linux/drivers/infiniband/core/umem.c
<<
>>
Prefs
   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/signal.h>
  38#include <linux/sched/mm.h>
  39#include <linux/export.h>
  40#include <linux/hugetlb.h>
  41#include <linux/slab.h>
  42#include <rdma/ib_umem_odp.h>
  43
  44#include "uverbs.h"
  45
  46
  47static 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->npages,
  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 */
  83struct 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        unsigned long dma_attrs = 0;
  96        struct scatterlist *sg, *sg_list_start;
  97        int need_release = 0;
  98        unsigned int gup_flags = FOLL_WRITE;
  99
 100        if (dmasync)
 101                dma_attrs |= DMA_ATTR_WRITE_BARRIER;
 102
 103        /*
 104         * If the combination of the addr and size requested for this memory
 105         * region causes an integer overflow, return error.
 106         */
 107        if (((addr + size) < addr) ||
 108            PAGE_ALIGN(addr + size) < (addr + size))
 109                return ERR_PTR(-EINVAL);
 110
 111        if (!can_do_mlock())
 112                return ERR_PTR(-EPERM);
 113
 114        umem = kzalloc(sizeof *umem, GFP_KERNEL);
 115        if (!umem)
 116                return ERR_PTR(-ENOMEM);
 117
 118        umem->context   = context;
 119        umem->length    = size;
 120        umem->address   = addr;
 121        umem->page_size = PAGE_SIZE;
 122        umem->pid       = get_task_pid(current, PIDTYPE_PID);
 123        /*
 124         * We ask for writable memory if any of the following
 125         * access flags are set.  "Local write" and "remote write"
 126         * obviously require write access.  "Remote atomic" can do
 127         * things like fetch and add, which will modify memory, and
 128         * "MW bind" can change permissions by binding a window.
 129         */
 130        umem->writable  = !!(access &
 131                (IB_ACCESS_LOCAL_WRITE   | IB_ACCESS_REMOTE_WRITE |
 132                 IB_ACCESS_REMOTE_ATOMIC | IB_ACCESS_MW_BIND));
 133
 134        if (access & IB_ACCESS_ON_DEMAND) {
 135                put_pid(umem->pid);
 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                put_pid(umem->pid);
 152                kfree(umem);
 153                return ERR_PTR(-ENOMEM);
 154        }
 155
 156        /*
 157         * if we can't alloc the vma_list, it's not so bad;
 158         * just assume the memory is not hugetlb memory
 159         */
 160        vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL);
 161        if (!vma_list)
 162                umem->hugetlb = 0;
 163
 164        npages = ib_umem_num_pages(umem);
 165
 166        down_write(&current->mm->mmap_sem);
 167
 168        locked     = npages + current->mm->pinned_vm;
 169        lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 170
 171        if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
 172                ret = -ENOMEM;
 173                goto out;
 174        }
 175
 176        cur_base = addr & PAGE_MASK;
 177
 178        if (npages == 0 || npages > UINT_MAX) {
 179                ret = -EINVAL;
 180                goto out;
 181        }
 182
 183        ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL);
 184        if (ret)
 185                goto out;
 186
 187        if (!umem->writable)
 188                gup_flags |= FOLL_FORCE;
 189
 190        need_release = 1;
 191        sg_list_start = umem->sg_head.sgl;
 192
 193        while (npages) {
 194                ret = get_user_pages(cur_base,
 195                                     min_t(unsigned long, npages,
 196                                           PAGE_SIZE / sizeof (struct page *)),
 197                                     gup_flags, page_list, vma_list);
 198
 199                if (ret < 0)
 200                        goto out;
 201
 202                umem->npages += ret;
 203                cur_base += ret * PAGE_SIZE;
 204                npages   -= ret;
 205
 206                for_each_sg(sg_list_start, sg, ret, i) {
 207                        if (vma_list && !is_vm_hugetlb_page(vma_list[i]))
 208                                umem->hugetlb = 0;
 209
 210                        sg_set_page(sg, page_list[i], PAGE_SIZE, 0);
 211                }
 212
 213                /* preparing for next loop */
 214                sg_list_start = sg;
 215        }
 216
 217        umem->nmap = ib_dma_map_sg_attrs(context->device,
 218                                  umem->sg_head.sgl,
 219                                  umem->npages,
 220                                  DMA_BIDIRECTIONAL,
 221                                  dma_attrs);
 222
 223        if (umem->nmap <= 0) {
 224                ret = -ENOMEM;
 225                goto out;
 226        }
 227
 228        ret = 0;
 229
 230out:
 231        if (ret < 0) {
 232                if (need_release)
 233                        __ib_umem_release(context->device, umem, 0);
 234                put_pid(umem->pid);
 235                kfree(umem);
 236        } else
 237                current->mm->pinned_vm = locked;
 238
 239        up_write(&current->mm->mmap_sem);
 240        if (vma_list)
 241                free_page((unsigned long) vma_list);
 242        free_page((unsigned long) page_list);
 243
 244        return ret < 0 ? ERR_PTR(ret) : umem;
 245}
 246EXPORT_SYMBOL(ib_umem_get);
 247
 248static void ib_umem_account(struct work_struct *work)
 249{
 250        struct ib_umem *umem = container_of(work, struct ib_umem, work);
 251
 252        down_write(&umem->mm->mmap_sem);
 253        umem->mm->pinned_vm -= umem->diff;
 254        up_write(&umem->mm->mmap_sem);
 255        mmput(umem->mm);
 256        kfree(umem);
 257}
 258
 259/**
 260 * ib_umem_release - release memory pinned with ib_umem_get
 261 * @umem: umem struct to release
 262 */
 263void ib_umem_release(struct ib_umem *umem)
 264{
 265        struct ib_ucontext *context = umem->context;
 266        struct mm_struct *mm;
 267        struct task_struct *task;
 268        unsigned long diff;
 269
 270        if (umem->odp_data) {
 271                ib_umem_odp_release(umem);
 272                return;
 273        }
 274
 275        __ib_umem_release(umem->context->device, umem, 1);
 276
 277        task = get_pid_task(umem->pid, PIDTYPE_PID);
 278        put_pid(umem->pid);
 279        if (!task)
 280                goto out;
 281        mm = get_task_mm(task);
 282        put_task_struct(task);
 283        if (!mm)
 284                goto out;
 285
 286        diff = ib_umem_num_pages(umem);
 287
 288        /*
 289         * We may be called with the mm's mmap_sem already held.  This
 290         * can happen when a userspace munmap() is the call that drops
 291         * the last reference to our file and calls our release
 292         * method.  If there are memory regions to destroy, we'll end
 293         * up here and not be able to take the mmap_sem.  In that case
 294         * we defer the vm_locked accounting to the system workqueue.
 295         */
 296        if (context->closing) {
 297                if (!down_write_trylock(&mm->mmap_sem)) {
 298                        INIT_WORK(&umem->work, ib_umem_account);
 299                        umem->mm   = mm;
 300                        umem->diff = diff;
 301
 302                        queue_work(ib_wq, &umem->work);
 303                        return;
 304                }
 305        } else
 306                down_write(&mm->mmap_sem);
 307
 308        mm->pinned_vm -= diff;
 309        up_write(&mm->mmap_sem);
 310        mmput(mm);
 311out:
 312        kfree(umem);
 313}
 314EXPORT_SYMBOL(ib_umem_release);
 315
 316int ib_umem_page_count(struct ib_umem *umem)
 317{
 318        int shift;
 319        int i;
 320        int n;
 321        struct scatterlist *sg;
 322
 323        if (umem->odp_data)
 324                return ib_umem_num_pages(umem);
 325
 326        shift = ilog2(umem->page_size);
 327
 328        n = 0;
 329        for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i)
 330                n += sg_dma_len(sg) >> shift;
 331
 332        return n;
 333}
 334EXPORT_SYMBOL(ib_umem_page_count);
 335
 336/*
 337 * Copy from the given ib_umem's pages to the given buffer.
 338 *
 339 * umem - the umem to copy from
 340 * offset - offset to start copying from
 341 * dst - destination buffer
 342 * length - buffer length
 343 *
 344 * Returns 0 on success, or an error code.
 345 */
 346int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 347                      size_t length)
 348{
 349        size_t end = offset + length;
 350        int ret;
 351
 352        if (offset > umem->length || length > umem->length - offset) {
 353                pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n",
 354                       offset, umem->length, end);
 355                return -EINVAL;
 356        }
 357
 358        ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->nmap, dst, length,
 359                                 offset + ib_umem_offset(umem));
 360
 361        if (ret < 0)
 362                return ret;
 363        else if (ret != length)
 364                return -EINVAL;
 365        else
 366                return 0;
 367}
 368EXPORT_SYMBOL(ib_umem_copy_from);
 369