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/writeback.h>
  22 #include <linux/frontswap.h>
  23 #include <linux/blkdev.h>
  24 #include <linux/psi.h>
  25 #include <linux/uio.h>
  26 #include <linux/sched/task.h>
  27 #include <linux/delayacct.h>
  28 #include "swap.h"
  29
  30 static void __end_swap_bio_write(struct bio *bio)
  31 {
  32         struct folio *folio = bio_first_folio_all(bio);
  33
  34         if (bio->bi_status) {
  35                 /*
  36                  * We failed to write the page out to swap-space.
  37                  * Re-dirty the page in order to avoid it being reclaimed.
  38                  * Also print a dire warning that things will go BAD (tm)
  39                  * very quickly.
  40                  *
  41                  * Also clear PG_reclaim to avoid folio_rotate_reclaimable()
  42                  */
  43                 folio_mark_dirty(folio);
  44                 pr_alert_ratelimited("Write-error on swap-device (%u:%u:%llu)\n",
  45                                      MAJOR(bio_dev(bio)), MINOR(bio_dev(bio)),
  46                                      (unsigned long long)bio->bi_iter.bi_sector);
  47                 folio_clear_reclaim(folio);
  48         }
  49         folio_end_writeback(folio);
  50 }
  51
  52 static void end_swap_bio_write(struct bio *bio)
  53 {
  54         __end_swap_bio_write(bio);
  55         bio_put(bio);
  56 }
  57
  58 static void __end_swap_bio_read(struct bio *bio)
  59 {
  60         struct folio *folio = bio_first_folio_all(bio);
  61
  62         if (bio->bi_status) {
  63                 pr_alert_ratelimited("Read-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         } else {
  67                 folio_mark_uptodate(folio);
  68         }
  69         folio_unlock(folio);
  70 }
  71
  72 static void end_swap_bio_read(struct bio *bio)
  73 {
  74         __end_swap_bio_read(bio);
  75         bio_put(bio);
  76 }
  77
  78 int generic_swapfile_activate(struct swap_info_struct *sis,
  79                                 struct file *swap_file,
  80                                 sector_t *span)
  81 {
  82         struct address_space *mapping = swap_file->f_mapping;
  83         struct inode *inode = mapping->host;
  84         unsigned blocks_per_page;
  85         unsigned long page_no;
  86         unsigned blkbits;
  87         sector_t probe_block;
  88         sector_t last_block;
  89         sector_t lowest_block = -1;
  90         sector_t highest_block = 0;
  91         int nr_extents = 0;
  92         int ret;
  93
  94         blkbits = inode->i_blkbits;
  95         blocks_per_page = PAGE_SIZE >> blkbits;
  96
  97         /*
  98          * Map all the blocks into the extent tree.  This code doesn't try
  99          * to be very smart.
 100          */
 101         probe_block = 0;
 102         page_no = 0;
 103         last_block = i_size_read(inode) >> blkbits;
 104         while ((probe_block + blocks_per_page) <= last_block &&
 105                         page_no < sis->max) {
 106                 unsigned block_in_page;
 107                 sector_t first_block;
 108
 109                 cond_resched();
 110
 111                 first_block = probe_block;
 112                 ret = bmap(inode, &first_block);
 113                 if (ret || !first_block)
 114                         goto bad_bmap;
 115
 116                 /*
 117                  * It must be PAGE_SIZE aligned on-disk
 118                  */
 119                 if (first_block & (blocks_per_page - 1)) {
 120                         probe_block++;
 121                         goto reprobe;
 122                 }
 123
 124                 for (block_in_page = 1; block_in_page < blocks_per_page;
 125                                         block_in_page++) {
 126                         sector_t block;
 127
 128                         block = probe_block + block_in_page;
 129                         ret = bmap(inode, &block);
 130                         if (ret || !block)
 131                                 goto bad_bmap;
 132
 133                         if (block != first_block + block_in_page) {
 134                                 /* Discontiguity */
 135                                 probe_block++;
 136                                 goto reprobe;
 137                         }
 138                 }
 139
 140                 first_block >>= (PAGE_SHIFT - blkbits);
 141                 if (page_no) {  /* exclude the header page */
 142                         if (first_block < lowest_block)
 143                                 lowest_block = first_block;
 144                         if (first_block > highest_block)
 145                                 highest_block = first_block;
 146                 }
 147
 148                 /*
 149                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
 150                  */
 151                 ret = add_swap_extent(sis, page_no, 1, first_block);
 152                 if (ret < 0)
 153                         goto out;
 154                 nr_extents += ret;
 155                 page_no++;
 156                 probe_block += blocks_per_page;
 157 reprobe:
 158                 continue;
 159         }
 160         ret = nr_extents;
 161         *span = 1 + highest_block - lowest_block;
 162         if (page_no == 0)
 163                 page_no = 1;    /* force Empty message */
 164         sis->max = page_no;
 165         sis->pages = page_no - 1;
 166         sis->highest_bit = page_no - 1;
 167 out:
 168         return ret;
 169 bad_bmap:
 170         pr_err("swapon: swapfile has holes\n");
 171         ret = -EINVAL;
 172         goto out;
 173 }
 174
 175 /*
 176  * We may have stale swap cache pages in memory: notice
 177  * them here and get rid of the unnecessary final write.
 178  */
 179 int swap_writepage(struct page *page, struct writeback_control *wbc)
 180 {
 181         struct folio *folio = page_folio(page);
 182         int ret;
 183
 184         if (folio_free_swap(folio)) {
 185                 folio_unlock(folio);
 186                 return 0;
 187         }
 188         /*
 189          * Arch code may have to preserve more data than just the page
 190          * contents, e.g. memory tags.
 191          */
 192         ret = arch_prepare_to_swap(&folio->page);
 193         if (ret) {
 194                 folio_mark_dirty(folio);
 195                 folio_unlock(folio);
 196                 return ret;
 197         }
 198         if (frontswap_store(&folio->page) == 0) {
 199                 folio_start_writeback(folio);
 200                 folio_unlock(folio);
 201                 folio_end_writeback(folio);
 202                 return 0;
 203         }
 204         __swap_writepage(&folio->page, wbc);
 205         return 0;
 206 }
 207
 208 static inline void count_swpout_vm_event(struct page *page)
 209 {
 210 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 211         if (unlikely(PageTransHuge(page)))
 212                 count_vm_event(THP_SWPOUT);
 213 #endif
 214         count_vm_events(PSWPOUT, thp_nr_pages(page));
 215 }
 216
 217 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
 218 static void bio_associate_blkg_from_page(struct bio *bio, struct page *page)
 219 {
 220         struct cgroup_subsys_state *css;
 221         struct mem_cgroup *memcg;
 222
 223         memcg = page_memcg(page);
 224         if (!memcg)
 225                 return;
 226
 227         rcu_read_lock();
 228         css = cgroup_e_css(memcg->css.cgroup, &io_cgrp_subsys);
 229         bio_associate_blkg_from_css(bio, css);
 230         rcu_read_unlock();
 231 }
 232 #else
 233 #define bio_associate_blkg_from_page(bio, page)         do { } while (0)
 234 #endif /* CONFIG_MEMCG && CONFIG_BLK_CGROUP */
 235
 236 struct swap_iocb {
 237         struct kiocb            iocb;
 238         struct bio_vec          bvec[SWAP_CLUSTER_MAX];
 239         int                     pages;
 240         int                     len;
 241 };
 242 static mempool_t *sio_pool;
 243
 244 int sio_pool_init(void)
 245 {
 246         if (!sio_pool) {
 247                 mempool_t *pool = mempool_create_kmalloc_pool(
 248                         SWAP_CLUSTER_MAX, sizeof(struct swap_iocb));
 249                 if (cmpxchg(&sio_pool, NULL, pool))
 250                         mempool_destroy(pool);
 251         }
 252         if (!sio_pool)
 253                 return -ENOMEM;
 254         return 0;
 255 }
 256
 257 static void sio_write_complete(struct kiocb *iocb, long ret)
 258 {
 259         struct swap_iocb *sio = container_of(iocb, struct swap_iocb, iocb);
 260         struct page *page = sio->bvec[0].bv_page;
 261         int p;
 262
 263         if (ret != sio->len) {
 264                 /*
 265                  * In the case of swap-over-nfs, this can be a
 266                  * temporary failure if the system has limited
 267                  * memory for allocating transmit buffers.
 268                  * Mark the page dirty and avoid
 269                  * folio_rotate_reclaimable but rate-limit the
 270                  * messages but do not flag PageError like
 271                  * the normal direct-to-bio case as it could
 272                  * be temporary.
 273                  */
 274                 pr_err_ratelimited("Write error %ld on dio swapfile (%llu)\n",
 275                                    ret, page_file_offset(page));
 276                 for (p = 0; p < sio->pages; p++) {
 277                         page = sio->bvec[p].bv_page;
 278                         set_page_dirty(page);
 279                         ClearPageReclaim(page);
 280                 }
 281         } else {
 282                 for (p = 0; p < sio->pages; p++)
 283                         count_swpout_vm_event(sio->bvec[p].bv_page);
 284         }
 285
 286         for (p = 0; p < sio->pages; p++)
 287                 end_page_writeback(sio->bvec[p].bv_page);
 288
 289         mempool_free(sio, sio_pool);
 290 }
 291
 292 static void swap_writepage_fs(struct page *page, struct writeback_control *wbc)
 293 {
 294         struct swap_iocb *sio = NULL;
 295         struct swap_info_struct *sis = page_swap_info(page);
 296         struct file *swap_file = sis->swap_file;
 297         loff_t pos = page_file_offset(page);
 298
 299         set_page_writeback(page);
 300         unlock_page(page);
 301         if (wbc->swap_plug)
 302                 sio = *wbc->swap_plug;
 303         if (sio) {
 304                 if (sio->iocb.ki_filp != swap_file ||
 305                     sio->iocb.ki_pos + sio->len != pos) {
 306                         swap_write_unplug(sio);
 307                         sio = NULL;
 308                 }
 309         }
 310         if (!sio) {
 311                 sio = mempool_alloc(sio_pool, GFP_NOIO);
 312                 init_sync_kiocb(&sio->iocb, swap_file);
 313                 sio->iocb.ki_complete = sio_write_complete;
 314                 sio->iocb.ki_pos = pos;
 315                 sio->pages = 0;
 316                 sio->len = 0;
 317         }
 318         bvec_set_page(&sio->bvec[sio->pages], page, thp_size(page), 0);
 319         sio->len += thp_size(page);
 320         sio->pages += 1;
 321         if (sio->pages == ARRAY_SIZE(sio->bvec) || !wbc->swap_plug) {
 322                 swap_write_unplug(sio);
 323                 sio = NULL;
 324         }
 325         if (wbc->swap_plug)
 326                 *wbc->swap_plug = sio;
 327 }
 328
 329 static void swap_writepage_bdev_sync(struct page *page,
 330                 struct writeback_control *wbc, struct swap_info_struct *sis)
 331 {
 332         struct bio_vec bv;
 333         struct bio bio;
 334
 335         bio_init(&bio, sis->bdev, &bv, 1,
 336                  REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc));
 337         bio.bi_iter.bi_sector = swap_page_sector(page);
 338         __bio_add_page(&bio, page, thp_size(page), 0);
 339
 340         bio_associate_blkg_from_page(&bio, page);
 341         count_swpout_vm_event(page);
 342
 343         set_page_writeback(page);
 344         unlock_page(page);
 345
 346         submit_bio_wait(&bio);
 347         __end_swap_bio_write(&bio);
 348 }
 349
 350 static void swap_writepage_bdev_async(struct page *page,
 351                 struct writeback_control *wbc, struct swap_info_struct *sis)
 352 {
 353         struct bio *bio;
 354
 355         bio = bio_alloc(sis->bdev, 1,
 356                         REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc),
 357                         GFP_NOIO);
 358         bio->bi_iter.bi_sector = swap_page_sector(page);
 359         bio->bi_end_io = end_swap_bio_write;
 360         __bio_add_page(bio, page, thp_size(page), 0);
 361
 362         bio_associate_blkg_from_page(bio, page);
 363         count_swpout_vm_event(page);
 364         set_page_writeback(page);
 365         unlock_page(page);
 366         submit_bio(bio);
 367 }
 368
 369 void __swap_writepage(struct page *page, struct writeback_control *wbc)
 370 {
 371         struct swap_info_struct *sis = page_swap_info(page);
 372
 373         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 374         /*
 375          * ->flags can be updated non-atomicially (scan_swap_map_slots),
 376          * but that will never affect SWP_FS_OPS, so the data_race
 377          * is safe.
 378          */
 379         if (data_race(sis->flags & SWP_FS_OPS))
 380                 swap_writepage_fs(page, wbc);
 381         else if (sis->flags & SWP_SYNCHRONOUS_IO)
 382                 swap_writepage_bdev_sync(page, wbc, sis);
 383         else
 384                 swap_writepage_bdev_async(page, wbc, sis);
 385 }
 386
 387 void swap_write_unplug(struct swap_iocb *sio)
 388 {
 389         struct iov_iter from;
 390         struct address_space *mapping = sio->iocb.ki_filp->f_mapping;
 391         int ret;
 392
 393         iov_iter_bvec(&from, ITER_SOURCE, sio->bvec, sio->pages, sio->len);
 394         ret = mapping->a_ops->swap_rw(&sio->iocb, &from);
 395         if (ret != -EIOCBQUEUED)
 396                 sio_write_complete(&sio->iocb, ret);
 397 }
 398
 399 static void sio_read_complete(struct kiocb *iocb, long ret)
 400 {
 401         struct swap_iocb *sio = container_of(iocb, struct swap_iocb, iocb);
 402         int p;
 403
 404         if (ret == sio->len) {
 405                 for (p = 0; p < sio->pages; p++) {
 406                         struct page *page = sio->bvec[p].bv_page;
 407
 408                         SetPageUptodate(page);
 409                         unlock_page(page);
 410                 }
 411                 count_vm_events(PSWPIN, sio->pages);
 412         } else {
 413                 for (p = 0; p < sio->pages; p++) {
 414                         struct page *page = sio->bvec[p].bv_page;
 415
 416                         unlock_page(page);
 417                 }
 418                 pr_alert_ratelimited("Read-error on swap-device\n");
 419         }
 420         mempool_free(sio, sio_pool);
 421 }
 422
 423 static void swap_readpage_fs(struct page *page,
 424                              struct swap_iocb **plug)
 425 {
 426         struct swap_info_struct *sis = page_swap_info(page);
 427         struct swap_iocb *sio = NULL;
 428         loff_t pos = page_file_offset(page);
 429
 430         if (plug)
 431                 sio = *plug;
 432         if (sio) {
 433                 if (sio->iocb.ki_filp != sis->swap_file ||
 434                     sio->iocb.ki_pos + sio->len != pos) {
 435                         swap_read_unplug(sio);
 436                         sio = NULL;
 437                 }
 438         }
 439         if (!sio) {
 440                 sio = mempool_alloc(sio_pool, GFP_KERNEL);
 441                 init_sync_kiocb(&sio->iocb, sis->swap_file);
 442                 sio->iocb.ki_pos = pos;
 443                 sio->iocb.ki_complete = sio_read_complete;
 444                 sio->pages = 0;
 445                 sio->len = 0;
 446         }
 447         bvec_set_page(&sio->bvec[sio->pages], page, thp_size(page), 0);
 448         sio->len += thp_size(page);
 449         sio->pages += 1;
 450         if (sio->pages == ARRAY_SIZE(sio->bvec) || !plug) {
 451                 swap_read_unplug(sio);
 452                 sio = NULL;
 453         }
 454         if (plug)
 455                 *plug = sio;
 456 }
 457
 458 static void swap_readpage_bdev_sync(struct page *page,
 459                 struct swap_info_struct *sis)
 460 {
 461         struct bio_vec bv;
 462         struct bio bio;
 463
 464         bio_init(&bio, sis->bdev, &bv, 1, REQ_OP_READ);
 465         bio.bi_iter.bi_sector = swap_page_sector(page);
 466         __bio_add_page(&bio, page, thp_size(page), 0);
 467         /*
 468          * Keep this task valid during swap readpage because the oom killer may
 469          * attempt to access it in the page fault retry time check.
 470          */
 471         get_task_struct(current);
 472         count_vm_event(PSWPIN);
 473         submit_bio_wait(&bio);
 474         __end_swap_bio_read(&bio);
 475         put_task_struct(current);
 476 }
 477
 478 static void swap_readpage_bdev_async(struct page *page,
 479                 struct swap_info_struct *sis)
 480 {
 481         struct bio *bio;
 482
 483         bio = bio_alloc(sis->bdev, 1, REQ_OP_READ, GFP_KERNEL);
 484         bio->bi_iter.bi_sector = swap_page_sector(page);
 485         bio->bi_end_io = end_swap_bio_read;
 486         __bio_add_page(bio, page, thp_size(page), 0);
 487         count_vm_event(PSWPIN);
 488         submit_bio(bio);
 489 }
 490
 491 void swap_readpage(struct page *page, bool synchronous, struct swap_iocb **plug)
 492 {
 493         struct swap_info_struct *sis = page_swap_info(page);
 494         bool workingset = PageWorkingset(page);
 495         unsigned long pflags;
 496         bool in_thrashing;
 497
 498         VM_BUG_ON_PAGE(!PageSwapCache(page) && !synchronous, page);
 499         VM_BUG_ON_PAGE(!PageLocked(page), page);
 500         VM_BUG_ON_PAGE(PageUptodate(page), page);
 501
 502         /*
 503          * Count submission time as memory stall and delay. When the device
 504          * is congested, or the submitting cgroup IO-throttled, submission
 505          * can be a significant part of overall IO time.
 506          */
 507         if (workingset) {
 508                 delayacct_thrashing_start(&in_thrashing);
 509                 psi_memstall_enter(&pflags);
 510         }
 511         delayacct_swapin_start();
 512
 513         if (frontswap_load(page) == 0) {
 514                 SetPageUptodate(page);
 515                 unlock_page(page);
 516         } else if (data_race(sis->flags & SWP_FS_OPS)) {
 517                 swap_readpage_fs(page, plug);
 518         } else if (synchronous || (sis->flags & SWP_SYNCHRONOUS_IO)) {
 519                 swap_readpage_bdev_sync(page, sis);
 520         } else {
 521                 swap_readpage_bdev_async(page, sis);
 522         }
 523
 524         if (workingset) {
 525                 delayacct_thrashing_end(&in_thrashing);
 526                 psi_memstall_leave(&pflags);
 527         }
 528         delayacct_swapin_end();
 529 }
 530
 531 void __swap_read_unplug(struct swap_iocb *sio)
 532 {
 533         struct iov_iter from;
 534         struct address_space *mapping = sio->iocb.ki_filp->f_mapping;
 535         int ret;
 536
 537         iov_iter_bvec(&from, ITER_DEST, sio->bvec, sio->pages, sio->len);
 538         ret = mapping->a_ops->swap_rw(&sio->iocb, &from);
 539         if (ret != -EIOCBQUEUED)
 540                 sio_read_complete(&sio->iocb, ret);
 541 }