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, int err)
  47 {
  48         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
  49         struct page *page = bio->bi_io_vec[0].bv_page;
  50
  51         if (!uptodate) {
  52                 SetPageError(page);
  53                 /*
  54                  * We failed to write the page out to swap-space.
  55                  * Re-dirty the page in order to avoid it being reclaimed.
  56                  * Also print a dire warning that things will go BAD (tm)
  57                  * very quickly.
  58                  *
  59                  * Also clear PG_reclaim to avoid rotate_reclaimable_page()
  60                  */
  61                 set_page_dirty(page);
  62                 printk(KERN_ALERT "Write-error on swap-device (%u:%u:%Lu)\n",
  63                                 imajor(bio->bi_bdev->bd_inode),
  64                                 iminor(bio->bi_bdev->bd_inode),
  65                                 (unsigned long long)bio->bi_iter.bi_sector);
  66                 ClearPageReclaim(page);
  67         }
  68         end_page_writeback(page);
  69         bio_put(bio);
  70 }
  71
  72 void end_swap_bio_read(struct bio *bio, int err)
  73 {
  74         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
  75         struct page *page = bio->bi_io_vec[0].bv_page;
  76
  77         if (!uptodate) {
  78                 SetPageError(page);
  79                 ClearPageUptodate(page);
  80                 printk(KERN_ALERT "Read-error on swap-device (%u:%u:%Lu)\n",
  81                                 imajor(bio->bi_bdev->bd_inode),
  82                                 iminor(bio->bi_bdev->bd_inode),
  83                                 (unsigned long long)bio->bi_iter.bi_sector);
  84                 goto out;
  85         }
  86
  87         SetPageUptodate(page);
  88
  89         /*
  90          * There is no guarantee that the page is in swap cache - the software
  91          * suspend code (at least) uses end_swap_bio_read() against a non-
  92          * swapcache page.  So we must check PG_swapcache before proceeding with
  93          * this optimization.
  94          */
  95         if (likely(PageSwapCache(page))) {
  96                 struct swap_info_struct *sis;
  97
  98                 sis = page_swap_info(page);
  99                 if (sis->flags & SWP_BLKDEV) {
 100                         /*
 101                          * The swap subsystem performs lazy swap slot freeing,
 102                          * expecting that the page will be swapped out again.
 103                          * So we can avoid an unnecessary write if the page
 104                          * isn't redirtied.
 105                          * This is good for real swap storage because we can
 106                          * reduce unnecessary I/O and enhance wear-leveling
 107                          * if an SSD is used as the as swap device.
 108                          * But if in-memory swap device (eg zram) is used,
 109                          * this causes a duplicated copy between uncompressed
 110                          * data in VM-owned memory and compressed data in
 111                          * zram-owned memory.  So let's free zram-owned memory
 112                          * and make the VM-owned decompressed page *dirty*,
 113                          * so the page should be swapped out somewhere again if
 114                          * we again wish to reclaim it.
 115                          */
 116                         struct gendisk *disk = sis->bdev->bd_disk;
 117                         if (disk->fops->swap_slot_free_notify) {
 118                                 swp_entry_t entry;
 119                                 unsigned long offset;
 120
 121                                 entry.val = page_private(page);
 122                                 offset = swp_offset(entry);
 123
 124                                 SetPageDirty(page);
 125                                 disk->fops->swap_slot_free_notify(sis->bdev,
 126                                                 offset);
 127                         }
 128                 }
 129         }
 130
 131 out:
 132         unlock_page(page);
 133         bio_put(bio);
 134 }
 135
 136 int generic_swapfile_activate(struct swap_info_struct *sis,
 137                                 struct file *swap_file,
 138                                 sector_t *span)
 139 {
 140         struct address_space *mapping = swap_file->f_mapping;
 141         struct inode *inode = mapping->host;
 142         unsigned blocks_per_page;
 143         unsigned long page_no;
 144         unsigned blkbits;
 145         sector_t probe_block;
 146         sector_t last_block;
 147         sector_t lowest_block = -1;
 148         sector_t highest_block = 0;
 149         int nr_extents = 0;
 150         int ret;
 151
 152         blkbits = inode->i_blkbits;
 153         blocks_per_page = PAGE_SIZE >> blkbits;
 154
 155         /*
 156          * Map all the blocks into the extent list.  This code doesn't try
 157          * to be very smart.
 158          */
 159         probe_block = 0;
 160         page_no = 0;
 161         last_block = i_size_read(inode) >> blkbits;
 162         while ((probe_block + blocks_per_page) <= last_block &&
 163                         page_no < sis->max) {
 164                 unsigned block_in_page;
 165                 sector_t first_block;
 166
 167                 first_block = bmap(inode, probe_block);
 168                 if (first_block == 0)
 169                         goto bad_bmap;
 170
 171                 /*
 172                  * It must be PAGE_SIZE aligned on-disk
 173                  */
 174                 if (first_block & (blocks_per_page - 1)) {
 175                         probe_block++;
 176                         goto reprobe;
 177                 }
 178
 179                 for (block_in_page = 1; block_in_page < blocks_per_page;
 180                                         block_in_page++) {
 181                         sector_t block;
 182
 183                         block = bmap(inode, probe_block + block_in_page);
 184                         if (block == 0)
 185                                 goto bad_bmap;
 186                         if (block != first_block + block_in_page) {
 187                                 /* Discontiguity */
 188                                 probe_block++;
 189                                 goto reprobe;
 190                         }
 191                 }
 192
 193                 first_block >>= (PAGE_SHIFT - blkbits);
 194                 if (page_no) {  /* exclude the header page */
 195                         if (first_block < lowest_block)
 196                                 lowest_block = first_block;
 197                         if (first_block > highest_block)
 198                                 highest_block = first_block;
 199                 }
 200
 201                 /*
 202                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
 203                  */
 204                 ret = add_swap_extent(sis, page_no, 1, first_block);
 205                 if (ret < 0)
 206                         goto out;
 207                 nr_extents += ret;
 208                 page_no++;
 209                 probe_block += blocks_per_page;
 210 reprobe:
 211                 continue;
 212         }
 213         ret = nr_extents;
 214         *span = 1 + highest_block - lowest_block;
 215         if (page_no == 0)
 216                 page_no = 1;    /* force Empty message */
 217         sis->max = page_no;
 218         sis->pages = page_no - 1;
 219         sis->highest_bit = page_no - 1;
 220 out:
 221         return ret;
 222 bad_bmap:
 223         printk(KERN_ERR "swapon: swapfile has holes\n");
 224         ret = -EINVAL;
 225         goto out;
 226 }
 227
 228 /*
 229  * We may have stale swap cache pages in memory: notice
 230  * them here and get rid of the unnecessary final write.
 231  */
 232 int swap_writepage(struct page *page, struct writeback_control *wbc)
 233 {
 234         int ret = 0;
 235
 236         if (try_to_free_swap(page)) {
 237                 unlock_page(page);
 238                 goto out;
 239         }
 240         if (frontswap_store(page) == 0) {
 241                 set_page_writeback(page);
 242                 unlock_page(page);
 243                 end_page_writeback(page);
 244                 goto out;
 245         }
 246         ret = __swap_writepage(page, wbc, end_swap_bio_write);
 247 out:
 248         return ret;
 249 }
 250
 251 static sector_t swap_page_sector(struct page *page)
 252 {
 253         return (sector_t)__page_file_index(page) << (PAGE_CACHE_SHIFT - 9);
 254 }
 255
 256 int __swap_writepage(struct page *page, struct writeback_control *wbc,
 257         void (*end_write_func)(struct bio *, int))
 258 {
 259         struct bio *bio;
 260         int ret, rw = WRITE;
 261         struct swap_info_struct *sis = page_swap_info(page);
 262
 263         if (sis->flags & SWP_FILE) {
 264                 struct kiocb kiocb;
 265                 struct file *swap_file = sis->swap_file;
 266                 struct address_space *mapping = swap_file->f_mapping;
 267                 struct bio_vec bv = {
 268                         .bv_page = page,
 269                         .bv_len  = PAGE_SIZE,
 270                         .bv_offset = 0
 271                 };
 272                 struct iov_iter from;
 273
 274                 iov_iter_bvec(&from, ITER_BVEC | WRITE, &bv, 1, PAGE_SIZE);
 275                 init_sync_kiocb(&kiocb, swap_file);
 276                 kiocb.ki_pos = page_file_offset(page);
 277
 278                 set_page_writeback(page);
 279                 unlock_page(page);
 280                 ret = mapping->a_ops->direct_IO(&kiocb, &from, kiocb.ki_pos);
 281                 if (ret == PAGE_SIZE) {
 282                         count_vm_event(PSWPOUT);
 283                         ret = 0;
 284                 } else {
 285                         /*
 286                          * In the case of swap-over-nfs, this can be a
 287                          * temporary failure if the system has limited
 288                          * memory for allocating transmit buffers.
 289                          * Mark the page dirty and avoid
 290                          * rotate_reclaimable_page but rate-limit the
 291                          * messages but do not flag PageError like
 292                          * the normal direct-to-bio case as it could
 293                          * be temporary.
 294                          */
 295                         set_page_dirty(page);
 296                         ClearPageReclaim(page);
 297                         pr_err_ratelimited("Write error on dio swapfile (%Lu)\n",
 298                                 page_file_offset(page));
 299                 }
 300                 end_page_writeback(page);
 301                 return ret;
 302         }
 303
 304         ret = bdev_write_page(sis->bdev, swap_page_sector(page), page, wbc);
 305         if (!ret) {
 306                 count_vm_event(PSWPOUT);
 307                 return 0;
 308         }
 309
 310         ret = 0;
 311         bio = get_swap_bio(GFP_NOIO, page, end_write_func);
 312         if (bio == NULL) {
 313                 set_page_dirty(page);
 314                 unlock_page(page);
 315                 ret = -ENOMEM;
 316                 goto out;
 317         }
 318         if (wbc->sync_mode == WB_SYNC_ALL)
 319                 rw |= REQ_SYNC;
 320         count_vm_event(PSWPOUT);
 321         set_page_writeback(page);
 322         unlock_page(page);
 323         submit_bio(rw, bio);
 324 out:
 325         return ret;
 326 }
 327
 328 int swap_readpage(struct page *page)
 329 {
 330         struct bio *bio;
 331         int ret = 0;
 332         struct swap_info_struct *sis = page_swap_info(page);
 333
 334         VM_BUG_ON_PAGE(!PageLocked(page), page);
 335         VM_BUG_ON_PAGE(PageUptodate(page), page);
 336         if (frontswap_load(page) == 0) {
 337                 SetPageUptodate(page);
 338                 unlock_page(page);
 339                 goto out;
 340         }
 341
 342         if (sis->flags & SWP_FILE) {
 343                 struct file *swap_file = sis->swap_file;
 344                 struct address_space *mapping = swap_file->f_mapping;
 345
 346                 ret = mapping->a_ops->readpage(swap_file, page);
 347                 if (!ret)
 348                         count_vm_event(PSWPIN);
 349                 return ret;
 350         }
 351
 352         ret = bdev_read_page(sis->bdev, swap_page_sector(page), page);
 353         if (!ret) {
 354                 count_vm_event(PSWPIN);
 355                 return 0;
 356         }
 357
 358         ret = 0;
 359         bio = get_swap_bio(GFP_KERNEL, page, end_swap_bio_read);
 360         if (bio == NULL) {
 361                 unlock_page(page);
 362                 ret = -ENOMEM;
 363                 goto out;
 364         }
 365         count_vm_event(PSWPIN);
 366         submit_bio(READ, bio);
 367 out:
 368         return ret;
 369 }
 370
 371 int swap_set_page_dirty(struct page *page)
 372 {
 373         struct swap_info_struct *sis = page_swap_info(page);
 374
 375         if (sis->flags & SWP_FILE) {
 376                 struct address_space *mapping = sis->swap_file->f_mapping;
 377                 return mapping->a_ops->set_page_dirty(page);
 378         } else {
 379                 return __set_page_dirty_no_writeback(page);
 380         }
 381 }