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/export.h>
  39 #include <linux/hugetlb.h>
  40 #include <linux/dma-attrs.h>
  41 #include <linux/slab.h>
  42
  43 #include "uverbs.h"
  44
  45 #define IB_UMEM_MAX_PAGE_CHUNK                                          \
  46         ((PAGE_SIZE - offsetof(struct ib_umem_chunk, page_list)) /      \
  47          ((void *) &((struct ib_umem_chunk *) 0)->page_list[1] -        \
  48           (void *) &((struct ib_umem_chunk *) 0)->page_list[0]))
  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                         struct page *page = sg_page(&chunk->page_list[i]);
  60
  61                         if (umem->writable && dirty)
  62                                 set_page_dirty_lock(page);
  63                         put_page(page);
  64                 }
  65
  66                 kfree(chunk);
  67         }
  68 }
  69
  70 /**
  71  * ib_umem_get - Pin and DMA map userspace memory.
  72  * @context: userspace context to pin memory for
  73  * @addr: userspace virtual address to start at
  74  * @size: length of region to pin
  75  * @access: IB_ACCESS_xxx flags for memory being pinned
  76  * @dmasync: flush in-flight DMA when the memory region is written
  77  */
  78 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
  79                             size_t size, int access, int dmasync)
  80 {
  81         struct ib_umem *umem;
  82         struct page **page_list;
  83         struct vm_area_struct **vma_list;
  84         struct ib_umem_chunk *chunk;
  85         unsigned long locked;
  86         unsigned long lock_limit;
  87         unsigned long cur_base;
  88         unsigned long npages;
  89         int ret;
  90         int off;
  91         int i;
  92         DEFINE_DMA_ATTRS(attrs);
  93
  94         if (dmasync)
  95                 dma_set_attr(DMA_ATTR_WRITE_BARRIER, &attrs);
  96
  97         if (!can_do_mlock())
  98                 return ERR_PTR(-EPERM);
  99
 100         umem = kmalloc(sizeof *umem, GFP_KERNEL);
 101         if (!umem)
 102                 return ERR_PTR(-ENOMEM);
 103
 104         umem->context   = context;
 105         umem->length    = size;
 106         umem->offset    = addr & ~PAGE_MASK;
 107         umem->page_size = PAGE_SIZE;
 108         /*
 109          * We ask for writable memory if any access flags other than
 110          * "remote read" are set.  "Local write" and "remote write"
 111          * obviously require write access.  "Remote atomic" can do
 112          * things like fetch and add, which will modify memory, and
 113          * "MW bind" can change permissions by binding a window.
 114          */
 115         umem->writable  = !!(access & ~IB_ACCESS_REMOTE_READ);
 116
 117         /* We assume the memory is from hugetlb until proved otherwise */
 118         umem->hugetlb   = 1;
 119
 120         INIT_LIST_HEAD(&umem->chunk_list);
 121
 122         page_list = (struct page **) __get_free_page(GFP_KERNEL);
 123         if (!page_list) {
 124                 kfree(umem);
 125                 return ERR_PTR(-ENOMEM);
 126         }
 127
 128         /*
 129          * if we can't alloc the vma_list, it's not so bad;
 130          * just assume the memory is not hugetlb memory
 131          */
 132         vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL);
 133         if (!vma_list)
 134                 umem->hugetlb = 0;
 135
 136         npages = PAGE_ALIGN(size + umem->offset) >> PAGE_SHIFT;
 137
 138         down_write(&current->mm->mmap_sem);
 139
 140         locked     = npages + current->mm->pinned_vm;
 141         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 142
 143         if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
 144                 ret = -ENOMEM;
 145                 goto out;
 146         }
 147
 148         cur_base = addr & PAGE_MASK;
 149
 150         ret = 0;
 151         while (npages) {
 152                 ret = get_user_pages(current, current->mm, cur_base,
 153                                      min_t(unsigned long, npages,
 154                                            PAGE_SIZE / sizeof (struct page *)),
 155                                      1, !umem->writable, page_list, vma_list);
 156
 157                 if (ret < 0)
 158                         goto out;
 159
 160                 cur_base += ret * PAGE_SIZE;
 161                 npages   -= ret;
 162
 163                 off = 0;
 164
 165                 while (ret) {
 166                         chunk = kmalloc(sizeof *chunk + sizeof (struct scatterlist) *
 167                                         min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK),
 168                                         GFP_KERNEL);
 169                         if (!chunk) {
 170                                 ret = -ENOMEM;
 171                                 goto out;
 172                         }
 173
 174                         chunk->nents = min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK);
 175                         sg_init_table(chunk->page_list, chunk->nents);
 176                         for (i = 0; i < chunk->nents; ++i) {
 177                                 if (vma_list &&
 178                                     !is_vm_hugetlb_page(vma_list[i + off]))
 179                                         umem->hugetlb = 0;
 180                                 sg_set_page(&chunk->page_list[i], page_list[i + off], PAGE_SIZE, 0);
 181                         }
 182
 183                         chunk->nmap = ib_dma_map_sg_attrs(context->device,
 184                                                           &chunk->page_list[0],
 185                                                           chunk->nents,
 186                                                           DMA_BIDIRECTIONAL,
 187                                                           &attrs);
 188                         if (chunk->nmap <= 0) {
 189                                 for (i = 0; i < chunk->nents; ++i)
 190                                         put_page(sg_page(&chunk->page_list[i]));
 191                                 kfree(chunk);
 192
 193                                 ret = -ENOMEM;
 194                                 goto out;
 195                         }
 196
 197                         ret -= chunk->nents;
 198                         off += chunk->nents;
 199                         list_add_tail(&chunk->list, &umem->chunk_list);
 200                 }
 201
 202                 ret = 0;
 203         }
 204
 205 out:
 206         if (ret < 0) {
 207                 __ib_umem_release(context->device, umem, 0);
 208                 kfree(umem);
 209         } else
 210                 current->mm->pinned_vm = locked;
 211
 212         up_write(&current->mm->mmap_sem);
 213         if (vma_list)
 214                 free_page((unsigned long) vma_list);
 215         free_page((unsigned long) page_list);
 216
 217         return ret < 0 ? ERR_PTR(ret) : umem;
 218 }
 219 EXPORT_SYMBOL(ib_umem_get);
 220
 221 static void ib_umem_account(struct work_struct *work)
 222 {
 223         struct ib_umem *umem = container_of(work, struct ib_umem, work);
 224
 225         down_write(&umem->mm->mmap_sem);
 226         umem->mm->pinned_vm -= umem->diff;
 227         up_write(&umem->mm->mmap_sem);
 228         mmput(umem->mm);
 229         kfree(umem);
 230 }
 231
 232 /**
 233  * ib_umem_release - release memory pinned with ib_umem_get
 234  * @umem: umem struct to release
 235  */
 236 void ib_umem_release(struct ib_umem *umem)
 237 {
 238         struct ib_ucontext *context = umem->context;
 239         struct mm_struct *mm;
 240         unsigned long diff;
 241
 242         __ib_umem_release(umem->context->device, umem, 1);
 243
 244         mm = get_task_mm(current);
 245         if (!mm) {
 246                 kfree(umem);
 247                 return;
 248         }
 249
 250         diff = PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT;
 251
 252         /*
 253          * We may be called with the mm's mmap_sem already held.  This
 254          * can happen when a userspace munmap() is the call that drops
 255          * the last reference to our file and calls our release
 256          * method.  If there are memory regions to destroy, we'll end
 257          * up here and not be able to take the mmap_sem.  In that case
 258          * we defer the vm_locked accounting to the system workqueue.
 259          */
 260         if (context->closing) {
 261                 if (!down_write_trylock(&mm->mmap_sem)) {
 262                         INIT_WORK(&umem->work, ib_umem_account);
 263                         umem->mm   = mm;
 264                         umem->diff = diff;
 265
 266                         queue_work(ib_wq, &umem->work);
 267                         return;
 268                 }
 269         } else
 270                 down_write(&mm->mmap_sem);
 271
 272         current->mm->pinned_vm -= diff;
 273         up_write(&mm->mmap_sem);
 274         mmput(mm);
 275         kfree(umem);
 276 }
 277 EXPORT_SYMBOL(ib_umem_release);
 278
 279 int ib_umem_page_count(struct ib_umem *umem)
 280 {
 281         struct ib_umem_chunk *chunk;
 282         int shift;
 283         int i;
 284         int n;
 285
 286         shift = ilog2(umem->page_size);
 287
 288         n = 0;
 289         list_for_each_entry(chunk, &umem->chunk_list, list)
 290                 for (i = 0; i < chunk->nmap; ++i)
 291                         n += sg_dma_len(&chunk->page_list[i]) >> shift;
 292
 293         return n;
 294 }
 295 EXPORT_SYMBOL(ib_umem_page_count);