Merge branch 'master' of master.kernel.org:/pub/scm/linux/kernel/git/davem/sparc-2.6
[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                         struct page *page = sg_page(&chunk->page_list[i]);
59
60                         if (umem->writable && dirty)
61                                 set_page_dirty_lock(page);
62                         put_page(page);
63                 }
64
65                 kfree(chunk);
66         }
67 }
68
69 /**
70  * ib_umem_get - Pin and DMA map userspace memory.
71  * @context: userspace context to pin memory for
72  * @addr: userspace virtual address to start at
73  * @size: length of region to pin
74  * @access: IB_ACCESS_xxx flags for memory being pinned
75  */
76 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
77                             size_t size, int access)
78 {
79         struct ib_umem *umem;
80         struct page **page_list;
81         struct vm_area_struct **vma_list;
82         struct ib_umem_chunk *chunk;
83         unsigned long locked;
84         unsigned long lock_limit;
85         unsigned long cur_base;
86         unsigned long npages;
87         int ret;
88         int off;
89         int i;
90
91         if (!can_do_mlock())
92                 return ERR_PTR(-EPERM);
93
94         umem = kmalloc(sizeof *umem, GFP_KERNEL);
95         if (!umem)
96                 return ERR_PTR(-ENOMEM);
97
98         umem->context   = context;
99         umem->length    = size;
100         umem->offset    = addr & ~PAGE_MASK;
101         umem->page_size = PAGE_SIZE;
102         /*
103          * We ask for writable memory if any access flags other than
104          * "remote read" are set.  "Local write" and "remote write"
105          * obviously require write access.  "Remote atomic" can do
106          * things like fetch and add, which will modify memory, and
107          * "MW bind" can change permissions by binding a window.
108          */
109         umem->writable  = !!(access & ~IB_ACCESS_REMOTE_READ);
110
111         /* We assume the memory is from hugetlb until proved otherwise */
112         umem->hugetlb   = 1;
113
114         INIT_LIST_HEAD(&umem->chunk_list);
115
116         page_list = (struct page **) __get_free_page(GFP_KERNEL);
117         if (!page_list) {
118                 kfree(umem);
119                 return ERR_PTR(-ENOMEM);
120         }
121
122         /*
123          * if we can't alloc the vma_list, it's not so bad;
124          * just assume the memory is not hugetlb memory
125          */
126         vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL);
127         if (!vma_list)
128                 umem->hugetlb = 0;
129
130         npages = PAGE_ALIGN(size + umem->offset) >> PAGE_SHIFT;
131
132         down_write(&current->mm->mmap_sem);
133
134         locked     = npages + current->mm->locked_vm;
135         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
136
137         if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
138                 ret = -ENOMEM;
139                 goto out;
140         }
141
142         cur_base = addr & PAGE_MASK;
143
144         ret = 0;
145         while (npages) {
146                 ret = get_user_pages(current, current->mm, cur_base,
147                                      min_t(int, npages,
148                                            PAGE_SIZE / sizeof (struct page *)),
149                                      1, !umem->writable, page_list, vma_list);
150
151                 if (ret < 0)
152                         goto out;
153
154                 cur_base += ret * PAGE_SIZE;
155                 npages   -= ret;
156
157                 off = 0;
158
159                 while (ret) {
160                         chunk = kmalloc(sizeof *chunk + sizeof (struct scatterlist) *
161                                         min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK),
162                                         GFP_KERNEL);
163                         if (!chunk) {
164                                 ret = -ENOMEM;
165                                 goto out;
166                         }
167
168                         chunk->nents = min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK);
169                         sg_init_table(chunk->page_list, chunk->nents);
170                         for (i = 0; i < chunk->nents; ++i) {
171                                 if (vma_list &&
172                                     !is_vm_hugetlb_page(vma_list[i + off]))
173                                         umem->hugetlb = 0;
174                                 sg_set_page(&chunk->page_list[i], page_list[i + off]);
175                                 chunk->page_list[i].offset = 0;
176                                 chunk->page_list[i].length = PAGE_SIZE;
177                         }
178
179                         chunk->nmap = ib_dma_map_sg(context->device,
180                                                     &chunk->page_list[0],
181                                                     chunk->nents,
182                                                     DMA_BIDIRECTIONAL);
183                         if (chunk->nmap <= 0) {
184                                 for (i = 0; i < chunk->nents; ++i)
185                                         put_page(sg_page(&chunk->page_list[i]));
186                                 kfree(chunk);
187
188                                 ret = -ENOMEM;
189                                 goto out;
190                         }
191
192                         ret -= chunk->nents;
193                         off += chunk->nents;
194                         list_add_tail(&chunk->list, &umem->chunk_list);
195                 }
196
197                 ret = 0;
198         }
199
200 out:
201         if (ret < 0) {
202                 __ib_umem_release(context->device, umem, 0);
203                 kfree(umem);
204         } else
205                 current->mm->locked_vm = locked;
206
207         up_write(&current->mm->mmap_sem);
208         if (vma_list)
209                 free_page((unsigned long) vma_list);
210         free_page((unsigned long) page_list);
211
212         return ret < 0 ? ERR_PTR(ret) : umem;
213 }
214 EXPORT_SYMBOL(ib_umem_get);
215
216 static void ib_umem_account(struct work_struct *work)
217 {
218         struct ib_umem *umem = container_of(work, struct ib_umem, work);
219
220         down_write(&umem->mm->mmap_sem);
221         umem->mm->locked_vm -= umem->diff;
222         up_write(&umem->mm->mmap_sem);
223         mmput(umem->mm);
224         kfree(umem);
225 }
226
227 /**
228  * ib_umem_release - release memory pinned with ib_umem_get
229  * @umem: umem struct to release
230  */
231 void ib_umem_release(struct ib_umem *umem)
232 {
233         struct ib_ucontext *context = umem->context;
234         struct mm_struct *mm;
235         unsigned long diff;
236
237         __ib_umem_release(umem->context->device, umem, 1);
238
239         mm = get_task_mm(current);
240         if (!mm) {
241                 kfree(umem);
242                 return;
243         }
244
245         diff = PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT;
246
247         /*
248          * We may be called with the mm's mmap_sem already held.  This
249          * can happen when a userspace munmap() is the call that drops
250          * the last reference to our file and calls our release
251          * method.  If there are memory regions to destroy, we'll end
252          * up here and not be able to take the mmap_sem.  In that case
253          * we defer the vm_locked accounting to the system workqueue.
254          */
255         if (context->closing) {
256                 if (!down_write_trylock(&mm->mmap_sem)) {
257                         INIT_WORK(&umem->work, ib_umem_account);
258                         umem->mm   = mm;
259                         umem->diff = diff;
260
261                         schedule_work(&umem->work);
262                         return;
263                 }
264         } else
265                 down_write(&mm->mmap_sem);
266
267         current->mm->locked_vm -= diff;
268         up_write(&mm->mmap_sem);
269         mmput(mm);
270         kfree(umem);
271 }
272 EXPORT_SYMBOL(ib_umem_release);
273
274 int ib_umem_page_count(struct ib_umem *umem)
275 {
276         struct ib_umem_chunk *chunk;
277         int shift;
278         int i;
279         int n;
280
281         shift = ilog2(umem->page_size);
282
283         n = 0;
284         list_for_each_entry(chunk, &umem->chunk_list, list)
285                 for (i = 0; i < chunk->nmap; ++i)
286                         n += sg_dma_len(&chunk->page_list[i]) >> shift;
287
288         return n;
289 }
290 EXPORT_SYMBOL(ib_umem_page_count);