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