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