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 <linux/sched/task.h>
  26 #include <asm/pgtable.h>
  27
  28 static struct bio *get_swap_bio(gfp_t gfp_flags,
  29                                 struct page *page, bio_end_io_t end_io)
  30 {
  31         int i, nr = hpage_nr_pages(page);
  32         struct bio *bio;
  33
  34         bio = bio_alloc(gfp_flags, nr);
  35         if (bio) {
  36                 bio->bi_iter.bi_sector = map_swap_page(page, &bio->bi_bdev);
  37                 bio->bi_iter.bi_sector <<= PAGE_SHIFT - 9;
  38                 bio->bi_end_io = end_io;
  39
  40                 for (i = 0; i < nr; i++)
  41                         bio_add_page(bio, page + i, PAGE_SIZE, 0);
  42                 VM_BUG_ON(bio->bi_iter.bi_size != PAGE_SIZE * nr);
  43         }
  44         return bio;
  45 }
  46
  47 void end_swap_bio_write(struct bio *bio)
  48 {
  49         struct page *page = bio->bi_io_vec[0].bv_page;
  50
  51         if (bio->bi_status) {
  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                 pr_alert("Write-error on swap-device (%u:%u:%llu)\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 static void swap_slot_free_notify(struct page *page)
  73 {
  74         struct swap_info_struct *sis;
  75         struct gendisk *disk;
  76
  77         /*
  78          * There is no guarantee that the page is in swap cache - the software
  79          * suspend code (at least) uses end_swap_bio_read() against a non-
  80          * swapcache page.  So we must check PG_swapcache before proceeding with
  81          * this optimization.
  82          */
  83         if (unlikely(!PageSwapCache(page)))
  84                 return;
  85
  86         sis = page_swap_info(page);
  87         if (!(sis->flags & SWP_BLKDEV))
  88                 return;
  89
  90         /*
  91          * The swap subsystem performs lazy swap slot freeing,
  92          * expecting that the page will be swapped out again.
  93          * So we can avoid an unnecessary write if the page
  94          * isn't redirtied.
  95          * This is good for real swap storage because we can
  96          * reduce unnecessary I/O and enhance wear-leveling
  97          * if an SSD is used as the as swap device.
  98          * But if in-memory swap device (eg zram) is used,
  99          * this causes a duplicated copy between uncompressed
 100          * data in VM-owned memory and compressed data in
 101          * zram-owned memory.  So let's free zram-owned memory
 102          * and make the VM-owned decompressed page *dirty*,
 103          * so the page should be swapped out somewhere again if
 104          * we again wish to reclaim it.
 105          */
 106         disk = sis->bdev->bd_disk;
 107         if (disk->fops->swap_slot_free_notify) {
 108                 swp_entry_t entry;
 109                 unsigned long offset;
 110
 111                 entry.val = page_private(page);
 112                 offset = swp_offset(entry);
 113
 114                 SetPageDirty(page);
 115                 disk->fops->swap_slot_free_notify(sis->bdev,
 116                                 offset);
 117         }
 118 }
 119
 120 static void end_swap_bio_read(struct bio *bio)
 121 {
 122         struct page *page = bio->bi_io_vec[0].bv_page;
 123         struct task_struct *waiter = bio->bi_private;
 124
 125         if (bio->bi_status) {
 126                 SetPageError(page);
 127                 ClearPageUptodate(page);
 128                 pr_alert("Read-error on swap-device (%u:%u:%llu)\n",
 129                          imajor(bio->bi_bdev->bd_inode),
 130                          iminor(bio->bi_bdev->bd_inode),
 131                          (unsigned long long)bio->bi_iter.bi_sector);
 132                 goto out;
 133         }
 134
 135         SetPageUptodate(page);
 136         swap_slot_free_notify(page);
 137 out:
 138         unlock_page(page);
 139         WRITE_ONCE(bio->bi_private, NULL);
 140         bio_put(bio);
 141         wake_up_process(waiter);
 142         put_task_struct(waiter);
 143 }
 144
 145 int generic_swapfile_activate(struct swap_info_struct *sis,
 146                                 struct file *swap_file,
 147                                 sector_t *span)
 148 {
 149         struct address_space *mapping = swap_file->f_mapping;
 150         struct inode *inode = mapping->host;
 151         unsigned blocks_per_page;
 152         unsigned long page_no;
 153         unsigned blkbits;
 154         sector_t probe_block;
 155         sector_t last_block;
 156         sector_t lowest_block = -1;
 157         sector_t highest_block = 0;
 158         int nr_extents = 0;
 159         int ret;
 160
 161         blkbits = inode->i_blkbits;
 162         blocks_per_page = PAGE_SIZE >> blkbits;
 163
 164         /*
 165          * Map all the blocks into the extent list.  This code doesn't try
 166          * to be very smart.
 167          */
 168         probe_block = 0;
 169         page_no = 0;
 170         last_block = i_size_read(inode) >> blkbits;
 171         while ((probe_block + blocks_per_page) <= last_block &&
 172                         page_no < sis->max) {
 173                 unsigned block_in_page;
 174                 sector_t first_block;
 175
 176                 cond_resched();
 177
 178                 first_block = bmap(inode, probe_block);
 179                 if (first_block == 0)
 180                         goto bad_bmap;
 181
 182                 /*
 183                  * It must be PAGE_SIZE aligned on-disk
 184                  */
 185                 if (first_block & (blocks_per_page - 1)) {
 186                         probe_block++;
 187                         goto reprobe;
 188                 }
 189
 190                 for (block_in_page = 1; block_in_page < blocks_per_page;
 191                                         block_in_page++) {
 192                         sector_t block;
 193
 194                         block = bmap(inode, probe_block + block_in_page);
 195                         if (block == 0)
 196                                 goto bad_bmap;
 197                         if (block != first_block + block_in_page) {
 198                                 /* Discontiguity */
 199                                 probe_block++;
 200                                 goto reprobe;
 201                         }
 202                 }
 203
 204                 first_block >>= (PAGE_SHIFT - blkbits);
 205                 if (page_no) {  /* exclude the header page */
 206                         if (first_block < lowest_block)
 207                                 lowest_block = first_block;
 208                         if (first_block > highest_block)
 209                                 highest_block = first_block;
 210                 }
 211
 212                 /*
 213                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
 214                  */
 215                 ret = add_swap_extent(sis, page_no, 1, first_block);
 216                 if (ret < 0)
 217                         goto out;
 218                 nr_extents += ret;
 219                 page_no++;
 220                 probe_block += blocks_per_page;
 221 reprobe:
 222                 continue;
 223         }
 224         ret = nr_extents;
 225         *span = 1 + highest_block - lowest_block;
 226         if (page_no == 0)
 227                 page_no = 1;    /* force Empty message */
 228         sis->max = page_no;
 229         sis->pages = page_no - 1;
 230         sis->highest_bit = page_no - 1;
 231 out:
 232         return ret;
 233 bad_bmap:
 234         pr_err("swapon: swapfile has holes\n");
 235         ret = -EINVAL;
 236         goto out;
 237 }
 238
 239 /*
 240  * We may have stale swap cache pages in memory: notice
 241  * them here and get rid of the unnecessary final write.
 242  */
 243 int swap_writepage(struct page *page, struct writeback_control *wbc)
 244 {
 245         int ret = 0;
 246
 247         if (try_to_free_swap(page)) {
 248                 unlock_page(page);
 249                 goto out;
 250         }
 251         if (frontswap_store(page) == 0) {
 252                 set_page_writeback(page);
 253                 unlock_page(page);
 254                 end_page_writeback(page);
 255                 goto out;
 256         }
 257         ret = __swap_writepage(page, wbc, end_swap_bio_write);
 258 out:
 259         return ret;
 260 }
 261
 262 static sector_t swap_page_sector(struct page *page)
 263 {
 264         return (sector_t)__page_file_index(page) << (PAGE_SHIFT - 9);
 265 }
 266
 267 static inline void count_swpout_vm_event(struct page *page)
 268 {
 269 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 270         if (unlikely(PageTransHuge(page)))
 271                 count_vm_event(THP_SWPOUT);
 272 #endif
 273         count_vm_events(PSWPOUT, hpage_nr_pages(page));
 274 }
 275
 276 int __swap_writepage(struct page *page, struct writeback_control *wbc,
 277                 bio_end_io_t end_write_func)
 278 {
 279         struct bio *bio;
 280         int ret;
 281         struct swap_info_struct *sis = page_swap_info(page);
 282
 283         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 284         if (sis->flags & SWP_FILE) {
 285                 struct kiocb kiocb;
 286                 struct file *swap_file = sis->swap_file;
 287                 struct address_space *mapping = swap_file->f_mapping;
 288                 struct bio_vec bv = {
 289                         .bv_page = page,
 290                         .bv_len  = PAGE_SIZE,
 291                         .bv_offset = 0
 292                 };
 293                 struct iov_iter from;
 294
 295                 iov_iter_bvec(&from, ITER_BVEC | WRITE, &bv, 1, PAGE_SIZE);
 296                 init_sync_kiocb(&kiocb, swap_file);
 297                 kiocb.ki_pos = page_file_offset(page);
 298
 299                 set_page_writeback(page);
 300                 unlock_page(page);
 301                 ret = mapping->a_ops->direct_IO(&kiocb, &from);
 302                 if (ret == PAGE_SIZE) {
 303                         count_vm_event(PSWPOUT);
 304                         ret = 0;
 305                 } else {
 306                         /*
 307                          * In the case of swap-over-nfs, this can be a
 308                          * temporary failure if the system has limited
 309                          * memory for allocating transmit buffers.
 310                          * Mark the page dirty and avoid
 311                          * rotate_reclaimable_page but rate-limit the
 312                          * messages but do not flag PageError like
 313                          * the normal direct-to-bio case as it could
 314                          * be temporary.
 315                          */
 316                         set_page_dirty(page);
 317                         ClearPageReclaim(page);
 318                         pr_err_ratelimited("Write error on dio swapfile (%llu)\n",
 319                                            page_file_offset(page));
 320                 }
 321                 end_page_writeback(page);
 322                 return ret;
 323         }
 324
 325         ret = bdev_write_page(sis->bdev, swap_page_sector(page), page, wbc);
 326         if (!ret) {
 327                 count_swpout_vm_event(page);
 328                 return 0;
 329         }
 330
 331         ret = 0;
 332         bio = get_swap_bio(GFP_NOIO, page, end_write_func);
 333         if (bio == NULL) {
 334                 set_page_dirty(page);
 335                 unlock_page(page);
 336                 ret = -ENOMEM;
 337                 goto out;
 338         }
 339         bio->bi_opf = REQ_OP_WRITE | wbc_to_write_flags(wbc);
 340         count_swpout_vm_event(page);
 341         set_page_writeback(page);
 342         unlock_page(page);
 343         submit_bio(bio);
 344 out:
 345         return ret;
 346 }
 347
 348 int swap_readpage(struct page *page, bool do_poll)
 349 {
 350         struct bio *bio;
 351         int ret = 0;
 352         struct swap_info_struct *sis = page_swap_info(page);
 353         blk_qc_t qc;
 354         struct block_device *bdev;
 355
 356         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 357         VM_BUG_ON_PAGE(!PageLocked(page), page);
 358         VM_BUG_ON_PAGE(PageUptodate(page), page);
 359         if (frontswap_load(page) == 0) {
 360                 SetPageUptodate(page);
 361                 unlock_page(page);
 362                 goto out;
 363         }
 364
 365         if (sis->flags & SWP_FILE) {
 366                 struct file *swap_file = sis->swap_file;
 367                 struct address_space *mapping = swap_file->f_mapping;
 368
 369                 ret = mapping->a_ops->readpage(swap_file, page);
 370                 if (!ret)
 371                         count_vm_event(PSWPIN);
 372                 return ret;
 373         }
 374
 375         ret = bdev_read_page(sis->bdev, swap_page_sector(page), page);
 376         if (!ret) {
 377                 if (trylock_page(page)) {
 378                         swap_slot_free_notify(page);
 379                         unlock_page(page);
 380                 }
 381
 382                 count_vm_event(PSWPIN);
 383                 return 0;
 384         }
 385
 386         ret = 0;
 387         bio = get_swap_bio(GFP_KERNEL, page, end_swap_bio_read);
 388         if (bio == NULL) {
 389                 unlock_page(page);
 390                 ret = -ENOMEM;
 391                 goto out;
 392         }
 393         bdev = bio->bi_bdev;
 394         /*
 395          * Keep this task valid during swap readpage because the oom killer may
 396          * attempt to access it in the page fault retry time check.
 397          */
 398         get_task_struct(current);
 399         bio->bi_private = current;
 400         bio_set_op_attrs(bio, REQ_OP_READ, 0);
 401         count_vm_event(PSWPIN);
 402         bio_get(bio);
 403         qc = submit_bio(bio);
 404         while (do_poll) {
 405                 set_current_state(TASK_UNINTERRUPTIBLE);
 406                 if (!READ_ONCE(bio->bi_private))
 407                         break;
 408
 409                 if (!blk_mq_poll(bdev_get_queue(bdev), qc))
 410                         break;
 411         }
 412         __set_current_state(TASK_RUNNING);
 413         bio_put(bio);
 414
 415 out:
 416         return ret;
 417 }
 418
 419 int swap_set_page_dirty(struct page *page)
 420 {
 421         struct swap_info_struct *sis = page_swap_info(page);
 422
 423         if (sis->flags & SWP_FILE) {
 424                 struct address_space *mapping = sis->swap_file->f_mapping;
 425
 426                 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 427                 return mapping->a_ops->set_page_dirty(page);
 428         } else {
 429                 return __set_page_dirty_no_writeback(page);
 430         }
 431 }