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/psi.h>
  26 #include <linux/uio.h>
  27 #include <linux/sched/task.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, thp_size(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         swp_entry_t entry;
  77
  78         /*
  79          * There is no guarantee that the page is in swap cache - the software
  80          * suspend code (at least) uses end_swap_bio_read() against a non-
  81          * swapcache page.  So we must check PG_swapcache before proceeding with
  82          * this optimization.
  83          */
  84         if (unlikely(!PageSwapCache(page)))
  85                 return;
  86
  87         sis = page_swap_info(page);
  88         if (data_race(!(sis->flags & SWP_BLKDEV)))
  89                 return;
  90
  91         /*
  92          * The swap subsystem performs lazy swap slot freeing,
  93          * expecting that the page will be swapped out again.
  94          * So we can avoid an unnecessary write if the page
  95          * isn't redirtied.
  96          * This is good for real swap storage because we can
  97          * reduce unnecessary I/O and enhance wear-leveling
  98          * if an SSD is used as the as swap device.
  99          * But if in-memory swap device (eg zram) is used,
 100          * this causes a duplicated copy between uncompressed
 101          * data in VM-owned memory and compressed data in
 102          * zram-owned memory.  So let's free zram-owned memory
 103          * and make the VM-owned decompressed page *dirty*,
 104          * so the page should be swapped out somewhere again if
 105          * we again wish to reclaim it.
 106          */
 107         disk = sis->bdev->bd_disk;
 108         entry.val = page_private(page);
 109         if (disk->fops->swap_slot_free_notify && __swap_count(entry) == 1) {
 110                 unsigned long offset;
 111
 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         if (waiter) {
 141                 blk_wake_io_task(waiter);
 142                 put_task_struct(waiter);
 143         }
 144 }
 145
 146 int generic_swapfile_activate(struct swap_info_struct *sis,
 147                                 struct file *swap_file,
 148                                 sector_t *span)
 149 {
 150         struct address_space *mapping = swap_file->f_mapping;
 151         struct inode *inode = mapping->host;
 152         unsigned blocks_per_page;
 153         unsigned long page_no;
 154         unsigned blkbits;
 155         sector_t probe_block;
 156         sector_t last_block;
 157         sector_t lowest_block = -1;
 158         sector_t highest_block = 0;
 159         int nr_extents = 0;
 160         int ret;
 161
 162         blkbits = inode->i_blkbits;
 163         blocks_per_page = PAGE_SIZE >> blkbits;
 164
 165         /*
 166          * Map all the blocks into the extent tree.  This code doesn't try
 167          * to be very smart.
 168          */
 169         probe_block = 0;
 170         page_no = 0;
 171         last_block = i_size_read(inode) >> blkbits;
 172         while ((probe_block + blocks_per_page) <= last_block &&
 173                         page_no < sis->max) {
 174                 unsigned block_in_page;
 175                 sector_t first_block;
 176
 177                 cond_resched();
 178
 179                 first_block = probe_block;
 180                 ret = bmap(inode, &first_block);
 181                 if (ret || !first_block)
 182                         goto bad_bmap;
 183
 184                 /*
 185                  * It must be PAGE_SIZE aligned on-disk
 186                  */
 187                 if (first_block & (blocks_per_page - 1)) {
 188                         probe_block++;
 189                         goto reprobe;
 190                 }
 191
 192                 for (block_in_page = 1; block_in_page < blocks_per_page;
 193                                         block_in_page++) {
 194                         sector_t block;
 195
 196                         block = probe_block + block_in_page;
 197                         ret = bmap(inode, &block);
 198                         if (ret || !block)
 199                                 goto bad_bmap;
 200
 201                         if (block != first_block + block_in_page) {
 202                                 /* Discontiguity */
 203                                 probe_block++;
 204                                 goto reprobe;
 205                         }
 206                 }
 207
 208                 first_block >>= (PAGE_SHIFT - blkbits);
 209                 if (page_no) {  /* exclude the header page */
 210                         if (first_block < lowest_block)
 211                                 lowest_block = first_block;
 212                         if (first_block > highest_block)
 213                                 highest_block = first_block;
 214                 }
 215
 216                 /*
 217                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
 218                  */
 219                 ret = add_swap_extent(sis, page_no, 1, first_block);
 220                 if (ret < 0)
 221                         goto out;
 222                 nr_extents += ret;
 223                 page_no++;
 224                 probe_block += blocks_per_page;
 225 reprobe:
 226                 continue;
 227         }
 228         ret = nr_extents;
 229         *span = 1 + highest_block - lowest_block;
 230         if (page_no == 0)
 231                 page_no = 1;    /* force Empty message */
 232         sis->max = page_no;
 233         sis->pages = page_no - 1;
 234         sis->highest_bit = page_no - 1;
 235 out:
 236         return ret;
 237 bad_bmap:
 238         pr_err("swapon: swapfile has holes\n");
 239         ret = -EINVAL;
 240         goto out;
 241 }
 242
 243 /*
 244  * We may have stale swap cache pages in memory: notice
 245  * them here and get rid of the unnecessary final write.
 246  */
 247 int swap_writepage(struct page *page, struct writeback_control *wbc)
 248 {
 249         int ret = 0;
 250
 251         if (try_to_free_swap(page)) {
 252                 unlock_page(page);
 253                 goto out;
 254         }
 255         /*
 256          * Arch code may have to preserve more data than just the page
 257          * contents, e.g. memory tags.
 258          */
 259         ret = arch_prepare_to_swap(page);
 260         if (ret) {
 261                 set_page_dirty(page);
 262                 unlock_page(page);
 263                 goto out;
 264         }
 265         if (frontswap_store(page) == 0) {
 266                 set_page_writeback(page);
 267                 unlock_page(page);
 268                 end_page_writeback(page);
 269                 goto out;
 270         }
 271         ret = __swap_writepage(page, wbc, end_swap_bio_write);
 272 out:
 273         return ret;
 274 }
 275
 276 static sector_t swap_page_sector(struct page *page)
 277 {
 278         return (sector_t)__page_file_index(page) << (PAGE_SHIFT - 9);
 279 }
 280
 281 static inline void count_swpout_vm_event(struct page *page)
 282 {
 283 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 284         if (unlikely(PageTransHuge(page)))
 285                 count_vm_event(THP_SWPOUT);
 286 #endif
 287         count_vm_events(PSWPOUT, thp_nr_pages(page));
 288 }
 289
 290 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
 291 static void bio_associate_blkg_from_page(struct bio *bio, struct page *page)
 292 {
 293         struct cgroup_subsys_state *css;
 294
 295         if (!page->mem_cgroup)
 296                 return;
 297
 298         rcu_read_lock();
 299         css = cgroup_e_css(page->mem_cgroup->css.cgroup, &io_cgrp_subsys);
 300         bio_associate_blkg_from_css(bio, css);
 301         rcu_read_unlock();
 302 }
 303 #else
 304 #define bio_associate_blkg_from_page(bio, page)         do { } while (0)
 305 #endif /* CONFIG_MEMCG && CONFIG_BLK_CGROUP */
 306
 307 int __swap_writepage(struct page *page, struct writeback_control *wbc,
 308                 bio_end_io_t end_write_func)
 309 {
 310         struct bio *bio;
 311         int ret;
 312         struct swap_info_struct *sis = page_swap_info(page);
 313
 314         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 315         if (data_race(sis->flags & SWP_FS_OPS)) {
 316                 struct kiocb kiocb;
 317                 struct file *swap_file = sis->swap_file;
 318                 struct address_space *mapping = swap_file->f_mapping;
 319                 struct bio_vec bv = {
 320                         .bv_page = page,
 321                         .bv_len  = PAGE_SIZE,
 322                         .bv_offset = 0
 323                 };
 324                 struct iov_iter from;
 325
 326                 iov_iter_bvec(&from, WRITE, &bv, 1, PAGE_SIZE);
 327                 init_sync_kiocb(&kiocb, swap_file);
 328                 kiocb.ki_pos = page_file_offset(page);
 329
 330                 set_page_writeback(page);
 331                 unlock_page(page);
 332                 ret = mapping->a_ops->direct_IO(&kiocb, &from);
 333                 if (ret == PAGE_SIZE) {
 334                         count_vm_event(PSWPOUT);
 335                         ret = 0;
 336                 } else {
 337                         /*
 338                          * In the case of swap-over-nfs, this can be a
 339                          * temporary failure if the system has limited
 340                          * memory for allocating transmit buffers.
 341                          * Mark the page dirty and avoid
 342                          * rotate_reclaimable_page but rate-limit the
 343                          * messages but do not flag PageError like
 344                          * the normal direct-to-bio case as it could
 345                          * be temporary.
 346                          */
 347                         set_page_dirty(page);
 348                         ClearPageReclaim(page);
 349                         pr_err_ratelimited("Write error on dio swapfile (%llu)\n",
 350                                            page_file_offset(page));
 351                 }
 352                 end_page_writeback(page);
 353                 return ret;
 354         }
 355
 356         ret = bdev_write_page(sis->bdev, swap_page_sector(page), page, wbc);
 357         if (!ret) {
 358                 count_swpout_vm_event(page);
 359                 return 0;
 360         }
 361
 362         bio = get_swap_bio(GFP_NOIO, page, end_write_func);
 363         if (bio == NULL) {
 364                 set_page_dirty(page);
 365                 unlock_page(page);
 366                 return -ENOMEM;
 367         }
 368         bio->bi_opf = REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc);
 369         bio_associate_blkg_from_page(bio, page);
 370         count_swpout_vm_event(page);
 371         set_page_writeback(page);
 372         unlock_page(page);
 373         submit_bio(bio);
 374
 375         return 0;
 376 }
 377
 378 int swap_readpage(struct page *page, bool synchronous)
 379 {
 380         struct bio *bio;
 381         int ret = 0;
 382         struct swap_info_struct *sis = page_swap_info(page);
 383         blk_qc_t qc;
 384         struct gendisk *disk;
 385         unsigned long pflags;
 386
 387         VM_BUG_ON_PAGE(!PageSwapCache(page) && !synchronous, page);
 388         VM_BUG_ON_PAGE(!PageLocked(page), page);
 389         VM_BUG_ON_PAGE(PageUptodate(page), page);
 390
 391         /*
 392          * Count submission time as memory stall. When the device is congested,
 393          * or the submitting cgroup IO-throttled, submission can be a
 394          * significant part of overall IO time.
 395          */
 396         psi_memstall_enter(&pflags);
 397
 398         if (frontswap_load(page) == 0) {
 399                 SetPageUptodate(page);
 400                 unlock_page(page);
 401                 goto out;
 402         }
 403
 404         if (data_race(sis->flags & SWP_FS_OPS)) {
 405                 struct file *swap_file = sis->swap_file;
 406                 struct address_space *mapping = swap_file->f_mapping;
 407
 408                 ret = mapping->a_ops->readpage(swap_file, page);
 409                 if (!ret)
 410                         count_vm_event(PSWPIN);
 411                 goto out;
 412         }
 413
 414         if (sis->flags & SWP_SYNCHRONOUS_IO) {
 415                 ret = bdev_read_page(sis->bdev, swap_page_sector(page), page);
 416                 if (!ret) {
 417                         if (trylock_page(page)) {
 418                                 swap_slot_free_notify(page);
 419                                 unlock_page(page);
 420                         }
 421
 422                         count_vm_event(PSWPIN);
 423                         goto out;
 424                 }
 425         }
 426
 427         ret = 0;
 428         bio = get_swap_bio(GFP_KERNEL, page, end_swap_bio_read);
 429         if (bio == NULL) {
 430                 unlock_page(page);
 431                 ret = -ENOMEM;
 432                 goto out;
 433         }
 434         disk = bio->bi_disk;
 435         /*
 436          * Keep this task valid during swap readpage because the oom killer may
 437          * attempt to access it in the page fault retry time check.
 438          */
 439         bio_set_op_attrs(bio, REQ_OP_READ, 0);
 440         if (synchronous) {
 441                 bio->bi_opf |= REQ_HIPRI;
 442                 get_task_struct(current);
 443                 bio->bi_private = current;
 444         }
 445         count_vm_event(PSWPIN);
 446         bio_get(bio);
 447         qc = submit_bio(bio);
 448         while (synchronous) {
 449                 set_current_state(TASK_UNINTERRUPTIBLE);
 450                 if (!READ_ONCE(bio->bi_private))
 451                         break;
 452
 453                 if (!blk_poll(disk->queue, qc, true))
 454                         blk_io_schedule();
 455         }
 456         __set_current_state(TASK_RUNNING);
 457         bio_put(bio);
 458
 459 out:
 460         psi_memstall_leave(&pflags);
 461         return ret;
 462 }
 463
 464 int swap_set_page_dirty(struct page *page)
 465 {
 466         struct swap_info_struct *sis = page_swap_info(page);
 467
 468         if (data_race(sis->flags & SWP_FS_OPS)) {
 469                 struct address_space *mapping = sis->swap_file->f_mapping;
 470
 471                 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 472                 return mapping->a_ops->set_page_dirty(page);
 473         } else {
 474                 return __set_page_dirty_no_writeback(page);
 475         }
 476 }