drivers/staging/zcache/zcache.c

   1 /*
   2  * zcache.c
   3  *
   4  * Copyright (c) 2010,2011, Dan Magenheimer, Oracle Corp.
   5  * Copyright (c) 2010,2011, Nitin Gupta
   6  *
   7  * Zcache provides an in-kernel "host implementation" for transcendent memory
   8  * and, thus indirectly, for cleancache and frontswap.  Zcache includes two
   9  * page-accessible memory [1] interfaces, both utilizing lzo1x compression:
  10  * 1) "compression buddies" ("zbud") is used for ephemeral pages
  11  * 2) xvmalloc is used for persistent pages.
  12  * Xvmalloc (based on the TLSF allocator) has very low fragmentation
  13  * so maximizes space efficiency, while zbud allows pairs (and potentially,
  14  * in the future, more than a pair of) compressed pages to be closely linked
  15  * so that reclaiming can be done via the kernel's physical-page-oriented
  16  * "shrinker" interface.
  17  *
  18  * [1] For a definition of page-accessible memory (aka PAM), see:
  19  *   http://marc.info/?l=linux-mm&m=127811271605009
  20  */
  21
  22 #include <linux/cpu.h>
  23 #include <linux/highmem.h>
  24 #include <linux/list.h>
  25 #include <linux/lzo.h>
  26 #include <linux/slab.h>
  27 #include <linux/spinlock.h>
  28 #include <linux/types.h>
  29 #include <linux/atomic.h>
  30 #include "tmem.h"
  31
  32 #include "../zram/xvmalloc.h" /* if built in drivers/staging */
  33
  34 #if (!defined(CONFIG_CLEANCACHE) && !defined(CONFIG_FRONTSWAP))
  35 #error "zcache is useless without CONFIG_CLEANCACHE or CONFIG_FRONTSWAP"
  36 #endif
  37 #ifdef CONFIG_CLEANCACHE
  38 #include <linux/cleancache.h>
  39 #endif
  40 #ifdef CONFIG_FRONTSWAP
  41 #include <linux/frontswap.h>
  42 #endif
  43
  44 #if 0
  45 /* this is more aggressive but may cause other problems? */
  46 #define ZCACHE_GFP_MASK (GFP_ATOMIC | __GFP_NORETRY | __GFP_NOWARN)
  47 #else
  48 #define ZCACHE_GFP_MASK \
  49         (__GFP_FS | __GFP_NORETRY | __GFP_NOWARN | __GFP_NOMEMALLOC)
  50 #endif
  51
  52 /**********
  53  * Compression buddies ("zbud") provides for packing two (or, possibly
  54  * in the future, more) compressed ephemeral pages into a single "raw"
  55  * (physical) page and tracking them with data structures so that
  56  * the raw pages can be easily reclaimed.
  57  *
  58  * A zbud page ("zbpg") is an aligned page containing a list_head,
  59  * a lock, and two "zbud headers".  The remainder of the physical
  60  * page is divided up into aligned 64-byte "chunks" which contain
  61  * the compressed data for zero, one, or two zbuds.  Each zbpg
  62  * resides on: (1) an "unused list" if it has no zbuds; (2) a
  63  * "buddied" list if it is fully populated  with two zbuds; or
  64  * (3) one of PAGE_SIZE/64 "unbuddied" lists indexed by how many chunks
  65  * the one unbuddied zbud uses.  The data inside a zbpg cannot be
  66  * read or written unless the zbpg's lock is held.
  67  */
  68
  69 #define ZBH_SENTINEL  0x43214321
  70 #define ZBPG_SENTINEL  0xdeadbeef
  71
  72 #define ZBUD_MAX_BUDS 2
  73
  74 struct zbud_hdr {
  75         uint32_t pool_id;
  76         struct tmem_oid oid;
  77         uint32_t index;
  78         uint16_t size; /* compressed size in bytes, zero means unused */
  79         DECL_SENTINEL
  80 };
  81
  82 struct zbud_page {
  83         struct list_head bud_list;
  84         spinlock_t lock;
  85         struct zbud_hdr buddy[ZBUD_MAX_BUDS];
  86         DECL_SENTINEL
  87         /* followed by NUM_CHUNK aligned CHUNK_SIZE-byte chunks */
  88 };
  89
  90 #define CHUNK_SHIFT     6
  91 #define CHUNK_SIZE      (1 << CHUNK_SHIFT)
  92 #define CHUNK_MASK      (~(CHUNK_SIZE-1))
  93 #define NCHUNKS         (((PAGE_SIZE - sizeof(struct zbud_page)) & \
  94                                 CHUNK_MASK) >> CHUNK_SHIFT)
  95 #define MAX_CHUNK       (NCHUNKS-1)
  96
  97 static struct {
  98         struct list_head list;
  99         unsigned count;
 100 } zbud_unbuddied[NCHUNKS];
 101 /* list N contains pages with N chunks USED and NCHUNKS-N unused */
 102 /* element 0 is never used but optimizing that isn't worth it */
 103 static unsigned long zbud_cumul_chunk_counts[NCHUNKS];
 104
 105 struct list_head zbud_buddied_list;
 106 static unsigned long zcache_zbud_buddied_count;
 107
 108 /* protects the buddied list and all unbuddied lists */
 109 static DEFINE_SPINLOCK(zbud_budlists_spinlock);
 110
 111 static LIST_HEAD(zbpg_unused_list);
 112 static unsigned long zcache_zbpg_unused_list_count;
 113
 114 /* protects the unused page list */
 115 static DEFINE_SPINLOCK(zbpg_unused_list_spinlock);
 116
 117 static atomic_t zcache_zbud_curr_raw_pages;
 118 static atomic_t zcache_zbud_curr_zpages;
 119 static unsigned long zcache_zbud_curr_zbytes;
 120 static unsigned long zcache_zbud_cumul_zpages;
 121 static unsigned long zcache_zbud_cumul_zbytes;
 122 static unsigned long zcache_compress_poor;
 123
 124 /* forward references */
 125 static void *zcache_get_free_page(void);
 126 static void zcache_free_page(void *p);
 127
 128 /*
 129  * zbud helper functions
 130  */
 131
 132 static inline unsigned zbud_max_buddy_size(void)
 133 {
 134         return MAX_CHUNK << CHUNK_SHIFT;
 135 }
 136
 137 static inline unsigned zbud_size_to_chunks(unsigned size)
 138 {
 139         BUG_ON(size == 0 || size > zbud_max_buddy_size());
 140         return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
 141 }
 142
 143 static inline int zbud_budnum(struct zbud_hdr *zh)
 144 {
 145         unsigned offset = (unsigned long)zh & (PAGE_SIZE - 1);
 146         struct zbud_page *zbpg = NULL;
 147         unsigned budnum = -1U;
 148         int i;
 149
 150         for (i = 0; i < ZBUD_MAX_BUDS; i++)
 151                 if (offset == offsetof(typeof(*zbpg), buddy[i])) {
 152                         budnum = i;
 153                         break;
 154                 }
 155         BUG_ON(budnum == -1U);
 156         return budnum;
 157 }
 158
 159 static char *zbud_data(struct zbud_hdr *zh, unsigned size)
 160 {
 161         struct zbud_page *zbpg;
 162         char *p;
 163         unsigned budnum;
 164
 165         ASSERT_SENTINEL(zh, ZBH);
 166         budnum = zbud_budnum(zh);
 167         BUG_ON(size == 0 || size > zbud_max_buddy_size());
 168         zbpg = container_of(zh, struct zbud_page, buddy[budnum]);
 169         ASSERT_SPINLOCK(&zbpg->lock);
 170         p = (char *)zbpg;
 171         if (budnum == 0)
 172                 p += ((sizeof(struct zbud_page) + CHUNK_SIZE - 1) &
 173                                                         CHUNK_MASK);
 174         else if (budnum == 1)
 175                 p += PAGE_SIZE - ((size + CHUNK_SIZE - 1) & CHUNK_MASK);
 176         return p;
 177 }
 178
 179 /*
 180  * zbud raw page management
 181  */
 182
 183 static struct zbud_page *zbud_alloc_raw_page(void)
 184 {
 185         struct zbud_page *zbpg = NULL;
 186         struct zbud_hdr *zh0, *zh1;
 187         bool recycled = 0;
 188
 189         /* if any pages on the zbpg list, use one */
 190         spin_lock(&zbpg_unused_list_spinlock);
 191         if (!list_empty(&zbpg_unused_list)) {
 192                 zbpg = list_first_entry(&zbpg_unused_list,
 193                                 struct zbud_page, bud_list);
 194                 list_del_init(&zbpg->bud_list);
 195                 zcache_zbpg_unused_list_count--;
 196                 recycled = 1;
 197         }
 198         spin_unlock(&zbpg_unused_list_spinlock);
 199         if (zbpg == NULL)
 200                 /* none on zbpg list, try to get a kernel page */
 201                 zbpg = zcache_get_free_page();
 202         if (likely(zbpg != NULL)) {
 203                 INIT_LIST_HEAD(&zbpg->bud_list);
 204                 zh0 = &zbpg->buddy[0]; zh1 = &zbpg->buddy[1];
 205                 spin_lock_init(&zbpg->lock);
 206                 if (recycled) {
 207                         ASSERT_INVERTED_SENTINEL(zbpg, ZBPG);
 208                         SET_SENTINEL(zbpg, ZBPG);
 209                         BUG_ON(zh0->size != 0 || tmem_oid_valid(&zh0->oid));
 210                         BUG_ON(zh1->size != 0 || tmem_oid_valid(&zh1->oid));
 211                 } else {
 212                         atomic_inc(&zcache_zbud_curr_raw_pages);
 213                         INIT_LIST_HEAD(&zbpg->bud_list);
 214                         SET_SENTINEL(zbpg, ZBPG);
 215                         zh0->size = 0; zh1->size = 0;
 216                         tmem_oid_set_invalid(&zh0->oid);
 217                         tmem_oid_set_invalid(&zh1->oid);
 218                 }
 219         }
 220         return zbpg;
 221 }
 222
 223 static void zbud_free_raw_page(struct zbud_page *zbpg)
 224 {
 225         struct zbud_hdr *zh0 = &zbpg->buddy[0], *zh1 = &zbpg->buddy[1];
 226
 227         ASSERT_SENTINEL(zbpg, ZBPG);
 228         BUG_ON(!list_empty(&zbpg->bud_list));
 229         ASSERT_SPINLOCK(&zbpg->lock);
 230         BUG_ON(zh0->size != 0 || tmem_oid_valid(&zh0->oid));
 231         BUG_ON(zh1->size != 0 || tmem_oid_valid(&zh1->oid));
 232         INVERT_SENTINEL(zbpg, ZBPG);
 233         spin_unlock(&zbpg->lock);
 234         spin_lock(&zbpg_unused_list_spinlock);
 235         list_add(&zbpg->bud_list, &zbpg_unused_list);
 236         zcache_zbpg_unused_list_count++;
 237         spin_unlock(&zbpg_unused_list_spinlock);
 238 }
 239
 240 /*
 241  * core zbud handling routines
 242  */
 243
 244 static unsigned zbud_free(struct zbud_hdr *zh)
 245 {
 246         unsigned size;
 247
 248         ASSERT_SENTINEL(zh, ZBH);
 249         BUG_ON(!tmem_oid_valid(&zh->oid));
 250         size = zh->size;
 251         BUG_ON(zh->size == 0 || zh->size > zbud_max_buddy_size());
 252         zh->size = 0;
 253         tmem_oid_set_invalid(&zh->oid);
 254         INVERT_SENTINEL(zh, ZBH);
 255         zcache_zbud_curr_zbytes -= size;
 256         atomic_dec(&zcache_zbud_curr_zpages);
 257         return size;
 258 }
 259
 260 static void zbud_free_and_delist(struct zbud_hdr *zh)
 261 {
 262         unsigned chunks;
 263         struct zbud_hdr *zh_other;
 264         unsigned budnum = zbud_budnum(zh), size;
 265         struct zbud_page *zbpg =
 266                 container_of(zh, struct zbud_page, buddy[budnum]);
 267
 268         spin_lock(&zbpg->lock);
 269         if (list_empty(&zbpg->bud_list)) {
 270                 /* ignore zombie page... see zbud_evict_pages() */
 271                 spin_unlock(&zbpg->lock);
 272                 return;
 273         }
 274         size = zbud_free(zh);
 275         ASSERT_SPINLOCK(&zbpg->lock);
 276         zh_other = &zbpg->buddy[(budnum == 0) ? 1 : 0];
 277         if (zh_other->size == 0) { /* was unbuddied: unlist and free */
 278                 chunks = zbud_size_to_chunks(size) ;
 279                 spin_lock(&zbud_budlists_spinlock);
 280                 BUG_ON(list_empty(&zbud_unbuddied[chunks].list));
 281                 list_del_init(&zbpg->bud_list);
 282                 zbud_unbuddied[chunks].count--;
 283                 spin_unlock(&zbud_budlists_spinlock);
 284                 zbud_free_raw_page(zbpg);
 285         } else { /* was buddied: move remaining buddy to unbuddied list */
 286                 chunks = zbud_size_to_chunks(zh_other->size) ;
 287                 spin_lock(&zbud_budlists_spinlock);
 288                 list_del_init(&zbpg->bud_list);
 289                 zcache_zbud_buddied_count--;
 290                 list_add_tail(&zbpg->bud_list, &zbud_unbuddied[chunks].list);
 291                 zbud_unbuddied[chunks].count++;
 292                 spin_unlock(&zbud_budlists_spinlock);
 293                 spin_unlock(&zbpg->lock);
 294         }
 295 }
 296
 297 static struct zbud_hdr *zbud_create(uint32_t pool_id, struct tmem_oid *oid,
 298                                         uint32_t index, struct page *page,
 299                                         void *cdata, unsigned size)
 300 {
 301         struct zbud_hdr *zh0, *zh1, *zh = NULL;
 302         struct zbud_page *zbpg = NULL, *ztmp;
 303         unsigned nchunks;
 304         char *to;
 305         int i, found_good_buddy = 0;
 306
 307         nchunks = zbud_size_to_chunks(size) ;
 308         for (i = MAX_CHUNK - nchunks + 1; i > 0; i--) {
 309                 spin_lock(&zbud_budlists_spinlock);
 310                 if (!list_empty(&zbud_unbuddied[i].list)) {
 311                         list_for_each_entry_safe(zbpg, ztmp,
 312                                     &zbud_unbuddied[i].list, bud_list) {
 313                                 if (spin_trylock(&zbpg->lock)) {
 314                                         found_good_buddy = i;
 315                                         goto found_unbuddied;
 316                                 }
 317                         }
 318                 }
 319                 spin_unlock(&zbud_budlists_spinlock);
 320         }
 321         /* didn't find a good buddy, try allocating a new page */
 322         zbpg = zbud_alloc_raw_page();
 323         if (unlikely(zbpg == NULL))
 324                 goto out;
 325         /* ok, have a page, now compress the data before taking locks */
 326         spin_lock(&zbpg->lock);
 327         spin_lock(&zbud_budlists_spinlock);
 328         list_add_tail(&zbpg->bud_list, &zbud_unbuddied[nchunks].list);
 329         zbud_unbuddied[nchunks].count++;
 330         zh = &zbpg->buddy[0];
 331         goto init_zh;
 332
 333 found_unbuddied:
 334         ASSERT_SPINLOCK(&zbpg->lock);
 335         zh0 = &zbpg->buddy[0]; zh1 = &zbpg->buddy[1];
 336         BUG_ON(!((zh0->size == 0) ^ (zh1->size == 0)));
 337         if (zh0->size != 0) { /* buddy0 in use, buddy1 is vacant */
 338                 ASSERT_SENTINEL(zh0, ZBH);
 339                 zh = zh1;
 340         } else if (zh1->size != 0) { /* buddy1 in use, buddy0 is vacant */
 341                 ASSERT_SENTINEL(zh1, ZBH);
 342                 zh = zh0;
 343         } else
 344                 BUG();
 345         list_del_init(&zbpg->bud_list);
 346         zbud_unbuddied[found_good_buddy].count--;
 347         list_add_tail(&zbpg->bud_list, &zbud_buddied_list);
 348         zcache_zbud_buddied_count++;
 349
 350 init_zh:
 351         SET_SENTINEL(zh, ZBH);
 352         zh->size = size;
 353         zh->index = index;
 354         zh->oid = *oid;
 355         zh->pool_id = pool_id;
 356         /* can wait to copy the data until the list locks are dropped */
 357         spin_unlock(&zbud_budlists_spinlock);
 358
 359         to = zbud_data(zh, size);
 360         memcpy(to, cdata, size);
 361         spin_unlock(&zbpg->lock);
 362         zbud_cumul_chunk_counts[nchunks]++;
 363         atomic_inc(&zcache_zbud_curr_zpages);
 364         zcache_zbud_cumul_zpages++;
 365         zcache_zbud_curr_zbytes += size;
 366         zcache_zbud_cumul_zbytes += size;
 367 out:
 368         return zh;
 369 }
 370
 371 static int zbud_decompress(struct page *page, struct zbud_hdr *zh)
 372 {
 373         struct zbud_page *zbpg;
 374         unsigned budnum = zbud_budnum(zh);
 375         size_t out_len = PAGE_SIZE;
 376         char *to_va, *from_va;
 377         unsigned size;
 378         int ret = 0;
 379
 380         zbpg = container_of(zh, struct zbud_page, buddy[budnum]);
 381         spin_lock(&zbpg->lock);
 382         if (list_empty(&zbpg->bud_list)) {
 383                 /* ignore zombie page... see zbud_evict_pages() */
 384                 ret = -EINVAL;
 385                 goto out;
 386         }
 387         ASSERT_SENTINEL(zh, ZBH);
 388         BUG_ON(zh->size == 0 || zh->size > zbud_max_buddy_size());
 389         to_va = kmap_atomic(page, KM_USER0);
 390         size = zh->size;
 391         from_va = zbud_data(zh, size);
 392         ret = lzo1x_decompress_safe(from_va, size, to_va, &out_len);
 393         BUG_ON(ret != LZO_E_OK);
 394         BUG_ON(out_len != PAGE_SIZE);
 395         kunmap_atomic(to_va, KM_USER0);
 396 out:
 397         spin_unlock(&zbpg->lock);
 398         return ret;
 399 }
 400
 401 /*
 402  * The following routines handle shrinking of ephemeral pages by evicting
 403  * pages "least valuable" first.
 404  */
 405
 406 static unsigned long zcache_evicted_raw_pages;
 407 static unsigned long zcache_evicted_buddied_pages;
 408 static unsigned long zcache_evicted_unbuddied_pages;
 409
 410 static struct tmem_pool *zcache_get_pool_by_id(uint32_t poolid);
 411 static void zcache_put_pool(struct tmem_pool *pool);
 412
 413 /*
 414  * Flush and free all zbuds in a zbpg, then free the pageframe
 415  */
 416 static void zbud_evict_zbpg(struct zbud_page *zbpg)
 417 {
 418         struct zbud_hdr *zh;
 419         int i, j;
 420         uint32_t pool_id[ZBUD_MAX_BUDS], index[ZBUD_MAX_BUDS];
 421         struct tmem_oid oid[ZBUD_MAX_BUDS];
 422         struct tmem_pool *pool;
 423
 424         ASSERT_SPINLOCK(&zbpg->lock);
 425         BUG_ON(!list_empty(&zbpg->bud_list));
 426         for (i = 0, j = 0; i < ZBUD_MAX_BUDS; i++) {
 427                 zh = &zbpg->buddy[i];
 428                 if (zh->size) {
 429                         pool_id[j] = zh->pool_id;
 430                         oid[j] = zh->oid;
 431                         index[j] = zh->index;
 432                         j++;
 433                         zbud_free(zh);
 434                 }
 435         }
 436         spin_unlock(&zbpg->lock);
 437         for (i = 0; i < j; i++) {
 438                 pool = zcache_get_pool_by_id(pool_id[i]);
 439                 if (pool != NULL) {
 440                         tmem_flush_page(pool, &oid[i], index[i]);
 441                         zcache_put_pool(pool);
 442                 }
 443         }
 444         ASSERT_SENTINEL(zbpg, ZBPG);
 445         spin_lock(&zbpg->lock);
 446         zbud_free_raw_page(zbpg);
 447 }
 448
 449 /*
 450  * Free nr pages.  This code is funky because we want to hold the locks
 451  * protecting various lists for as short a time as possible, and in some
 452  * circumstances the list may change asynchronously when the list lock is
 453  * not held.  In some cases we also trylock not only to avoid waiting on a
 454  * page in use by another cpu, but also to avoid potential deadlock due to
 455  * lock inversion.
 456  */
 457 static void zbud_evict_pages(int nr)
 458 {
 459         struct zbud_page *zbpg;
 460         int i;
 461
 462         /* first try freeing any pages on unused list */
 463 retry_unused_list:
 464         spin_lock_bh(&zbpg_unused_list_spinlock);
 465         if (!list_empty(&zbpg_unused_list)) {
 466                 /* can't walk list here, since it may change when unlocked */
 467                 zbpg = list_first_entry(&zbpg_unused_list,
 468                                 struct zbud_page, bud_list);
 469                 list_del_init(&zbpg->bud_list);
 470                 zcache_zbpg_unused_list_count--;
 471                 atomic_dec(&zcache_zbud_curr_raw_pages);
 472                 spin_unlock_bh(&zbpg_unused_list_spinlock);
 473                 zcache_free_page(zbpg);
 474                 zcache_evicted_raw_pages++;
 475                 if (--nr <= 0)
 476                         goto out;
 477                 goto retry_unused_list;
 478         }
 479         spin_unlock_bh(&zbpg_unused_list_spinlock);
 480
 481         /* now try freeing unbuddied pages, starting with least space avail */
 482         for (i = 0; i < MAX_CHUNK; i++) {
 483 retry_unbud_list_i:
 484                 spin_lock_bh(&zbud_budlists_spinlock);
 485                 if (list_empty(&zbud_unbuddied[i].list)) {
 486                         spin_unlock_bh(&zbud_budlists_spinlock);
 487                         continue;
 488                 }
 489                 list_for_each_entry(zbpg, &zbud_unbuddied[i].list, bud_list) {
 490                         if (unlikely(!spin_trylock(&zbpg->lock)))
 491                                 continue;
 492                         list_del_init(&zbpg->bud_list);
 493                         zbud_unbuddied[i].count--;
 494                         spin_unlock(&zbud_budlists_spinlock);
 495                         zcache_evicted_unbuddied_pages++;
 496                         /* want budlists unlocked when doing zbpg eviction */
 497                         zbud_evict_zbpg(zbpg);
 498                         local_bh_enable();
 499                         if (--nr <= 0)
 500                                 goto out;
 501                         goto retry_unbud_list_i;
 502                 }
 503                 spin_unlock_bh(&zbud_budlists_spinlock);
 504         }
 505
 506         /* as a last resort, free buddied pages */
 507 retry_bud_list:
 508         spin_lock_bh(&zbud_budlists_spinlock);
 509         if (list_empty(&zbud_buddied_list)) {
 510                 spin_unlock_bh(&zbud_budlists_spinlock);
 511                 goto out;
 512         }
 513         list_for_each_entry(zbpg, &zbud_buddied_list, bud_list) {
 514                 if (unlikely(!spin_trylock(&zbpg->lock)))
 515                         continue;
 516                 list_del_init(&zbpg->bud_list);
 517                 zcache_zbud_buddied_count--;
 518                 spin_unlock(&zbud_budlists_spinlock);
 519                 zcache_evicted_buddied_pages++;
 520                 /* want budlists unlocked when doing zbpg eviction */
 521                 zbud_evict_zbpg(zbpg);
 522                 local_bh_enable();
 523                 if (--nr <= 0)
 524                         goto out;
 525                 goto retry_bud_list;
 526         }
 527         spin_unlock_bh(&zbud_budlists_spinlock);
 528 out:
 529         return;
 530 }
 531
 532 static void zbud_init(void)
 533 {
 534         int i;
 535
 536         INIT_LIST_HEAD(&zbud_buddied_list);
 537         zcache_zbud_buddied_count = 0;
 538         for (i = 0; i < NCHUNKS; i++) {
 539                 INIT_LIST_HEAD(&zbud_unbuddied[i].list);
 540                 zbud_unbuddied[i].count = 0;
 541         }
 542 }
 543
 544 #ifdef CONFIG_SYSFS
 545 /*
 546  * These sysfs routines show a nice distribution of how many zbpg's are
 547  * currently (and have ever been placed) in each unbuddied list.  It's fun
 548  * to watch but can probably go away before final merge.
 549  */
 550 static int zbud_show_unbuddied_list_counts(char *buf)
 551 {
 552         int i;
 553         char *p = buf;
 554
 555         for (i = 0; i < NCHUNKS - 1; i++)
 556                 p += sprintf(p, "%u ", zbud_unbuddied[i].count);
 557         p += sprintf(p, "%d\n", zbud_unbuddied[i].count);
 558         return p - buf;
 559 }
 560
 561 static int zbud_show_cumul_chunk_counts(char *buf)
 562 {
 563         unsigned long i, chunks = 0, total_chunks = 0, sum_total_chunks = 0;
 564         unsigned long total_chunks_lte_21 = 0, total_chunks_lte_32 = 0;
 565         unsigned long total_chunks_lte_42 = 0;
 566         char *p = buf;
 567
 568         for (i = 0; i < NCHUNKS; i++) {
 569                 p += sprintf(p, "%lu ", zbud_cumul_chunk_counts[i]);
 570                 chunks += zbud_cumul_chunk_counts[i];
 571                 total_chunks += zbud_cumul_chunk_counts[i];
 572                 sum_total_chunks += i * zbud_cumul_chunk_counts[i];
 573                 if (i == 21)
 574                         total_chunks_lte_21 = total_chunks;
 575                 if (i == 32)
 576                         total_chunks_lte_32 = total_chunks;
 577                 if (i == 42)
 578                         total_chunks_lte_42 = total_chunks;
 579         }
 580         p += sprintf(p, "<=21:%lu <=32:%lu <=42:%lu, mean:%lu\n",
 581                 total_chunks_lte_21, total_chunks_lte_32, total_chunks_lte_42,
 582                 chunks == 0 ? 0 : sum_total_chunks / chunks);
 583         return p - buf;
 584 }
 585 #endif
 586
 587 /**********
 588  * This "zv" PAM implementation combines the TLSF-based xvMalloc
 589  * with lzo1x compression to maximize the amount of data that can
 590  * be packed into a physical page.
 591  *
 592  * Zv represents a PAM page with the index and object (plus a "size" value
 593  * necessary for decompression) immediately preceding the compressed data.
 594  */
 595
 596 #define ZVH_SENTINEL  0x43214321
 597
 598 struct zv_hdr {
 599         uint32_t pool_id;
 600         struct tmem_oid oid;
 601         uint32_t index;
 602         DECL_SENTINEL
 603 };
 604
 605 static const int zv_max_page_size = (PAGE_SIZE / 8) * 7;
 606
 607 static struct zv_hdr *zv_create(struct xv_pool *xvpool, uint32_t pool_id,
 608                                 struct tmem_oid *oid, uint32_t index,
 609                                 void *cdata, unsigned clen)
 610 {
 611         struct page *page;
 612         struct zv_hdr *zv = NULL;
 613         uint32_t offset;
 614         int ret;
 615
 616         BUG_ON(!irqs_disabled());
 617         ret = xv_malloc(xvpool, clen + sizeof(struct zv_hdr),
 618                         &page, &offset, ZCACHE_GFP_MASK);
 619         if (unlikely(ret))
 620                 goto out;
 621         zv = kmap_atomic(page, KM_USER0) + offset;
 622         zv->index = index;
 623         zv->oid = *oid;
 624         zv->pool_id = pool_id;
 625         SET_SENTINEL(zv, ZVH);
 626         memcpy((char *)zv + sizeof(struct zv_hdr), cdata, clen);
 627         kunmap_atomic(zv, KM_USER0);
 628 out:
 629         return zv;
 630 }
 631
 632 static void zv_free(struct xv_pool *xvpool, struct zv_hdr *zv)
 633 {
 634         unsigned long flags;
 635         struct page *page;
 636         uint32_t offset;
 637         uint16_t size;
 638
 639         ASSERT_SENTINEL(zv, ZVH);
 640         size = xv_get_object_size(zv) - sizeof(*zv);
 641         BUG_ON(size == 0 || size > zv_max_page_size);
 642         INVERT_SENTINEL(zv, ZVH);
 643         page = virt_to_page(zv);
 644         offset = (unsigned long)zv & ~PAGE_MASK;
 645         local_irq_save(flags);
 646         xv_free(xvpool, page, offset);
 647         local_irq_restore(flags);
 648 }
 649
 650 static void zv_decompress(struct page *page, struct zv_hdr *zv)
 651 {
 652         size_t clen = PAGE_SIZE;
 653         char *to_va;
 654         unsigned size;
 655         int ret;
 656
 657         ASSERT_SENTINEL(zv, ZVH);
 658         size = xv_get_object_size(zv) - sizeof(*zv);
 659         BUG_ON(size == 0 || size > zv_max_page_size);
 660         to_va = kmap_atomic(page, KM_USER0);
 661         ret = lzo1x_decompress_safe((char *)zv + sizeof(*zv),
 662                                         size, to_va, &clen);
 663         kunmap_atomic(to_va, KM_USER0);
 664         BUG_ON(ret != LZO_E_OK);
 665         BUG_ON(clen != PAGE_SIZE);
 666 }
 667
 668 /*
 669  * zcache core code starts here
 670  */
 671
 672 /* useful stats not collected by cleancache or frontswap */
 673 static unsigned long zcache_flush_total;
 674 static unsigned long zcache_flush_found;
 675 static unsigned long zcache_flobj_total;
 676 static unsigned long zcache_flobj_found;
 677 static unsigned long zcache_failed_eph_puts;
 678 static unsigned long zcache_failed_pers_puts;
 679
 680 #define MAX_POOLS_PER_CLIENT 16
 681
 682 static struct {
 683         struct tmem_pool *tmem_pools[MAX_POOLS_PER_CLIENT];
 684         struct xv_pool *xvpool;
 685 } zcache_client;
 686
 687 /*
 688  * Tmem operations assume the poolid implies the invoking client.
 689  * Zcache only has one client (the kernel itself), so translate
 690  * the poolid into the tmem_pool allocated for it.  A KVM version
 691  * of zcache would have one client per guest and each client might
 692  * have a poolid==N.
 693  */
 694 static struct tmem_pool *zcache_get_pool_by_id(uint32_t poolid)
 695 {
 696         struct tmem_pool *pool = NULL;
 697
 698         if (poolid >= 0) {
 699                 pool = zcache_client.tmem_pools[poolid];
 700                 if (pool != NULL)
 701                         atomic_inc(&pool->refcount);
 702         }
 703         return pool;
 704 }
 705
 706 static void zcache_put_pool(struct tmem_pool *pool)
 707 {
 708         if (pool != NULL)
 709                 atomic_dec(&pool->refcount);
 710 }
 711
 712 /* counters for debugging */
 713 static unsigned long zcache_failed_get_free_pages;
 714 static unsigned long zcache_failed_alloc;
 715 static unsigned long zcache_put_to_flush;
 716 static unsigned long zcache_aborted_preload;
 717 static unsigned long zcache_aborted_shrink;
 718
 719 /*
 720  * Ensure that memory allocation requests in zcache don't result
 721  * in direct reclaim requests via the shrinker, which would cause
 722  * an infinite loop.  Maybe a GFP flag would be better?
 723  */
 724 static DEFINE_SPINLOCK(zcache_direct_reclaim_lock);
 725
 726 /*
 727  * for now, used named slabs so can easily track usage; later can
 728  * either just use kmalloc, or perhaps add a slab-like allocator
 729  * to more carefully manage total memory utilization
 730  */
 731 static struct kmem_cache *zcache_objnode_cache;
 732 static struct kmem_cache *zcache_obj_cache;
 733 static atomic_t zcache_curr_obj_count = ATOMIC_INIT(0);
 734 static unsigned long zcache_curr_obj_count_max;
 735 static atomic_t zcache_curr_objnode_count = ATOMIC_INIT(0);
 736 static unsigned long zcache_curr_objnode_count_max;
 737
 738 /*
 739  * to avoid memory allocation recursion (e.g. due to direct reclaim), we
 740  * preload all necessary data structures so the hostops callbacks never
 741  * actually do a malloc
 742  */
 743 struct zcache_preload {
 744         void *page;
 745         struct tmem_obj *obj;
 746         int nr;
 747         struct tmem_objnode *objnodes[OBJNODE_TREE_MAX_PATH];
 748 };
 749 static DEFINE_PER_CPU(struct zcache_preload, zcache_preloads) = { 0, };
 750
 751 static int zcache_do_preload(struct tmem_pool *pool)
 752 {
 753         struct zcache_preload *kp;
 754         struct tmem_objnode *objnode;
 755         struct tmem_obj *obj;
 756         void *page;
 757         int ret = -ENOMEM;
 758
 759         if (unlikely(zcache_objnode_cache == NULL))
 760                 goto out;
 761         if (unlikely(zcache_obj_cache == NULL))
 762                 goto out;
 763         if (!spin_trylock(&zcache_direct_reclaim_lock)) {
 764                 zcache_aborted_preload++;
 765                 goto out;
 766         }
 767         preempt_disable();
 768         kp = &__get_cpu_var(zcache_preloads);
 769         while (kp->nr < ARRAY_SIZE(kp->objnodes)) {
 770                 preempt_enable_no_resched();
 771                 objnode = kmem_cache_alloc(zcache_objnode_cache,
 772                                 ZCACHE_GFP_MASK);
 773                 if (unlikely(objnode == NULL)) {
 774                         zcache_failed_alloc++;
 775                         goto unlock_out;
 776                 }
 777                 preempt_disable();
 778                 kp = &__get_cpu_var(zcache_preloads);
 779                 if (kp->nr < ARRAY_SIZE(kp->objnodes))
 780                         kp->objnodes[kp->nr++] = objnode;
 781                 else
 782                         kmem_cache_free(zcache_objnode_cache, objnode);
 783         }
 784         preempt_enable_no_resched();
 785         obj = kmem_cache_alloc(zcache_obj_cache, ZCACHE_GFP_MASK);
 786         if (unlikely(obj == NULL)) {
 787                 zcache_failed_alloc++;
 788                 goto unlock_out;
 789         }
 790         page = (void *)__get_free_page(ZCACHE_GFP_MASK);
 791         if (unlikely(page == NULL)) {
 792                 zcache_failed_get_free_pages++;
 793                 kmem_cache_free(zcache_obj_cache, obj);
 794                 goto unlock_out;
 795         }
 796         preempt_disable();
 797         kp = &__get_cpu_var(zcache_preloads);
 798         if (kp->obj == NULL)
 799                 kp->obj = obj;
 800         else
 801                 kmem_cache_free(zcache_obj_cache, obj);
 802         if (kp->page == NULL)
 803                 kp->page = page;
 804         else
 805                 free_page((unsigned long)page);
 806         ret = 0;
 807 unlock_out:
 808         spin_unlock(&zcache_direct_reclaim_lock);
 809 out:
 810         return ret;
 811 }
 812
 813 static void *zcache_get_free_page(void)
 814 {
 815         struct zcache_preload *kp;
 816         void *page;
 817
 818         kp = &__get_cpu_var(zcache_preloads);
 819         page = kp->page;
 820         BUG_ON(page == NULL);
 821         kp->page = NULL;
 822         return page;
 823 }
 824
 825 static void zcache_free_page(void *p)
 826 {
 827         free_page((unsigned long)p);
 828 }
 829
 830 /*
 831  * zcache implementation for tmem host ops
 832  */
 833
 834 static struct tmem_objnode *zcache_objnode_alloc(struct tmem_pool *pool)
 835 {
 836         struct tmem_objnode *objnode = NULL;
 837         unsigned long count;
 838         struct zcache_preload *kp;
 839
 840         kp = &__get_cpu_var(zcache_preloads);
 841         if (kp->nr <= 0)
 842                 goto out;
 843         objnode = kp->objnodes[kp->nr - 1];
 844         BUG_ON(objnode == NULL);
 845         kp->objnodes[kp->nr - 1] = NULL;
 846         kp->nr--;
 847         count = atomic_inc_return(&zcache_curr_objnode_count);
 848         if (count > zcache_curr_objnode_count_max)
 849                 zcache_curr_objnode_count_max = count;
 850 out:
 851         return objnode;
 852 }
 853
 854 static void zcache_objnode_free(struct tmem_objnode *objnode,
 855                                         struct tmem_pool *pool)
 856 {
 857         atomic_dec(&zcache_curr_objnode_count);
 858         BUG_ON(atomic_read(&zcache_curr_objnode_count) < 0);
 859         kmem_cache_free(zcache_objnode_cache, objnode);
 860 }
 861
 862 static struct tmem_obj *zcache_obj_alloc(struct tmem_pool *pool)
 863 {
 864         struct tmem_obj *obj = NULL;
 865         unsigned long count;
 866         struct zcache_preload *kp;
 867
 868         kp = &__get_cpu_var(zcache_preloads);
 869         obj = kp->obj;
 870         BUG_ON(obj == NULL);
 871         kp->obj = NULL;
 872         count = atomic_inc_return(&zcache_curr_obj_count);
 873         if (count > zcache_curr_obj_count_max)
 874                 zcache_curr_obj_count_max = count;
 875         return obj;
 876 }
 877
 878 static void zcache_obj_free(struct tmem_obj *obj, struct tmem_pool *pool)
 879 {
 880         atomic_dec(&zcache_curr_obj_count);
 881         BUG_ON(atomic_read(&zcache_curr_obj_count) < 0);
 882         kmem_cache_free(zcache_obj_cache, obj);
 883 }
 884
 885 static struct tmem_hostops zcache_hostops = {
 886         .obj_alloc = zcache_obj_alloc,
 887         .obj_free = zcache_obj_free,
 888         .objnode_alloc = zcache_objnode_alloc,
 889         .objnode_free = zcache_objnode_free,
 890 };
 891
 892 /*
 893  * zcache implementations for PAM page descriptor ops
 894  */
 895
 896 static atomic_t zcache_curr_eph_pampd_count = ATOMIC_INIT(0);
 897 static unsigned long zcache_curr_eph_pampd_count_max;
 898 static atomic_t zcache_curr_pers_pampd_count = ATOMIC_INIT(0);
 899 static unsigned long zcache_curr_pers_pampd_count_max;
 900
 901 /* forward reference */
 902 static int zcache_compress(struct page *from, void **out_va, size_t *out_len);
 903
 904 static void *zcache_pampd_create(struct tmem_pool *pool, struct tmem_oid *oid,
 905                                  uint32_t index, struct page *page)
 906 {
 907         void *pampd = NULL, *cdata;
 908         size_t clen;
 909         int ret;
 910         bool ephemeral = is_ephemeral(pool);
 911         unsigned long count;
 912
 913         if (ephemeral) {
 914                 ret = zcache_compress(page, &cdata, &clen);
 915                 if (ret == 0)
 916
 917                         goto out;
 918                 if (clen == 0 || clen > zbud_max_buddy_size()) {
 919                         zcache_compress_poor++;
 920                         goto out;
 921                 }
 922                 pampd = (void *)zbud_create(pool->pool_id, oid, index,
 923                                                 page, cdata, clen);
 924                 if (pampd != NULL) {
 925                         count = atomic_inc_return(&zcache_curr_eph_pampd_count);
 926                         if (count > zcache_curr_eph_pampd_count_max)
 927                                 zcache_curr_eph_pampd_count_max = count;
 928                 }
 929         } else {
 930                 /*
 931                  * FIXME: This is all the "policy" there is for now.
 932                  * 3/4 totpages should allow ~37% of RAM to be filled with
 933                  * compressed frontswap pages
 934                  */
 935                 if (atomic_read(&zcache_curr_pers_pampd_count) >
 936                                                         3 * totalram_pages / 4)
 937                         goto out;
 938                 ret = zcache_compress(page, &cdata, &clen);
 939                 if (ret == 0)
 940                         goto out;
 941                 if (clen > zv_max_page_size) {
 942                         zcache_compress_poor++;
 943                         goto out;
 944                 }
 945                 pampd = (void *)zv_create(zcache_client.xvpool, pool->pool_id,
 946                                                 oid, index, cdata, clen);
 947                 if (pampd == NULL)
 948                         goto out;
 949                 count = atomic_inc_return(&zcache_curr_pers_pampd_count);
 950                 if (count > zcache_curr_pers_pampd_count_max)
 951                         zcache_curr_pers_pampd_count_max = count;
 952         }
 953 out:
 954         return pampd;
 955 }
 956
 957 /*
 958  * fill the pageframe corresponding to the struct page with the data
 959  * from the passed pampd
 960  */
 961 static int zcache_pampd_get_data(struct page *page, void *pampd,
 962                                                 struct tmem_pool *pool)
 963 {
 964         int ret = 0;
 965
 966         if (is_ephemeral(pool))
 967                 ret = zbud_decompress(page, pampd);
 968         else
 969                 zv_decompress(page, pampd);
 970         return ret;
 971 }
 972
 973 /*
 974  * free the pampd and remove it from any zcache lists
 975  * pampd must no longer be pointed to from any tmem data structures!
 976  */
 977 static void zcache_pampd_free(void *pampd, struct tmem_pool *pool)
 978 {
 979         if (is_ephemeral(pool)) {
 980                 zbud_free_and_delist((struct zbud_hdr *)pampd);
 981                 atomic_dec(&zcache_curr_eph_pampd_count);
 982                 BUG_ON(atomic_read(&zcache_curr_eph_pampd_count) < 0);
 983         } else {
 984                 zv_free(zcache_client.xvpool, (struct zv_hdr *)pampd);
 985                 atomic_dec(&zcache_curr_pers_pampd_count);
 986                 BUG_ON(atomic_read(&zcache_curr_pers_pampd_count) < 0);
 987         }
 988 }
 989
 990 static struct tmem_pamops zcache_pamops = {
 991         .create = zcache_pampd_create,
 992         .get_data = zcache_pampd_get_data,
 993         .free = zcache_pampd_free,
 994 };
 995
 996 /*
 997  * zcache compression/decompression and related per-cpu stuff
 998  */
 999
1000 #define LZO_WORKMEM_BYTES LZO1X_1_MEM_COMPRESS
1001 #define LZO_DSTMEM_PAGE_ORDER 1
1002 static DEFINE_PER_CPU(unsigned char *, zcache_workmem);
1003 static DEFINE_PER_CPU(unsigned char *, zcache_dstmem);
1004
1005 static int zcache_compress(struct page *from, void **out_va, size_t *out_len)
1006 {
1007         int ret = 0;
1008         unsigned char *dmem = __get_cpu_var(zcache_dstmem);
1009         unsigned char *wmem = __get_cpu_var(zcache_workmem);
1010         char *from_va;
1011
1012         BUG_ON(!irqs_disabled());
1013         if (unlikely(dmem == NULL || wmem == NULL))
1014                 goto out;  /* no buffer, so can't compress */
1015         from_va = kmap_atomic(from, KM_USER0);
1016         mb();
1017         ret = lzo1x_1_compress(from_va, PAGE_SIZE, dmem, out_len, wmem);
1018         BUG_ON(ret != LZO_E_OK);
1019         *out_va = dmem;
1020         kunmap_atomic(from_va, KM_USER0);
1021         ret = 1;
1022 out:
1023         return ret;
1024 }
1025
1026
1027 static int zcache_cpu_notifier(struct notifier_block *nb,
1028                                 unsigned long action, void *pcpu)
1029 {
1030         int cpu = (long)pcpu;
1031         struct zcache_preload *kp;
1032
1033         switch (action) {
1034         case CPU_UP_PREPARE:
1035                 per_cpu(zcache_dstmem, cpu) = (void *)__get_free_pages(
1036                         GFP_KERNEL | __GFP_REPEAT,
1037                         LZO_DSTMEM_PAGE_ORDER),
1038                 per_cpu(zcache_workmem, cpu) =
1039                         kzalloc(LZO1X_MEM_COMPRESS,
1040                                 GFP_KERNEL | __GFP_REPEAT);
1041                 break;
1042         case CPU_DEAD:
1043         case CPU_UP_CANCELED:
1044                 free_pages((unsigned long)per_cpu(zcache_dstmem, cpu),
1045                                 LZO_DSTMEM_PAGE_ORDER);
1046                 per_cpu(zcache_dstmem, cpu) = NULL;
1047                 kfree(per_cpu(zcache_workmem, cpu));
1048                 per_cpu(zcache_workmem, cpu) = NULL;
1049                 kp = &per_cpu(zcache_preloads, cpu);
1050                 while (kp->nr) {
1051                         kmem_cache_free(zcache_objnode_cache,
1052                                         kp->objnodes[kp->nr - 1]);
1053                         kp->objnodes[kp->nr - 1] = NULL;
1054                         kp->nr--;
1055                 }
1056                 kmem_cache_free(zcache_obj_cache, kp->obj);
1057                 free_page((unsigned long)kp->page);
1058                 break;
1059         default:
1060                 break;
1061         }
1062         return NOTIFY_OK;
1063 }
1064
1065 static struct notifier_block zcache_cpu_notifier_block = {
1066         .notifier_call = zcache_cpu_notifier
1067 };
1068
1069 #ifdef CONFIG_SYSFS
1070 #define ZCACHE_SYSFS_RO(_name) \
1071         static ssize_t zcache_##_name##_show(struct kobject *kobj, \
1072                                 struct kobj_attribute *attr, char *buf) \
1073         { \
1074                 return sprintf(buf, "%lu\n", zcache_##_name); \
1075         } \
1076         static struct kobj_attribute zcache_##_name##_attr = { \
1077                 .attr = { .name = __stringify(_name), .mode = 0444 }, \
1078                 .show = zcache_##_name##_show, \
1079         }
1080
1081 #define ZCACHE_SYSFS_RO_ATOMIC(_name) \
1082         static ssize_t zcache_##_name##_show(struct kobject *kobj, \
1083                                 struct kobj_attribute *attr, char *buf) \
1084         { \
1085             return sprintf(buf, "%d\n", atomic_read(&zcache_##_name)); \
1086         } \
1087         static struct kobj_attribute zcache_##_name##_attr = { \
1088                 .attr = { .name = __stringify(_name), .mode = 0444 }, \
1089                 .show = zcache_##_name##_show, \
1090         }
1091
1092 #define ZCACHE_SYSFS_RO_CUSTOM(_name, _func) \
1093         static ssize_t zcache_##_name##_show(struct kobject *kobj, \
1094                                 struct kobj_attribute *attr, char *buf) \
1095         { \
1096             return _func(buf); \
1097         } \
1098         static struct kobj_attribute zcache_##_name##_attr = { \
1099                 .attr = { .name = __stringify(_name), .mode = 0444 }, \
1100                 .show = zcache_##_name##_show, \
1101         }
1102
1103 ZCACHE_SYSFS_RO(curr_obj_count_max);
1104 ZCACHE_SYSFS_RO(curr_objnode_count_max);
1105 ZCACHE_SYSFS_RO(flush_total);
1106 ZCACHE_SYSFS_RO(flush_found);
1107 ZCACHE_SYSFS_RO(flobj_total);
1108 ZCACHE_SYSFS_RO(flobj_found);
1109 ZCACHE_SYSFS_RO(failed_eph_puts);
1110 ZCACHE_SYSFS_RO(failed_pers_puts);
1111 ZCACHE_SYSFS_RO(zbud_curr_zbytes);
1112 ZCACHE_SYSFS_RO(zbud_cumul_zpages);
1113 ZCACHE_SYSFS_RO(zbud_cumul_zbytes);
1114 ZCACHE_SYSFS_RO(zbud_buddied_count);
1115 ZCACHE_SYSFS_RO(zbpg_unused_list_count);
1116 ZCACHE_SYSFS_RO(evicted_raw_pages);
1117 ZCACHE_SYSFS_RO(evicted_unbuddied_pages);
1118 ZCACHE_SYSFS_RO(evicted_buddied_pages);
1119 ZCACHE_SYSFS_RO(failed_get_free_pages);
1120 ZCACHE_SYSFS_RO(failed_alloc);
1121 ZCACHE_SYSFS_RO(put_to_flush);
1122 ZCACHE_SYSFS_RO(aborted_preload);
1123 ZCACHE_SYSFS_RO(aborted_shrink);
1124 ZCACHE_SYSFS_RO(compress_poor);
1125 ZCACHE_SYSFS_RO_ATOMIC(zbud_curr_raw_pages);
1126 ZCACHE_SYSFS_RO_ATOMIC(zbud_curr_zpages);
1127 ZCACHE_SYSFS_RO_ATOMIC(curr_obj_count);
1128 ZCACHE_SYSFS_RO_ATOMIC(curr_objnode_count);
1129 ZCACHE_SYSFS_RO_CUSTOM(zbud_unbuddied_list_counts,
1130                         zbud_show_unbuddied_list_counts);
1131 ZCACHE_SYSFS_RO_CUSTOM(zbud_cumul_chunk_counts,
1132                         zbud_show_cumul_chunk_counts);
1133
1134 static struct attribute *zcache_attrs[] = {
1135         &zcache_curr_obj_count_attr.attr,
1136         &zcache_curr_obj_count_max_attr.attr,
1137         &zcache_curr_objnode_count_attr.attr,
1138         &zcache_curr_objnode_count_max_attr.attr,
1139         &zcache_flush_total_attr.attr,
1140         &zcache_flobj_total_attr.attr,
1141         &zcache_flush_found_attr.attr,
1142         &zcache_flobj_found_attr.attr,
1143         &zcache_failed_eph_puts_attr.attr,
1144         &zcache_failed_pers_puts_attr.attr,
1145         &zcache_compress_poor_attr.attr,
1146         &zcache_zbud_curr_raw_pages_attr.attr,
1147         &zcache_zbud_curr_zpages_attr.attr,
1148         &zcache_zbud_curr_zbytes_attr.attr,
1149         &zcache_zbud_cumul_zpages_attr.attr,
1150         &zcache_zbud_cumul_zbytes_attr.attr,
1151         &zcache_zbud_buddied_count_attr.attr,
1152         &zcache_zbpg_unused_list_count_attr.attr,
1153         &zcache_evicted_raw_pages_attr.attr,
1154         &zcache_evicted_unbuddied_pages_attr.attr,
1155         &zcache_evicted_buddied_pages_attr.attr,
1156         &zcache_failed_get_free_pages_attr.attr,
1157         &zcache_failed_alloc_attr.attr,
1158         &zcache_put_to_flush_attr.attr,
1159         &zcache_aborted_preload_attr.attr,
1160         &zcache_aborted_shrink_attr.attr,
1161         &zcache_zbud_unbuddied_list_counts_attr.attr,
1162         &zcache_zbud_cumul_chunk_counts_attr.attr,
1163         NULL,
1164 };
1165
1166 static struct attribute_group zcache_attr_group = {
1167         .attrs = zcache_attrs,
1168         .name = "zcache",
1169 };
1170
1171 #endif /* CONFIG_SYSFS */
1172 /*
1173  * When zcache is disabled ("frozen"), pools can be created and destroyed,
1174  * but all puts (and thus all other operations that require memory allocation)
1175  * must fail.  If zcache is unfrozen, accepts puts, then frozen again,
1176  * data consistency requires all puts while frozen to be converted into
1177  * flushes.
1178  */
1179 static bool zcache_freeze;
1180
1181 /*
1182  * zcache shrinker interface (only useful for ephemeral pages, so zbud only)
1183  */
1184 static int shrink_zcache_memory(struct shrinker *shrink, int nr, gfp_t gfp_mask)
1185 {
1186         int ret = -1;
1187
1188         if (nr >= 0) {
1189                 if (!(gfp_mask & __GFP_FS))
1190                         /* does this case really need to be skipped? */
1191                         goto out;
1192                 if (spin_trylock(&zcache_direct_reclaim_lock)) {
1193                         zbud_evict_pages(nr);
1194                         spin_unlock(&zcache_direct_reclaim_lock);
1195                 } else
1196                         zcache_aborted_shrink++;
1197         }
1198         ret = (int)atomic_read(&zcache_zbud_curr_raw_pages);
1199 out:
1200         return ret;
1201 }
1202
1203 static struct shrinker zcache_shrinker = {
1204         .shrink = shrink_zcache_memory,
1205         .seeks = DEFAULT_SEEKS,
1206 };
1207
1208 /*
1209  * zcache shims between cleancache/frontswap ops and tmem
1210  */
1211
1212 static int zcache_put_page(int pool_id, struct tmem_oid *oidp,
1213                                 uint32_t index, struct page *page)
1214 {
1215         struct tmem_pool *pool;
1216         int ret = -1;
1217
1218         BUG_ON(!irqs_disabled());
1219         pool = zcache_get_pool_by_id(pool_id);
1220         if (unlikely(pool == NULL))
1221                 goto out;
1222         if (!zcache_freeze && zcache_do_preload(pool) == 0) {
1223                 /* preload does preempt_disable on success */
1224                 ret = tmem_put(pool, oidp, index, page);
1225                 if (ret < 0) {
1226                         if (is_ephemeral(pool))
1227                                 zcache_failed_eph_puts++;
1228                         else
1229                                 zcache_failed_pers_puts++;
1230                 }
1231                 zcache_put_pool(pool);
1232                 preempt_enable_no_resched();
1233         } else {
1234                 zcache_put_to_flush++;
1235                 if (atomic_read(&pool->obj_count) > 0)
1236                         /* the put fails whether the flush succeeds or not */
1237                         (void)tmem_flush_page(pool, oidp, index);
1238                 zcache_put_pool(pool);
1239         }
1240 out:
1241         return ret;
1242 }
1243
1244 static int zcache_get_page(int pool_id, struct tmem_oid *oidp,
1245                                 uint32_t index, struct page *page)
1246 {
1247         struct tmem_pool *pool;
1248         int ret = -1;
1249         unsigned long flags;
1250
1251         local_irq_save(flags);
1252         pool = zcache_get_pool_by_id(pool_id);
1253         if (likely(pool != NULL)) {
1254                 if (atomic_read(&pool->obj_count) > 0)
1255                         ret = tmem_get(pool, oidp, index, page);
1256                 zcache_put_pool(pool);
1257         }
1258         local_irq_restore(flags);
1259         return ret;
1260 }
1261
1262 static int zcache_flush_page(int pool_id, struct tmem_oid *oidp, uint32_t index)
1263 {
1264         struct tmem_pool *pool;
1265         int ret = -1;
1266         unsigned long flags;
1267
1268         local_irq_save(flags);
1269         zcache_flush_total++;
1270         pool = zcache_get_pool_by_id(pool_id);
1271         if (likely(pool != NULL)) {
1272                 if (atomic_read(&pool->obj_count) > 0)
1273                         ret = tmem_flush_page(pool, oidp, index);
1274                 zcache_put_pool(pool);
1275         }
1276         if (ret >= 0)
1277                 zcache_flush_found++;
1278         local_irq_restore(flags);
1279         return ret;
1280 }
1281
1282 static int zcache_flush_object(int pool_id, struct tmem_oid *oidp)
1283 {
1284         struct tmem_pool *pool;
1285         int ret = -1;
1286         unsigned long flags;
1287
1288         local_irq_save(flags);
1289         zcache_flobj_total++;
1290         pool = zcache_get_pool_by_id(pool_id);
1291         if (likely(pool != NULL)) {
1292                 if (atomic_read(&pool->obj_count) > 0)
1293                         ret = tmem_flush_object(pool, oidp);
1294                 zcache_put_pool(pool);
1295         }
1296         if (ret >= 0)
1297                 zcache_flobj_found++;
1298         local_irq_restore(flags);
1299         return ret;
1300 }
1301
1302 static int zcache_destroy_pool(int pool_id)
1303 {
1304         struct tmem_pool *pool = NULL;
1305         int ret = -1;
1306
1307         if (pool_id < 0)
1308                 goto out;
1309         pool = zcache_client.tmem_pools[pool_id];
1310         if (pool == NULL)
1311                 goto out;
1312         zcache_client.tmem_pools[pool_id] = NULL;
1313         /* wait for pool activity on other cpus to quiesce */
1314         while (atomic_read(&pool->refcount) != 0)
1315                 ;
1316         local_bh_disable();
1317         ret = tmem_destroy_pool(pool);
1318         local_bh_enable();
1319         kfree(pool);
1320         pr_info("zcache: destroyed pool id=%d\n", pool_id);
1321 out:
1322         return ret;
1323 }
1324
1325 static int zcache_new_pool(uint32_t flags)
1326 {
1327         int poolid = -1;
1328         struct tmem_pool *pool;
1329
1330         pool = kmalloc(sizeof(struct tmem_pool), GFP_KERNEL);
1331         if (pool == NULL) {
1332                 pr_info("zcache: pool creation failed: out of memory\n");
1333                 goto out;
1334         }
1335
1336         for (poolid = 0; poolid < MAX_POOLS_PER_CLIENT; poolid++)
1337                 if (zcache_client.tmem_pools[poolid] == NULL)
1338                         break;
1339         if (poolid >= MAX_POOLS_PER_CLIENT) {
1340                 pr_info("zcache: pool creation failed: max exceeded\n");
1341                 kfree(pool);
1342                 poolid = -1;
1343                 goto out;
1344         }
1345         atomic_set(&pool->refcount, 0);
1346         pool->client = &zcache_client;
1347         pool->pool_id = poolid;
1348         tmem_new_pool(pool, flags);
1349         zcache_client.tmem_pools[poolid] = pool;
1350         pr_info("zcache: created %s tmem pool, id=%d\n",
1351                 flags & TMEM_POOL_PERSIST ? "persistent" : "ephemeral",
1352                 poolid);
1353 out:
1354         return poolid;
1355 }
1356
1357 /**********
1358  * Two kernel functionalities currently can be layered on top of tmem.
1359  * These are "cleancache" which is used as a second-chance cache for clean
1360  * page cache pages; and "frontswap" which is used for swap pages
1361  * to avoid writes to disk.  A generic "shim" is provided here for each
1362  * to translate in-kernel semantics to zcache semantics.
1363  */
1364
1365 #ifdef CONFIG_CLEANCACHE
1366 static void zcache_cleancache_put_page(int pool_id,
1367                                         struct cleancache_filekey key,
1368                                         pgoff_t index, struct page *page)
1369 {
1370         u32 ind = (u32) index;
1371         struct tmem_oid oid = *(struct tmem_oid *)&key;
1372
1373         if (likely(ind == index))
1374                 (void)zcache_put_page(pool_id, &oid, index, page);
1375 }
1376
1377 static int zcache_cleancache_get_page(int pool_id,
1378                                         struct cleancache_filekey key,
1379                                         pgoff_t index, struct page *page)
1380 {
1381         u32 ind = (u32) index;
1382         struct tmem_oid oid = *(struct tmem_oid *)&key;
1383         int ret = -1;
1384
1385         if (likely(ind == index))
1386                 ret = zcache_get_page(pool_id, &oid, index, page);
1387         return ret;
1388 }
1389
1390 static void zcache_cleancache_flush_page(int pool_id,
1391                                         struct cleancache_filekey key,
1392                                         pgoff_t index)
1393 {
1394         u32 ind = (u32) index;
1395         struct tmem_oid oid = *(struct tmem_oid *)&key;
1396
1397         if (likely(ind == index))
1398                 (void)zcache_flush_page(pool_id, &oid, ind);
1399 }
1400
1401 static void zcache_cleancache_flush_inode(int pool_id,
1402                                         struct cleancache_filekey key)
1403 {
1404         struct tmem_oid oid = *(struct tmem_oid *)&key;
1405
1406         (void)zcache_flush_object(pool_id, &oid);
1407 }
1408
1409 static void zcache_cleancache_flush_fs(int pool_id)
1410 {
1411         if (pool_id >= 0)
1412                 (void)zcache_destroy_pool(pool_id);
1413 }
1414
1415 static int zcache_cleancache_init_fs(size_t pagesize)
1416 {
1417         BUG_ON(sizeof(struct cleancache_filekey) !=
1418                                 sizeof(struct tmem_oid));
1419         BUG_ON(pagesize != PAGE_SIZE);
1420         return zcache_new_pool(0);
1421 }
1422
1423 static int zcache_cleancache_init_shared_fs(char *uuid, size_t pagesize)
1424 {
1425         /* shared pools are unsupported and map to private */
1426         BUG_ON(sizeof(struct cleancache_filekey) !=
1427                                 sizeof(struct tmem_oid));
1428         BUG_ON(pagesize != PAGE_SIZE);
1429         return zcache_new_pool(0);
1430 }
1431
1432 static struct cleancache_ops zcache_cleancache_ops = {
1433         .put_page = zcache_cleancache_put_page,
1434         .get_page = zcache_cleancache_get_page,
1435         .flush_page = zcache_cleancache_flush_page,
1436         .flush_inode = zcache_cleancache_flush_inode,
1437         .flush_fs = zcache_cleancache_flush_fs,
1438         .init_shared_fs = zcache_cleancache_init_shared_fs,
1439         .init_fs = zcache_cleancache_init_fs
1440 };
1441
1442 struct cleancache_ops zcache_cleancache_register_ops(void)
1443 {
1444         struct cleancache_ops old_ops =
1445                 cleancache_register_ops(&zcache_cleancache_ops);
1446
1447         return old_ops;
1448 }
1449 #endif
1450
1451 #ifdef CONFIG_FRONTSWAP
1452 /* a single tmem poolid is used for all frontswap "types" (swapfiles) */
1453 static int zcache_frontswap_poolid = -1;
1454
1455 /*
1456  * Swizzling increases objects per swaptype, increasing tmem concurrency
1457  * for heavy swaploads.  Later, larger nr_cpus -> larger SWIZ_BITS
1458  */
1459 #define SWIZ_BITS               4
1460 #define SWIZ_MASK               ((1 << SWIZ_BITS) - 1)
1461 #define _oswiz(_type, _ind)     ((_type << SWIZ_BITS) | (_ind & SWIZ_MASK))
1462 #define iswiz(_ind)             (_ind >> SWIZ_BITS)
1463
1464 static inline struct tmem_oid oswiz(unsigned type, u32 ind)
1465 {
1466         struct tmem_oid oid = { .oid = { 0 } };
1467         oid.oid[0] = _oswiz(type, ind);
1468         return oid;
1469 }
1470
1471 static int zcache_frontswap_put_page(unsigned type, pgoff_t offset,
1472                                    struct page *page)
1473 {
1474         u64 ind64 = (u64)offset;
1475         u32 ind = (u32)offset;
1476         struct tmem_oid oid = oswiz(type, ind);
1477         int ret = -1;
1478         unsigned long flags;
1479
1480         BUG_ON(!PageLocked(page));
1481         if (likely(ind64 == ind)) {
1482                 local_irq_save(flags);
1483                 ret = zcache_put_page(zcache_frontswap_poolid, &oid,
1484                                         iswiz(ind), page);
1485                 local_irq_restore(flags);
1486         }
1487         return ret;
1488 }
1489
1490 /* returns 0 if the page was successfully gotten from frontswap, -1 if
1491  * was not present (should never happen!) */
1492 static int zcache_frontswap_get_page(unsigned type, pgoff_t offset,
1493                                    struct page *page)
1494 {
1495         u64 ind64 = (u64)offset;
1496         u32 ind = (u32)offset;
1497         struct tmem_oid oid = oswiz(type, ind);
1498         int ret = -1;
1499
1500         BUG_ON(!PageLocked(page));
1501         if (likely(ind64 == ind))
1502                 ret = zcache_get_page(zcache_frontswap_poolid, &oid,
1503                                         iswiz(ind), page);
1504         return ret;
1505 }
1506
1507 /* flush a single page from frontswap */
1508 static void zcache_frontswap_flush_page(unsigned type, pgoff_t offset)
1509 {
1510         u64 ind64 = (u64)offset;
1511         u32 ind = (u32)offset;
1512         struct tmem_oid oid = oswiz(type, ind);
1513
1514         if (likely(ind64 == ind))
1515                 (void)zcache_flush_page(zcache_frontswap_poolid, &oid,
1516                                         iswiz(ind));
1517 }
1518
1519 /* flush all pages from the passed swaptype */
1520 static void zcache_frontswap_flush_area(unsigned type)
1521 {
1522         struct tmem_oid oid;
1523         int ind;
1524
1525         for (ind = SWIZ_MASK; ind >= 0; ind--) {
1526                 oid = oswiz(type, ind);
1527                 (void)zcache_flush_object(zcache_frontswap_poolid, &oid);
1528         }
1529 }
1530
1531 static void zcache_frontswap_init(unsigned ignored)
1532 {
1533         /* a single tmem poolid is used for all frontswap "types" (swapfiles) */
1534         if (zcache_frontswap_poolid < 0)
1535                 zcache_frontswap_poolid = zcache_new_pool(TMEM_POOL_PERSIST);
1536 }
1537
1538 static struct frontswap_ops zcache_frontswap_ops = {
1539         .put_page = zcache_frontswap_put_page,
1540         .get_page = zcache_frontswap_get_page,
1541         .flush_page = zcache_frontswap_flush_page,
1542         .flush_area = zcache_frontswap_flush_area,
1543         .init = zcache_frontswap_init
1544 };
1545
1546 struct frontswap_ops zcache_frontswap_register_ops(void)
1547 {
1548         struct frontswap_ops old_ops =
1549                 frontswap_register_ops(&zcache_frontswap_ops);
1550
1551         return old_ops;
1552 }
1553 #endif
1554
1555 /*
1556  * zcache initialization
1557  * NOTE FOR NOW zcache MUST BE PROVIDED AS A KERNEL BOOT PARAMETER OR
1558  * NOTHING HAPPENS!
1559  */
1560
1561 static int zcache_enabled;
1562
1563 static int __init enable_zcache(char *s)
1564 {
1565         zcache_enabled = 1;
1566         return 1;
1567 }
1568 __setup("zcache", enable_zcache);
1569
1570 /* allow independent dynamic disabling of cleancache and frontswap */
1571
1572 static int use_cleancache = 1;
1573
1574 static int __init no_cleancache(char *s)
1575 {
1576         use_cleancache = 0;
1577         return 1;
1578 }
1579
1580 __setup("nocleancache", no_cleancache);
1581
1582 static int use_frontswap = 1;
1583
1584 static int __init no_frontswap(char *s)
1585 {
1586         use_frontswap = 0;
1587         return 1;
1588 }
1589
1590 __setup("nofrontswap", no_frontswap);
1591
1592 static int __init zcache_init(void)
1593 {
1594 #ifdef CONFIG_SYSFS
1595         int ret = 0;
1596
1597         ret = sysfs_create_group(mm_kobj, &zcache_attr_group);
1598         if (ret) {
1599                 pr_err("zcache: can't create sysfs\n");
1600                 goto out;
1601         }
1602 #endif /* CONFIG_SYSFS */
1603 #if defined(CONFIG_CLEANCACHE) || defined(CONFIG_FRONTSWAP)
1604         if (zcache_enabled) {
1605                 unsigned int cpu;
1606
1607                 tmem_register_hostops(&zcache_hostops);
1608                 tmem_register_pamops(&zcache_pamops);
1609                 ret = register_cpu_notifier(&zcache_cpu_notifier_block);
1610                 if (ret) {
1611                         pr_err("zcache: can't register cpu notifier\n");
1612                         goto out;
1613                 }
1614                 for_each_online_cpu(cpu) {
1615                         void *pcpu = (void *)(long)cpu;
1616                         zcache_cpu_notifier(&zcache_cpu_notifier_block,
1617                                 CPU_UP_PREPARE, pcpu);
1618                 }
1619         }
1620         zcache_objnode_cache = kmem_cache_create("zcache_objnode",
1621                                 sizeof(struct tmem_objnode), 0, 0, NULL);
1622         zcache_obj_cache = kmem_cache_create("zcache_obj",
1623                                 sizeof(struct tmem_obj), 0, 0, NULL);
1624 #endif
1625 #ifdef CONFIG_CLEANCACHE
1626         if (zcache_enabled && use_cleancache) {
1627                 struct cleancache_ops old_ops;
1628
1629                 zbud_init();
1630                 register_shrinker(&zcache_shrinker);
1631                 old_ops = zcache_cleancache_register_ops();
1632                 pr_info("zcache: cleancache enabled using kernel "
1633                         "transcendent memory and compression buddies\n");
1634                 if (old_ops.init_fs != NULL)
1635                         pr_warning("zcache: cleancache_ops overridden");
1636         }
1637 #endif
1638 #ifdef CONFIG_FRONTSWAP
1639         if (zcache_enabled && use_frontswap) {
1640                 struct frontswap_ops old_ops;
1641
1642                 zcache_client.xvpool = xv_create_pool();
1643                 if (zcache_client.xvpool == NULL) {
1644                         pr_err("zcache: can't create xvpool\n");
1645                         goto out;
1646                 }
1647                 old_ops = zcache_frontswap_register_ops();
1648                 pr_info("zcache: frontswap enabled using kernel "
1649                         "transcendent memory and xvmalloc\n");
1650                 if (old_ops.init != NULL)
1651                         pr_warning("ktmem: frontswap_ops overridden");
1652         }
1653 #endif
1654 out:
1655         return ret;
1656 }
1657
1658 module_init(zcache_init)