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/blkdev.h>
  23 #include <linux/psi.h>
  24 #include <linux/uio.h>
  25 #include <linux/sched/task.h>
  26 #include <linux/delayacct.h>
  27 #include <linux/zswap.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 (zswap_store(folio)) {
 199                 folio_start_writeback(folio);
 200                 folio_unlock(folio);
 201                 folio_end_writeback(folio);
 202                 return 0;
 203         }
 204         if (!mem_cgroup_zswap_writeback_enabled(folio_memcg(folio))) {
 205                 folio_mark_dirty(folio);
 206                 return AOP_WRITEPAGE_ACTIVATE;
 207         }
 208
 209         __swap_writepage(folio, wbc);
 210         return 0;
 211 }
 212
 213 static inline void count_swpout_vm_event(struct folio *folio)
 214 {
 215 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 216         if (unlikely(folio_test_pmd_mappable(folio))) {
 217                 count_memcg_folio_events(folio, THP_SWPOUT, 1);
 218                 count_vm_event(THP_SWPOUT);
 219         }
 220 #endif
 221         count_vm_events(PSWPOUT, folio_nr_pages(folio));
 222 }
 223
 224 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
 225 static void bio_associate_blkg_from_page(struct bio *bio, struct folio *folio)
 226 {
 227         struct cgroup_subsys_state *css;
 228         struct mem_cgroup *memcg;
 229
 230         memcg = folio_memcg(folio);
 231         if (!memcg)
 232                 return;
 233
 234         rcu_read_lock();
 235         css = cgroup_e_css(memcg->css.cgroup, &io_cgrp_subsys);
 236         bio_associate_blkg_from_css(bio, css);
 237         rcu_read_unlock();
 238 }
 239 #else
 240 #define bio_associate_blkg_from_page(bio, folio)                do { } while (0)
 241 #endif /* CONFIG_MEMCG && CONFIG_BLK_CGROUP */
 242
 243 struct swap_iocb {
 244         struct kiocb            iocb;
 245         struct bio_vec          bvec[SWAP_CLUSTER_MAX];
 246         int                     pages;
 247         int                     len;
 248 };
 249 static mempool_t *sio_pool;
 250
 251 int sio_pool_init(void)
 252 {
 253         if (!sio_pool) {
 254                 mempool_t *pool = mempool_create_kmalloc_pool(
 255                         SWAP_CLUSTER_MAX, sizeof(struct swap_iocb));
 256                 if (cmpxchg(&sio_pool, NULL, pool))
 257                         mempool_destroy(pool);
 258         }
 259         if (!sio_pool)
 260                 return -ENOMEM;
 261         return 0;
 262 }
 263
 264 static void sio_write_complete(struct kiocb *iocb, long ret)
 265 {
 266         struct swap_iocb *sio = container_of(iocb, struct swap_iocb, iocb);
 267         struct page *page = sio->bvec[0].bv_page;
 268         int p;
 269
 270         if (ret != sio->len) {
 271                 /*
 272                  * In the case of swap-over-nfs, this can be a
 273                  * temporary failure if the system has limited
 274                  * memory for allocating transmit buffers.
 275                  * Mark the page dirty and avoid
 276                  * folio_rotate_reclaimable but rate-limit the
 277                  * messages but do not flag PageError like
 278                  * the normal direct-to-bio case as it could
 279                  * be temporary.
 280                  */
 281                 pr_err_ratelimited("Write error %ld on dio swapfile (%llu)\n",
 282                                    ret, page_file_offset(page));
 283                 for (p = 0; p < sio->pages; p++) {
 284                         page = sio->bvec[p].bv_page;
 285                         set_page_dirty(page);
 286                         ClearPageReclaim(page);
 287                 }
 288         }
 289
 290         for (p = 0; p < sio->pages; p++)
 291                 end_page_writeback(sio->bvec[p].bv_page);
 292
 293         mempool_free(sio, sio_pool);
 294 }
 295
 296 static void swap_writepage_fs(struct folio *folio, struct writeback_control *wbc)
 297 {
 298         struct swap_iocb *sio = NULL;
 299         struct swap_info_struct *sis = swp_swap_info(folio->swap);
 300         struct file *swap_file = sis->swap_file;
 301         loff_t pos = folio_file_pos(folio);
 302
 303         count_swpout_vm_event(folio);
 304         folio_start_writeback(folio);
 305         folio_unlock(folio);
 306         if (wbc->swap_plug)
 307                 sio = *wbc->swap_plug;
 308         if (sio) {
 309                 if (sio->iocb.ki_filp != swap_file ||
 310                     sio->iocb.ki_pos + sio->len != pos) {
 311                         swap_write_unplug(sio);
 312                         sio = NULL;
 313                 }
 314         }
 315         if (!sio) {
 316                 sio = mempool_alloc(sio_pool, GFP_NOIO);
 317                 init_sync_kiocb(&sio->iocb, swap_file);
 318                 sio->iocb.ki_complete = sio_write_complete;
 319                 sio->iocb.ki_pos = pos;
 320                 sio->pages = 0;
 321                 sio->len = 0;
 322         }
 323         bvec_set_folio(&sio->bvec[sio->pages], folio, folio_size(folio), 0);
 324         sio->len += folio_size(folio);
 325         sio->pages += 1;
 326         if (sio->pages == ARRAY_SIZE(sio->bvec) || !wbc->swap_plug) {
 327                 swap_write_unplug(sio);
 328                 sio = NULL;
 329         }
 330         if (wbc->swap_plug)
 331                 *wbc->swap_plug = sio;
 332 }
 333
 334 static void swap_writepage_bdev_sync(struct folio *folio,
 335                 struct writeback_control *wbc, struct swap_info_struct *sis)
 336 {
 337         struct bio_vec bv;
 338         struct bio bio;
 339
 340         bio_init(&bio, sis->bdev, &bv, 1,
 341                  REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc));
 342         bio.bi_iter.bi_sector = swap_folio_sector(folio);
 343         bio_add_folio_nofail(&bio, folio, folio_size(folio), 0);
 344
 345         bio_associate_blkg_from_page(&bio, folio);
 346         count_swpout_vm_event(folio);
 347
 348         folio_start_writeback(folio);
 349         folio_unlock(folio);
 350
 351         submit_bio_wait(&bio);
 352         __end_swap_bio_write(&bio);
 353 }
 354
 355 static void swap_writepage_bdev_async(struct folio *folio,
 356                 struct writeback_control *wbc, struct swap_info_struct *sis)
 357 {
 358         struct bio *bio;
 359
 360         bio = bio_alloc(sis->bdev, 1,
 361                         REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc),
 362                         GFP_NOIO);
 363         bio->bi_iter.bi_sector = swap_folio_sector(folio);
 364         bio->bi_end_io = end_swap_bio_write;
 365         bio_add_folio_nofail(bio, folio, folio_size(folio), 0);
 366
 367         bio_associate_blkg_from_page(bio, folio);
 368         count_swpout_vm_event(folio);
 369         folio_start_writeback(folio);
 370         folio_unlock(folio);
 371         submit_bio(bio);
 372 }
 373
 374 void __swap_writepage(struct folio *folio, struct writeback_control *wbc)
 375 {
 376         struct swap_info_struct *sis = swp_swap_info(folio->swap);
 377
 378         VM_BUG_ON_FOLIO(!folio_test_swapcache(folio), folio);
 379         /*
 380          * ->flags can be updated non-atomicially (scan_swap_map_slots),
 381          * but that will never affect SWP_FS_OPS, so the data_race
 382          * is safe.
 383          */
 384         if (data_race(sis->flags & SWP_FS_OPS))
 385                 swap_writepage_fs(folio, wbc);
 386         else if (sis->flags & SWP_SYNCHRONOUS_IO)
 387                 swap_writepage_bdev_sync(folio, wbc, sis);
 388         else
 389                 swap_writepage_bdev_async(folio, wbc, sis);
 390 }
 391
 392 void swap_write_unplug(struct swap_iocb *sio)
 393 {
 394         struct iov_iter from;
 395         struct address_space *mapping = sio->iocb.ki_filp->f_mapping;
 396         int ret;
 397
 398         iov_iter_bvec(&from, ITER_SOURCE, sio->bvec, sio->pages, sio->len);
 399         ret = mapping->a_ops->swap_rw(&sio->iocb, &from);
 400         if (ret != -EIOCBQUEUED)
 401                 sio_write_complete(&sio->iocb, ret);
 402 }
 403
 404 static void sio_read_complete(struct kiocb *iocb, long ret)
 405 {
 406         struct swap_iocb *sio = container_of(iocb, struct swap_iocb, iocb);
 407         int p;
 408
 409         if (ret == sio->len) {
 410                 for (p = 0; p < sio->pages; p++) {
 411                         struct folio *folio = page_folio(sio->bvec[p].bv_page);
 412
 413                         folio_mark_uptodate(folio);
 414                         folio_unlock(folio);
 415                 }
 416                 count_vm_events(PSWPIN, sio->pages);
 417         } else {
 418                 for (p = 0; p < sio->pages; p++) {
 419                         struct folio *folio = page_folio(sio->bvec[p].bv_page);
 420
 421                         folio_unlock(folio);
 422                 }
 423                 pr_alert_ratelimited("Read-error on swap-device\n");
 424         }
 425         mempool_free(sio, sio_pool);
 426 }
 427
 428 static void swap_read_folio_fs(struct folio *folio, struct swap_iocb **plug)
 429 {
 430         struct swap_info_struct *sis = swp_swap_info(folio->swap);
 431         struct swap_iocb *sio = NULL;
 432         loff_t pos = folio_file_pos(folio);
 433
 434         if (plug)
 435                 sio = *plug;
 436         if (sio) {
 437                 if (sio->iocb.ki_filp != sis->swap_file ||
 438                     sio->iocb.ki_pos + sio->len != pos) {
 439                         swap_read_unplug(sio);
 440                         sio = NULL;
 441                 }
 442         }
 443         if (!sio) {
 444                 sio = mempool_alloc(sio_pool, GFP_KERNEL);
 445                 init_sync_kiocb(&sio->iocb, sis->swap_file);
 446                 sio->iocb.ki_pos = pos;
 447                 sio->iocb.ki_complete = sio_read_complete;
 448                 sio->pages = 0;
 449                 sio->len = 0;
 450         }
 451         bvec_set_folio(&sio->bvec[sio->pages], folio, folio_size(folio), 0);
 452         sio->len += folio_size(folio);
 453         sio->pages += 1;
 454         if (sio->pages == ARRAY_SIZE(sio->bvec) || !plug) {
 455                 swap_read_unplug(sio);
 456                 sio = NULL;
 457         }
 458         if (plug)
 459                 *plug = sio;
 460 }
 461
 462 static void swap_read_folio_bdev_sync(struct folio *folio,
 463                 struct swap_info_struct *sis)
 464 {
 465         struct bio_vec bv;
 466         struct bio bio;
 467
 468         bio_init(&bio, sis->bdev, &bv, 1, REQ_OP_READ);
 469         bio.bi_iter.bi_sector = swap_folio_sector(folio);
 470         bio_add_folio_nofail(&bio, folio, folio_size(folio), 0);
 471         /*
 472          * Keep this task valid during swap readpage because the oom killer may
 473          * attempt to access it in the page fault retry time check.
 474          */
 475         get_task_struct(current);
 476         count_vm_event(PSWPIN);
 477         submit_bio_wait(&bio);
 478         __end_swap_bio_read(&bio);
 479         put_task_struct(current);
 480 }
 481
 482 static void swap_read_folio_bdev_async(struct folio *folio,
 483                 struct swap_info_struct *sis)
 484 {
 485         struct bio *bio;
 486
 487         bio = bio_alloc(sis->bdev, 1, REQ_OP_READ, GFP_KERNEL);
 488         bio->bi_iter.bi_sector = swap_folio_sector(folio);
 489         bio->bi_end_io = end_swap_bio_read;
 490         bio_add_folio_nofail(bio, folio, folio_size(folio), 0);
 491         count_vm_event(PSWPIN);
 492         submit_bio(bio);
 493 }
 494
 495 void swap_read_folio(struct folio *folio, bool synchronous,
 496                 struct swap_iocb **plug)
 497 {
 498         struct swap_info_struct *sis = swp_swap_info(folio->swap);
 499         bool workingset = folio_test_workingset(folio);
 500         unsigned long pflags;
 501         bool in_thrashing;
 502
 503         VM_BUG_ON_FOLIO(!folio_test_swapcache(folio) && !synchronous, folio);
 504         VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
 505         VM_BUG_ON_FOLIO(folio_test_uptodate(folio), folio);
 506
 507         /*
 508          * Count submission time as memory stall and delay. When the device
 509          * is congested, or the submitting cgroup IO-throttled, submission
 510          * can be a significant part of overall IO time.
 511          */
 512         if (workingset) {
 513                 delayacct_thrashing_start(&in_thrashing);
 514                 psi_memstall_enter(&pflags);
 515         }
 516         delayacct_swapin_start();
 517
 518         if (zswap_load(folio)) {
 519                 folio_mark_uptodate(folio);
 520                 folio_unlock(folio);
 521         } else if (data_race(sis->flags & SWP_FS_OPS)) {
 522                 swap_read_folio_fs(folio, plug);
 523         } else if (synchronous || (sis->flags & SWP_SYNCHRONOUS_IO)) {
 524                 swap_read_folio_bdev_sync(folio, sis);
 525         } else {
 526                 swap_read_folio_bdev_async(folio, sis);
 527         }
 528
 529         if (workingset) {
 530                 delayacct_thrashing_end(&in_thrashing);
 531                 psi_memstall_leave(&pflags);
 532         }
 533         delayacct_swapin_end();
 534 }
 535
 536 void __swap_read_unplug(struct swap_iocb *sio)
 537 {
 538         struct iov_iter from;
 539         struct address_space *mapping = sio->iocb.ki_filp->f_mapping;
 540         int ret;
 541
 542         iov_iter_bvec(&from, ITER_DEST, sio->bvec, sio->pages, sio->len);
 543         ret = mapping->a_ops->swap_rw(&sio->iocb, &from);
 544         if (ret != -EIOCBQUEUED)
 545                 sio_read_complete(&sio->iocb, ret);
 546 }