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 page *page = bio_first_page_all(bio); 33 34 if (bio->bi_status) { 35 SetPageError(page); 36 /* 37 * We failed to write the page out to swap-space. 38 * Re-dirty the page in order to avoid it being reclaimed. 39 * Also print a dire warning that things will go BAD (tm) 40 * very quickly. 41 * 42 * Also clear PG_reclaim to avoid folio_rotate_reclaimable() 43 */ 44 set_page_dirty(page); 45 pr_alert_ratelimited("Write-error on swap-device (%u:%u:%llu)\n", 46 MAJOR(bio_dev(bio)), MINOR(bio_dev(bio)), 47 (unsigned long long)bio->bi_iter.bi_sector); 48 ClearPageReclaim(page); 49 } 50 end_page_writeback(page); 51 } 52 53 static void end_swap_bio_write(struct bio *bio) 54 { 55 __end_swap_bio_write(bio); 56 bio_put(bio); 57 } 58 59 static void __end_swap_bio_read(struct bio *bio) 60 { 61 struct page *page = bio_first_page_all(bio); 62 63 if (bio->bi_status) { 64 SetPageError(page); 65 pr_alert_ratelimited("Read-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 } else { 69 SetPageUptodate(page); 70 } 71 unlock_page(page); 72 } 73 74 static void end_swap_bio_read(struct bio *bio) 75 { 76 __end_swap_bio_read(bio); 77 bio_put(bio); 78 } 79 80 int generic_swapfile_activate(struct swap_info_struct *sis, 81 struct file *swap_file, 82 sector_t *span) 83 { 84 struct address_space *mapping = swap_file->f_mapping; 85 struct inode *inode = mapping->host; 86 unsigned blocks_per_page; 87 unsigned long page_no; 88 unsigned blkbits; 89 sector_t probe_block; 90 sector_t last_block; 91 sector_t lowest_block = -1; 92 sector_t highest_block = 0; 93 int nr_extents = 0; 94 int ret; 95 96 blkbits = inode->i_blkbits; 97 blocks_per_page = PAGE_SIZE >> blkbits; 98 99 /* 100 * Map all the blocks into the extent tree. This code doesn't try 101 * to be very smart. 102 */ 103 probe_block = 0; 104 page_no = 0; 105 last_block = i_size_read(inode) >> blkbits; 106 while ((probe_block + blocks_per_page) <= last_block && 107 page_no < sis->max) { 108 unsigned block_in_page; 109 sector_t first_block; 110 111 cond_resched(); 112 113 first_block = probe_block; 114 ret = bmap(inode, &first_block); 115 if (ret || !first_block) 116 goto bad_bmap; 117 118 /* 119 * It must be PAGE_SIZE aligned on-disk 120 */ 121 if (first_block & (blocks_per_page - 1)) { 122 probe_block++; 123 goto reprobe; 124 } 125 126 for (block_in_page = 1; block_in_page < blocks_per_page; 127 block_in_page++) { 128 sector_t block; 129 130 block = probe_block + block_in_page; 131 ret = bmap(inode, &block); 132 if (ret || !block) 133 goto bad_bmap; 134 135 if (block != first_block + block_in_page) { 136 /* Discontiguity */ 137 probe_block++; 138 goto reprobe; 139 } 140 } 141 142 first_block >>= (PAGE_SHIFT - blkbits); 143 if (page_no) { /* exclude the header page */ 144 if (first_block < lowest_block) 145 lowest_block = first_block; 146 if (first_block > highest_block) 147 highest_block = first_block; 148 } 149 150 /* 151 * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks 152 */ 153 ret = add_swap_extent(sis, page_no, 1, first_block); 154 if (ret < 0) 155 goto out; 156 nr_extents += ret; 157 page_no++; 158 probe_block += blocks_per_page; 159 reprobe: 160 continue; 161 } 162 ret = nr_extents; 163 *span = 1 + highest_block - lowest_block; 164 if (page_no == 0) 165 page_no = 1; /* force Empty message */ 166 sis->max = page_no; 167 sis->pages = page_no - 1; 168 sis->highest_bit = page_no - 1; 169 out: 170 return ret; 171 bad_bmap: 172 pr_err("swapon: swapfile has holes\n"); 173 ret = -EINVAL; 174 goto out; 175 } 176 177 /* 178 * We may have stale swap cache pages in memory: notice 179 * them here and get rid of the unnecessary final write. 180 */ 181 int swap_writepage(struct page *page, struct writeback_control *wbc) 182 { 183 struct folio *folio = page_folio(page); 184 int ret; 185 186 if (folio_free_swap(folio)) { 187 folio_unlock(folio); 188 return 0; 189 } 190 /* 191 * Arch code may have to preserve more data than just the page 192 * contents, e.g. memory tags. 193 */ 194 ret = arch_prepare_to_swap(&folio->page); 195 if (ret) { 196 folio_mark_dirty(folio); 197 folio_unlock(folio); 198 return ret; 199 } 200 if (frontswap_store(&folio->page) == 0) { 201 folio_start_writeback(folio); 202 folio_unlock(folio); 203 folio_end_writeback(folio); 204 return 0; 205 } 206 __swap_writepage(&folio->page, wbc); 207 return 0; 208 } 209 210 static inline void count_swpout_vm_event(struct page *page) 211 { 212 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 213 if (unlikely(PageTransHuge(page))) 214 count_vm_event(THP_SWPOUT); 215 #endif 216 count_vm_events(PSWPOUT, thp_nr_pages(page)); 217 } 218 219 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP) 220 static void bio_associate_blkg_from_page(struct bio *bio, struct page *page) 221 { 222 struct cgroup_subsys_state *css; 223 struct mem_cgroup *memcg; 224 225 memcg = page_memcg(page); 226 if (!memcg) 227 return; 228 229 rcu_read_lock(); 230 css = cgroup_e_css(memcg->css.cgroup, &io_cgrp_subsys); 231 bio_associate_blkg_from_css(bio, css); 232 rcu_read_unlock(); 233 } 234 #else 235 #define bio_associate_blkg_from_page(bio, page) do { } while (0) 236 #endif /* CONFIG_MEMCG && CONFIG_BLK_CGROUP */ 237 238 struct swap_iocb { 239 struct kiocb iocb; 240 struct bio_vec bvec[SWAP_CLUSTER_MAX]; 241 int pages; 242 int len; 243 }; 244 static mempool_t *sio_pool; 245 246 int sio_pool_init(void) 247 { 248 if (!sio_pool) { 249 mempool_t *pool = mempool_create_kmalloc_pool( 250 SWAP_CLUSTER_MAX, sizeof(struct swap_iocb)); 251 if (cmpxchg(&sio_pool, NULL, pool)) 252 mempool_destroy(pool); 253 } 254 if (!sio_pool) 255 return -ENOMEM; 256 return 0; 257 } 258 259 static void sio_write_complete(struct kiocb *iocb, long ret) 260 { 261 struct swap_iocb *sio = container_of(iocb, struct swap_iocb, iocb); 262 struct page *page = sio->bvec[0].bv_page; 263 int p; 264 265 if (ret != sio->len) { 266 /* 267 * In the case of swap-over-nfs, this can be a 268 * temporary failure if the system has limited 269 * memory for allocating transmit buffers. 270 * Mark the page dirty and avoid 271 * folio_rotate_reclaimable but rate-limit the 272 * messages but do not flag PageError like 273 * the normal direct-to-bio case as it could 274 * be temporary. 275 */ 276 pr_err_ratelimited("Write error %ld on dio swapfile (%llu)\n", 277 ret, page_file_offset(page)); 278 for (p = 0; p < sio->pages; p++) { 279 page = sio->bvec[p].bv_page; 280 set_page_dirty(page); 281 ClearPageReclaim(page); 282 } 283 } else { 284 for (p = 0; p < sio->pages; p++) 285 count_swpout_vm_event(sio->bvec[p].bv_page); 286 } 287 288 for (p = 0; p < sio->pages; p++) 289 end_page_writeback(sio->bvec[p].bv_page); 290 291 mempool_free(sio, sio_pool); 292 } 293 294 static void swap_writepage_fs(struct page *page, struct writeback_control *wbc) 295 { 296 struct swap_iocb *sio = NULL; 297 struct swap_info_struct *sis = page_swap_info(page); 298 struct file *swap_file = sis->swap_file; 299 loff_t pos = page_file_offset(page); 300 301 set_page_writeback(page); 302 unlock_page(page); 303 if (wbc->swap_plug) 304 sio = *wbc->swap_plug; 305 if (sio) { 306 if (sio->iocb.ki_filp != swap_file || 307 sio->iocb.ki_pos + sio->len != pos) { 308 swap_write_unplug(sio); 309 sio = NULL; 310 } 311 } 312 if (!sio) { 313 sio = mempool_alloc(sio_pool, GFP_NOIO); 314 init_sync_kiocb(&sio->iocb, swap_file); 315 sio->iocb.ki_complete = sio_write_complete; 316 sio->iocb.ki_pos = pos; 317 sio->pages = 0; 318 sio->len = 0; 319 } 320 bvec_set_page(&sio->bvec[sio->pages], page, thp_size(page), 0); 321 sio->len += thp_size(page); 322 sio->pages += 1; 323 if (sio->pages == ARRAY_SIZE(sio->bvec) || !wbc->swap_plug) { 324 swap_write_unplug(sio); 325 sio = NULL; 326 } 327 if (wbc->swap_plug) 328 *wbc->swap_plug = sio; 329 } 330 331 static void swap_writepage_bdev_sync(struct page *page, 332 struct writeback_control *wbc, struct swap_info_struct *sis) 333 { 334 struct bio_vec bv; 335 struct bio bio; 336 337 bio_init(&bio, sis->bdev, &bv, 1, 338 REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc)); 339 bio.bi_iter.bi_sector = swap_page_sector(page); 340 __bio_add_page(&bio, page, thp_size(page), 0); 341 342 bio_associate_blkg_from_page(&bio, page); 343 count_swpout_vm_event(page); 344 345 set_page_writeback(page); 346 unlock_page(page); 347 348 submit_bio_wait(&bio); 349 __end_swap_bio_write(&bio); 350 } 351 352 static void swap_writepage_bdev_async(struct page *page, 353 struct writeback_control *wbc, struct swap_info_struct *sis) 354 { 355 struct bio *bio; 356 357 bio = bio_alloc(sis->bdev, 1, 358 REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc), 359 GFP_NOIO); 360 bio->bi_iter.bi_sector = swap_page_sector(page); 361 bio->bi_end_io = end_swap_bio_write; 362 __bio_add_page(bio, page, thp_size(page), 0); 363 364 bio_associate_blkg_from_page(bio, page); 365 count_swpout_vm_event(page); 366 set_page_writeback(page); 367 unlock_page(page); 368 submit_bio(bio); 369 } 370 371 void __swap_writepage(struct page *page, struct writeback_control *wbc) 372 { 373 struct swap_info_struct *sis = page_swap_info(page); 374 375 VM_BUG_ON_PAGE(!PageSwapCache(page), page); 376 /* 377 * ->flags can be updated non-atomicially (scan_swap_map_slots), 378 * but that will never affect SWP_FS_OPS, so the data_race 379 * is safe. 380 */ 381 if (data_race(sis->flags & SWP_FS_OPS)) 382 swap_writepage_fs(page, wbc); 383 else if (sis->flags & SWP_SYNCHRONOUS_IO) 384 swap_writepage_bdev_sync(page, wbc, sis); 385 else 386 swap_writepage_bdev_async(page, wbc, sis); 387 } 388 389 void swap_write_unplug(struct swap_iocb *sio) 390 { 391 struct iov_iter from; 392 struct address_space *mapping = sio->iocb.ki_filp->f_mapping; 393 int ret; 394 395 iov_iter_bvec(&from, ITER_SOURCE, sio->bvec, sio->pages, sio->len); 396 ret = mapping->a_ops->swap_rw(&sio->iocb, &from); 397 if (ret != -EIOCBQUEUED) 398 sio_write_complete(&sio->iocb, ret); 399 } 400 401 static void sio_read_complete(struct kiocb *iocb, long ret) 402 { 403 struct swap_iocb *sio = container_of(iocb, struct swap_iocb, iocb); 404 int p; 405 406 if (ret == sio->len) { 407 for (p = 0; p < sio->pages; p++) { 408 struct page *page = sio->bvec[p].bv_page; 409 410 SetPageUptodate(page); 411 unlock_page(page); 412 } 413 count_vm_events(PSWPIN, sio->pages); 414 } else { 415 for (p = 0; p < sio->pages; p++) { 416 struct page *page = sio->bvec[p].bv_page; 417 418 SetPageError(page); 419 unlock_page(page); 420 } 421 pr_alert_ratelimited("Read-error on swap-device\n"); 422 } 423 mempool_free(sio, sio_pool); 424 } 425 426 static void swap_readpage_fs(struct page *page, 427 struct swap_iocb **plug) 428 { 429 struct swap_info_struct *sis = page_swap_info(page); 430 struct swap_iocb *sio = NULL; 431 loff_t pos = page_file_offset(page); 432 433 if (plug) 434 sio = *plug; 435 if (sio) { 436 if (sio->iocb.ki_filp != sis->swap_file || 437 sio->iocb.ki_pos + sio->len != pos) { 438 swap_read_unplug(sio); 439 sio = NULL; 440 } 441 } 442 if (!sio) { 443 sio = mempool_alloc(sio_pool, GFP_KERNEL); 444 init_sync_kiocb(&sio->iocb, sis->swap_file); 445 sio->iocb.ki_pos = pos; 446 sio->iocb.ki_complete = sio_read_complete; 447 sio->pages = 0; 448 sio->len = 0; 449 } 450 bvec_set_page(&sio->bvec[sio->pages], page, thp_size(page), 0); 451 sio->len += thp_size(page); 452 sio->pages += 1; 453 if (sio->pages == ARRAY_SIZE(sio->bvec) || !plug) { 454 swap_read_unplug(sio); 455 sio = NULL; 456 } 457 if (plug) 458 *plug = sio; 459 } 460 461 static void swap_readpage_bdev_sync(struct page *page, 462 struct swap_info_struct *sis) 463 { 464 struct bio_vec bv; 465 struct bio bio; 466 467 bio_init(&bio, sis->bdev, &bv, 1, REQ_OP_READ); 468 bio.bi_iter.bi_sector = swap_page_sector(page); 469 __bio_add_page(&bio, page, thp_size(page), 0); 470 /* 471 * Keep this task valid during swap readpage because the oom killer may 472 * attempt to access it in the page fault retry time check. 473 */ 474 get_task_struct(current); 475 count_vm_event(PSWPIN); 476 submit_bio_wait(&bio); 477 __end_swap_bio_read(&bio); 478 put_task_struct(current); 479 } 480 481 static void swap_readpage_bdev_async(struct page *page, 482 struct swap_info_struct *sis) 483 { 484 struct bio *bio; 485 486 bio = bio_alloc(sis->bdev, 1, REQ_OP_READ, GFP_KERNEL); 487 bio->bi_iter.bi_sector = swap_page_sector(page); 488 bio->bi_end_io = end_swap_bio_read; 489 __bio_add_page(bio, page, thp_size(page), 0); 490 count_vm_event(PSWPIN); 491 submit_bio(bio); 492 } 493 494 void swap_readpage(struct page *page, bool synchronous, struct swap_iocb **plug) 495 { 496 struct swap_info_struct *sis = page_swap_info(page); 497 bool workingset = PageWorkingset(page); 498 unsigned long pflags; 499 bool in_thrashing; 500 501 VM_BUG_ON_PAGE(!PageSwapCache(page) && !synchronous, page); 502 VM_BUG_ON_PAGE(!PageLocked(page), page); 503 VM_BUG_ON_PAGE(PageUptodate(page), page); 504 505 /* 506 * Count submission time as memory stall and delay. When the device 507 * is congested, or the submitting cgroup IO-throttled, submission 508 * can be a significant part of overall IO time. 509 */ 510 if (workingset) { 511 delayacct_thrashing_start(&in_thrashing); 512 psi_memstall_enter(&pflags); 513 } 514 delayacct_swapin_start(); 515 516 if (frontswap_load(page) == 0) { 517 SetPageUptodate(page); 518 unlock_page(page); 519 } else if (data_race(sis->flags & SWP_FS_OPS)) { 520 swap_readpage_fs(page, plug); 521 } else if (synchronous || (sis->flags & SWP_SYNCHRONOUS_IO)) { 522 swap_readpage_bdev_sync(page, sis); 523 } else { 524 swap_readpage_bdev_async(page, sis); 525 } 526 527 if (workingset) { 528 delayacct_thrashing_end(&in_thrashing); 529 psi_memstall_leave(&pflags); 530 } 531 delayacct_swapin_end(); 532 } 533 534 void __swap_read_unplug(struct swap_iocb *sio) 535 { 536 struct iov_iter from; 537 struct address_space *mapping = sio->iocb.ki_filp->f_mapping; 538 int ret; 539 540 iov_iter_bvec(&from, ITER_DEST, sio->bvec, sio->pages, sio->len); 541 ret = mapping->a_ops->swap_rw(&sio->iocb, &from); 542 if (ret != -EIOCBQUEUED) 543 sio_read_complete(&sio->iocb, ret); 544 } 545