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/uio.h>
  26 #include <linux/sched/task.h>
  27 #include <asm/pgtable.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         int i, nr = hpage_nr_pages(page);
  33         struct bio *bio;
  34
  35         bio = bio_alloc(gfp_flags, nr);
  36         if (bio) {
  37                 struct block_device *bdev;
  38
  39                 bio->bi_iter.bi_sector = map_swap_page(page, &bdev);
  40                 bio_set_dev(bio, bdev);
  41                 bio->bi_end_io = end_io;
  42
  43                 for (i = 0; i < nr; i++)
  44                         bio_add_page(bio, page + i, PAGE_SIZE, 0);
  45                 VM_BUG_ON(bio->bi_iter.bi_size != PAGE_SIZE * nr);
  46         }
  47         return bio;
  48 }
  49
  50 void end_swap_bio_write(struct bio *bio)
  51 {
  52         struct page *page = bio_first_page_all(bio);
  53
  54         if (bio->bi_status) {
  55                 SetPageError(page);
  56                 /*
  57                  * We failed to write the page out to swap-space.
  58                  * Re-dirty the page in order to avoid it being reclaimed.
  59                  * Also print a dire warning that things will go BAD (tm)
  60                  * very quickly.
  61                  *
  62                  * Also clear PG_reclaim to avoid rotate_reclaimable_page()
  63                  */
  64                 set_page_dirty(page);
  65                 pr_alert("Write-error on swap-device (%u:%u:%llu)\n",
  66                          MAJOR(bio_dev(bio)), MINOR(bio_dev(bio)),
  67                          (unsigned long long)bio->bi_iter.bi_sector);
  68                 ClearPageReclaim(page);
  69         }
  70         end_page_writeback(page);
  71         bio_put(bio);
  72 }
  73
  74 static void end_swap_bio_read(struct bio *bio)
  75 {
  76         struct page *page = bio_first_page_all(bio);
  77         struct task_struct *waiter = bio->bi_private;
  78
  79         if (bio->bi_status) {
  80                 SetPageError(page);
  81                 ClearPageUptodate(page);
  82                 pr_alert("Read-error on swap-device (%u:%u:%llu)\n",
  83                          MAJOR(bio_dev(bio)), MINOR(bio_dev(bio)),
  84                          (unsigned long long)bio->bi_iter.bi_sector);
  85                 goto out;
  86         }
  87
  88         SetPageUptodate(page);
  89 out:
  90         unlock_page(page);
  91         WRITE_ONCE(bio->bi_private, NULL);
  92         bio_put(bio);
  93         wake_up_process(waiter);
  94         put_task_struct(waiter);
  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 list.  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 = bmap(inode, probe_block);
 131                 if (first_block == 0)
 132                         goto bad_bmap;
 133
 134                 /*
 135                  * It must be PAGE_SIZE aligned on-disk
 136                  */
 137                 if (first_block & (blocks_per_page - 1)) {
 138                         probe_block++;
 139                         goto reprobe;
 140                 }
 141
 142                 for (block_in_page = 1; block_in_page < blocks_per_page;
 143                                         block_in_page++) {
 144                         sector_t block;
 145
 146                         block = bmap(inode, probe_block + block_in_page);
 147                         if (block == 0)
 148                                 goto bad_bmap;
 149                         if (block != first_block + block_in_page) {
 150                                 /* Discontiguity */
 151                                 probe_block++;
 152                                 goto reprobe;
 153                         }
 154                 }
 155
 156                 first_block >>= (PAGE_SHIFT - blkbits);
 157                 if (page_no) {  /* exclude the header page */
 158                         if (first_block < lowest_block)
 159                                 lowest_block = first_block;
 160                         if (first_block > highest_block)
 161                                 highest_block = first_block;
 162                 }
 163
 164                 /*
 165                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
 166                  */
 167                 ret = add_swap_extent(sis, page_no, 1, first_block);
 168                 if (ret < 0)
 169                         goto out;
 170                 nr_extents += ret;
 171                 page_no++;
 172                 probe_block += blocks_per_page;
 173 reprobe:
 174                 continue;
 175         }
 176         ret = nr_extents;
 177         *span = 1 + highest_block - lowest_block;
 178         if (page_no == 0)
 179                 page_no = 1;    /* force Empty message */
 180         sis->max = page_no;
 181         sis->pages = page_no - 1;
 182         sis->highest_bit = page_no - 1;
 183 out:
 184         return ret;
 185 bad_bmap:
 186         pr_err("swapon: swapfile has holes\n");
 187         ret = -EINVAL;
 188         goto out;
 189 }
 190
 191 /*
 192  * We may have stale swap cache pages in memory: notice
 193  * them here and get rid of the unnecessary final write.
 194  */
 195 int swap_writepage(struct page *page, struct writeback_control *wbc)
 196 {
 197         int ret = 0;
 198
 199         if (try_to_free_swap(page)) {
 200                 unlock_page(page);
 201                 goto out;
 202         }
 203         if (frontswap_store(page) == 0) {
 204                 set_page_writeback(page);
 205                 unlock_page(page);
 206                 end_page_writeback(page);
 207                 goto out;
 208         }
 209         ret = __swap_writepage(page, wbc, end_swap_bio_write);
 210 out:
 211         return ret;
 212 }
 213
 214 static inline void count_swpout_vm_event(struct page *page)
 215 {
 216 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 217         if (unlikely(PageTransHuge(page)))
 218                 count_vm_event(THP_SWPOUT);
 219 #endif
 220         count_vm_events(PSWPOUT, hpage_nr_pages(page));
 221 }
 222
 223 int __swap_writepage(struct page *page, struct writeback_control *wbc,
 224                 bio_end_io_t end_write_func)
 225 {
 226         struct bio *bio;
 227         int ret;
 228         struct swap_info_struct *sis = page_swap_info(page);
 229
 230         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 231         if (sis->flags & SWP_FILE) {
 232                 struct kiocb kiocb;
 233                 struct file *swap_file = sis->swap_file;
 234                 struct address_space *mapping = swap_file->f_mapping;
 235                 struct bio_vec bv = {
 236                         .bv_page = page,
 237                         .bv_len  = PAGE_SIZE,
 238                         .bv_offset = 0
 239                 };
 240                 struct iov_iter from;
 241
 242                 iov_iter_bvec(&from, ITER_BVEC | WRITE, &bv, 1, PAGE_SIZE);
 243                 init_sync_kiocb(&kiocb, swap_file);
 244                 kiocb.ki_pos = page_file_offset(page);
 245
 246                 set_page_writeback(page);
 247                 unlock_page(page);
 248                 ret = mapping->a_ops->direct_IO(&kiocb, &from);
 249                 if (ret == PAGE_SIZE) {
 250                         count_vm_event(PSWPOUT);
 251                         ret = 0;
 252                 } else {
 253                         /*
 254                          * In the case of swap-over-nfs, this can be a
 255                          * temporary failure if the system has limited
 256                          * memory for allocating transmit buffers.
 257                          * Mark the page dirty and avoid
 258                          * rotate_reclaimable_page but rate-limit the
 259                          * messages but do not flag PageError like
 260                          * the normal direct-to-bio case as it could
 261                          * be temporary.
 262                          */
 263                         set_page_dirty(page);
 264                         ClearPageReclaim(page);
 265                         pr_err_ratelimited("Write error on dio swapfile (%llu)\n",
 266                                            page_file_offset(page));
 267                 }
 268                 end_page_writeback(page);
 269                 return ret;
 270         }
 271
 272         ret = bdev_write_page(sis->bdev, map_swap_page(page, &sis->bdev),
 273                               page, wbc);
 274         if (!ret) {
 275                 count_swpout_vm_event(page);
 276                 return 0;
 277         }
 278
 279         ret = 0;
 280         bio = get_swap_bio(GFP_NOIO, page, end_write_func);
 281         if (bio == NULL) {
 282                 set_page_dirty(page);
 283                 unlock_page(page);
 284                 ret = -ENOMEM;
 285                 goto out;
 286         }
 287         bio->bi_opf = REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc);
 288         bio_associate_blkcg_from_page(bio, page);
 289         count_swpout_vm_event(page);
 290         set_page_writeback(page);
 291         unlock_page(page);
 292         submit_bio(bio);
 293 out:
 294         return ret;
 295 }
 296
 297 int swap_readpage(struct page *page, bool synchronous)
 298 {
 299         struct bio *bio;
 300         int ret = 0;
 301         struct swap_info_struct *sis = page_swap_info(page);
 302         blk_qc_t qc;
 303         struct gendisk *disk;
 304
 305         VM_BUG_ON_PAGE(!PageSwapCache(page) && !synchronous, page);
 306         VM_BUG_ON_PAGE(!PageLocked(page), page);
 307         VM_BUG_ON_PAGE(PageUptodate(page), page);
 308         if (frontswap_load(page) == 0) {
 309                 SetPageUptodate(page);
 310                 unlock_page(page);
 311                 goto out;
 312         }
 313
 314         if (sis->flags & SWP_FILE) {
 315                 struct file *swap_file = sis->swap_file;
 316                 struct address_space *mapping = swap_file->f_mapping;
 317
 318                 ret = mapping->a_ops->readpage(swap_file, page);
 319                 if (!ret)
 320                         count_vm_event(PSWPIN);
 321                 return ret;
 322         }
 323
 324         ret = bdev_read_page(sis->bdev, map_swap_page(page, &sis->bdev), page);
 325         if (!ret) {
 326                 count_vm_event(PSWPIN);
 327                 return 0;
 328         }
 329
 330         ret = 0;
 331         bio = get_swap_bio(GFP_KERNEL, page, end_swap_bio_read);
 332         if (bio == NULL) {
 333                 unlock_page(page);
 334                 ret = -ENOMEM;
 335                 goto out;
 336         }
 337         disk = bio->bi_disk;
 338         /*
 339          * Keep this task valid during swap readpage because the oom killer may
 340          * attempt to access it in the page fault retry time check.
 341          */
 342         get_task_struct(current);
 343         bio->bi_private = current;
 344         bio_set_op_attrs(bio, REQ_OP_READ, 0);
 345         count_vm_event(PSWPIN);
 346         bio_get(bio);
 347         qc = submit_bio(bio);
 348         while (synchronous) {
 349                 set_current_state(TASK_UNINTERRUPTIBLE);
 350                 if (!READ_ONCE(bio->bi_private))
 351                         break;
 352
 353                 if (!blk_poll(disk->queue, qc))
 354                         break;
 355         }
 356         __set_current_state(TASK_RUNNING);
 357         bio_put(bio);
 358
 359 out:
 360         return ret;
 361 }
 362
 363 int swap_set_page_dirty(struct page *page)
 364 {
 365         struct swap_info_struct *sis = page_swap_info(page);
 366
 367         if (sis->flags & SWP_FILE) {
 368                 struct address_space *mapping = sis->swap_file->f_mapping;
 369
 370                 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 371                 return mapping->a_ops->set_page_dirty(page);
 372         } else {
 373                 return __set_page_dirty_no_writeback(page);
 374         }
 375 }