drivers/infiniband/core/umem.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
  40 #include "uverbs.h"
  41
  42 struct ib_umem_account_work {
  43         struct work_struct work;
  44         struct mm_struct  *mm;
  45         unsigned long      diff;
  46 };
  47
  48
  49 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
  50 {
  51         struct ib_umem_chunk *chunk, *tmp;
  52         int i;
  53
  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);
  61                 }
  62
  63                 kfree(chunk);
  64         }
  65 }
  66
  67 /**
  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
  73  */
  74 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
  75                             size_t size, int access)
  76 {
  77         struct ib_umem *umem;
  78         struct page **page_list;
  79         struct ib_umem_chunk *chunk;
  80         unsigned long locked;
  81         unsigned long lock_limit;
  82         unsigned long cur_base;
  83         unsigned long npages;
  84         int ret;
  85         int off;
  86         int i;
  87
  88         if (!can_do_mlock())
  89                 return ERR_PTR(-EPERM);
  90
  91         umem = kmalloc(sizeof *umem, GFP_KERNEL);
  92         if (!umem)
  93                 return ERR_PTR(-ENOMEM);
  94
  95         umem->context   = context;
  96         umem->length    = size;
  97         umem->offset    = addr & ~PAGE_MASK;
  98         umem->page_size = PAGE_SIZE;
  99         /*
 100          * We ask for writable memory if any access flags other than
 101          * "remote read" are set.  "Local write" and "remote write"
 102          * obviously require write access.  "Remote atomic" can do
 103          * things like fetch and add, which will modify memory, and
 104          * "MW bind" can change permissions by binding a window.
 105          */
 106         umem->writable  = !!(access & ~IB_ACCESS_REMOTE_READ);
 107
 108         INIT_LIST_HEAD(&umem->chunk_list);
 109
 110         page_list = (struct page **) __get_free_page(GFP_KERNEL);
 111         if (!page_list) {
 112                 kfree(umem);
 113                 return ERR_PTR(-ENOMEM);
 114         }
 115
 116         npages = PAGE_ALIGN(size + umem->offset) >> PAGE_SHIFT;
 117
 118         down_write(&current->mm->mmap_sem);
 119
 120         locked     = npages + current->mm->locked_vm;
 121         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
 122
 123         if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
 124                 ret = -ENOMEM;
 125                 goto out;
 126         }
 127
 128         cur_base = addr & PAGE_MASK;
 129
 130         while (npages) {
 131                 ret = get_user_pages(current, current->mm, cur_base,
 132                                      min_t(int, npages,
 133                                            PAGE_SIZE / sizeof (struct page *)),
 134                                      1, !umem->writable, page_list, NULL);
 135
 136                 if (ret < 0)
 137                         goto out;
 138
 139                 cur_base += ret * PAGE_SIZE;
 140                 npages   -= ret;
 141
 142                 off = 0;
 143
 144                 while (ret) {
 145                         chunk = kmalloc(sizeof *chunk + sizeof (struct scatterlist) *
 146                                         min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK),
 147                                         GFP_KERNEL);
 148                         if (!chunk) {
 149                                 ret = -ENOMEM;
 150                                 goto out;
 151                         }
 152
 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;
 158                         }
 159
 160                         chunk->nmap = ib_dma_map_sg(context->device,
 161                                                     &chunk->page_list[0],
 162                                                     chunk->nents,
 163                                                     DMA_BIDIRECTIONAL);
 164                         if (chunk->nmap <= 0) {
 165                                 for (i = 0; i < chunk->nents; ++i)
 166                                         put_page(chunk->page_list[i].page);
 167                                 kfree(chunk);
 168
 169                                 ret = -ENOMEM;
 170                                 goto out;
 171                         }
 172
 173                         ret -= chunk->nents;
 174                         off += chunk->nents;
 175                         list_add_tail(&chunk->list, &umem->chunk_list);
 176                 }
 177
 178                 ret = 0;
 179         }
 180
 181 out:
 182         if (ret < 0) {
 183                 __ib_umem_release(context->device, umem, 0);
 184                 kfree(umem);
 185         } else
 186                 current->mm->locked_vm = locked;
 187
 188         up_write(&current->mm->mmap_sem);
 189         free_page((unsigned long) page_list);
 190
 191         return ret < 0 ? ERR_PTR(ret) : umem;
 192 }
 193 EXPORT_SYMBOL(ib_umem_get);
 194
 195 static void ib_umem_account(struct work_struct *_work)
 196 {
 197         struct ib_umem_account_work *work =
 198                 container_of(_work, struct ib_umem_account_work, work);
 199
 200         down_write(&work->mm->mmap_sem);
 201         work->mm->locked_vm -= work->diff;
 202         up_write(&work->mm->mmap_sem);
 203         mmput(work->mm);
 204         kfree(work);
 205 }
 206
 207 /**
 208  * ib_umem_release - release memory pinned with ib_umem_get
 209  * @umem: umem struct to release
 210  */
 211 void ib_umem_release(struct ib_umem *umem)
 212 {
 213         struct ib_umem_account_work *work;
 214         struct ib_ucontext *context = umem->context;
 215         struct mm_struct *mm;
 216         unsigned long diff;
 217
 218         __ib_umem_release(umem->context->device, umem, 1);
 219
 220         mm = get_task_mm(current);
 221         if (!mm)
 222                 return;
 223
 224         diff = PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT;
 225         kfree(umem);
 226
 227         /*
 228          * We may be called with the mm's mmap_sem already held.  This
 229          * can happen when a userspace munmap() is the call that drops
 230          * the last reference to our file and calls our release
 231          * method.  If there are memory regions to destroy, we'll end
 232          * up here and not be able to take the mmap_sem.  In that case
 233          * we defer the vm_locked accounting to the system workqueue.
 234          */
 235         if (context->closing && !down_write_trylock(&mm->mmap_sem)) {
 236                 work = kmalloc(sizeof *work, GFP_KERNEL);
 237                 if (!work) {
 238                         mmput(mm);
 239                         return;
 240                 }
 241
 242                 INIT_WORK(&work->work, ib_umem_account);
 243                 work->mm   = mm;
 244                 work->diff = diff;
 245
 246                 schedule_work(&work->work);
 247                 return;
 248         } else
 249                 down_write(&mm->mmap_sem);
 250
 251         current->mm->locked_vm -= diff;
 252         up_write(&mm->mmap_sem);
 253         mmput(mm);
 254 }
 255 EXPORT_SYMBOL(ib_umem_release);
 256
 257 int ib_umem_page_count(struct ib_umem *umem)
 258 {
 259         struct ib_umem_chunk *chunk;
 260         int shift;
 261         int i;
 262         int n;
 263
 264         shift = ilog2(umem->page_size);
 265
 266         n = 0;
 267         list_for_each_entry(chunk, &umem->chunk_list, list)
 268                 for (i = 0; i < chunk->nmap; ++i)
 269                         n += sg_dma_len(&chunk->page_list[i]) >> shift;
 270
 271         return n;
 272 }
 273 EXPORT_SYMBOL(ib_umem_page_count);