Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland...
[linux-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
35 #include <linux/mm.h>
36 #include <linux/dma-mapping.h>
37 #include <linux/sched.h>
38 #include <linux/hugetlb.h>
39 #include <linux/dma-attrs.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                         struct page *page = sg_page(&chunk->page_list[i]);
58
59                         if (umem->writable && dirty)
60                                 set_page_dirty_lock(page);
61                         put_page(page);
62                 }
63
64                 kfree(chunk);
65         }
66 }
67
68 /**
69  * ib_umem_get - Pin and DMA map userspace memory.
70  * @context: userspace context to pin memory for
71  * @addr: userspace virtual address to start at
72  * @size: length of region to pin
73  * @access: IB_ACCESS_xxx flags for memory being pinned
74  * @dmasync: flush in-flight DMA when the memory region is written
75  */
76 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
77                             size_t size, int access, int dmasync)
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         DEFINE_DMA_ATTRS(attrs);
91
92         if (dmasync)
93                 dma_set_attr(DMA_ATTR_WRITE_BARRIER, &attrs);
94
95         if (!can_do_mlock())
96                 return ERR_PTR(-EPERM);
97
98         umem = kmalloc(sizeof *umem, GFP_KERNEL);
99         if (!umem)
100                 return ERR_PTR(-ENOMEM);
101
102         umem->context   = context;
103         umem->length    = size;
104         umem->offset    = addr & ~PAGE_MASK;
105         umem->page_size = PAGE_SIZE;
106         /*
107          * We ask for writable memory if any access flags other than
108          * "remote read" are set.  "Local write" and "remote write"
109          * obviously require write access.  "Remote atomic" can do
110          * things like fetch and add, which will modify memory, and
111          * "MW bind" can change permissions by binding a window.
112          */
113         umem->writable  = !!(access & ~IB_ACCESS_REMOTE_READ);
114
115         /* We assume the memory is from hugetlb until proved otherwise */
116         umem->hugetlb   = 1;
117
118         INIT_LIST_HEAD(&umem->chunk_list);
119
120         page_list = (struct page **) __get_free_page(GFP_KERNEL);
121         if (!page_list) {
122                 kfree(umem);
123                 return ERR_PTR(-ENOMEM);
124         }
125
126         /*
127          * if we can't alloc the vma_list, it's not so bad;
128          * just assume the memory is not hugetlb memory
129          */
130         vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL);
131         if (!vma_list)
132                 umem->hugetlb = 0;
133
134         npages = PAGE_ALIGN(size + umem->offset) >> PAGE_SHIFT;
135
136         down_write(&current->mm->mmap_sem);
137
138         locked     = npages + current->mm->locked_vm;
139         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
140
141         if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
142                 ret = -ENOMEM;
143                 goto out;
144         }
145
146         cur_base = addr & PAGE_MASK;
147
148         ret = 0;
149         while (npages) {
150                 ret = get_user_pages(current, current->mm, cur_base,
151                                      min_t(unsigned long, npages,
152                                            PAGE_SIZE / sizeof (struct page *)),
153                                      1, !umem->writable, page_list, vma_list);
154
155                 if (ret < 0)
156                         goto out;
157
158                 cur_base += ret * PAGE_SIZE;
159                 npages   -= ret;
160
161                 off = 0;
162
163                 while (ret) {
164                         chunk = kmalloc(sizeof *chunk + sizeof (struct scatterlist) *
165                                         min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK),
166                                         GFP_KERNEL);
167                         if (!chunk) {
168                                 ret = -ENOMEM;
169                                 goto out;
170                         }
171
172                         chunk->nents = min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK);
173                         sg_init_table(chunk->page_list, chunk->nents);
174                         for (i = 0; i < chunk->nents; ++i) {
175                                 if (vma_list &&
176                                     !is_vm_hugetlb_page(vma_list[i + off]))
177                                         umem->hugetlb = 0;
178                                 sg_set_page(&chunk->page_list[i], page_list[i + off], PAGE_SIZE, 0);
179                         }
180
181                         chunk->nmap = ib_dma_map_sg_attrs(context->device,
182                                                           &chunk->page_list[0],
183                                                           chunk->nents,
184                                                           DMA_BIDIRECTIONAL,
185                                                           &attrs);
186                         if (chunk->nmap <= 0) {
187                                 for (i = 0; i < chunk->nents; ++i)
188                                         put_page(sg_page(&chunk->page_list[i]));
189                                 kfree(chunk);
190
191                                 ret = -ENOMEM;
192                                 goto out;
193                         }
194
195                         ret -= chunk->nents;
196                         off += chunk->nents;
197                         list_add_tail(&chunk->list, &umem->chunk_list);
198                 }
199
200                 ret = 0;
201         }
202
203 out:
204         if (ret < 0) {
205                 __ib_umem_release(context->device, umem, 0);
206                 kfree(umem);
207         } else
208                 current->mm->locked_vm = locked;
209
210         up_write(&current->mm->mmap_sem);
211         if (vma_list)
212                 free_page((unsigned long) vma_list);
213         free_page((unsigned long) page_list);
214
215         return ret < 0 ? ERR_PTR(ret) : umem;
216 }
217 EXPORT_SYMBOL(ib_umem_get);
218
219 static void ib_umem_account(struct work_struct *work)
220 {
221         struct ib_umem *umem = container_of(work, struct ib_umem, work);
222
223         down_write(&umem->mm->mmap_sem);
224         umem->mm->locked_vm -= umem->diff;
225         up_write(&umem->mm->mmap_sem);
226         mmput(umem->mm);
227         kfree(umem);
228 }
229
230 /**
231  * ib_umem_release - release memory pinned with ib_umem_get
232  * @umem: umem struct to release
233  */
234 void ib_umem_release(struct ib_umem *umem)
235 {
236         struct ib_ucontext *context = umem->context;
237         struct mm_struct *mm;
238         unsigned long diff;
239
240         __ib_umem_release(umem->context->device, umem, 1);
241
242         mm = get_task_mm(current);
243         if (!mm) {
244                 kfree(umem);
245                 return;
246         }
247
248         diff = PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT;
249
250         /*
251          * We may be called with the mm's mmap_sem already held.  This
252          * can happen when a userspace munmap() is the call that drops
253          * the last reference to our file and calls our release
254          * method.  If there are memory regions to destroy, we'll end
255          * up here and not be able to take the mmap_sem.  In that case
256          * we defer the vm_locked accounting to the system workqueue.
257          */
258         if (context->closing) {
259                 if (!down_write_trylock(&mm->mmap_sem)) {
260                         INIT_WORK(&umem->work, ib_umem_account);
261                         umem->mm   = mm;
262                         umem->diff = diff;
263
264                         schedule_work(&umem->work);
265                         return;
266                 }
267         } else
268                 down_write(&mm->mmap_sem);
269
270         current->mm->locked_vm -= diff;
271         up_write(&mm->mmap_sem);
272         mmput(mm);
273         kfree(umem);
274 }
275 EXPORT_SYMBOL(ib_umem_release);
276
277 int ib_umem_page_count(struct ib_umem *umem)
278 {
279         struct ib_umem_chunk *chunk;
280         int shift;
281         int i;
282         int n;
283
284         shift = ilog2(umem->page_size);
285
286         n = 0;
287         list_for_each_entry(chunk, &umem->chunk_list, list)
288                 for (i = 0; i < chunk->nmap; ++i)
289                         n += sg_dma_len(&chunk->page_list[i]) >> shift;
290
291         return n;
292 }
293 EXPORT_SYMBOL(ib_umem_page_count);