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 #include <linux/delayacct.h>
  29
  30 void end_swap_bio_write(struct bio *bio)
  31 {
  32         struct page *page = bio_first_page_all(bio);
  33
  34         if (bio->bi_status) {
  35                 SetPageError(page);
  36                 /*
  37                  * We failed to write the page out to swap-space.
  38                  * Re-dirty the page in order to avoid it being reclaimed.
  39                  * Also print a dire warning that things will go BAD (tm)
  40                  * very quickly.
  41                  *
  42                  * Also clear PG_reclaim to avoid folio_rotate_reclaimable()
  43                  */
  44                 set_page_dirty(page);
  45                 pr_alert_ratelimited("Write-error on swap-device (%u:%u:%llu)\n",
  46                                      MAJOR(bio_dev(bio)), MINOR(bio_dev(bio)),
  47                                      (unsigned long long)bio->bi_iter.bi_sector);
  48                 ClearPageReclaim(page);
  49         }
  50         end_page_writeback(page);
  51         bio_put(bio);
  52 }
  53
  54 static void swap_slot_free_notify(struct page *page)
  55 {
  56         struct swap_info_struct *sis;
  57         struct gendisk *disk;
  58         swp_entry_t entry;
  59
  60         /*
  61          * There is no guarantee that the page is in swap cache - the software
  62          * suspend code (at least) uses end_swap_bio_read() against a non-
  63          * swapcache page.  So we must check PG_swapcache before proceeding with
  64          * this optimization.
  65          */
  66         if (unlikely(!PageSwapCache(page)))
  67                 return;
  68
  69         sis = page_swap_info(page);
  70         if (data_race(!(sis->flags & SWP_BLKDEV)))
  71                 return;
  72
  73         /*
  74          * The swap subsystem performs lazy swap slot freeing,
  75          * expecting that the page will be swapped out again.
  76          * So we can avoid an unnecessary write if the page
  77          * isn't redirtied.
  78          * This is good for real swap storage because we can
  79          * reduce unnecessary I/O and enhance wear-leveling
  80          * if an SSD is used as the as swap device.
  81          * But if in-memory swap device (eg zram) is used,
  82          * this causes a duplicated copy between uncompressed
  83          * data in VM-owned memory and compressed data in
  84          * zram-owned memory.  So let's free zram-owned memory
  85          * and make the VM-owned decompressed page *dirty*,
  86          * so the page should be swapped out somewhere again if
  87          * we again wish to reclaim it.
  88          */
  89         disk = sis->bdev->bd_disk;
  90         entry.val = page_private(page);
  91         if (disk->fops->swap_slot_free_notify && __swap_count(entry) == 1) {
  92                 unsigned long offset;
  93
  94                 offset = swp_offset(entry);
  95
  96                 SetPageDirty(page);
  97                 disk->fops->swap_slot_free_notify(sis->bdev,
  98                                 offset);
  99         }
 100 }
 101
 102 static void end_swap_bio_read(struct bio *bio)
 103 {
 104         struct page *page = bio_first_page_all(bio);
 105         struct task_struct *waiter = bio->bi_private;
 106
 107         if (bio->bi_status) {
 108                 SetPageError(page);
 109                 ClearPageUptodate(page);
 110                 pr_alert_ratelimited("Read-error on swap-device (%u:%u:%llu)\n",
 111                                      MAJOR(bio_dev(bio)), MINOR(bio_dev(bio)),
 112                                      (unsigned long long)bio->bi_iter.bi_sector);
 113                 goto out;
 114         }
 115
 116         SetPageUptodate(page);
 117         swap_slot_free_notify(page);
 118 out:
 119         unlock_page(page);
 120         WRITE_ONCE(bio->bi_private, NULL);
 121         bio_put(bio);
 122         if (waiter) {
 123                 blk_wake_io_task(waiter);
 124                 put_task_struct(waiter);
 125         }
 126 }
 127
 128 int generic_swapfile_activate(struct swap_info_struct *sis,
 129                                 struct file *swap_file,
 130                                 sector_t *span)
 131 {
 132         struct address_space *mapping = swap_file->f_mapping;
 133         struct inode *inode = mapping->host;
 134         unsigned blocks_per_page;
 135         unsigned long page_no;
 136         unsigned blkbits;
 137         sector_t probe_block;
 138         sector_t last_block;
 139         sector_t lowest_block = -1;
 140         sector_t highest_block = 0;
 141         int nr_extents = 0;
 142         int ret;
 143
 144         blkbits = inode->i_blkbits;
 145         blocks_per_page = PAGE_SIZE >> blkbits;
 146
 147         /*
 148          * Map all the blocks into the extent tree.  This code doesn't try
 149          * to be very smart.
 150          */
 151         probe_block = 0;
 152         page_no = 0;
 153         last_block = i_size_read(inode) >> blkbits;
 154         while ((probe_block + blocks_per_page) <= last_block &&
 155                         page_no < sis->max) {
 156                 unsigned block_in_page;
 157                 sector_t first_block;
 158
 159                 cond_resched();
 160
 161                 first_block = probe_block;
 162                 ret = bmap(inode, &first_block);
 163                 if (ret || !first_block)
 164                         goto bad_bmap;
 165
 166                 /*
 167                  * It must be PAGE_SIZE aligned on-disk
 168                  */
 169                 if (first_block & (blocks_per_page - 1)) {
 170                         probe_block++;
 171                         goto reprobe;
 172                 }
 173
 174                 for (block_in_page = 1; block_in_page < blocks_per_page;
 175                                         block_in_page++) {
 176                         sector_t block;
 177
 178                         block = probe_block + block_in_page;
 179                         ret = bmap(inode, &block);
 180                         if (ret || !block)
 181                                 goto bad_bmap;
 182
 183                         if (block != first_block + block_in_page) {
 184                                 /* Discontiguity */
 185                                 probe_block++;
 186                                 goto reprobe;
 187                         }
 188                 }
 189
 190                 first_block >>= (PAGE_SHIFT - blkbits);
 191                 if (page_no) {  /* exclude the header page */
 192                         if (first_block < lowest_block)
 193                                 lowest_block = first_block;
 194                         if (first_block > highest_block)
 195                                 highest_block = first_block;
 196                 }
 197
 198                 /*
 199                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
 200                  */
 201                 ret = add_swap_extent(sis, page_no, 1, first_block);
 202                 if (ret < 0)
 203                         goto out;
 204                 nr_extents += ret;
 205                 page_no++;
 206                 probe_block += blocks_per_page;
 207 reprobe:
 208                 continue;
 209         }
 210         ret = nr_extents;
 211         *span = 1 + highest_block - lowest_block;
 212         if (page_no == 0)
 213                 page_no = 1;    /* force Empty message */
 214         sis->max = page_no;
 215         sis->pages = page_no - 1;
 216         sis->highest_bit = page_no - 1;
 217 out:
 218         return ret;
 219 bad_bmap:
 220         pr_err("swapon: swapfile has holes\n");
 221         ret = -EINVAL;
 222         goto out;
 223 }
 224
 225 /*
 226  * We may have stale swap cache pages in memory: notice
 227  * them here and get rid of the unnecessary final write.
 228  */
 229 int swap_writepage(struct page *page, struct writeback_control *wbc)
 230 {
 231         int ret = 0;
 232
 233         if (try_to_free_swap(page)) {
 234                 unlock_page(page);
 235                 goto out;
 236         }
 237         /*
 238          * Arch code may have to preserve more data than just the page
 239          * contents, e.g. memory tags.
 240          */
 241         ret = arch_prepare_to_swap(page);
 242         if (ret) {
 243                 set_page_dirty(page);
 244                 unlock_page(page);
 245                 goto out;
 246         }
 247         if (frontswap_store(page) == 0) {
 248                 set_page_writeback(page);
 249                 unlock_page(page);
 250                 end_page_writeback(page);
 251                 goto out;
 252         }
 253         ret = __swap_writepage(page, wbc, end_swap_bio_write);
 254 out:
 255         return ret;
 256 }
 257
 258 static inline void count_swpout_vm_event(struct page *page)
 259 {
 260 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 261         if (unlikely(PageTransHuge(page)))
 262                 count_vm_event(THP_SWPOUT);
 263 #endif
 264         count_vm_events(PSWPOUT, thp_nr_pages(page));
 265 }
 266
 267 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
 268 static void bio_associate_blkg_from_page(struct bio *bio, struct page *page)
 269 {
 270         struct cgroup_subsys_state *css;
 271         struct mem_cgroup *memcg;
 272
 273         memcg = page_memcg(page);
 274         if (!memcg)
 275                 return;
 276
 277         rcu_read_lock();
 278         css = cgroup_e_css(memcg->css.cgroup, &io_cgrp_subsys);
 279         bio_associate_blkg_from_css(bio, css);
 280         rcu_read_unlock();
 281 }
 282 #else
 283 #define bio_associate_blkg_from_page(bio, page)         do { } while (0)
 284 #endif /* CONFIG_MEMCG && CONFIG_BLK_CGROUP */
 285
 286 int __swap_writepage(struct page *page, struct writeback_control *wbc,
 287                 bio_end_io_t end_write_func)
 288 {
 289         struct bio *bio;
 290         int ret;
 291         struct swap_info_struct *sis = page_swap_info(page);
 292
 293         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 294         if (data_race(sis->flags & SWP_FS_OPS)) {
 295                 struct kiocb kiocb;
 296                 struct file *swap_file = sis->swap_file;
 297                 struct address_space *mapping = swap_file->f_mapping;
 298                 struct bio_vec bv = {
 299                         .bv_page = page,
 300                         .bv_len  = PAGE_SIZE,
 301                         .bv_offset = 0
 302                 };
 303                 struct iov_iter from;
 304
 305                 iov_iter_bvec(&from, WRITE, &bv, 1, PAGE_SIZE);
 306                 init_sync_kiocb(&kiocb, swap_file);
 307                 kiocb.ki_pos = page_file_offset(page);
 308
 309                 set_page_writeback(page);
 310                 unlock_page(page);
 311                 ret = mapping->a_ops->direct_IO(&kiocb, &from);
 312                 if (ret == PAGE_SIZE) {
 313                         count_vm_event(PSWPOUT);
 314                         ret = 0;
 315                 } else {
 316                         /*
 317                          * In the case of swap-over-nfs, this can be a
 318                          * temporary failure if the system has limited
 319                          * memory for allocating transmit buffers.
 320                          * Mark the page dirty and avoid
 321                          * folio_rotate_reclaimable but rate-limit the
 322                          * messages but do not flag PageError like
 323                          * the normal direct-to-bio case as it could
 324                          * be temporary.
 325                          */
 326                         set_page_dirty(page);
 327                         ClearPageReclaim(page);
 328                         pr_err_ratelimited("Write error on dio swapfile (%llu)\n",
 329                                            page_file_offset(page));
 330                 }
 331                 end_page_writeback(page);
 332                 return ret;
 333         }
 334
 335         ret = bdev_write_page(sis->bdev, swap_page_sector(page), page, wbc);
 336         if (!ret) {
 337                 count_swpout_vm_event(page);
 338                 return 0;
 339         }
 340
 341         bio = bio_alloc(sis->bdev, 1,
 342                         REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc),
 343                         GFP_NOIO);
 344         bio->bi_iter.bi_sector = swap_page_sector(page);
 345         bio->bi_end_io = end_write_func;
 346         bio_add_page(bio, page, thp_size(page), 0);
 347
 348         bio_associate_blkg_from_page(bio, page);
 349         count_swpout_vm_event(page);
 350         set_page_writeback(page);
 351         unlock_page(page);
 352         submit_bio(bio);
 353
 354         return 0;
 355 }
 356
 357 int swap_readpage(struct page *page, bool synchronous)
 358 {
 359         struct bio *bio;
 360         int ret = 0;
 361         struct swap_info_struct *sis = page_swap_info(page);
 362         bool workingset = PageWorkingset(page);
 363         unsigned long pflags;
 364
 365         VM_BUG_ON_PAGE(!PageSwapCache(page) && !synchronous, page);
 366         VM_BUG_ON_PAGE(!PageLocked(page), page);
 367         VM_BUG_ON_PAGE(PageUptodate(page), page);
 368
 369         /*
 370          * Count submission time as memory stall. When the device is congested,
 371          * or the submitting cgroup IO-throttled, submission can be a
 372          * significant part of overall IO time.
 373          */
 374         if (workingset)
 375                 psi_memstall_enter(&pflags);
 376         delayacct_swapin_start();
 377
 378         if (frontswap_load(page) == 0) {
 379                 SetPageUptodate(page);
 380                 unlock_page(page);
 381                 goto out;
 382         }
 383
 384         if (data_race(sis->flags & SWP_FS_OPS)) {
 385                 struct file *swap_file = sis->swap_file;
 386                 struct address_space *mapping = swap_file->f_mapping;
 387
 388                 ret = mapping->a_ops->readpage(swap_file, page);
 389                 if (!ret)
 390                         count_vm_event(PSWPIN);
 391                 goto out;
 392         }
 393
 394         if (sis->flags & SWP_SYNCHRONOUS_IO) {
 395                 ret = bdev_read_page(sis->bdev, swap_page_sector(page), page);
 396                 if (!ret) {
 397                         if (trylock_page(page)) {
 398                                 swap_slot_free_notify(page);
 399                                 unlock_page(page);
 400                         }
 401
 402                         count_vm_event(PSWPIN);
 403                         goto out;
 404                 }
 405         }
 406
 407         ret = 0;
 408         bio = bio_alloc(sis->bdev, 1, REQ_OP_READ, GFP_KERNEL);
 409         bio->bi_iter.bi_sector = swap_page_sector(page);
 410         bio->bi_end_io = end_swap_bio_read;
 411         bio_add_page(bio, page, thp_size(page), 0);
 412         /*
 413          * Keep this task valid during swap readpage because the oom killer may
 414          * attempt to access it in the page fault retry time check.
 415          */
 416         if (synchronous) {
 417                 bio->bi_opf |= REQ_POLLED;
 418                 get_task_struct(current);
 419                 bio->bi_private = current;
 420         }
 421         count_vm_event(PSWPIN);
 422         bio_get(bio);
 423         submit_bio(bio);
 424         while (synchronous) {
 425                 set_current_state(TASK_UNINTERRUPTIBLE);
 426                 if (!READ_ONCE(bio->bi_private))
 427                         break;
 428
 429                 if (!bio_poll(bio, NULL, 0))
 430                         blk_io_schedule();
 431         }
 432         __set_current_state(TASK_RUNNING);
 433         bio_put(bio);
 434
 435 out:
 436         if (workingset)
 437                 psi_memstall_leave(&pflags);
 438         delayacct_swapin_end();
 439         return ret;
 440 }
 441
 442 bool swap_dirty_folio(struct address_space *mapping, struct folio *folio)
 443 {
 444         struct swap_info_struct *sis = swp_swap_info(folio_swap_entry(folio));
 445
 446         if (data_race(sis->flags & SWP_FS_OPS)) {
 447                 const struct address_space_operations *aops;
 448
 449                 mapping = sis->swap_file->f_mapping;
 450                 aops = mapping->a_ops;
 451
 452                 VM_BUG_ON_FOLIO(!folio_test_swapcache(folio), folio);
 453                 return aops->dirty_folio(mapping, folio);
 454         } else {
 455                 return noop_dirty_folio(mapping, folio);
 456         }
 457 }