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
36 #include <linux/dma-mapping.h>
37 #include <linux/sched.h>
38 #include <linux/export.h>
39 #include <linux/hugetlb.h>
40 #include <linux/dma-attrs.h>
41 #include <linux/slab.h>
42 #include <rdma/ib_umem_odp.h>
47 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
49 struct scatterlist *sg;
54 ib_dma_unmap_sg(dev, umem->sg_head.sgl,
58 for_each_sg(umem->sg_head.sgl, sg, umem->npages, i) {
61 if (umem->writable && dirty)
62 set_page_dirty_lock(page);
66 sg_free_table(&umem->sg_head);
72 * ib_umem_get - Pin and DMA map userspace memory.
74 * If access flags indicate ODP memory, avoid pinning. Instead, stores
75 * the mm for future page fault handling in conjunction with MMU notifiers.
77 * @context: userspace context to pin memory for
78 * @addr: userspace virtual address to start at
79 * @size: length of region to pin
80 * @access: IB_ACCESS_xxx flags for memory being pinned
81 * @dmasync: flush in-flight DMA when the memory region is written
83 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
84 size_t size, int access, int dmasync)
87 struct page **page_list;
88 struct vm_area_struct **vma_list;
90 unsigned long lock_limit;
91 unsigned long cur_base;
95 DEFINE_DMA_ATTRS(attrs);
96 struct scatterlist *sg, *sg_list_start;
100 dma_set_attr(DMA_ATTR_WRITE_BARRIER, &attrs);
103 * If the combination of the addr and size requested for this memory
104 * region causes an integer overflow, return error.
106 if ((PAGE_ALIGN(addr + size) <= size) ||
107 (PAGE_ALIGN(addr + size) <= addr))
108 return ERR_PTR(-EINVAL);
111 return ERR_PTR(-EPERM);
113 umem = kzalloc(sizeof *umem, GFP_KERNEL);
115 return ERR_PTR(-ENOMEM);
117 umem->context = context;
119 umem->address = addr;
120 umem->page_size = PAGE_SIZE;
121 umem->pid = get_task_pid(current, PIDTYPE_PID);
123 * We ask for writable memory if any of the following
124 * access flags are set. "Local write" and "remote write"
125 * obviously require write access. "Remote atomic" can do
126 * things like fetch and add, which will modify memory, and
127 * "MW bind" can change permissions by binding a window.
129 umem->writable = !!(access &
130 (IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE |
131 IB_ACCESS_REMOTE_ATOMIC | IB_ACCESS_MW_BIND));
133 if (access & IB_ACCESS_ON_DEMAND) {
134 ret = ib_umem_odp_get(context, umem);
142 umem->odp_data = NULL;
144 /* We assume the memory is from hugetlb until proved otherwise */
147 page_list = (struct page **) __get_free_page(GFP_KERNEL);
150 return ERR_PTR(-ENOMEM);
154 * if we can't alloc the vma_list, it's not so bad;
155 * just assume the memory is not hugetlb memory
157 vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL);
161 npages = ib_umem_num_pages(umem);
163 down_write(¤t->mm->mmap_sem);
165 locked = npages + current->mm->pinned_vm;
166 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
168 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
173 cur_base = addr & PAGE_MASK;
180 ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL);
185 sg_list_start = umem->sg_head.sgl;
188 ret = get_user_pages(current, current->mm, cur_base,
189 min_t(unsigned long, npages,
190 PAGE_SIZE / sizeof (struct page *)),
191 1, !umem->writable, page_list, vma_list);
197 cur_base += ret * PAGE_SIZE;
200 for_each_sg(sg_list_start, sg, ret, i) {
201 if (vma_list && !is_vm_hugetlb_page(vma_list[i]))
204 sg_set_page(sg, page_list[i], PAGE_SIZE, 0);
207 /* preparing for next loop */
211 umem->nmap = ib_dma_map_sg_attrs(context->device,
217 if (umem->nmap <= 0) {
227 __ib_umem_release(context->device, umem, 0);
231 current->mm->pinned_vm = locked;
233 up_write(¤t->mm->mmap_sem);
235 free_page((unsigned long) vma_list);
236 free_page((unsigned long) page_list);
238 return ret < 0 ? ERR_PTR(ret) : umem;
240 EXPORT_SYMBOL(ib_umem_get);
242 static void ib_umem_account(struct work_struct *work)
244 struct ib_umem *umem = container_of(work, struct ib_umem, work);
246 down_write(&umem->mm->mmap_sem);
247 umem->mm->pinned_vm -= umem->diff;
248 up_write(&umem->mm->mmap_sem);
254 * ib_umem_release - release memory pinned with ib_umem_get
255 * @umem: umem struct to release
257 void ib_umem_release(struct ib_umem *umem)
259 struct ib_ucontext *context = umem->context;
260 struct mm_struct *mm;
261 struct task_struct *task;
264 if (umem->odp_data) {
265 ib_umem_odp_release(umem);
269 __ib_umem_release(umem->context->device, umem, 1);
271 task = get_pid_task(umem->pid, PIDTYPE_PID);
275 mm = get_task_mm(task);
276 put_task_struct(task);
280 diff = ib_umem_num_pages(umem);
283 * We may be called with the mm's mmap_sem already held. This
284 * can happen when a userspace munmap() is the call that drops
285 * the last reference to our file and calls our release
286 * method. If there are memory regions to destroy, we'll end
287 * up here and not be able to take the mmap_sem. In that case
288 * we defer the vm_locked accounting to the system workqueue.
290 if (context->closing) {
291 if (!down_write_trylock(&mm->mmap_sem)) {
292 INIT_WORK(&umem->work, ib_umem_account);
296 queue_work(ib_wq, &umem->work);
300 down_write(&mm->mmap_sem);
302 mm->pinned_vm -= diff;
303 up_write(&mm->mmap_sem);
308 EXPORT_SYMBOL(ib_umem_release);
310 int ib_umem_page_count(struct ib_umem *umem)
315 struct scatterlist *sg;
318 return ib_umem_num_pages(umem);
320 shift = ilog2(umem->page_size);
323 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i)
324 n += sg_dma_len(sg) >> shift;
328 EXPORT_SYMBOL(ib_umem_page_count);
331 * Copy from the given ib_umem's pages to the given buffer.
333 * umem - the umem to copy from
334 * offset - offset to start copying from
335 * dst - destination buffer
336 * length - buffer length
338 * Returns 0 on success, or an error code.
340 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
343 size_t end = offset + length;
346 if (offset > umem->length || length > umem->length - offset) {
347 pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n",
348 offset, umem->length, end);
352 ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->nmap, dst, length,
353 offset + ib_umem_offset(umem));
357 else if (ret != length)
362 EXPORT_SYMBOL(ib_umem_copy_from);