1/*
   2 * zsmalloc memory allocator
   3 *
   4 * Copyright (C) 2011  Nitin Gupta
   5 * Copyright (C) 2012, 2013 Minchan Kim
   6 *
   7 * This code is released using a dual license strategy: BSD/GPL
   8 * You can choose the license that better fits your requirements.
   9 *
  10 * Released under the terms of 3-clause BSD License
  11 * Released under the terms of GNU General Public License Version 2.0
  12 */
  13
  14/*
  15 * This allocator is designed for use with zram. Thus, the allocator is
  16 * supposed to work well under low memory conditions. In particular, it
  17 * never attempts higher order page allocation which is very likely to
  18 * fail under memory pressure. On the other hand, if we just use single
  19 * (0-order) pages, it would suffer from very high fragmentation --
  20 * any object of size PAGE_SIZE/2 or larger would occupy an entire page.
  21 * This was one of the major issues with its predecessor (xvmalloc).
  22 *
  23 * To overcome these issues, zsmalloc allocates a bunch of 0-order pages
  24 * and links them together using various 'struct page' fields. These linked
  25 * pages act as a single higher-order page i.e. an object can span 0-order
  26 * page boundaries. The code refers to these linked pages as a single entity
  27 * called zspage.
  28 *
  29 * For simplicity, zsmalloc can only allocate objects of size up to PAGE_SIZE
  30 * since this satisfies the requirements of all its current users (in the
  31 * worst case, page is incompressible and is thus stored "as-is" i.e. in
  32 * uncompressed form). For allocation requests larger than this size, failure
  33 * is returned (see zs_malloc).
  34 *
  35 * Additionally, zs_malloc() does not return a dereferenceable pointer.
  36 * Instead, it returns an opaque handle (unsigned long) which encodes actual
  37 * location of the allocated object. The reason for this indirection is that
  38 * zsmalloc does not keep zspages permanently mapped since that would cause
  39 * issues on 32-bit systems where the VA region for kernel space mappings
  40 * is very small. So, before using the allocating memory, the object has to
  41 * be mapped using zs_map_object() to get a usable pointer and subsequently
  42 * unmapped using zs_unmap_object().
  43 *
  44 * Following is how we use various fields and flags of underlying
  45 * struct page(s) to form a zspage.
  46 *
  47 * Usage of struct page fields:
  48 *	page->first_page: points to the first component (0-order) page
  49 *	page->index (union with page->freelist): offset of the first object
  50 *		starting in this page. For the first page, this is
  51 *		always 0, so we use this field (aka freelist) to point
  52 *		to the first free object in zspage.
  53 *	page->lru: links together all component pages (except the first page)
  54 *		of a zspage
  55 *
  56 *	For _first_ page only:
  57 *
  58 *	page->private (union with page->first_page): refers to the
  59 *		component page after the first page
  60 *	page->freelist: points to the first free object in zspage.
  61 *		Free objects are linked together using in-place
  62 *		metadata.
  63 *	page->objects: maximum number of objects we can store in this
  64 *		zspage (class->zspage_order * PAGE_SIZE / class->size)
  65 *	page->lru: links together first pages of various zspages.
  66 *		Basically forming list of zspages in a fullness group.
  67 *	page->mapping: class index and fullness group of the zspage
  68 *
  69 * Usage of struct page flags:
  70 *	PG_private: identifies the first component page
  71 *	PG_private2: identifies the last component page
  72 *
  73 */
  74
  75#ifdef CONFIG_ZSMALLOC_DEBUG
  76#define DEBUG
  77#endif
  78
  79#include <linux/module.h>
  80#include <linux/kernel.h>
 
 
  81#include <linux/bitops.h>
  82#include <linux/errno.h>
  83#include <linux/highmem.h>
  84#include <linux/string.h>
  85#include <linux/slab.h>
  86#include <asm/tlbflush.h>
  87#include <asm/pgtable.h>
  88#include <linux/cpumask.h>
  89#include <linux/cpu.h>
  90#include <linux/vmalloc.h>
  91#include <linux/hardirq.h>
  92#include <linux/spinlock.h>
 
  93#include <linux/types.h>
 
  94#include <linux/zsmalloc.h>
 
 
 
 
 
 
 
 
 
  95
  96/*
  97 * This must be power of 2 and greater than of equal to sizeof(link_free).
  98 * These two conditions ensure that any 'struct link_free' itself doesn't
  99 * span more than 1 page which avoids complex case of mapping 2 pages simply
 100 * to restore link_free pointer values.
 101 */
 102#define ZS_ALIGN		8
 103
 104/*
 105 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
 106 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
 107 */
 108#define ZS_MAX_ZSPAGE_ORDER 2
 109#define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
 110
 
 
 111/*
 112 * Object location (<PFN>, <obj_idx>) is encoded as
 113 * as single (unsigned long) handle value.
 114 *
 115 * Note that object index <obj_idx> is relative to system
 116 * page <PFN> it is stored in, so for each sub-page belonging
 117 * to a zspage, obj_idx starts with 0.
 118 *
 119 * This is made more complicated by various memory models and PAE.
 120 */
 121
 122#ifndef MAX_PHYSMEM_BITS
 123#ifdef CONFIG_HIGHMEM64G
 124#define MAX_PHYSMEM_BITS 36
 125#else /* !CONFIG_HIGHMEM64G */
 126/*
 127 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
 128 * be PAGE_SHIFT
 129 */
 130#define MAX_PHYSMEM_BITS BITS_PER_LONG
 131#endif
 132#endif
 133#define _PFN_BITS		(MAX_PHYSMEM_BITS - PAGE_SHIFT)
 134#define OBJ_INDEX_BITS	(BITS_PER_LONG - _PFN_BITS)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 135#define OBJ_INDEX_MASK	((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
 136
 
 
 
 
 
 137#define MAX(a, b) ((a) >= (b) ? (a) : (b))
 138/* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
 139#define ZS_MIN_ALLOC_SIZE \
 140	MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
 
 141#define ZS_MAX_ALLOC_SIZE	PAGE_SIZE
 142
 143/*
 144 * On systems with 4K page size, this gives 254 size classes! There is a
 145 * trader-off here:
 146 *  - Large number of size classes is potentially wasteful as free page are
 147 *    spread across these classes
 148 *  - Small number of size classes causes large internal fragmentation
 149 *  - Probably its better to use specific size classes (empirically
 150 *    determined). NOTE: all those class sizes must be set as multiple of
 151 *    ZS_ALIGN to make sure link_free itself never has to span 2 pages.
 152 *
 153 *  ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
 154 *  (reason above)
 155 */
 156#define ZS_SIZE_CLASS_DELTA	(PAGE_SIZE >> 8)
 157#define ZS_SIZE_CLASSES		((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / \
 158					ZS_SIZE_CLASS_DELTA + 1)
 159
 160/*
 161 * We do not maintain any list for completely empty or full pages
 162 */
 163enum fullness_group {
 164	ZS_ALMOST_FULL,
 165	ZS_ALMOST_EMPTY,
 166	_ZS_NR_FULLNESS_GROUPS,
 
 
 
 167
 168	ZS_EMPTY,
 169	ZS_FULL
 
 
 
 
 
 
 
 
 
 
 170};
 171
 
 
 
 
 
 
 
 
 172/*
 173 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
 174 *	n <= N / f, where
 175 * n = number of allocated objects
 176 * N = total number of objects zspage can store
 177 * f = 1/fullness_threshold_frac
 178 *
 179 * Similarly, we assign zspage to:
 180 *	ZS_ALMOST_FULL	when n > N / f
 181 *	ZS_EMPTY	when n == 0
 182 *	ZS_FULL		when n == N
 183 *
 184 * (see: fix_fullness_group())
 185 */
 186static const int fullness_threshold_frac = 4;
 
 187
 188struct size_class {
 
 
 189	/*
 190	 * Size of objects stored in this class. Must be multiple
 191	 * of ZS_ALIGN.
 192	 */
 193	int size;
 194	unsigned int index;
 195
 196	/* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
 197	int pages_per_zspage;
 198
 199	spinlock_t lock;
 
 
 200
 201	/* stats */
 202	u64 pages_allocated;
 
 
 
 203
 204	struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS];
 205};
 
 
 
 
 
 
 
 206
 207/*
 208 * Placed within free objects to form a singly linked list.
 209 * For every zspage, first_page->freelist gives head of this list.
 210 *
 211 * This must be power of 2 and less than or equal to ZS_ALIGN
 212 */
 213struct link_free {
 214	/* Handle of next free chunk (encodes <PFN, obj_idx>) */
 215	void *next;
 
 
 
 
 
 
 
 
 
 216};
 217
 218struct zs_pool {
 219	struct size_class size_class[ZS_SIZE_CLASSES];
 
 
 
 
 
 
 
 
 220
 221	gfp_t flags;	/* allocation flags used when growing pool */
 
 
 
 
 
 
 
 
 
 
 
 
 
 222};
 223
 224/*
 225 * A zspage's class index and fullness group
 226 * are encoded in its (first)page->mapping
 227 */
 228#define CLASS_IDX_BITS	28
 229#define FULLNESS_BITS	4
 230#define CLASS_IDX_MASK	((1 << CLASS_IDX_BITS) - 1)
 231#define FULLNESS_MASK	((1 << FULLNESS_BITS) - 1)
 
 
 
 
 
 
 
 232
 233struct mapping_area {
 234#ifdef CONFIG_PGTABLE_MAPPING
 235	struct vm_struct *vm; /* vm area for mapping object that span pages */
 236#else
 237	char *vm_buf; /* copy buffer for objects that span pages */
 238#endif
 239	char *vm_addr; /* address of kmap_atomic()'ed pages */
 240	enum zs_mapmode vm_mm; /* mapping mode */
 241};
 242
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 243
 244/* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
 245static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
 246
 247static int is_first_page(struct page *page)
 
 
 
 
 
 248{
 249	return PagePrivate(page);
 250}
 251
 252static int is_last_page(struct page *page)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 253{
 254	return PagePrivate2(page);
 255}
 256
 257static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
 
 
 
 
 
 
 
 
 
 
 
 258				enum fullness_group *fullness)
 259{
 260	unsigned long m;
 261	BUG_ON(!is_first_page(page));
 262
 263	m = (unsigned long)page->mapping;
 264	*fullness = m & FULLNESS_MASK;
 265	*class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
 266}
 267
 268static void set_zspage_mapping(struct page *page, unsigned int class_idx,
 
 269				enum fullness_group fullness)
 270{
 271	unsigned long m;
 272	BUG_ON(!is_first_page(page));
 273
 274	m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
 275			(fullness & FULLNESS_MASK);
 276	page->mapping = (struct address_space *)m;
 277}
 278
 279/*
 280 * zsmalloc divides the pool into various size classes where each
 281 * class maintains a list of zspages where each zspage is divided
 282 * into equal sized chunks. Each allocation falls into one of these
 283 * classes depending on its size. This function returns index of the
 284 * size class which has chunk size big enough to hold the give size.
 285 */
 286static int get_size_class_index(int size)
 287{
 288	int idx = 0;
 289
 290	if (likely(size > ZS_MIN_ALLOC_SIZE))
 291		idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
 292				ZS_SIZE_CLASS_DELTA);
 293
 294	return idx;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 295}
 296
 
 
 
 
 
 
 297/*
 298 * For each size class, zspages are divided into different groups
 299 * depending on how "full" they are. This was done so that we could
 300 * easily find empty or nearly empty zspages when we try to shrink
 301 * the pool (not yet implemented). This function returns fullness
 302 * status of the given page.
 303 */
 304static enum fullness_group get_fullness_group(struct page *page)
 
 305{
 306	int inuse, max_objects;
 307	enum fullness_group fg;
 308	BUG_ON(!is_first_page(page));
 309
 310	inuse = page->inuse;
 311	max_objects = page->objects;
 312
 313	if (inuse == 0)
 314		fg = ZS_EMPTY;
 315	else if (inuse == max_objects)
 316		fg = ZS_FULL;
 317	else if (inuse <= max_objects / fullness_threshold_frac)
 318		fg = ZS_ALMOST_EMPTY;
 319	else
 320		fg = ZS_ALMOST_FULL;
 321
 322	return fg;
 323}
 324
 325/*
 326 * Each size class maintains various freelists and zspages are assigned
 327 * to one of these freelists based on the number of live objects they
 328 * have. This functions inserts the given zspage into the freelist
 329 * identified by <class, fullness_group>.
 330 */
 331static void insert_zspage(struct page *page, struct size_class *class,
 
 332				enum fullness_group fullness)
 333{
 334	struct page **head;
 335
 336	BUG_ON(!is_first_page(page));
 337
 338	if (fullness >= _ZS_NR_FULLNESS_GROUPS)
 339		return;
 340
 341	head = &class->fullness_list[fullness];
 342	if (*head)
 343		list_add_tail(&page->lru, &(*head)->lru);
 344
 345	*head = page;
 
 
 
 
 
 
 
 
 
 
 
 
 
 346}
 347
 348/*
 349 * This function removes the given zspage from the freelist identified
 350 * by <class, fullness_group>.
 351 */
 352static void remove_zspage(struct page *page, struct size_class *class,
 
 353				enum fullness_group fullness)
 354{
 355	struct page **head;
 356
 357	BUG_ON(!is_first_page(page));
 358
 359	if (fullness >= _ZS_NR_FULLNESS_GROUPS)
 360		return;
 361
 362	head = &class->fullness_list[fullness];
 363	BUG_ON(!*head);
 364	if (list_empty(&(*head)->lru))
 365		*head = NULL;
 366	else if (*head == page)
 367		*head = (struct page *)list_entry((*head)->lru.next,
 368					struct page, lru);
 369
 370	list_del_init(&page->lru);
 371}
 372
 373/*
 374 * Each size class maintains zspages in different fullness groups depending
 375 * on the number of live objects they contain. When allocating or freeing
 376 * objects, the fullness status of the page can change, say, from ALMOST_FULL
 377 * to ALMOST_EMPTY when freeing an object. This function checks if such
 378 * a status change has occurred for the given page and accordingly moves the
 379 * page from the freelist of the old fullness group to that of the new
 380 * fullness group.
 381 */
 382static enum fullness_group fix_fullness_group(struct zs_pool *pool,
 383						struct page *page)
 384{
 385	int class_idx;
 386	struct size_class *class;
 387	enum fullness_group currfg, newfg;
 388
 389	BUG_ON(!is_first_page(page));
 390
 391	get_zspage_mapping(page, &class_idx, &currfg);
 392	newfg = get_fullness_group(page);
 393	if (newfg == currfg)
 394		goto out;
 395
 396	class = &pool->size_class[class_idx];
 397	remove_zspage(page, class, currfg);
 398	insert_zspage(page, class, newfg);
 399	set_zspage_mapping(page, class_idx, newfg);
 
 
 400
 401out:
 402	return newfg;
 403}
 404
 405/*
 406 * We have to decide on how many pages to link together
 407 * to form a zspage for each size class. This is important
 408 * to reduce wastage due to unusable space left at end of
 409 * each zspage which is given as:
 410 *	wastage = Zp - Zp % size_class
 
 411 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
 412 *
 413 * For example, for size class of 3/8 * PAGE_SIZE, we should
 414 * link together 3 PAGE_SIZE sized pages to form a zspage
 415 * since then we can perfectly fit in 8 such objects.
 416 */
 417static int get_pages_per_zspage(int class_size)
 418{
 419	int i, max_usedpc = 0;
 420	/* zspage order which gives maximum used size per KB */
 421	int max_usedpc_order = 1;
 422
 423	for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
 424		int zspage_size;
 425		int waste, usedpc;
 426
 427		zspage_size = i * PAGE_SIZE;
 428		waste = zspage_size % class_size;
 429		usedpc = (zspage_size - waste) * 100 / zspage_size;
 430
 431		if (usedpc > max_usedpc) {
 432			max_usedpc = usedpc;
 433			max_usedpc_order = i;
 434		}
 435	}
 436
 437	return max_usedpc_order;
 438}
 439
 440/*
 441 * A single 'zspage' is composed of many system pages which are
 442 * linked together using fields in struct page. This function finds
 443 * the first/head page, given any component page of a zspage.
 444 */
 445static struct page *get_first_page(struct page *page)
 446{
 447	if (is_first_page(page))
 448		return page;
 449	else
 450		return page->first_page;
 451}
 452
 453static struct page *get_next_page(struct page *page)
 454{
 455	struct page *next;
 
 456
 457	if (is_last_page(page))
 458		next = NULL;
 459	else if (is_first_page(page))
 460		next = (struct page *)page_private(page);
 461	else
 462		next = list_entry(page->lru.next, struct page, lru);
 463
 464	return next;
 
 
 
 
 
 
 
 
 
 
 
 465}
 466
 467/*
 468 * Encode <page, obj_idx> as a single handle value.
 469 * On hardware platforms with physical memory starting at 0x0 the pfn
 470 * could be 0 so we ensure that the handle will never be 0 by adjusting the
 471 * encoded obj_idx value before encoding.
 472 */
 473static void *obj_location_to_handle(struct page *page, unsigned long obj_idx)
 474{
 475	unsigned long handle;
 476
 477	if (!page) {
 478		BUG_ON(obj_idx);
 479		return NULL;
 480	}
 481
 482	handle = page_to_pfn(page) << OBJ_INDEX_BITS;
 483	handle |= ((obj_idx + 1) & OBJ_INDEX_MASK);
 484
 485	return (void *)handle;
 
 
 486}
 487
 488/*
 489 * Decode <page, obj_idx> pair from the given object handle. We adjust the
 490 * decoded obj_idx back to its original value since it was adjusted in
 491 * obj_location_to_handle().
 492 */
 493static void obj_handle_to_location(unsigned long handle, struct page **page,
 494				unsigned long *obj_idx)
 495{
 496	*page = pfn_to_page(handle >> OBJ_INDEX_BITS);
 497	*obj_idx = (handle & OBJ_INDEX_MASK) - 1;
 
 
 
 498}
 499
 500static unsigned long obj_idx_to_offset(struct page *page,
 501				unsigned long obj_idx, int class_size)
 502{
 503	unsigned long off = 0;
 
 504
 505	if (!is_first_page(page))
 506		off = page->index;
 
 
 507
 508	return off + obj_idx * class_size;
 
 
 
 
 
 
 
 509}
 510
 511static void reset_page(struct page *page)
 512{
 513	clear_bit(PG_private, &page->flags);
 514	clear_bit(PG_private_2, &page->flags);
 515	set_page_private(page, 0);
 516	page->mapping = NULL;
 517	page->freelist = NULL;
 518	page_mapcount_reset(page);
 
 
 519}
 520
 521static void free_zspage(struct page *first_page)
 522{
 523	struct page *nextp, *tmp, *head_extra;
 524
 525	BUG_ON(!is_first_page(first_page));
 526	BUG_ON(first_page->inuse);
 
 
 
 
 
 527
 528	head_extra = (struct page *)page_private(first_page);
 
 
 
 
 529
 530	reset_page(first_page);
 531	__free_page(first_page);
 532
 533	/* zspage with only 1 system page */
 534	if (!head_extra)
 535		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 536
 537	list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
 538		list_del(&nextp->lru);
 539		reset_page(nextp);
 540		__free_page(nextp);
 
 
 
 
 
 
 
 
 
 
 
 
 541	}
 542	reset_page(head_extra);
 543	__free_page(head_extra);
 
 544}
 545
 546/* Initialize a newly allocated zspage */
 547static void init_zspage(struct page *first_page, struct size_class *class)
 548{
 
 549	unsigned long off = 0;
 550	struct page *page = first_page;
 551
 552	BUG_ON(!is_first_page(first_page));
 553	while (page) {
 554		struct page *next_page;
 555		struct link_free *link;
 556		unsigned int i, objs_on_page;
 557
 558		/*
 559		 * page->index stores offset of first object starting
 560		 * in the page. For the first page, this is always 0,
 561		 * so we use first_page->index (aka ->freelist) to store
 562		 * head of corresponding zspage's freelist.
 563		 */
 564		if (page != first_page)
 565			page->index = off;
 566
 567		link = (struct link_free *)kmap_atomic(page) +
 568						off / sizeof(*link);
 569		objs_on_page = (PAGE_SIZE - off) / class->size;
 570
 571		for (i = 1; i <= objs_on_page; i++) {
 572			off += class->size;
 573			if (off < PAGE_SIZE) {
 574				link->next = obj_location_to_handle(page, i);
 575				link += class->size / sizeof(*link);
 576			}
 577		}
 578
 579		/*
 580		 * We now come to the last (full or partial) object on this
 581		 * page, which must point to the first object on the next
 582		 * page (if present)
 583		 */
 584		next_page = get_next_page(page);
 585		link->next = obj_location_to_handle(next_page, 0);
 586		kunmap_atomic(link);
 
 
 
 
 
 
 
 
 587		page = next_page;
 588		off = (off + class->size) % PAGE_SIZE;
 589	}
 
 
 590}
 591
 592/*
 593 * Allocate a zspage for the given size class
 594 */
 595static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
 596{
 597	int i, error;
 598	struct page *first_page = NULL, *uninitialized_var(prev_page);
 
 
 599
 600	/*
 601	 * Allocate individual pages and link them together as:
 602	 * 1. first page->private = first sub-page
 603	 * 2. all sub-pages are linked together using page->lru
 604	 * 3. each sub-page is linked to the first page using page->first_page
 605	 *
 606	 * For each size class, First/Head pages are linked together using
 607	 * page->lru. Also, we set PG_private to identify the first page
 608	 * (i.e. no other sub-page has this flag set) and PG_private_2 to
 609	 * identify the last page.
 610	 */
 611	error = -ENOMEM;
 612	for (i = 0; i < class->pages_per_zspage; i++) {
 613		struct page *page;
 614
 615		page = alloc_page(flags);
 616		if (!page)
 617			goto cleanup;
 618
 619		INIT_LIST_HEAD(&page->lru);
 620		if (i == 0) {	/* first page */
 621			SetPagePrivate(page);
 622			set_page_private(page, 0);
 623			first_page = page;
 624			first_page->inuse = 0;
 625		}
 626		if (i == 1)
 627			set_page_private(first_page, (unsigned long)page);
 628		if (i >= 1)
 629			page->first_page = first_page;
 630		if (i >= 2)
 631			list_add(&page->lru, &prev_page->lru);
 632		if (i == class->pages_per_zspage - 1)	/* last page */
 633			SetPagePrivate2(page);
 634		prev_page = page;
 635	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 636
 637	init_zspage(first_page, class);
 
 
 
 
 
 
 
 
 638
 639	first_page->freelist = obj_location_to_handle(first_page, 0);
 640	/* Maximum number of objects we can store in this zspage */
 641	first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
 642
 643	error = 0; /* Success */
 644
 645cleanup:
 646	if (unlikely(error) && first_page) {
 647		free_zspage(first_page);
 648		first_page = NULL;
 649	}
 650
 651	return first_page;
 
 
 
 652}
 653
 654static struct page *find_get_zspage(struct size_class *class)
 655{
 656	int i;
 657	struct page *page;
 658
 659	for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
 660		page = class->fullness_list[i];
 661		if (page)
 
 662			break;
 663	}
 664
 665	return page;
 666}
 667
 668#ifdef CONFIG_PGTABLE_MAPPING
 669static inline int __zs_cpu_up(struct mapping_area *area)
 670{
 671	/*
 672	 * Make sure we don't leak memory if a cpu UP notification
 673	 * and zs_init() race and both call zs_cpu_up() on the same cpu
 674	 */
 675	if (area->vm)
 676		return 0;
 677	area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
 678	if (!area->vm)
 679		return -ENOMEM;
 680	return 0;
 681}
 682
 683static inline void __zs_cpu_down(struct mapping_area *area)
 684{
 685	if (area->vm)
 686		free_vm_area(area->vm);
 687	area->vm = NULL;
 688}
 689
 690static inline void *__zs_map_object(struct mapping_area *area,
 691				struct page *pages[2], int off, int size)
 692{
 693	BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, &pages));
 694	area->vm_addr = area->vm->addr;
 695	return area->vm_addr + off;
 696}
 697
 698static inline void __zs_unmap_object(struct mapping_area *area,
 699				struct page *pages[2], int off, int size)
 700{
 701	unsigned long addr = (unsigned long)area->vm_addr;
 702
 703	unmap_kernel_range(addr, PAGE_SIZE * 2);
 704}
 705
 706#else /* CONFIG_PGTABLE_MAPPING */
 707
 708static inline int __zs_cpu_up(struct mapping_area *area)
 709{
 710	/*
 711	 * Make sure we don't leak memory if a cpu UP notification
 712	 * and zs_init() race and both call zs_cpu_up() on the same cpu
 713	 */
 714	if (area->vm_buf)
 715		return 0;
 716	area->vm_buf = (char *)__get_free_page(GFP_KERNEL);
 717	if (!area->vm_buf)
 718		return -ENOMEM;
 719	return 0;
 720}
 721
 722static inline void __zs_cpu_down(struct mapping_area *area)
 723{
 724	if (area->vm_buf)
 725		free_page((unsigned long)area->vm_buf);
 726	area->vm_buf = NULL;
 727}
 728
 729static void *__zs_map_object(struct mapping_area *area,
 730			struct page *pages[2], int off, int size)
 731{
 732	int sizes[2];
 733	void *addr;
 734	char *buf = area->vm_buf;
 735
 736	/* disable page faults to match kmap_atomic() return conditions */
 737	pagefault_disable();
 738
 739	/* no read fastpath */
 740	if (area->vm_mm == ZS_MM_WO)
 741		goto out;
 742
 743	sizes[0] = PAGE_SIZE - off;
 744	sizes[1] = size - sizes[0];
 745
 746	/* copy object to per-cpu buffer */
 747	addr = kmap_atomic(pages[0]);
 748	memcpy(buf, addr + off, sizes[0]);
 749	kunmap_atomic(addr);
 750	addr = kmap_atomic(pages[1]);
 751	memcpy(buf + sizes[0], addr, sizes[1]);
 752	kunmap_atomic(addr);
 753out:
 754	return area->vm_buf;
 755}
 756
 757static void __zs_unmap_object(struct mapping_area *area,
 758			struct page *pages[2], int off, int size)
 759{
 760	int sizes[2];
 761	void *addr;
 762	char *buf = area->vm_buf;
 763
 764	/* no write fastpath */
 765	if (area->vm_mm == ZS_MM_RO)
 766		goto out;
 767
 
 
 
 
 
 768	sizes[0] = PAGE_SIZE - off;
 769	sizes[1] = size - sizes[0];
 770
 771	/* copy per-cpu buffer to object */
 772	addr = kmap_atomic(pages[0]);
 773	memcpy(addr + off, buf, sizes[0]);
 774	kunmap_atomic(addr);
 775	addr = kmap_atomic(pages[1]);
 776	memcpy(addr, buf + sizes[0], sizes[1]);
 777	kunmap_atomic(addr);
 778
 779out:
 780	/* enable page faults to match kunmap_atomic() return conditions */
 781	pagefault_enable();
 782}
 783
 784#endif /* CONFIG_PGTABLE_MAPPING */
 785
 786static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
 787				void *pcpu)
 788{
 789	int ret, cpu = (long)pcpu;
 790	struct mapping_area *area;
 791
 792	switch (action) {
 793	case CPU_UP_PREPARE:
 794		area = &per_cpu(zs_map_area, cpu);
 795		ret = __zs_cpu_up(area);
 796		if (ret)
 797			return notifier_from_errno(ret);
 798		break;
 799	case CPU_DEAD:
 800	case CPU_UP_CANCELED:
 801		area = &per_cpu(zs_map_area, cpu);
 802		__zs_cpu_down(area);
 803		break;
 804	}
 805
 806	return NOTIFY_OK;
 807}
 808
 809static struct notifier_block zs_cpu_nb = {
 810	.notifier_call = zs_cpu_notifier
 811};
 812
 813static void zs_exit(void)
 814{
 815	int cpu;
 816
 817	cpu_notifier_register_begin();
 818
 819	for_each_online_cpu(cpu)
 820		zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
 821	__unregister_cpu_notifier(&zs_cpu_nb);
 822
 823	cpu_notifier_register_done();
 
 
 824}
 825
 826static int zs_init(void)
 
 827{
 828	int cpu, ret;
 829
 830	cpu_notifier_register_begin();
 831
 832	__register_cpu_notifier(&zs_cpu_nb);
 833	for_each_online_cpu(cpu) {
 834		ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
 835		if (notifier_to_errno(ret)) {
 836			cpu_notifier_register_done();
 837			goto fail;
 838		}
 839	}
 840
 841	cpu_notifier_register_done();
 
 
 
 842
 843	return 0;
 844fail:
 845	zs_exit();
 846	return notifier_to_errno(ret);
 847}
 
 848
 849/**
 850 * zs_create_pool - Creates an allocation pool to work from.
 851 * @flags: allocation flags used to allocate pool metadata
 
 
 852 *
 853 * This function must be called before anything when using
 854 * the zsmalloc allocator.
 
 855 *
 856 * On success, a pointer to the newly created pool is returned,
 857 * otherwise NULL.
 
 
 858 */
 859struct zs_pool *zs_create_pool(gfp_t flags)
 
 860{
 861	int i, ovhd_size;
 862	struct zs_pool *pool;
 
 
 863
 864	ovhd_size = roundup(sizeof(*pool), PAGE_SIZE);
 865	pool = kzalloc(ovhd_size, GFP_KERNEL);
 866	if (!pool)
 867		return NULL;
 
 
 868
 869	for (i = 0; i < ZS_SIZE_CLASSES; i++) {
 870		int size;
 871		struct size_class *class;
 
 
 
 872
 873		size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
 874		if (size > ZS_MAX_ALLOC_SIZE)
 875			size = ZS_MAX_ALLOC_SIZE;
 876
 877		class = &pool->size_class[i];
 878		class->size = size;
 879		class->index = i;
 880		spin_lock_init(&class->lock);
 881		class->pages_per_zspage = get_pages_per_zspage(size);
 
 
 
 
 
 882
 
 
 
 
 
 
 
 883	}
 884
 885	pool->flags = flags;
 
 
 
 
 
 
 
 
 886
 887	return pool;
 888}
 889EXPORT_SYMBOL_GPL(zs_create_pool);
 890
 891void zs_destroy_pool(struct zs_pool *pool)
 892{
 893	int i;
 
 
 
 894
 895	for (i = 0; i < ZS_SIZE_CLASSES; i++) {
 896		int fg;
 897		struct size_class *class = &pool->size_class[i];
 
 898
 899		for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
 900			if (class->fullness_list[fg]) {
 901				pr_info("Freeing non-empty class with size %db, fullness group %d\n",
 902					class->size, fg);
 903			}
 904		}
 
 
 
 
 
 
 
 
 
 
 
 
 905	}
 906	kfree(pool);
 
 
 
 907}
 908EXPORT_SYMBOL_GPL(zs_destroy_pool);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 909
 910/**
 911 * zs_malloc - Allocate block of given size from pool.
 912 * @pool: pool to allocate from
 913 * @size: size of block to allocate
 
 914 *
 915 * On success, handle to the allocated object is returned,
 916 * otherwise 0.
 917 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
 918 */
 919unsigned long zs_malloc(struct zs_pool *pool, size_t size)
 920{
 921	unsigned long obj;
 922	struct link_free *link;
 923	int class_idx;
 924	struct size_class *class;
 925
 926	struct page *first_page, *m_page;
 927	unsigned long m_objidx, m_offset;
 928
 929	if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
 930		return 0;
 931
 932	class_idx = get_size_class_index(size);
 933	class = &pool->size_class[class_idx];
 934	BUG_ON(class_idx != class->index);
 
 
 
 
 935
 936	spin_lock(&class->lock);
 937	first_page = find_get_zspage(class);
 938
 939	if (!first_page) {
 
 
 
 940		spin_unlock(&class->lock);
 941		first_page = alloc_zspage(class, pool->flags);
 942		if (unlikely(!first_page))
 943			return 0;
 944
 945		set_zspage_mapping(first_page, class->index, ZS_EMPTY);
 946		spin_lock(&class->lock);
 947		class->pages_allocated += class->pages_per_zspage;
 948	}
 949
 950	obj = (unsigned long)first_page->freelist;
 951	obj_handle_to_location(obj, &m_page, &m_objidx);
 952	m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
 953
 954	link = (struct link_free *)kmap_atomic(m_page) +
 955					m_offset / sizeof(*link);
 956	first_page->freelist = link->next;
 957	memset(link, POISON_INUSE, sizeof(*link));
 958	kunmap_atomic(link);
 959
 960	first_page->inuse++;
 961	/* Now move the zspage to another fullness group, if required */
 962	fix_fullness_group(pool, first_page);
 963	spin_unlock(&class->lock);
 964
 965	return obj;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 966}
 967EXPORT_SYMBOL_GPL(zs_malloc);
 968
 969void zs_free(struct zs_pool *pool, unsigned long obj)
 970{
 971	struct link_free *link;
 972	struct page *first_page, *f_page;
 973	unsigned long f_objidx, f_offset;
 
 
 
 
 
 
 
 
 974
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 975	int class_idx;
 976	struct size_class *class;
 977	enum fullness_group fullness;
 
 978
 979	if (unlikely(!obj))
 980		return;
 981
 982	obj_handle_to_location(obj, &f_page, &f_objidx);
 983	first_page = get_first_page(f_page);
 
 
 
 
 984
 985	get_zspage_mapping(first_page, &class_idx, &fullness);
 986	class = &pool->size_class[class_idx];
 987	f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
 988
 989	spin_lock(&class->lock);
 
 
 
 
 
 
 990
 991	/* Insert this object in containing zspage's freelist */
 992	link = (struct link_free *)((unsigned char *)kmap_atomic(f_page)
 993							+ f_offset);
 994	link->next = first_page->freelist;
 995	kunmap_atomic(link);
 996	first_page->freelist = (void *)obj;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 997
 998	first_page->inuse--;
 999	fullness = fix_fullness_group(pool, first_page);
1000
1001	if (fullness == ZS_EMPTY)
1002		class->pages_allocated -= class->pages_per_zspage;
1003
1004	spin_unlock(&class->lock);
 
 
 
 
 
 
 
 
 
 
1005
1006	if (fullness == ZS_EMPTY)
1007		free_zspage(first_page);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1008}
1009EXPORT_SYMBOL_GPL(zs_free);
1010
1011/**
1012 * zs_map_object - get address of allocated object from handle.
1013 * @pool: pool from which the object was allocated
1014 * @handle: handle returned from zs_malloc
1015 *
1016 * Before using an object allocated from zs_malloc, it must be mapped using
1017 * this function. When done with the object, it must be unmapped using
1018 * zs_unmap_object.
1019 *
1020 * Only one object can be mapped per cpu at a time. There is no protection
1021 * against nested mappings.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1022 *
1023 * This function returns with preemption and page faults disabled.
1024 */
1025void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1026			enum zs_mapmode mm)
1027{
1028	struct page *page;
1029	unsigned long obj_idx, off;
1030
1031	unsigned int class_idx;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1032	enum fullness_group fg;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1033	struct size_class *class;
1034	struct mapping_area *area;
1035	struct page *pages[2];
 
 
1036
1037	BUG_ON(!handle);
 
 
 
 
 
 
 
1038
1039	/*
1040	 * Because we use per-cpu mapping areas shared among the
1041	 * pools/users, we can't allow mapping in interrupt context
1042	 * because it can corrupt another users mappings.
1043	 */
1044	BUG_ON(in_interrupt());
 
 
 
1045
1046	obj_handle_to_location(handle, &page, &obj_idx);
1047	get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1048	class = &pool->size_class[class_idx];
1049	off = obj_idx_to_offset(page, obj_idx, class->size);
 
1050
1051	area = &get_cpu_var(zs_map_area);
1052	area->vm_mm = mm;
1053	if (off + class->size <= PAGE_SIZE) {
1054		/* this object is contained entirely within a page */
1055		area->vm_addr = kmap_atomic(page);
1056		return area->vm_addr + off;
1057	}
1058
1059	/* this object spans two pages */
1060	pages[0] = page;
1061	pages[1] = get_next_page(page);
1062	BUG_ON(!pages[1]);
 
 
 
 
 
1063
1064	return __zs_map_object(area, pages, off, class->size);
 
 
 
1065}
1066EXPORT_SYMBOL_GPL(zs_map_object);
1067
1068void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
 
1069{
1070	struct page *page;
1071	unsigned long obj_idx, off;
 
 
 
 
 
 
 
 
 
 
1072
1073	unsigned int class_idx;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1074	enum fullness_group fg;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1075	struct size_class *class;
1076	struct mapping_area *area;
 
 
 
 
 
1077
1078	BUG_ON(!handle);
 
 
 
1079
1080	obj_handle_to_location(handle, &page, &obj_idx);
1081	get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1082	class = &pool->size_class[class_idx];
1083	off = obj_idx_to_offset(page, obj_idx, class->size);
1084
1085	area = &__get_cpu_var(zs_map_area);
1086	if (off + class->size <= PAGE_SIZE)
1087		kunmap_atomic(area->vm_addr);
1088	else {
1089		struct page *pages[2];
1090
1091		pages[0] = page;
1092		pages[1] = get_next_page(page);
1093		BUG_ON(!pages[1]);
 
 
 
 
 
 
 
 
 
1094
1095		__zs_unmap_object(area, pages, off, class->size);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1096	}
1097	put_cpu_var(zs_map_area);
 
 
 
 
1098}
1099EXPORT_SYMBOL_GPL(zs_unmap_object);
1100
1101u64 zs_get_total_size_bytes(struct zs_pool *pool)
1102{
1103	int i;
1104	u64 npages = 0;
 
 
 
 
 
 
 
 
 
1105
1106	for (i = 0; i < ZS_SIZE_CLASSES; i++)
1107		npages += pool->size_class[i].pages_allocated;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1108
1109	return npages << PAGE_SHIFT;
1110}
1111EXPORT_SYMBOL_GPL(zs_get_total_size_bytes);
1112
1113module_init(zs_init);
1114module_exit(zs_exit);
1115
1116MODULE_LICENSE("Dual BSD/GPL");
1117MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");