drivers/infiniband/hw/ipath/ipath_user_pages.c

   1 /*
   2  * Copyright (c) 2006, 2007 QLogic Corporation. 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 #include <linux/sched.h>
  37
  38 #include "ipath_kernel.h"
  39
  40 static void __ipath_release_user_pages(struct page **p, size_t num_pages,
  41                                    int dirty)
  42 {
  43         size_t i;
  44
  45         for (i = 0; i < num_pages; i++) {
  46                 ipath_cdbg(MM, "%lu/%lu put_page %p\n", (unsigned long) i,
  47                            (unsigned long) num_pages, p[i]);
  48                 if (dirty)
  49                         set_page_dirty_lock(p[i]);
  50                 put_page(p[i]);
  51         }
  52 }
  53
  54 /* call with current->mm->mmap_sem held */
  55 static int __get_user_pages(unsigned long start_page, size_t num_pages,
  56                         struct page **p, struct vm_area_struct **vma)
  57 {
  58         unsigned long lock_limit;
  59         size_t got;
  60         int ret;
  61
  62         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur >>
  63                 PAGE_SHIFT;
  64
  65         if (num_pages > lock_limit) {
  66                 ret = -ENOMEM;
  67                 goto bail;
  68         }
  69
  70         ipath_cdbg(VERBOSE, "pin %lx pages from vaddr %lx\n",
  71                    (unsigned long) num_pages, start_page);
  72
  73         for (got = 0; got < num_pages; got += ret) {
  74                 ret = get_user_pages(current, current->mm,
  75                                      start_page + got * PAGE_SIZE,
  76                                      num_pages - got, 1, 1,
  77                                      p + got, vma);
  78                 if (ret < 0)
  79                         goto bail_release;
  80         }
  81
  82         current->mm->locked_vm += num_pages;
  83
  84         ret = 0;
  85         goto bail;
  86
  87 bail_release:
  88         __ipath_release_user_pages(p, got, 0);
  89 bail:
  90         return ret;
  91 }
  92
  93 /**
  94  * ipath_map_page - a safety wrapper around pci_map_page()
  95  *
  96  * A dma_addr of all 0's is interpreted by the chip as "disabled".
  97  * Unfortunately, it can also be a valid dma_addr returned on some
  98  * architectures.
  99  *
 100  * The powerpc iommu assigns dma_addrs in ascending order, so we don't
 101  * have to bother with retries or mapping a dummy page to insure we
 102  * don't just get the same mapping again.
 103  *
 104  * I'm sure we won't be so lucky with other iommu's, so FIXME.
 105  */
 106 dma_addr_t ipath_map_page(struct pci_dev *hwdev, struct page *page,
 107         unsigned long offset, size_t size, int direction)
 108 {
 109         dma_addr_t phys;
 110
 111         phys = pci_map_page(hwdev, page, offset, size, direction);
 112
 113         if (phys == 0) {
 114                 pci_unmap_page(hwdev, phys, size, direction);
 115                 phys = pci_map_page(hwdev, page, offset, size, direction);
 116                 /*
 117                  * FIXME: If we get 0 again, we should keep this page,
 118                  * map another, then free the 0 page.
 119                  */
 120         }
 121
 122         return phys;
 123 }
 124
 125 /**
 126  * ipath_map_single - a safety wrapper around pci_map_single()
 127  *
 128  * Same idea as ipath_map_page().
 129  */
 130 dma_addr_t ipath_map_single(struct pci_dev *hwdev, void *ptr, size_t size,
 131         int direction)
 132 {
 133         dma_addr_t phys;
 134
 135         phys = pci_map_single(hwdev, ptr, size, direction);
 136
 137         if (phys == 0) {
 138                 pci_unmap_single(hwdev, phys, size, direction);
 139                 phys = pci_map_single(hwdev, ptr, size, direction);
 140                 /*
 141                  * FIXME: If we get 0 again, we should keep this page,
 142                  * map another, then free the 0 page.
 143                  */
 144         }
 145
 146         return phys;
 147 }
 148
 149 /**
 150  * ipath_get_user_pages - lock user pages into memory
 151  * @start_page: the start page
 152  * @num_pages: the number of pages
 153  * @p: the output page structures
 154  *
 155  * This function takes a given start page (page aligned user virtual
 156  * address) and pins it and the following specified number of pages.  For
 157  * now, num_pages is always 1, but that will probably change at some point
 158  * (because caller is doing expected sends on a single virtually contiguous
 159  * buffer, so we can do all pages at once).
 160  */
 161 int ipath_get_user_pages(unsigned long start_page, size_t num_pages,
 162                          struct page **p)
 163 {
 164         int ret;
 165
 166         down_write(&current->mm->mmap_sem);
 167
 168         ret = __get_user_pages(start_page, num_pages, p, NULL);
 169
 170         up_write(&current->mm->mmap_sem);
 171
 172         return ret;
 173 }
 174
 175 void ipath_release_user_pages(struct page **p, size_t num_pages)
 176 {
 177         down_write(&current->mm->mmap_sem);
 178
 179         __ipath_release_user_pages(p, num_pages, 1);
 180
 181         current->mm->locked_vm -= num_pages;
 182
 183         up_write(&current->mm->mmap_sem);
 184 }
 185
 186 struct ipath_user_pages_work {
 187         struct work_struct work;
 188         struct mm_struct *mm;
 189         unsigned long num_pages;
 190 };
 191
 192 static void user_pages_account(struct work_struct *_work)
 193 {
 194         struct ipath_user_pages_work *work =
 195                 container_of(_work, struct ipath_user_pages_work, work);
 196
 197         down_write(&work->mm->mmap_sem);
 198         work->mm->locked_vm -= work->num_pages;
 199         up_write(&work->mm->mmap_sem);
 200         mmput(work->mm);
 201         kfree(work);
 202 }
 203
 204 void ipath_release_user_pages_on_close(struct page **p, size_t num_pages)
 205 {
 206         struct ipath_user_pages_work *work;
 207         struct mm_struct *mm;
 208
 209         __ipath_release_user_pages(p, num_pages, 1);
 210
 211         mm = get_task_mm(current);
 212         if (!mm)
 213                 return;
 214
 215         work = kmalloc(sizeof(*work), GFP_KERNEL);
 216         if (!work)
 217                 goto bail_mm;
 218
 219         INIT_WORK(&work->work, user_pages_account);
 220         work->mm = mm;
 221         work->num_pages = num_pages;
 222
 223         schedule_work(&work->work);
 224         return;
 225
 226 bail_mm:
 227         mmput(mm);
 228         return;
 229 }