mm/page_io.c

   1 /*
   2  *  linux/mm/page_io.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  *
   6  *  Swap reorganised 29.12.95,
   7  *  Asynchronous swapping added 30.12.95. Stephen Tweedie
   8  *  Removed race in async swapping. 14.4.1996. Bruno Haible
   9  *  Add swap of shared pages through the page cache. 20.2.1998. Stephen Tweedie
  10  *  Always use brw_page, life becomes simpler. 12 May 1998 Eric Biederman
  11  */
  12
  13 #include <linux/mm.h>
  14 #include <linux/kernel_stat.h>
  15 #include <linux/gfp.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/swap.h>
  18 #include <linux/bio.h>
  19 #include <linux/swapops.h>
  20 #include <linux/buffer_head.h>
  21 #include <linux/writeback.h>
  22 #include <linux/frontswap.h>
  23 #include <linux/blkdev.h>
  24 #include <linux/uio.h>
  25 #include <asm/pgtable.h>
  26
  27 static struct bio *get_swap_bio(gfp_t gfp_flags,
  28                                 struct page *page, bio_end_io_t end_io)
  29 {
  30         struct bio *bio;
  31
  32         bio = bio_alloc(gfp_flags, 1);
  33         if (bio) {
  34                 bio->bi_iter.bi_sector = map_swap_page(page, &bio->bi_bdev);
  35                 bio->bi_iter.bi_sector <<= PAGE_SHIFT - 9;
  36                 bio->bi_io_vec[0].bv_page = page;
  37                 bio->bi_io_vec[0].bv_len = PAGE_SIZE;
  38                 bio->bi_io_vec[0].bv_offset = 0;
  39                 bio->bi_vcnt = 1;
  40                 bio->bi_iter.bi_size = PAGE_SIZE;
  41                 bio->bi_end_io = end_io;
  42         }
  43         return bio;
  44 }
  45
  46 void end_swap_bio_write(struct bio *bio)
  47 {
  48         struct page *page = bio->bi_io_vec[0].bv_page;
  49
  50         if (bio->bi_error) {
  51                 SetPageError(page);
  52                 /*
  53                  * We failed to write the page out to swap-space.
  54                  * Re-dirty the page in order to avoid it being reclaimed.
  55                  * Also print a dire warning that things will go BAD (tm)
  56                  * very quickly.
  57                  *
  58                  * Also clear PG_reclaim to avoid rotate_reclaimable_page()
  59                  */
  60                 set_page_dirty(page);
  61                 printk(KERN_ALERT "Write-error on swap-device (%u:%u:%Lu)\n",
  62                                 imajor(bio->bi_bdev->bd_inode),
  63                                 iminor(bio->bi_bdev->bd_inode),
  64                                 (unsigned long long)bio->bi_iter.bi_sector);
  65                 ClearPageReclaim(page);
  66         }
  67         end_page_writeback(page);
  68         bio_put(bio);
  69 }
  70
  71 static void end_swap_bio_read(struct bio *bio)
  72 {
  73         struct page *page = bio->bi_io_vec[0].bv_page;
  74
  75         if (bio->bi_error) {
  76                 SetPageError(page);
  77                 ClearPageUptodate(page);
  78                 printk(KERN_ALERT "Read-error on swap-device (%u:%u:%Lu)\n",
  79                                 imajor(bio->bi_bdev->bd_inode),
  80                                 iminor(bio->bi_bdev->bd_inode),
  81                                 (unsigned long long)bio->bi_iter.bi_sector);
  82                 goto out;
  83         }
  84
  85         SetPageUptodate(page);
  86
  87         /*
  88          * There is no guarantee that the page is in swap cache - the software
  89          * suspend code (at least) uses end_swap_bio_read() against a non-
  90          * swapcache page.  So we must check PG_swapcache before proceeding with
  91          * this optimization.
  92          */
  93         if (likely(PageSwapCache(page))) {
  94                 struct swap_info_struct *sis;
  95
  96                 sis = page_swap_info(page);
  97                 if (sis->flags & SWP_BLKDEV) {
  98                         /*
  99                          * The swap subsystem performs lazy swap slot freeing,
 100                          * expecting that the page will be swapped out again.
 101                          * So we can avoid an unnecessary write if the page
 102                          * isn't redirtied.
 103                          * This is good for real swap storage because we can
 104                          * reduce unnecessary I/O and enhance wear-leveling
 105                          * if an SSD is used as the as swap device.
 106                          * But if in-memory swap device (eg zram) is used,
 107                          * this causes a duplicated copy between uncompressed
 108                          * data in VM-owned memory and compressed data in
 109                          * zram-owned memory.  So let's free zram-owned memory
 110                          * and make the VM-owned decompressed page *dirty*,
 111                          * so the page should be swapped out somewhere again if
 112                          * we again wish to reclaim it.
 113                          */
 114                         struct gendisk *disk = sis->bdev->bd_disk;
 115                         if (disk->fops->swap_slot_free_notify) {
 116                                 swp_entry_t entry;
 117                                 unsigned long offset;
 118
 119                                 entry.val = page_private(page);
 120                                 offset = swp_offset(entry);
 121
 122                                 SetPageDirty(page);
 123                                 disk->fops->swap_slot_free_notify(sis->bdev,
 124                                                 offset);
 125                         }
 126                 }
 127         }
 128
 129 out:
 130         unlock_page(page);
 131         bio_put(bio);
 132 }
 133
 134 int generic_swapfile_activate(struct swap_info_struct *sis,
 135                                 struct file *swap_file,
 136                                 sector_t *span)
 137 {
 138         struct address_space *mapping = swap_file->f_mapping;
 139         struct inode *inode = mapping->host;
 140         unsigned blocks_per_page;
 141         unsigned long page_no;
 142         unsigned blkbits;
 143         sector_t probe_block;
 144         sector_t last_block;
 145         sector_t lowest_block = -1;
 146         sector_t highest_block = 0;
 147         int nr_extents = 0;
 148         int ret;
 149
 150         blkbits = inode->i_blkbits;
 151         blocks_per_page = PAGE_SIZE >> blkbits;
 152
 153         /*
 154          * Map all the blocks into the extent list.  This code doesn't try
 155          * to be very smart.
 156          */
 157         probe_block = 0;
 158         page_no = 0;
 159         last_block = i_size_read(inode) >> blkbits;
 160         while ((probe_block + blocks_per_page) <= last_block &&
 161                         page_no < sis->max) {
 162                 unsigned block_in_page;
 163                 sector_t first_block;
 164
 165                 first_block = bmap(inode, probe_block);
 166                 if (first_block == 0)
 167                         goto bad_bmap;
 168
 169                 /*
 170                  * It must be PAGE_SIZE aligned on-disk
 171                  */
 172                 if (first_block & (blocks_per_page - 1)) {
 173                         probe_block++;
 174                         goto reprobe;
 175                 }
 176
 177                 for (block_in_page = 1; block_in_page < blocks_per_page;
 178                                         block_in_page++) {
 179                         sector_t block;
 180
 181                         block = bmap(inode, probe_block + block_in_page);
 182                         if (block == 0)
 183                                 goto bad_bmap;
 184                         if (block != first_block + block_in_page) {
 185                                 /* Discontiguity */
 186                                 probe_block++;
 187                                 goto reprobe;
 188                         }
 189                 }
 190
 191                 first_block >>= (PAGE_SHIFT - blkbits);
 192                 if (page_no) {  /* exclude the header page */
 193                         if (first_block < lowest_block)
 194                                 lowest_block = first_block;
 195                         if (first_block > highest_block)
 196                                 highest_block = first_block;
 197                 }
 198
 199                 /*
 200                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
 201                  */
 202                 ret = add_swap_extent(sis, page_no, 1, first_block);
 203                 if (ret < 0)
 204                         goto out;
 205                 nr_extents += ret;
 206                 page_no++;
 207                 probe_block += blocks_per_page;
 208 reprobe:
 209                 continue;
 210         }
 211         ret = nr_extents;
 212         *span = 1 + highest_block - lowest_block;
 213         if (page_no == 0)
 214                 page_no = 1;    /* force Empty message */
 215         sis->max = page_no;
 216         sis->pages = page_no - 1;
 217         sis->highest_bit = page_no - 1;
 218 out:
 219         return ret;
 220 bad_bmap:
 221         printk(KERN_ERR "swapon: swapfile has holes\n");
 222         ret = -EINVAL;
 223         goto out;
 224 }
 225
 226 /*
 227  * We may have stale swap cache pages in memory: notice
 228  * them here and get rid of the unnecessary final write.
 229  */
 230 int swap_writepage(struct page *page, struct writeback_control *wbc)
 231 {
 232         int ret = 0;
 233
 234         if (try_to_free_swap(page)) {
 235                 unlock_page(page);
 236                 goto out;
 237         }
 238         if (frontswap_store(page) == 0) {
 239                 set_page_writeback(page);
 240                 unlock_page(page);
 241                 end_page_writeback(page);
 242                 goto out;
 243         }
 244         ret = __swap_writepage(page, wbc, end_swap_bio_write);
 245 out:
 246         return ret;
 247 }
 248
 249 static sector_t swap_page_sector(struct page *page)
 250 {
 251         return (sector_t)__page_file_index(page) << (PAGE_CACHE_SHIFT - 9);
 252 }
 253
 254 int __swap_writepage(struct page *page, struct writeback_control *wbc,
 255                 bio_end_io_t end_write_func)
 256 {
 257         struct bio *bio;
 258         int ret, rw = WRITE;
 259         struct swap_info_struct *sis = page_swap_info(page);
 260
 261         if (sis->flags & SWP_FILE) {
 262                 struct kiocb kiocb;
 263                 struct file *swap_file = sis->swap_file;
 264                 struct address_space *mapping = swap_file->f_mapping;
 265                 struct bio_vec bv = {
 266                         .bv_page = page,
 267                         .bv_len  = PAGE_SIZE,
 268                         .bv_offset = 0
 269                 };
 270                 struct iov_iter from;
 271
 272                 iov_iter_bvec(&from, ITER_BVEC | WRITE, &bv, 1, PAGE_SIZE);
 273                 init_sync_kiocb(&kiocb, swap_file);
 274                 kiocb.ki_pos = page_file_offset(page);
 275
 276                 set_page_writeback(page);
 277                 unlock_page(page);
 278                 ret = mapping->a_ops->direct_IO(&kiocb, &from, kiocb.ki_pos);
 279                 if (ret == PAGE_SIZE) {
 280                         count_vm_event(PSWPOUT);
 281                         ret = 0;
 282                 } else {
 283                         /*
 284                          * In the case of swap-over-nfs, this can be a
 285                          * temporary failure if the system has limited
 286                          * memory for allocating transmit buffers.
 287                          * Mark the page dirty and avoid
 288                          * rotate_reclaimable_page but rate-limit the
 289                          * messages but do not flag PageError like
 290                          * the normal direct-to-bio case as it could
 291                          * be temporary.
 292                          */
 293                         set_page_dirty(page);
 294                         ClearPageReclaim(page);
 295                         pr_err_ratelimited("Write error on dio swapfile (%Lu)\n",
 296                                 page_file_offset(page));
 297                 }
 298                 end_page_writeback(page);
 299                 return ret;
 300         }
 301
 302         ret = bdev_write_page(sis->bdev, swap_page_sector(page), page, wbc);
 303         if (!ret) {
 304                 count_vm_event(PSWPOUT);
 305                 return 0;
 306         }
 307
 308         ret = 0;
 309         bio = get_swap_bio(GFP_NOIO, page, end_write_func);
 310         if (bio == NULL) {
 311                 set_page_dirty(page);
 312                 unlock_page(page);
 313                 ret = -ENOMEM;
 314                 goto out;
 315         }
 316         if (wbc->sync_mode == WB_SYNC_ALL)
 317                 rw |= REQ_SYNC;
 318         count_vm_event(PSWPOUT);
 319         set_page_writeback(page);
 320         unlock_page(page);
 321         submit_bio(rw, bio);
 322 out:
 323         return ret;
 324 }
 325
 326 int swap_readpage(struct page *page)
 327 {
 328         struct bio *bio;
 329         int ret = 0;
 330         struct swap_info_struct *sis = page_swap_info(page);
 331
 332         VM_BUG_ON_PAGE(!PageLocked(page), page);
 333         VM_BUG_ON_PAGE(PageUptodate(page), page);
 334         if (frontswap_load(page) == 0) {
 335                 SetPageUptodate(page);
 336                 unlock_page(page);
 337                 goto out;
 338         }
 339
 340         if (sis->flags & SWP_FILE) {
 341                 struct file *swap_file = sis->swap_file;
 342                 struct address_space *mapping = swap_file->f_mapping;
 343
 344                 ret = mapping->a_ops->readpage(swap_file, page);
 345                 if (!ret)
 346                         count_vm_event(PSWPIN);
 347                 return ret;
 348         }
 349
 350         ret = bdev_read_page(sis->bdev, swap_page_sector(page), page);
 351         if (!ret) {
 352                 count_vm_event(PSWPIN);
 353                 return 0;
 354         }
 355
 356         ret = 0;
 357         bio = get_swap_bio(GFP_KERNEL, page, end_swap_bio_read);
 358         if (bio == NULL) {
 359                 unlock_page(page);
 360                 ret = -ENOMEM;
 361                 goto out;
 362         }
 363         count_vm_event(PSWPIN);
 364         submit_bio(READ, bio);
 365 out:
 366         return ret;
 367 }
 368
 369 int swap_set_page_dirty(struct page *page)
 370 {
 371         struct swap_info_struct *sis = page_swap_info(page);
 372
 373         if (sis->flags & SWP_FILE) {
 374                 struct address_space *mapping = sis->swap_file->f_mapping;
 375                 return mapping->a_ops->set_page_dirty(page);
 376         } else {
 377                 return __set_page_dirty_no_writeback(page);
 378         }
 379 }