mm/fadvise.c

   1 /*
   2  * mm/fadvise.c
   3  *
   4  * Copyright (C) 2002, Linus Torvalds
   5  *
   6  * 11Jan2003    Andrew Morton
   7  *              Initial version.
   8  */
   9
  10 #include <linux/kernel.h>
  11 #include <linux/file.h>
  12 #include <linux/fs.h>
  13 #include <linux/mm.h>
  14 #include <linux/pagemap.h>
  15 #include <linux/backing-dev.h>
  16 #include <linux/pagevec.h>
  17 #include <linux/fadvise.h>
  18 #include <linux/writeback.h>
  19 #include <linux/syscalls.h>
  20 #include <linux/swap.h>
  21
  22 #include <asm/unistd.h>
  23
  24 /*
  25  * POSIX_FADV_WILLNEED could set PG_Referenced, and POSIX_FADV_NOREUSE could
  26  * deactivate the pages and clear PG_Referenced.
  27  */
  28 SYSCALL_DEFINE4(fadvise64_64, int, fd, loff_t, offset, loff_t, len, int, advice)
  29 {
  30         struct fd f = fdget(fd);
  31         struct address_space *mapping;
  32         struct backing_dev_info *bdi;
  33         loff_t endbyte;                 /* inclusive */
  34         pgoff_t start_index;
  35         pgoff_t end_index;
  36         unsigned long nrpages;
  37         int ret = 0;
  38
  39         if (!f.file)
  40                 return -EBADF;
  41
  42         if (S_ISFIFO(file_inode(f.file)->i_mode)) {
  43                 ret = -ESPIPE;
  44                 goto out;
  45         }
  46
  47         mapping = f.file->f_mapping;
  48         if (!mapping || len < 0) {
  49                 ret = -EINVAL;
  50                 goto out;
  51         }
  52
  53         if (mapping->a_ops->get_xip_mem) {
  54                 switch (advice) {
  55                 case POSIX_FADV_NORMAL:
  56                 case POSIX_FADV_RANDOM:
  57                 case POSIX_FADV_SEQUENTIAL:
  58                 case POSIX_FADV_WILLNEED:
  59                 case POSIX_FADV_NOREUSE:
  60                 case POSIX_FADV_DONTNEED:
  61                         /* no bad return value, but ignore advice */
  62                         break;
  63                 default:
  64                         ret = -EINVAL;
  65                 }
  66                 goto out;
  67         }
  68
  69         /* Careful about overflows. Len == 0 means "as much as possible" */
  70         endbyte = offset + len;
  71         if (!len || endbyte < len)
  72                 endbyte = -1;
  73         else
  74                 endbyte--;              /* inclusive */
  75
  76         bdi = mapping->backing_dev_info;
  77
  78         switch (advice) {
  79         case POSIX_FADV_NORMAL:
  80                 f.file->f_ra.ra_pages = bdi->ra_pages;
  81                 spin_lock(&f.file->f_lock);
  82                 f.file->f_mode &= ~FMODE_RANDOM;
  83                 spin_unlock(&f.file->f_lock);
  84                 break;
  85         case POSIX_FADV_RANDOM:
  86                 spin_lock(&f.file->f_lock);
  87                 f.file->f_mode |= FMODE_RANDOM;
  88                 spin_unlock(&f.file->f_lock);
  89                 break;
  90         case POSIX_FADV_SEQUENTIAL:
  91                 f.file->f_ra.ra_pages = bdi->ra_pages * 2;
  92                 spin_lock(&f.file->f_lock);
  93                 f.file->f_mode &= ~FMODE_RANDOM;
  94                 spin_unlock(&f.file->f_lock);
  95                 break;
  96         case POSIX_FADV_WILLNEED:
  97                 /* First and last PARTIAL page! */
  98                 start_index = offset >> PAGE_CACHE_SHIFT;
  99                 end_index = endbyte >> PAGE_CACHE_SHIFT;
 100
 101                 /* Careful about overflow on the "+1" */
 102                 nrpages = end_index - start_index + 1;
 103                 if (!nrpages)
 104                         nrpages = ~0UL;
 105
 106                 /*
 107                  * Ignore return value because fadvise() shall return
 108                  * success even if filesystem can't retrieve a hint,
 109                  */
 110                 force_page_cache_readahead(mapping, f.file, start_index,
 111                                            nrpages);
 112                 break;
 113         case POSIX_FADV_NOREUSE:
 114                 break;
 115         case POSIX_FADV_DONTNEED:
 116                 if (!bdi_write_congested(mapping->backing_dev_info))
 117                         __filemap_fdatawrite_range(mapping, offset, endbyte,
 118                                                    WB_SYNC_NONE);
 119
 120                 /*
 121                  * First and last FULL page! Partial pages are deliberately
 122                  * preserved on the expectation that it is better to preserve
 123                  * needed memory than to discard unneeded memory.
 124                  */
 125                 start_index = (offset+(PAGE_CACHE_SIZE-1)) >> PAGE_CACHE_SHIFT;
 126                 end_index = (endbyte >> PAGE_CACHE_SHIFT);
 127
 128                 if (end_index >= start_index) {
 129                         unsigned long count = invalidate_mapping_pages(mapping,
 130                                                 start_index, end_index);
 131
 132                         /*
 133                          * If fewer pages were invalidated than expected then
 134                          * it is possible that some of the pages were on
 135                          * a per-cpu pagevec for a remote CPU. Drain all
 136                          * pagevecs and try again.
 137                          */
 138                         if (count < (end_index - start_index + 1)) {
 139                                 lru_add_drain_all();
 140                                 invalidate_mapping_pages(mapping, start_index,
 141                                                 end_index);
 142                         }
 143                 }
 144                 break;
 145         default:
 146                 ret = -EINVAL;
 147         }
 148 out:
 149         fdput(f);
 150         return ret;
 151 }
 152
 153 #ifdef __ARCH_WANT_SYS_FADVISE64
 154
 155 SYSCALL_DEFINE4(fadvise64, int, fd, loff_t, offset, size_t, len, int, advice)
 156 {
 157         return sys_fadvise64_64(fd, offset, len, advice);
 158 }
 159
 160 #endif