mm/frontswap.c

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