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 void end_swap_bio_write(struct bio *bio)
  30 {
  31         struct page *page = bio_first_page_all(bio);
  32
  33         if (bio->bi_status) {
  34                 SetPageError(page);
  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 rotate_reclaimable_page()
  42                  */
  43                 set_page_dirty(page);
  44                 pr_alert("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                 ClearPageReclaim(page);
  48         }
  49         end_page_writeback(page);
  50         bio_put(bio);
  51 }
  52
  53 static void swap_slot_free_notify(struct page *page)
  54 {
  55         struct swap_info_struct *sis;
  56         struct gendisk *disk;
  57         swp_entry_t entry;
  58
  59         /*
  60          * There is no guarantee that the page is in swap cache - the software
  61          * suspend code (at least) uses end_swap_bio_read() against a non-
  62          * swapcache page.  So we must check PG_swapcache before proceeding with
  63          * this optimization.
  64          */
  65         if (unlikely(!PageSwapCache(page)))
  66                 return;
  67
  68         sis = page_swap_info(page);
  69         if (data_race(!(sis->flags & SWP_BLKDEV)))
  70                 return;
  71
  72         /*
  73          * The swap subsystem performs lazy swap slot freeing,
  74          * expecting that the page will be swapped out again.
  75          * So we can avoid an unnecessary write if the page
  76          * isn't redirtied.
  77          * This is good for real swap storage because we can
  78          * reduce unnecessary I/O and enhance wear-leveling
  79          * if an SSD is used as the as swap device.
  80          * But if in-memory swap device (eg zram) is used,
  81          * this causes a duplicated copy between uncompressed
  82          * data in VM-owned memory and compressed data in
  83          * zram-owned memory.  So let's free zram-owned memory
  84          * and make the VM-owned decompressed page *dirty*,
  85          * so the page should be swapped out somewhere again if
  86          * we again wish to reclaim it.
  87          */
  88         disk = sis->bdev->bd_disk;
  89         entry.val = page_private(page);
  90         if (disk->fops->swap_slot_free_notify && __swap_count(entry) == 1) {
  91                 unsigned long offset;
  92
  93                 offset = swp_offset(entry);
  94
  95                 SetPageDirty(page);
  96                 disk->fops->swap_slot_free_notify(sis->bdev,
  97                                 offset);
  98         }
  99 }
 100
 101 static void end_swap_bio_read(struct bio *bio)
 102 {
 103         struct page *page = bio_first_page_all(bio);
 104         struct task_struct *waiter = bio->bi_private;
 105
 106         if (bio->bi_status) {
 107                 SetPageError(page);
 108                 ClearPageUptodate(page);
 109                 pr_alert("Read-error on swap-device (%u:%u:%llu)\n",
 110                          MAJOR(bio_dev(bio)), MINOR(bio_dev(bio)),
 111                          (unsigned long long)bio->bi_iter.bi_sector);
 112                 goto out;
 113         }
 114
 115         SetPageUptodate(page);
 116         swap_slot_free_notify(page);
 117 out:
 118         unlock_page(page);
 119         WRITE_ONCE(bio->bi_private, NULL);
 120         bio_put(bio);
 121         if (waiter) {
 122                 blk_wake_io_task(waiter);
 123                 put_task_struct(waiter);
 124         }
 125 }
 126
 127 int generic_swapfile_activate(struct swap_info_struct *sis,
 128                                 struct file *swap_file,
 129                                 sector_t *span)
 130 {
 131         struct address_space *mapping = swap_file->f_mapping;
 132         struct inode *inode = mapping->host;
 133         unsigned blocks_per_page;
 134         unsigned long page_no;
 135         unsigned blkbits;
 136         sector_t probe_block;
 137         sector_t last_block;
 138         sector_t lowest_block = -1;
 139         sector_t highest_block = 0;
 140         int nr_extents = 0;
 141         int ret;
 142
 143         blkbits = inode->i_blkbits;
 144         blocks_per_page = PAGE_SIZE >> blkbits;
 145
 146         /*
 147          * Map all the blocks into the extent tree.  This code doesn't try
 148          * to be very smart.
 149          */
 150         probe_block = 0;
 151         page_no = 0;
 152         last_block = i_size_read(inode) >> blkbits;
 153         while ((probe_block + blocks_per_page) <= last_block &&
 154                         page_no < sis->max) {
 155                 unsigned block_in_page;
 156                 sector_t first_block;
 157
 158                 cond_resched();
 159
 160                 first_block = probe_block;
 161                 ret = bmap(inode, &first_block);
 162                 if (ret || !first_block)
 163                         goto bad_bmap;
 164
 165                 /*
 166                  * It must be PAGE_SIZE aligned on-disk
 167                  */
 168                 if (first_block & (blocks_per_page - 1)) {
 169                         probe_block++;
 170                         goto reprobe;
 171                 }
 172
 173                 for (block_in_page = 1; block_in_page < blocks_per_page;
 174                                         block_in_page++) {
 175                         sector_t block;
 176
 177                         block = probe_block + block_in_page;
 178                         ret = bmap(inode, &block);
 179                         if (ret || !block)
 180                                 goto bad_bmap;
 181
 182                         if (block != first_block + block_in_page) {
 183                                 /* Discontiguity */
 184                                 probe_block++;
 185                                 goto reprobe;
 186                         }
 187                 }
 188
 189                 first_block >>= (PAGE_SHIFT - blkbits);
 190                 if (page_no) {  /* exclude the header page */
 191                         if (first_block < lowest_block)
 192                                 lowest_block = first_block;
 193                         if (first_block > highest_block)
 194                                 highest_block = first_block;
 195                 }
 196
 197                 /*
 198                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
 199                  */
 200                 ret = add_swap_extent(sis, page_no, 1, first_block);
 201                 if (ret < 0)
 202                         goto out;
 203                 nr_extents += ret;
 204                 page_no++;
 205                 probe_block += blocks_per_page;
 206 reprobe:
 207                 continue;
 208         }
 209         ret = nr_extents;
 210         *span = 1 + highest_block - lowest_block;
 211         if (page_no == 0)
 212                 page_no = 1;    /* force Empty message */
 213         sis->max = page_no;
 214         sis->pages = page_no - 1;
 215         sis->highest_bit = page_no - 1;
 216 out:
 217         return ret;
 218 bad_bmap:
 219         pr_err("swapon: swapfile has holes\n");
 220         ret = -EINVAL;
 221         goto out;
 222 }
 223
 224 /*
 225  * We may have stale swap cache pages in memory: notice
 226  * them here and get rid of the unnecessary final write.
 227  */
 228 int swap_writepage(struct page *page, struct writeback_control *wbc)
 229 {
 230         int ret = 0;
 231
 232         if (try_to_free_swap(page)) {
 233                 unlock_page(page);
 234                 goto out;
 235         }
 236         /*
 237          * Arch code may have to preserve more data than just the page
 238          * contents, e.g. memory tags.
 239          */
 240         ret = arch_prepare_to_swap(page);
 241         if (ret) {
 242                 set_page_dirty(page);
 243                 unlock_page(page);
 244                 goto out;
 245         }
 246         if (frontswap_store(page) == 0) {
 247                 set_page_writeback(page);
 248                 unlock_page(page);
 249                 end_page_writeback(page);
 250                 goto out;
 251         }
 252         ret = __swap_writepage(page, wbc, end_swap_bio_write);
 253 out:
 254         return ret;
 255 }
 256
 257 static sector_t swap_page_sector(struct page *page)
 258 {
 259         return (sector_t)__page_file_index(page) << (PAGE_SHIFT - 9);
 260 }
 261
 262 static inline void count_swpout_vm_event(struct page *page)
 263 {
 264 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 265         if (unlikely(PageTransHuge(page)))
 266                 count_vm_event(THP_SWPOUT);
 267 #endif
 268         count_vm_events(PSWPOUT, thp_nr_pages(page));
 269 }
 270
 271 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
 272 static void bio_associate_blkg_from_page(struct bio *bio, struct page *page)
 273 {
 274         struct cgroup_subsys_state *css;
 275         struct mem_cgroup *memcg;
 276
 277         memcg = page_memcg(page);
 278         if (!memcg)
 279                 return;
 280
 281         rcu_read_lock();
 282         css = cgroup_e_css(memcg->css.cgroup, &io_cgrp_subsys);
 283         bio_associate_blkg_from_css(bio, css);
 284         rcu_read_unlock();
 285 }
 286 #else
 287 #define bio_associate_blkg_from_page(bio, page)         do { } while (0)
 288 #endif /* CONFIG_MEMCG && CONFIG_BLK_CGROUP */
 289
 290 int __swap_writepage(struct page *page, struct writeback_control *wbc,
 291                 bio_end_io_t end_write_func)
 292 {
 293         struct bio *bio;
 294         int ret;
 295         struct swap_info_struct *sis = page_swap_info(page);
 296
 297         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 298         if (data_race(sis->flags & SWP_FS_OPS)) {
 299                 struct kiocb kiocb;
 300                 struct file *swap_file = sis->swap_file;
 301                 struct address_space *mapping = swap_file->f_mapping;
 302                 struct bio_vec bv = {
 303                         .bv_page = page,
 304                         .bv_len  = PAGE_SIZE,
 305                         .bv_offset = 0
 306                 };
 307                 struct iov_iter from;
 308
 309                 iov_iter_bvec(&from, WRITE, &bv, 1, PAGE_SIZE);
 310                 init_sync_kiocb(&kiocb, swap_file);
 311                 kiocb.ki_pos = page_file_offset(page);
 312
 313                 set_page_writeback(page);
 314                 unlock_page(page);
 315                 ret = mapping->a_ops->direct_IO(&kiocb, &from);
 316                 if (ret == PAGE_SIZE) {
 317                         count_vm_event(PSWPOUT);
 318                         ret = 0;
 319                 } else {
 320                         /*
 321                          * In the case of swap-over-nfs, this can be a
 322                          * temporary failure if the system has limited
 323                          * memory for allocating transmit buffers.
 324                          * Mark the page dirty and avoid
 325                          * rotate_reclaimable_page but rate-limit the
 326                          * messages but do not flag PageError like
 327                          * the normal direct-to-bio case as it could
 328                          * be temporary.
 329                          */
 330                         set_page_dirty(page);
 331                         ClearPageReclaim(page);
 332                         pr_err_ratelimited("Write error on dio swapfile (%llu)\n",
 333                                            page_file_offset(page));
 334                 }
 335                 end_page_writeback(page);
 336                 return ret;
 337         }
 338
 339         ret = bdev_write_page(sis->bdev, swap_page_sector(page), page, wbc);
 340         if (!ret) {
 341                 count_swpout_vm_event(page);
 342                 return 0;
 343         }
 344
 345         bio = bio_alloc(GFP_NOIO, 1);
 346         bio_set_dev(bio, sis->bdev);
 347         bio->bi_iter.bi_sector = swap_page_sector(page);
 348         bio->bi_opf = REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc);
 349         bio->bi_end_io = end_write_func;
 350         bio_add_page(bio, page, thp_size(page), 0);
 351
 352         bio_associate_blkg_from_page(bio, page);
 353         count_swpout_vm_event(page);
 354         set_page_writeback(page);
 355         unlock_page(page);
 356         submit_bio(bio);
 357
 358         return 0;
 359 }
 360
 361 int swap_readpage(struct page *page, bool synchronous)
 362 {
 363         struct bio *bio;
 364         int ret = 0;
 365         struct swap_info_struct *sis = page_swap_info(page);
 366         blk_qc_t qc;
 367         struct gendisk *disk;
 368         unsigned long pflags;
 369
 370         VM_BUG_ON_PAGE(!PageSwapCache(page) && !synchronous, page);
 371         VM_BUG_ON_PAGE(!PageLocked(page), page);
 372         VM_BUG_ON_PAGE(PageUptodate(page), page);
 373
 374         /*
 375          * Count submission time as memory stall. When the device is congested,
 376          * or the submitting cgroup IO-throttled, submission can be a
 377          * significant part of overall IO time.
 378          */
 379         psi_memstall_enter(&pflags);
 380
 381         if (frontswap_load(page) == 0) {
 382                 SetPageUptodate(page);
 383                 unlock_page(page);
 384                 goto out;
 385         }
 386
 387         if (data_race(sis->flags & SWP_FS_OPS)) {
 388                 struct file *swap_file = sis->swap_file;
 389                 struct address_space *mapping = swap_file->f_mapping;
 390
 391                 ret = mapping->a_ops->readpage(swap_file, page);
 392                 if (!ret)
 393                         count_vm_event(PSWPIN);
 394                 goto out;
 395         }
 396
 397         if (sis->flags & SWP_SYNCHRONOUS_IO) {
 398                 ret = bdev_read_page(sis->bdev, swap_page_sector(page), page);
 399                 if (!ret) {
 400                         if (trylock_page(page)) {
 401                                 swap_slot_free_notify(page);
 402                                 unlock_page(page);
 403                         }
 404
 405                         count_vm_event(PSWPIN);
 406                         goto out;
 407                 }
 408         }
 409
 410         ret = 0;
 411         bio = bio_alloc(GFP_KERNEL, 1);
 412         bio_set_dev(bio, sis->bdev);
 413         bio->bi_opf = REQ_OP_READ;
 414         bio->bi_iter.bi_sector = swap_page_sector(page);
 415         bio->bi_end_io = end_swap_bio_read;
 416         bio_add_page(bio, page, thp_size(page), 0);
 417
 418         disk = bio->bi_bdev->bd_disk;
 419         /*
 420          * Keep this task valid during swap readpage because the oom killer may
 421          * attempt to access it in the page fault retry time check.
 422          */
 423         if (synchronous) {
 424                 bio->bi_opf |= REQ_HIPRI;
 425                 get_task_struct(current);
 426                 bio->bi_private = current;
 427         }
 428         count_vm_event(PSWPIN);
 429         bio_get(bio);
 430         qc = submit_bio(bio);
 431         while (synchronous) {
 432                 set_current_state(TASK_UNINTERRUPTIBLE);
 433                 if (!READ_ONCE(bio->bi_private))
 434                         break;
 435
 436                 if (!blk_poll(disk->queue, qc, true))
 437                         blk_io_schedule();
 438         }
 439         __set_current_state(TASK_RUNNING);
 440         bio_put(bio);
 441
 442 out:
 443         psi_memstall_leave(&pflags);
 444         return ret;
 445 }
 446
 447 int swap_set_page_dirty(struct page *page)
 448 {
 449         struct swap_info_struct *sis = page_swap_info(page);
 450
 451         if (data_race(sis->flags & SWP_FS_OPS)) {
 452                 struct address_space *mapping = sis->swap_file->f_mapping;
 453
 454                 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 455                 return mapping->a_ops->set_page_dirty(page);
 456         } else {
 457                 return __set_page_dirty_no_writeback(page);
 458         }
 459 }