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.
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:
12 * Redistribution and use in source and binary forms, with or
13 * without modification, are permitted provided that the following
16 * - Redistributions of source code must retain the above
17 * copyright notice, this list of conditions and the following
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.
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
34 * $Id: uverbs_mem.c 2743 2005-06-28 22:27:59Z roland $
38 #include <linux/dma-mapping.h>
42 struct ib_umem_account_work {
43 struct work_struct work;
49 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
51 struct ib_umem_chunk *chunk, *tmp;
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);
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
74 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
75 size_t size, int access)
78 struct page **page_list;
79 struct ib_umem_chunk *chunk;
81 unsigned long lock_limit;
82 unsigned long cur_base;
89 return ERR_PTR(-EPERM);
91 umem = kmalloc(sizeof *umem, GFP_KERNEL);
93 return ERR_PTR(-ENOMEM);
95 umem->context = context;
97 umem->offset = addr & ~PAGE_MASK;
98 umem->page_size = PAGE_SIZE;
100 * We ask for writable memory if any access flags other than
101 * "remote read" are set. "Local write" and "remote write"
102 * obviously require write access. "Remote atomic" can do
103 * things like fetch and add, which will modify memory, and
104 * "MW bind" can change permissions by binding a window.
106 umem->writable = !!(access & ~IB_ACCESS_REMOTE_READ);
108 INIT_LIST_HEAD(&umem->chunk_list);
110 page_list = (struct page **) __get_free_page(GFP_KERNEL);
113 return ERR_PTR(-ENOMEM);
116 npages = PAGE_ALIGN(size + umem->offset) >> PAGE_SHIFT;
118 down_write(¤t->mm->mmap_sem);
120 locked = npages + current->mm->locked_vm;
121 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
123 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
128 cur_base = addr & PAGE_MASK;
131 ret = get_user_pages(current, current->mm, cur_base,
133 PAGE_SIZE / sizeof (struct page *)),
134 1, !umem->writable, page_list, NULL);
139 cur_base += ret * PAGE_SIZE;
145 chunk = kmalloc(sizeof *chunk + sizeof (struct scatterlist) *
146 min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK),
153 chunk->nents = min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK);
154 for (i = 0; i < chunk->nents; ++i) {
155 chunk->page_list[i].page = page_list[i + off];
156 chunk->page_list[i].offset = 0;
157 chunk->page_list[i].length = PAGE_SIZE;
160 chunk->nmap = ib_dma_map_sg(context->device,
161 &chunk->page_list[0],
164 if (chunk->nmap <= 0) {
165 for (i = 0; i < chunk->nents; ++i)
166 put_page(chunk->page_list[i].page);
175 list_add_tail(&chunk->list, &umem->chunk_list);
183 __ib_umem_release(context->device, umem, 0);
186 current->mm->locked_vm = locked;
188 up_write(¤t->mm->mmap_sem);
189 free_page((unsigned long) page_list);
191 return ret < 0 ? ERR_PTR(ret) : umem;
193 EXPORT_SYMBOL(ib_umem_get);
195 static void ib_umem_account(struct work_struct *_work)
197 struct ib_umem_account_work *work =
198 container_of(_work, struct ib_umem_account_work, work);
200 down_write(&work->mm->mmap_sem);
201 work->mm->locked_vm -= work->diff;
202 up_write(&work->mm->mmap_sem);
208 * ib_umem_release - release memory pinned with ib_umem_get
209 * @umem: umem struct to release
211 void ib_umem_release(struct ib_umem *umem)
213 struct ib_umem_account_work *work;
214 struct ib_ucontext *context = umem->context;
215 struct mm_struct *mm;
218 __ib_umem_release(umem->context->device, umem, 1);
220 mm = get_task_mm(current);
224 diff = PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT;
228 * We may be called with the mm's mmap_sem already held. This
229 * can happen when a userspace munmap() is the call that drops
230 * the last reference to our file and calls our release
231 * method. If there are memory regions to destroy, we'll end
232 * up here and not be able to take the mmap_sem. In that case
233 * we defer the vm_locked accounting to the system workqueue.
235 if (context->closing && !down_write_trylock(&mm->mmap_sem)) {
236 work = kmalloc(sizeof *work, GFP_KERNEL);
242 INIT_WORK(&work->work, ib_umem_account);
246 schedule_work(&work->work);
249 down_write(&mm->mmap_sem);
251 current->mm->locked_vm -= diff;
252 up_write(&mm->mmap_sem);
255 EXPORT_SYMBOL(ib_umem_release);
257 int ib_umem_page_count(struct ib_umem *umem)
259 struct ib_umem_chunk *chunk;
264 shift = ilog2(umem->page_size);
267 list_for_each_entry(chunk, &umem->chunk_list, list)
268 for (i = 0; i < chunk->nmap; ++i)
269 n += sg_dma_len(&chunk->page_list[i]) >> shift;
273 EXPORT_SYMBOL(ib_umem_page_count);