Pool for compression pages
Subsystems (2)
David Sterba
maintainercover Pool for compression pages
Add wrappers for alloc/free of pages used for compresion and keep a small cache (for all filesystems, ie. per module). This can speed up short compression IO when we don't need to go to the allocator. Shrinker is attached to free the pages if there's memory pressure.
David Sterba (2):
btrfs: use page alloc/free wrapeprs for compression pages
btrfs: use shrinker for compression page pool
fs/btrfs/compression.c | 118 ++++++++++++++++++++++++++++++++++++++++- fs/btrfs/compression.h | 5 ++ fs/btrfs/inode.c | 4 +- fs/btrfs/lzo.c | 4 +- fs/btrfs/zlib.c | 6 +-- fs/btrfs/zstd.c | 7 ++- 6 files changed, 132 insertions(+), 12 deletions(-)
Signature · 1 line
2.42.1
David Sterba
maintainerPATCH 1/2 btrfs: use page alloc/free wrapeprs for compression pages
This is a preparation for managing compression pages in a cache-like manner, instead of asking the allocator each time. The common allocation and free wrappers are introduced and are functionally equivalent to the current code.
The freeing helpers need to be carefully placed where the last reference is dropped. This is either after directly allocating (error handling) or when there are no other users of the pages (after copying the contents).
It's safe to not use the helper and use put_page() that will handle the reference count. Not using the helper means there's lower number of pages that could be reused without passing them back to allocator.
--- fs/btrfs/compression.c | 16 +++++++++++++++- fs/btrfs/compression.h | 5 +++++ fs/btrfs/inode.c | 4 ++-- fs/btrfs/lzo.c | 4 ++-- fs/btrfs/zlib.c | 6 +++--- fs/btrfs/zstd.c | 7 +++---- 6 files changed, 30 insertions(+), 12 deletions(-)
fs/btrfs/compression.c +15 -1
@@ -163,12 +163,26 @@ static int compression_decompress(int type, struct list_head *ws,static void btrfs_free_compressed_pages (struct compressed_bio *cb) {for (unsigned int i =0 ; i < cb->nr_pages; i++) -put_page (cb->compressed_pages[i]); +btrfs_free_compr_page (cb->compressed_pages[i]);kfree (cb->compressed_pages); }static int btrfs_decompress_bio (struct compressed_bio *cb); +/* +* Common wrappers for page allocation from compression wrappers +*/ +struct page *btrfs_alloc_compr_page (void ) +{ +return alloc_page (GFP_NOFS ); +} + +void btrfs_free_compr_page (struct page *page) +{ +ASSERT (page_ref_count (page) ==1 ); +put_page (page); +} +static void end_compressed_bio_read (struct btrfs_bio *bbio) {struct compressed_bio *cb =to_compressed_bio (bbio);
fs/btrfs/compression.h +5 -0
@@ -32,6 +32,8 @@ static_assert((BTRFS_MAX_COMPRESSED % PAGE_SIZE) == 0);#define BTRFS_ZLIB_DEFAULT_LEVEL 3 +struct page ; +struct compressed_bio {/* Number of compressed pages in the array */ unsigned int nr_pages;
@@ -96,6 +98,9 @@ void btrfs_submit_compressed_read(struct btrfs_bio *bbio);unsigned int btrfs_compress_str2level (unsigned int type,const char *str); +struct page *btrfs_alloc_compr_page (void ); +void btrfs_free_compr_page (struct page *page); +enum btrfs_compression_type {BTRFS_COMPRESS_NONE =0 ,BTRFS_COMPRESS_ZLIB =1 ,
fs/btrfs/inode.c +2 -2
@@ -1037,7 +1037,7 @@ static void compress_file_range(struct btrfs_work *work)if (pages) {for (i =0 ; i < nr_pages; i++) {WARN_ON (pages[i]->mapping); -put_page (pages[i]); +btrfs_free_compr_page (pages[i]); }kfree (pages); }
@@ -1052,7 +1052,7 @@ static void free_async_extent_pages(struct async_extent *async_extent)for (i =0 ; i < async_extent->nr_pages; i++) {WARN_ON (async_extent->pages[i]->mapping); -put_page (async_extent->pages[i]); +btrfs_free_compr_page (async_extent->pages[i]); }kfree (async_extent->pages); async_extent->nr_pages =0 ;
fs/btrfs/lzo.c +2 -2
@@ -152,7 +152,7 @@ static int copy_compressed_data_to_page(char *compressed_data, cur_page = out_pages[*cur_out /PAGE_SIZE ];/* Allocate a new page */ if (!cur_page) { - cur_page =alloc_page (GFP_NOFS ); + cur_page =btrfs_alloc_compr_page ();if (!cur_page)return -ENOMEM ; out_pages[*cur_out /PAGE_SIZE ] = cur_page;
@@ -178,7 +178,7 @@ static int copy_compressed_data_to_page(char *compressed_data, cur_page = out_pages[*cur_out /PAGE_SIZE ];/* Allocate a new page */ if (!cur_page) { - cur_page =alloc_page (GFP_NOFS ); + cur_page =btrfs_alloc_compr_page ();if (!cur_page)return -ENOMEM ; out_pages[*cur_out /PAGE_SIZE ] = cur_page;
fs/btrfs/zlib.c +3 -3
@@ -121,7 +121,7 @@ int zlib_compress_pages(struct list_head *ws, struct address_space *mapping, workspace->strm.total_in =0 ; workspace->strm.total_out =0 ; - out_page =alloc_page (GFP_NOFS ); + out_page =btrfs_alloc_compr_page ();if (out_page ==NULL ) { ret = -ENOMEM ; goto out;
@@ -200,7 +200,7 @@ int zlib_compress_pages(struct list_head *ws, struct address_space *mapping, ret = -E2BIG ; goto out; } - out_page =alloc_page (GFP_NOFS ); + out_page =btrfs_alloc_compr_page ();if (out_page ==NULL ) { ret = -ENOMEM ; goto out;
@@ -236,7 +236,7 @@ int zlib_compress_pages(struct list_head *ws, struct address_space *mapping, ret = -E2BIG ; goto out; } - out_page =alloc_page (GFP_NOFS ); + out_page =btrfs_alloc_compr_page ();if (out_page ==NULL ) { ret = -ENOMEM ; goto out;
fs/btrfs/zstd.c +3 -4
@@ -410,9 +410,8 @@ int zstd_compress_pages(struct list_head *ws, struct address_space *mapping, workspace->in_buf.pos =0 ; workspace->in_buf.size =min_t (size_t, len,PAGE_SIZE ); -/* Allocate and map in the output buffer */ - out_page =alloc_page (GFP_NOFS ); + out_page =btrfs_alloc_compr_page ();if (out_page ==NULL ) { ret = -ENOMEM ; goto out;
@@ -457,7 +456,7 @@ int zstd_compress_pages(struct list_head *ws, struct address_space *mapping, ret = -E2BIG ; goto out; } - out_page =alloc_page (GFP_NOFS ); + out_page =btrfs_alloc_compr_page ();if (out_page ==NULL ) { ret = -ENOMEM ; goto out;
@@ -514,7 +513,7 @@ int zstd_compress_pages(struct list_head *ws, struct address_space *mapping, ret = -E2BIG ; goto out; } - out_page =alloc_page (GFP_NOFS ); + out_page =btrfs_alloc_compr_page ();if (out_page ==NULL ) { ret = -ENOMEM ; goto out;
Signature · 1 line
2.42.1
David Sterba
maintainerPATCH 2/2 btrfs: use shrinker for compression page pool
In reply to David SterbaThe pages are now allocated and freed centrally, so we can extend the logic to manage the lifetime. The main idea is to keep a few recently used pages and hand them to all writers. Ideally we won't have to go to allocator at all (a slight performance gain) and also raise chance that we'll have the pages available (slightly increased reliability).
In order to avoid gathering too many pages, the shrinker is attached to the cache so we can free them on when MM demands that. The first implementation will drain the whole cache. Further this can be refined to keep some minimal number of pages for emergency purposes. The ultimate goal to avoid memory allocation failures on the write out path from the compression.
The pool threshold is set to cover full BTRFS_MAX_COMPRESSED / PAGE_SIZE for minimal thread pool, which is 8 (btrfs_init_fs_info()). This is 128K / 4K * 8 = 256 pages at maximum, which is 1MiB.
This is for all filesystems currently mounted, with heavy use of compression IO the allocator is still needed. The cache helps for short burst IO.
--- fs/btrfs/compression.c | 102 +++++++++++++++++++++++++++++++++++++++++ 1 file changed, 102 insertions(+)
fs/btrfs/compression.c +102 -0
@@ -20,6 +20,7 @@#include <linux/slab.h> #include <linux/sched/mm.h> #include <linux/log2.h> +#include <linux/shrinker.h> #include <crypto/hash.h> #include "misc.h" #include "ctree.h"
@@ -169,16 +170,96 @@ static void btrfs_free_compressed_pages(struct compressed_bio *cb)static int btrfs_decompress_bio (struct compressed_bio *cb); +/* +* Global cache of last unused pages for compression/decompression. +*/ +static struct btrfs_compr_pool { +struct shrinker *shrinker; +spinlock_t lock; +struct list_head list; +int count; +int thresh; +} compr_pool; + +static unsigned long btrfs_compr_pool_count (struct shrinker *sh,struct shrink_control *sc) +{ +int ret; + +/* +* We must not read the values more than once if 'ret' gets expanded in +* the return statement so we don't accidentally return a negative +* number, even if the first condition finds it positive. +*/ + ret =READ_ONCE (compr_pool.count) -READ_ONCE (compr_pool.thresh); + +return ret >0 ? ret :0 ; +} + +static unsigned long btrfs_compr_pool_scan (struct shrinker *sh,struct shrink_control *sc) +{ +struct list_head remove; +struct list_head *tmp, *next; +int freed; + +if (compr_pool.count ==0 ) +return SHRINK_STOP ; + +INIT_LIST_HEAD (&remove); + +/* For now, just simply drain the whole list. */ +spin_lock (&compr_pool.lock); +list_splice_init (&compr_pool.list, &remove); + freed = compr_pool.count; + compr_pool.count =0 ; +spin_unlock (&compr_pool.lock); + +list_for_each_safe (tmp, next, &remove) { + struct page *page = list_entry(tmp, struct page, lru); + +ASSERT (page_ref_count (page) ==1 ); +put_page (page); + } + +return freed; +} +/* * Common wrappers for page allocation from compression wrappers */ struct page *btrfs_alloc_compr_page (void ) { +struct page *page =NULL ; + +spin_lock (&compr_pool.lock); +if (compr_pool.count >0 ) { + page = list_first_entry(&compr_pool.list, struct page, lru); +list_del_init (&page->lru); + compr_pool.count--; + } +spin_unlock (&compr_pool.lock); + +if (page) +return page; +return alloc_page (GFP_NOFS ); }void btrfs_free_compr_page (struct page *page) { +bool do_free = false; + +spin_lock (&compr_pool.lock); +if (compr_pool.count > compr_pool.thresh) { + do_free = true; + }else { +list_add (&page->lru, &compr_pool.list); + compr_pool.count++; + } +spin_unlock (&compr_pool.lock); + +if (!do_free) +return ; +ASSERT (page_ref_count (page) ==1 );put_page (page); }
@@ -974,15 +1055,36 @@ int __init btrfs_init_compress(void) offsetof(struct compressed_bio, bbio.bio), BIOSET_NEED_BVECS))return -ENOMEM ; + + compr_pool.shrinker =shrinker_alloc (SHRINKER_NONSLAB ,"btrfs-compr-pages" ); +if (!compr_pool.shrinker) +return -ENOMEM ; +btrfs_init_workspace_manager (BTRFS_COMPRESS_NONE );btrfs_init_workspace_manager (BTRFS_COMPRESS_ZLIB );btrfs_init_workspace_manager (BTRFS_COMPRESS_LZO );zstd_init_workspace_manager (); + +spin_lock_init (&compr_pool.lock); +INIT_LIST_HEAD (&compr_pool.list); + compr_pool.count =0 ; +/* 128K / 4K = 32, for 8 threads is 256 pages. */ + compr_pool.thresh =BTRFS_MAX_COMPRESSED /PAGE_SIZE *8 ; + compr_pool.shrinker->count_objects = btrfs_compr_pool_count; + compr_pool.shrinker->scan_objects = btrfs_compr_pool_scan; + compr_pool.shrinker->batch =32 ; + compr_pool.shrinker->seeks =DEFAULT_SEEKS ; +shrinker_register (compr_pool.shrinker); +return 0 ; }void __cold btrfs_exit_compress (void ) { +/* For now scan drains all pages and does not touch the parameters. */ +btrfs_compr_pool_scan (NULL ,NULL ); +shrinker_free (compr_pool.shrinker); +btrfs_cleanup_workspace_manager (BTRFS_COMPRESS_NONE );btrfs_cleanup_workspace_manager (BTRFS_COMPRESS_ZLIB );btrfs_cleanup_workspace_manager (BTRFS_COMPRESS_LZO );
Signature · 1 line
2.42.1
Re: [PATCH 1/2] btrfs: use page alloc/free wrapeprs for compression pages
In reply to David Sterbafs/btrfs/compression.c static int compression_decompress(int type, struct list_head *ws, · line 183
Quoted 4 lines · David Sterba
fs/btrfs/compression.c +4
-0+void btrfs_free_compr_page (struct page *page) +{ +ASSERT (page_ref_count (page) ==1 ); +put_page (page);
Out of curiosity, why the ASSERT()?
Re: [PATCH 1/2] btrfs: use page alloc/free wrapeprs for compression pages
Quoted 7 lines · Johannes Thumshirn
fs/btrfs/compression.c static int compression_decompress(int type, struct list_head *ws, · line 183
Quoted 4 lines · David Sterba
fs/btrfs/compression.c +4
-0+void btrfs_free_compr_page (struct page *page) +{ +ASSERT (page_ref_count (page) ==1 ); +put_page (page);Out of curiosity, why the
ASSERT()?
To verify that it's the last reference and put_page is going to free it , so we don't put an actually used page into the cache.
Quoted 9 lines · David Sterba
Add wrappers for alloc/free of pages used for compresion and keep a small cache (for all filesystems, ie. per module). This can speed up short compression IO when we don't need to go to the allocator. Shrinker is attached to free the pages if there's memory pressure.
David Sterba (2):
btrfs: use page alloc/free wrapeprs for compression pages
btrfs: use shrinker for compression page pool
Sorry I thought I responded to this last week,
Thanks,
Josef