blob: b645e558876f89e49797c0971b4f55a9799c3522 [file] [log] [blame]
Roland Dreiereb8ffbf2005-07-07 17:57:14 -07001/*
2 * Copyright (c) 2005 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Cisco Systems. All rights reserved.
Roland Dreier2a1d9b72005-08-10 23:03:10 -07004 * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
Roland Dreiereb8ffbf2005-07-07 17:57:14 -07005 *
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.
Roland Dreiereb8ffbf2005-07-07 17:57:14 -070033 */
34
35#include <linux/mm.h>
36#include <linux/dma-mapping.h>
Alexey Dobriyane8edc6e2007-05-21 01:22:52 +040037#include <linux/sched.h>
Joachim Fenkesc8d8bee2007-09-13 18:15:28 +020038#include <linux/hugetlb.h>
Arthur Kepnercb9fbc52008-04-29 01:00:34 -070039#include <linux/dma-attrs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090040#include <linux/slab.h>
Roland Dreiereb8ffbf2005-07-07 17:57:14 -070041
42#include "uverbs.h"
43
Dotan Barak92ddc442007-08-01 13:33:56 +030044#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
Roland Dreiereb8ffbf2005-07-07 17:57:14 -070049static 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) {
Ralph Campbell15271062006-12-12 14:28:30 -080055 ib_dma_unmap_sg(dev, chunk->page_list,
56 chunk->nents, DMA_BIDIRECTIONAL);
Roland Dreiereb8ffbf2005-07-07 17:57:14 -070057 for (i = 0; i < chunk->nents; ++i) {
Jens Axboe45711f12007-10-22 21:19:53 +020058 struct page *page = sg_page(&chunk->page_list[i]);
59
Roland Dreiereb8ffbf2005-07-07 17:57:14 -070060 if (umem->writable && dirty)
Jens Axboe45711f12007-10-22 21:19:53 +020061 set_page_dirty_lock(page);
62 put_page(page);
Roland Dreiereb8ffbf2005-07-07 17:57:14 -070063 }
64
65 kfree(chunk);
66 }
67}
68
Roland Dreierf7c6a7b2007-03-04 16:15:11 -080069/**
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
Arthur Kepnercb9fbc52008-04-29 01:00:34 -070075 * @dmasync: flush in-flight DMA when the memory region is written
Roland Dreierf7c6a7b2007-03-04 16:15:11 -080076 */
77struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
Arthur Kepnercb9fbc52008-04-29 01:00:34 -070078 size_t size, int access, int dmasync)
Roland Dreiereb8ffbf2005-07-07 17:57:14 -070079{
Roland Dreierf7c6a7b2007-03-04 16:15:11 -080080 struct ib_umem *umem;
Roland Dreiereb8ffbf2005-07-07 17:57:14 -070081 struct page **page_list;
Joachim Fenkesc8d8bee2007-09-13 18:15:28 +020082 struct vm_area_struct **vma_list;
Roland Dreiereb8ffbf2005-07-07 17:57:14 -070083 struct ib_umem_chunk *chunk;
84 unsigned long locked;
85 unsigned long lock_limit;
86 unsigned long cur_base;
87 unsigned long npages;
Roland Dreierf7c6a7b2007-03-04 16:15:11 -080088 int ret;
Roland Dreiereb8ffbf2005-07-07 17:57:14 -070089 int off;
90 int i;
Arthur Kepnercb9fbc52008-04-29 01:00:34 -070091 DEFINE_DMA_ATTRS(attrs);
92
93 if (dmasync)
94 dma_set_attr(DMA_ATTR_WRITE_BARRIER, &attrs);
Roland Dreiereb8ffbf2005-07-07 17:57:14 -070095
96 if (!can_do_mlock())
Roland Dreierf7c6a7b2007-03-04 16:15:11 -080097 return ERR_PTR(-EPERM);
98
99 umem = kmalloc(sizeof *umem, GFP_KERNEL);
100 if (!umem)
101 return ERR_PTR(-ENOMEM);
102
103 umem->context = context;
104 umem->length = size;
105 umem->offset = addr & ~PAGE_MASK;
106 umem->page_size = PAGE_SIZE;
107 /*
108 * We ask for writable memory if any access flags other than
109 * "remote read" are set. "Local write" and "remote write"
110 * obviously require write access. "Remote atomic" can do
111 * things like fetch and add, which will modify memory, and
112 * "MW bind" can change permissions by binding a window.
113 */
114 umem->writable = !!(access & ~IB_ACCESS_REMOTE_READ);
115
Joachim Fenkesc8d8bee2007-09-13 18:15:28 +0200116 /* We assume the memory is from hugetlb until proved otherwise */
117 umem->hugetlb = 1;
118
Roland Dreierf7c6a7b2007-03-04 16:15:11 -0800119 INIT_LIST_HEAD(&umem->chunk_list);
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700120
121 page_list = (struct page **) __get_free_page(GFP_KERNEL);
Roland Dreierf7c6a7b2007-03-04 16:15:11 -0800122 if (!page_list) {
123 kfree(umem);
124 return ERR_PTR(-ENOMEM);
125 }
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700126
Joachim Fenkesc8d8bee2007-09-13 18:15:28 +0200127 /*
128 * if we can't alloc the vma_list, it's not so bad;
129 * just assume the memory is not hugetlb memory
130 */
131 vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL);
132 if (!vma_list)
133 umem->hugetlb = 0;
134
Roland Dreierf7c6a7b2007-03-04 16:15:11 -0800135 npages = PAGE_ALIGN(size + umem->offset) >> PAGE_SHIFT;
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700136
137 down_write(&current->mm->mmap_sem);
138
139 locked = npages + current->mm->locked_vm;
Jiri Slabyccbe9f02010-02-11 15:40:48 -0800140 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700141
142 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
143 ret = -ENOMEM;
144 goto out;
145 }
146
Roland Dreierf7c6a7b2007-03-04 16:15:11 -0800147 cur_base = addr & PAGE_MASK;
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700148
Andrew Morton1d3f4b92007-06-08 16:29:43 -0700149 ret = 0;
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700150 while (npages) {
151 ret = get_user_pages(current, current->mm, cur_base,
Roland Dreier8079ffa2008-06-06 21:38:37 -0700152 min_t(unsigned long, npages,
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700153 PAGE_SIZE / sizeof (struct page *)),
Joachim Fenkesc8d8bee2007-09-13 18:15:28 +0200154 1, !umem->writable, page_list, vma_list);
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700155
156 if (ret < 0)
157 goto out;
158
159 cur_base += ret * PAGE_SIZE;
160 npages -= ret;
161
162 off = 0;
163
164 while (ret) {
165 chunk = kmalloc(sizeof *chunk + sizeof (struct scatterlist) *
166 min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK),
167 GFP_KERNEL);
168 if (!chunk) {
169 ret = -ENOMEM;
170 goto out;
171 }
172
173 chunk->nents = min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK);
Jens Axboe45711f12007-10-22 21:19:53 +0200174 sg_init_table(chunk->page_list, chunk->nents);
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700175 for (i = 0; i < chunk->nents; ++i) {
Joachim Fenkesc8d8bee2007-09-13 18:15:28 +0200176 if (vma_list &&
177 !is_vm_hugetlb_page(vma_list[i + off]))
178 umem->hugetlb = 0;
Jens Axboe642f1492007-10-24 11:20:47 +0200179 sg_set_page(&chunk->page_list[i], page_list[i + off], PAGE_SIZE, 0);
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700180 }
181
Arthur Kepnercb9fbc52008-04-29 01:00:34 -0700182 chunk->nmap = ib_dma_map_sg_attrs(context->device,
183 &chunk->page_list[0],
184 chunk->nents,
185 DMA_BIDIRECTIONAL,
186 &attrs);
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700187 if (chunk->nmap <= 0) {
188 for (i = 0; i < chunk->nents; ++i)
Jens Axboe45711f12007-10-22 21:19:53 +0200189 put_page(sg_page(&chunk->page_list[i]));
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700190 kfree(chunk);
191
192 ret = -ENOMEM;
193 goto out;
194 }
195
196 ret -= chunk->nents;
197 off += chunk->nents;
Roland Dreierf7c6a7b2007-03-04 16:15:11 -0800198 list_add_tail(&chunk->list, &umem->chunk_list);
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700199 }
200
201 ret = 0;
202 }
203
204out:
Roland Dreierf7c6a7b2007-03-04 16:15:11 -0800205 if (ret < 0) {
206 __ib_umem_release(context->device, umem, 0);
207 kfree(umem);
208 } else
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700209 current->mm->locked_vm = locked;
210
211 up_write(&current->mm->mmap_sem);
Joachim Fenkesc8d8bee2007-09-13 18:15:28 +0200212 if (vma_list)
213 free_page((unsigned long) vma_list);
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700214 free_page((unsigned long) page_list);
215
Roland Dreierf7c6a7b2007-03-04 16:15:11 -0800216 return ret < 0 ? ERR_PTR(ret) : umem;
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700217}
Roland Dreierf7c6a7b2007-03-04 16:15:11 -0800218EXPORT_SYMBOL(ib_umem_get);
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700219
Roland Dreier1bf66a32007-04-18 20:20:28 -0700220static void ib_umem_account(struct work_struct *work)
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700221{
Roland Dreier1bf66a32007-04-18 20:20:28 -0700222 struct ib_umem *umem = container_of(work, struct ib_umem, work);
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700223
Roland Dreier1bf66a32007-04-18 20:20:28 -0700224 down_write(&umem->mm->mmap_sem);
225 umem->mm->locked_vm -= umem->diff;
226 up_write(&umem->mm->mmap_sem);
227 mmput(umem->mm);
228 kfree(umem);
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700229}
230
Roland Dreierf7c6a7b2007-03-04 16:15:11 -0800231/**
232 * ib_umem_release - release memory pinned with ib_umem_get
233 * @umem: umem struct to release
234 */
235void ib_umem_release(struct ib_umem *umem)
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700236{
Roland Dreierf7c6a7b2007-03-04 16:15:11 -0800237 struct ib_ucontext *context = umem->context;
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700238 struct mm_struct *mm;
Roland Dreierf7c6a7b2007-03-04 16:15:11 -0800239 unsigned long diff;
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700240
Roland Dreierf7c6a7b2007-03-04 16:15:11 -0800241 __ib_umem_release(umem->context->device, umem, 1);
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700242
243 mm = get_task_mm(current);
Eli Cohen7b82cd82007-05-14 11:35:43 +0300244 if (!mm) {
245 kfree(umem);
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700246 return;
Eli Cohen7b82cd82007-05-14 11:35:43 +0300247 }
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700248
Roland Dreierf7c6a7b2007-03-04 16:15:11 -0800249 diff = PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT;
Roland Dreierf7c6a7b2007-03-04 16:15:11 -0800250
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700251 /*
252 * We may be called with the mm's mmap_sem already held. This
253 * can happen when a userspace munmap() is the call that drops
254 * the last reference to our file and calls our release
255 * method. If there are memory regions to destroy, we'll end
Roland Dreierf7c6a7b2007-03-04 16:15:11 -0800256 * up here and not be able to take the mmap_sem. In that case
257 * we defer the vm_locked accounting to the system workqueue.
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700258 */
Roland Dreier24bce502007-06-21 11:05:58 -0700259 if (context->closing) {
260 if (!down_write_trylock(&mm->mmap_sem)) {
261 INIT_WORK(&umem->work, ib_umem_account);
262 umem->mm = mm;
263 umem->diff = diff;
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700264
Tejun Heof0626712010-10-19 15:24:36 +0000265 queue_work(ib_wq, &umem->work);
Roland Dreier24bce502007-06-21 11:05:58 -0700266 return;
267 }
Roland Dreierf7c6a7b2007-03-04 16:15:11 -0800268 } else
269 down_write(&mm->mmap_sem);
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700270
Roland Dreierf7c6a7b2007-03-04 16:15:11 -0800271 current->mm->locked_vm -= diff;
272 up_write(&mm->mmap_sem);
273 mmput(mm);
Roland Dreier1bf66a32007-04-18 20:20:28 -0700274 kfree(umem);
Roland Dreiereb8ffbf2005-07-07 17:57:14 -0700275}
Roland Dreierf7c6a7b2007-03-04 16:15:11 -0800276EXPORT_SYMBOL(ib_umem_release);
277
278int ib_umem_page_count(struct ib_umem *umem)
279{
280 struct ib_umem_chunk *chunk;
281 int shift;
282 int i;
283 int n;
284
285 shift = ilog2(umem->page_size);
286
287 n = 0;
288 list_for_each_entry(chunk, &umem->chunk_list, list)
289 for (i = 0; i < chunk->nmap; ++i)
290 n += sg_dma_len(&chunk->page_list[i]) >> shift;
291
292 return n;
293}
294EXPORT_SYMBOL(ib_umem_page_count);