From 6b5f0b0b73a5898bd4f8b2b1d7cf4f693534d176 Mon Sep 17 00:00:00 2001 From: Joonsoo Kim Date: Fri, 12 May 2017 11:30:02 +0900 Subject: [PATCH] zram: implement deduplication in zram This patch implements deduplication feature in zram. The purpose of this work is naturally to save amount of memory usage by zram. Android is one of the biggest users to use zram as swap and it's really important to save amount of memory usage. There is a paper that reports that duplication ratio of Android's memory content is rather high [1]. And, there is a similar work on zswap that also reports that experiments has shown that around 10-15% of pages stored in zswp are duplicates and deduplicate them provides some benefits [2]. Also, there is a different kind of workload that uses zram as blockdev and store build outputs into it to reduce wear-out problem of real blockdev. In this workload, deduplication hit is very high due to temporary files and intermediate object files. Detailed analysis is on the bottom of this description. Anyway, if we can detect duplicated content and avoid to store duplicated content at different memory space, we can save memory. This patch tries to do that. Implementation is almost simple and intuitive but I should note one thing about implementation detail. To check duplication, this patch uses checksum of the page and collision of this checksum could be possible. There would be many choices to handle this situation but this patch chooses to allow entry with duplicated checksum to be added to the hash, but, not to compare all entries with duplicated checksum when checking duplication. I guess that checksum collision is quite rare event and we don't need to pay any attention to such a case. Therefore, I decided the most simplest way to implement the feature. If there is a different opinion, I can accept and go that way. Following is the result of this patch. Test result #1 (Swap): Android Marshmallow, emulator, x86_64, Backporting to kernel v3.18 orig_data_size: 145297408 compr_data_size: 32408125 mem_used_total: 32276480 dup_data_size: 3188134 meta_data_size: 1444272 Last two metrics added to mm_stat are related to this work. First one, dup_data_size, is amount of saved memory by avoiding to store duplicated page. Later one, meta_data_size, is the amount of data structure to support deduplication. If dup > meta, we can judge that the patch improves memory usage. In Adnroid, we can save 5% of memory usage by this work. Test result #2 (Blockdev): build the kernel and store output to ext4 FS on zram Elapsed time: 249 s mm_stat: 430845952 191014886 196898816 0 196898816 28320 0 0 0 Elapsed time: 250 s mm_stat: 430505984 190971334 148365312 0 148365312 28404 0 47287038 3945792 There is no performance degradation and save 23% memory. Test result #3 (Blockdev): copy android build output dir(out/host) to ext4 FS on zram Elapsed time: out/host: 88 s mm_stat: 8834420736 3658184579 3834208256 0 3834208256 32889 0 0 0 Elapsed time: out/host: 100 s mm_stat: 8832929792 3657329322 2832015360 0 2832015360 32609 0 952568877 80880336 It shows performance degradation roughly 13% and save 24% memory. Maybe, it is due to overhead of calculating checksum and comparison. Test result #4 (Blockdev): copy android build output dir(out/target/common) to ext4 FS on zram Elapsed time: out/host: 203 s mm_stat: 4041678848 2310355010 2346577920 0 2346582016 500 4 0 0 Elapsed time: out/host: 201 s mm_stat: 4041666560 2310488276 1338150912 0 1338150912 476 0 989088794 24564336 Memory is saved by 42% and performance is the same. Even if there is overhead of calculating checksum and comparison, large hit ratio compensate it since hit leads to less compression attempt. I checked the detailed reason of savings on kernel build workload and there are some cases that deduplication happens. 1) *.cmd Build command is usually similar in one directory so content of these file are very similar. In my system, more than 789 lines in fs/ext4/.namei.o.cmd and fs/ext4/.inode.o.cmd are the same in 944 and 938 lines of the file, respectively. 2) intermediate object files built-in.o and temporary object file have the similar contents. More than 50% of fs/ext4/ext4.o is the same with fs/ext4/built-in.o. 3) vmlinux .tmp_vmlinux1 and .tmp_vmlinux2 and arch/x86/boo/compressed/vmlinux.bin have the similar contents. Android test has similar case that some of object files(.class and .so) are similar with another ones. (./host/linux-x86/lib/libartd.so and ./host/linux-x86-lib/libartd-comiler.so) Anyway, benefit seems to be largely dependent on the workload so following patch will make this feature optional. However, this feature can help some usecases so is deserved to be merged. [1]: MemScope: Analyzing Memory Duplication on Android Systems, dl.acm.org/citation.cfm?id=2797023 [2]: zswap: Optimize compressed pool memory utilization, lkml.kernel.org/r/1341407574.7551.1471584870761.JavaMail.weblogic@epwas3p2 Change-Id: I8fe80c956c33f88a6af337d50d9e210e5c35ce37 Reviewed-by: Sergey Senozhatsky Acked-by: Minchan Kim Signed-off-by: Joonsoo Kim Link: https://lore.kernel.org/patchwork/patch/787162/ Patch-mainline: linux-kernel@ Thu, 11 May 2017 22:30:26 Signed-off-by: Charan Teja Reddy Signed-off-by: Park Ju Hyung Signed-off-by: Alexander Winkowski --- drivers/block/zram/Makefile | 2 +- drivers/block/zram/zram_dedup.c | 204 ++++++++++++++++++++++++++++++++ drivers/block/zram/zram_dedup.h | 22 ++++ drivers/block/zram/zram_drv.c | 38 +++++- drivers/block/zram/zram_drv.h | 20 ++++ 5 files changed, 279 insertions(+), 7 deletions(-) create mode 100644 drivers/block/zram/zram_dedup.c create mode 100644 drivers/block/zram/zram_dedup.h diff --git a/drivers/block/zram/Makefile b/drivers/block/zram/Makefile index de9e457907b1e..72cc66ffb5e8a 100644 --- a/drivers/block/zram/Makefile +++ b/drivers/block/zram/Makefile @@ -1,4 +1,4 @@ # SPDX-License-Identifier: GPL-2.0-only -zram-y := zcomp.o zram_drv.o +zram-y := zcomp.o zram_drv.o zram_dedup.o obj-$(CONFIG_ZRAM) += zram.o diff --git a/drivers/block/zram/zram_dedup.c b/drivers/block/zram/zram_dedup.c new file mode 100644 index 0000000000000..a8427f75b6ea9 --- /dev/null +++ b/drivers/block/zram/zram_dedup.c @@ -0,0 +1,204 @@ +/* + * Copyright (C) 2017 Joonsoo Kim. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation; either version + * 2 of the License, or (at your option) any later version. + */ + +#include +#include +#include + +#include "zram_drv.h" + +/* One slot will contain 128 pages theoretically */ +#define ZRAM_HASH_SHIFT 7 +#define ZRAM_HASH_SIZE_MIN (1 << 10) +#define ZRAM_HASH_SIZE_MAX (1 << 31) + +u64 zram_dedup_dup_size(struct zram *zram) +{ + return (u64)atomic64_read(&zram->stats.dup_data_size); +} + +u64 zram_dedup_meta_size(struct zram *zram) +{ + return (u64)atomic64_read(&zram->stats.meta_data_size); +} + +static u32 zram_dedup_checksum(unsigned char *mem) +{ + return jhash(mem, PAGE_SIZE, 0); +} + +void zram_dedup_insert(struct zram *zram, struct zram_entry *new, + u32 checksum) +{ + struct zram_hash *hash; + struct rb_root *rb_root; + struct rb_node **rb_node, *parent = NULL; + struct zram_entry *entry; + + new->checksum = checksum; + hash = &zram->hash[checksum % zram->hash_size]; + rb_root = &hash->rb_root; + + spin_lock(&hash->lock); + rb_node = &rb_root->rb_node; + while (*rb_node) { + parent = *rb_node; + entry = rb_entry(parent, struct zram_entry, rb_node); + if (checksum < entry->checksum) + rb_node = &parent->rb_left; + else if (checksum > entry->checksum) + rb_node = &parent->rb_right; + else + rb_node = &parent->rb_left; + } + + rb_link_node(&new->rb_node, parent, rb_node); + rb_insert_color(&new->rb_node, rb_root); + spin_unlock(&hash->lock); +} + +static bool zram_dedup_match(struct zram *zram, struct zram_entry *entry, + unsigned char *mem) +{ + bool match = false; + unsigned char *cmem; + struct zcomp_strm *zstrm; + + cmem = zs_map_object(zram->mem_pool, entry->handle, ZS_MM_RO); + if (entry->len == PAGE_SIZE) { + match = !memcmp(mem, cmem, PAGE_SIZE); + } else { + zstrm = zcomp_stream_get(zram->comp); + if (!zcomp_decompress(zstrm, cmem, entry->len, zstrm->buffer)) + match = !memcmp(mem, zstrm->buffer, PAGE_SIZE); + zcomp_stream_put(zram->comp); + } + zs_unmap_object(zram->mem_pool, entry->handle); + + return match; +} + +static unsigned long zram_dedup_put(struct zram *zram, + struct zram_entry *entry) +{ + struct zram_hash *hash; + u32 checksum; + + checksum = entry->checksum; + hash = &zram->hash[checksum % zram->hash_size]; + + spin_lock(&hash->lock); + + entry->refcount--; + if (!entry->refcount) + rb_erase(&entry->rb_node, &hash->rb_root); + else + atomic64_sub(entry->len, &zram->stats.dup_data_size); + + spin_unlock(&hash->lock); + + return entry->refcount; +} + +static struct zram_entry *zram_dedup_get(struct zram *zram, + unsigned char *mem, u32 checksum) +{ + struct zram_hash *hash; + struct zram_entry *entry; + struct rb_node *rb_node; + + hash = &zram->hash[checksum % zram->hash_size]; + + spin_lock(&hash->lock); + rb_node = hash->rb_root.rb_node; + while (rb_node) { + entry = rb_entry(rb_node, struct zram_entry, rb_node); + if (checksum == entry->checksum) { + entry->refcount++; + atomic64_add(entry->len, &zram->stats.dup_data_size); + spin_unlock(&hash->lock); + + if (zram_dedup_match(zram, entry, mem)) + return entry; + + zram_entry_free(zram, entry); + + return NULL; + } + + if (checksum < entry->checksum) + rb_node = rb_node->rb_left; + else + rb_node = rb_node->rb_right; + } + spin_unlock(&hash->lock); + + return NULL; +} + +struct zram_entry *zram_dedup_find(struct zram *zram, struct page *page, + u32 *checksum) +{ + void *mem; + struct zram_entry *entry; + + mem = kmap_atomic(page); + *checksum = zram_dedup_checksum(mem); + + entry = zram_dedup_get(zram, mem, *checksum); + kunmap_atomic(mem); + + return entry; +} + +void zram_dedup_init_entry(struct zram *zram, struct zram_entry *entry, + unsigned long handle, unsigned int len) +{ + entry->handle = handle; + entry->refcount = 1; + entry->len = len; +} + +bool zram_dedup_put_entry(struct zram *zram, struct zram_entry *entry) +{ + if (zram_dedup_put(zram, entry)) + return false; + + return true; +} + +int zram_dedup_init(struct zram *zram, size_t num_pages) +{ + int i; + struct zram_hash *hash; + + zram->hash_size = num_pages >> ZRAM_HASH_SHIFT; + zram->hash_size = min_t(size_t, ZRAM_HASH_SIZE_MAX, zram->hash_size); + zram->hash_size = max_t(size_t, ZRAM_HASH_SIZE_MIN, zram->hash_size); + zram->hash = vzalloc(zram->hash_size * sizeof(struct zram_hash)); + if (!zram->hash) { + pr_err("Error allocating zram entry hash\n"); + return -ENOMEM; + } + + for (i = 0; i < zram->hash_size; i++) { + hash = &zram->hash[i]; + spin_lock_init(&hash->lock); + hash->rb_root = RB_ROOT; + } + + return 0; +} + +void zram_dedup_fini(struct zram *zram) +{ + vfree(zram->hash); + zram->hash = NULL; + zram->hash_size = 0; +} diff --git a/drivers/block/zram/zram_dedup.h b/drivers/block/zram/zram_dedup.h new file mode 100644 index 0000000000000..ebe6bff6c0dac --- /dev/null +++ b/drivers/block/zram/zram_dedup.h @@ -0,0 +1,22 @@ +#ifndef _ZRAM_DEDUP_H_ +#define _ZRAM_DEDUP_H_ + +struct zram; +struct zram_entry; + +u64 zram_dedup_dup_size(struct zram *zram); +u64 zram_dedup_meta_size(struct zram *zram); + +void zram_dedup_insert(struct zram *zram, struct zram_entry *new, + u32 checksum); +struct zram_entry *zram_dedup_find(struct zram *zram, struct page *page, + u32 *checksum); + +void zram_dedup_init_entry(struct zram *zram, struct zram_entry *entry, + unsigned long handle, unsigned int len); +bool zram_dedup_put_entry(struct zram *zram, struct zram_entry *entry); + +int zram_dedup_init(struct zram *zram, size_t num_pages); +void zram_dedup_fini(struct zram *zram); + +#endif /* _ZRAM_DEDUP_H_ */ diff --git a/drivers/block/zram/zram_drv.c b/drivers/block/zram/zram_drv.c index 4efed2c0ef510..31c355166ca1e 100644 --- a/drivers/block/zram/zram_drv.c +++ b/drivers/block/zram/zram_drv.c @@ -1112,7 +1112,7 @@ static ssize_t mm_stat_show(struct device *dev, max_used = atomic_long_read(&zram->stats.max_used_pages); ret = scnprintf(buf, PAGE_SIZE, - "%8llu %8llu %8llu %8lu %8ld %8llu %8lu %8llu %8llu\n", + "%8llu %8llu %8llu %8lu %8ld %8llu %8lu %8llu %8llu %8llu %8llu\n", orig_size << PAGE_SHIFT, (u64)atomic64_read(&zram->stats.compr_data_size), mem_used << PAGE_SHIFT, @@ -1121,7 +1121,9 @@ static ssize_t mm_stat_show(struct device *dev, (u64)atomic64_read(&zram->stats.same_pages), atomic_long_read(&pool_stats.pages_compacted), (u64)atomic64_read(&zram->stats.huge_pages), - (u64)atomic64_read(&zram->stats.huge_pages_since)); + (u64)atomic64_read(&zram->stats.huge_pages_since), + zram_dedup_dup_size(zram), + zram_dedup_meta_size(zram)); up_read(&zram->init_lock); return ret; @@ -1176,26 +1178,34 @@ static struct zram_entry *zram_entry_alloc(struct zram *zram, unsigned int len, gfp_t flags) { struct zram_entry *entry; + unsigned long handle; entry = kzalloc(sizeof(*entry), flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE)); if (!entry) return NULL; - entry->handle = zs_malloc(zram->mem_pool, len, flags); - if (!entry->handle) { + handle = zs_malloc(zram->mem_pool, len, flags); + if (!handle) { kfree(entry); return NULL; } + zram_dedup_init_entry(zram, entry, handle, len); + atomic64_add(sizeof(*entry), &zram->stats.meta_data_size); + return entry; } -static inline void zram_entry_free(struct zram *zram, - struct zram_entry *entry) +void zram_entry_free(struct zram *zram, struct zram_entry *entry) { + if (!zram_dedup_put_entry(zram, entry)) + return; + zs_free(zram->mem_pool, entry->handle); kfree(entry); + + atomic64_sub(sizeof(*entry), &zram->stats.meta_data_size); } static void zram_meta_free(struct zram *zram, u64 disksize) @@ -1208,6 +1218,7 @@ static void zram_meta_free(struct zram *zram, u64 disksize) zram_free_page(zram, index); zs_destroy_pool(zram->mem_pool); + zram_dedup_fini(zram); vfree(zram->table); } @@ -1228,6 +1239,13 @@ static bool zram_meta_alloc(struct zram *zram, u64 disksize) if (!huge_class_size) huge_class_size = zs_huge_class_size(zram->mem_pool); + + if (zram_dedup_init(zram, num_pages)) { + vfree(zram->table); + zs_destroy_pool(zram->mem_pool); + return false; + } + return true; } @@ -1389,6 +1407,7 @@ static int __zram_bvec_write(struct zram *zram, struct bio_vec *bvec, void *src, *dst, *mem; struct zcomp_strm *zstrm; struct page *page = bvec->bv_page; + u32 checksum; unsigned long element = 0; enum zram_pageflags flags = 0; @@ -1402,6 +1421,12 @@ static int __zram_bvec_write(struct zram *zram, struct bio_vec *bvec, } kunmap_atomic(mem); + entry = zram_dedup_find(zram, page, &checksum); + if (entry) { + comp_len = entry->len; + goto out; + } + compress_again: zstrm = zcomp_stream_get(zram->comp); src = kmap_atomic(page); @@ -1469,6 +1494,7 @@ static int __zram_bvec_write(struct zram *zram, struct bio_vec *bvec, zcomp_stream_put(zram->comp); zs_unmap_object(zram->mem_pool, entry->handle); atomic64_add(comp_len, &zram->stats.compr_data_size); + zram_dedup_insert(zram, entry, checksum); out: /* * Free memory associated with this sector diff --git a/drivers/block/zram/zram_drv.h b/drivers/block/zram/zram_drv.h index b00bc8efc2f4f..1c8bd0b2b438d 100644 --- a/drivers/block/zram/zram_drv.h +++ b/drivers/block/zram/zram_drv.h @@ -18,8 +18,10 @@ #include #include #include +#include #include "zcomp.h" +#include "zram_dedup.h" #define SECTORS_PER_PAGE_SHIFT (PAGE_SHIFT - SECTOR_SHIFT) #define SECTORS_PER_PAGE (1 << SECTORS_PER_PAGE_SHIFT) @@ -57,6 +59,10 @@ enum zram_pageflags { /*-- Data structures */ struct zram_entry { + struct rb_node rb_node; + u32 len; + u32 checksum; + unsigned long refcount; unsigned long handle; }; @@ -92,6 +98,16 @@ struct zram_stats { atomic64_t bd_reads; /* no. of reads from backing device */ atomic64_t bd_writes; /* no. of writes from backing device */ #endif + atomic64_t dup_data_size; /* + * compressed size of pages + * duplicated + */ + atomic64_t meta_data_size; /* size of zram_entries */ +}; + +struct zram_hash { + spinlock_t lock; + struct rb_root rb_root; }; struct zram { @@ -99,6 +115,8 @@ struct zram { struct zs_pool *mem_pool; struct zcomp *comp; struct gendisk *disk; + struct zram_hash *hash; + size_t hash_size; /* Prevent concurrent execution of device init */ struct rw_semaphore init_lock; /* @@ -130,4 +148,6 @@ struct zram { struct dentry *debugfs_dir; #endif }; + +void zram_entry_free(struct zram *zram, struct zram_entry *entry); #endif