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>
40 #include <linux/hugetlb.h>
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]))
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 vm_area_struct **vma_list;
80 struct ib_umem_chunk *chunk;
82 unsigned long lock_limit;
83 unsigned long cur_base;
90 return ERR_PTR(-EPERM);
92 umem = kmalloc(sizeof *umem, GFP_KERNEL);
94 return ERR_PTR(-ENOMEM);
96 umem->context = context;
98 umem->offset = addr & ~PAGE_MASK;
99 umem->page_size = PAGE_SIZE;
101 * We ask for writable memory if any access flags other than
102 * "remote read" are set. "Local write" and "remote write"
103 * obviously require write access. "Remote atomic" can do
104 * things like fetch and add, which will modify memory, and
105 * "MW bind" can change permissions by binding a window.
107 umem->writable = !!(access & ~IB_ACCESS_REMOTE_READ);
109 /* We assume the memory is from hugetlb until proved otherwise */
112 INIT_LIST_HEAD(&umem->chunk_list);
114 page_list = (struct page **) __get_free_page(GFP_KERNEL);
117 return ERR_PTR(-ENOMEM);
121 * if we can't alloc the vma_list, it's not so bad;
122 * just assume the memory is not hugetlb memory
124 vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL);
128 npages = PAGE_ALIGN(size + umem->offset) >> PAGE_SHIFT;
130 down_write(¤t->mm->mmap_sem);
132 locked = npages + current->mm->locked_vm;
133 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
135 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
140 cur_base = addr & PAGE_MASK;
144 ret = get_user_pages(current, current->mm, cur_base,
146 PAGE_SIZE / sizeof (struct page *)),
147 1, !umem->writable, page_list, vma_list);
152 cur_base += ret * PAGE_SIZE;
158 chunk = kmalloc(sizeof *chunk + sizeof (struct scatterlist) *
159 min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK),
166 chunk->nents = min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK);
167 for (i = 0; i < chunk->nents; ++i) {
169 !is_vm_hugetlb_page(vma_list[i + off]))
171 chunk->page_list[i].page = page_list[i + off];
172 chunk->page_list[i].offset = 0;
173 chunk->page_list[i].length = PAGE_SIZE;
176 chunk->nmap = ib_dma_map_sg(context->device,
177 &chunk->page_list[0],
180 if (chunk->nmap <= 0) {
181 for (i = 0; i < chunk->nents; ++i)
182 put_page(chunk->page_list[i].page);
191 list_add_tail(&chunk->list, &umem->chunk_list);
199 __ib_umem_release(context->device, umem, 0);
202 current->mm->locked_vm = locked;
204 up_write(¤t->mm->mmap_sem);
206 free_page((unsigned long) vma_list);
207 free_page((unsigned long) page_list);
209 return ret < 0 ? ERR_PTR(ret) : umem;
211 EXPORT_SYMBOL(ib_umem_get);
213 static void ib_umem_account(struct work_struct *work)
215 struct ib_umem *umem = container_of(work, struct ib_umem, work);
217 down_write(&umem->mm->mmap_sem);
218 umem->mm->locked_vm -= umem->diff;
219 up_write(&umem->mm->mmap_sem);
225 * ib_umem_release - release memory pinned with ib_umem_get
226 * @umem: umem struct to release
228 void ib_umem_release(struct ib_umem *umem)
230 struct ib_ucontext *context = umem->context;
231 struct mm_struct *mm;
234 __ib_umem_release(umem->context->device, umem, 1);
236 mm = get_task_mm(current);
242 diff = PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT;
245 * We may be called with the mm's mmap_sem already held. This
246 * can happen when a userspace munmap() is the call that drops
247 * the last reference to our file and calls our release
248 * method. If there are memory regions to destroy, we'll end
249 * up here and not be able to take the mmap_sem. In that case
250 * we defer the vm_locked accounting to the system workqueue.
252 if (context->closing) {
253 if (!down_write_trylock(&mm->mmap_sem)) {
254 INIT_WORK(&umem->work, ib_umem_account);
258 schedule_work(&umem->work);
262 down_write(&mm->mmap_sem);
264 current->mm->locked_vm -= diff;
265 up_write(&mm->mmap_sem);
269 EXPORT_SYMBOL(ib_umem_release);
271 int ib_umem_page_count(struct ib_umem *umem)
273 struct ib_umem_chunk *chunk;
278 shift = ilog2(umem->page_size);
281 list_for_each_entry(chunk, &umem->chunk_list, list)
282 for (i = 0; i < chunk->nmap; ++i)
283 n += sg_dma_len(&chunk->page_list[i]) >> shift;
287 EXPORT_SYMBOL(ib_umem_page_count);