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