Message ID | Pine.LNX.4.64.0811141115100.29826@melkki.cs.Helsinki.FI |
---|---|
State | Superseded, archived |
Headers | show |
On Nov 14, 2008 11:17 +0200, Pekka J Enberg wrote: > As spotted by kmemtrace, struct ext2_sb_info is 17024 bytes and ext3_sb_info is > 17152 bytes on 64-bit which makes them a very bad fit for SLAB allocators. In > fact, both allocations are round up to the next available page size of > order 3 which is 32 KB. > > The culprit if the wasted memory is the ->s_blockgroup_lock which can be as > big as 16 KB when CONFIG_NR_CPUS is set to 32. As struct blockgroup_lock is a > perfect fit for order 2 page in the worst case, allocate ->s_blockgroup_lock > separately to avoid wasting space. > > The change shrinks struct ext2_sb_info to 592 bytes and struct ext3_sb_info to > 640 bytes which fits into a 1024 byte slab cache so now we allocate 16 KB + 1 > KB instead of 32 KB saving 15 KB of memory! > > Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi> This looks very reasonable, with some minor comments below. Could you please also include a patch for ext4. Also, Andrew prefers that the patches for ext2/ext3/ext4 are in separate emails. > --- a/include/linux/blockgroup_lock.h > +++ b/include/linux/blockgroup_lock.h > #define sb_bgl_lock(sb, block_group) \ > - (&(sb)->s_blockgroup_lock.locks[(block_group) & (NR_BG_LOCKS-1)].lock) > + (&(sb)->s_blockgroup_lock->locks[(block_group) & (NR_BG_LOCKS-1)].lock) How the struct is allocated seems like an implementation detail that doesn't belong in blockgroup_lock.h at all, because "sb" is not "struct superblock" but rather "struct ext[23]_sb_info". In fact, changing this without also patching ext4 would cause ext4 to break. I would suggest to change this to take the s_blockgroup_lock as a parameter, #define bgl_lock_ptr(bgl, block_group) (bgl->locks[(block_group) & (NR_BG_LOCKS - 1)].lock) and then in ext[234]_fs_sb.h add a new helper in the same (first) patch: #define sb_bgl_lock(sbi, block_group) bgl_lock_ptr(&sbi->s_blockgroup_lock, block_group) and remove sb_bgl_lock() from blockgroup_lock.h entirely. As part of the later patches to change the s_blockgroup_lock allocations for each of ext[234] this changes in ext[234]_fs_sb.h to: #define sb_bgl_lock(sbi, block_group) bgl_lock_ptr(sbi->s_blockgroup_lock, block_group) This allows each of the later patches to be landed separately without breaking the build. Cheers, Andreas -- Andreas Dilger Sr. Staff Engineer, Lustre Group Sun Microsystems of Canada, Inc. -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, 2008-11-14 at 11:17 +0200, Pekka J Enberg wrote: > From: Pekka Enberg <penberg@cs.helsinki.fi> > > As spotted by kmemtrace, struct ext2_sb_info is 17024 bytes and ext3_sb_info is > 17152 bytes on 64-bit which makes them a very bad fit for SLAB allocators. In > fact, both allocations are round up to the next available page size of > order 3 which is 32 KB. > > The culprit if the wasted memory is the ->s_blockgroup_lock which can be as big > as 16 KB when CONFIG_NR_CPUS is set to 32. As struct blockgroup_lock is a > perfect fit for order 2 page in the worst case, allocate ->s_blockgroup_lock > separately to avoid wasting space. And here I was thinking that NR_CPUS=4096 is currently our worst case ;-) -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Peter Zijlstra wrote: > On Fri, 2008-11-14 at 11:17 +0200, Pekka J Enberg wrote: >> From: Pekka Enberg <penberg@cs.helsinki.fi> >> >> As spotted by kmemtrace, struct ext2_sb_info is 17024 bytes and ext3_sb_info is >> 17152 bytes on 64-bit which makes them a very bad fit for SLAB allocators. In >> fact, both allocations are round up to the next available page size of >> order 3 which is 32 KB. >> >> The culprit if the wasted memory is the ->s_blockgroup_lock which can be as big >> as 16 KB when CONFIG_NR_CPUS is set to 32. As struct blockgroup_lock is a >> perfect fit for order 2 page in the worst case, allocate ->s_blockgroup_lock >> separately to avoid wasting space. > > And here I was thinking that NR_CPUS=4096 is currently our worst > case ;-) Sure but look at <linux/blockgroup_lock.h>. NR_BG_LOCKS is capped to 128 for >= 32 CPUs. -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/fs/ext2/super.c b/fs/ext2/super.c index 647cd88..da8bdea 100644 --- a/fs/ext2/super.c +++ b/fs/ext2/super.c @@ -132,6 +132,7 @@ static void ext2_put_super (struct super_block * sb) percpu_counter_destroy(&sbi->s_dirs_counter); brelse (sbi->s_sbh); sb->s_fs_info = NULL; + kfree(sbi->s_blockgroup_lock); kfree(sbi); return; @@ -756,6 +757,13 @@ static int ext2_fill_super(struct super_block *sb, void *data, int silent) sbi = kzalloc(sizeof(*sbi), GFP_KERNEL); if (!sbi) return -ENOMEM; + + sbi->s_blockgroup_lock = + kzalloc(sizeof(struct blockgroup_lock), GFP_KERNEL); + if (!sbi->s_blockgroup_lock) { + kfree(sbi); + return -ENOMEM; + } sb->s_fs_info = sbi; sbi->s_sb_block = sb_block; @@ -983,7 +991,7 @@ static int ext2_fill_super(struct super_block *sb, void *data, int silent) printk ("EXT2-fs: not enough memory\n"); goto failed_mount; } - bgl_lock_init(&sbi->s_blockgroup_lock); + bgl_lock_init(sbi->s_blockgroup_lock); sbi->s_debts = kcalloc(sbi->s_groups_count, sizeof(*sbi->s_debts), GFP_KERNEL); if (!sbi->s_debts) { printk ("EXT2-fs: not enough memory\n"); diff --git a/fs/ext3/super.c b/fs/ext3/super.c index f6c94f2..f41df22 100644 --- a/fs/ext3/super.c +++ b/fs/ext3/super.c @@ -439,6 +439,7 @@ static void ext3_put_super (struct super_block * sb) ext3_blkdev_remove(sbi); } sb->s_fs_info = NULL; + kfree(sbi->s_blockgroup_lock); kfree(sbi); return; } @@ -1548,6 +1549,13 @@ static int ext3_fill_super (struct super_block *sb, void *data, int silent) sbi = kzalloc(sizeof(*sbi), GFP_KERNEL); if (!sbi) return -ENOMEM; + + sbi->s_blockgroup_lock = + kzalloc(sizeof(struct blockgroup_lock), GFP_KERNEL); + if (!sbi->s_blockgroup_lock) { + kfree(sbi); + return -ENOMEM; + } sb->s_fs_info = sbi; sbi->s_mount_opt = 0; sbi->s_resuid = EXT3_DEF_RESUID; @@ -1788,7 +1796,7 @@ static int ext3_fill_super (struct super_block *sb, void *data, int silent) goto failed_mount; } - bgl_lock_init(&sbi->s_blockgroup_lock); + bgl_lock_init(sbi->s_blockgroup_lock); for (i = 0; i < db_count; i++) { block = descriptor_loc(sb, logic_sb_block, i); diff --git a/include/linux/blockgroup_lock.h b/include/linux/blockgroup_lock.h index 8607312..d6d4787 100644 --- a/include/linux/blockgroup_lock.h +++ b/include/linux/blockgroup_lock.h @@ -54,6 +54,6 @@ static inline void bgl_lock_init(struct blockgroup_lock *bgl) * superblock types */ #define sb_bgl_lock(sb, block_group) \ - (&(sb)->s_blockgroup_lock.locks[(block_group) & (NR_BG_LOCKS-1)].lock) + (&(sb)->s_blockgroup_lock->locks[(block_group) & (NR_BG_LOCKS-1)].lock) #endif diff --git a/include/linux/ext2_fs_sb.h b/include/linux/ext2_fs_sb.h index f273415..7e61de9 100644 --- a/include/linux/ext2_fs_sb.h +++ b/include/linux/ext2_fs_sb.h @@ -101,7 +101,7 @@ struct ext2_sb_info { struct percpu_counter s_freeblocks_counter; struct percpu_counter s_freeinodes_counter; struct percpu_counter s_dirs_counter; - struct blockgroup_lock s_blockgroup_lock; + struct blockgroup_lock *s_blockgroup_lock; /* root of the per fs reservation window tree */ spinlock_t s_rsv_window_lock; struct rb_root s_rsv_window_root; diff --git a/include/linux/ext3_fs_sb.h b/include/linux/ext3_fs_sb.h index b65f028..ec10d96 100644 --- a/include/linux/ext3_fs_sb.h +++ b/include/linux/ext3_fs_sb.h @@ -60,7 +60,7 @@ struct ext3_sb_info { struct percpu_counter s_freeblocks_counter; struct percpu_counter s_freeinodes_counter; struct percpu_counter s_dirs_counter; - struct blockgroup_lock s_blockgroup_lock; + struct blockgroup_lock *s_blockgroup_lock; /* root of the per fs reservation window tree */ spinlock_t s_rsv_window_lock;