drivers/staging/zcache/zcache-main.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 the crypto compression
  10  * API:
  11  * 1) "compression buddies" ("zbud") is used for ephemeral pages
  12  * 2) zsmalloc is used for persistent pages.
  13  * Xvmalloc (based on the TLSF allocator) has very low fragmentation
  14  * so maximizes space efficiency, while zbud allows pairs (and potentially,
  15  * in the future, more than a pair of) compressed pages to be closely linked
  16  * so that reclaiming can be done via the kernel's physical-page-oriented
  17  * "shrinker" interface.
  18  *
  19  * [1] For a definition of page-accessible memory (aka PAM), see:
  20  *   http://marc.info/?l=linux-mm&m=127811271605009
  21  */
  22
  23 #include <linux/module.h>
  24 #include <linux/cpu.h>
  25 #include <linux/highmem.h>
  26 #include <linux/list.h>
  27 #include <linux/slab.h>
  28 #include <linux/spinlock.h>
  29 #include <linux/types.h>
  30 #include <linux/atomic.h>
  31 #include <linux/math64.h>
  32 #include <linux/crypto.h>
  33 #include <linux/string.h>
  34 #include <linux/idr.h>
  35 #include "tmem.h"
  36
  37 #include "../zsmalloc/zsmalloc.h"
  38
  39 #ifdef CONFIG_CLEANCACHE
  40 #include <linux/cleancache.h>
  41 #endif
  42 #ifdef CONFIG_FRONTSWAP
  43 #include <linux/frontswap.h>
  44 #endif
  45
  46 #if 0
  47 /* this is more aggressive but may cause other problems? */
  48 #define ZCACHE_GFP_MASK (GFP_ATOMIC | __GFP_NORETRY | __GFP_NOWARN)
  49 #else
  50 #define ZCACHE_GFP_MASK \
  51         (__GFP_FS | __GFP_NORETRY | __GFP_NOWARN | __GFP_NOMEMALLOC)
  52 #endif
  53
  54 #define MAX_CLIENTS 16
  55 #define LOCAL_CLIENT ((uint16_t)-1)
  56
  57 MODULE_LICENSE("GPL");
  58
  59 struct zcache_client {
  60         struct idr tmem_pools;
  61         struct zs_pool *zspool;
  62         bool allocated;
  63         atomic_t refcount;
  64 };
  65
  66 static struct zcache_client zcache_host;
  67 static struct zcache_client zcache_clients[MAX_CLIENTS];
  68
  69 static inline uint16_t get_client_id_from_client(struct zcache_client *cli)
  70 {
  71         BUG_ON(cli == NULL);
  72         if (cli == &zcache_host)
  73                 return LOCAL_CLIENT;
  74         return cli - &zcache_clients[0];
  75 }
  76
  77 static inline bool is_local_client(struct zcache_client *cli)
  78 {
  79         return cli == &zcache_host;
  80 }
  81
  82 /* crypto API for zcache  */
  83 #define ZCACHE_COMP_NAME_SZ CRYPTO_MAX_ALG_NAME
  84 static char zcache_comp_name[ZCACHE_COMP_NAME_SZ];
  85 static struct crypto_comp * __percpu *zcache_comp_pcpu_tfms;
  86
  87 enum comp_op {
  88         ZCACHE_COMPOP_COMPRESS,
  89         ZCACHE_COMPOP_DECOMPRESS
  90 };
  91
  92 static inline int zcache_comp_op(enum comp_op op,
  93                                 const u8 *src, unsigned int slen,
  94                                 u8 *dst, unsigned int *dlen)
  95 {
  96         struct crypto_comp *tfm;
  97         int ret;
  98
  99         BUG_ON(!zcache_comp_pcpu_tfms);
 100         tfm = *per_cpu_ptr(zcache_comp_pcpu_tfms, get_cpu());
 101         BUG_ON(!tfm);
 102         switch (op) {
 103         case ZCACHE_COMPOP_COMPRESS:
 104                 ret = crypto_comp_compress(tfm, src, slen, dst, dlen);
 105                 break;
 106         case ZCACHE_COMPOP_DECOMPRESS:
 107                 ret = crypto_comp_decompress(tfm, src, slen, dst, dlen);
 108                 break;
 109         default:
 110                 ret = -EINVAL;
 111         }
 112         put_cpu();
 113         return ret;
 114 }
 115
 116 /**********
 117  * Compression buddies ("zbud") provides for packing two (or, possibly
 118  * in the future, more) compressed ephemeral pages into a single "raw"
 119  * (physical) page and tracking them with data structures so that
 120  * the raw pages can be easily reclaimed.
 121  *
 122  * A zbud page ("zbpg") is an aligned page containing a list_head,
 123  * a lock, and two "zbud headers".  The remainder of the physical
 124  * page is divided up into aligned 64-byte "chunks" which contain
 125  * the compressed data for zero, one, or two zbuds.  Each zbpg
 126  * resides on: (1) an "unused list" if it has no zbuds; (2) a
 127  * "buddied" list if it is fully populated  with two zbuds; or
 128  * (3) one of PAGE_SIZE/64 "unbuddied" lists indexed by how many chunks
 129  * the one unbuddied zbud uses.  The data inside a zbpg cannot be
 130  * read or written unless the zbpg's lock is held.
 131  */
 132
 133 #define ZBH_SENTINEL  0x43214321
 134 #define ZBPG_SENTINEL  0xdeadbeef
 135
 136 #define ZBUD_MAX_BUDS 2
 137
 138 struct zbud_hdr {
 139         uint16_t client_id;
 140         uint16_t pool_id;
 141         struct tmem_oid oid;
 142         uint32_t index;
 143         uint16_t size; /* compressed size in bytes, zero means unused */
 144         DECL_SENTINEL
 145 };
 146
 147 struct zbud_page {
 148         struct list_head bud_list;
 149         spinlock_t lock;
 150         struct zbud_hdr buddy[ZBUD_MAX_BUDS];
 151         DECL_SENTINEL
 152         /* followed by NUM_CHUNK aligned CHUNK_SIZE-byte chunks */
 153 };
 154
 155 #define CHUNK_SHIFT     6
 156 #define CHUNK_SIZE      (1 << CHUNK_SHIFT)
 157 #define CHUNK_MASK      (~(CHUNK_SIZE-1))
 158 #define NCHUNKS         (((PAGE_SIZE - sizeof(struct zbud_page)) & \
 159                                 CHUNK_MASK) >> CHUNK_SHIFT)
 160 #define MAX_CHUNK       (NCHUNKS-1)
 161
 162 static struct {
 163         struct list_head list;
 164         unsigned count;
 165 } zbud_unbuddied[NCHUNKS];
 166 /* list N contains pages with N chunks USED and NCHUNKS-N unused */
 167 /* element 0 is never used but optimizing that isn't worth it */
 168 static unsigned long zbud_cumul_chunk_counts[NCHUNKS];
 169
 170 struct list_head zbud_buddied_list;
 171 static unsigned long zcache_zbud_buddied_count;
 172
 173 /* protects the buddied list and all unbuddied lists */
 174 static DEFINE_SPINLOCK(zbud_budlists_spinlock);
 175
 176 static LIST_HEAD(zbpg_unused_list);
 177 static unsigned long zcache_zbpg_unused_list_count;
 178
 179 /* protects the unused page list */
 180 static DEFINE_SPINLOCK(zbpg_unused_list_spinlock);
 181
 182 static atomic_t zcache_zbud_curr_raw_pages;
 183 static atomic_t zcache_zbud_curr_zpages;
 184 static unsigned long zcache_zbud_curr_zbytes;
 185 static unsigned long zcache_zbud_cumul_zpages;
 186 static unsigned long zcache_zbud_cumul_zbytes;
 187 static unsigned long zcache_compress_poor;
 188 static unsigned long zcache_mean_compress_poor;
 189
 190 /* forward references */
 191 static void *zcache_get_free_page(void);
 192 static void zcache_free_page(void *p);
 193
 194 /*
 195  * zbud helper functions
 196  */
 197
 198 static inline unsigned zbud_max_buddy_size(void)
 199 {
 200         return MAX_CHUNK << CHUNK_SHIFT;
 201 }
 202
 203 static inline unsigned zbud_size_to_chunks(unsigned size)
 204 {
 205         BUG_ON(size == 0 || size > zbud_max_buddy_size());
 206         return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
 207 }
 208
 209 static inline int zbud_budnum(struct zbud_hdr *zh)
 210 {
 211         unsigned offset = (unsigned long)zh & (PAGE_SIZE - 1);
 212         struct zbud_page *zbpg = NULL;
 213         unsigned budnum = -1U;
 214         int i;
 215
 216         for (i = 0; i < ZBUD_MAX_BUDS; i++)
 217                 if (offset == offsetof(typeof(*zbpg), buddy[i])) {
 218                         budnum = i;
 219                         break;
 220                 }
 221         BUG_ON(budnum == -1U);
 222         return budnum;
 223 }
 224
 225 static char *zbud_data(struct zbud_hdr *zh, unsigned size)
 226 {
 227         struct zbud_page *zbpg;
 228         char *p;
 229         unsigned budnum;
 230
 231         ASSERT_SENTINEL(zh, ZBH);
 232         budnum = zbud_budnum(zh);
 233         BUG_ON(size == 0 || size > zbud_max_buddy_size());
 234         zbpg = container_of(zh, struct zbud_page, buddy[budnum]);
 235         ASSERT_SPINLOCK(&zbpg->lock);
 236         p = (char *)zbpg;
 237         if (budnum == 0)
 238                 p += ((sizeof(struct zbud_page) + CHUNK_SIZE - 1) &
 239                                                         CHUNK_MASK);
 240         else if (budnum == 1)
 241                 p += PAGE_SIZE - ((size + CHUNK_SIZE - 1) & CHUNK_MASK);
 242         return p;
 243 }
 244
 245 /*
 246  * zbud raw page management
 247  */
 248
 249 static struct zbud_page *zbud_alloc_raw_page(void)
 250 {
 251         struct zbud_page *zbpg = NULL;
 252         struct zbud_hdr *zh0, *zh1;
 253         bool recycled = 0;
 254
 255         /* if any pages on the zbpg list, use one */
 256         spin_lock(&zbpg_unused_list_spinlock);
 257         if (!list_empty(&zbpg_unused_list)) {
 258                 zbpg = list_first_entry(&zbpg_unused_list,
 259                                 struct zbud_page, bud_list);
 260                 list_del_init(&zbpg->bud_list);
 261                 zcache_zbpg_unused_list_count--;
 262                 recycled = 1;
 263         }
 264         spin_unlock(&zbpg_unused_list_spinlock);
 265         if (zbpg == NULL)
 266                 /* none on zbpg list, try to get a kernel page */
 267                 zbpg = zcache_get_free_page();
 268         if (likely(zbpg != NULL)) {
 269                 INIT_LIST_HEAD(&zbpg->bud_list);
 270                 zh0 = &zbpg->buddy[0]; zh1 = &zbpg->buddy[1];
 271                 spin_lock_init(&zbpg->lock);
 272                 if (recycled) {
 273                         ASSERT_INVERTED_SENTINEL(zbpg, ZBPG);
 274                         SET_SENTINEL(zbpg, ZBPG);
 275                         BUG_ON(zh0->size != 0 || tmem_oid_valid(&zh0->oid));
 276                         BUG_ON(zh1->size != 0 || tmem_oid_valid(&zh1->oid));
 277                 } else {
 278                         atomic_inc(&zcache_zbud_curr_raw_pages);
 279                         INIT_LIST_HEAD(&zbpg->bud_list);
 280                         SET_SENTINEL(zbpg, ZBPG);
 281                         zh0->size = 0; zh1->size = 0;
 282                         tmem_oid_set_invalid(&zh0->oid);
 283                         tmem_oid_set_invalid(&zh1->oid);
 284                 }
 285         }
 286         return zbpg;
 287 }
 288
 289 static void zbud_free_raw_page(struct zbud_page *zbpg)
 290 {
 291         struct zbud_hdr *zh0 = &zbpg->buddy[0], *zh1 = &zbpg->buddy[1];
 292
 293         ASSERT_SENTINEL(zbpg, ZBPG);
 294         BUG_ON(!list_empty(&zbpg->bud_list));
 295         ASSERT_SPINLOCK(&zbpg->lock);
 296         BUG_ON(zh0->size != 0 || tmem_oid_valid(&zh0->oid));
 297         BUG_ON(zh1->size != 0 || tmem_oid_valid(&zh1->oid));
 298         INVERT_SENTINEL(zbpg, ZBPG);
 299         spin_unlock(&zbpg->lock);
 300         spin_lock(&zbpg_unused_list_spinlock);
 301         list_add(&zbpg->bud_list, &zbpg_unused_list);
 302         zcache_zbpg_unused_list_count++;
 303         spin_unlock(&zbpg_unused_list_spinlock);
 304 }
 305
 306 /*
 307  * core zbud handling routines
 308  */
 309
 310 static unsigned zbud_free(struct zbud_hdr *zh)
 311 {
 312         unsigned size;
 313
 314         ASSERT_SENTINEL(zh, ZBH);
 315         BUG_ON(!tmem_oid_valid(&zh->oid));
 316         size = zh->size;
 317         BUG_ON(zh->size == 0 || zh->size > zbud_max_buddy_size());
 318         zh->size = 0;
 319         tmem_oid_set_invalid(&zh->oid);
 320         INVERT_SENTINEL(zh, ZBH);
 321         zcache_zbud_curr_zbytes -= size;
 322         atomic_dec(&zcache_zbud_curr_zpages);
 323         return size;
 324 }
 325
 326 static void zbud_free_and_delist(struct zbud_hdr *zh)
 327 {
 328         unsigned chunks;
 329         struct zbud_hdr *zh_other;
 330         unsigned budnum = zbud_budnum(zh), size;
 331         struct zbud_page *zbpg =
 332                 container_of(zh, struct zbud_page, buddy[budnum]);
 333
 334         spin_lock(&zbud_budlists_spinlock);
 335         spin_lock(&zbpg->lock);
 336         if (list_empty(&zbpg->bud_list)) {
 337                 /* ignore zombie page... see zbud_evict_pages() */
 338                 spin_unlock(&zbpg->lock);
 339                 spin_unlock(&zbud_budlists_spinlock);
 340                 return;
 341         }
 342         size = zbud_free(zh);
 343         ASSERT_SPINLOCK(&zbpg->lock);
 344         zh_other = &zbpg->buddy[(budnum == 0) ? 1 : 0];
 345         if (zh_other->size == 0) { /* was unbuddied: unlist and free */
 346                 chunks = zbud_size_to_chunks(size) ;
 347                 BUG_ON(list_empty(&zbud_unbuddied[chunks].list));
 348                 list_del_init(&zbpg->bud_list);
 349                 zbud_unbuddied[chunks].count--;
 350                 spin_unlock(&zbud_budlists_spinlock);
 351                 zbud_free_raw_page(zbpg);
 352         } else { /* was buddied: move remaining buddy to unbuddied list */
 353                 chunks = zbud_size_to_chunks(zh_other->size) ;
 354                 list_del_init(&zbpg->bud_list);
 355                 zcache_zbud_buddied_count--;
 356                 list_add_tail(&zbpg->bud_list, &zbud_unbuddied[chunks].list);
 357                 zbud_unbuddied[chunks].count++;
 358                 spin_unlock(&zbud_budlists_spinlock);
 359                 spin_unlock(&zbpg->lock);
 360         }
 361 }
 362
 363 static struct zbud_hdr *zbud_create(uint16_t client_id, uint16_t pool_id,
 364                                         struct tmem_oid *oid,
 365                                         uint32_t index, struct page *page,
 366                                         void *cdata, unsigned size)
 367 {
 368         struct zbud_hdr *zh0, *zh1, *zh = NULL;
 369         struct zbud_page *zbpg = NULL, *ztmp;
 370         unsigned nchunks;
 371         char *to;
 372         int i, found_good_buddy = 0;
 373
 374         nchunks = zbud_size_to_chunks(size) ;
 375         for (i = MAX_CHUNK - nchunks + 1; i > 0; i--) {
 376                 spin_lock(&zbud_budlists_spinlock);
 377                 if (!list_empty(&zbud_unbuddied[i].list)) {
 378                         list_for_each_entry_safe(zbpg, ztmp,
 379                                     &zbud_unbuddied[i].list, bud_list) {
 380                                 if (spin_trylock(&zbpg->lock)) {
 381                                         found_good_buddy = i;
 382                                         goto found_unbuddied;
 383                                 }
 384                         }
 385                 }
 386                 spin_unlock(&zbud_budlists_spinlock);
 387         }
 388         /* didn't find a good buddy, try allocating a new page */
 389         zbpg = zbud_alloc_raw_page();
 390         if (unlikely(zbpg == NULL))
 391                 goto out;
 392         /* ok, have a page, now compress the data before taking locks */
 393         spin_lock(&zbud_budlists_spinlock);
 394         spin_lock(&zbpg->lock);
 395         list_add_tail(&zbpg->bud_list, &zbud_unbuddied[nchunks].list);
 396         zbud_unbuddied[nchunks].count++;
 397         zh = &zbpg->buddy[0];
 398         goto init_zh;
 399
 400 found_unbuddied:
 401         ASSERT_SPINLOCK(&zbpg->lock);
 402         zh0 = &zbpg->buddy[0]; zh1 = &zbpg->buddy[1];
 403         BUG_ON(!((zh0->size == 0) ^ (zh1->size == 0)));
 404         if (zh0->size != 0) { /* buddy0 in use, buddy1 is vacant */
 405                 ASSERT_SENTINEL(zh0, ZBH);
 406                 zh = zh1;
 407         } else if (zh1->size != 0) { /* buddy1 in use, buddy0 is vacant */
 408                 ASSERT_SENTINEL(zh1, ZBH);
 409                 zh = zh0;
 410         } else
 411                 BUG();
 412         list_del_init(&zbpg->bud_list);
 413         zbud_unbuddied[found_good_buddy].count--;
 414         list_add_tail(&zbpg->bud_list, &zbud_buddied_list);
 415         zcache_zbud_buddied_count++;
 416
 417 init_zh:
 418         SET_SENTINEL(zh, ZBH);
 419         zh->size = size;
 420         zh->index = index;
 421         zh->oid = *oid;
 422         zh->pool_id = pool_id;
 423         zh->client_id = client_id;
 424         to = zbud_data(zh, size);
 425         memcpy(to, cdata, size);
 426         spin_unlock(&zbpg->lock);
 427         spin_unlock(&zbud_budlists_spinlock);
 428
 429         zbud_cumul_chunk_counts[nchunks]++;
 430         atomic_inc(&zcache_zbud_curr_zpages);
 431         zcache_zbud_cumul_zpages++;
 432         zcache_zbud_curr_zbytes += size;
 433         zcache_zbud_cumul_zbytes += size;
 434 out:
 435         return zh;
 436 }
 437
 438 static int zbud_decompress(struct page *page, struct zbud_hdr *zh)
 439 {
 440         struct zbud_page *zbpg;
 441         unsigned budnum = zbud_budnum(zh);
 442         unsigned int out_len = PAGE_SIZE;
 443         char *to_va, *from_va;
 444         unsigned size;
 445         int ret = 0;
 446
 447         zbpg = container_of(zh, struct zbud_page, buddy[budnum]);
 448         spin_lock(&zbpg->lock);
 449         if (list_empty(&zbpg->bud_list)) {
 450                 /* ignore zombie page... see zbud_evict_pages() */
 451                 ret = -EINVAL;
 452                 goto out;
 453         }
 454         ASSERT_SENTINEL(zh, ZBH);
 455         BUG_ON(zh->size == 0 || zh->size > zbud_max_buddy_size());
 456         to_va = kmap_atomic(page);
 457         size = zh->size;
 458         from_va = zbud_data(zh, size);
 459         ret = zcache_comp_op(ZCACHE_COMPOP_DECOMPRESS, from_va, size,
 460                                 to_va, &out_len);
 461         BUG_ON(ret);
 462         BUG_ON(out_len != PAGE_SIZE);
 463         kunmap_atomic(to_va);
 464 out:
 465         spin_unlock(&zbpg->lock);
 466         return ret;
 467 }
 468
 469 /*
 470  * The following routines handle shrinking of ephemeral pages by evicting
 471  * pages "least valuable" first.
 472  */
 473
 474 static unsigned long zcache_evicted_raw_pages;
 475 static unsigned long zcache_evicted_buddied_pages;
 476 static unsigned long zcache_evicted_unbuddied_pages;
 477
 478 static struct tmem_pool *zcache_get_pool_by_id(uint16_t cli_id,
 479                                                 uint16_t poolid);
 480 static void zcache_put_pool(struct tmem_pool *pool);
 481
 482 /*
 483  * Flush and free all zbuds in a zbpg, then free the pageframe
 484  */
 485 static void zbud_evict_zbpg(struct zbud_page *zbpg)
 486 {
 487         struct zbud_hdr *zh;
 488         int i, j;
 489         uint32_t pool_id[ZBUD_MAX_BUDS], client_id[ZBUD_MAX_BUDS];
 490         uint32_t index[ZBUD_MAX_BUDS];
 491         struct tmem_oid oid[ZBUD_MAX_BUDS];
 492         struct tmem_pool *pool;
 493
 494         ASSERT_SPINLOCK(&zbpg->lock);
 495         BUG_ON(!list_empty(&zbpg->bud_list));
 496         for (i = 0, j = 0; i < ZBUD_MAX_BUDS; i++) {
 497                 zh = &zbpg->buddy[i];
 498                 if (zh->size) {
 499                         client_id[j] = zh->client_id;
 500                         pool_id[j] = zh->pool_id;
 501                         oid[j] = zh->oid;
 502                         index[j] = zh->index;
 503                         j++;
 504                         zbud_free(zh);
 505                 }
 506         }
 507         spin_unlock(&zbpg->lock);
 508         for (i = 0; i < j; i++) {
 509                 pool = zcache_get_pool_by_id(client_id[i], pool_id[i]);
 510                 if (pool != NULL) {
 511                         tmem_flush_page(pool, &oid[i], index[i]);
 512                         zcache_put_pool(pool);
 513                 }
 514         }
 515         ASSERT_SENTINEL(zbpg, ZBPG);
 516         spin_lock(&zbpg->lock);
 517         zbud_free_raw_page(zbpg);
 518 }
 519
 520 /*
 521  * Free nr pages.  This code is funky because we want to hold the locks
 522  * protecting various lists for as short a time as possible, and in some
 523  * circumstances the list may change asynchronously when the list lock is
 524  * not held.  In some cases we also trylock not only to avoid waiting on a
 525  * page in use by another cpu, but also to avoid potential deadlock due to
 526  * lock inversion.
 527  */
 528 static void zbud_evict_pages(int nr)
 529 {
 530         struct zbud_page *zbpg;
 531         int i;
 532
 533         /* first try freeing any pages on unused list */
 534 retry_unused_list:
 535         spin_lock_bh(&zbpg_unused_list_spinlock);
 536         if (!list_empty(&zbpg_unused_list)) {
 537                 /* can't walk list here, since it may change when unlocked */
 538                 zbpg = list_first_entry(&zbpg_unused_list,
 539                                 struct zbud_page, bud_list);
 540                 list_del_init(&zbpg->bud_list);
 541                 zcache_zbpg_unused_list_count--;
 542                 atomic_dec(&zcache_zbud_curr_raw_pages);
 543                 spin_unlock_bh(&zbpg_unused_list_spinlock);
 544                 zcache_free_page(zbpg);
 545                 zcache_evicted_raw_pages++;
 546                 if (--nr <= 0)
 547                         goto out;
 548                 goto retry_unused_list;
 549         }
 550         spin_unlock_bh(&zbpg_unused_list_spinlock);
 551
 552         /* now try freeing unbuddied pages, starting with least space avail */
 553         for (i = 0; i < MAX_CHUNK; i++) {
 554 retry_unbud_list_i:
 555                 spin_lock_bh(&zbud_budlists_spinlock);
 556                 if (list_empty(&zbud_unbuddied[i].list)) {
 557                         spin_unlock_bh(&zbud_budlists_spinlock);
 558                         continue;
 559                 }
 560                 list_for_each_entry(zbpg, &zbud_unbuddied[i].list, bud_list) {
 561                         if (unlikely(!spin_trylock(&zbpg->lock)))
 562                                 continue;
 563                         list_del_init(&zbpg->bud_list);
 564                         zbud_unbuddied[i].count--;
 565                         spin_unlock(&zbud_budlists_spinlock);
 566                         zcache_evicted_unbuddied_pages++;
 567                         /* want budlists unlocked when doing zbpg eviction */
 568                         zbud_evict_zbpg(zbpg);
 569                         local_bh_enable();
 570                         if (--nr <= 0)
 571                                 goto out;
 572                         goto retry_unbud_list_i;
 573                 }
 574                 spin_unlock_bh(&zbud_budlists_spinlock);
 575         }
 576
 577         /* as a last resort, free buddied pages */
 578 retry_bud_list:
 579         spin_lock_bh(&zbud_budlists_spinlock);
 580         if (list_empty(&zbud_buddied_list)) {
 581                 spin_unlock_bh(&zbud_budlists_spinlock);
 582                 goto out;
 583         }
 584         list_for_each_entry(zbpg, &zbud_buddied_list, bud_list) {
 585                 if (unlikely(!spin_trylock(&zbpg->lock)))
 586                         continue;
 587                 list_del_init(&zbpg->bud_list);
 588                 zcache_zbud_buddied_count--;
 589                 spin_unlock(&zbud_budlists_spinlock);
 590                 zcache_evicted_buddied_pages++;
 591                 /* want budlists unlocked when doing zbpg eviction */
 592                 zbud_evict_zbpg(zbpg);
 593                 local_bh_enable();
 594                 if (--nr <= 0)
 595                         goto out;
 596                 goto retry_bud_list;
 597         }
 598         spin_unlock_bh(&zbud_budlists_spinlock);
 599 out:
 600         return;
 601 }
 602
 603 static void __init zbud_init(void)
 604 {
 605         int i;
 606
 607         INIT_LIST_HEAD(&zbud_buddied_list);
 608
 609         for (i = 0; i < NCHUNKS; i++)
 610                 INIT_LIST_HEAD(&zbud_unbuddied[i].list);
 611 }
 612
 613 #ifdef CONFIG_SYSFS
 614 /*
 615  * These sysfs routines show a nice distribution of how many zbpg's are
 616  * currently (and have ever been placed) in each unbuddied list.  It's fun
 617  * to watch but can probably go away before final merge.
 618  */
 619 static int zbud_show_unbuddied_list_counts(char *buf)
 620 {
 621         int i;
 622         char *p = buf;
 623
 624         for (i = 0; i < NCHUNKS; i++)
 625                 p += sprintf(p, "%u ", zbud_unbuddied[i].count);
 626         return p - buf;
 627 }
 628
 629 static int zbud_show_cumul_chunk_counts(char *buf)
 630 {
 631         unsigned long i, chunks = 0, total_chunks = 0, sum_total_chunks = 0;
 632         unsigned long total_chunks_lte_21 = 0, total_chunks_lte_32 = 0;
 633         unsigned long total_chunks_lte_42 = 0;
 634         char *p = buf;
 635
 636         for (i = 0; i < NCHUNKS; i++) {
 637                 p += sprintf(p, "%lu ", zbud_cumul_chunk_counts[i]);
 638                 chunks += zbud_cumul_chunk_counts[i];
 639                 total_chunks += zbud_cumul_chunk_counts[i];
 640                 sum_total_chunks += i * zbud_cumul_chunk_counts[i];
 641                 if (i == 21)
 642                         total_chunks_lte_21 = total_chunks;
 643                 if (i == 32)
 644                         total_chunks_lte_32 = total_chunks;
 645                 if (i == 42)
 646                         total_chunks_lte_42 = total_chunks;
 647         }
 648         p += sprintf(p, "<=21:%lu <=32:%lu <=42:%lu, mean:%lu\n",
 649                 total_chunks_lte_21, total_chunks_lte_32, total_chunks_lte_42,
 650                 chunks == 0 ? 0 : sum_total_chunks / chunks);
 651         return p - buf;
 652 }
 653 #endif
 654
 655 /**********
 656  * This "zv" PAM implementation combines the slab-based zsmalloc
 657  * with the crypto compression API to maximize the amount of data that can
 658  * be packed into a physical page.
 659  *
 660  * Zv represents a PAM page with the index and object (plus a "size" value
 661  * necessary for decompression) immediately preceding the compressed data.
 662  */
 663
 664 #define ZVH_SENTINEL  0x43214321
 665
 666 struct zv_hdr {
 667         uint32_t pool_id;
 668         struct tmem_oid oid;
 669         uint32_t index;
 670         size_t size;
 671         DECL_SENTINEL
 672 };
 673
 674 /* rudimentary policy limits */
 675 /* total number of persistent pages may not exceed this percentage */
 676 static unsigned int zv_page_count_policy_percent = 75;
 677 /*
 678  * byte count defining poor compression; pages with greater zsize will be
 679  * rejected
 680  */
 681 static unsigned int zv_max_zsize = (PAGE_SIZE / 8) * 7;
 682 /*
 683  * byte count defining poor *mean* compression; pages with greater zsize
 684  * will be rejected until sufficient better-compressed pages are accepted
 685  * driving the mean below this threshold
 686  */
 687 static unsigned int zv_max_mean_zsize = (PAGE_SIZE / 8) * 5;
 688
 689 static atomic_t zv_curr_dist_counts[NCHUNKS];
 690 static atomic_t zv_cumul_dist_counts[NCHUNKS];
 691
 692 static unsigned long zv_create(struct zs_pool *pool, uint32_t pool_id,
 693                                 struct tmem_oid *oid, uint32_t index,
 694                                 void *cdata, unsigned clen)
 695 {
 696         struct zv_hdr *zv;
 697         u32 size = clen + sizeof(struct zv_hdr);
 698         int chunks = (size + (CHUNK_SIZE - 1)) >> CHUNK_SHIFT;
 699         unsigned long handle = 0;
 700
 701         BUG_ON(!irqs_disabled());
 702         BUG_ON(chunks >= NCHUNKS);
 703         handle = zs_malloc(pool, size);
 704         if (!handle)
 705                 goto out;
 706         atomic_inc(&zv_curr_dist_counts[chunks]);
 707         atomic_inc(&zv_cumul_dist_counts[chunks]);
 708         zv = zs_map_object(pool, handle);
 709         zv->index = index;
 710         zv->oid = *oid;
 711         zv->pool_id = pool_id;
 712         zv->size = clen;
 713         SET_SENTINEL(zv, ZVH);
 714         memcpy((char *)zv + sizeof(struct zv_hdr), cdata, clen);
 715         zs_unmap_object(pool, handle);
 716 out:
 717         return handle;
 718 }
 719
 720 static void zv_free(struct zs_pool *pool, unsigned long handle)
 721 {
 722         unsigned long flags;
 723         struct zv_hdr *zv;
 724         uint16_t size;
 725         int chunks;
 726
 727         zv = zs_map_object(pool, handle);
 728         ASSERT_SENTINEL(zv, ZVH);
 729         size = zv->size + sizeof(struct zv_hdr);
 730         INVERT_SENTINEL(zv, ZVH);
 731         zs_unmap_object(pool, handle);
 732
 733         chunks = (size + (CHUNK_SIZE - 1)) >> CHUNK_SHIFT;
 734         BUG_ON(chunks >= NCHUNKS);
 735         atomic_dec(&zv_curr_dist_counts[chunks]);
 736
 737         local_irq_save(flags);
 738         zs_free(pool, handle);
 739         local_irq_restore(flags);
 740 }
 741
 742 static void zv_decompress(struct page *page, unsigned long handle)
 743 {
 744         unsigned int clen = PAGE_SIZE;
 745         char *to_va;
 746         int ret;
 747         struct zv_hdr *zv;
 748
 749         zv = zs_map_object(zcache_host.zspool, handle);
 750         BUG_ON(zv->size == 0);
 751         ASSERT_SENTINEL(zv, ZVH);
 752         to_va = kmap_atomic(page);
 753         ret = zcache_comp_op(ZCACHE_COMPOP_DECOMPRESS, (char *)zv + sizeof(*zv),
 754                                 zv->size, to_va, &clen);
 755         kunmap_atomic(to_va);
 756         zs_unmap_object(zcache_host.zspool, handle);
 757         BUG_ON(ret);
 758         BUG_ON(clen != PAGE_SIZE);
 759 }
 760
 761 #ifdef CONFIG_SYSFS
 762 /*
 763  * show a distribution of compression stats for zv pages.
 764  */
 765
 766 static int zv_curr_dist_counts_show(char *buf)
 767 {
 768         unsigned long i, n, chunks = 0, sum_total_chunks = 0;
 769         char *p = buf;
 770
 771         for (i = 0; i < NCHUNKS; i++) {
 772                 n = atomic_read(&zv_curr_dist_counts[i]);
 773                 p += sprintf(p, "%lu ", n);
 774                 chunks += n;
 775                 sum_total_chunks += i * n;
 776         }
 777         p += sprintf(p, "mean:%lu\n",
 778                 chunks == 0 ? 0 : sum_total_chunks / chunks);
 779         return p - buf;
 780 }
 781
 782 static int zv_cumul_dist_counts_show(char *buf)
 783 {
 784         unsigned long i, n, chunks = 0, sum_total_chunks = 0;
 785         char *p = buf;
 786
 787         for (i = 0; i < NCHUNKS; i++) {
 788                 n = atomic_read(&zv_cumul_dist_counts[i]);
 789                 p += sprintf(p, "%lu ", n);
 790                 chunks += n;
 791                 sum_total_chunks += i * n;
 792         }
 793         p += sprintf(p, "mean:%lu\n",
 794                 chunks == 0 ? 0 : sum_total_chunks / chunks);
 795         return p - buf;
 796 }
 797
 798 /*
 799  * setting zv_max_zsize via sysfs causes all persistent (e.g. swap)
 800  * pages that don't compress to less than this value (including metadata
 801  * overhead) to be rejected.  We don't allow the value to get too close
 802  * to PAGE_SIZE.
 803  */
 804 static ssize_t zv_max_zsize_show(struct kobject *kobj,
 805                                     struct kobj_attribute *attr,
 806                                     char *buf)
 807 {
 808         return sprintf(buf, "%u\n", zv_max_zsize);
 809 }
 810
 811 static ssize_t zv_max_zsize_store(struct kobject *kobj,
 812                                     struct kobj_attribute *attr,
 813                                     const char *buf, size_t count)
 814 {
 815         unsigned long val;
 816         int err;
 817
 818         if (!capable(CAP_SYS_ADMIN))
 819                 return -EPERM;
 820
 821         err = kstrtoul(buf, 10, &val);
 822         if (err || (val == 0) || (val > (PAGE_SIZE / 8) * 7))
 823                 return -EINVAL;
 824         zv_max_zsize = val;
 825         return count;
 826 }
 827
 828 /*
 829  * setting zv_max_mean_zsize via sysfs causes all persistent (e.g. swap)
 830  * pages that don't compress to less than this value (including metadata
 831  * overhead) to be rejected UNLESS the mean compression is also smaller
 832  * than this value.  In other words, we are load-balancing-by-zsize the
 833  * accepted pages.  Again, we don't allow the value to get too close
 834  * to PAGE_SIZE.
 835  */
 836 static ssize_t zv_max_mean_zsize_show(struct kobject *kobj,
 837                                     struct kobj_attribute *attr,
 838                                     char *buf)
 839 {
 840         return sprintf(buf, "%u\n", zv_max_mean_zsize);
 841 }
 842
 843 static ssize_t zv_max_mean_zsize_store(struct kobject *kobj,
 844                                     struct kobj_attribute *attr,
 845                                     const char *buf, size_t count)
 846 {
 847         unsigned long val;
 848         int err;
 849
 850         if (!capable(CAP_SYS_ADMIN))
 851                 return -EPERM;
 852
 853         err = kstrtoul(buf, 10, &val);
 854         if (err || (val == 0) || (val > (PAGE_SIZE / 8) * 7))
 855                 return -EINVAL;
 856         zv_max_mean_zsize = val;
 857         return count;
 858 }
 859
 860 /*
 861  * setting zv_page_count_policy_percent via sysfs sets an upper bound of
 862  * persistent (e.g. swap) pages that will be retained according to:
 863  *     (zv_page_count_policy_percent * totalram_pages) / 100)
 864  * when that limit is reached, further puts will be rejected (until
 865  * some pages have been flushed).  Note that, due to compression,
 866  * this number may exceed 100; it defaults to 75 and we set an
 867  * arbitary limit of 150.  A poor choice will almost certainly result
 868  * in OOM's, so this value should only be changed prudently.
 869  */
 870 static ssize_t zv_page_count_policy_percent_show(struct kobject *kobj,
 871                                                  struct kobj_attribute *attr,
 872                                                  char *buf)
 873 {
 874         return sprintf(buf, "%u\n", zv_page_count_policy_percent);
 875 }
 876
 877 static ssize_t zv_page_count_policy_percent_store(struct kobject *kobj,
 878                                                   struct kobj_attribute *attr,
 879                                                   const char *buf, size_t count)
 880 {
 881         unsigned long val;
 882         int err;
 883
 884         if (!capable(CAP_SYS_ADMIN))
 885                 return -EPERM;
 886
 887         err = kstrtoul(buf, 10, &val);
 888         if (err || (val == 0) || (val > 150))
 889                 return -EINVAL;
 890         zv_page_count_policy_percent = val;
 891         return count;
 892 }
 893
 894 static struct kobj_attribute zcache_zv_max_zsize_attr = {
 895                 .attr = { .name = "zv_max_zsize", .mode = 0644 },
 896                 .show = zv_max_zsize_show,
 897                 .store = zv_max_zsize_store,
 898 };
 899
 900 static struct kobj_attribute zcache_zv_max_mean_zsize_attr = {
 901                 .attr = { .name = "zv_max_mean_zsize", .mode = 0644 },
 902                 .show = zv_max_mean_zsize_show,
 903                 .store = zv_max_mean_zsize_store,
 904 };
 905
 906 static struct kobj_attribute zcache_zv_page_count_policy_percent_attr = {
 907                 .attr = { .name = "zv_page_count_policy_percent",
 908                           .mode = 0644 },
 909                 .show = zv_page_count_policy_percent_show,
 910                 .store = zv_page_count_policy_percent_store,
 911 };
 912 #endif
 913
 914 /*
 915  * zcache core code starts here
 916  */
 917
 918 /* useful stats not collected by cleancache or frontswap */
 919 static unsigned long zcache_flush_total;
 920 static unsigned long zcache_flush_found;
 921 static unsigned long zcache_flobj_total;
 922 static unsigned long zcache_flobj_found;
 923 static unsigned long zcache_failed_eph_puts;
 924 static unsigned long zcache_failed_pers_puts;
 925
 926 /*
 927  * Tmem operations assume the poolid implies the invoking client.
 928  * Zcache only has one client (the kernel itself): LOCAL_CLIENT.
 929  * RAMster has each client numbered by cluster node, and a KVM version
 930  * of zcache would have one client per guest and each client might
 931  * have a poolid==N.
 932  */
 933 static struct tmem_pool *zcache_get_pool_by_id(uint16_t cli_id, uint16_t poolid)
 934 {
 935         struct tmem_pool *pool = NULL;
 936         struct zcache_client *cli = NULL;
 937
 938         if (cli_id == LOCAL_CLIENT)
 939                 cli = &zcache_host;
 940         else {
 941                 if (cli_id >= MAX_CLIENTS)
 942                         goto out;
 943                 cli = &zcache_clients[cli_id];
 944                 if (cli == NULL)
 945                         goto out;
 946         }
 947
 948         atomic_inc(&cli->refcount);
 949         pool = idr_find(&cli->tmem_pools, poolid);
 950         if (pool != NULL)
 951                 atomic_inc(&pool->refcount);
 952 out:
 953         return pool;
 954 }
 955
 956 static void zcache_put_pool(struct tmem_pool *pool)
 957 {
 958         struct zcache_client *cli = NULL;
 959
 960         if (pool == NULL)
 961                 BUG();
 962         cli = pool->client;
 963         atomic_dec(&pool->refcount);
 964         atomic_dec(&cli->refcount);
 965 }
 966
 967 int zcache_new_client(uint16_t cli_id)
 968 {
 969         struct zcache_client *cli = NULL;
 970         int ret = -1;
 971
 972         if (cli_id == LOCAL_CLIENT)
 973                 cli = &zcache_host;
 974         else if ((unsigned int)cli_id < MAX_CLIENTS)
 975                 cli = &zcache_clients[cli_id];
 976         if (cli == NULL)
 977                 goto out;
 978         if (cli->allocated)
 979                 goto out;
 980         cli->allocated = 1;
 981 #ifdef CONFIG_FRONTSWAP
 982         cli->zspool = zs_create_pool("zcache", ZCACHE_GFP_MASK);
 983         if (cli->zspool == NULL)
 984                 goto out;
 985         idr_init(&cli->tmem_pools);
 986 #endif
 987         ret = 0;
 988 out:
 989         return ret;
 990 }
 991
 992 /* counters for debugging */
 993 static unsigned long zcache_failed_get_free_pages;
 994 static unsigned long zcache_failed_alloc;
 995 static unsigned long zcache_put_to_flush;
 996
 997 /*
 998  * for now, used named slabs so can easily track usage; later can
 999  * either just use kmalloc, or perhaps add a slab-like allocator
1000  * to more carefully manage total memory utilization
1001  */
1002 static struct kmem_cache *zcache_objnode_cache;
1003 static struct kmem_cache *zcache_obj_cache;
1004 static atomic_t zcache_curr_obj_count = ATOMIC_INIT(0);
1005 static unsigned long zcache_curr_obj_count_max;
1006 static atomic_t zcache_curr_objnode_count = ATOMIC_INIT(0);
1007 static unsigned long zcache_curr_objnode_count_max;
1008
1009 /*
1010  * to avoid memory allocation recursion (e.g. due to direct reclaim), we
1011  * preload all necessary data structures so the hostops callbacks never
1012  * actually do a malloc
1013  */
1014 struct zcache_preload {
1015         void *page;
1016         struct tmem_obj *obj;
1017         int nr;
1018         struct tmem_objnode *objnodes[OBJNODE_TREE_MAX_PATH];
1019 };
1020 static DEFINE_PER_CPU(struct zcache_preload, zcache_preloads) = { 0, };
1021
1022 static int zcache_do_preload(struct tmem_pool *pool)
1023 {
1024         struct zcache_preload *kp;
1025         struct tmem_objnode *objnode;
1026         struct tmem_obj *obj;
1027         void *page;
1028         int ret = -ENOMEM;
1029
1030         if (unlikely(zcache_objnode_cache == NULL))
1031                 goto out;
1032         if (unlikely(zcache_obj_cache == NULL))
1033                 goto out;
1034
1035         /* IRQ has already been disabled. */
1036         kp = &__get_cpu_var(zcache_preloads);
1037         while (kp->nr < ARRAY_SIZE(kp->objnodes)) {
1038                 objnode = kmem_cache_alloc(zcache_objnode_cache,
1039                                 ZCACHE_GFP_MASK);
1040                 if (unlikely(objnode == NULL)) {
1041                         zcache_failed_alloc++;
1042                         goto out;
1043                 }
1044
1045                 kp->objnodes[kp->nr++] = objnode;
1046         }
1047
1048         obj = kmem_cache_alloc(zcache_obj_cache, ZCACHE_GFP_MASK);
1049         if (unlikely(obj == NULL)) {
1050                 zcache_failed_alloc++;
1051                 goto out;
1052         }
1053
1054         page = (void *)__get_free_page(ZCACHE_GFP_MASK);
1055         if (unlikely(page == NULL)) {
1056                 zcache_failed_get_free_pages++;
1057                 kmem_cache_free(zcache_obj_cache, obj);
1058                 goto out;
1059         }
1060
1061         if (kp->obj == NULL)
1062                 kp->obj = obj;
1063         else
1064                 kmem_cache_free(zcache_obj_cache, obj);
1065
1066         if (kp->page == NULL)
1067                 kp->page = page;
1068         else
1069                 free_page((unsigned long)page);
1070
1071         ret = 0;
1072 out:
1073         return ret;
1074 }
1075
1076 static void *zcache_get_free_page(void)
1077 {
1078         struct zcache_preload *kp;
1079         void *page;
1080
1081         kp = &__get_cpu_var(zcache_preloads);
1082         page = kp->page;
1083         BUG_ON(page == NULL);
1084         kp->page = NULL;
1085         return page;
1086 }
1087
1088 static void zcache_free_page(void *p)
1089 {
1090         free_page((unsigned long)p);
1091 }
1092
1093 /*
1094  * zcache implementation for tmem host ops
1095  */
1096
1097 static struct tmem_objnode *zcache_objnode_alloc(struct tmem_pool *pool)
1098 {
1099         struct tmem_objnode *objnode = NULL;
1100         unsigned long count;
1101         struct zcache_preload *kp;
1102
1103         kp = &__get_cpu_var(zcache_preloads);
1104         if (kp->nr <= 0)
1105                 goto out;
1106         objnode = kp->objnodes[kp->nr - 1];
1107         BUG_ON(objnode == NULL);
1108         kp->objnodes[kp->nr - 1] = NULL;
1109         kp->nr--;
1110         count = atomic_inc_return(&zcache_curr_objnode_count);
1111         if (count > zcache_curr_objnode_count_max)
1112                 zcache_curr_objnode_count_max = count;
1113 out:
1114         return objnode;
1115 }
1116
1117 static void zcache_objnode_free(struct tmem_objnode *objnode,
1118                                         struct tmem_pool *pool)
1119 {
1120         atomic_dec(&zcache_curr_objnode_count);
1121         BUG_ON(atomic_read(&zcache_curr_objnode_count) < 0);
1122         kmem_cache_free(zcache_objnode_cache, objnode);
1123 }
1124
1125 static struct tmem_obj *zcache_obj_alloc(struct tmem_pool *pool)
1126 {
1127         struct tmem_obj *obj = NULL;
1128         unsigned long count;
1129         struct zcache_preload *kp;
1130
1131         kp = &__get_cpu_var(zcache_preloads);
1132         obj = kp->obj;
1133         BUG_ON(obj == NULL);
1134         kp->obj = NULL;
1135         count = atomic_inc_return(&zcache_curr_obj_count);
1136         if (count > zcache_curr_obj_count_max)
1137                 zcache_curr_obj_count_max = count;
1138         return obj;
1139 }
1140
1141 static void zcache_obj_free(struct tmem_obj *obj, struct tmem_pool *pool)
1142 {
1143         atomic_dec(&zcache_curr_obj_count);
1144         BUG_ON(atomic_read(&zcache_curr_obj_count) < 0);
1145         kmem_cache_free(zcache_obj_cache, obj);
1146 }
1147
1148 static struct tmem_hostops zcache_hostops = {
1149         .obj_alloc = zcache_obj_alloc,
1150         .obj_free = zcache_obj_free,
1151         .objnode_alloc = zcache_objnode_alloc,
1152         .objnode_free = zcache_objnode_free,
1153 };
1154
1155 /*
1156  * zcache implementations for PAM page descriptor ops
1157  */
1158
1159 static atomic_t zcache_curr_eph_pampd_count = ATOMIC_INIT(0);
1160 static unsigned long zcache_curr_eph_pampd_count_max;
1161 static atomic_t zcache_curr_pers_pampd_count = ATOMIC_INIT(0);
1162 static unsigned long zcache_curr_pers_pampd_count_max;
1163
1164 /* forward reference */
1165 static int zcache_compress(struct page *from, void **out_va, unsigned *out_len);
1166
1167 static void *zcache_pampd_create(char *data, size_t size, bool raw, int eph,
1168                                 struct tmem_pool *pool, struct tmem_oid *oid,
1169                                  uint32_t index)
1170 {
1171         void *pampd = NULL, *cdata;
1172         unsigned clen;
1173         int ret;
1174         unsigned long count;
1175         struct page *page = (struct page *)(data);
1176         struct zcache_client *cli = pool->client;
1177         uint16_t client_id = get_client_id_from_client(cli);
1178         unsigned long zv_mean_zsize;
1179         unsigned long curr_pers_pampd_count;
1180         u64 total_zsize;
1181
1182         if (eph) {
1183                 ret = zcache_compress(page, &cdata, &clen);
1184                 if (ret == 0)
1185                         goto out;
1186                 if (clen == 0 || clen > zbud_max_buddy_size()) {
1187                         zcache_compress_poor++;
1188                         goto out;
1189                 }
1190                 pampd = (void *)zbud_create(client_id, pool->pool_id, oid,
1191                                                 index, page, cdata, clen);
1192                 if (pampd != NULL) {
1193                         count = atomic_inc_return(&zcache_curr_eph_pampd_count);
1194                         if (count > zcache_curr_eph_pampd_count_max)
1195                                 zcache_curr_eph_pampd_count_max = count;
1196                 }
1197         } else {
1198                 curr_pers_pampd_count =
1199                         atomic_read(&zcache_curr_pers_pampd_count);
1200                 if (curr_pers_pampd_count >
1201                     (zv_page_count_policy_percent * totalram_pages) / 100)
1202                         goto out;
1203                 ret = zcache_compress(page, &cdata, &clen);
1204                 if (ret == 0)
1205                         goto out;
1206                 /* reject if compression is too poor */
1207                 if (clen > zv_max_zsize) {
1208                         zcache_compress_poor++;
1209                         goto out;
1210                 }
1211                 /* reject if mean compression is too poor */
1212                 if ((clen > zv_max_mean_zsize) && (curr_pers_pampd_count > 0)) {
1213                         total_zsize = zs_get_total_size_bytes(cli->zspool);
1214                         zv_mean_zsize = div_u64(total_zsize,
1215                                                 curr_pers_pampd_count);
1216                         if (zv_mean_zsize > zv_max_mean_zsize) {
1217                                 zcache_mean_compress_poor++;
1218                                 goto out;
1219                         }
1220                 }
1221                 pampd = (void *)zv_create(cli->zspool, pool->pool_id,
1222                                                 oid, index, cdata, clen);
1223                 if (pampd == NULL)
1224                         goto out;
1225                 count = atomic_inc_return(&zcache_curr_pers_pampd_count);
1226                 if (count > zcache_curr_pers_pampd_count_max)
1227                         zcache_curr_pers_pampd_count_max = count;
1228         }
1229 out:
1230         return pampd;
1231 }
1232
1233 /*
1234  * fill the pageframe corresponding to the struct page with the data
1235  * from the passed pampd
1236  */
1237 static int zcache_pampd_get_data(char *data, size_t *bufsize, bool raw,
1238                                         void *pampd, struct tmem_pool *pool,
1239                                         struct tmem_oid *oid, uint32_t index)
1240 {
1241         int ret = 0;
1242
1243         BUG_ON(is_ephemeral(pool));
1244         zv_decompress((struct page *)(data), (unsigned long)pampd);
1245         return ret;
1246 }
1247
1248 /*
1249  * fill the pageframe corresponding to the struct page with the data
1250  * from the passed pampd
1251  */
1252 static int zcache_pampd_get_data_and_free(char *data, size_t *bufsize, bool raw,
1253                                         void *pampd, struct tmem_pool *pool,
1254                                         struct tmem_oid *oid, uint32_t index)
1255 {
1256         int ret = 0;
1257
1258         BUG_ON(!is_ephemeral(pool));
1259         zbud_decompress((struct page *)(data), pampd);
1260         zbud_free_and_delist((struct zbud_hdr *)pampd);
1261         atomic_dec(&zcache_curr_eph_pampd_count);
1262         return ret;
1263 }
1264
1265 /*
1266  * free the pampd and remove it from any zcache lists
1267  * pampd must no longer be pointed to from any tmem data structures!
1268  */
1269 static void zcache_pampd_free(void *pampd, struct tmem_pool *pool,
1270                                 struct tmem_oid *oid, uint32_t index)
1271 {
1272         struct zcache_client *cli = pool->client;
1273
1274         if (is_ephemeral(pool)) {
1275                 zbud_free_and_delist((struct zbud_hdr *)pampd);
1276                 atomic_dec(&zcache_curr_eph_pampd_count);
1277                 BUG_ON(atomic_read(&zcache_curr_eph_pampd_count) < 0);
1278         } else {
1279                 zv_free(cli->zspool, (unsigned long)pampd);
1280                 atomic_dec(&zcache_curr_pers_pampd_count);
1281                 BUG_ON(atomic_read(&zcache_curr_pers_pampd_count) < 0);
1282         }
1283 }
1284
1285 static void zcache_pampd_free_obj(struct tmem_pool *pool, struct tmem_obj *obj)
1286 {
1287 }
1288
1289 static void zcache_pampd_new_obj(struct tmem_obj *obj)
1290 {
1291 }
1292
1293 static int zcache_pampd_replace_in_obj(void *pampd, struct tmem_obj *obj)
1294 {
1295         return -1;
1296 }
1297
1298 static bool zcache_pampd_is_remote(void *pampd)
1299 {
1300         return 0;
1301 }
1302
1303 static struct tmem_pamops zcache_pamops = {
1304         .create = zcache_pampd_create,
1305         .get_data = zcache_pampd_get_data,
1306         .get_data_and_free = zcache_pampd_get_data_and_free,
1307         .free = zcache_pampd_free,
1308         .free_obj = zcache_pampd_free_obj,
1309         .new_obj = zcache_pampd_new_obj,
1310         .replace_in_obj = zcache_pampd_replace_in_obj,
1311         .is_remote = zcache_pampd_is_remote,
1312 };
1313
1314 /*
1315  * zcache compression/decompression and related per-cpu stuff
1316  */
1317
1318 static DEFINE_PER_CPU(unsigned char *, zcache_dstmem);
1319 #define ZCACHE_DSTMEM_ORDER 1
1320
1321 static int zcache_compress(struct page *from, void **out_va, unsigned *out_len)
1322 {
1323         int ret = 0;
1324         unsigned char *dmem = __get_cpu_var(zcache_dstmem);
1325         char *from_va;
1326
1327         BUG_ON(!irqs_disabled());
1328         if (unlikely(dmem == NULL))
1329                 goto out;  /* no buffer or no compressor so can't compress */
1330         *out_len = PAGE_SIZE << ZCACHE_DSTMEM_ORDER;
1331         from_va = kmap_atomic(from);
1332         mb();
1333         ret = zcache_comp_op(ZCACHE_COMPOP_COMPRESS, from_va, PAGE_SIZE, dmem,
1334                                 out_len);
1335         BUG_ON(ret);
1336         *out_va = dmem;
1337         kunmap_atomic(from_va);
1338         ret = 1;
1339 out:
1340         return ret;
1341 }
1342
1343 static int zcache_comp_cpu_up(int cpu)
1344 {
1345         struct crypto_comp *tfm;
1346
1347         tfm = crypto_alloc_comp(zcache_comp_name, 0, 0);
1348         if (IS_ERR(tfm))
1349                 return NOTIFY_BAD;
1350         *per_cpu_ptr(zcache_comp_pcpu_tfms, cpu) = tfm;
1351         return NOTIFY_OK;
1352 }
1353
1354 static void zcache_comp_cpu_down(int cpu)
1355 {
1356         struct crypto_comp *tfm;
1357
1358         tfm = *per_cpu_ptr(zcache_comp_pcpu_tfms, cpu);
1359         crypto_free_comp(tfm);
1360         *per_cpu_ptr(zcache_comp_pcpu_tfms, cpu) = NULL;
1361 }
1362
1363 static int zcache_cpu_notifier(struct notifier_block *nb,
1364                                 unsigned long action, void *pcpu)
1365 {
1366         int ret, cpu = (long)pcpu;
1367         struct zcache_preload *kp;
1368
1369         switch (action) {
1370         case CPU_UP_PREPARE:
1371                 ret = zcache_comp_cpu_up(cpu);
1372                 if (ret != NOTIFY_OK) {
1373                         pr_err("zcache: can't allocate compressor transform\n");
1374                         return ret;
1375                 }
1376                 per_cpu(zcache_dstmem, cpu) = (void *)__get_free_pages(
1377                         GFP_KERNEL | __GFP_REPEAT, ZCACHE_DSTMEM_ORDER);
1378                 break;
1379         case CPU_DEAD:
1380         case CPU_UP_CANCELED:
1381                 zcache_comp_cpu_down(cpu);
1382                 free_pages((unsigned long)per_cpu(zcache_dstmem, cpu),
1383                         ZCACHE_DSTMEM_ORDER);
1384                 per_cpu(zcache_dstmem, cpu) = NULL;
1385                 kp = &per_cpu(zcache_preloads, cpu);
1386                 while (kp->nr) {
1387                         kmem_cache_free(zcache_objnode_cache,
1388                                         kp->objnodes[kp->nr - 1]);
1389                         kp->objnodes[kp->nr - 1] = NULL;
1390                         kp->nr--;
1391                 }
1392                 if (kp->obj) {
1393                         kmem_cache_free(zcache_obj_cache, kp->obj);
1394                         kp->obj = NULL;
1395                 }
1396                 if (kp->page) {
1397                         free_page((unsigned long)kp->page);
1398                         kp->page = NULL;
1399                 }
1400                 break;
1401         default:
1402                 break;
1403         }
1404         return NOTIFY_OK;
1405 }
1406
1407 static struct notifier_block zcache_cpu_notifier_block = {
1408         .notifier_call = zcache_cpu_notifier
1409 };
1410
1411 #ifdef CONFIG_SYSFS
1412 #define ZCACHE_SYSFS_RO(_name) \
1413         static ssize_t zcache_##_name##_show(struct kobject *kobj, \
1414                                 struct kobj_attribute *attr, char *buf) \
1415         { \
1416                 return sprintf(buf, "%lu\n", zcache_##_name); \
1417         } \
1418         static struct kobj_attribute zcache_##_name##_attr = { \
1419                 .attr = { .name = __stringify(_name), .mode = 0444 }, \
1420                 .show = zcache_##_name##_show, \
1421         }
1422
1423 #define ZCACHE_SYSFS_RO_ATOMIC(_name) \
1424         static ssize_t zcache_##_name##_show(struct kobject *kobj, \
1425                                 struct kobj_attribute *attr, char *buf) \
1426         { \
1427             return sprintf(buf, "%d\n", atomic_read(&zcache_##_name)); \
1428         } \
1429         static struct kobj_attribute zcache_##_name##_attr = { \
1430                 .attr = { .name = __stringify(_name), .mode = 0444 }, \
1431                 .show = zcache_##_name##_show, \
1432         }
1433
1434 #define ZCACHE_SYSFS_RO_CUSTOM(_name, _func) \
1435         static ssize_t zcache_##_name##_show(struct kobject *kobj, \
1436                                 struct kobj_attribute *attr, char *buf) \
1437         { \
1438             return _func(buf); \
1439         } \
1440         static struct kobj_attribute zcache_##_name##_attr = { \
1441                 .attr = { .name = __stringify(_name), .mode = 0444 }, \
1442                 .show = zcache_##_name##_show, \
1443         }
1444
1445 ZCACHE_SYSFS_RO(curr_obj_count_max);
1446 ZCACHE_SYSFS_RO(curr_objnode_count_max);
1447 ZCACHE_SYSFS_RO(flush_total);
1448 ZCACHE_SYSFS_RO(flush_found);
1449 ZCACHE_SYSFS_RO(flobj_total);
1450 ZCACHE_SYSFS_RO(flobj_found);
1451 ZCACHE_SYSFS_RO(failed_eph_puts);
1452 ZCACHE_SYSFS_RO(failed_pers_puts);
1453 ZCACHE_SYSFS_RO(zbud_curr_zbytes);
1454 ZCACHE_SYSFS_RO(zbud_cumul_zpages);
1455 ZCACHE_SYSFS_RO(zbud_cumul_zbytes);
1456 ZCACHE_SYSFS_RO(zbud_buddied_count);
1457 ZCACHE_SYSFS_RO(zbpg_unused_list_count);
1458 ZCACHE_SYSFS_RO(evicted_raw_pages);
1459 ZCACHE_SYSFS_RO(evicted_unbuddied_pages);
1460 ZCACHE_SYSFS_RO(evicted_buddied_pages);
1461 ZCACHE_SYSFS_RO(failed_get_free_pages);
1462 ZCACHE_SYSFS_RO(failed_alloc);
1463 ZCACHE_SYSFS_RO(put_to_flush);
1464 ZCACHE_SYSFS_RO(compress_poor);
1465 ZCACHE_SYSFS_RO(mean_compress_poor);
1466 ZCACHE_SYSFS_RO_ATOMIC(zbud_curr_raw_pages);
1467 ZCACHE_SYSFS_RO_ATOMIC(zbud_curr_zpages);
1468 ZCACHE_SYSFS_RO_ATOMIC(curr_obj_count);
1469 ZCACHE_SYSFS_RO_ATOMIC(curr_objnode_count);
1470 ZCACHE_SYSFS_RO_CUSTOM(zbud_unbuddied_list_counts,
1471                         zbud_show_unbuddied_list_counts);
1472 ZCACHE_SYSFS_RO_CUSTOM(zbud_cumul_chunk_counts,
1473                         zbud_show_cumul_chunk_counts);
1474 ZCACHE_SYSFS_RO_CUSTOM(zv_curr_dist_counts,
1475                         zv_curr_dist_counts_show);
1476 ZCACHE_SYSFS_RO_CUSTOM(zv_cumul_dist_counts,
1477                         zv_cumul_dist_counts_show);
1478
1479 static struct attribute *zcache_attrs[] = {
1480         &zcache_curr_obj_count_attr.attr,
1481         &zcache_curr_obj_count_max_attr.attr,
1482         &zcache_curr_objnode_count_attr.attr,
1483         &zcache_curr_objnode_count_max_attr.attr,
1484         &zcache_flush_total_attr.attr,
1485         &zcache_flobj_total_attr.attr,
1486         &zcache_flush_found_attr.attr,
1487         &zcache_flobj_found_attr.attr,
1488         &zcache_failed_eph_puts_attr.attr,
1489         &zcache_failed_pers_puts_attr.attr,
1490         &zcache_compress_poor_attr.attr,
1491         &zcache_mean_compress_poor_attr.attr,
1492         &zcache_zbud_curr_raw_pages_attr.attr,
1493         &zcache_zbud_curr_zpages_attr.attr,
1494         &zcache_zbud_curr_zbytes_attr.attr,
1495         &zcache_zbud_cumul_zpages_attr.attr,
1496         &zcache_zbud_cumul_zbytes_attr.attr,
1497         &zcache_zbud_buddied_count_attr.attr,
1498         &zcache_zbpg_unused_list_count_attr.attr,
1499         &zcache_evicted_raw_pages_attr.attr,
1500         &zcache_evicted_unbuddied_pages_attr.attr,
1501         &zcache_evicted_buddied_pages_attr.attr,
1502         &zcache_failed_get_free_pages_attr.attr,
1503         &zcache_failed_alloc_attr.attr,
1504         &zcache_put_to_flush_attr.attr,
1505         &zcache_zbud_unbuddied_list_counts_attr.attr,
1506         &zcache_zbud_cumul_chunk_counts_attr.attr,
1507         &zcache_zv_curr_dist_counts_attr.attr,
1508         &zcache_zv_cumul_dist_counts_attr.attr,
1509         &zcache_zv_max_zsize_attr.attr,
1510         &zcache_zv_max_mean_zsize_attr.attr,
1511         &zcache_zv_page_count_policy_percent_attr.attr,
1512         NULL,
1513 };
1514
1515 static struct attribute_group zcache_attr_group = {
1516         .attrs = zcache_attrs,
1517         .name = "zcache",
1518 };
1519
1520 #endif /* CONFIG_SYSFS */
1521 /*
1522  * When zcache is disabled ("frozen"), pools can be created and destroyed,
1523  * but all puts (and thus all other operations that require memory allocation)
1524  * must fail.  If zcache is unfrozen, accepts puts, then frozen again,
1525  * data consistency requires all puts while frozen to be converted into
1526  * flushes.
1527  */
1528 static bool zcache_freeze;
1529
1530 /*
1531  * zcache shrinker interface (only useful for ephemeral pages, so zbud only)
1532  */
1533 static int shrink_zcache_memory(struct shrinker *shrink,
1534                                 struct shrink_control *sc)
1535 {
1536         int ret = -1;
1537         int nr = sc->nr_to_scan;
1538         gfp_t gfp_mask = sc->gfp_mask;
1539
1540         if (nr >= 0) {
1541                 if (!(gfp_mask & __GFP_FS))
1542                         /* does this case really need to be skipped? */
1543                         goto out;
1544                 zbud_evict_pages(nr);
1545         }
1546         ret = (int)atomic_read(&zcache_zbud_curr_raw_pages);
1547 out:
1548         return ret;
1549 }
1550
1551 static struct shrinker zcache_shrinker = {
1552         .shrink = shrink_zcache_memory,
1553         .seeks = DEFAULT_SEEKS,
1554 };
1555
1556 /*
1557  * zcache shims between cleancache/frontswap ops and tmem
1558  */
1559
1560 static int zcache_put_page(int cli_id, int pool_id, struct tmem_oid *oidp,
1561                                 uint32_t index, struct page *page)
1562 {
1563         struct tmem_pool *pool;
1564         int ret = -1;
1565
1566         BUG_ON(!irqs_disabled());
1567         pool = zcache_get_pool_by_id(cli_id, pool_id);
1568         if (unlikely(pool == NULL))
1569                 goto out;
1570         if (!zcache_freeze && zcache_do_preload(pool) == 0) {
1571                 /* preload does preempt_disable on success */
1572                 ret = tmem_put(pool, oidp, index, (char *)(page),
1573                                 PAGE_SIZE, 0, is_ephemeral(pool));
1574                 if (ret < 0) {
1575                         if (is_ephemeral(pool))
1576                                 zcache_failed_eph_puts++;
1577                         else
1578                                 zcache_failed_pers_puts++;
1579                 }
1580                 zcache_put_pool(pool);
1581         } else {
1582                 zcache_put_to_flush++;
1583                 if (atomic_read(&pool->obj_count) > 0)
1584                         /* the put fails whether the flush succeeds or not */
1585                         (void)tmem_flush_page(pool, oidp, index);
1586                 zcache_put_pool(pool);
1587         }
1588 out:
1589         return ret;
1590 }
1591
1592 static int zcache_get_page(int cli_id, int pool_id, struct tmem_oid *oidp,
1593                                 uint32_t index, struct page *page)
1594 {
1595         struct tmem_pool *pool;
1596         int ret = -1;
1597         unsigned long flags;
1598         size_t size = PAGE_SIZE;
1599
1600         local_irq_save(flags);
1601         pool = zcache_get_pool_by_id(cli_id, pool_id);
1602         if (likely(pool != NULL)) {
1603                 if (atomic_read(&pool->obj_count) > 0)
1604                         ret = tmem_get(pool, oidp, index, (char *)(page),
1605                                         &size, 0, is_ephemeral(pool));
1606                 zcache_put_pool(pool);
1607         }
1608         local_irq_restore(flags);
1609         return ret;
1610 }
1611
1612 static int zcache_flush_page(int cli_id, int pool_id,
1613                                 struct tmem_oid *oidp, uint32_t index)
1614 {
1615         struct tmem_pool *pool;
1616         int ret = -1;
1617         unsigned long flags;
1618
1619         local_irq_save(flags);
1620         zcache_flush_total++;
1621         pool = zcache_get_pool_by_id(cli_id, pool_id);
1622         if (likely(pool != NULL)) {
1623                 if (atomic_read(&pool->obj_count) > 0)
1624                         ret = tmem_flush_page(pool, oidp, index);
1625                 zcache_put_pool(pool);
1626         }
1627         if (ret >= 0)
1628                 zcache_flush_found++;
1629         local_irq_restore(flags);
1630         return ret;
1631 }
1632
1633 static int zcache_flush_object(int cli_id, int pool_id,
1634                                 struct tmem_oid *oidp)
1635 {
1636         struct tmem_pool *pool;
1637         int ret = -1;
1638         unsigned long flags;
1639
1640         local_irq_save(flags);
1641         zcache_flobj_total++;
1642         pool = zcache_get_pool_by_id(cli_id, pool_id);
1643         if (likely(pool != NULL)) {
1644                 if (atomic_read(&pool->obj_count) > 0)
1645                         ret = tmem_flush_object(pool, oidp);
1646                 zcache_put_pool(pool);
1647         }
1648         if (ret >= 0)
1649                 zcache_flobj_found++;
1650         local_irq_restore(flags);
1651         return ret;
1652 }
1653
1654 static int zcache_destroy_pool(int cli_id, int pool_id)
1655 {
1656         struct tmem_pool *pool = NULL;
1657         struct zcache_client *cli = NULL;
1658         int ret = -1;
1659
1660         if (pool_id < 0)
1661                 goto out;
1662         if (cli_id == LOCAL_CLIENT)
1663                 cli = &zcache_host;
1664         else if ((unsigned int)cli_id < MAX_CLIENTS)
1665                 cli = &zcache_clients[cli_id];
1666         if (cli == NULL)
1667                 goto out;
1668         atomic_inc(&cli->refcount);
1669         pool = idr_find(&cli->tmem_pools, pool_id);
1670         if (pool == NULL)
1671                 goto out;
1672         idr_remove(&cli->tmem_pools, pool_id);
1673         /* wait for pool activity on other cpus to quiesce */
1674         while (atomic_read(&pool->refcount) != 0)
1675                 ;
1676         atomic_dec(&cli->refcount);
1677         local_bh_disable();
1678         ret = tmem_destroy_pool(pool);
1679         local_bh_enable();
1680         kfree(pool);
1681         pr_info("zcache: destroyed pool id=%d, cli_id=%d\n",
1682                         pool_id, cli_id);
1683 out:
1684         return ret;
1685 }
1686
1687 static int zcache_new_pool(uint16_t cli_id, uint32_t flags)
1688 {
1689         int poolid = -1;
1690         struct tmem_pool *pool;
1691         struct zcache_client *cli = NULL;
1692         int r;
1693
1694         if (cli_id == LOCAL_CLIENT)
1695                 cli = &zcache_host;
1696         else if ((unsigned int)cli_id < MAX_CLIENTS)
1697                 cli = &zcache_clients[cli_id];
1698         if (cli == NULL)
1699                 goto out;
1700         atomic_inc(&cli->refcount);
1701         pool = kmalloc(sizeof(struct tmem_pool), GFP_ATOMIC);
1702         if (pool == NULL) {
1703                 pr_info("zcache: pool creation failed: out of memory\n");
1704                 goto out;
1705         }
1706
1707         do {
1708                 r = idr_pre_get(&cli->tmem_pools, GFP_ATOMIC);
1709                 if (r != 1) {
1710                         kfree(pool);
1711                         pr_info("zcache: pool creation failed: out of memory\n");
1712                         goto out;
1713                 }
1714                 r = idr_get_new(&cli->tmem_pools, pool, &poolid);
1715         } while (r == -EAGAIN);
1716         if (r) {
1717                 pr_info("zcache: pool creation failed: error %d\n", r);
1718                 kfree(pool);
1719                 goto out;
1720         }
1721
1722         atomic_set(&pool->refcount, 0);
1723         pool->client = cli;
1724         pool->pool_id = poolid;
1725         tmem_new_pool(pool, flags);
1726         pr_info("zcache: created %s tmem pool, id=%d, client=%d\n",
1727                 flags & TMEM_POOL_PERSIST ? "persistent" : "ephemeral",
1728                 poolid, cli_id);
1729 out:
1730         if (cli != NULL)
1731                 atomic_dec(&cli->refcount);
1732         return poolid;
1733 }
1734
1735 /**********
1736  * Two kernel functionalities currently can be layered on top of tmem.
1737  * These are "cleancache" which is used as a second-chance cache for clean
1738  * page cache pages; and "frontswap" which is used for swap pages
1739  * to avoid writes to disk.  A generic "shim" is provided here for each
1740  * to translate in-kernel semantics to zcache semantics.
1741  */
1742
1743 #ifdef CONFIG_CLEANCACHE
1744 static void zcache_cleancache_put_page(int pool_id,
1745                                         struct cleancache_filekey key,
1746                                         pgoff_t index, struct page *page)
1747 {
1748         u32 ind = (u32) index;
1749         struct tmem_oid oid = *(struct tmem_oid *)&key;
1750
1751         if (likely(ind == index))
1752                 (void)zcache_put_page(LOCAL_CLIENT, pool_id, &oid, index, page);
1753 }
1754
1755 static int zcache_cleancache_get_page(int pool_id,
1756                                         struct cleancache_filekey key,
1757                                         pgoff_t index, struct page *page)
1758 {
1759         u32 ind = (u32) index;
1760         struct tmem_oid oid = *(struct tmem_oid *)&key;
1761         int ret = -1;
1762
1763         if (likely(ind == index))
1764                 ret = zcache_get_page(LOCAL_CLIENT, pool_id, &oid, index, page);
1765         return ret;
1766 }
1767
1768 static void zcache_cleancache_flush_page(int pool_id,
1769                                         struct cleancache_filekey key,
1770                                         pgoff_t index)
1771 {
1772         u32 ind = (u32) index;
1773         struct tmem_oid oid = *(struct tmem_oid *)&key;
1774
1775         if (likely(ind == index))
1776                 (void)zcache_flush_page(LOCAL_CLIENT, pool_id, &oid, ind);
1777 }
1778
1779 static void zcache_cleancache_flush_inode(int pool_id,
1780                                         struct cleancache_filekey key)
1781 {
1782         struct tmem_oid oid = *(struct tmem_oid *)&key;
1783
1784         (void)zcache_flush_object(LOCAL_CLIENT, pool_id, &oid);
1785 }
1786
1787 static void zcache_cleancache_flush_fs(int pool_id)
1788 {
1789         if (pool_id >= 0)
1790                 (void)zcache_destroy_pool(LOCAL_CLIENT, pool_id);
1791 }
1792
1793 static int zcache_cleancache_init_fs(size_t pagesize)
1794 {
1795         BUG_ON(sizeof(struct cleancache_filekey) !=
1796                                 sizeof(struct tmem_oid));
1797         BUG_ON(pagesize != PAGE_SIZE);
1798         return zcache_new_pool(LOCAL_CLIENT, 0);
1799 }
1800
1801 static int zcache_cleancache_init_shared_fs(char *uuid, size_t pagesize)
1802 {
1803         /* shared pools are unsupported and map to private */
1804         BUG_ON(sizeof(struct cleancache_filekey) !=
1805                                 sizeof(struct tmem_oid));
1806         BUG_ON(pagesize != PAGE_SIZE);
1807         return zcache_new_pool(LOCAL_CLIENT, 0);
1808 }
1809
1810 static struct cleancache_ops zcache_cleancache_ops = {
1811         .put_page = zcache_cleancache_put_page,
1812         .get_page = zcache_cleancache_get_page,
1813         .invalidate_page = zcache_cleancache_flush_page,
1814         .invalidate_inode = zcache_cleancache_flush_inode,
1815         .invalidate_fs = zcache_cleancache_flush_fs,
1816         .init_shared_fs = zcache_cleancache_init_shared_fs,
1817         .init_fs = zcache_cleancache_init_fs
1818 };
1819
1820 struct cleancache_ops zcache_cleancache_register_ops(void)
1821 {
1822         struct cleancache_ops old_ops =
1823                 cleancache_register_ops(&zcache_cleancache_ops);
1824
1825         return old_ops;
1826 }
1827 #endif
1828
1829 #ifdef CONFIG_FRONTSWAP
1830 /* a single tmem poolid is used for all frontswap "types" (swapfiles) */
1831 static int zcache_frontswap_poolid = -1;
1832
1833 /*
1834  * Swizzling increases objects per swaptype, increasing tmem concurrency
1835  * for heavy swaploads.  Later, larger nr_cpus -> larger SWIZ_BITS
1836  * Setting SWIZ_BITS to 27 basically reconstructs the swap entry from
1837  * frontswap_load(), but has side-effects. Hence using 8.
1838  */
1839 #define SWIZ_BITS               8
1840 #define SWIZ_MASK               ((1 << SWIZ_BITS) - 1)
1841 #define _oswiz(_type, _ind)     ((_type << SWIZ_BITS) | (_ind & SWIZ_MASK))
1842 #define iswiz(_ind)             (_ind >> SWIZ_BITS)
1843
1844 static inline struct tmem_oid oswiz(unsigned type, u32 ind)
1845 {
1846         struct tmem_oid oid = { .oid = { 0 } };
1847         oid.oid[0] = _oswiz(type, ind);
1848         return oid;
1849 }
1850
1851 static int zcache_frontswap_store(unsigned type, pgoff_t offset,
1852                                    struct page *page)
1853 {
1854         u64 ind64 = (u64)offset;
1855         u32 ind = (u32)offset;
1856         struct tmem_oid oid = oswiz(type, ind);
1857         int ret = -1;
1858         unsigned long flags;
1859
1860         BUG_ON(!PageLocked(page));
1861         if (likely(ind64 == ind)) {
1862                 local_irq_save(flags);
1863                 ret = zcache_put_page(LOCAL_CLIENT, zcache_frontswap_poolid,
1864                                         &oid, iswiz(ind), page);
1865                 local_irq_restore(flags);
1866         }
1867         return ret;
1868 }
1869
1870 /* returns 0 if the page was successfully gotten from frontswap, -1 if
1871  * was not present (should never happen!) */
1872 static int zcache_frontswap_load(unsigned type, pgoff_t offset,
1873                                    struct page *page)
1874 {
1875         u64 ind64 = (u64)offset;
1876         u32 ind = (u32)offset;
1877         struct tmem_oid oid = oswiz(type, ind);
1878         int ret = -1;
1879
1880         BUG_ON(!PageLocked(page));
1881         if (likely(ind64 == ind))
1882                 ret = zcache_get_page(LOCAL_CLIENT, zcache_frontswap_poolid,
1883                                         &oid, iswiz(ind), page);
1884         return ret;
1885 }
1886
1887 /* flush a single page from frontswap */
1888 static void zcache_frontswap_flush_page(unsigned type, pgoff_t offset)
1889 {
1890         u64 ind64 = (u64)offset;
1891         u32 ind = (u32)offset;
1892         struct tmem_oid oid = oswiz(type, ind);
1893
1894         if (likely(ind64 == ind))
1895                 (void)zcache_flush_page(LOCAL_CLIENT, zcache_frontswap_poolid,
1896                                         &oid, iswiz(ind));
1897 }
1898
1899 /* flush all pages from the passed swaptype */
1900 static void zcache_frontswap_flush_area(unsigned type)
1901 {
1902         struct tmem_oid oid;
1903         int ind;
1904
1905         for (ind = SWIZ_MASK; ind >= 0; ind--) {
1906                 oid = oswiz(type, ind);
1907                 (void)zcache_flush_object(LOCAL_CLIENT,
1908                                                 zcache_frontswap_poolid, &oid);
1909         }
1910 }
1911
1912 static void zcache_frontswap_init(unsigned ignored)
1913 {
1914         /* a single tmem poolid is used for all frontswap "types" (swapfiles) */
1915         if (zcache_frontswap_poolid < 0)
1916                 zcache_frontswap_poolid =
1917                         zcache_new_pool(LOCAL_CLIENT, TMEM_POOL_PERSIST);
1918 }
1919
1920 static struct frontswap_ops zcache_frontswap_ops = {
1921         .store = zcache_frontswap_store,
1922         .load = zcache_frontswap_load,
1923         .invalidate_page = zcache_frontswap_flush_page,
1924         .invalidate_area = zcache_frontswap_flush_area,
1925         .init = zcache_frontswap_init
1926 };
1927
1928 struct frontswap_ops zcache_frontswap_register_ops(void)
1929 {
1930         struct frontswap_ops old_ops =
1931                 frontswap_register_ops(&zcache_frontswap_ops);
1932
1933         return old_ops;
1934 }
1935 #endif
1936
1937 /*
1938  * zcache initialization
1939  * NOTE FOR NOW zcache MUST BE PROVIDED AS A KERNEL BOOT PARAMETER OR
1940  * NOTHING HAPPENS!
1941  */
1942
1943 static int zcache_enabled;
1944
1945 static int __init enable_zcache(char *s)
1946 {
1947         zcache_enabled = 1;
1948         return 1;
1949 }
1950 __setup("zcache", enable_zcache);
1951
1952 /* allow independent dynamic disabling of cleancache and frontswap */
1953
1954 static int use_cleancache = 1;
1955
1956 static int __init no_cleancache(char *s)
1957 {
1958         use_cleancache = 0;
1959         return 1;
1960 }
1961
1962 __setup("nocleancache", no_cleancache);
1963
1964 static int use_frontswap = 1;
1965
1966 static int __init no_frontswap(char *s)
1967 {
1968         use_frontswap = 0;
1969         return 1;
1970 }
1971
1972 __setup("nofrontswap", no_frontswap);
1973
1974 static int __init enable_zcache_compressor(char *s)
1975 {
1976         strncpy(zcache_comp_name, s, ZCACHE_COMP_NAME_SZ);
1977         zcache_enabled = 1;
1978         return 1;
1979 }
1980 __setup("zcache=", enable_zcache_compressor);
1981
1982
1983 static int __init zcache_comp_init(void)
1984 {
1985         int ret = 0;
1986
1987         /* check crypto algorithm */
1988         if (*zcache_comp_name != '\0') {
1989                 ret = crypto_has_comp(zcache_comp_name, 0, 0);
1990                 if (!ret)
1991                         pr_info("zcache: %s not supported\n",
1992                                         zcache_comp_name);
1993         }
1994         if (!ret)
1995                 strcpy(zcache_comp_name, "lzo");
1996         ret = crypto_has_comp(zcache_comp_name, 0, 0);
1997         if (!ret) {
1998                 ret = 1;
1999                 goto out;
2000         }
2001         pr_info("zcache: using %s compressor\n", zcache_comp_name);
2002
2003         /* alloc percpu transforms */
2004         ret = 0;
2005         zcache_comp_pcpu_tfms = alloc_percpu(struct crypto_comp *);
2006         if (!zcache_comp_pcpu_tfms)
2007                 ret = 1;
2008 out:
2009         return ret;
2010 }
2011
2012 static int __init zcache_init(void)
2013 {
2014         int ret = 0;
2015
2016 #ifdef CONFIG_SYSFS
2017         ret = sysfs_create_group(mm_kobj, &zcache_attr_group);
2018         if (ret) {
2019                 pr_err("zcache: can't create sysfs\n");
2020                 goto out;
2021         }
2022 #endif /* CONFIG_SYSFS */
2023
2024         if (zcache_enabled) {
2025                 unsigned int cpu;
2026
2027                 tmem_register_hostops(&zcache_hostops);
2028                 tmem_register_pamops(&zcache_pamops);
2029                 ret = register_cpu_notifier(&zcache_cpu_notifier_block);
2030                 if (ret) {
2031                         pr_err("zcache: can't register cpu notifier\n");
2032                         goto out;
2033                 }
2034                 ret = zcache_comp_init();
2035                 if (ret) {
2036                         pr_err("zcache: compressor initialization failed\n");
2037                         goto out;
2038                 }
2039                 for_each_online_cpu(cpu) {
2040                         void *pcpu = (void *)(long)cpu;
2041                         zcache_cpu_notifier(&zcache_cpu_notifier_block,
2042                                 CPU_UP_PREPARE, pcpu);
2043                 }
2044         }
2045         zcache_objnode_cache = kmem_cache_create("zcache_objnode",
2046                                 sizeof(struct tmem_objnode), 0, 0, NULL);
2047         zcache_obj_cache = kmem_cache_create("zcache_obj",
2048                                 sizeof(struct tmem_obj), 0, 0, NULL);
2049         ret = zcache_new_client(LOCAL_CLIENT);
2050         if (ret) {
2051                 pr_err("zcache: can't create client\n");
2052                 goto out;
2053         }
2054
2055 #ifdef CONFIG_CLEANCACHE
2056         if (zcache_enabled && use_cleancache) {
2057                 struct cleancache_ops old_ops;
2058
2059                 zbud_init();
2060                 register_shrinker(&zcache_shrinker);
2061                 old_ops = zcache_cleancache_register_ops();
2062                 pr_info("zcache: cleancache enabled using kernel "
2063                         "transcendent memory and compression buddies\n");
2064                 if (old_ops.init_fs != NULL)
2065                         pr_warning("zcache: cleancache_ops overridden");
2066         }
2067 #endif
2068 #ifdef CONFIG_FRONTSWAP
2069         if (zcache_enabled && use_frontswap) {
2070                 struct frontswap_ops old_ops;
2071
2072                 old_ops = zcache_frontswap_register_ops();
2073                 pr_info("zcache: frontswap enabled using kernel "
2074                         "transcendent memory and zsmalloc\n");
2075                 if (old_ops.init != NULL)
2076                         pr_warning("zcache: frontswap_ops overridden");
2077         }
2078 #endif
2079 out:
2080         return ret;
2081 }
2082
2083 module_init(zcache_init)