[f2fs-dev][PATCH] f2fs: avoid to use slab memory in f2fs_issue_flush for efficiency

From: Chao Yu
Date: Thu May 08 2014 - 05:01:33 EST


If we use slab memory in f2fs_issue_flush(), we will face memory pressure and
latency time caused by racing of kmem_cache_{alloc,free}.

Let's alloc memory in stack instead of slab.

Signed-off-by: Chao Yu <chao2.yu@xxxxxxxxxxx>
---
fs/f2fs/segment.c | 28 +++++++++-------------------
1 file changed, 9 insertions(+), 19 deletions(-)

diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
index 2ecbffb..f25f0e0 100644
--- a/fs/f2fs/segment.c
+++ b/fs/f2fs/segment.c
@@ -25,7 +25,6 @@
#define __reverse_ffz(x) __reverse_ffs(~(x))

static struct kmem_cache *discard_entry_slab;
-static struct kmem_cache *flush_cmd_slab;

/*
* __reverse_ffs is copied from include/asm-generic/bitops/__ffs.h since
@@ -238,30 +237,28 @@ repeat:
int f2fs_issue_flush(struct f2fs_sb_info *sbi)
{
struct flush_cmd_control *fcc = SM_I(sbi)->cmd_control_info;
- struct flush_cmd *cmd;
- int ret;
+ struct flush_cmd cmd;

if (!test_opt(sbi, FLUSH_MERGE))
return blkdev_issue_flush(sbi->sb->s_bdev, GFP_KERNEL, NULL);

- cmd = f2fs_kmem_cache_alloc(flush_cmd_slab, GFP_ATOMIC | __GFP_ZERO);
- init_completion(&cmd->wait);
+ init_completion(&cmd.wait);
+ cmd.next = NULL;

spin_lock(&fcc->issue_lock);
if (fcc->issue_list)
- fcc->issue_tail->next = cmd;
+ fcc->issue_tail->next = &cmd;
else
- fcc->issue_list = cmd;
- fcc->issue_tail = cmd;
+ fcc->issue_list = &cmd;
+ fcc->issue_tail = &cmd;
spin_unlock(&fcc->issue_lock);

if (!fcc->dispatch_list)
wake_up(&fcc->flush_wait_queue);

- wait_for_completion(&cmd->wait);
- ret = cmd->ret;
- kmem_cache_free(flush_cmd_slab, cmd);
- return ret;
+ wait_for_completion(&cmd.wait);
+
+ return cmd.ret;
}

int create_flush_cmd_control(struct f2fs_sb_info *sbi)
@@ -2036,17 +2033,10 @@ int __init create_segment_manager_caches(void)
sizeof(struct discard_entry));
if (!discard_entry_slab)
return -ENOMEM;
- flush_cmd_slab = f2fs_kmem_cache_create("flush_command",
- sizeof(struct flush_cmd));
- if (!flush_cmd_slab) {
- kmem_cache_destroy(discard_entry_slab);
- return -ENOMEM;
- }
return 0;
}

void destroy_segment_manager_caches(void)
{
kmem_cache_destroy(discard_entry_slab);
- kmem_cache_destroy(flush_cmd_slab);
}
--
1.7.10.4


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/