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/slab.h>
  41 #include <rdma/ib_umem_odp.h>
  42
  43 #include "uverbs.h"
  44
  45
  46 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
  47 {
  48         struct scatterlist *sg;
  49         struct page *page;
  50         int i;
  51
  52         if (umem->nmap > 0)
  53                 ib_dma_unmap_sg(dev, umem->sg_head.sgl,
  54                                 umem->nmap,
  55                                 DMA_BIDIRECTIONAL);
  56
  57         for_each_sg(umem->sg_head.sgl, sg, umem->npages, i) {
  58
  59                 page = sg_page(sg);
  60                 if (umem->writable && dirty)
  61                         set_page_dirty_lock(page);
  62                 put_page(page);
  63         }
  64
  65         sg_free_table(&umem->sg_head);
  66         return;
  67
  68 }
  69
  70 /**
  71  * ib_umem_get - Pin and DMA map userspace memory.
  72  *
  73  * If access flags indicate ODP memory, avoid pinning. Instead, stores
  74  * the mm for future page fault handling in conjunction with MMU notifiers.
  75  *
  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
  81  */
  82 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
  83                             size_t size, int access, int dmasync)
  84 {
  85         struct ib_umem *umem;
  86         struct page **page_list;
  87         struct vm_area_struct **vma_list;
  88         unsigned long locked;
  89         unsigned long lock_limit;
  90         unsigned long cur_base;
  91         unsigned long npages;
  92         int ret;
  93         int i;
  94         unsigned long dma_attrs = 0;
  95         struct scatterlist *sg, *sg_list_start;
  96         int need_release = 0;
  97
  98         if (dmasync)
  99                 dma_attrs |= DMA_ATTR_WRITE_BARRIER;
 100
 101         if (!size)
 102                 return ERR_PTR(-EINVAL);
 103
 104         /*
 105          * If the combination of the addr and size requested for this memory
 106          * region causes an integer overflow, return error.
 107          */
 108         if (((addr + size) < addr) ||
 109             PAGE_ALIGN(addr + size) < (addr + size))
 110                 return ERR_PTR(-EINVAL);
 111
 112         if (!can_do_mlock())
 113                 return ERR_PTR(-EPERM);
 114
 115         umem = kzalloc(sizeof *umem, GFP_KERNEL);
 116         if (!umem)
 117                 return ERR_PTR(-ENOMEM);
 118
 119         umem->context   = context;
 120         umem->length    = size;
 121         umem->address   = addr;
 122         umem->page_size = PAGE_SIZE;
 123         umem->pid       = get_task_pid(current, PIDTYPE_PID);
 124         /*
 125          * We ask for writable memory if any of the following
 126          * access flags are set.  "Local write" and "remote write"
 127          * obviously require write access.  "Remote atomic" can do
 128          * things like fetch and add, which will modify memory, and
 129          * "MW bind" can change permissions by binding a window.
 130          */
 131         umem->writable  = !!(access &
 132                 (IB_ACCESS_LOCAL_WRITE   | IB_ACCESS_REMOTE_WRITE |
 133                  IB_ACCESS_REMOTE_ATOMIC | IB_ACCESS_MW_BIND));
 134
 135         if (access & IB_ACCESS_ON_DEMAND) {
 136                 ret = ib_umem_odp_get(context, umem);
 137                 if (ret) {
 138                         kfree(umem);
 139                         return ERR_PTR(ret);
 140                 }
 141                 return umem;
 142         }
 143
 144         umem->odp_data = NULL;
 145
 146         /* We assume the memory is from hugetlb until proved otherwise */
 147         umem->hugetlb   = 1;
 148
 149         page_list = (struct page **) __get_free_page(GFP_KERNEL);
 150         if (!page_list) {
 151                 kfree(umem);
 152                 return ERR_PTR(-ENOMEM);
 153         }
 154
 155         /*
 156          * if we can't alloc the vma_list, it's not so bad;
 157          * just assume the memory is not hugetlb memory
 158          */
 159         vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL);
 160         if (!vma_list)
 161                 umem->hugetlb = 0;
 162
 163         npages = ib_umem_num_pages(umem);
 164
 165         down_write(&current->mm->mmap_sem);
 166
 167         locked     = npages + current->mm->pinned_vm;
 168         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 169
 170         if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
 171                 ret = -ENOMEM;
 172                 goto out;
 173         }
 174
 175         cur_base = addr & PAGE_MASK;
 176
 177         if (npages == 0) {
 178                 ret = -EINVAL;
 179                 goto out;
 180         }
 181
 182         ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL);
 183         if (ret)
 184                 goto out;
 185
 186         need_release = 1;
 187         sg_list_start = umem->sg_head.sgl;
 188
 189         while (npages) {
 190                 ret = get_user_pages(cur_base,
 191                                      min_t(unsigned long, npages,
 192                                            PAGE_SIZE / sizeof (struct page *)),
 193                                      1, !umem->writable, page_list, vma_list);
 194
 195                 if (ret < 0)
 196                         goto out;
 197
 198                 umem->npages += ret;
 199                 cur_base += ret * PAGE_SIZE;
 200                 npages   -= ret;
 201
 202                 for_each_sg(sg_list_start, sg, ret, i) {
 203                         if (vma_list && !is_vm_hugetlb_page(vma_list[i]))
 204                                 umem->hugetlb = 0;
 205
 206                         sg_set_page(sg, page_list[i], PAGE_SIZE, 0);
 207                 }
 208
 209                 /* preparing for next loop */
 210                 sg_list_start = sg;
 211         }
 212
 213         umem->nmap = ib_dma_map_sg_attrs(context->device,
 214                                   umem->sg_head.sgl,
 215                                   umem->npages,
 216                                   DMA_BIDIRECTIONAL,
 217                                   dma_attrs);
 218
 219         if (umem->nmap <= 0) {
 220                 ret = -ENOMEM;
 221                 goto out;
 222         }
 223
 224         ret = 0;
 225
 226 out:
 227         if (ret < 0) {
 228                 if (need_release)
 229                         __ib_umem_release(context->device, umem, 0);
 230                 put_pid(umem->pid);
 231                 kfree(umem);
 232         } else
 233                 current->mm->pinned_vm = locked;
 234
 235         up_write(&current->mm->mmap_sem);
 236         if (vma_list)
 237                 free_page((unsigned long) vma_list);
 238         free_page((unsigned long) page_list);
 239
 240         return ret < 0 ? ERR_PTR(ret) : umem;
 241 }
 242 EXPORT_SYMBOL(ib_umem_get);
 243
 244 static void ib_umem_account(struct work_struct *work)
 245 {
 246         struct ib_umem *umem = container_of(work, struct ib_umem, work);
 247
 248         down_write(&umem->mm->mmap_sem);
 249         umem->mm->pinned_vm -= umem->diff;
 250         up_write(&umem->mm->mmap_sem);
 251         mmput(umem->mm);
 252         kfree(umem);
 253 }
 254
 255 /**
 256  * ib_umem_release - release memory pinned with ib_umem_get
 257  * @umem: umem struct to release
 258  */
 259 void ib_umem_release(struct ib_umem *umem)
 260 {
 261         struct ib_ucontext *context = umem->context;
 262         struct mm_struct *mm;
 263         struct task_struct *task;
 264         unsigned long diff;
 265
 266         if (umem->odp_data) {
 267                 ib_umem_odp_release(umem);
 268                 return;
 269         }
 270
 271         __ib_umem_release(umem->context->device, umem, 1);
 272
 273         task = get_pid_task(umem->pid, PIDTYPE_PID);
 274         put_pid(umem->pid);
 275         if (!task)
 276                 goto out;
 277         mm = get_task_mm(task);
 278         put_task_struct(task);
 279         if (!mm)
 280                 goto out;
 281
 282         diff = ib_umem_num_pages(umem);
 283
 284         /*
 285          * We may be called with the mm's mmap_sem already held.  This
 286          * can happen when a userspace munmap() is the call that drops
 287          * the last reference to our file and calls our release
 288          * method.  If there are memory regions to destroy, we'll end
 289          * up here and not be able to take the mmap_sem.  In that case
 290          * we defer the vm_locked accounting to the system workqueue.
 291          */
 292         if (context->closing) {
 293                 if (!down_write_trylock(&mm->mmap_sem)) {
 294                         INIT_WORK(&umem->work, ib_umem_account);
 295                         umem->mm   = mm;
 296                         umem->diff = diff;
 297
 298                         queue_work(ib_wq, &umem->work);
 299                         return;
 300                 }
 301         } else
 302                 down_write(&mm->mmap_sem);
 303
 304         mm->pinned_vm -= diff;
 305         up_write(&mm->mmap_sem);
 306         mmput(mm);
 307 out:
 308         kfree(umem);
 309 }
 310 EXPORT_SYMBOL(ib_umem_release);
 311
 312 int ib_umem_page_count(struct ib_umem *umem)
 313 {
 314         int shift;
 315         int i;
 316         int n;
 317         struct scatterlist *sg;
 318
 319         if (umem->odp_data)
 320                 return ib_umem_num_pages(umem);
 321
 322         shift = ilog2(umem->page_size);
 323
 324         n = 0;
 325         for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i)
 326                 n += sg_dma_len(sg) >> shift;
 327
 328         return n;
 329 }
 330 EXPORT_SYMBOL(ib_umem_page_count);
 331
 332 /*
 333  * Copy from the given ib_umem's pages to the given buffer.
 334  *
 335  * umem - the umem to copy from
 336  * offset - offset to start copying from
 337  * dst - destination buffer
 338  * length - buffer length
 339  *
 340  * Returns 0 on success, or an error code.
 341  */
 342 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 343                       size_t length)
 344 {
 345         size_t end = offset + length;
 346         int ret;
 347
 348         if (offset > umem->length || length > umem->length - offset) {
 349                 pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n",
 350                        offset, umem->length, end);
 351                 return -EINVAL;
 352         }
 353
 354         ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->nmap, dst, length,
 355                                  offset + ib_umem_offset(umem));
 356
 357         if (ret < 0)
 358                 return ret;
 359         else if (ret != length)
 360                 return -EINVAL;
 361         else
 362                 return 0;
 363 }
 364 EXPORT_SYMBOL(ib_umem_copy_from);