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
  35 #include <linux/mm.h>
  36 #include <linux/dma-mapping.h>
  37 #include <linux/sched.h>
  38 #include <linux/hugetlb.h>
  39 #include <linux/dma-attrs.h>
  40 #include <linux/slab.h>
  41
  42 #include "uverbs.h"
  43
  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]))
  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                         struct page *page = sg_page(&chunk->page_list[i]);
  59
  60                         if (umem->writable && dirty)
  61                                 set_page_dirty_lock(page);
  62                         put_page(page);
  63                 }
  64
  65                 kfree(chunk);
  66         }
  67 }
  68
  69 /**
  70  * ib_umem_get - Pin and DMA map userspace memory.
  71  * @context: userspace context to pin memory for
  72  * @addr: userspace virtual address to start at
  73  * @size: length of region to pin
  74  * @access: IB_ACCESS_xxx flags for memory being pinned
  75  * @dmasync: flush in-flight DMA when the memory region is written
  76  */
  77 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
  78                             size_t size, int access, int dmasync)
  79 {
  80         struct ib_umem *umem;
  81         struct page **page_list;
  82         struct vm_area_struct **vma_list;
  83         struct ib_umem_chunk *chunk;
  84         unsigned long locked;
  85         unsigned long lock_limit;
  86         unsigned long cur_base;
  87         unsigned long npages;
  88         int ret;
  89         int off;
  90         int i;
  91         DEFINE_DMA_ATTRS(attrs);
  92
  93         if (dmasync)
  94                 dma_set_attr(DMA_ATTR_WRITE_BARRIER, &attrs);
  95
  96         if (!can_do_mlock())
  97                 return ERR_PTR(-EPERM);
  98
  99         umem = kmalloc(sizeof *umem, GFP_KERNEL);
 100         if (!umem)
 101                 return ERR_PTR(-ENOMEM);
 102
 103         umem->context   = context;
 104         umem->length    = size;
 105         umem->offset    = addr & ~PAGE_MASK;
 106         umem->page_size = PAGE_SIZE;
 107         /*
 108          * We ask for writable memory if any access flags other than
 109          * "remote read" are set.  "Local write" and "remote write"
 110          * obviously require write access.  "Remote atomic" can do
 111          * things like fetch and add, which will modify memory, and
 112          * "MW bind" can change permissions by binding a window.
 113          */
 114         umem->writable  = !!(access & ~IB_ACCESS_REMOTE_READ);
 115
 116         /* We assume the memory is from hugetlb until proved otherwise */
 117         umem->hugetlb   = 1;
 118
 119         INIT_LIST_HEAD(&umem->chunk_list);
 120
 121         page_list = (struct page **) __get_free_page(GFP_KERNEL);
 122         if (!page_list) {
 123                 kfree(umem);
 124                 return ERR_PTR(-ENOMEM);
 125         }
 126
 127         /*
 128          * if we can't alloc the vma_list, it's not so bad;
 129          * just assume the memory is not hugetlb memory
 130          */
 131         vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL);
 132         if (!vma_list)
 133                 umem->hugetlb = 0;
 134
 135         npages = PAGE_ALIGN(size + umem->offset) >> PAGE_SHIFT;
 136
 137         down_write(&current->mm->mmap_sem);
 138
 139         locked     = npages + current->mm->locked_vm;
 140         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 141
 142         if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
 143                 ret = -ENOMEM;
 144                 goto out;
 145         }
 146
 147         cur_base = addr & PAGE_MASK;
 148
 149         ret = 0;
 150         while (npages) {
 151                 ret = get_user_pages(current, current->mm, cur_base,
 152                                      min_t(unsigned long, npages,
 153                                            PAGE_SIZE / sizeof (struct page *)),
 154                                      1, !umem->writable, page_list, vma_list);
 155
 156                 if (ret < 0)
 157                         goto out;
 158
 159                 cur_base += ret * PAGE_SIZE;
 160                 npages   -= ret;
 161
 162                 off = 0;
 163
 164                 while (ret) {
 165                         chunk = kmalloc(sizeof *chunk + sizeof (struct scatterlist) *
 166                                         min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK),
 167                                         GFP_KERNEL);
 168                         if (!chunk) {
 169                                 ret = -ENOMEM;
 170                                 goto out;
 171                         }
 172
 173                         chunk->nents = min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK);
 174                         sg_init_table(chunk->page_list, chunk->nents);
 175                         for (i = 0; i < chunk->nents; ++i) {
 176                                 if (vma_list &&
 177                                     !is_vm_hugetlb_page(vma_list[i + off]))
 178                                         umem->hugetlb = 0;
 179                                 sg_set_page(&chunk->page_list[i], page_list[i + off], PAGE_SIZE, 0);
 180                         }
 181
 182                         chunk->nmap = ib_dma_map_sg_attrs(context->device,
 183                                                           &chunk->page_list[0],
 184                                                           chunk->nents,
 185                                                           DMA_BIDIRECTIONAL,
 186                                                           &attrs);
 187                         if (chunk->nmap <= 0) {
 188                                 for (i = 0; i < chunk->nents; ++i)
 189                                         put_page(sg_page(&chunk->page_list[i]));
 190                                 kfree(chunk);
 191
 192                                 ret = -ENOMEM;
 193                                 goto out;
 194                         }
 195
 196                         ret -= chunk->nents;
 197                         off += chunk->nents;
 198                         list_add_tail(&chunk->list, &umem->chunk_list);
 199                 }
 200
 201                 ret = 0;
 202         }
 203
 204 out:
 205         if (ret < 0) {
 206                 __ib_umem_release(context->device, umem, 0);
 207                 kfree(umem);
 208         } else
 209                 current->mm->locked_vm = locked;
 210
 211         up_write(&current->mm->mmap_sem);
 212         if (vma_list)
 213                 free_page((unsigned long) vma_list);
 214         free_page((unsigned long) page_list);
 215
 216         return ret < 0 ? ERR_PTR(ret) : umem;
 217 }
 218 EXPORT_SYMBOL(ib_umem_get);
 219
 220 static void ib_umem_account(struct work_struct *work)
 221 {
 222         struct ib_umem *umem = container_of(work, struct ib_umem, work);
 223
 224         down_write(&umem->mm->mmap_sem);
 225         umem->mm->locked_vm -= umem->diff;
 226         up_write(&umem->mm->mmap_sem);
 227         mmput(umem->mm);
 228         kfree(umem);
 229 }
 230
 231 /**
 232  * ib_umem_release - release memory pinned with ib_umem_get
 233  * @umem: umem struct to release
 234  */
 235 void ib_umem_release(struct ib_umem *umem)
 236 {
 237         struct ib_ucontext *context = umem->context;
 238         struct mm_struct *mm;
 239         unsigned long diff;
 240
 241         __ib_umem_release(umem->context->device, umem, 1);
 242
 243         mm = get_task_mm(current);
 244         if (!mm) {
 245                 kfree(umem);
 246                 return;
 247         }
 248
 249         diff = PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT;
 250
 251         /*
 252          * We may be called with the mm's mmap_sem already held.  This
 253          * can happen when a userspace munmap() is the call that drops
 254          * the last reference to our file and calls our release
 255          * method.  If there are memory regions to destroy, we'll end
 256          * up here and not be able to take the mmap_sem.  In that case
 257          * we defer the vm_locked accounting to the system workqueue.
 258          */
 259         if (context->closing) {
 260                 if (!down_write_trylock(&mm->mmap_sem)) {
 261                         INIT_WORK(&umem->work, ib_umem_account);
 262                         umem->mm   = mm;
 263                         umem->diff = diff;
 264
 265                         queue_work(ib_wq, &umem->work);
 266                         return;
 267                 }
 268         } else
 269                 down_write(&mm->mmap_sem);
 270
 271         current->mm->locked_vm -= diff;
 272         up_write(&mm->mmap_sem);
 273         mmput(mm);
 274         kfree(umem);
 275 }
 276 EXPORT_SYMBOL(ib_umem_release);
 277
 278 int ib_umem_page_count(struct ib_umem *umem)
 279 {
 280         struct ib_umem_chunk *chunk;
 281         int shift;
 282         int i;
 283         int n;
 284
 285         shift = ilog2(umem->page_size);
 286
 287         n = 0;
 288         list_for_each_entry(chunk, &umem->chunk_list, list)
 289                 for (i = 0; i < chunk->nmap; ++i)
 290                         n += sg_dma_len(&chunk->page_list[i]) >> shift;
 291
 292         return n;
 293 }
 294 EXPORT_SYMBOL(ib_umem_page_count);