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>
39 #include <linux/sched.h>
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]))
48 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
50 struct ib_umem_chunk *chunk, *tmp;
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 if (umem->writable && dirty)
58 set_page_dirty_lock(chunk->page_list[i].page);
59 put_page(chunk->page_list[i].page);
67 * ib_umem_get - Pin and DMA map userspace memory.
68 * @context: userspace context to pin memory for
69 * @addr: userspace virtual address to start at
70 * @size: length of region to pin
71 * @access: IB_ACCESS_xxx flags for memory being pinned
73 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
74 size_t size, int access)
77 struct page **page_list;
78 struct ib_umem_chunk *chunk;
80 unsigned long lock_limit;
81 unsigned long cur_base;
88 return ERR_PTR(-EPERM);
90 umem = kmalloc(sizeof *umem, GFP_KERNEL);
92 return ERR_PTR(-ENOMEM);
94 umem->context = context;
96 umem->offset = addr & ~PAGE_MASK;
97 umem->page_size = PAGE_SIZE;
99 * We ask for writable memory if any access flags other than
100 * "remote read" are set. "Local write" and "remote write"
101 * obviously require write access. "Remote atomic" can do
102 * things like fetch and add, which will modify memory, and
103 * "MW bind" can change permissions by binding a window.
105 umem->writable = !!(access & ~IB_ACCESS_REMOTE_READ);
107 INIT_LIST_HEAD(&umem->chunk_list);
109 page_list = (struct page **) __get_free_page(GFP_KERNEL);
112 return ERR_PTR(-ENOMEM);
115 npages = PAGE_ALIGN(size + umem->offset) >> PAGE_SHIFT;
117 down_write(¤t->mm->mmap_sem);
119 locked = npages + current->mm->locked_vm;
120 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
122 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
127 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 *umem = container_of(work, struct ib_umem, work);
199 down_write(&umem->mm->mmap_sem);
200 umem->mm->locked_vm -= umem->diff;
201 up_write(&umem->mm->mmap_sem);
207 * ib_umem_release - release memory pinned with ib_umem_get
208 * @umem: umem struct to release
210 void ib_umem_release(struct ib_umem *umem)
212 struct ib_ucontext *context = umem->context;
213 struct mm_struct *mm;
216 __ib_umem_release(umem->context->device, umem, 1);
218 mm = get_task_mm(current);
224 diff = PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT;
227 * We may be called with the mm's mmap_sem already held. This
228 * can happen when a userspace munmap() is the call that drops
229 * the last reference to our file and calls our release
230 * method. If there are memory regions to destroy, we'll end
231 * up here and not be able to take the mmap_sem. In that case
232 * we defer the vm_locked accounting to the system workqueue.
234 if (context->closing) {
235 if (!down_write_trylock(&mm->mmap_sem)) {
236 INIT_WORK(&umem->work, ib_umem_account);
240 schedule_work(&umem->work);
244 down_write(&mm->mmap_sem);
246 current->mm->locked_vm -= diff;
247 up_write(&mm->mmap_sem);
251 EXPORT_SYMBOL(ib_umem_release);
253 int ib_umem_page_count(struct ib_umem *umem)
255 struct ib_umem_chunk *chunk;
260 shift = ilog2(umem->page_size);
263 list_for_each_entry(chunk, &umem->chunk_list, list)
264 for (i = 0; i < chunk->nmap; ++i)
265 n += sg_dma_len(&chunk->page_list[i]) >> shift;
269 EXPORT_SYMBOL(ib_umem_page_count);