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 #include <rdma/ib_umem_odp.h>
  43
  44 #include "uverbs.h"
  45
  46
  47 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
  48 {
  49         struct scatterlist *sg;
  50         struct page *page;
  51         int i;
  52
  53         if (umem->nmap > 0)
  54                 ib_dma_unmap_sg(dev, umem->sg_head.sgl,
  55                                 umem->nmap,
  56                                 DMA_BIDIRECTIONAL);
  57
  58         for_each_sg(umem->sg_head.sgl, sg, umem->npages, i) {
  59
  60                 page = sg_page(sg);
  61                 if (umem->writable && dirty)
  62                         set_page_dirty_lock(page);
  63                 put_page(page);
  64         }
  65
  66         sg_free_table(&umem->sg_head);
  67         return;
  68
  69 }
  70
  71 /**
  72  * ib_umem_get - Pin and DMA map userspace memory.
  73  *
  74  * If access flags indicate ODP memory, avoid pinning. Instead, stores
  75  * the mm for future page fault handling in conjunction with MMU notifiers.
  76  *
  77  * @context: userspace context to pin memory for
  78  * @addr: userspace virtual address to start at
  79  * @size: length of region to pin
  80  * @access: IB_ACCESS_xxx flags for memory being pinned
  81  * @dmasync: flush in-flight DMA when the memory region is written
  82  */
  83 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
  84                             size_t size, int access, int dmasync)
  85 {
  86         struct ib_umem *umem;
  87         struct page **page_list;
  88         struct vm_area_struct **vma_list;
  89         unsigned long locked;
  90         unsigned long lock_limit;
  91         unsigned long cur_base;
  92         unsigned long npages;
  93         int ret;
  94         int i;
  95         DEFINE_DMA_ATTRS(attrs);
  96         struct scatterlist *sg, *sg_list_start;
  97         int need_release = 0;
  98
  99         if (dmasync)
 100                 dma_set_attr(DMA_ATTR_WRITE_BARRIER, &attrs);
 101
 102         if (!can_do_mlock())
 103                 return ERR_PTR(-EPERM);
 104
 105         umem = kzalloc(sizeof *umem, GFP_KERNEL);
 106         if (!umem)
 107                 return ERR_PTR(-ENOMEM);
 108
 109         umem->context   = context;
 110         umem->length    = size;
 111         umem->address   = addr;
 112         umem->page_size = PAGE_SIZE;
 113         umem->pid       = get_task_pid(current, PIDTYPE_PID);
 114         /*
 115          * We ask for writable memory if any of the following
 116          * access flags are set.  "Local write" and "remote write"
 117          * obviously require write access.  "Remote atomic" can do
 118          * things like fetch and add, which will modify memory, and
 119          * "MW bind" can change permissions by binding a window.
 120          */
 121         umem->writable  = !!(access &
 122                 (IB_ACCESS_LOCAL_WRITE   | IB_ACCESS_REMOTE_WRITE |
 123                  IB_ACCESS_REMOTE_ATOMIC | IB_ACCESS_MW_BIND));
 124
 125         if (access & IB_ACCESS_ON_DEMAND) {
 126                 ret = ib_umem_odp_get(context, umem);
 127                 if (ret) {
 128                         kfree(umem);
 129                         return ERR_PTR(ret);
 130                 }
 131                 return umem;
 132         }
 133
 134         umem->odp_data = NULL;
 135
 136         /* We assume the memory is from hugetlb until proved otherwise */
 137         umem->hugetlb   = 1;
 138
 139         page_list = (struct page **) __get_free_page(GFP_KERNEL);
 140         if (!page_list) {
 141                 kfree(umem);
 142                 return ERR_PTR(-ENOMEM);
 143         }
 144
 145         /*
 146          * if we can't alloc the vma_list, it's not so bad;
 147          * just assume the memory is not hugetlb memory
 148          */
 149         vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL);
 150         if (!vma_list)
 151                 umem->hugetlb = 0;
 152
 153         npages = ib_umem_num_pages(umem);
 154
 155         down_write(&current->mm->mmap_sem);
 156
 157         locked     = npages + current->mm->pinned_vm;
 158         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 159
 160         if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
 161                 ret = -ENOMEM;
 162                 goto out;
 163         }
 164
 165         cur_base = addr & PAGE_MASK;
 166
 167         if (npages == 0) {
 168                 ret = -EINVAL;
 169                 goto out;
 170         }
 171
 172         ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL);
 173         if (ret)
 174                 goto out;
 175
 176         need_release = 1;
 177         sg_list_start = umem->sg_head.sgl;
 178
 179         while (npages) {
 180                 ret = get_user_pages(current, current->mm, cur_base,
 181                                      min_t(unsigned long, npages,
 182                                            PAGE_SIZE / sizeof (struct page *)),
 183                                      1, !umem->writable, page_list, vma_list);
 184
 185                 if (ret < 0)
 186                         goto out;
 187
 188                 umem->npages += ret;
 189                 cur_base += ret * PAGE_SIZE;
 190                 npages   -= ret;
 191
 192                 for_each_sg(sg_list_start, sg, ret, i) {
 193                         if (vma_list && !is_vm_hugetlb_page(vma_list[i]))
 194                                 umem->hugetlb = 0;
 195
 196                         sg_set_page(sg, page_list[i], PAGE_SIZE, 0);
 197                 }
 198
 199                 /* preparing for next loop */
 200                 sg_list_start = sg;
 201         }
 202
 203         umem->nmap = ib_dma_map_sg_attrs(context->device,
 204                                   umem->sg_head.sgl,
 205                                   umem->npages,
 206                                   DMA_BIDIRECTIONAL,
 207                                   &attrs);
 208
 209         if (umem->nmap <= 0) {
 210                 ret = -ENOMEM;
 211                 goto out;
 212         }
 213
 214         ret = 0;
 215
 216 out:
 217         if (ret < 0) {
 218                 if (need_release)
 219                         __ib_umem_release(context->device, umem, 0);
 220                 put_pid(umem->pid);
 221                 kfree(umem);
 222         } else
 223                 current->mm->pinned_vm = locked;
 224
 225         up_write(&current->mm->mmap_sem);
 226         if (vma_list)
 227                 free_page((unsigned long) vma_list);
 228         free_page((unsigned long) page_list);
 229
 230         return ret < 0 ? ERR_PTR(ret) : umem;
 231 }
 232 EXPORT_SYMBOL(ib_umem_get);
 233
 234 static void ib_umem_account(struct work_struct *work)
 235 {
 236         struct ib_umem *umem = container_of(work, struct ib_umem, work);
 237
 238         down_write(&umem->mm->mmap_sem);
 239         umem->mm->pinned_vm -= umem->diff;
 240         up_write(&umem->mm->mmap_sem);
 241         mmput(umem->mm);
 242         kfree(umem);
 243 }
 244
 245 /**
 246  * ib_umem_release - release memory pinned with ib_umem_get
 247  * @umem: umem struct to release
 248  */
 249 void ib_umem_release(struct ib_umem *umem)
 250 {
 251         struct ib_ucontext *context = umem->context;
 252         struct mm_struct *mm;
 253         struct task_struct *task;
 254         unsigned long diff;
 255
 256         if (umem->odp_data) {
 257                 ib_umem_odp_release(umem);
 258                 return;
 259         }
 260
 261         __ib_umem_release(umem->context->device, umem, 1);
 262
 263         task = get_pid_task(umem->pid, PIDTYPE_PID);
 264         put_pid(umem->pid);
 265         if (!task)
 266                 goto out;
 267         mm = get_task_mm(task);
 268         put_task_struct(task);
 269         if (!mm)
 270                 goto out;
 271
 272         diff = ib_umem_num_pages(umem);
 273
 274         /*
 275          * We may be called with the mm's mmap_sem already held.  This
 276          * can happen when a userspace munmap() is the call that drops
 277          * the last reference to our file and calls our release
 278          * method.  If there are memory regions to destroy, we'll end
 279          * up here and not be able to take the mmap_sem.  In that case
 280          * we defer the vm_locked accounting to the system workqueue.
 281          */
 282         if (context->closing) {
 283                 if (!down_write_trylock(&mm->mmap_sem)) {
 284                         INIT_WORK(&umem->work, ib_umem_account);
 285                         umem->mm   = mm;
 286                         umem->diff = diff;
 287
 288                         queue_work(ib_wq, &umem->work);
 289                         return;
 290                 }
 291         } else
 292                 down_write(&mm->mmap_sem);
 293
 294         mm->pinned_vm -= diff;
 295         up_write(&mm->mmap_sem);
 296         mmput(mm);
 297 out:
 298         kfree(umem);
 299 }
 300 EXPORT_SYMBOL(ib_umem_release);
 301
 302 int ib_umem_page_count(struct ib_umem *umem)
 303 {
 304         int shift;
 305         int i;
 306         int n;
 307         struct scatterlist *sg;
 308
 309         if (umem->odp_data)
 310                 return ib_umem_num_pages(umem);
 311
 312         shift = ilog2(umem->page_size);
 313
 314         n = 0;
 315         for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i)
 316                 n += sg_dma_len(sg) >> shift;
 317
 318         return n;
 319 }
 320 EXPORT_SYMBOL(ib_umem_page_count);
 321
 322 /*
 323  * Copy from the given ib_umem's pages to the given buffer.
 324  *
 325  * umem - the umem to copy from
 326  * offset - offset to start copying from
 327  * dst - destination buffer
 328  * length - buffer length
 329  *
 330  * Returns 0 on success, or an error code.
 331  */
 332 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 333                       size_t length)
 334 {
 335         size_t end = offset + length;
 336         int ret;
 337
 338         if (offset > umem->length || length > umem->length - offset) {
 339                 pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n",
 340                        offset, umem->length, end);
 341                 return -EINVAL;
 342         }
 343
 344         ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->nmap, dst, length,
 345                                  offset + ib_umem_offset(umem));
 346
 347         if (ret < 0)
 348                 return ret;
 349         else if (ret != length)
 350                 return -EINVAL;
 351         else
 352                 return 0;
 353 }
 354 EXPORT_SYMBOL(ib_umem_copy_from);