_PTR(-EPERM);
91
92         umem = kmalloc(sizeof *umem, GFP_KERNEL);
93         if (!umem)
94                 return ERR_PTR(-ENOMEM);
95
96         umem->context   = context;
97         umem->length    = size;
98         umem->offset    = addr & ~PAGE_MASK;
99         umem->page_size = PAGE_SIZE;
100         /*
101          * We ask for writable memory if any access flags other than
102          * "remote read" are set.  "Local write" and "remote write"
103          * obviously require write access.  "Remote atomic" can do
104          * things like fetch and add, which will modify memory, and
105          * "MW bind" can change permissions by binding a window.
106          */
107         umem->writable  = !!(access & ~IB_ACCESS_REMOTE_READ);
108
109         /* We assume the memory is from hugetlb until proved otherwise */
110         umem->hugetlb   = 1;
111
112         INIT_LIST_HEAD(&umem->chunk_list);
113
114         page_list = (struct page **) __get_free_page(GFP_KERNEL);
115         if (!page_list) {
116                 kfree(umem);
117                 return ERR_PTR(-ENOMEM);
118         }
119
120         /*
121          * if we can't alloc the vma_list, it's not so bad;
122          * just assume the memory is not hugetlb memory
123          */
124         vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL);
125         if (!vma_list)
126                 umem->hugetlb = 0;
127
128         npages = PAGE_ALIGN(size + umem->offset) >> PAGE_SHIFT;
129
130         down_write(&current->mm->mmap_sem);
131
132         locked     = npages + current->mm->locked_vm;
133         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
134
135         if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
136                 ret = -ENOMEM;
137                 goto out;
138         }
139
140         cur_base = addr & PAGE_MASK;
141
142         ret = 0;
143         while (npages) {
144                 ret = get_user_pages(current, current->mm, cur_base,
145                                      min_t(int, npages,
146                                            PAGE_SIZE / sizeof (struct page *)),
147                                      1, !umem->writable, page_list, vma_list);
148
149                 if (ret < 0)
150                         goto out;
151
152                 cur_base += ret * PAGE_SIZE;
153                 npages   -= ret;
154
155                 off = 0;
156
157                 while (ret) {
158                         chunk = kmalloc(sizeof *chunk + sizeof (struct scatterlist) *
159                                         min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK),
160                                         GFP_KERNEL);
161                         if (!chunk) {
162                                 ret = -ENOMEM;
163                                 goto out;
164                         }
165
166                         chunk->nents = min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK);
167                         for (i = 0; i < chunk->nents; ++i) {
168                                 if (vma_list &&
169                                     !is_vm_hugetlb_page(vma_list[i + off]))
170                                         umem->hugetlb = 0;
171                                 chunk->page_list[i].page   = page_list[i + off];
172                                 chunk->page_list[i].offset = 0;
173                                 chunk->page_list[i].length = PAGE_SIZE;
174                         }
175
176                         chunk->nmap = ib_dma_map_sg(context->device,
177                                                     &chunk->page_list[0],
178                                                     chunk->nents,
179                                                     DMA_BIDIRECTIONAL);
180                         if (chunk->nmap <= 0) {
181                                 for (i = 0; i < chunk->nents; ++i)
182                                         put_page(chunk->page_list[i].page);
183                                 kfree(chunk);
184
185                                 ret = -ENOMEM;
186                                 goto out;
187                         }
188
189                         ret -= chunk->nents;
190                         off += chunk->nents;
191                         list_add_tail(&chunk->list, &umem->chunk_list);
192                 }
193
194                 ret = 0;
195         }
196
197 out:
198         if (ret < 0) {
199                 __ib_umem_release(context->device, umem, 0);
200                 kfree(umem);
201         } else
202                 current->mm->locked_vm = locked;
203
204         up_write(&current->mm->mmap_sem);
205         if (vma_list)
206                 free_page((unsigned long) vma_list);
207         free_page((unsigned long) page_list);
208
209         return ret < 0 ? ERR_PTR(ret) : umem;
210 }
211 EXPORT_SYMBOL(ib_umem_get);
212
213 static void ib_umem_account(struct work_struct *work)
214 {
215         struct ib_umem *umem = container_of(work, struct ib_umem, work);
216
217         down_write(&umem->mm->mmap_sem);
218         umem->mm->locked_vm -= umem->diff;
219         up_write(&umem->mm->mmap_sem);
220         mmput(umem->mm);
221         kfree(umem);
222 }
223
224 /**
225  * ib_umem_release - release memory pinned with ib_umem_get
226  * @umem: umem struct to release
227  */
228 void ib_umem_release(struct ib_umem *umem)
229 {
230         struct ib_ucontext *context = umem->context;
231         struct mm_struct *mm;
232         unsigned long diff;
233
234         __ib_umem_release(umem->context->device, umem, 1);
235
236         mm = get_task_mm(current);
237         if (!mm) {
238                 kfree(umem);
239                 return;
240         }
241
242         diff = PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT;
243
244         /*
245          * We may be called with the mm's mmap_sem already held.  This
246          * can happen when a userspace munmap() is the call that drops
247          * the last reference to our file and calls our release
248          * method.  If there are memory regions to destroy, we'll end
249          * up here and not be able to take the mmap_sem.  In that case
250          * we defer the vm_locked accounting to the system workqueue.
251          */
252         if (context->closing) {
253                 if (!down_write_trylock(&mm->mmap_sem)) {
254                         INIT_WORK(&umem->work, ib_umem_account);
255                         umem->mm   = mm;
256                         umem->diff = diff;
257
258                         schedule_work(&umem->work);
259                         return;
260                 }
261         } else
262                 down_write(&mm->mmap_sem);
263
264         current->mm->locked_vm -= diff;
265         up_write(&mm->mmap_sem);
266         mmput(mm);
267         kfree(umem);
268 }
269 EXPORT_SYMBOL(ib_umem_release);
270
271 int ib_umem_page_count(struct ib_umem *umem)
272 {
273         struct ib_umem_chunk *chunk;
274         int shift;
275         int i;
276         int n;
277
278         shift = ilog2(umem->page_size);
279
280         n = 0;
281         list_for_each_entry(chunk, &umem->chunk_list, list)
282                 for (i = 0; i < chunk->nmap; ++i)
283                         n += sg_dma_len(&chunk->page_list[i]) >> shift;
284
285         return n;
286 }
287 EXPORT_SYMBOL(ib_umem_page_count);