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                 pr_alert("Write-error on swap-device (%u:%u:%llu)\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 swap_slot_free_notify(struct page *page)
  69 {
  70         struct swap_info_struct *sis;
  71         struct gendisk *disk;
  72
  73         /*
  74          * There is no guarantee that the page is in swap cache - the software
  75          * suspend code (at least) uses end_swap_bio_read() against a non-
  76          * swapcache page.  So we must check PG_swapcache before proceeding with
  77          * this optimization.
  78          */
  79         if (unlikely(!PageSwapCache(page)))
  80                 return;
  81
  82         sis = page_swap_info(page);
  83         if (!(sis->flags & SWP_BLKDEV))
  84                 return;
  85
  86         /*
  87          * The swap subsystem performs lazy swap slot freeing,
  88          * expecting that the page will be swapped out again.
  89          * So we can avoid an unnecessary write if the page
  90          * isn't redirtied.
  91          * This is good for real swap storage because we can
  92          * reduce unnecessary I/O and enhance wear-leveling
  93          * if an SSD is used as the as swap device.
  94          * But if in-memory swap device (eg zram) is used,
  95          * this causes a duplicated copy between uncompressed
  96          * data in VM-owned memory and compressed data in
  97          * zram-owned memory.  So let's free zram-owned memory
  98          * and make the VM-owned decompressed page *dirty*,
  99          * so the page should be swapped out somewhere again if
 100          * we again wish to reclaim it.
 101          */
 102         disk = sis->bdev->bd_disk;
 103         if (disk->fops->swap_slot_free_notify) {
 104                 swp_entry_t entry;
 105                 unsigned long offset;
 106
 107                 entry.val = page_private(page);
 108                 offset = swp_offset(entry);
 109
 110                 SetPageDirty(page);
 111                 disk->fops->swap_slot_free_notify(sis->bdev,
 112                                 offset);
 113         }
 114 }
 115
 116 static void end_swap_bio_read(struct bio *bio)
 117 {
 118         struct page *page = bio->bi_io_vec[0].bv_page;
 119
 120         if (bio->bi_error) {
 121                 SetPageError(page);
 122                 ClearPageUptodate(page);
 123                 pr_alert("Read-error on swap-device (%u:%u:%llu)\n",
 124                          imajor(bio->bi_bdev->bd_inode),
 125                          iminor(bio->bi_bdev->bd_inode),
 126                          (unsigned long long)bio->bi_iter.bi_sector);
 127                 goto out;
 128         }
 129
 130         SetPageUptodate(page);
 131         swap_slot_free_notify(page);
 132 out:
 133         unlock_page(page);
 134         bio_put(bio);
 135 }
 136
 137 int generic_swapfile_activate(struct swap_info_struct *sis,
 138                                 struct file *swap_file,
 139                                 sector_t *span)
 140 {
 141         struct address_space *mapping = swap_file->f_mapping;
 142         struct inode *inode = mapping->host;
 143         unsigned blocks_per_page;
 144         unsigned long page_no;
 145         unsigned blkbits;
 146         sector_t probe_block;
 147         sector_t last_block;
 148         sector_t lowest_block = -1;
 149         sector_t highest_block = 0;
 150         int nr_extents = 0;
 151         int ret;
 152
 153         blkbits = inode->i_blkbits;
 154         blocks_per_page = PAGE_SIZE >> blkbits;
 155
 156         /*
 157          * Map all the blocks into the extent list.  This code doesn't try
 158          * to be very smart.
 159          */
 160         probe_block = 0;
 161         page_no = 0;
 162         last_block = i_size_read(inode) >> blkbits;
 163         while ((probe_block + blocks_per_page) <= last_block &&
 164                         page_no < sis->max) {
 165                 unsigned block_in_page;
 166                 sector_t first_block;
 167
 168                 cond_resched();
 169
 170                 first_block = bmap(inode, probe_block);
 171                 if (first_block == 0)
 172                         goto bad_bmap;
 173
 174                 /*
 175                  * It must be PAGE_SIZE aligned on-disk
 176                  */
 177                 if (first_block & (blocks_per_page - 1)) {
 178                         probe_block++;
 179                         goto reprobe;
 180                 }
 181
 182                 for (block_in_page = 1; block_in_page < blocks_per_page;
 183                                         block_in_page++) {
 184                         sector_t block;
 185
 186                         block = bmap(inode, probe_block + block_in_page);
 187                         if (block == 0)
 188                                 goto bad_bmap;
 189                         if (block != first_block + block_in_page) {
 190                                 /* Discontiguity */
 191                                 probe_block++;
 192                                 goto reprobe;
 193                         }
 194                 }
 195
 196                 first_block >>= (PAGE_SHIFT - blkbits);
 197                 if (page_no) {  /* exclude the header page */
 198                         if (first_block < lowest_block)
 199                                 lowest_block = first_block;
 200                         if (first_block > highest_block)
 201                                 highest_block = first_block;
 202                 }
 203
 204                 /*
 205                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
 206                  */
 207                 ret = add_swap_extent(sis, page_no, 1, first_block);
 208                 if (ret < 0)
 209                         goto out;
 210                 nr_extents += ret;
 211                 page_no++;
 212                 probe_block += blocks_per_page;
 213 reprobe:
 214                 continue;
 215         }
 216         ret = nr_extents;
 217         *span = 1 + highest_block - lowest_block;
 218         if (page_no == 0)
 219                 page_no = 1;    /* force Empty message */
 220         sis->max = page_no;
 221         sis->pages = page_no - 1;
 222         sis->highest_bit = page_no - 1;
 223 out:
 224         return ret;
 225 bad_bmap:
 226         pr_err("swapon: swapfile has holes\n");
 227         ret = -EINVAL;
 228         goto out;
 229 }
 230
 231 /*
 232  * We may have stale swap cache pages in memory: notice
 233  * them here and get rid of the unnecessary final write.
 234  */
 235 int swap_writepage(struct page *page, struct writeback_control *wbc)
 236 {
 237         int ret = 0;
 238
 239         if (try_to_free_swap(page)) {
 240                 unlock_page(page);
 241                 goto out;
 242         }
 243         if (frontswap_store(page) == 0) {
 244                 set_page_writeback(page);
 245                 unlock_page(page);
 246                 end_page_writeback(page);
 247                 goto out;
 248         }
 249         ret = __swap_writepage(page, wbc, end_swap_bio_write);
 250 out:
 251         return ret;
 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;
 259         struct swap_info_struct *sis = page_swap_info(page);
 260
 261         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 262         if (sis->flags & SWP_FILE) {
 263                 struct kiocb kiocb;
 264                 struct file *swap_file = sis->swap_file;
 265                 struct address_space *mapping = swap_file->f_mapping;
 266                 struct bio_vec bv = {
 267                         .bv_page = page,
 268                         .bv_len  = PAGE_SIZE,
 269                         .bv_offset = 0
 270                 };
 271                 struct iov_iter from;
 272
 273                 iov_iter_bvec(&from, ITER_BVEC | WRITE, &bv, 1, PAGE_SIZE);
 274                 init_sync_kiocb(&kiocb, swap_file);
 275                 kiocb.ki_pos = page_file_offset(page);
 276
 277                 set_page_writeback(page);
 278                 unlock_page(page);
 279                 ret = mapping->a_ops->direct_IO(&kiocb, &from);
 280                 if (ret == PAGE_SIZE) {
 281                         count_vm_event(PSWPOUT);
 282                         ret = 0;
 283                 } else {
 284                         /*
 285                          * In the case of swap-over-nfs, this can be a
 286                          * temporary failure if the system has limited
 287                          * memory for allocating transmit buffers.
 288                          * Mark the page dirty and avoid
 289                          * rotate_reclaimable_page but rate-limit the
 290                          * messages but do not flag PageError like
 291                          * the normal direct-to-bio case as it could
 292                          * be temporary.
 293                          */
 294                         set_page_dirty(page);
 295                         ClearPageReclaim(page);
 296                         pr_err_ratelimited("Write error on dio swapfile (%llu)\n",
 297                                            page_file_offset(page));
 298                 }
 299                 end_page_writeback(page);
 300                 return ret;
 301         }
 302
 303         ret = bdev_write_page(sis->bdev, map_swap_page(page, &sis->bdev),
 304                               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                 bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC);
 320         else
 321                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
 322         count_vm_event(PSWPOUT);
 323         set_page_writeback(page);
 324         unlock_page(page);
 325         submit_bio(bio);
 326 out:
 327         return ret;
 328 }
 329
 330 int swap_readpage(struct page *page)
 331 {
 332         struct bio *bio;
 333         int ret = 0;
 334         struct swap_info_struct *sis = page_swap_info(page);
 335
 336         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 337         VM_BUG_ON_PAGE(!PageLocked(page), page);
 338         VM_BUG_ON_PAGE(PageUptodate(page), page);
 339         if (frontswap_load(page) == 0) {
 340                 SetPageUptodate(page);
 341                 unlock_page(page);
 342                 goto out;
 343         }
 344
 345         if (sis->flags & SWP_FILE) {
 346                 struct file *swap_file = sis->swap_file;
 347                 struct address_space *mapping = swap_file->f_mapping;
 348
 349                 ret = mapping->a_ops->readpage(swap_file, page);
 350                 if (!ret)
 351                         count_vm_event(PSWPIN);
 352                 return ret;
 353         }
 354
 355         ret = bdev_read_page(sis->bdev, map_swap_page(page, &sis->bdev), page);
 356         if (!ret) {
 357                 if (trylock_page(page)) {
 358                         swap_slot_free_notify(page);
 359                         unlock_page(page);
 360                 }
 361
 362                 count_vm_event(PSWPIN);
 363                 return 0;
 364         }
 365
 366         ret = 0;
 367         bio = get_swap_bio(GFP_KERNEL, page, end_swap_bio_read);
 368         if (bio == NULL) {
 369                 unlock_page(page);
 370                 ret = -ENOMEM;
 371                 goto out;
 372         }
 373         bio_set_op_attrs(bio, REQ_OP_READ, 0);
 374         count_vm_event(PSWPIN);
 375         submit_bio(bio);
 376 out:
 377         return ret;
 378 }
 379
 380 int swap_set_page_dirty(struct page *page)
 381 {
 382         struct swap_info_struct *sis = page_swap_info(page);
 383
 384         if (sis->flags & SWP_FILE) {
 385                 struct address_space *mapping = sis->swap_file->f_mapping;
 386
 387                 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 388                 return mapping->a_ops->set_page_dirty(page);
 389         } else {
 390                 return __set_page_dirty_no_writeback(page);
 391         }
 392 }