mm/page_io.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  linux/mm/page_io.c
   4  *
   5  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   6  *
   7  *  Swap reorganised 29.12.95,
   8  *  Asynchronous swapping added 30.12.95. Stephen Tweedie
   9  *  Removed race in async swapping. 14.4.1996. Bruno Haible
  10  *  Add swap of shared pages through the page cache. 20.2.1998. Stephen Tweedie
  11  *  Always use brw_page, life becomes simpler. 12 May 1998 Eric Biederman
  12  */
  13
  14 #include <linux/mm.h>
  15 #include <linux/kernel_stat.h>
  16 #include <linux/gfp.h>
  17 #include <linux/pagemap.h>
  18 #include <linux/swap.h>
  19 #include <linux/bio.h>
  20 #include <linux/swapops.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/writeback.h>
  23 #include <linux/frontswap.h>
  24 #include <linux/blkdev.h>
  25 #include <linux/uio.h>
  26 #include <linux/sched/task.h>
  27 #include <asm/pgtable.h>
  28
  29 static struct bio *get_swap_bio(gfp_t gfp_flags,
  30                                 struct page *page, bio_end_io_t end_io)
  31 {
  32         struct bio *bio;
  33
  34         bio = bio_alloc(gfp_flags, 1);
  35         if (bio) {
  36                 struct block_device *bdev;
  37
  38                 bio->bi_iter.bi_sector = map_swap_page(page, &bdev);
  39                 bio_set_dev(bio, bdev);
  40                 bio->bi_iter.bi_sector <<= PAGE_SHIFT - 9;
  41                 bio->bi_end_io = end_io;
  42
  43                 bio_add_page(bio, page, PAGE_SIZE * hpage_nr_pages(page), 0);
  44         }
  45         return bio;
  46 }
  47
  48 void end_swap_bio_write(struct bio *bio)
  49 {
  50         struct page *page = bio_first_page_all(bio);
  51
  52         if (bio->bi_status) {
  53                 SetPageError(page);
  54                 /*
  55                  * We failed to write the page out to swap-space.
  56                  * Re-dirty the page in order to avoid it being reclaimed.
  57                  * Also print a dire warning that things will go BAD (tm)
  58                  * very quickly.
  59                  *
  60                  * Also clear PG_reclaim to avoid rotate_reclaimable_page()
  61                  */
  62                 set_page_dirty(page);
  63                 pr_alert("Write-error on swap-device (%u:%u:%llu)\n",
  64                          MAJOR(bio_dev(bio)), MINOR(bio_dev(bio)),
  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_first_page_all(bio);
 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                          MAJOR(bio_dev(bio)), MINOR(bio_dev(bio)),
 130                          (unsigned long long)bio->bi_iter.bi_sector);
 131                 goto out;
 132         }
 133
 134         SetPageUptodate(page);
 135         swap_slot_free_notify(page);
 136 out:
 137         unlock_page(page);
 138         WRITE_ONCE(bio->bi_private, NULL);
 139         bio_put(bio);
 140         blk_wake_io_task(waiter);
 141         put_task_struct(waiter);
 142 }
 143
 144 int generic_swapfile_activate(struct swap_info_struct *sis,
 145                                 struct file *swap_file,
 146                                 sector_t *span)
 147 {
 148         struct address_space *mapping = swap_file->f_mapping;
 149         struct inode *inode = mapping->host;
 150         unsigned blocks_per_page;
 151         unsigned long page_no;
 152         unsigned blkbits;
 153         sector_t probe_block;
 154         sector_t last_block;
 155         sector_t lowest_block = -1;
 156         sector_t highest_block = 0;
 157         int nr_extents = 0;
 158         int ret;
 159
 160         blkbits = inode->i_blkbits;
 161         blocks_per_page = PAGE_SIZE >> blkbits;
 162
 163         /*
 164          * Map all the blocks into the extent list.  This code doesn't try
 165          * to be very smart.
 166          */
 167         probe_block = 0;
 168         page_no = 0;
 169         last_block = i_size_read(inode) >> blkbits;
 170         while ((probe_block + blocks_per_page) <= last_block &&
 171                         page_no < sis->max) {
 172                 unsigned block_in_page;
 173                 sector_t first_block;
 174
 175                 cond_resched();
 176
 177                 first_block = bmap(inode, probe_block);
 178                 if (first_block == 0)
 179                         goto bad_bmap;
 180
 181                 /*
 182                  * It must be PAGE_SIZE aligned on-disk
 183                  */
 184                 if (first_block & (blocks_per_page - 1)) {
 185                         probe_block++;
 186                         goto reprobe;
 187                 }
 188
 189                 for (block_in_page = 1; block_in_page < blocks_per_page;
 190                                         block_in_page++) {
 191                         sector_t block;
 192
 193                         block = bmap(inode, probe_block + block_in_page);
 194                         if (block == 0)
 195                                 goto bad_bmap;
 196                         if (block != first_block + block_in_page) {
 197                                 /* Discontiguity */
 198                                 probe_block++;
 199                                 goto reprobe;
 200                         }
 201                 }
 202
 203                 first_block >>= (PAGE_SHIFT - blkbits);
 204                 if (page_no) {  /* exclude the header page */
 205                         if (first_block < lowest_block)
 206                                 lowest_block = first_block;
 207                         if (first_block > highest_block)
 208                                 highest_block = first_block;
 209                 }
 210
 211                 /*
 212                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
 213                  */
 214                 ret = add_swap_extent(sis, page_no, 1, first_block);
 215                 if (ret < 0)
 216                         goto out;
 217                 nr_extents += ret;
 218                 page_no++;
 219                 probe_block += blocks_per_page;
 220 reprobe:
 221                 continue;
 222         }
 223         ret = nr_extents;
 224         *span = 1 + highest_block - lowest_block;
 225         if (page_no == 0)
 226                 page_no = 1;    /* force Empty message */
 227         sis->max = page_no;
 228         sis->pages = page_no - 1;
 229         sis->highest_bit = page_no - 1;
 230 out:
 231         return ret;
 232 bad_bmap:
 233         pr_err("swapon: swapfile has holes\n");
 234         ret = -EINVAL;
 235         goto out;
 236 }
 237
 238 /*
 239  * We may have stale swap cache pages in memory: notice
 240  * them here and get rid of the unnecessary final write.
 241  */
 242 int swap_writepage(struct page *page, struct writeback_control *wbc)
 243 {
 244         int ret = 0;
 245
 246         if (try_to_free_swap(page)) {
 247                 unlock_page(page);
 248                 goto out;
 249         }
 250         if (frontswap_store(page) == 0) {
 251                 set_page_writeback(page);
 252                 unlock_page(page);
 253                 end_page_writeback(page);
 254                 goto out;
 255         }
 256         ret = __swap_writepage(page, wbc, end_swap_bio_write);
 257 out:
 258         return ret;
 259 }
 260
 261 static sector_t swap_page_sector(struct page *page)
 262 {
 263         return (sector_t)__page_file_index(page) << (PAGE_SHIFT - 9);
 264 }
 265
 266 static inline void count_swpout_vm_event(struct page *page)
 267 {
 268 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 269         if (unlikely(PageTransHuge(page)))
 270                 count_vm_event(THP_SWPOUT);
 271 #endif
 272         count_vm_events(PSWPOUT, hpage_nr_pages(page));
 273 }
 274
 275 int __swap_writepage(struct page *page, struct writeback_control *wbc,
 276                 bio_end_io_t end_write_func)
 277 {
 278         struct bio *bio;
 279         int ret;
 280         struct swap_info_struct *sis = page_swap_info(page);
 281
 282         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 283         if (sis->flags & SWP_FS) {
 284                 struct kiocb kiocb;
 285                 struct file *swap_file = sis->swap_file;
 286                 struct address_space *mapping = swap_file->f_mapping;
 287                 struct bio_vec bv = {
 288                         .bv_page = page,
 289                         .bv_len  = PAGE_SIZE,
 290                         .bv_offset = 0
 291                 };
 292                 struct iov_iter from;
 293
 294                 iov_iter_bvec(&from, WRITE, &bv, 1, PAGE_SIZE);
 295                 init_sync_kiocb(&kiocb, swap_file);
 296                 kiocb.ki_pos = page_file_offset(page);
 297
 298                 set_page_writeback(page);
 299                 unlock_page(page);
 300                 ret = mapping->a_ops->direct_IO(&kiocb, &from);
 301                 if (ret == PAGE_SIZE) {
 302                         count_vm_event(PSWPOUT);
 303                         ret = 0;
 304                 } else {
 305                         /*
 306                          * In the case of swap-over-nfs, this can be a
 307                          * temporary failure if the system has limited
 308                          * memory for allocating transmit buffers.
 309                          * Mark the page dirty and avoid
 310                          * rotate_reclaimable_page but rate-limit the
 311                          * messages but do not flag PageError like
 312                          * the normal direct-to-bio case as it could
 313                          * be temporary.
 314                          */
 315                         set_page_dirty(page);
 316                         ClearPageReclaim(page);
 317                         pr_err_ratelimited("Write error on dio swapfile (%llu)\n",
 318                                            page_file_offset(page));
 319                 }
 320                 end_page_writeback(page);
 321                 return ret;
 322         }
 323
 324         ret = bdev_write_page(sis->bdev, swap_page_sector(page), page, wbc);
 325         if (!ret) {
 326                 count_swpout_vm_event(page);
 327                 return 0;
 328         }
 329
 330         ret = 0;
 331         bio = get_swap_bio(GFP_NOIO, page, end_write_func);
 332         if (bio == NULL) {
 333                 set_page_dirty(page);
 334                 unlock_page(page);
 335                 ret = -ENOMEM;
 336                 goto out;
 337         }
 338         bio->bi_opf = REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc);
 339         bio_associate_blkg_from_page(bio, page);
 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 synchronous)
 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 gendisk *disk;
 355
 356         VM_BUG_ON_PAGE(!PageSwapCache(page) && !synchronous, 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_FS) {
 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         disk = bio->bi_disk;
 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         if (synchronous)
 402                 bio->bi_opf |= REQ_HIPRI;
 403         count_vm_event(PSWPIN);
 404         bio_get(bio);
 405         qc = submit_bio(bio);
 406         while (synchronous) {
 407                 set_current_state(TASK_UNINTERRUPTIBLE);
 408                 if (!READ_ONCE(bio->bi_private))
 409                         break;
 410
 411                 if (!blk_poll(disk->queue, qc, true))
 412                         io_schedule();
 413         }
 414         __set_current_state(TASK_RUNNING);
 415         bio_put(bio);
 416
 417 out:
 418         return ret;
 419 }
 420
 421 int swap_set_page_dirty(struct page *page)
 422 {
 423         struct swap_info_struct *sis = page_swap_info(page);
 424
 425         if (sis->flags & SWP_FS) {
 426                 struct address_space *mapping = sis->swap_file->f_mapping;
 427
 428                 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 429                 return mapping->a_ops->set_page_dirty(page);
 430         } else {
 431                 return __set_page_dirty_no_writeback(page);
 432         }
 433 }