drivers/infiniband/hw/ipath/ipath_user_pages.c

   1 /*
   2  * Copyright (c) 2006 QLogic, Inc. All rights reserved.
   3  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <linux/mm.h>
  35 #include <linux/device.h>
  36
  37 #include "ipath_kernel.h"
  38
  39 static void __ipath_release_user_pages(struct page **p, size_t num_pages,
  40                                    int dirty)
  41 {
  42         size_t i;
  43
  44         for (i = 0; i < num_pages; i++) {
  45                 ipath_cdbg(MM, "%lu/%lu put_page %p\n", (unsigned long) i,
  46                            (unsigned long) num_pages, p[i]);
  47                 if (dirty)
  48                         set_page_dirty_lock(p[i]);
  49                 put_page(p[i]);
  50         }
  51 }
  52
  53 /* call with current->mm->mmap_sem held */
  54 static int __get_user_pages(unsigned long start_page, size_t num_pages,
  55                         struct page **p, struct vm_area_struct **vma)
  56 {
  57         unsigned long lock_limit;
  58         size_t got;
  59         int ret;
  60
  61         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur >>
  62                 PAGE_SHIFT;
  63
  64         if (num_pages > lock_limit) {
  65                 ret = -ENOMEM;
  66                 goto bail;
  67         }
  68
  69         ipath_cdbg(VERBOSE, "pin %lx pages from vaddr %lx\n",
  70                    (unsigned long) num_pages, start_page);
  71
  72         for (got = 0; got < num_pages; got += ret) {
  73                 ret = get_user_pages(current, current->mm,
  74                                      start_page + got * PAGE_SIZE,
  75                                      num_pages - got, 1, 1,
  76                                      p + got, vma);
  77                 if (ret < 0)
  78                         goto bail_release;
  79         }
  80
  81         current->mm->locked_vm += num_pages;
  82
  83         ret = 0;
  84         goto bail;
  85
  86 bail_release:
  87         __ipath_release_user_pages(p, got, 0);
  88 bail:
  89         return ret;
  90 }
  91
  92 /**
  93  * ipath_map_page - a safety wrapper around pci_map_page()
  94  *
  95  * A dma_addr of all 0's is interpreted by the chip as "disabled".
  96  * Unfortunately, it can also be a valid dma_addr returned on some
  97  * architectures.
  98  *
  99  * The powerpc iommu assigns dma_addrs in ascending order, so we don't
 100  * have to bother with retries or mapping a dummy page to insure we
 101  * don't just get the same mapping again.
 102  *
 103  * I'm sure we won't be so lucky with other iommu's, so FIXME.
 104  */
 105 dma_addr_t ipath_map_page(struct pci_dev *hwdev, struct page *page,
 106         unsigned long offset, size_t size, int direction)
 107 {
 108         dma_addr_t phys;
 109
 110         phys = pci_map_page(hwdev, page, offset, size, direction);
 111
 112         if (phys == 0) {
 113                 pci_unmap_page(hwdev, phys, size, direction);
 114                 phys = pci_map_page(hwdev, page, offset, size, direction);
 115                 /*
 116                  * FIXME: If we get 0 again, we should keep this page,
 117                  * map another, then free the 0 page.
 118                  */
 119         }
 120
 121         return phys;
 122 }
 123
 124 /**
 125  * ipath_map_single - a safety wrapper around pci_map_single()
 126  *
 127  * Same idea as ipath_map_page().
 128  */
 129 dma_addr_t ipath_map_single(struct pci_dev *hwdev, void *ptr, size_t size,
 130         int direction)
 131 {
 132         dma_addr_t phys;
 133
 134         phys = pci_map_single(hwdev, ptr, size, direction);
 135
 136         if (phys == 0) {
 137                 pci_unmap_single(hwdev, phys, size, direction);
 138                 phys = pci_map_single(hwdev, ptr, size, direction);
 139                 /*
 140                  * FIXME: If we get 0 again, we should keep this page,
 141                  * map another, then free the 0 page.
 142                  */
 143         }
 144
 145         return phys;
 146 }
 147
 148 /**
 149  * ipath_get_user_pages - lock user pages into memory
 150  * @start_page: the start page
 151  * @num_pages: the number of pages
 152  * @p: the output page structures
 153  *
 154  * This function takes a given start page (page aligned user virtual
 155  * address) and pins it and the following specified number of pages.  For
 156  * now, num_pages is always 1, but that will probably change at some point
 157  * (because caller is doing expected sends on a single virtually contiguous
 158  * buffer, so we can do all pages at once).
 159  */
 160 int ipath_get_user_pages(unsigned long start_page, size_t num_pages,
 161                          struct page **p)
 162 {
 163         int ret;
 164
 165         down_write(&current->mm->mmap_sem);
 166
 167         ret = __get_user_pages(start_page, num_pages, p, NULL);
 168
 169         up_write(&current->mm->mmap_sem);
 170
 171         return ret;
 172 }
 173
 174 /**
 175  * ipath_get_user_pages_nocopy - lock a single page for I/O and mark shared
 176  * @start_page: the page to lock
 177  * @p: the output page structure
 178  *
 179  * This is similar to ipath_get_user_pages, but it's always one page, and we
 180  * mark the page as locked for I/O, and shared.  This is used for the user
 181  * process page that contains the destination address for the rcvhdrq tail
 182  * update, so we need to have the vma. If we don't do this, the page can be
 183  * taken away from us on fork, even if the child never touches it, and then
 184  * the user process never sees the tail register updates.
 185  */
 186 int ipath_get_user_pages_nocopy(unsigned long page, struct page **p)
 187 {
 188         struct vm_area_struct *vma;
 189         int ret;
 190
 191         down_write(&current->mm->mmap_sem);
 192
 193         ret = __get_user_pages(page, 1, p, &vma);
 194
 195         up_write(&current->mm->mmap_sem);
 196
 197         return ret;
 198 }
 199
 200 void ipath_release_user_pages(struct page **p, size_t num_pages)
 201 {
 202         down_write(&current->mm->mmap_sem);
 203
 204         __ipath_release_user_pages(p, num_pages, 1);
 205
 206         current->mm->locked_vm -= num_pages;
 207
 208         up_write(&current->mm->mmap_sem);
 209 }
 210
 211 struct ipath_user_pages_work {
 212         struct work_struct work;
 213         struct mm_struct *mm;
 214         unsigned long num_pages;
 215 };
 216
 217 static void user_pages_account(struct work_struct *_work)
 218 {
 219         struct ipath_user_pages_work *work =
 220                 container_of(_work, struct ipath_user_pages_work, work);
 221
 222         down_write(&work->mm->mmap_sem);
 223         work->mm->locked_vm -= work->num_pages;
 224         up_write(&work->mm->mmap_sem);
 225         mmput(work->mm);
 226         kfree(work);
 227 }
 228
 229 void ipath_release_user_pages_on_close(struct page **p, size_t num_pages)
 230 {
 231         struct ipath_user_pages_work *work;
 232         struct mm_struct *mm;
 233
 234         __ipath_release_user_pages(p, num_pages, 1);
 235
 236         mm = get_task_mm(current);
 237         if (!mm)
 238                 goto bail;
 239
 240         work = kmalloc(sizeof(*work), GFP_KERNEL);
 241         if (!work)
 242                 goto bail_mm;
 243
 244         goto bail;
 245
 246         INIT_WORK(&work->work, user_pages_account);
 247         work->mm = mm;
 248         work->num_pages = num_pages;
 249
 250 bail_mm:
 251         mmput(mm);
 252 bail:
 253         return;
 254 }