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/slab.h>
41 #include <rdma/ib_umem_odp.h>
46 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
48 struct scatterlist *sg;
53 ib_dma_unmap_sg(dev, umem->sg_head.sgl,
57 for_each_sg(umem->sg_head.sgl, sg, umem->npages, i) {
60 if (umem->writable && dirty)
61 set_page_dirty_lock(page);
65 sg_free_table(&umem->sg_head);
71 * ib_umem_get - Pin and DMA map userspace memory.
73 * If access flags indicate ODP memory, avoid pinning. Instead, stores
74 * the mm for future page fault handling in conjunction with MMU notifiers.
76 * @context: userspace context to pin memory for
77 * @addr: userspace virtual address to start at
78 * @size: length of region to pin
79 * @access: IB_ACCESS_xxx flags for memory being pinned
80 * @dmasync: flush in-flight DMA when the memory region is written
82 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
83 size_t size, int access, int dmasync)
86 struct page **page_list;
87 struct vm_area_struct **vma_list;
89 unsigned long lock_limit;
90 unsigned long cur_base;
94 unsigned long dma_attrs = 0;
95 struct scatterlist *sg, *sg_list_start;
97 unsigned int gup_flags = FOLL_WRITE;
100 dma_attrs |= DMA_ATTR_WRITE_BARRIER;
103 return ERR_PTR(-EINVAL);
106 * If the combination of the addr and size requested for this memory
107 * region causes an integer overflow, return error.
109 if (((addr + size) < addr) ||
110 PAGE_ALIGN(addr + size) < (addr + size))
111 return ERR_PTR(-EINVAL);
114 return ERR_PTR(-EPERM);
116 umem = kzalloc(sizeof *umem, GFP_KERNEL);
118 return ERR_PTR(-ENOMEM);
120 umem->context = context;
122 umem->address = addr;
123 umem->page_size = PAGE_SIZE;
124 umem->pid = get_task_pid(current, PIDTYPE_PID);
125 umem->writable = ib_access_writable(access);
127 if (access & IB_ACCESS_ON_DEMAND) {
129 ret = ib_umem_odp_get(context, umem);
137 umem->odp_data = NULL;
139 /* We assume the memory is from hugetlb until proved otherwise */
142 page_list = (struct page **) __get_free_page(GFP_KERNEL);
146 return ERR_PTR(-ENOMEM);
150 * if we can't alloc the vma_list, it's not so bad;
151 * just assume the memory is not hugetlb memory
153 vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL);
157 npages = ib_umem_num_pages(umem);
159 down_write(¤t->mm->mmap_sem);
161 locked = npages + current->mm->pinned_vm;
162 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
164 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
169 cur_base = addr & PAGE_MASK;
171 if (npages == 0 || npages > UINT_MAX) {
176 ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL);
181 gup_flags |= FOLL_FORCE;
184 sg_list_start = umem->sg_head.sgl;
187 ret = get_user_pages_longterm(cur_base,
188 min_t(unsigned long, npages,
189 PAGE_SIZE / sizeof (struct page *)),
190 gup_flags, page_list, vma_list);
196 cur_base += ret * PAGE_SIZE;
199 for_each_sg(sg_list_start, sg, ret, i) {
200 if (vma_list && !is_vm_hugetlb_page(vma_list[i]))
203 sg_set_page(sg, page_list[i], PAGE_SIZE, 0);
206 /* preparing for next loop */
210 umem->nmap = ib_dma_map_sg_attrs(context->device,
216 if (umem->nmap <= 0) {
226 __ib_umem_release(context->device, umem, 0);
230 current->mm->pinned_vm = locked;
232 up_write(¤t->mm->mmap_sem);
234 free_page((unsigned long) vma_list);
235 free_page((unsigned long) page_list);
237 return ret < 0 ? ERR_PTR(ret) : umem;
239 EXPORT_SYMBOL(ib_umem_get);
241 static void ib_umem_account(struct work_struct *work)
243 struct ib_umem *umem = container_of(work, struct ib_umem, work);
245 down_write(&umem->mm->mmap_sem);
246 umem->mm->pinned_vm -= umem->diff;
247 up_write(&umem->mm->mmap_sem);
253 * ib_umem_release - release memory pinned with ib_umem_get
254 * @umem: umem struct to release
256 void ib_umem_release(struct ib_umem *umem)
258 struct ib_ucontext *context = umem->context;
259 struct mm_struct *mm;
260 struct task_struct *task;
263 if (umem->odp_data) {
264 ib_umem_odp_release(umem);
268 __ib_umem_release(umem->context->device, umem, 1);
270 task = get_pid_task(umem->pid, PIDTYPE_PID);
274 mm = get_task_mm(task);
275 put_task_struct(task);
279 diff = ib_umem_num_pages(umem);
282 * We may be called with the mm's mmap_sem already held. This
283 * can happen when a userspace munmap() is the call that drops
284 * the last reference to our file and calls our release
285 * method. If there are memory regions to destroy, we'll end
286 * up here and not be able to take the mmap_sem. In that case
287 * we defer the vm_locked accounting to the system workqueue.
289 if (context->closing) {
290 if (!down_write_trylock(&mm->mmap_sem)) {
291 INIT_WORK(&umem->work, ib_umem_account);
295 queue_work(ib_wq, &umem->work);
299 down_write(&mm->mmap_sem);
301 mm->pinned_vm -= diff;
302 up_write(&mm->mmap_sem);
307 EXPORT_SYMBOL(ib_umem_release);
309 int ib_umem_page_count(struct ib_umem *umem)
314 struct scatterlist *sg;
317 return ib_umem_num_pages(umem);
319 shift = ilog2(umem->page_size);
322 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i)
323 n += sg_dma_len(sg) >> shift;
327 EXPORT_SYMBOL(ib_umem_page_count);
330 * Copy from the given ib_umem's pages to the given buffer.
332 * umem - the umem to copy from
333 * offset - offset to start copying from
334 * dst - destination buffer
335 * length - buffer length
337 * Returns 0 on success, or an error code.
339 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
342 size_t end = offset + length;
345 if (offset > umem->length || length > umem->length - offset) {
346 pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n",
347 offset, umem->length, end);
351 ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->npages, dst, length,
352 offset + ib_umem_offset(umem));
356 else if (ret != length)
361 EXPORT_SYMBOL(ib_umem_copy_from);