mm/frontswap.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Frontswap frontend
   4  *
   5  * This code provides the generic "frontend" layer to call a matching
   6  * "backend" driver implementation of frontswap.  See
   7  * Documentation/vm/frontswap.rst for more information.
   8  *
   9  * Copyright (C) 2009-2012 Oracle Corp.  All rights reserved.
  10  * Author: Dan Magenheimer
  11  */
  12
  13 #include <linux/mman.h>
  14 #include <linux/swap.h>
  15 #include <linux/swapops.h>
  16 #include <linux/security.h>
  17 #include <linux/module.h>
  18 #include <linux/debugfs.h>
  19 #include <linux/frontswap.h>
  20 #include <linux/swapfile.h>
  21
  22 DEFINE_STATIC_KEY_FALSE(frontswap_enabled_key);
  23
  24 /*
  25  * frontswap_ops are added by frontswap_register_ops, and provide the
  26  * frontswap "backend" implementation functions.  Multiple implementations
  27  * may be registered, but implementations can never deregister.  This
  28  * is a simple singly-linked list of all registered implementations.
  29  */
  30 static struct frontswap_ops *frontswap_ops __read_mostly;
  31
  32 #define for_each_frontswap_ops(ops)             \
  33         for ((ops) = frontswap_ops; (ops); (ops) = (ops)->next)
  34
  35 /*
  36  * If enabled, frontswap_store will return failure even on success.  As
  37  * a result, the swap subsystem will always write the page to swap, in
  38  * effect converting frontswap into a writethrough cache.  In this mode,
  39  * there is no direct reduction in swap writes, but a frontswap backend
  40  * can unilaterally "reclaim" any pages in use with no data loss, thus
  41  * providing increases control over maximum memory usage due to frontswap.
  42  */
  43 static bool frontswap_writethrough_enabled __read_mostly;
  44
  45 /*
  46  * If enabled, the underlying tmem implementation is capable of doing
  47  * exclusive gets, so frontswap_load, on a successful tmem_get must
  48  * mark the page as no longer in frontswap AND mark it dirty.
  49  */
  50 static bool frontswap_tmem_exclusive_gets_enabled __read_mostly;
  51
  52 #ifdef CONFIG_DEBUG_FS
  53 /*
  54  * Counters available via /sys/kernel/debug/frontswap (if debugfs is
  55  * properly configured).  These are for information only so are not protected
  56  * against increment races.
  57  */
  58 static u64 frontswap_loads;
  59 static u64 frontswap_succ_stores;
  60 static u64 frontswap_failed_stores;
  61 static u64 frontswap_invalidates;
  62
  63 static inline void inc_frontswap_loads(void) {
  64         data_race(frontswap_loads++);
  65 }
  66 static inline void inc_frontswap_succ_stores(void) {
  67         data_race(frontswap_succ_stores++);
  68 }
  69 static inline void inc_frontswap_failed_stores(void) {
  70         data_race(frontswap_failed_stores++);
  71 }
  72 static inline void inc_frontswap_invalidates(void) {
  73         data_race(frontswap_invalidates++);
  74 }
  75 #else
  76 static inline void inc_frontswap_loads(void) { }
  77 static inline void inc_frontswap_succ_stores(void) { }
  78 static inline void inc_frontswap_failed_stores(void) { }
  79 static inline void inc_frontswap_invalidates(void) { }
  80 #endif
  81
  82 /*
  83  * Due to the asynchronous nature of the backends loading potentially
  84  * _after_ the swap system has been activated, we have chokepoints
  85  * on all frontswap functions to not call the backend until the backend
  86  * has registered.
  87  *
  88  * This would not guards us against the user deciding to call swapoff right as
  89  * we are calling the backend to initialize (so swapon is in action).
  90  * Fortunately for us, the swapon_mutex has been taken by the callee so we are
  91  * OK. The other scenario where calls to frontswap_store (called via
  92  * swap_writepage) is racing with frontswap_invalidate_area (called via
  93  * swapoff) is again guarded by the swap subsystem.
  94  *
  95  * While no backend is registered all calls to frontswap_[store|load|
  96  * invalidate_area|invalidate_page] are ignored or fail.
  97  *
  98  * The time between the backend being registered and the swap file system
  99  * calling the backend (via the frontswap_* functions) is indeterminate as
 100  * frontswap_ops is not atomic_t (or a value guarded by a spinlock).
 101  * That is OK as we are comfortable missing some of these calls to the newly
 102  * registered backend.
 103  *
 104  * Obviously the opposite (unloading the backend) must be done after all
 105  * the frontswap_[store|load|invalidate_area|invalidate_page] start
 106  * ignoring or failing the requests.  However, there is currently no way
 107  * to unload a backend once it is registered.
 108  */
 109
 110 /*
 111  * Register operations for frontswap
 112  */
 113 void frontswap_register_ops(struct frontswap_ops *ops)
 114 {
 115         DECLARE_BITMAP(a, MAX_SWAPFILES);
 116         DECLARE_BITMAP(b, MAX_SWAPFILES);
 117         struct swap_info_struct *si;
 118         unsigned int i;
 119
 120         bitmap_zero(a, MAX_SWAPFILES);
 121         bitmap_zero(b, MAX_SWAPFILES);
 122
 123         spin_lock(&swap_lock);
 124         plist_for_each_entry(si, &swap_active_head, list) {
 125                 if (!WARN_ON(!si->frontswap_map))
 126                         set_bit(si->type, a);
 127         }
 128         spin_unlock(&swap_lock);
 129
 130         /* the new ops needs to know the currently active swap devices */
 131         for_each_set_bit(i, a, MAX_SWAPFILES)
 132                 ops->init(i);
 133
 134         /*
 135          * Setting frontswap_ops must happen after the ops->init() calls
 136          * above; cmpxchg implies smp_mb() which will ensure the init is
 137          * complete at this point.
 138          */
 139         do {
 140                 ops->next = frontswap_ops;
 141         } while (cmpxchg(&frontswap_ops, ops->next, ops) != ops->next);
 142
 143         static_branch_inc(&frontswap_enabled_key);
 144
 145         spin_lock(&swap_lock);
 146         plist_for_each_entry(si, &swap_active_head, list) {
 147                 if (si->frontswap_map)
 148                         set_bit(si->type, b);
 149         }
 150         spin_unlock(&swap_lock);
 151
 152         /*
 153          * On the very unlikely chance that a swap device was added or
 154          * removed between setting the "a" list bits and the ops init
 155          * calls, we re-check and do init or invalidate for any changed
 156          * bits.
 157          */
 158         if (unlikely(!bitmap_equal(a, b, MAX_SWAPFILES))) {
 159                 for (i = 0; i < MAX_SWAPFILES; i++) {
 160                         if (!test_bit(i, a) && test_bit(i, b))
 161                                 ops->init(i);
 162                         else if (test_bit(i, a) && !test_bit(i, b))
 163                                 ops->invalidate_area(i);
 164                 }
 165         }
 166 }
 167 EXPORT_SYMBOL(frontswap_register_ops);
 168
 169 /*
 170  * Enable/disable frontswap writethrough (see above).
 171  */
 172 void frontswap_writethrough(bool enable)
 173 {
 174         frontswap_writethrough_enabled = enable;
 175 }
 176 EXPORT_SYMBOL(frontswap_writethrough);
 177
 178 /*
 179  * Enable/disable frontswap exclusive gets (see above).
 180  */
 181 void frontswap_tmem_exclusive_gets(bool enable)
 182 {
 183         frontswap_tmem_exclusive_gets_enabled = enable;
 184 }
 185 EXPORT_SYMBOL(frontswap_tmem_exclusive_gets);
 186
 187 /*
 188  * Called when a swap device is swapon'd.
 189  */
 190 void __frontswap_init(unsigned type, unsigned long *map)
 191 {
 192         struct swap_info_struct *sis = swap_info[type];
 193         struct frontswap_ops *ops;
 194
 195         VM_BUG_ON(sis == NULL);
 196
 197         /*
 198          * p->frontswap is a bitmap that we MUST have to figure out which page
 199          * has gone in frontswap. Without it there is no point of continuing.
 200          */
 201         if (WARN_ON(!map))
 202                 return;
 203         /*
 204          * Irregardless of whether the frontswap backend has been loaded
 205          * before this function or it will be later, we _MUST_ have the
 206          * p->frontswap set to something valid to work properly.
 207          */
 208         frontswap_map_set(sis, map);
 209
 210         for_each_frontswap_ops(ops)
 211                 ops->init(type);
 212 }
 213 EXPORT_SYMBOL(__frontswap_init);
 214
 215 bool __frontswap_test(struct swap_info_struct *sis,
 216                                 pgoff_t offset)
 217 {
 218         if (sis->frontswap_map)
 219                 return test_bit(offset, sis->frontswap_map);
 220         return false;
 221 }
 222 EXPORT_SYMBOL(__frontswap_test);
 223
 224 static inline void __frontswap_set(struct swap_info_struct *sis,
 225                                    pgoff_t offset)
 226 {
 227         set_bit(offset, sis->frontswap_map);
 228         atomic_inc(&sis->frontswap_pages);
 229 }
 230
 231 static inline void __frontswap_clear(struct swap_info_struct *sis,
 232                                      pgoff_t offset)
 233 {
 234         clear_bit(offset, sis->frontswap_map);
 235         atomic_dec(&sis->frontswap_pages);
 236 }
 237
 238 /*
 239  * "Store" data from a page to frontswap and associate it with the page's
 240  * swaptype and offset.  Page must be locked and in the swap cache.
 241  * If frontswap already contains a page with matching swaptype and
 242  * offset, the frontswap implementation may either overwrite the data and
 243  * return success or invalidate the page from frontswap and return failure.
 244  */
 245 int __frontswap_store(struct page *page)
 246 {
 247         int ret = -1;
 248         swp_entry_t entry = { .val = page_private(page), };
 249         int type = swp_type(entry);
 250         struct swap_info_struct *sis = swap_info[type];
 251         pgoff_t offset = swp_offset(entry);
 252         struct frontswap_ops *ops;
 253
 254         VM_BUG_ON(!frontswap_ops);
 255         VM_BUG_ON(!PageLocked(page));
 256         VM_BUG_ON(sis == NULL);
 257
 258         /*
 259          * If a dup, we must remove the old page first; we can't leave the
 260          * old page no matter if the store of the new page succeeds or fails,
 261          * and we can't rely on the new page replacing the old page as we may
 262          * not store to the same implementation that contains the old page.
 263          */
 264         if (__frontswap_test(sis, offset)) {
 265                 __frontswap_clear(sis, offset);
 266                 for_each_frontswap_ops(ops)
 267                         ops->invalidate_page(type, offset);
 268         }
 269
 270         /* Try to store in each implementation, until one succeeds. */
 271         for_each_frontswap_ops(ops) {
 272                 ret = ops->store(type, offset, page);
 273                 if (!ret) /* successful store */
 274                         break;
 275         }
 276         if (ret == 0) {
 277                 __frontswap_set(sis, offset);
 278                 inc_frontswap_succ_stores();
 279         } else {
 280                 inc_frontswap_failed_stores();
 281         }
 282         if (frontswap_writethrough_enabled)
 283                 /* report failure so swap also writes to swap device */
 284                 ret = -1;
 285         return ret;
 286 }
 287 EXPORT_SYMBOL(__frontswap_store);
 288
 289 /*
 290  * "Get" data from frontswap associated with swaptype and offset that were
 291  * specified when the data was put to frontswap and use it to fill the
 292  * specified page with data. Page must be locked and in the swap cache.
 293  */
 294 int __frontswap_load(struct page *page)
 295 {
 296         int ret = -1;
 297         swp_entry_t entry = { .val = page_private(page), };
 298         int type = swp_type(entry);
 299         struct swap_info_struct *sis = swap_info[type];
 300         pgoff_t offset = swp_offset(entry);
 301         struct frontswap_ops *ops;
 302
 303         VM_BUG_ON(!frontswap_ops);
 304         VM_BUG_ON(!PageLocked(page));
 305         VM_BUG_ON(sis == NULL);
 306
 307         if (!__frontswap_test(sis, offset))
 308                 return -1;
 309
 310         /* Try loading from each implementation, until one succeeds. */
 311         for_each_frontswap_ops(ops) {
 312                 ret = ops->load(type, offset, page);
 313                 if (!ret) /* successful load */
 314                         break;
 315         }
 316         if (ret == 0) {
 317                 inc_frontswap_loads();
 318                 if (frontswap_tmem_exclusive_gets_enabled) {
 319                         SetPageDirty(page);
 320                         __frontswap_clear(sis, offset);
 321                 }
 322         }
 323         return ret;
 324 }
 325 EXPORT_SYMBOL(__frontswap_load);
 326
 327 /*
 328  * Invalidate any data from frontswap associated with the specified swaptype
 329  * and offset so that a subsequent "get" will fail.
 330  */
 331 void __frontswap_invalidate_page(unsigned type, pgoff_t offset)
 332 {
 333         struct swap_info_struct *sis = swap_info[type];
 334         struct frontswap_ops *ops;
 335
 336         VM_BUG_ON(!frontswap_ops);
 337         VM_BUG_ON(sis == NULL);
 338
 339         if (!__frontswap_test(sis, offset))
 340                 return;
 341
 342         for_each_frontswap_ops(ops)
 343                 ops->invalidate_page(type, offset);
 344         __frontswap_clear(sis, offset);
 345         inc_frontswap_invalidates();
 346 }
 347 EXPORT_SYMBOL(__frontswap_invalidate_page);
 348
 349 /*
 350  * Invalidate all data from frontswap associated with all offsets for the
 351  * specified swaptype.
 352  */
 353 void __frontswap_invalidate_area(unsigned type)
 354 {
 355         struct swap_info_struct *sis = swap_info[type];
 356         struct frontswap_ops *ops;
 357
 358         VM_BUG_ON(!frontswap_ops);
 359         VM_BUG_ON(sis == NULL);
 360
 361         if (sis->frontswap_map == NULL)
 362                 return;
 363
 364         for_each_frontswap_ops(ops)
 365                 ops->invalidate_area(type);
 366         atomic_set(&sis->frontswap_pages, 0);
 367         bitmap_zero(sis->frontswap_map, sis->max);
 368 }
 369 EXPORT_SYMBOL(__frontswap_invalidate_area);
 370
 371 static unsigned long __frontswap_curr_pages(void)
 372 {
 373         unsigned long totalpages = 0;
 374         struct swap_info_struct *si = NULL;
 375
 376         assert_spin_locked(&swap_lock);
 377         plist_for_each_entry(si, &swap_active_head, list)
 378                 totalpages += atomic_read(&si->frontswap_pages);
 379         return totalpages;
 380 }
 381
 382 static int __frontswap_unuse_pages(unsigned long total, unsigned long *unused,
 383                                         int *swapid)
 384 {
 385         int ret = -EINVAL;
 386         struct swap_info_struct *si = NULL;
 387         int si_frontswap_pages;
 388         unsigned long total_pages_to_unuse = total;
 389         unsigned long pages = 0, pages_to_unuse = 0;
 390
 391         assert_spin_locked(&swap_lock);
 392         plist_for_each_entry(si, &swap_active_head, list) {
 393                 si_frontswap_pages = atomic_read(&si->frontswap_pages);
 394                 if (total_pages_to_unuse < si_frontswap_pages) {
 395                         pages = pages_to_unuse = total_pages_to_unuse;
 396                 } else {
 397                         pages = si_frontswap_pages;
 398                         pages_to_unuse = 0; /* unuse all */
 399                 }
 400                 /* ensure there is enough RAM to fetch pages from frontswap */
 401                 if (security_vm_enough_memory_mm(current->mm, pages)) {
 402                         ret = -ENOMEM;
 403                         continue;
 404                 }
 405                 vm_unacct_memory(pages);
 406                 *unused = pages_to_unuse;
 407                 *swapid = si->type;
 408                 ret = 0;
 409                 break;
 410         }
 411
 412         return ret;
 413 }
 414
 415 /*
 416  * Used to check if it's necessary and feasible to unuse pages.
 417  * Return 1 when nothing to do, 0 when need to shrink pages,
 418  * error code when there is an error.
 419  */
 420 static int __frontswap_shrink(unsigned long target_pages,
 421                                 unsigned long *pages_to_unuse,
 422                                 int *type)
 423 {
 424         unsigned long total_pages = 0, total_pages_to_unuse;
 425
 426         assert_spin_locked(&swap_lock);
 427
 428         total_pages = __frontswap_curr_pages();
 429         if (total_pages <= target_pages) {
 430                 /* Nothing to do */
 431                 *pages_to_unuse = 0;
 432                 return 1;
 433         }
 434         total_pages_to_unuse = total_pages - target_pages;
 435         return __frontswap_unuse_pages(total_pages_to_unuse, pages_to_unuse, type);
 436 }
 437
 438 /*
 439  * Frontswap, like a true swap device, may unnecessarily retain pages
 440  * under certain circumstances; "shrink" frontswap is essentially a
 441  * "partial swapoff" and works by calling try_to_unuse to attempt to
 442  * unuse enough frontswap pages to attempt to -- subject to memory
 443  * constraints -- reduce the number of pages in frontswap to the
 444  * number given in the parameter target_pages.
 445  */
 446 void frontswap_shrink(unsigned long target_pages)
 447 {
 448         unsigned long pages_to_unuse = 0;
 449         int type, ret;
 450
 451         /*
 452          * we don't want to hold swap_lock while doing a very
 453          * lengthy try_to_unuse, but swap_list may change
 454          * so restart scan from swap_active_head each time
 455          */
 456         spin_lock(&swap_lock);
 457         ret = __frontswap_shrink(target_pages, &pages_to_unuse, &type);
 458         spin_unlock(&swap_lock);
 459         if (ret == 0)
 460                 try_to_unuse(type, true, pages_to_unuse);
 461         return;
 462 }
 463 EXPORT_SYMBOL(frontswap_shrink);
 464
 465 /*
 466  * Count and return the number of frontswap pages across all
 467  * swap devices.  This is exported so that backend drivers can
 468  * determine current usage without reading debugfs.
 469  */
 470 unsigned long frontswap_curr_pages(void)
 471 {
 472         unsigned long totalpages = 0;
 473
 474         spin_lock(&swap_lock);
 475         totalpages = __frontswap_curr_pages();
 476         spin_unlock(&swap_lock);
 477
 478         return totalpages;
 479 }
 480 EXPORT_SYMBOL(frontswap_curr_pages);
 481
 482 static int __init init_frontswap(void)
 483 {
 484 #ifdef CONFIG_DEBUG_FS
 485         struct dentry *root = debugfs_create_dir("frontswap", NULL);
 486         if (root == NULL)
 487                 return -ENXIO;
 488         debugfs_create_u64("loads", 0444, root, &frontswap_loads);
 489         debugfs_create_u64("succ_stores", 0444, root, &frontswap_succ_stores);
 490         debugfs_create_u64("failed_stores", 0444, root,
 491                            &frontswap_failed_stores);
 492         debugfs_create_u64("invalidates", 0444, root, &frontswap_invalidates);
 493 #endif
 494         return 0;
 495 }
 496
 497 module_init(init_frontswap);