udf: add extent cache support in case of file reading
Subsystems (2)
PATCH 1/1 udf: add extent cache support in case of file reading
From: Namjae Jeon <[email protected]>
This patch implements extent caching in case of file reading. While reading a file, currently, UDF reads metadata serially which takes a lot of time depending on the number of extents present in the file. Caching last accessd extent improves metadata read time. Instead of reading file metadata from start, now we read from the cached extent.
This patch considerably improves the time spent by CPU in kernel mode. For example, while reading a 10.9 GB file using dd:
Time before applying patch:
11677022208 bytes (10.9GB) copied, 1529.748921 seconds, 7.3MB/s
real 25m 29.85s user 0m 12.41s sys 15m 34.75s
Time after applying patch:
11677022208 bytes (10.9GB) copied, 1469.338231 seconds, 7.6MB/s
real 24m 29.44s user 0m 15.73s sys 3m 27.61s
--- fs/udf/ialloc.c | 4 +++ fs/udf/inode.c | 79 +++++++++++++++++++++++++++++++++++++++++++++++++----- fs/udf/udf_i.h | 16 +++++++++++ fs/udf/udfdecl.h | 10 +++---- 4 files changed, 98 insertions(+), 11 deletions(-)
fs/udf/ialloc.c +4 -0
@@ -117,6 +117,10 @@ struct inode *udf_new_inode(struct inode *dir, umode_t mode, int *err) iinfo->i_lenAlloc =0 ; iinfo->i_use =0 ; iinfo->i_checkpoint =1 ; +memset (&iinfo->cached_extent,0 ,sizeof (struct udf_ext_cache )); +spin_lock_init (&(iinfo->i_extent_cache_lock)); +/* Mark extent cache as invalid for now */ + iinfo->cached_extent.lstart =-1 ;if (UDF_QUERY_FLAG (inode->i_sb,UDF_FLAG_USE_AD_IN_ICB )) iinfo->i_alloc_type =ICBTAG_FLAG_AD_IN_ICB ;else if (UDF_QUERY_FLAG (inode->i_sb,UDF_FLAG_USE_SHORT_AD ))
fs/udf/inode.c +73 -6
131 lines not on this page · raw message
fs/udf/udf_i.h +16 -0
@@ -1,6 +1,19 @@#ifndef _UDF_I_H#define _UDF_I_H +struct extent_position { +struct buffer_head *bh; +uint32_t offset; +struct kernel_lb_addr block; +}; + +struct udf_ext_cache { +/* Extent position */ +struct extent_position epos; +/* Start logical offset in bytes */ +loff_t lstart; +}; +/* * The i_data_sem and i_mutex serve for protection of allocation information * of a regular files and symlinks. This includes all extents belonging to
@@ -35,6 +48,9 @@ struct udf_inode_info {__u8 *i_data; } i_ext;struct rw_semaphore i_data_sem; +struct udf_ext_cache cached_extent; +/* Spinlock for protecting extent cache */ +spinlock_t i_extent_cache_lock;struct inode vfs_inode; };
fs/udf/udfdecl.h +5 -5
@@ -113,11 +113,6 @@ struct ustr {uint8_t u_len; }; -struct extent_position { -struct buffer_head *bh; -uint32_t offset; -struct kernel_lb_addr block; -};/* super.c */
@@ -164,6 +159,11 @@ extern int8_t udf_next_aext(struct inode *, struct extent_position *, struct kernel_lb_addr *, uint32_t *, int);extern int8_t udf_current_aext (struct inode *,struct extent_position *,struct kernel_lb_addr *,uint32_t *,int ); +int udf_read_extent_cache (struct inode *inode,loff_t bcount,loff_t *lbcount, +struct extent_position *pos); +void udf_update_extent_cache (struct inode *inode,loff_t estart, +struct extent_position *pos,int next_epos); +void udf_clear_extent_cache (struct udf_inode_info *iinfo);/* misc.c */ extern struct buffer_head *udf_tgetblk (struct super_block *,int );
Signature · 1 line
1.7.9.5
Quoted 17 lines · Namjae Jeon
From: Namjae Jeon <[email protected]>
This patch implements extent caching in case of file reading. While reading a file, currently, UDF reads metadata serially which takes a lot of time depending on the number of extents present in the file. Caching last accessd extent improves metadata read time. Instead of reading file metadata from start, now we read from the cached extent.
This patch considerably improves the time spent by CPU in kernel mode. For example, while reading a 10.9 GB file using dd:
Time before applying patch:
11677022208 bytes (10.9GB) copied, 1529.748921 seconds, 7.3MB/sreal 25m 29.85s user 0m 12.41s sys 15m 34.75s
Quoted 5 lines
Time after applying patch:
11677022208 bytes (10.9GB) copied, 1469.338231 seconds, 7.6MB/sreal 24m 29.44s user 0m 15.73s sys 3m 27.61s
did you have any test on lots of small files?
- cong
Quoted 240 lines
Signed-off-by Namjae Jeon <[email protected]>Signed-off-by Ashish Sangwan <[email protected]>Signed-off-by Bonggil Bak <[email protected]>--- fs/udf/ialloc.c | 4 +++ fs/udf/inode.c | 79 +++++++++++++++++++++++++++++++++++++++++++++++++----- fs/udf/udf_i.h | 16 +++++++++++ fs/udf/udfdecl.h | 10 +++---- 4 files changed, 98 insertions(+), 11 deletions(-)fs/udf/ialloc.c +4
-0@@ -117,6 +117,10 @@ struct inode *udf_new_inode(struct inode *dir, umode_t mode, int *err) iinfo->i_lenAlloc =0 ; iinfo->i_use =0 ; iinfo->i_checkpoint =1 ; +memset (&iinfo->cached_extent,0 ,sizeof (struct udf_ext_cache )); +spin_lock_init (&(iinfo->i_extent_cache_lock)); +/* Mark extent cache as invalid for now */ + iinfo->cached_extent.lstart =-1 ;if (UDF_QUERY_FLAG (inode->i_sb,UDF_FLAG_USE_AD_IN_ICB )) iinfo->i_alloc_type =ICBTAG_FLAG_AD_IN_ICB ;else if (UDF_QUERY_FLAG (inode->i_sb,UDF_FLAG_USE_SHORT_AD ))fs/udf/inode.c +73
-6@@ -91,6 +91,7 @@ void udf_evict_inode(struct inode *inode) }kfree (iinfo->i_ext.i_data); iinfo->i_ext.i_data =NULL ; +udf_clear_extent_cache (iinfo);if (want_delete) {udf_free_inode (inode); }@@ -106,6 +107,7 @@ static void udf_write_failed(struct address_space *mapping, loff_t to)truncate_pagecache (inode, to, isize);if (iinfo->i_alloc_type !=ICBTAG_FLAG_AD_IN_ICB ) {down_write (&iinfo->i_data_sem); +udf_clear_extent_cache (iinfo);udf_truncate_extents (inode);up_write (&iinfo->i_data_sem); }@@ -373,7 +375,7 @@ static int udf_get_block(struct inode *inode, sector_t block, iinfo->i_next_alloc_goal++; } - +udf_clear_extent_cache (iinfo); phys =inode_getblk (inode, block, &err, &new);if (!phys) goto abort;@@ -1172,6 +1174,7 @@ set_size: }else {if (iinfo->i_alloc_type ==ICBTAG_FLAG_AD_IN_ICB ) {down_write (&iinfo->i_data_sem); +udf_clear_extent_cache (iinfo);memset (iinfo->i_ext.i_data + iinfo->i_lenEAttr + newsize,0x00 , bsize - newsize -udf_file_entry_alloc_offset (inode));@@ -1185,6 +1188,7 @@ set_size:if (err)return err;down_write (&iinfo->i_data_sem); +udf_clear_extent_cache (iinfo);truncate_setsize (inode, newsize);udf_truncate_extents (inode);up_write (&iinfo->i_data_sem);@@ -1302,6 +1306,9 @@ static void udf_fill_inode(struct inode *inode, struct buffer_head *bh) iinfo->i_lenAlloc =0 ; iinfo->i_next_alloc_block =0 ; iinfo->i_next_alloc_goal =0 ; +memset (&iinfo->cached_extent,0 ,sizeof (struct udf_ext_cache )); +spin_lock_init (&(iinfo->i_extent_cache_lock)); + iinfo->cached_extent.lstart =-1 ;if (fe->descTag.tagIdent ==cpu_to_le16 (TAG_IDENT_EFE )) { iinfo->i_efe =1 ; iinfo->i_use =0 ;@@ -2157,11 +2164,12 @@ int8_t inode_bmap(struct inode *inode, sector_t block,struct udf_inode_info *iinfo; iinfo =UDF_I (inode); - pos->offset =0 ; - pos->block = iinfo->i_location; - pos->bh =NULL ; +if (!udf_read_extent_cache (inode, bcount, &lbcount, pos)) { + pos->offset =0 ; + pos->block = iinfo->i_location; + pos->bh =NULL ; + } *elen =0 ; - do { etype = udf_next_aext(inode, pos, eloc, elen, 1); if (etype == -1) {@@ -2171,7 +2179,8 @@ int8_t inode_bmap(struct inode *inode, sector_t block, } lbcount += *elen; }while (lbcount <= bcount); - +/* update extent cache */ +udf_update_extent_cache (inode, lbcount - *elen, pos,1 ); *offset = (bcount + *elen - lbcount) >> blocksize_bits;return etype;@@ -2201,3 +2210,61 @@ long udf_block_map(struct inode *inode, sector_t block)else return ret; } + +int udf_read_extent_cache (struct inode *inode,loff_t bcount, +loff_t *lbcount,struct extent_position *pos) +{ +struct udf_inode_info *iinfo =UDF_I (inode); + +spin_lock (&iinfo->i_extent_cache_lock); +if ((iinfo->cached_extent.lstart <= bcount) && + (iinfo->cached_extent.lstart !=-1 )) { +/* Cache hit */ + *lbcount = iinfo->cached_extent.lstart; +memcpy (pos, &iinfo->cached_extent.epos, +sizeof (struct extent_position )); +spin_unlock (&iinfo->i_extent_cache_lock); +return 1 ; + }else +udf_clear_extent_cache (iinfo); +spin_unlock (&iinfo->i_extent_cache_lock); +return 0 ; +} + +void udf_update_extent_cache (struct inode *inode,loff_t estart, +struct extent_position *pos,int next_epos) +{ +struct udf_inode_info *iinfo =UDF_I (inode); + +spin_lock (&iinfo->i_extent_cache_lock); +if (pos->bh !=NULL ) +/* Increase ref count */ +get_bh (pos->bh); +memcpy (&iinfo->cached_extent.epos, pos, +sizeof (struct extent_position )); + iinfo->cached_extent.lstart = estart; +if (next_epos) +switch (iinfo->i_alloc_type) { +case ICBTAG_FLAG_AD_SHORT : + iinfo->cached_extent.epos.offset -= +sizeof (struct short_ad ); +break ; +case ICBTAG_FLAG_AD_LONG : + iinfo->cached_extent.epos.offset -= +sizeof (struct long_ad ); + } +spin_unlock (&iinfo->i_extent_cache_lock); +} + +/* This function should be called after i_data_sem is +* held for writing OR i_extent_cache_lock is taken. +*/ +void udf_clear_extent_cache (struct udf_inode_info *iinfo) +{ +if (iinfo->cached_extent.lstart !=-1 ) { +brelse (iinfo->cached_extent.epos.bh); +memset (&iinfo->cached_extent,0 ,sizeof (struct udf_ext_cache )); + iinfo->cached_extent.lstart =-1 ; + } +} +fs/udf/udf_i.h +16
-0@@ -1,6 +1,19 @@#ifndef _UDF_I_H#define _UDF_I_H +struct extent_position { +struct buffer_head *bh; +uint32_t offset; +struct kernel_lb_addr block; +}; + +struct udf_ext_cache { +/* Extent position */ +struct extent_position epos; +/* Start logical offset in bytes */ +loff_t lstart; +}; +/* * The i_data_sem and i_mutex serve for protection of allocation information * of a regular files and symlinks. This includes all extents belonging to @@ -35,6 +48,9 @@ struct udf_inode_info {__u8 *i_data; } i_ext;struct rw_semaphore i_data_sem; +struct udf_ext_cache cached_extent; +/* Spinlock for protecting extent cache */ +spinlock_t i_extent_cache_lock;struct inode vfs_inode; };fs/udf/udfdecl.h +5
-5@@ -113,11 +113,6 @@ struct ustr {uint8_t u_len; }; -struct extent_position { -struct buffer_head *bh; -uint32_t offset; -struct kernel_lb_addr block; -};/* super.c */ @@ -164,6 +159,11 @@ extern int8_t udf_next_aext(struct inode *, struct extent_position *, struct kernel_lb_addr *, uint32_t *, int);extern int8_t udf_current_aext (struct inode *,struct extent_position *,struct kernel_lb_addr *,uint32_t *,int ); +int udf_read_extent_cache (struct inode *inode,loff_t bcount,loff_t *lbcount, +struct extent_position *pos); +void udf_update_extent_cache (struct inode *inode,loff_t estart, +struct extent_position *pos,int next_epos); +void udf_clear_extent_cache (struct udf_inode_info *iinfo);/* misc.c */ extern struct buffer_head *udf_tgetblk (struct super_block *,int );Signature · 7 lines
1.7.9.5 -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
2013/1/19, Cong Ding <[email protected]>:
Quoted 18 lines · Namjae Jeon
Quoted 17 lines
From: Namjae Jeon <[email protected]>
This patch implements extent caching in case of file reading. While reading a file, currently, UDF reads metadata serially which takes a lot of time depending on the number of extents present in the file. Caching last accessd extent improves metadata read time. Instead of reading file metadata from start, now we read from the cached extent.
This patch considerably improves the time spent by CPU in kernel mode. For example, while reading a 10.9 GB file using dd:
Time before applying patch:
11677022208 bytes (10.9GB) copied, 1529.748921 seconds, 7.3MB/sreal 25m 29.85s user 0m 12.41s sys 15m 34.75s
Quoted 6 lines
Quoted 5 lines
Time after applying patch:
11677022208 bytes (10.9GB) copied, 1469.338231 seconds, 7.6MB/sreal 24m 29.44s user 0m 15.73s sys 3m 27.61sdid you have any test on lots of small files?
Hi. Cong.
I created 2048 files of each 512KB size for testing performance drpping by extent cache feature.
Used this script to read every file:
index=0
while [ $index != 2048 ]
do
dd if=file.$index of=/dev/zero 1> /dev/null 2>/dev/null index=$(($index + 1))
done
Performance without patch:
VDLinux#> echo 3 > /proc/sys/vm/drop_caches
VDLinux#> time ./script2.sh
real 0m 55.13s user 0m 1.40s sys 0m 25.17s
Performace with patch =>
VDLinux#> time ./script2.sh
real 0m 53.70s user 0m 1.60s sys 0m 25.11s
I can not find any performance dropping with extent cache patch. Thanks.
Quoted 256 lines
- cong
Quoted 253 lines
Signed-off-by Namjae Jeon <[email protected]>Signed-off-by Ashish Sangwan <[email protected]>Signed-off-by Bonggil Bak <[email protected]>--- fs/udf/ialloc.c | 4 +++ fs/udf/inode.c | 79 +++++++++++++++++++++++++++++++++++++++++++++++++----- fs/udf/udf_i.h | 16 +++++++++++ fs/udf/udfdecl.h | 10 +++---- 4 files changed, 98 insertions(+), 11 deletions(-)fs/udf/ialloc.c +4
-0@@ -117,6 +117,10 @@ struct inode *udf_new_inode(struct inode *dir, umode_t mode, int *err) iinfo->i_lenAlloc = 0; iinfo->i_use =0 ; iinfo->i_checkpoint =1 ; +memset (&iinfo->cached_extent,0 ,sizeof (struct udf_ext_cache )); +spin_lock_init (&(iinfo->i_extent_cache_lock)); +/* Mark extent cache as invalid for now */ + iinfo->cached_extent.lstart =-1 ;if (UDF_QUERY_FLAG (inode->i_sb,UDF_FLAG_USE_AD_IN_ICB )) iinfo->i_alloc_type =ICBTAG_FLAG_AD_IN_ICB ;fs/udf/ialloc.c +0
-0else if (UDF_QUERY_FLAG (inode->i_sb,UDF_FLAG_USE_SHORT_AD ))fs/udf/inode.c +2
-0@@ -91,6 +91,7 @@ void udf_evict_inode(struct inode *inode) }kfree (iinfo->i_ext.i_data); iinfo->i_ext.i_data =NULL ; +udf_clear_extent_cache (iinfo);if (want_delete) {udf_free_inode (inode); }@@ -106,6 +107,7 @@ static void udf_write_failed(struct address_space *mapping, loff_t to)truncate_pagecache (inode, to, isize);if (iinfo->i_alloc_type !=ICBTAG_FLAG_AD_IN_ICB ) {down_write (&iinfo->i_data_sem); +udf_clear_extent_cache (iinfo);udf_truncate_extents (inode);up_write (&iinfo->i_data_sem);fs/udf/inode.c +1
-1}@@ -373,7 +375,7 @@ static int udf_get_block(struct inode *inode, sector_t block, iinfo->i_next_alloc_goal++; } - +udf_clear_extent_cache (iinfo); phys =inode_getblk (inode, block, &err, &new);if (!phys)fs/udf/inode.c +5
-0goto abort;@@ -1172,6 +1174,7 @@ set_size: }else {if (iinfo->i_alloc_type ==ICBTAG_FLAG_AD_IN_ICB ) {down_write (&iinfo->i_data_sem); +udf_clear_extent_cache (iinfo);memset (iinfo->i_ext.i_data + iinfo->i_lenEAttr + newsize,0x00 , bsize - newsize -udf_file_entry_alloc_offset (inode));@@ -1185,6 +1188,7 @@ set_size:if (err)return err;down_write (&iinfo->i_data_sem); +udf_clear_extent_cache (iinfo);truncate_setsize (inode, newsize);udf_truncate_extents (inode);up_write (&iinfo->i_data_sem);@@ -1302,6 +1306,9 @@ static void udf_fill_inode(struct inode *inode, struct buffer_head *bh) iinfo->i_lenAlloc = 0; iinfo->i_next_alloc_block =0 ; iinfo->i_next_alloc_goal =0 ; +memset (&iinfo->cached_extent,0 ,sizeof (struct udf_ext_cache )); +spin_lock_init (&(iinfo->i_extent_cache_lock)); + iinfo->cached_extent.lstart =-1 ;if (fe->descTag.tagIdent ==cpu_to_le16 (TAG_IDENT_EFE )) { iinfo->i_efe =1 ;fs/udf/inode.c +5
-4iinfo->i_use =0 ;@@ -2157,11 +2164,12 @@ int8_t inode_bmap(struct inode *inode, sector_t block, struct udf_inode_info *iinfo; iinfo =UDF_I (inode); - pos->offset =0 ; - pos->block = iinfo->i_location; - pos->bh =NULL ; +if (!udf_read_extent_cache (inode, bcount, &lbcount, pos)) { + pos->offset =0 ; + pos->block = iinfo->i_location; + pos->bh =NULL ; + } *elen =0 ; - do { etype = udf_next_aext(inode, pos, eloc, elen, 1);fs/udf/inode.c +2
-1if (etype ==-1 ) {@@ -2171,7 +2179,8 @@ int8_t inode_bmap(struct inode *inode, sector_t block, } lbcount += *elen; }while (lbcount <= bcount); - +/* update extent cache */ +udf_update_extent_cache (inode, lbcount - *elen, pos,1 ); *offset = (bcount + *elen - lbcount) >> blocksize_bits;fs/udf/inode.c +58
-0return etype;@@ -2201,3 +2210,61 @@ long udf_block_map(struct inode *inode, sector_t block) elsereturn ret; } + +int udf_read_extent_cache (struct inode *inode,loff_t bcount, +loff_t *lbcount,struct extent_position *pos) +{ +struct udf_inode_info *iinfo =UDF_I (inode); + +spin_lock (&iinfo->i_extent_cache_lock); +if ((iinfo->cached_extent.lstart <= bcount) && + (iinfo->cached_extent.lstart !=-1 )) { +/* Cache hit */ + *lbcount = iinfo->cached_extent.lstart; +memcpy (pos, &iinfo->cached_extent.epos, +sizeof (struct extent_position )); +spin_unlock (&iinfo->i_extent_cache_lock); +return 1 ; + }else +udf_clear_extent_cache (iinfo); +spin_unlock (&iinfo->i_extent_cache_lock); +return 0 ; +} + +void udf_update_extent_cache (struct inode *inode,loff_t estart, +struct extent_position *pos,int next_epos) +{ +struct udf_inode_info *iinfo =UDF_I (inode); + +spin_lock (&iinfo->i_extent_cache_lock); +if (pos->bh !=NULL ) +/* Increase ref count */ +get_bh (pos->bh); +memcpy (&iinfo->cached_extent.epos, pos, +sizeof (struct extent_position )); + iinfo->cached_extent.lstart = estart; +if (next_epos) +switch (iinfo->i_alloc_type) { +case ICBTAG_FLAG_AD_SHORT : + iinfo->cached_extent.epos.offset -= +sizeof (struct short_ad ); +break ; +case ICBTAG_FLAG_AD_LONG : + iinfo->cached_extent.epos.offset -= +sizeof (struct long_ad ); + } +spin_unlock (&iinfo->i_extent_cache_lock); +} + +/* This function should be called after i_data_sem is +* held for writing OR i_extent_cache_lock is taken. +*/ +void udf_clear_extent_cache (struct udf_inode_info *iinfo) +{ +if (iinfo->cached_extent.lstart !=-1 ) { +brelse (iinfo->cached_extent.epos.bh); +memset (&iinfo->cached_extent,0 ,sizeof (struct udf_ext_cache )); + iinfo->cached_extent.lstart =-1 ; + } +} +fs/udf/udf_i.h +13
-0@@ -1,6 +1,19 @@#ifndef _UDF_I_H#define _UDF_I_H +struct extent_position { +struct buffer_head *bh; +uint32_t offset; +struct kernel_lb_addr block; +}; + +struct udf_ext_cache { +/* Extent position */ +struct extent_position epos; +/* Start logical offset in bytes */ +loff_t lstart; +}; +/* * The i_data_sem and i_mutex serve for protection of allocation information
* of a regular files and symlinks. This includes all extents belonging
tofs/udf/udf_i.h +3
-0@@ -35,6 +48,9 @@ struct udf_inode_info {__u8 *i_data; } i_ext;struct rw_semaphore i_data_sem; +struct udf_ext_cache cached_extent; +/* Spinlock for protecting extent cache */ +spinlock_t i_extent_cache_lock;struct inode vfs_inode; };fs/udf/udfdecl.h +5
-5@@ -113,11 +113,6 @@ struct ustr {uint8_t u_len; }; -struct extent_position { -struct buffer_head *bh; -uint32_t offset; -struct kernel_lb_addr block; -};/* super.c */ @@ -164,6 +159,11 @@ extern int8_t udf_next_aext(struct inode *, struct extent_position *, struct kernel_lb_addr *, uint32_t *, int);extern int8_t udf_current_aext (struct inode *,struct extent_position *,struct kernel_lb_addr *,uint32_t *,int ); +int udf_read_extent_cache (struct inode *inode,loff_t bcount,loff_t *lbcount, +struct extent_position *pos); +void udf_update_extent_cache (struct inode *inode,loff_t estart, +struct extent_position *pos,int next_epos); +void udf_clear_extent_cache (struct udf_inode_info *iinfo);/*
misc.c*/
extern structbuffer_head*udf_tgetblk(structsuper_block*, int);
--
1.7.9.5--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Jan Kara
reviewerQuoted 19 lines · Namjae Jeon
From: Namjae Jeon <[email protected]>
This patch implements extent caching in case of file reading. While reading a file, currently, UDF reads metadata serially which takes a lot of time depending on the number of extents present in the file. Caching last accessd extent improves metadata read time. Instead of reading file metadata from start, now we read from the cached extent.
This patch considerably improves the time spent by CPU in kernel mode. For example, while reading a 10.9 GB file using dd:
Time before applying patch:
11677022208 bytes (10.9GB) copied, 1529.748921 seconds, 7.3MB/sreal 25m 29.85s user 0m 12.41s sys 15m 34.75sTime after applying patch:
11677022208 bytes (10.9GB) copied, 1469.338231 seconds, 7.6MB/s
Quoted 7 lines
real 24m 29.44s user 0m 15.73s sys 3m 27.61sSigned-off-by Namjae Jeon <[email protected]>Signed-off-by Ashish Sangwan <[email protected]>Signed-off-by Bonggil Bak <[email protected]>
Thanks for the patch Namjae. I did a few more changes to the patch.
Please check them whether you think they are OK.
fs/udf/ialloc.c +0 -4
@@ -117,10 +117,6 @@ struct inode *udf_new_inode(struct inode *dir, umode_t mode, int *err) iinfo->i_lenAlloc =0 ; iinfo->i_use =0 ; iinfo->i_checkpoint =1 ; -memset (&iinfo->cached_extent,0 ,sizeof (struct udf_ext_cache )); -spin_lock_init (&(iinfo->i_extent_cache_lock)); -/* Mark extent cache as invalid for now */ - iinfo->cached_extent.lstart =-1 ;if (UDF_QUERY_FLAG (inode->i_sb,UDF_FLAG_USE_AD_IN_ICB )) iinfo->i_alloc_type =ICBTAG_FLAG_AD_IN_ICB ;else if (UDF_QUERY_FLAG (inode->i_sb,UDF_FLAG_USE_SHORT_AD ))
fs/udf/inode.c +0 -3
@@ -1305,9 +1305,6 @@ static void udf_fill_inode(struct inode *inode, struct buffer_head *bh) iinfo->i_lenAlloc =0 ; iinfo->i_next_alloc_block =0 ; iinfo->i_next_alloc_goal =0 ; -memset (&iinfo->cached_extent,0 ,sizeof (struct udf_ext_cache )); -spin_lock_init (&(iinfo->i_extent_cache_lock)); - iinfo->cached_extent.lstart =-1 ;if (fe->descTag.tagIdent ==cpu_to_le16 (TAG_IDENT_EFE )) { iinfo->i_efe =1 ; iinfo->i_use =0 ;
Initialization now happens in udf_alloc_inode(). Also it's not necessary
to initialized cached_extent.epos when lstart == -1 - noone should look at that.
@@ -2222,6 +2219,8 @@ int udf_read_extent_cache(struct inode *inode, loff_t bcount,
*lbcount = iinfo->cached_extent.lstart;
memcpy(pos, &iinfo->cached_extent.epos,
sizeof(struct extent_position));
+ if (pos->bh)
+ get_bh(pos->bh);
spin_unlock(&iinfo->i_extent_cache_lock);
return 1;
} else
This is the most important - we should give buffer reference to pos->bh.
Caller will eventually free it right?
@@ -2236,8 +2235,7 @@ void udf_update_extent_cache(struct inode *inode, loff_t estart,
struct udf_inode_info *iinfo = UDF_I(inode);
spin_lock(&iinfo->i_extent_cache_lock);
- if (pos->bh != NULL)
- /* Increase ref count */
+ if (pos->bh)
get_bh(pos->bh);
memcpy(&iinfo->cached_extent.epos, pos,
sizeof(struct extent_position));
@@ -2266,4 +2264,3 @@ void udf_clear_extent_cache(struct udf_inode_info *iinfo)
iinfo->cached_extent.lstart = -1;
}
}
-
fs/udf/super.c +2 -0
@@ -134,6 +134,8 @@ static struct inode *udf_alloc_inode(struct super_block *sb) ei->i_next_alloc_goal =0 ; ei->i_strat4096 =0 ;init_rwsem (&ei->i_data_sem); + ei->cached_extent.lstart =-1 ; +spin_lock_init (&ei->i_extent_cache_lock);return &ei->vfs_inode; }
Honza
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR
2013/1/21, Jan Kara <[email protected]>:
Quoted 60 lines · Namjae Jeon
Quoted 19 lines
From: Namjae Jeon <[email protected]>
This patch implements extent caching in case of file reading. While reading a file, currently, UDF reads metadata serially which takes a lot of time depending on the number of extents present in the file. Caching last accessd extent improves metadata read time. Instead of reading file metadata from start, now we read from the cached extent.
This patch considerably improves the time spent by CPU in kernel mode. For example, while reading a 10.9 GB file using dd:
Time before applying patch:
11677022208 bytes (10.9GB) copied, 1529.748921 seconds, 7.3MB/sreal 25m 29.85s user 0m 12.41s sys 15m 34.75sTime after applying patch:
11677022208 bytes (10.9GB) copied, 1469.338231 seconds, 7.6MB/sQuoted 7 lines
real 24m 29.44s user 0m 15.73s sys 3m 27.61sSigned-off-by Namjae Jeon <[email protected]>Signed-off-by Ashish Sangwan <[email protected]>Signed-off-by Bonggil Bak <[email protected]>Thanks for the patch Namjae. I did a few more changes to the patch.
Please check them whether you think they are OK.fs/udf/ialloc.c +0
-4@@ -117,10 +117,6 @@ struct inode *udf_new_inode(struct inode *dir, umode_t mode, int *err) iinfo->i_lenAlloc =0 ; iinfo->i_use =0 ; iinfo->i_checkpoint =1 ; -memset (&iinfo->cached_extent,0 ,sizeof (struct udf_ext_cache )); -spin_lock_init (&(iinfo->i_extent_cache_lock)); -/* Mark extent cache as invalid for now */ - iinfo->cached_extent.lstart =-1 ;if (UDF_QUERY_FLAG (inode->i_sb,UDF_FLAG_USE_AD_IN_ICB )) iinfo->i_alloc_type =ICBTAG_FLAG_AD_IN_ICB ;fs/udf/ialloc.c +0
-0else if (UDF_QUERY_FLAG (inode->i_sb,UDF_FLAG_USE_SHORT_AD ))fs/udf/inode.c +0
-3@@ -1305,9 +1305,6 @@ static void udf_fill_inode(struct inode *inode, struct buffer_head *bh) iinfo->i_lenAlloc = 0; iinfo->i_next_alloc_block =0 ; iinfo->i_next_alloc_goal =0 ; -memset (&iinfo->cached_extent,0 ,sizeof (struct udf_ext_cache )); -spin_lock_init (&(iinfo->i_extent_cache_lock)); - iinfo->cached_extent.lstart =-1 ;if (fe->descTag.tagIdent ==cpu_to_le16 (TAG_IDENT_EFE )) { iinfo->i_efe =1 ;
Quoted 2 lines
iinfo->i_use= 0;
Hi Jan.
Quoted 3 lines
Initialization now happens in
udf_alloc_inode(). Also it's not necessary
to initialized cached_extent.epos when lstart == -1 - noone should look at that.
This change is okay.
Quoted 10 lines
fs/udf/inode.c +2
-0@@ -2222,6 +2219,8 @@ int udf_read_extent_cache(struct inode *inode, loff_t bcount, *lbcount = iinfo->cached_extent.lstart;memcpy (pos, &iinfo->cached_extent.epos,sizeof (struct extent_position )); +if (pos->bh) +get_bh (pos->bh);spin_unlock (&iinfo->i_extent_cache_lock);return 1 ;
Quoted 3 lines
} else
This is the most important - we should give buffer reference topos->bh.
Caller will eventually free it right?
This change is not required as we give buffer reference to pos->bh at the time of cache update.
When we start reading a file, first we try to read the cache which will lead to cache miss.
So, we would really access the pos->bh in udf_update_extent_cache for the first time, and this is where the buffer reference is incremented. Calling get_bh at 2 places will eventually lead to mem leak. Let me know your opinion.
Thanks for review and change!
Quoted 38 lines
fs/udf/inode.c +1
-2@@ -2236,8 +2235,7 @@ void udf_update_extent_cache(struct inode *inode,loff_t estart, struct udf_inode_info *iinfo = UDF_I(inode);spin_lock (&iinfo->i_extent_cache_lock); -if (pos->bh !=NULL ) -/* Increase ref count */ +if (pos->bh)get_bh (pos->bh); memcpy(&iinfo->cached_extent.epos, pos,fs/udf/inode.c +0
-1sizeof(struct extent_position));@@ -2266,4 +2264,3 @@ void udf_clear_extent_cache(struct udf_inode_info *iinfo) iinfo->cached_extent.lstart =-1 ; } } -fs/udf/super.c +2
-0@@ -134,6 +134,8 @@ static struct inode *udf_alloc_inode(struct super_block *sb) ei->i_next_alloc_goal =0 ; ei->i_strat4096 =0 ;init_rwsem (&ei->i_data_sem); + ei->cached_extent.lstart =-1 ; +spin_lock_init (&ei->i_extent_cache_lock);return &ei->vfs_inode;}
Honza
--
Jan Kara <[email protected]>
SUSE Labs, CR
Jan Kara
reviewerQuoted 13 lines · Namjae Jeon
2013/1/21, Jan Kara <[email protected]>:
Quoted 9 lines
(quoted diff) +2
-0@@ -2222,6 +2219,8 @@ int udf_read_extent_cache(struct inode *inode, loff_t bcount, *lbcount = iinfo->cached_extent.lstart; memcpy(pos, &iinfo->cached_extent.epos, sizeof(struct extent_position)); + if (pos->bh) + get_bh(pos->bh); spin_unlock(&iinfo->i_extent_cache_lock); return 1;Quoted 3 lines
} else
This is the most important - we should give buffer reference topos->bh.
Caller will eventually free it right?
Quoted 8 lines
This change is not required as we give buffer reference to
pos->bhat the time of cache update.
When we start reading a file, first we try to read the cache which will lead to cache miss.
So, we would really access thepos->bhinudf_update_extent_cachefor the first time, and this is where the buffer reference is incremented. Callingget_bhat 2 places will eventually lead to mem leak. Let me know your opinion.
Yes, udf_update_extent_cache() gets its own reference to bh but that is
dropped in udf_clear_extent_cache(). So I think udf_read_extent_cache() needs to get a reference to the caller (as the caller will eventually free the bh via brelse(epos.bh) e.g. in udf_extend_file(). Also I realized udf_update_extent_cache() needs to first clear the cache if it is valid. Otherwise it just overwrites bh pointer and reference is leaked. Is it clearer now?
I've also changed locking of udf_clear_extent_cache() so thati_extent_cache_lock is always taken for that function - it makes the locking rules obvious at the first sight.
Attached is the patch I currently carry.
Honza
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR
Quoted 1 line
From
99600051b04bc4ec8bd4d16a8bf993ca54042db6Mon Sep 17 00:00:00 2001
From: Namjae Jeon <[email protected]>
Date: Sat, 19 Jan 2013 11:17:14 +0900
Subject: [PATCH] udf: add extent cache support in case of file reading
This patch implements extent caching in case of file reading. While reading a file, currently, UDF reads metadata serially which takes a lot of time depending on the number of extents present in the file. Caching last accessd extent improves metadata read time. Instead of reading file metadata from start, now we read from the cached extent.
This patch considerably improves the time spent by CPU in kernel mode. For example, while reading a 10.9 GB file using dd:
Time before applying patch:
11677022208 bytes (10.9GB) copied, 1529.748921 seconds, 7.3MB/s
real 25m 29.85s user 0m 12.41s sys 15m 34.75s
Time after applying patch:
11677022208 bytes (10.9GB) copied, 1469.338231 seconds, 7.6MB/s
real 24m 29.44s user 0m 15.73s sys 3m 27.61s
[JK: Fix bh refcounting issues, simplify initialization]
--- fs/udf/inode.c | 86 ++++++++++++++++++++++++++++++++++++++++++++++++++---- fs/udf/super.c | 2 + fs/udf/udf_i.h | 16 ++++++++++ fs/udf/udfdecl.h | 5 --- 4 files changed, 98 insertions(+), 11 deletions(-)
fs/udf/inode.c +80 -6
@@ -67,6 +67,74 @@ static void udf_update_extents(struct inode *, struct extent_position *);static int udf_get_block (struct inode *,sector_t ,struct buffer_head *,int ); +static void __udf_clear_extent_cache (struct inode *inode) +{ +struct udf_inode_info *iinfo =UDF_I (inode); + +if (iinfo->cached_extent.lstart !=-1 ) { +brelse (iinfo->cached_extent.epos.bh); + iinfo->cached_extent.lstart =-1 ; + } +} + +/* Invalidate extent cache */ +static void udf_clear_extent_cache (struct inode *inode) +{ +struct udf_inode_info *iinfo =UDF_I (inode); + +spin_lock (&iinfo->i_extent_cache_lock); +__udf_clear_extent_cache (inode); +spin_unlock (&iinfo->i_extent_cache_lock); +} + +/* Return contents of extent cache */ +static int udf_read_extent_cache (struct inode *inode,loff_t bcount, +loff_t *lbcount,struct extent_position *pos) +{ +struct udf_inode_info *iinfo =UDF_I (inode); +int ret =0 ; + +spin_lock (&iinfo->i_extent_cache_lock); +if ((iinfo->cached_extent.lstart <= bcount) && + (iinfo->cached_extent.lstart !=-1 )) { +/* Cache hit */ + *lbcount = iinfo->cached_extent.lstart; +memcpy (pos, &iinfo->cached_extent.epos, +sizeof (struct extent_position )); +if (pos->bh) +get_bh (pos->bh); + ret =1 ; + } +spin_unlock (&iinfo->i_extent_cache_lock); +return ret; +} + +/* Add extent to extent cache */ +static void udf_update_extent_cache (struct inode *inode,loff_t estart, +struct extent_position *pos,int next_epos) +{ +struct udf_inode_info *iinfo =UDF_I (inode); + +spin_lock (&iinfo->i_extent_cache_lock); +/* Invalidate previously cached extent */ +__udf_clear_extent_cache (inode); +if (pos->bh) +get_bh (pos->bh); +memcpy (&iinfo->cached_extent.epos, pos, +sizeof (struct extent_position )); + iinfo->cached_extent.lstart = estart; +if (next_epos) +switch (iinfo->i_alloc_type) { +case ICBTAG_FLAG_AD_SHORT : + iinfo->cached_extent.epos.offset -= +sizeof (struct short_ad ); +break ; +case ICBTAG_FLAG_AD_LONG : + iinfo->cached_extent.epos.offset -= +sizeof (struct long_ad ); + } +spin_unlock (&iinfo->i_extent_cache_lock); +}void udf_evict_inode (struct inode *inode) {
@@ -90,6 +158,7 @@ void udf_evict_inode(struct inode *inode) }kfree (iinfo->i_ext.i_data); iinfo->i_ext.i_data =NULL ; +udf_clear_extent_cache (inode);if (want_delete) {udf_free_inode (inode); }
@@ -105,6 +174,7 @@ static void udf_write_failed(struct address_space *mapping, loff_t to)truncate_pagecache (inode, to, isize);if (iinfo->i_alloc_type !=ICBTAG_FLAG_AD_IN_ICB ) {down_write (&iinfo->i_data_sem); +udf_clear_extent_cache (inode);udf_truncate_extents (inode);up_write (&iinfo->i_data_sem); }
@@ -372,7 +442,7 @@ static int udf_get_block(struct inode *inode, sector_t block, iinfo->i_next_alloc_goal++; } - +udf_clear_extent_cache (inode); phys =inode_getblk (inode, block, &err, &new);if (!phys) goto abort;
@@ -1171,6 +1241,7 @@ set_size: }else {if (iinfo->i_alloc_type ==ICBTAG_FLAG_AD_IN_ICB ) {down_write (&iinfo->i_data_sem); +udf_clear_extent_cache (inode);memset (iinfo->i_ext.i_data + iinfo->i_lenEAttr + newsize,0x00 , bsize - newsize -udf_file_entry_alloc_offset (inode));
@@ -1184,6 +1255,7 @@ set_size:if (err)return err;down_write (&iinfo->i_data_sem); +udf_clear_extent_cache (inode);truncate_setsize (inode, newsize);udf_truncate_extents (inode);up_write (&iinfo->i_data_sem);
@@ -2156,11 +2228,12 @@ int8_t inode_bmap(struct inode *inode, sector_t block,struct udf_inode_info *iinfo; iinfo =UDF_I (inode); - pos->offset =0 ; - pos->block = iinfo->i_location; - pos->bh =NULL ; +if (!udf_read_extent_cache (inode, bcount, &lbcount, pos)) { + pos->offset =0 ; + pos->block = iinfo->i_location; + pos->bh =NULL ; + } *elen =0 ; - do { etype = udf_next_aext(inode, pos, eloc, elen, 1); if (etype == -1) {
@@ -2170,7 +2243,8 @@ int8_t inode_bmap(struct inode *inode, sector_t block, } lbcount += *elen; }while (lbcount <= bcount); - +/* update extent cache */ +udf_update_extent_cache (inode, lbcount - *elen, pos,1 ); *offset = (bcount + *elen - lbcount) >> blocksize_bits;return etype;
fs/udf/super.c +2 -0
@@ -134,6 +134,8 @@ static struct inode *udf_alloc_inode(struct super_block *sb) ei->i_next_alloc_goal =0 ; ei->i_strat4096 =0 ;init_rwsem (&ei->i_data_sem); + ei->cached_extent.lstart =-1 ; +spin_lock_init (&ei->i_extent_cache_lock);return &ei->vfs_inode; }
fs/udf/udf_i.h +16 -0
@@ -1,6 +1,19 @@#ifndef _UDF_I_H#define _UDF_I_H +struct extent_position { +struct buffer_head *bh; +uint32_t offset; +struct kernel_lb_addr block; +}; + +struct udf_ext_cache { +/* Extent position */ +struct extent_position epos; +/* Start logical offset in bytes */ +loff_t lstart; +}; +/* * The i_data_sem and i_mutex serve for protection of allocation information * of a regular files and symlinks. This includes all extents belonging to
@@ -35,6 +48,9 @@ struct udf_inode_info {__u8 *i_data; } i_ext;struct rw_semaphore i_data_sem; +struct udf_ext_cache cached_extent; +/* Spinlock for protecting extent cache */ +spinlock_t i_extent_cache_lock;struct inode vfs_inode; };
fs/udf/udfdecl.h +0 -5
@@ -113,11 +113,6 @@ struct ustr {uint8_t u_len; }; -struct extent_position { -struct buffer_head *bh; -uint32_t offset; -struct kernel_lb_addr block; -};/* super.c */
Signature · 1 line
1.7.1
2013/1/22, Jan Kara <[email protected]>:
Quoted 24 lines · Namjae Jeon
Quoted 15 lines
2013/1/21, Jan Kara <[email protected]>:
Quoted 12 lines
(quoted diff) +0
-0@@ -2222,6 +2219,8 @@ int udf_read_extent_cache(struct inode *inode,
loff_t
bcount,(quoted diff) +2
-0*lbcount = iinfo->cached_extent.lstart; memcpy(pos, &iinfo->cached_extent.epos, sizeof(struct extent_position)); + if (pos->bh) + get_bh(pos->bh); spin_unlock(&iinfo->i_extent_cache_lock); return 1; } else This is the most important - we should give buffer reference toQuoted 2 lines
pos->bh.
Caller will eventually free it right?Quoted 8 lines
This change is not required as we give buffer reference to
pos->bhat the time of cache update.
When we start reading a file, first we try to read the cache which will lead to cache miss.
So, we would really access thepos->bhinudf_update_extent_cachefor the first time, and this is where the buffer reference is incremented. Callingget_bhat 2 places will eventually lead to mem leak. Let me know your opinion.
Quoted 7 lines
Yes,
udf_update_extent_cache()gets its own reference to bh but that is
dropped inudf_clear_extent_cache(). So I thinkudf_read_extent_cache()needs to get a reference to the caller (as the caller will eventually free the bh via brelse(epos.bh) e.g. inudf_extend_file(). Also I realizedudf_update_extent_cache()needs to first clear the cache if it is valid. Otherwise it just overwrites bh pointer and reference is leaked. Is it clearer now?
Yes, you're right. Also, this patch looks good to me.
Quoted 4 lines
I've also changed locking of
udf_clear_extent_cache()so thati_extent_cache_lockis always taken for that function - it makes the locking rules obvious at the first sight.
Yes, right. it is needed.
When we test with this patch, working fine.
Thanks Jan!
Quoted 9 lines
Attached is the patch I currently carry.
Honza
--
Jan Kara <[email protected]>
SUSE Labs, CR
Hi. Jan.
Sorry for interrupt.
Have you taken this patch to your tree ? I can not find it.. or Is there any issue regarding this patch ?
Thanks!
2013/1/22, Namjae Jeon <[email protected]>:
Quoted 51 lines
2013/1/22, Jan Kara <[email protected]>:
Quoted 24 lines · Namjae Jeon
Quoted 15 lines
2013/1/21, Jan Kara <[email protected]>:
Quoted 12 lines
(quoted diff) +0
-0@@ -2222,6 +2219,8 @@ int udf_read_extent_cache(struct inode *inode,
loff_t
bcount,(quoted diff) +2
-0*lbcount = iinfo->cached_extent.lstart; memcpy(pos, &iinfo->cached_extent.epos, sizeof(struct extent_position)); + if (pos->bh) + get_bh(pos->bh); spin_unlock(&iinfo->i_extent_cache_lock); return 1; } else This is the most important - we should give buffer reference toQuoted 2 lines
pos->bh.
Caller will eventually free it right?Quoted 8 lines
This change is not required as we give buffer reference to
pos->bhat the time of cache update.
When we start reading a file, first we try to read the cache which will lead to cache miss.
So, we would really access thepos->bhinudf_update_extent_cachefor the first time, and this is where the buffer reference is incremented. Callingget_bhat 2 places will eventually lead to mem leak. Let me know your opinion.Quoted 8 lines
Yes,
udf_update_extent_cache()gets its own reference to bh but that is
dropped inudf_clear_extent_cache(). So I thinkudf_read_extent_cache()needs to get a reference to the caller (as the caller will eventually free
the bh via brelse(epos.bh) e.g. inudf_extend_file(). Also I realizedudf_update_extent_cache()needs to first clear the cache if it is valid. Otherwise it just overwrites bh pointer and reference is leaked. Is it clearer now?Yes, you're right. Also, this patch looks good to me.
Quoted 4 lines
I've also changed locking of
udf_clear_extent_cache()so thati_extent_cache_lockis always taken for that function - it makes the locking rules obvious at the first sight.Yes, right. it is needed.
When we test with this patch, working fine.
Thanks Jan!Quoted 9 lines
Attached is the patch I currently carry.
Honza
--
Jan Kara <[email protected]>
SUSE Labs, CR
Jan Kara
reviewerQuoted 5 lines · Namjae Jeon
Hi. Jan.
Sorry for interrupt.
Have you taken this patch to your tree ? I can not find it.. or Is there any issue regarding this patch ?
I had it in my tree but not in the for_next branch. Did it now so you
should see the patch in tomorrow's linux-next.
Honza
Quoted 52 lines
2013/1/22, Namjae Jeon <[email protected]>:
Quoted 51 lines
2013/1/22, Jan Kara <[email protected]>:
Quoted 24 lines · Namjae Jeon
Quoted 15 lines
2013/1/21, Jan Kara <[email protected]>:
Quoted 12 lines
(quoted diff) +0
-0@@ -2222,6 +2219,8 @@ int udf_read_extent_cache(struct inode *inode,
loff_t
bcount,(quoted diff) +2
-0*lbcount = iinfo->cached_extent.lstart; memcpy(pos, &iinfo->cached_extent.epos, sizeof(struct extent_position)); + if (pos->bh) + get_bh(pos->bh); spin_unlock(&iinfo->i_extent_cache_lock); return 1; } else This is the most important - we should give buffer reference toQuoted 2 lines
pos->bh.
Caller will eventually free it right?Quoted 8 lines
This change is not required as we give buffer reference to
pos->bhat the time of cache update.
When we start reading a file, first we try to read the cache which will lead to cache miss.
So, we would really access thepos->bhinudf_update_extent_cachefor the first time, and this is where the buffer reference is incremented. Callingget_bhat 2 places will eventually lead to mem leak. Let me know your opinion.Quoted 8 lines
Yes,
udf_update_extent_cache()gets its own reference to bh but that is
dropped inudf_clear_extent_cache(). So I thinkudf_read_extent_cache()needs to get a reference to the caller (as the caller will eventually free
the bh via brelse(epos.bh) e.g. inudf_extend_file(). Also I realizedudf_update_extent_cache()needs to first clear the cache if it is valid. Otherwise it just overwrites bh pointer and reference is leaked. Is it clearer now?Yes, you're right. Also, this patch looks good to me.
Quoted 4 lines
I've also changed locking of
udf_clear_extent_cache()so thati_extent_cache_lockis always taken for that function - it makes the locking rules obvious at the first sight.Yes, right. it is needed.
When we test with this patch, working fine.
Thanks Jan!Quoted 9 lines
Attached is the patch I currently carry.
Honza
--
Jan Kara <[email protected]>
SUSE Labs, CR
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR
2013/2/4, Jan Kara <[email protected]>:
Quoted 8 lines · Namjae Jeon
Quoted 5 lines
Hi. Jan.
Sorry for interrupt.
Have you taken this patch to your tree ? I can not find it.. or Is there any issue regarding this patch ?I had it in my tree but not in the
for_nextbranch. Did it now so you
should see the patch in tomorrow's linux-next.
Okay, I see.
Thanks Jan!
Quoted 63 lines
Honza
Quoted 57 lines
2013/1/22, Namjae Jeon <[email protected]>:
Quoted 56 lines
2013/1/22, Jan Kara <[email protected]>:
Quoted 26 lines · Namjae Jeon
Quoted 17 lines
2013/1/21, Jan Kara <[email protected]>:
Quoted 13 lines
(quoted diff) +0
-0@@ -2222,6 +2219,8 @@ int udf_read_extent_cache(struct inode *inode,
loff_t
bcount,(quoted diff) +2
-0*lbcount = iinfo->cached_extent.lstart; memcpy(pos, &iinfo->cached_extent.epos, sizeof(struct extent_position)); + if (pos->bh) + get_bh(pos->bh); spin_unlock(&iinfo->i_extent_cache_lock); return 1; } else This is the most important - we should give buffer reference toQuoted 2 lines
pos->bh.
Caller will eventually free it right?This change is not required as we give buffer reference to
pos->bhatQuoted 8 lines
the time of cache update.
When we start reading a file, first we try to read the cache which will lead to cache miss.
So, we would really access thepos->bhinudf_update_extent_cachefor the first time, and this is where the buffer reference is incremented.
Callingget_bhat 2 places will eventually lead to mem leak. Let me know your opinion.Quoted 11 lines
Yes,
udf_update_extent_cache()gets its own reference to bh but that
is
dropped inudf_clear_extent_cache(). So I thinkudf_read_extent_cache()
needs to get a reference to the caller (as the caller will eventually free
the bh via brelse(epos.bh) e.g. inudf_extend_file(). Also I realizedudf_update_extent_cache()needs to first clear the cache if it is valid.
Otherwise it just overwrites bh pointer and reference is leaked. Is it clearer now?Yes, you're right. Also, this patch looks good to me.
Quoted 4 lines
I've also changed locking of
udf_clear_extent_cache()so thati_extent_cache_lockis always taken for that function - it makes the locking rules obvious at the first sight.Yes, right. it is needed.
When we test with this patch, working fine.
Thanks Jan!Quoted 9 lines
Attached is the patch I currently carry.
Honza
--
Jan Kara <[email protected]>
SUSE Labs, CR--
Jan Kara <[email protected]>
SUSE Labs, CR