Merge tag 'sunxi-dt-for-3.11-2' of git://github.com/mripard/linux into next/dt
[sfrench/cifs-2.6.git] / drivers / staging / zcache / zcache-main.c
1 /*
2  * zcache.c
3  *
4  * Copyright (c) 2010-2012, Dan Magenheimer, Oracle Corp.
5  * Copyright (c) 2010,2011, Nitin Gupta
6  *
7  * Zcache provides an in-kernel "host implementation" for transcendent memory
8  * ("tmem") and, thus indirectly, for cleancache and frontswap.  Zcache uses
9  * lzo1x compression to improve density and an embedded allocator called
10  * "zbud" which "buddies" two compressed pages semi-optimally in each physical
11  * pageframe.  Zbud is integrally tied into tmem to allow pageframes to
12  * be "reclaimed" efficiently.
13  */
14
15 #include <linux/module.h>
16 #include <linux/cpu.h>
17 #include <linux/highmem.h>
18 #include <linux/list.h>
19 #include <linux/slab.h>
20 #include <linux/spinlock.h>
21 #include <linux/types.h>
22 #include <linux/string.h>
23 #include <linux/atomic.h>
24 #include <linux/math64.h>
25 #include <linux/crypto.h>
26 #include <linux/swap.h>
27 #include <linux/swapops.h>
28 #include <linux/pagemap.h>
29 #include <linux/writeback.h>
30
31 #include <linux/cleancache.h>
32 #include <linux/frontswap.h>
33 #include "tmem.h"
34 #include "zcache.h"
35 #include "zbud.h"
36 #include "ramster.h"
37 #include "debug.h"
38 #ifdef CONFIG_RAMSTER
39 static bool ramster_enabled __read_mostly;
40 static int disable_frontswap_selfshrink;
41 #else
42 #define ramster_enabled false
43 #define disable_frontswap_selfshrink 0
44 #endif
45
46 #ifndef __PG_WAS_ACTIVE
47 static inline bool PageWasActive(struct page *page)
48 {
49         return true;
50 }
51
52 static inline void SetPageWasActive(struct page *page)
53 {
54 }
55 #endif
56
57 #ifdef FRONTSWAP_HAS_EXCLUSIVE_GETS
58 static bool frontswap_has_exclusive_gets __read_mostly = true;
59 #else
60 static bool frontswap_has_exclusive_gets __read_mostly;
61 static inline void frontswap_tmem_exclusive_gets(bool b)
62 {
63 }
64 #endif
65
66 /*
67  * mark pampd to special value in order that later
68  * retrieve will identify zero-filled pages
69  */
70 #define ZERO_FILLED 0x2
71
72 /* enable (or fix code) when Seth's patches are accepted upstream */
73 #define zcache_writeback_enabled 0
74
75 static bool zcache_enabled __read_mostly;
76 static bool disable_cleancache __read_mostly;
77 static bool disable_frontswap __read_mostly;
78 static bool disable_frontswap_ignore_nonactive __read_mostly;
79 static bool disable_cleancache_ignore_nonactive __read_mostly;
80 static char *namestr __read_mostly = "zcache";
81
82 #define ZCACHE_GFP_MASK \
83         (__GFP_FS | __GFP_NORETRY | __GFP_NOWARN | __GFP_NOMEMALLOC)
84
85 /* crypto API for zcache  */
86 #ifdef CONFIG_ZCACHE_MODULE
87 static char *zcache_comp_name = "lzo";
88 #else
89 #define ZCACHE_COMP_NAME_SZ CRYPTO_MAX_ALG_NAME
90 static char zcache_comp_name[ZCACHE_COMP_NAME_SZ] __read_mostly;
91 #endif
92 static struct crypto_comp * __percpu *zcache_comp_pcpu_tfms __read_mostly;
93
94 enum comp_op {
95         ZCACHE_COMPOP_COMPRESS,
96         ZCACHE_COMPOP_DECOMPRESS
97 };
98
99 static inline int zcache_comp_op(enum comp_op op,
100                                 const u8 *src, unsigned int slen,
101                                 u8 *dst, unsigned int *dlen)
102 {
103         struct crypto_comp *tfm;
104         int ret = -1;
105
106         BUG_ON(!zcache_comp_pcpu_tfms);
107         tfm = *per_cpu_ptr(zcache_comp_pcpu_tfms, get_cpu());
108         BUG_ON(!tfm);
109         switch (op) {
110         case ZCACHE_COMPOP_COMPRESS:
111                 ret = crypto_comp_compress(tfm, src, slen, dst, dlen);
112                 break;
113         case ZCACHE_COMPOP_DECOMPRESS:
114                 ret = crypto_comp_decompress(tfm, src, slen, dst, dlen);
115                 break;
116         default:
117                 ret = -EINVAL;
118         }
119         put_cpu();
120         return ret;
121 }
122
123 /*
124  * policy parameters
125  */
126
127 /*
128  * byte count defining poor compression; pages with greater zsize will be
129  * rejected
130  */
131 static unsigned int zbud_max_zsize __read_mostly = (PAGE_SIZE / 8) * 7;
132 /*
133  * byte count defining poor *mean* compression; pages with greater zsize
134  * will be rejected until sufficient better-compressed pages are accepted
135  * driving the mean below this threshold
136  */
137 static unsigned int zbud_max_mean_zsize __read_mostly = (PAGE_SIZE / 8) * 5;
138
139 /*
140  * for now, used named slabs so can easily track usage; later can
141  * either just use kmalloc, or perhaps add a slab-like allocator
142  * to more carefully manage total memory utilization
143  */
144 static struct kmem_cache *zcache_objnode_cache;
145 static struct kmem_cache *zcache_obj_cache;
146
147 static DEFINE_PER_CPU(struct zcache_preload, zcache_preloads) = { 0, };
148
149 /* Used by debug.c */
150 ssize_t zcache_pers_zpages;
151 u64 zcache_pers_zbytes;
152 ssize_t zcache_eph_pageframes;
153 ssize_t zcache_pers_pageframes;
154
155 /* Used by this code. */
156 ssize_t zcache_last_active_file_pageframes;
157 ssize_t zcache_last_inactive_file_pageframes;
158 ssize_t zcache_last_active_anon_pageframes;
159 ssize_t zcache_last_inactive_anon_pageframes;
160 #ifdef CONFIG_ZCACHE_WRITEBACK
161 ssize_t zcache_writtenback_pages;
162 ssize_t zcache_outstanding_writeback_pages;
163 #endif
164 /*
165  * zcache core code starts here
166  */
167
168 static struct zcache_client zcache_host;
169 static struct zcache_client zcache_clients[MAX_CLIENTS];
170
171 static inline bool is_local_client(struct zcache_client *cli)
172 {
173         return cli == &zcache_host;
174 }
175
176 static struct zcache_client *zcache_get_client_by_id(uint16_t cli_id)
177 {
178         struct zcache_client *cli = &zcache_host;
179
180         if (cli_id != LOCAL_CLIENT) {
181                 if (cli_id >= MAX_CLIENTS)
182                         goto out;
183                 cli = &zcache_clients[cli_id];
184         }
185 out:
186         return cli;
187 }
188
189 /*
190  * Tmem operations assume the poolid implies the invoking client.
191  * Zcache only has one client (the kernel itself): LOCAL_CLIENT.
192  * RAMster has each client numbered by cluster node, and a KVM version
193  * of zcache would have one client per guest and each client might
194  * have a poolid==N.
195  */
196 struct tmem_pool *zcache_get_pool_by_id(uint16_t cli_id, uint16_t poolid)
197 {
198         struct tmem_pool *pool = NULL;
199         struct zcache_client *cli = NULL;
200
201         cli = zcache_get_client_by_id(cli_id);
202         if (cli == NULL)
203                 goto out;
204         if (!is_local_client(cli))
205                 atomic_inc(&cli->refcount);
206         if (poolid < MAX_POOLS_PER_CLIENT) {
207                 pool = cli->tmem_pools[poolid];
208                 if (pool != NULL)
209                         atomic_inc(&pool->refcount);
210         }
211 out:
212         return pool;
213 }
214
215 void zcache_put_pool(struct tmem_pool *pool)
216 {
217         struct zcache_client *cli = NULL;
218
219         if (pool == NULL)
220                 BUG();
221         cli = pool->client;
222         atomic_dec(&pool->refcount);
223         if (!is_local_client(cli))
224                 atomic_dec(&cli->refcount);
225 }
226
227 int zcache_new_client(uint16_t cli_id)
228 {
229         struct zcache_client *cli;
230         int ret = -1;
231
232         cli = zcache_get_client_by_id(cli_id);
233         if (cli == NULL)
234                 goto out;
235         if (cli->allocated)
236                 goto out;
237         cli->allocated = 1;
238         ret = 0;
239 out:
240         return ret;
241 }
242
243 /*
244  * zcache implementation for tmem host ops
245  */
246
247 static struct tmem_objnode *zcache_objnode_alloc(struct tmem_pool *pool)
248 {
249         struct tmem_objnode *objnode = NULL;
250         struct zcache_preload *kp;
251         int i;
252
253         kp = &__get_cpu_var(zcache_preloads);
254         for (i = 0; i < ARRAY_SIZE(kp->objnodes); i++) {
255                 objnode = kp->objnodes[i];
256                 if (objnode != NULL) {
257                         kp->objnodes[i] = NULL;
258                         break;
259                 }
260         }
261         BUG_ON(objnode == NULL);
262         inc_zcache_objnode_count();
263         return objnode;
264 }
265
266 static void zcache_objnode_free(struct tmem_objnode *objnode,
267                                         struct tmem_pool *pool)
268 {
269         dec_zcache_objnode_count();
270         kmem_cache_free(zcache_objnode_cache, objnode);
271 }
272
273 static struct tmem_obj *zcache_obj_alloc(struct tmem_pool *pool)
274 {
275         struct tmem_obj *obj = NULL;
276         struct zcache_preload *kp;
277
278         kp = &__get_cpu_var(zcache_preloads);
279         obj = kp->obj;
280         BUG_ON(obj == NULL);
281         kp->obj = NULL;
282         inc_zcache_obj_count();
283         return obj;
284 }
285
286 static void zcache_obj_free(struct tmem_obj *obj, struct tmem_pool *pool)
287 {
288         dec_zcache_obj_count();
289         kmem_cache_free(zcache_obj_cache, obj);
290 }
291
292 /*
293  * Compressing zero-filled pages will waste memory and introduce
294  * serious fragmentation, skip it to avoid overhead.
295  */
296 static bool page_is_zero_filled(struct page *p)
297 {
298         unsigned int pos;
299         char *page;
300
301         page = kmap_atomic(p);
302         for (pos = 0; pos < PAGE_SIZE / sizeof(*page); pos++) {
303                 if (page[pos]) {
304                         kunmap_atomic(page);
305                         return false;
306                 }
307         }
308         kunmap_atomic(page);
309
310         return true;
311 }
312
313 static void handle_zero_filled_page(void *p)
314 {
315         void *user_mem;
316         struct page *page = (struct page *)p;
317
318         user_mem = kmap_atomic(page);
319         memset(user_mem, 0, PAGE_SIZE);
320         kunmap_atomic(user_mem);
321
322         flush_dcache_page(page);
323 }
324
325 static struct tmem_hostops zcache_hostops = {
326         .obj_alloc = zcache_obj_alloc,
327         .obj_free = zcache_obj_free,
328         .objnode_alloc = zcache_objnode_alloc,
329         .objnode_free = zcache_objnode_free,
330 };
331
332 static struct page *zcache_alloc_page(void)
333 {
334         struct page *page = alloc_page(ZCACHE_GFP_MASK);
335
336         if (page != NULL)
337                 inc_zcache_pageframes_alloced();
338         return page;
339 }
340
341 static void zcache_free_page(struct page *page)
342 {
343         long curr_pageframes;
344         static long max_pageframes, min_pageframes;
345
346         if (page == NULL)
347                 BUG();
348         __free_page(page);
349         inc_zcache_pageframes_freed();
350         curr_pageframes = curr_pageframes_count();
351         if (curr_pageframes > max_pageframes)
352                 max_pageframes = curr_pageframes;
353         if (curr_pageframes < min_pageframes)
354                 min_pageframes = curr_pageframes;
355 #ifdef CONFIG_ZCACHE_DEBUG
356         if (curr_pageframes > 2L || curr_pageframes < -2L) {
357                 /* pr_info here */
358         }
359 #endif
360 }
361
362 /*
363  * zcache implementations for PAM page descriptor ops
364  */
365
366 /* forward reference */
367 static void zcache_compress(struct page *from,
368                                 void **out_va, unsigned *out_len);
369
370 static struct page *zcache_evict_eph_pageframe(void);
371
372 static void *zcache_pampd_eph_create(char *data, size_t size, bool raw,
373                                         struct tmem_handle *th)
374 {
375         void *pampd = NULL, *cdata = data;
376         unsigned clen = size;
377         bool zero_filled = false;
378         struct page *page = (struct page *)(data), *newpage;
379
380         if (page_is_zero_filled(page)) {
381                 clen = 0;
382                 zero_filled = true;
383                 inc_zcache_zero_filled_pages();
384                 goto got_pampd;
385         }
386
387         if (!raw) {
388                 zcache_compress(page, &cdata, &clen);
389                 if (clen > zbud_max_buddy_size()) {
390                         inc_zcache_compress_poor();
391                         goto out;
392                 }
393         } else {
394                 BUG_ON(clen > zbud_max_buddy_size());
395         }
396
397         /* look for space via an existing match first */
398         pampd = (void *)zbud_match_prep(th, true, cdata, clen);
399         if (pampd != NULL)
400                 goto got_pampd;
401
402         /* no match, now we need to find (or free up) a full page */
403         newpage = zcache_alloc_page();
404         if (newpage != NULL)
405                 goto create_in_new_page;
406
407         inc_zcache_failed_getfreepages();
408         /* can't allocate a page, evict an ephemeral page via LRU */
409         newpage = zcache_evict_eph_pageframe();
410         if (newpage == NULL) {
411                 inc_zcache_eph_ate_tail_failed();
412                 goto out;
413         }
414         inc_zcache_eph_ate_tail();
415
416 create_in_new_page:
417         pampd = (void *)zbud_create_prep(th, true, cdata, clen, newpage);
418         BUG_ON(pampd == NULL);
419         inc_zcache_eph_pageframes();
420
421 got_pampd:
422         inc_zcache_eph_zbytes(clen);
423         inc_zcache_eph_zpages();
424         if (ramster_enabled && raw && !zero_filled)
425                 ramster_count_foreign_pages(true, 1);
426         if (zero_filled)
427                 pampd = (void *)ZERO_FILLED;
428 out:
429         return pampd;
430 }
431
432 static void *zcache_pampd_pers_create(char *data, size_t size, bool raw,
433                                         struct tmem_handle *th)
434 {
435         void *pampd = NULL, *cdata = data;
436         unsigned clen = size;
437         bool zero_filled = false;
438         struct page *page = (struct page *)(data), *newpage;
439         unsigned long zbud_mean_zsize;
440         unsigned long curr_pers_zpages, total_zsize;
441
442         if (data == NULL) {
443                 BUG_ON(!ramster_enabled);
444                 goto create_pampd;
445         }
446
447         if (page_is_zero_filled(page)) {
448                 clen = 0;
449                 zero_filled = true;
450                 inc_zcache_zero_filled_pages();
451                 goto got_pampd;
452         }
453
454         curr_pers_zpages = zcache_pers_zpages;
455 /* FIXME CONFIG_RAMSTER... subtract atomic remote_pers_pages here? */
456         if (!raw)
457                 zcache_compress(page, &cdata, &clen);
458         /* reject if compression is too poor */
459         if (clen > zbud_max_zsize) {
460                 inc_zcache_compress_poor();
461                 goto out;
462         }
463         /* reject if mean compression is too poor */
464         if ((clen > zbud_max_mean_zsize) && (curr_pers_zpages > 0)) {
465                 total_zsize = zcache_pers_zbytes;
466                 if ((long)total_zsize < 0)
467                         total_zsize = 0;
468                 zbud_mean_zsize = div_u64(total_zsize,
469                                         curr_pers_zpages);
470                 if (zbud_mean_zsize > zbud_max_mean_zsize) {
471                         inc_zcache_mean_compress_poor();
472                         goto out;
473                 }
474         }
475
476 create_pampd:
477         /* look for space via an existing match first */
478         pampd = (void *)zbud_match_prep(th, false, cdata, clen);
479         if (pampd != NULL)
480                 goto got_pampd;
481
482         /* no match, now we need to find (or free up) a full page */
483         newpage = zcache_alloc_page();
484         if (newpage != NULL)
485                 goto create_in_new_page;
486         /*
487          * FIXME do the following only if eph is oversized?
488          * if (zcache_eph_pageframes >
489          * (global_page_state(NR_LRU_BASE + LRU_ACTIVE_FILE) +
490          * global_page_state(NR_LRU_BASE + LRU_INACTIVE_FILE)))
491          */
492         inc_zcache_failed_getfreepages();
493         /* can't allocate a page, evict an ephemeral page via LRU */
494         newpage = zcache_evict_eph_pageframe();
495         if (newpage == NULL) {
496                 inc_zcache_pers_ate_eph_failed();
497                 goto out;
498         }
499         inc_zcache_pers_ate_eph();
500
501 create_in_new_page:
502         pampd = (void *)zbud_create_prep(th, false, cdata, clen, newpage);
503         BUG_ON(pampd == NULL);
504         inc_zcache_pers_pageframes();
505
506 got_pampd:
507         inc_zcache_pers_zpages();
508         inc_zcache_pers_zbytes(clen);
509         if (ramster_enabled && raw && !zero_filled)
510                 ramster_count_foreign_pages(false, 1);
511         if (zero_filled)
512                 pampd = (void *)ZERO_FILLED;
513 out:
514         return pampd;
515 }
516
517 /*
518  * This is called directly from zcache_put_page to pre-allocate space
519  * to store a zpage.
520  */
521 void *zcache_pampd_create(char *data, unsigned int size, bool raw,
522                                         int eph, struct tmem_handle *th)
523 {
524         void *pampd = NULL;
525         struct zcache_preload *kp;
526         struct tmem_objnode *objnode;
527         struct tmem_obj *obj;
528         int i;
529
530         BUG_ON(!irqs_disabled());
531         /* pre-allocate per-cpu metadata */
532         BUG_ON(zcache_objnode_cache == NULL);
533         BUG_ON(zcache_obj_cache == NULL);
534         kp = &__get_cpu_var(zcache_preloads);
535         for (i = 0; i < ARRAY_SIZE(kp->objnodes); i++) {
536                 objnode = kp->objnodes[i];
537                 if (objnode == NULL) {
538                         objnode = kmem_cache_alloc(zcache_objnode_cache,
539                                                         ZCACHE_GFP_MASK);
540                         if (unlikely(objnode == NULL)) {
541                                 inc_zcache_failed_alloc();
542                                 goto out;
543                         }
544                         kp->objnodes[i] = objnode;
545                 }
546         }
547         if (kp->obj == NULL) {
548                 obj = kmem_cache_alloc(zcache_obj_cache, ZCACHE_GFP_MASK);
549                 kp->obj = obj;
550         }
551         if (unlikely(kp->obj == NULL)) {
552                 inc_zcache_failed_alloc();
553                 goto out;
554         }
555         /*
556          * ok, have all the metadata pre-allocated, now do the data
557          * but since how we allocate the data is dependent on ephemeral
558          * or persistent, we split the call here to different sub-functions
559          */
560         if (eph)
561                 pampd = zcache_pampd_eph_create(data, size, raw, th);
562         else
563                 pampd = zcache_pampd_pers_create(data, size, raw, th);
564 out:
565         return pampd;
566 }
567
568 /*
569  * This is a pamops called via tmem_put and is necessary to "finish"
570  * a pampd creation.
571  */
572 void zcache_pampd_create_finish(void *pampd, bool eph)
573 {
574         if (pampd != (void *)ZERO_FILLED)
575                 zbud_create_finish((struct zbudref *)pampd, eph);
576 }
577
578 /*
579  * This is passed as a function parameter to zbud_decompress so that
580  * zbud need not be familiar with the details of crypto. It assumes that
581  * the bytes from_va and to_va through from_va+size-1 and to_va+size-1 are
582  * kmapped.  It must be successful, else there is a logic bug somewhere.
583  */
584 static void zcache_decompress(char *from_va, unsigned int size, char *to_va)
585 {
586         int ret;
587         unsigned int outlen = PAGE_SIZE;
588
589         ret = zcache_comp_op(ZCACHE_COMPOP_DECOMPRESS, from_va, size,
590                                 to_va, &outlen);
591         BUG_ON(ret);
592         BUG_ON(outlen != PAGE_SIZE);
593 }
594
595 /*
596  * Decompress from the kernel va to a pageframe
597  */
598 void zcache_decompress_to_page(char *from_va, unsigned int size,
599                                         struct page *to_page)
600 {
601         char *to_va = kmap_atomic(to_page);
602         zcache_decompress(from_va, size, to_va);
603         kunmap_atomic(to_va);
604 }
605
606 /*
607  * fill the pageframe corresponding to the struct page with the data
608  * from the passed pampd
609  */
610 static int zcache_pampd_get_data(char *data, size_t *sizep, bool raw,
611                                         void *pampd, struct tmem_pool *pool,
612                                         struct tmem_oid *oid, uint32_t index)
613 {
614         int ret;
615         bool eph = !is_persistent(pool);
616
617         BUG_ON(preemptible());
618         BUG_ON(eph);    /* fix later if shared pools get implemented */
619         BUG_ON(pampd_is_remote(pampd));
620
621         if (pampd == (void *)ZERO_FILLED) {
622                 handle_zero_filled_page(data);
623                 if (!raw)
624                         *sizep = PAGE_SIZE;
625                 return 0;
626         }
627
628         if (raw)
629                 ret = zbud_copy_from_zbud(data, (struct zbudref *)pampd,
630                                                 sizep, eph);
631         else {
632                 ret = zbud_decompress((struct page *)(data),
633                                         (struct zbudref *)pampd, false,
634                                         zcache_decompress);
635                 *sizep = PAGE_SIZE;
636         }
637         return ret;
638 }
639
640 /*
641  * fill the pageframe corresponding to the struct page with the data
642  * from the passed pampd
643  */
644 static int zcache_pampd_get_data_and_free(char *data, size_t *sizep, bool raw,
645                                         void *pampd, struct tmem_pool *pool,
646                                         struct tmem_oid *oid, uint32_t index)
647 {
648         int ret = 0;
649         bool eph = !is_persistent(pool), zero_filled = false;
650         struct page *page = NULL;
651         unsigned int zsize, zpages;
652
653         BUG_ON(preemptible());
654         BUG_ON(pampd_is_remote(pampd));
655
656         if (pampd == (void *)ZERO_FILLED) {
657                 handle_zero_filled_page(data);
658                 zero_filled = true;
659                 zsize = 0;
660                 zpages = 1;
661                 if (!raw)
662                         *sizep = PAGE_SIZE;
663                 dec_zcache_zero_filled_pages();
664                 goto zero_fill;
665         }
666
667         if (raw)
668                 ret = zbud_copy_from_zbud(data, (struct zbudref *)pampd,
669                                                 sizep, eph);
670         else {
671                 ret = zbud_decompress((struct page *)(data),
672                                         (struct zbudref *)pampd, eph,
673                                         zcache_decompress);
674                 *sizep = PAGE_SIZE;
675         }
676         page = zbud_free_and_delist((struct zbudref *)pampd, eph,
677                                         &zsize, &zpages);
678 zero_fill:
679         if (eph) {
680                 if (page)
681                         dec_zcache_eph_pageframes();
682                 dec_zcache_eph_zpages(zpages);
683                 dec_zcache_eph_zbytes(zsize);
684         } else {
685                 if (page)
686                         dec_zcache_pers_pageframes();
687                 dec_zcache_pers_zpages(zpages);
688                 dec_zcache_pers_zbytes(zsize);
689         }
690         if (!is_local_client(pool->client) && !zero_filled)
691                 ramster_count_foreign_pages(eph, -1);
692         if (page && !zero_filled)
693                 zcache_free_page(page);
694         return ret;
695 }
696
697 /*
698  * free the pampd and remove it from any zcache lists
699  * pampd must no longer be pointed to from any tmem data structures!
700  */
701 static void zcache_pampd_free(void *pampd, struct tmem_pool *pool,
702                               struct tmem_oid *oid, uint32_t index, bool acct)
703 {
704         struct page *page = NULL;
705         unsigned int zsize, zpages;
706         bool zero_filled = false;
707
708         BUG_ON(preemptible());
709
710         if (pampd == (void *)ZERO_FILLED) {
711                 zero_filled = true;
712                 zsize = 0;
713                 zpages = 1;
714                 dec_zcache_zero_filled_pages();
715         }
716
717         if (pampd_is_remote(pampd) && !zero_filled) {
718                 BUG_ON(!ramster_enabled);
719                 pampd = ramster_pampd_free(pampd, pool, oid, index, acct);
720                 if (pampd == NULL)
721                         return;
722         }
723         if (is_ephemeral(pool)) {
724                 if (!zero_filled)
725                         page = zbud_free_and_delist((struct zbudref *)pampd,
726                                                 true, &zsize, &zpages);
727                 if (page)
728                         dec_zcache_eph_pageframes();
729                 dec_zcache_eph_zpages(zpages);
730                 dec_zcache_eph_zbytes(zsize);
731                 /* FIXME CONFIG_RAMSTER... check acct parameter? */
732         } else {
733                 if (!zero_filled)
734                         page = zbud_free_and_delist((struct zbudref *)pampd,
735                                                 false, &zsize, &zpages);
736                 if (page)
737                         dec_zcache_pers_pageframes();
738                 dec_zcache_pers_zpages(zpages);
739                 dec_zcache_pers_zbytes(zsize);
740         }
741         if (!is_local_client(pool->client) && !zero_filled)
742                 ramster_count_foreign_pages(is_ephemeral(pool), -1);
743         if (page && !zero_filled)
744                 zcache_free_page(page);
745 }
746
747 static struct tmem_pamops zcache_pamops = {
748         .create_finish = zcache_pampd_create_finish,
749         .get_data = zcache_pampd_get_data,
750         .get_data_and_free = zcache_pampd_get_data_and_free,
751         .free = zcache_pampd_free,
752 };
753
754 /*
755  * zcache compression/decompression and related per-cpu stuff
756  */
757
758 static DEFINE_PER_CPU(unsigned char *, zcache_dstmem);
759 #define ZCACHE_DSTMEM_ORDER 1
760
761 static void zcache_compress(struct page *from, void **out_va, unsigned *out_len)
762 {
763         int ret;
764         unsigned char *dmem = __get_cpu_var(zcache_dstmem);
765         char *from_va;
766
767         BUG_ON(!irqs_disabled());
768         /* no buffer or no compressor so can't compress */
769         BUG_ON(dmem == NULL);
770         *out_len = PAGE_SIZE << ZCACHE_DSTMEM_ORDER;
771         from_va = kmap_atomic(from);
772         mb();
773         ret = zcache_comp_op(ZCACHE_COMPOP_COMPRESS, from_va, PAGE_SIZE, dmem,
774                                 out_len);
775         BUG_ON(ret);
776         *out_va = dmem;
777         kunmap_atomic(from_va);
778 }
779
780 static int zcache_comp_cpu_up(int cpu)
781 {
782         struct crypto_comp *tfm;
783
784         tfm = crypto_alloc_comp(zcache_comp_name, 0, 0);
785         if (IS_ERR(tfm))
786                 return NOTIFY_BAD;
787         *per_cpu_ptr(zcache_comp_pcpu_tfms, cpu) = tfm;
788         return NOTIFY_OK;
789 }
790
791 static void zcache_comp_cpu_down(int cpu)
792 {
793         struct crypto_comp *tfm;
794
795         tfm = *per_cpu_ptr(zcache_comp_pcpu_tfms, cpu);
796         crypto_free_comp(tfm);
797         *per_cpu_ptr(zcache_comp_pcpu_tfms, cpu) = NULL;
798 }
799
800 static int zcache_cpu_notifier(struct notifier_block *nb,
801                                 unsigned long action, void *pcpu)
802 {
803         int ret, i, cpu = (long)pcpu;
804         struct zcache_preload *kp;
805
806         switch (action) {
807         case CPU_UP_PREPARE:
808                 ret = zcache_comp_cpu_up(cpu);
809                 if (ret != NOTIFY_OK) {
810                         pr_err("%s: can't allocate compressor xform\n",
811                                 namestr);
812                         return ret;
813                 }
814                 per_cpu(zcache_dstmem, cpu) = (void *)__get_free_pages(
815                         GFP_KERNEL | __GFP_REPEAT, ZCACHE_DSTMEM_ORDER);
816                 if (ramster_enabled)
817                         ramster_cpu_up(cpu);
818                 break;
819         case CPU_DEAD:
820         case CPU_UP_CANCELED:
821                 zcache_comp_cpu_down(cpu);
822                 free_pages((unsigned long)per_cpu(zcache_dstmem, cpu),
823                         ZCACHE_DSTMEM_ORDER);
824                 per_cpu(zcache_dstmem, cpu) = NULL;
825                 kp = &per_cpu(zcache_preloads, cpu);
826                 for (i = 0; i < ARRAY_SIZE(kp->objnodes); i++) {
827                         if (kp->objnodes[i])
828                                 kmem_cache_free(zcache_objnode_cache,
829                                                 kp->objnodes[i]);
830                 }
831                 if (kp->obj) {
832                         kmem_cache_free(zcache_obj_cache, kp->obj);
833                         kp->obj = NULL;
834                 }
835                 if (ramster_enabled)
836                         ramster_cpu_down(cpu);
837                 break;
838         default:
839                 break;
840         }
841         return NOTIFY_OK;
842 }
843
844 static struct notifier_block zcache_cpu_notifier_block = {
845         .notifier_call = zcache_cpu_notifier
846 };
847
848 /*
849  * The following code interacts with the zbud eviction and zbud
850  * zombify code to access LRU pages
851  */
852
853 static struct page *zcache_evict_eph_pageframe(void)
854 {
855         struct page *page;
856         unsigned int zsize = 0, zpages = 0;
857
858         page = zbud_evict_pageframe_lru(&zsize, &zpages);
859         if (page == NULL)
860                 goto out;
861         dec_zcache_eph_zbytes(zsize);
862         dec_zcache_eph_zpages(zpages);
863         inc_zcache_evicted_eph_zpages(zpages);
864         dec_zcache_eph_pageframes();
865         inc_zcache_evicted_eph_pageframes();
866 out:
867         return page;
868 }
869
870 #ifdef CONFIG_ZCACHE_WRITEBACK
871
872 static atomic_t zcache_outstanding_writeback_pages_atomic = ATOMIC_INIT(0);
873
874 static inline void inc_zcache_outstanding_writeback_pages(void)
875 {
876         zcache_outstanding_writeback_pages =
877             atomic_inc_return(&zcache_outstanding_writeback_pages_atomic);
878 }
879 static inline void dec_zcache_outstanding_writeback_pages(void)
880 {
881         zcache_outstanding_writeback_pages =
882           atomic_dec_return(&zcache_outstanding_writeback_pages_atomic);
883 };
884 static void unswiz(struct tmem_oid oid, u32 index,
885                                 unsigned *type, pgoff_t *offset);
886
887 /*
888  *  Choose an LRU persistent pageframe and attempt to write it back to
889  *  the backing swap disk by calling frontswap_writeback on both zpages.
890  *
891  *  This is work-in-progress.
892  */
893
894 static void zcache_end_swap_write(struct bio *bio, int err)
895 {
896         end_swap_bio_write(bio, err);
897         dec_zcache_outstanding_writeback_pages();
898         zcache_writtenback_pages++;
899 }
900
901 /*
902  * zcache_get_swap_cache_page
903  *
904  * This is an adaption of read_swap_cache_async()
905  *
906  * If success, page is returned in retpage
907  * Returns 0 if page was already in the swap cache, page is not locked
908  * Returns 1 if the new page needs to be populated, page is locked
909  */
910 static int zcache_get_swap_cache_page(int type, pgoff_t offset,
911                                 struct page *new_page)
912 {
913         struct page *found_page;
914         swp_entry_t entry = swp_entry(type, offset);
915         int err;
916
917         BUG_ON(new_page == NULL);
918         do {
919                 /*
920                  * First check the swap cache.  Since this is normally
921                  * called after lookup_swap_cache() failed, re-calling
922                  * that would confuse statistics.
923                  */
924                 found_page = find_get_page(&swapper_space, entry.val);
925                 if (found_page)
926                         return 0;
927
928                 /*
929                  * call radix_tree_preload() while we can wait.
930                  */
931                 err = radix_tree_preload(GFP_KERNEL);
932                 if (err)
933                         break;
934
935                 /*
936                  * Swap entry may have been freed since our caller observed it.
937                  */
938                 err = swapcache_prepare(entry);
939                 if (err == -EEXIST) { /* seems racy */
940                         radix_tree_preload_end();
941                         continue;
942                 }
943                 if (err) { /* swp entry is obsolete ? */
944                         radix_tree_preload_end();
945                         break;
946                 }
947
948                 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
949                 __set_page_locked(new_page);
950                 SetPageSwapBacked(new_page);
951                 err = __add_to_swap_cache(new_page, entry);
952                 if (likely(!err)) {
953                         radix_tree_preload_end();
954                         lru_cache_add_anon(new_page);
955                         return 1;
956                 }
957                 radix_tree_preload_end();
958                 ClearPageSwapBacked(new_page);
959                 __clear_page_locked(new_page);
960                 /*
961                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
962                  * clear SWAP_HAS_CACHE flag.
963                  */
964                 swapcache_free(entry, NULL);
965                 /* FIXME: is it possible to get here without err==-ENOMEM?
966                  * If not, we can dispense with the do loop, use goto retry */
967         } while (err != -ENOMEM);
968
969         return -ENOMEM;
970 }
971
972 /*
973  * Given a frontswap zpage in zcache (identified by type/offset) and
974  * an empty page, put the page into the swap cache, use frontswap
975  * to get the page from zcache into the empty page, then give it
976  * to the swap subsystem to send to disk (carefully avoiding the
977  * possibility that frontswap might snatch it back).
978  * Returns < 0 if error, 0 if successful, and 1 if successful but
979  * the newpage passed in not needed and should be freed.
980  */
981 static int zcache_frontswap_writeback_zpage(int type, pgoff_t offset,
982                                         struct page *newpage)
983 {
984         struct page *page = newpage;
985         int ret;
986         struct writeback_control wbc = {
987                 .sync_mode = WB_SYNC_NONE,
988         };
989
990         ret = zcache_get_swap_cache_page(type, offset, page);
991         if (ret < 0)
992                 return ret;
993         else if (ret == 0) {
994                 /* more uptodate page is already in swapcache */
995                 __frontswap_invalidate_page(type, offset);
996                 return 1;
997         }
998
999         BUG_ON(!frontswap_has_exclusive_gets); /* load must also invalidate */
1000         /* FIXME: how is it possible to get here when page is unlocked? */
1001         __frontswap_load(page);
1002         SetPageUptodate(page);  /* above does SetPageDirty, is that enough? */
1003
1004         /* start writeback */
1005         SetPageReclaim(page);
1006         /*
1007          * Return value is ignored here because it doesn't change anything
1008          * for us.  Page is returned unlocked.
1009          */
1010         (void)__swap_writepage(page, &wbc, zcache_end_swap_write);
1011         page_cache_release(page);
1012         inc_zcache_outstanding_writeback_pages();
1013
1014         return 0;
1015 }
1016
1017 /*
1018  * The following is still a magic number... we want to allow forward progress
1019  * for writeback because it clears out needed RAM when under pressure, but
1020  * we don't want to allow writeback to absorb and queue too many GFP_KERNEL
1021  * pages if the swap device is very slow.
1022  */
1023 #define ZCACHE_MAX_OUTSTANDING_WRITEBACK_PAGES 6400
1024
1025 /*
1026  * Try to allocate two free pages, first using a non-aggressive alloc,
1027  * then by evicting zcache ephemeral (clean pagecache) pages, and last
1028  * by aggressive GFP_KERNEL alloc.  We allow zbud to choose a pageframe
1029  * consisting of 1-2 zbuds/zpages, then call the writeback_zpage helper
1030  * function above for each.
1031  */
1032 static int zcache_frontswap_writeback(void)
1033 {
1034         struct tmem_handle th[2];
1035         int ret = 0;
1036         int nzbuds, writeback_ret;
1037         unsigned type;
1038         struct page *znewpage1 = NULL, *znewpage2 = NULL;
1039         struct page *evictpage1 = NULL, *evictpage2 = NULL;
1040         struct page *newpage1 = NULL, *newpage2 = NULL;
1041         struct page *page1 = NULL, *page2 = NULL;
1042         pgoff_t offset;
1043
1044         znewpage1 = alloc_page(ZCACHE_GFP_MASK);
1045         znewpage2 = alloc_page(ZCACHE_GFP_MASK);
1046         if (znewpage1 == NULL)
1047                 evictpage1 = zcache_evict_eph_pageframe();
1048         if (znewpage2 == NULL)
1049                 evictpage2 = zcache_evict_eph_pageframe();
1050
1051         if ((evictpage1 == NULL || evictpage2 == NULL) &&
1052             atomic_read(&zcache_outstanding_writeback_pages_atomic) >
1053                                 ZCACHE_MAX_OUTSTANDING_WRITEBACK_PAGES) {
1054                 goto free_and_out;
1055         }
1056         if (znewpage1 == NULL && evictpage1 == NULL)
1057                 newpage1 = alloc_page(GFP_KERNEL);
1058         if (znewpage2 == NULL && evictpage2 == NULL)
1059                 newpage2 = alloc_page(GFP_KERNEL);
1060         if (newpage1 == NULL || newpage2 == NULL)
1061                         goto free_and_out;
1062
1063         /* ok, we have two pageframes pre-allocated, get a pair of zbuds */
1064         nzbuds = zbud_make_zombie_lru(&th[0], NULL, NULL, false);
1065         if (nzbuds == 0) {
1066                 ret = -ENOENT;
1067                 goto free_and_out;
1068         }
1069
1070         /* process the first zbud */
1071         unswiz(th[0].oid, th[0].index, &type, &offset);
1072         page1 = (znewpage1 != NULL) ? znewpage1 :
1073                         ((newpage1 != NULL) ? newpage1 : evictpage1);
1074         writeback_ret = zcache_frontswap_writeback_zpage(type, offset, page1);
1075         if (writeback_ret < 0) {
1076                 ret = -ENOMEM;
1077                 goto free_and_out;
1078         }
1079         if (evictpage1 != NULL)
1080                 zcache_pageframes_freed =
1081                         atomic_inc_return(&zcache_pageframes_freed_atomic);
1082         if (writeback_ret == 0) {
1083                 /* zcache_get_swap_cache_page will free, don't double free */
1084                 znewpage1 = NULL;
1085                 newpage1 = NULL;
1086                 evictpage1 = NULL;
1087         }
1088         if (nzbuds < 2)
1089                 goto free_and_out;
1090
1091         /* if there is a second zbud, process it */
1092         unswiz(th[1].oid, th[1].index, &type, &offset);
1093         page2 = (znewpage2 != NULL) ? znewpage2 :
1094                         ((newpage2 != NULL) ? newpage2 : evictpage2);
1095         writeback_ret = zcache_frontswap_writeback_zpage(type, offset, page2);
1096         if (writeback_ret < 0) {
1097                 ret = -ENOMEM;
1098                 goto free_and_out;
1099         }
1100         if (evictpage2 != NULL)
1101                 zcache_pageframes_freed =
1102                         atomic_inc_return(&zcache_pageframes_freed_atomic);
1103         if (writeback_ret == 0) {
1104                 znewpage2 = NULL;
1105                 newpage2 = NULL;
1106                 evictpage2 = NULL;
1107         }
1108
1109 free_and_out:
1110         if (znewpage1 != NULL)
1111                 page_cache_release(znewpage1);
1112         if (znewpage2 != NULL)
1113                 page_cache_release(znewpage2);
1114         if (newpage1 != NULL)
1115                 page_cache_release(newpage1);
1116         if (newpage2 != NULL)
1117                 page_cache_release(newpage2);
1118         if (evictpage1 != NULL)
1119                 zcache_free_page(evictpage1);
1120         if (evictpage2 != NULL)
1121                 zcache_free_page(evictpage2);
1122         return ret;
1123 }
1124 #endif /* CONFIG_ZCACHE_WRITEBACK */
1125
1126 /*
1127  * When zcache is disabled ("frozen"), pools can be created and destroyed,
1128  * but all puts (and thus all other operations that require memory allocation)
1129  * must fail.  If zcache is unfrozen, accepts puts, then frozen again,
1130  * data consistency requires all puts while frozen to be converted into
1131  * flushes.
1132  */
1133 static bool zcache_freeze;
1134
1135 /*
1136  * This zcache shrinker interface reduces the number of ephemeral pageframes
1137  * used by zcache to approximately the same as the total number of LRU_FILE
1138  * pageframes in use, and now also reduces the number of persistent pageframes
1139  * used by zcache to approximately the same as the total number of LRU_ANON
1140  * pageframes in use.  FIXME POLICY: Probably the writeback should only occur
1141  * if the eviction doesn't free enough pages.
1142  */
1143 static int shrink_zcache_memory(struct shrinker *shrink,
1144                                 struct shrink_control *sc)
1145 {
1146         static bool in_progress;
1147         int ret = -1;
1148         int nr = sc->nr_to_scan;
1149         int nr_evict = 0;
1150         int nr_writeback = 0;
1151         struct page *page;
1152         int  file_pageframes_inuse, anon_pageframes_inuse;
1153
1154         if (nr <= 0)
1155                 goto skip_evict;
1156
1157         /* don't allow more than one eviction thread at a time */
1158         if (in_progress)
1159                 goto skip_evict;
1160
1161         in_progress = true;
1162
1163         /* we are going to ignore nr, and target a different value */
1164         zcache_last_active_file_pageframes =
1165                 global_page_state(NR_LRU_BASE + LRU_ACTIVE_FILE);
1166         zcache_last_inactive_file_pageframes =
1167                 global_page_state(NR_LRU_BASE + LRU_INACTIVE_FILE);
1168         file_pageframes_inuse = zcache_last_active_file_pageframes +
1169                                 zcache_last_inactive_file_pageframes;
1170         if (zcache_eph_pageframes > file_pageframes_inuse)
1171                 nr_evict = zcache_eph_pageframes - file_pageframes_inuse;
1172         else
1173                 nr_evict = 0;
1174         while (nr_evict-- > 0) {
1175                 page = zcache_evict_eph_pageframe();
1176                 if (page == NULL)
1177                         break;
1178                 zcache_free_page(page);
1179         }
1180
1181         zcache_last_active_anon_pageframes =
1182                 global_page_state(NR_LRU_BASE + LRU_ACTIVE_ANON);
1183         zcache_last_inactive_anon_pageframes =
1184                 global_page_state(NR_LRU_BASE + LRU_INACTIVE_ANON);
1185         anon_pageframes_inuse = zcache_last_active_anon_pageframes +
1186                                 zcache_last_inactive_anon_pageframes;
1187         if (zcache_pers_pageframes > anon_pageframes_inuse)
1188                 nr_writeback = zcache_pers_pageframes - anon_pageframes_inuse;
1189         else
1190                 nr_writeback = 0;
1191         while (nr_writeback-- > 0) {
1192 #ifdef CONFIG_ZCACHE_WRITEBACK
1193                 int writeback_ret;
1194                 writeback_ret = zcache_frontswap_writeback();
1195                 if (writeback_ret == -ENOMEM)
1196 #endif
1197                         break;
1198         }
1199         in_progress = false;
1200
1201 skip_evict:
1202         /* resample: has changed, but maybe not all the way yet */
1203         zcache_last_active_file_pageframes =
1204                 global_page_state(NR_LRU_BASE + LRU_ACTIVE_FILE);
1205         zcache_last_inactive_file_pageframes =
1206                 global_page_state(NR_LRU_BASE + LRU_INACTIVE_FILE);
1207         ret = zcache_eph_pageframes - zcache_last_active_file_pageframes +
1208                 zcache_last_inactive_file_pageframes;
1209         if (ret < 0)
1210                 ret = 0;
1211         return ret;
1212 }
1213
1214 static struct shrinker zcache_shrinker = {
1215         .shrink = shrink_zcache_memory,
1216         .seeks = DEFAULT_SEEKS,
1217 };
1218
1219 /*
1220  * zcache shims between cleancache/frontswap ops and tmem
1221  */
1222
1223 /* FIXME rename these core routines to zcache_tmemput etc? */
1224 int zcache_put_page(int cli_id, int pool_id, struct tmem_oid *oidp,
1225                                 uint32_t index, void *page,
1226                                 unsigned int size, bool raw, int ephemeral)
1227 {
1228         struct tmem_pool *pool;
1229         struct tmem_handle th;
1230         int ret = -1;
1231         void *pampd = NULL;
1232
1233         BUG_ON(!irqs_disabled());
1234         pool = zcache_get_pool_by_id(cli_id, pool_id);
1235         if (unlikely(pool == NULL))
1236                 goto out;
1237         if (!zcache_freeze) {
1238                 ret = 0;
1239                 th.client_id = cli_id;
1240                 th.pool_id = pool_id;
1241                 th.oid = *oidp;
1242                 th.index = index;
1243                 pampd = zcache_pampd_create((char *)page, size, raw,
1244                                 ephemeral, &th);
1245                 if (pampd == NULL) {
1246                         ret = -ENOMEM;
1247                         if (ephemeral)
1248                                 inc_zcache_failed_eph_puts();
1249                         else
1250                                 inc_zcache_failed_pers_puts();
1251                 } else {
1252                         if (ramster_enabled)
1253                                 ramster_do_preload_flnode(pool);
1254                         ret = tmem_put(pool, oidp, index, 0, pampd);
1255                         if (ret < 0)
1256                                 BUG();
1257                 }
1258                 zcache_put_pool(pool);
1259         } else {
1260                 inc_zcache_put_to_flush();
1261                 if (ramster_enabled)
1262                         ramster_do_preload_flnode(pool);
1263                 if (atomic_read(&pool->obj_count) > 0)
1264                         /* the put fails whether the flush succeeds or not */
1265                         (void)tmem_flush_page(pool, oidp, index);
1266                 zcache_put_pool(pool);
1267         }
1268 out:
1269         return ret;
1270 }
1271
1272 int zcache_get_page(int cli_id, int pool_id, struct tmem_oid *oidp,
1273                                 uint32_t index, void *page,
1274                                 size_t *sizep, bool raw, int get_and_free)
1275 {
1276         struct tmem_pool *pool;
1277         int ret = -1;
1278         bool eph;
1279
1280         if (!raw) {
1281                 BUG_ON(irqs_disabled());
1282                 BUG_ON(in_softirq());
1283         }
1284         pool = zcache_get_pool_by_id(cli_id, pool_id);
1285         eph = is_ephemeral(pool);
1286         if (likely(pool != NULL)) {
1287                 if (atomic_read(&pool->obj_count) > 0)
1288                         ret = tmem_get(pool, oidp, index, (char *)(page),
1289                                         sizep, raw, get_and_free);
1290                 zcache_put_pool(pool);
1291         }
1292         WARN_ONCE((!is_ephemeral(pool) && (ret != 0)),
1293                         "zcache_get fails on persistent pool, "
1294                         "bad things are very likely to happen soon\n");
1295 #ifdef RAMSTER_TESTING
1296         if (ret != 0 && ret != -1 && !(ret == -EINVAL && is_ephemeral(pool)))
1297                 pr_err("TESTING zcache_get tmem_get returns ret=%d\n", ret);
1298 #endif
1299         return ret;
1300 }
1301
1302 int zcache_flush_page(int cli_id, int pool_id,
1303                                 struct tmem_oid *oidp, uint32_t index)
1304 {
1305         struct tmem_pool *pool;
1306         int ret = -1;
1307         unsigned long flags;
1308
1309         local_irq_save(flags);
1310         inc_zcache_flush_total();
1311         pool = zcache_get_pool_by_id(cli_id, pool_id);
1312         if (ramster_enabled)
1313                 ramster_do_preload_flnode(pool);
1314         if (likely(pool != NULL)) {
1315                 if (atomic_read(&pool->obj_count) > 0)
1316                         ret = tmem_flush_page(pool, oidp, index);
1317                 zcache_put_pool(pool);
1318         }
1319         if (ret >= 0)
1320                 inc_zcache_flush_found();
1321         local_irq_restore(flags);
1322         return ret;
1323 }
1324
1325 int zcache_flush_object(int cli_id, int pool_id,
1326                                 struct tmem_oid *oidp)
1327 {
1328         struct tmem_pool *pool;
1329         int ret = -1;
1330         unsigned long flags;
1331
1332         local_irq_save(flags);
1333         inc_zcache_flobj_total();
1334         pool = zcache_get_pool_by_id(cli_id, pool_id);
1335         if (ramster_enabled)
1336                 ramster_do_preload_flnode(pool);
1337         if (likely(pool != NULL)) {
1338                 if (atomic_read(&pool->obj_count) > 0)
1339                         ret = tmem_flush_object(pool, oidp);
1340                 zcache_put_pool(pool);
1341         }
1342         if (ret >= 0)
1343                 inc_zcache_flobj_found();
1344         local_irq_restore(flags);
1345         return ret;
1346 }
1347
1348 static int zcache_client_destroy_pool(int cli_id, int pool_id)
1349 {
1350         struct tmem_pool *pool = NULL;
1351         struct zcache_client *cli = NULL;
1352         int ret = -1;
1353
1354         if (pool_id < 0)
1355                 goto out;
1356         if (cli_id == LOCAL_CLIENT)
1357                 cli = &zcache_host;
1358         else if ((unsigned int)cli_id < MAX_CLIENTS)
1359                 cli = &zcache_clients[cli_id];
1360         if (cli == NULL)
1361                 goto out;
1362         atomic_inc(&cli->refcount);
1363         pool = cli->tmem_pools[pool_id];
1364         if (pool == NULL)
1365                 goto out;
1366         cli->tmem_pools[pool_id] = NULL;
1367         /* wait for pool activity on other cpus to quiesce */
1368         while (atomic_read(&pool->refcount) != 0)
1369                 ;
1370         atomic_dec(&cli->refcount);
1371         local_bh_disable();
1372         ret = tmem_destroy_pool(pool);
1373         local_bh_enable();
1374         kfree(pool);
1375         if (cli_id == LOCAL_CLIENT)
1376                 pr_info("%s: destroyed local pool id=%d\n", namestr, pool_id);
1377         else
1378                 pr_info("%s: destroyed pool id=%d, client=%d\n",
1379                                 namestr, pool_id, cli_id);
1380 out:
1381         return ret;
1382 }
1383
1384 int zcache_new_pool(uint16_t cli_id, uint32_t flags)
1385 {
1386         int poolid = -1;
1387         struct tmem_pool *pool;
1388         struct zcache_client *cli = NULL;
1389
1390         if (cli_id == LOCAL_CLIENT)
1391                 cli = &zcache_host;
1392         else if ((unsigned int)cli_id < MAX_CLIENTS)
1393                 cli = &zcache_clients[cli_id];
1394         if (cli == NULL)
1395                 goto out;
1396         atomic_inc(&cli->refcount);
1397         pool = kmalloc(sizeof(struct tmem_pool), GFP_ATOMIC);
1398         if (pool == NULL)
1399                 goto out;
1400
1401         for (poolid = 0; poolid < MAX_POOLS_PER_CLIENT; poolid++)
1402                 if (cli->tmem_pools[poolid] == NULL)
1403                         break;
1404         if (poolid >= MAX_POOLS_PER_CLIENT) {
1405                 pr_info("%s: pool creation failed: max exceeded\n", namestr);
1406                 kfree(pool);
1407                 poolid = -1;
1408                 goto out;
1409         }
1410         atomic_set(&pool->refcount, 0);
1411         pool->client = cli;
1412         pool->pool_id = poolid;
1413         tmem_new_pool(pool, flags);
1414         cli->tmem_pools[poolid] = pool;
1415         if (cli_id == LOCAL_CLIENT)
1416                 pr_info("%s: created %s local tmem pool, id=%d\n", namestr,
1417                         flags & TMEM_POOL_PERSIST ? "persistent" : "ephemeral",
1418                         poolid);
1419         else
1420                 pr_info("%s: created %s tmem pool, id=%d, client=%d\n", namestr,
1421                         flags & TMEM_POOL_PERSIST ? "persistent" : "ephemeral",
1422                         poolid, cli_id);
1423 out:
1424         if (cli != NULL)
1425                 atomic_dec(&cli->refcount);
1426         return poolid;
1427 }
1428
1429 static int zcache_local_new_pool(uint32_t flags)
1430 {
1431         return zcache_new_pool(LOCAL_CLIENT, flags);
1432 }
1433
1434 int zcache_autocreate_pool(unsigned int cli_id, unsigned int pool_id, bool eph)
1435 {
1436         struct tmem_pool *pool;
1437         struct zcache_client *cli = NULL;
1438         uint32_t flags = eph ? 0 : TMEM_POOL_PERSIST;
1439         int ret = -1;
1440
1441         BUG_ON(!ramster_enabled);
1442         if (cli_id == LOCAL_CLIENT)
1443                 goto out;
1444         if (pool_id >= MAX_POOLS_PER_CLIENT)
1445                 goto out;
1446         if (cli_id >= MAX_CLIENTS)
1447                 goto out;
1448
1449         cli = &zcache_clients[cli_id];
1450         if ((eph && disable_cleancache) || (!eph && disable_frontswap)) {
1451                 pr_err("zcache_autocreate_pool: pool type disabled\n");
1452                 goto out;
1453         }
1454         if (!cli->allocated) {
1455                 if (zcache_new_client(cli_id)) {
1456                         pr_err("zcache_autocreate_pool: can't create client\n");
1457                         goto out;
1458                 }
1459                 cli = &zcache_clients[cli_id];
1460         }
1461         atomic_inc(&cli->refcount);
1462         pool = cli->tmem_pools[pool_id];
1463         if (pool != NULL) {
1464                 if (pool->persistent && eph) {
1465                         pr_err("zcache_autocreate_pool: type mismatch\n");
1466                         goto out;
1467                 }
1468                 ret = 0;
1469                 goto out;
1470         }
1471         pool = kmalloc(sizeof(struct tmem_pool), GFP_KERNEL);
1472         if (pool == NULL)
1473                 goto out;
1474
1475         atomic_set(&pool->refcount, 0);
1476         pool->client = cli;
1477         pool->pool_id = pool_id;
1478         tmem_new_pool(pool, flags);
1479         cli->tmem_pools[pool_id] = pool;
1480         pr_info("%s: AUTOcreated %s tmem poolid=%d, for remote client=%d\n",
1481                 namestr, flags & TMEM_POOL_PERSIST ? "persistent" : "ephemeral",
1482                 pool_id, cli_id);
1483         ret = 0;
1484 out:
1485         if (cli != NULL)
1486                 atomic_dec(&cli->refcount);
1487         return ret;
1488 }
1489
1490 /**********
1491  * Two kernel functionalities currently can be layered on top of tmem.
1492  * These are "cleancache" which is used as a second-chance cache for clean
1493  * page cache pages; and "frontswap" which is used for swap pages
1494  * to avoid writes to disk.  A generic "shim" is provided here for each
1495  * to translate in-kernel semantics to zcache semantics.
1496  */
1497
1498 static void zcache_cleancache_put_page(int pool_id,
1499                                         struct cleancache_filekey key,
1500                                         pgoff_t index, struct page *page)
1501 {
1502         u32 ind = (u32) index;
1503         struct tmem_oid oid = *(struct tmem_oid *)&key;
1504
1505         if (!disable_cleancache_ignore_nonactive && !PageWasActive(page)) {
1506                 inc_zcache_eph_nonactive_puts_ignored();
1507                 return;
1508         }
1509         if (likely(ind == index))
1510                 (void)zcache_put_page(LOCAL_CLIENT, pool_id, &oid, index,
1511                                         page, PAGE_SIZE, false, 1);
1512 }
1513
1514 static int zcache_cleancache_get_page(int pool_id,
1515                                         struct cleancache_filekey key,
1516                                         pgoff_t index, struct page *page)
1517 {
1518         u32 ind = (u32) index;
1519         struct tmem_oid oid = *(struct tmem_oid *)&key;
1520         size_t size;
1521         int ret = -1;
1522
1523         if (likely(ind == index)) {
1524                 ret = zcache_get_page(LOCAL_CLIENT, pool_id, &oid, index,
1525                                         page, &size, false, 0);
1526                 BUG_ON(ret >= 0 && size != PAGE_SIZE);
1527                 if (ret == 0)
1528                         SetPageWasActive(page);
1529         }
1530         return ret;
1531 }
1532
1533 static void zcache_cleancache_flush_page(int pool_id,
1534                                         struct cleancache_filekey key,
1535                                         pgoff_t index)
1536 {
1537         u32 ind = (u32) index;
1538         struct tmem_oid oid = *(struct tmem_oid *)&key;
1539
1540         if (likely(ind == index))
1541                 (void)zcache_flush_page(LOCAL_CLIENT, pool_id, &oid, ind);
1542 }
1543
1544 static void zcache_cleancache_flush_inode(int pool_id,
1545                                         struct cleancache_filekey key)
1546 {
1547         struct tmem_oid oid = *(struct tmem_oid *)&key;
1548
1549         (void)zcache_flush_object(LOCAL_CLIENT, pool_id, &oid);
1550 }
1551
1552 static void zcache_cleancache_flush_fs(int pool_id)
1553 {
1554         if (pool_id >= 0)
1555                 (void)zcache_client_destroy_pool(LOCAL_CLIENT, pool_id);
1556 }
1557
1558 static int zcache_cleancache_init_fs(size_t pagesize)
1559 {
1560         BUG_ON(sizeof(struct cleancache_filekey) !=
1561                                 sizeof(struct tmem_oid));
1562         BUG_ON(pagesize != PAGE_SIZE);
1563         return zcache_local_new_pool(0);
1564 }
1565
1566 static int zcache_cleancache_init_shared_fs(char *uuid, size_t pagesize)
1567 {
1568         /* shared pools are unsupported and map to private */
1569         BUG_ON(sizeof(struct cleancache_filekey) !=
1570                                 sizeof(struct tmem_oid));
1571         BUG_ON(pagesize != PAGE_SIZE);
1572         return zcache_local_new_pool(0);
1573 }
1574
1575 static struct cleancache_ops zcache_cleancache_ops = {
1576         .put_page = zcache_cleancache_put_page,
1577         .get_page = zcache_cleancache_get_page,
1578         .invalidate_page = zcache_cleancache_flush_page,
1579         .invalidate_inode = zcache_cleancache_flush_inode,
1580         .invalidate_fs = zcache_cleancache_flush_fs,
1581         .init_shared_fs = zcache_cleancache_init_shared_fs,
1582         .init_fs = zcache_cleancache_init_fs
1583 };
1584
1585 struct cleancache_ops *zcache_cleancache_register_ops(void)
1586 {
1587         struct cleancache_ops *old_ops =
1588                 cleancache_register_ops(&zcache_cleancache_ops);
1589
1590         return old_ops;
1591 }
1592
1593 /* a single tmem poolid is used for all frontswap "types" (swapfiles) */
1594 static int zcache_frontswap_poolid __read_mostly = -1;
1595
1596 /*
1597  * Swizzling increases objects per swaptype, increasing tmem concurrency
1598  * for heavy swaploads.  Later, larger nr_cpus -> larger SWIZ_BITS
1599  * Setting SWIZ_BITS to 27 basically reconstructs the swap entry from
1600  * frontswap_get_page(), but has side-effects. Hence using 8.
1601  */
1602 #define SWIZ_BITS               8
1603 #define SWIZ_MASK               ((1 << SWIZ_BITS) - 1)
1604 #define _oswiz(_type, _ind)     ((_type << SWIZ_BITS) | (_ind & SWIZ_MASK))
1605 #define iswiz(_ind)             (_ind >> SWIZ_BITS)
1606
1607 static inline struct tmem_oid oswiz(unsigned type, u32 ind)
1608 {
1609         struct tmem_oid oid = { .oid = { 0 } };
1610         oid.oid[0] = _oswiz(type, ind);
1611         return oid;
1612 }
1613
1614 #ifdef CONFIG_ZCACHE_WRITEBACK
1615 static void unswiz(struct tmem_oid oid, u32 index,
1616                                 unsigned *type, pgoff_t *offset)
1617 {
1618         *type = (unsigned)(oid.oid[0] >> SWIZ_BITS);
1619         *offset = (pgoff_t)((index << SWIZ_BITS) |
1620                         (oid.oid[0] & SWIZ_MASK));
1621 }
1622 #endif
1623
1624 static int zcache_frontswap_put_page(unsigned type, pgoff_t offset,
1625                                         struct page *page)
1626 {
1627         u64 ind64 = (u64)offset;
1628         u32 ind = (u32)offset;
1629         struct tmem_oid oid = oswiz(type, ind);
1630         int ret = -1;
1631         unsigned long flags;
1632
1633         BUG_ON(!PageLocked(page));
1634         if (!disable_frontswap_ignore_nonactive && !PageWasActive(page)) {
1635                 inc_zcache_pers_nonactive_puts_ignored();
1636                 ret = -ERANGE;
1637                 goto out;
1638         }
1639         if (likely(ind64 == ind)) {
1640                 local_irq_save(flags);
1641                 ret = zcache_put_page(LOCAL_CLIENT, zcache_frontswap_poolid,
1642                                         &oid, iswiz(ind),
1643                                         page, PAGE_SIZE, false, 0);
1644                 local_irq_restore(flags);
1645         }
1646 out:
1647         return ret;
1648 }
1649
1650 /* returns 0 if the page was successfully gotten from frontswap, -1 if
1651  * was not present (should never happen!) */
1652 static int zcache_frontswap_get_page(unsigned type, pgoff_t offset,
1653                                         struct page *page)
1654 {
1655         u64 ind64 = (u64)offset;
1656         u32 ind = (u32)offset;
1657         struct tmem_oid oid = oswiz(type, ind);
1658         size_t size;
1659         int ret = -1, get_and_free;
1660
1661         if (frontswap_has_exclusive_gets)
1662                 get_and_free = 1;
1663         else
1664                 get_and_free = -1;
1665         BUG_ON(!PageLocked(page));
1666         if (likely(ind64 == ind)) {
1667                 ret = zcache_get_page(LOCAL_CLIENT, zcache_frontswap_poolid,
1668                                         &oid, iswiz(ind),
1669                                         page, &size, false, get_and_free);
1670                 BUG_ON(ret >= 0 && size != PAGE_SIZE);
1671         }
1672         return ret;
1673 }
1674
1675 /* flush a single page from frontswap */
1676 static void zcache_frontswap_flush_page(unsigned type, pgoff_t offset)
1677 {
1678         u64 ind64 = (u64)offset;
1679         u32 ind = (u32)offset;
1680         struct tmem_oid oid = oswiz(type, ind);
1681
1682         if (likely(ind64 == ind))
1683                 (void)zcache_flush_page(LOCAL_CLIENT, zcache_frontswap_poolid,
1684                                         &oid, iswiz(ind));
1685 }
1686
1687 /* flush all pages from the passed swaptype */
1688 static void zcache_frontswap_flush_area(unsigned type)
1689 {
1690         struct tmem_oid oid;
1691         int ind;
1692
1693         for (ind = SWIZ_MASK; ind >= 0; ind--) {
1694                 oid = oswiz(type, ind);
1695                 (void)zcache_flush_object(LOCAL_CLIENT,
1696                                                 zcache_frontswap_poolid, &oid);
1697         }
1698 }
1699
1700 static void zcache_frontswap_init(unsigned ignored)
1701 {
1702         /* a single tmem poolid is used for all frontswap "types" (swapfiles) */
1703         if (zcache_frontswap_poolid < 0)
1704                 zcache_frontswap_poolid =
1705                         zcache_local_new_pool(TMEM_POOL_PERSIST);
1706 }
1707
1708 static struct frontswap_ops zcache_frontswap_ops = {
1709         .store = zcache_frontswap_put_page,
1710         .load = zcache_frontswap_get_page,
1711         .invalidate_page = zcache_frontswap_flush_page,
1712         .invalidate_area = zcache_frontswap_flush_area,
1713         .init = zcache_frontswap_init
1714 };
1715
1716 struct frontswap_ops *zcache_frontswap_register_ops(void)
1717 {
1718         struct frontswap_ops *old_ops =
1719                 frontswap_register_ops(&zcache_frontswap_ops);
1720
1721         return old_ops;
1722 }
1723
1724 /*
1725  * zcache initialization
1726  * NOTE FOR NOW zcache or ramster MUST BE PROVIDED AS A KERNEL BOOT PARAMETER
1727  * OR NOTHING HAPPENS!
1728  */
1729
1730 #ifndef CONFIG_ZCACHE_MODULE
1731 static int __init enable_zcache(char *s)
1732 {
1733         zcache_enabled = true;
1734         return 1;
1735 }
1736 __setup("zcache", enable_zcache);
1737
1738 static int __init enable_ramster(char *s)
1739 {
1740         zcache_enabled = true;
1741 #ifdef CONFIG_RAMSTER
1742         ramster_enabled = true;
1743 #endif
1744         return 1;
1745 }
1746 __setup("ramster", enable_ramster);
1747
1748 /* allow independent dynamic disabling of cleancache and frontswap */
1749
1750 static int __init no_cleancache(char *s)
1751 {
1752         disable_cleancache = true;
1753         return 1;
1754 }
1755
1756 __setup("nocleancache", no_cleancache);
1757
1758 static int __init no_frontswap(char *s)
1759 {
1760         disable_frontswap = true;
1761         return 1;
1762 }
1763
1764 __setup("nofrontswap", no_frontswap);
1765
1766 static int __init no_frontswap_exclusive_gets(char *s)
1767 {
1768         frontswap_has_exclusive_gets = false;
1769         return 1;
1770 }
1771
1772 __setup("nofrontswapexclusivegets", no_frontswap_exclusive_gets);
1773
1774 static int __init no_frontswap_ignore_nonactive(char *s)
1775 {
1776         disable_frontswap_ignore_nonactive = true;
1777         return 1;
1778 }
1779
1780 __setup("nofrontswapignorenonactive", no_frontswap_ignore_nonactive);
1781
1782 static int __init no_cleancache_ignore_nonactive(char *s)
1783 {
1784         disable_cleancache_ignore_nonactive = true;
1785         return 1;
1786 }
1787
1788 __setup("nocleancacheignorenonactive", no_cleancache_ignore_nonactive);
1789
1790 static int __init enable_zcache_compressor(char *s)
1791 {
1792         strlcpy(zcache_comp_name, s, sizeof(zcache_comp_name));
1793         zcache_enabled = true;
1794         return 1;
1795 }
1796 __setup("zcache=", enable_zcache_compressor);
1797 #endif
1798
1799
1800 static int zcache_comp_init(void)
1801 {
1802         int ret = 0;
1803
1804         /* check crypto algorithm */
1805 #ifdef CONFIG_ZCACHE_MODULE
1806         ret = crypto_has_comp(zcache_comp_name, 0, 0);
1807         if (!ret) {
1808                 ret = -1;
1809                 goto out;
1810         }
1811 #else
1812         if (*zcache_comp_name != '\0') {
1813                 ret = crypto_has_comp(zcache_comp_name, 0, 0);
1814                 if (!ret)
1815                         pr_info("zcache: %s not supported\n",
1816                                         zcache_comp_name);
1817                 goto out;
1818         }
1819         if (!ret)
1820                 strcpy(zcache_comp_name, "lzo");
1821         ret = crypto_has_comp(zcache_comp_name, 0, 0);
1822         if (!ret) {
1823                 ret = 1;
1824                 goto out;
1825         }
1826 #endif
1827         pr_info("zcache: using %s compressor\n", zcache_comp_name);
1828
1829         /* alloc percpu transforms */
1830         ret = 0;
1831         zcache_comp_pcpu_tfms = alloc_percpu(struct crypto_comp *);
1832         if (!zcache_comp_pcpu_tfms)
1833                 ret = 1;
1834 out:
1835         return ret;
1836 }
1837
1838 static int zcache_init(void)
1839 {
1840         int ret = 0;
1841
1842 #ifdef CONFIG_ZCACHE_MODULE
1843         zcache_enabled = 1;
1844 #endif
1845         if (ramster_enabled) {
1846                 namestr = "ramster";
1847                 ramster_register_pamops(&zcache_pamops);
1848         }
1849         zcache_debugfs_init();
1850         if (zcache_enabled) {
1851                 unsigned int cpu;
1852
1853                 tmem_register_hostops(&zcache_hostops);
1854                 tmem_register_pamops(&zcache_pamops);
1855                 ret = register_cpu_notifier(&zcache_cpu_notifier_block);
1856                 if (ret) {
1857                         pr_err("%s: can't register cpu notifier\n", namestr);
1858                         goto out;
1859                 }
1860                 ret = zcache_comp_init();
1861                 if (ret) {
1862                         pr_err("%s: compressor initialization failed\n",
1863                                 namestr);
1864                         goto out;
1865                 }
1866                 for_each_online_cpu(cpu) {
1867                         void *pcpu = (void *)(long)cpu;
1868                         zcache_cpu_notifier(&zcache_cpu_notifier_block,
1869                                 CPU_UP_PREPARE, pcpu);
1870                 }
1871         }
1872         zcache_objnode_cache = kmem_cache_create("zcache_objnode",
1873                                 sizeof(struct tmem_objnode), 0, 0, NULL);
1874         zcache_obj_cache = kmem_cache_create("zcache_obj",
1875                                 sizeof(struct tmem_obj), 0, 0, NULL);
1876         ret = zcache_new_client(LOCAL_CLIENT);
1877         if (ret) {
1878                 pr_err("%s: can't create client\n", namestr);
1879                 goto out;
1880         }
1881         zbud_init();
1882         if (zcache_enabled && !disable_cleancache) {
1883                 struct cleancache_ops *old_ops;
1884
1885                 register_shrinker(&zcache_shrinker);
1886                 old_ops = zcache_cleancache_register_ops();
1887                 pr_info("%s: cleancache enabled using kernel transcendent "
1888                         "memory and compression buddies\n", namestr);
1889 #ifdef CONFIG_ZCACHE_DEBUG
1890                 pr_info("%s: cleancache: ignorenonactive = %d\n",
1891                         namestr, !disable_cleancache_ignore_nonactive);
1892 #endif
1893                 if (old_ops != NULL)
1894                         pr_warn("%s: cleancache_ops overridden\n", namestr);
1895         }
1896         if (zcache_enabled && !disable_frontswap) {
1897                 struct frontswap_ops *old_ops;
1898
1899                 old_ops = zcache_frontswap_register_ops();
1900                 if (frontswap_has_exclusive_gets)
1901                         frontswap_tmem_exclusive_gets(true);
1902                 pr_info("%s: frontswap enabled using kernel transcendent "
1903                         "memory and compression buddies\n", namestr);
1904 #ifdef CONFIG_ZCACHE_DEBUG
1905                 pr_info("%s: frontswap: excl gets = %d active only = %d\n",
1906                         namestr, frontswap_has_exclusive_gets,
1907                         !disable_frontswap_ignore_nonactive);
1908 #endif
1909                 if (IS_ERR(old_ops) || old_ops) {
1910                         if (IS_ERR(old_ops))
1911                                 return PTR_RET(old_ops);
1912                         pr_warn("%s: frontswap_ops overridden\n", namestr);
1913                 }
1914         }
1915         if (ramster_enabled)
1916                 ramster_init(!disable_cleancache, !disable_frontswap,
1917                                 frontswap_has_exclusive_gets,
1918                                 !disable_frontswap_selfshrink);
1919 out:
1920         return ret;
1921 }
1922
1923 #ifdef CONFIG_ZCACHE_MODULE
1924 #ifdef CONFIG_RAMSTER
1925 module_param(ramster_enabled, bool, S_IRUGO);
1926 module_param(disable_frontswap_selfshrink, int, S_IRUGO);
1927 #endif
1928 module_param(disable_cleancache, bool, S_IRUGO);
1929 module_param(disable_frontswap, bool, S_IRUGO);
1930 #ifdef FRONTSWAP_HAS_EXCLUSIVE_GETS
1931 module_param(frontswap_has_exclusive_gets, bool, S_IRUGO);
1932 #endif
1933 module_param(disable_frontswap_ignore_nonactive, bool, S_IRUGO);
1934 module_param(zcache_comp_name, charp, S_IRUGO);
1935 module_init(zcache_init);
1936 MODULE_LICENSE("GPL");
1937 MODULE_AUTHOR("Dan Magenheimer <dan.magenheimer@oracle.com>");
1938 MODULE_DESCRIPTION("In-kernel compression of cleancache/frontswap pages");
1939 #else
1940 late_initcall(zcache_init);
1941 #endif