drivers/infiniband/core/uverbs_mem.c

   1 /*
   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.
   5  *
   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:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  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.
  24  *
  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
  32  * SOFTWARE.
  33  *
  34  * $Id: uverbs_mem.c 2743 2005-06-28 22:27:59Z roland $
  35  */
  36
  37 #include <linux/mm.h>
  38 #include <linux/dma-mapping.h>
  39 #include <linux/vs_memory.h>
  40
  41 #include "uverbs.h"
  42
  43 struct ib_umem_account_work {
  44         struct work_struct work;
  45         struct mm_struct  *mm;
  46         unsigned long      diff;
  47 };
  48
  49
  50 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
  51 {
  52         struct ib_umem_chunk *chunk, *tmp;
  53         int i;
  54
  55         list_for_each_entry_safe(chunk, tmp, &umem->chunk_list, list) {
  56                 ib_dma_unmap_sg(dev, chunk->page_list,
  57                                 chunk->nents, DMA_BIDIRECTIONAL);
  58                 for (i = 0; i < chunk->nents; ++i) {
  59                         if (umem->writable && dirty)
  60                                 set_page_dirty_lock(chunk->page_list[i].page);
  61                         put_page(chunk->page_list[i].page);
  62                 }
  63
  64                 kfree(chunk);
  65         }
  66 }
  67
  68 int ib_umem_get(struct ib_device *dev, struct ib_umem *mem,
  69                 void *addr, size_t size, int write)
  70 {
  71         struct page **page_list;
  72         struct ib_umem_chunk *chunk;
  73         unsigned long locked;
  74         unsigned long lock_limit;
  75         unsigned long cur_base;
  76         unsigned long npages;
  77         int ret = 0;
  78         int off;
  79         int i;
  80
  81         if (!can_do_mlock())
  82                 return -EPERM;
  83
  84         page_list = (struct page **) __get_free_page(GFP_KERNEL);
  85         if (!page_list)
  86                 return -ENOMEM;
  87
  88         mem->user_base = (unsigned long) addr;
  89         mem->length    = size;
  90         mem->offset    = (unsigned long) addr & ~PAGE_MASK;
  91         mem->page_size = PAGE_SIZE;
  92         mem->writable  = write;
  93
  94         INIT_LIST_HEAD(&mem->chunk_list);
  95
  96         npages = PAGE_ALIGN(size + mem->offset) >> PAGE_SHIFT;
  97
  98         down_write(&current->mm->mmap_sem);
  99
 100         locked     = npages + current->mm->locked_vm;
 101         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
 102
 103         if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
 104                 ret = -ENOMEM;
 105                 goto out;
 106         }
 107
 108         cur_base = (unsigned long) addr & PAGE_MASK;
 109
 110         while (npages) {
 111                 ret = get_user_pages(current, current->mm, cur_base,
 112                                      min_t(int, npages,
 113                                            PAGE_SIZE / sizeof (struct page *)),
 114                                      1, !write, page_list, NULL);
 115
 116                 if (ret < 0)
 117                         goto out;
 118
 119                 cur_base += ret * PAGE_SIZE;
 120                 npages   -= ret;
 121
 122                 off = 0;
 123
 124                 while (ret) {
 125                         chunk = kmalloc(sizeof *chunk + sizeof (struct scatterlist) *
 126                                         min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK),
 127                                         GFP_KERNEL);
 128                         if (!chunk) {
 129                                 ret = -ENOMEM;
 130                                 goto out;
 131                         }
 132
 133                         chunk->nents = min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK);
 134                         for (i = 0; i < chunk->nents; ++i) {
 135                                 chunk->page_list[i].page   = page_list[i + off];
 136                                 chunk->page_list[i].offset = 0;
 137                                 chunk->page_list[i].length = PAGE_SIZE;
 138                         }
 139
 140                         chunk->nmap = ib_dma_map_sg(dev,
 141                                                     &chunk->page_list[0],
 142                                                     chunk->nents,
 143                                                     DMA_BIDIRECTIONAL);
 144                         if (chunk->nmap <= 0) {
 145                                 for (i = 0; i < chunk->nents; ++i)
 146                                         put_page(chunk->page_list[i].page);
 147                                 kfree(chunk);
 148
 149                                 ret = -ENOMEM;
 150                                 goto out;
 151                         }
 152
 153                         ret -= chunk->nents;
 154                         off += chunk->nents;
 155                         list_add_tail(&chunk->list, &mem->chunk_list);
 156                 }
 157
 158                 ret = 0;
 159         }
 160
 161 out:
 162         if (ret < 0)
 163                 __ib_umem_release(dev, mem, 0);
 164         else
 165                 vx_vmlocked_sub(current->mm, current->mm->locked_vm - locked);
 166
 167         up_write(&current->mm->mmap_sem);
 168         free_page((unsigned long) page_list);
 169
 170         return ret;
 171 }
 172
 173 void ib_umem_release(struct ib_device *dev, struct ib_umem *umem)
 174 {
 175         __ib_umem_release(dev, umem, 1);
 176
 177         down_write(&current->mm->mmap_sem);
 178         vx_vmlocked_sub(current->mm,
 179                 PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT);
 180         up_write(&current->mm->mmap_sem);
 181 }
 182
 183 static void ib_umem_account(struct work_struct *_work)
 184 {
 185         struct ib_umem_account_work *work =
 186                 container_of(_work, struct ib_umem_account_work, work);
 187
 188         down_write(&work->mm->mmap_sem);
 189         vx_vmlocked_sub(work->mm, work->diff);
 190         up_write(&work->mm->mmap_sem);
 191         mmput(work->mm);
 192         kfree(work);
 193 }
 194
 195 void ib_umem_release_on_close(struct ib_device *dev, struct ib_umem *umem)
 196 {
 197         struct ib_umem_account_work *work;
 198         struct mm_struct *mm;
 199
 200         __ib_umem_release(dev, umem, 1);
 201
 202         mm = get_task_mm(current);
 203         if (!mm)
 204                 return;
 205
 206         /*
 207          * We may be called with the mm's mmap_sem already held.  This
 208          * can happen when a userspace munmap() is the call that drops
 209          * the last reference to our file and calls our release
 210          * method.  If there are memory regions to destroy, we'll end
 211          * up here and not be able to take the mmap_sem.  Therefore we
 212          * defer the vm_locked accounting to the system workqueue.
 213          */
 214
 215         work = kmalloc(sizeof *work, GFP_KERNEL);
 216         if (!work) {
 217                 mmput(mm);
 218                 return;
 219         }
 220
 221         INIT_WORK(&work->work, ib_umem_account);
 222         work->mm   = mm;
 223         work->diff = PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT;
 224
 225         schedule_work(&work->work);
 226 }