PATCH] fix potential latency issues in JBD's journal code
From: Arjan van de Ven <[email protected]>
Date: Fri, 4 Jul 2008 19:26:16 -0700
Subject: [PATCH] fix potential latency issues in JBD's journal code
CC: [email protected]
CC: [email protected]
This is a follow-on to commit 18ce3751ccd488c78d3827e9f6bf54e6322676fb (Properly notify block layer of sync writes)
which fixed some severe latency issues due to a "submit IO and then wait for it" pattern, which got fixed by properly informing the block layer that the IOs in question are going to be waited on immediately after a batch submission.
In the JBD layer, some of the core journal routines have the exact same pattern of code, and... surprising (or not)... they're also not using the WRITE_SYNC variant to inform the blocklayer.
This patch modifies two key places that submit IO that then immediately will get waited on. The JBD code is slightly convoluted, but after some chasing of abstraction layers, these instances seem to really be of this pattern. There's one case in checkpoint.c which is another candidate, but I've not been able to get my head around the code enough to verify that that one really is of this pattern as well, so for now I'll leave that one as is.
--- fs/jbd/commit.c | 4 ++-- fs/jbd/revoke.c | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-)
fs/jbd/commit.c +2 -2
@@ -165,7 +165,7 @@ static void journal_do_submit_data(struct buffer_head **wbuf, int bufs)for (i =0 ; i < bufs; i++) { wbuf[i]->b_end_io = end_buffer_write_sync;/* We use-up our safety reference in submit_bh() */ -submit_bh (WRITE , wbuf[i]); +submit_bh (WRITE_SYNC , wbuf[i]); } }
@@ -622,7 +622,7 @@ start_journal_io:clear_buffer_dirty (bh);set_buffer_uptodate (bh); bh->b_end_io = journal_end_buffer_io_sync; -submit_bh (WRITE , bh); +submit_bh (WRITE_SYNC , bh); }cond_resched ();
fs/jbd/revoke.c +1 -1
@@ -616,7 +616,7 @@ static void flush_descriptor(journal_t *journal,set_buffer_jwrite (bh);BUFFER_TRACE (bh,"write" );set_buffer_dirty (bh); - ll_rw_block(SWRITE, 1, &bh); + ll_rw_block(SWRITE_SYNC, 1, &bh); }#endif
Signature · 7 lines
1.5.5.1 -- If you want to reach me at my work email, use [email protected] For development, discussion and tips for power savings, visit http://www.lesswatts.org
Helps if I send the patch that compiles ;)
From: Arjan van de Ven <[email protected]>
Subject: [PATCH] fix potential latency issues in JBD's journal code
This is a follow-on to commit 18ce3751ccd488c78d3827e9f6bf54e6322676fb
(Properly notify block layer of sync writes)
which fixed some severe latency issues due to a "submit IO and then wait for it" pattern, which got fixed by properly informing the block layer that the IOs in question are going to be waited on immediately after a batch submission.
In the JBD layer, some of the core journal routines have the exact same pattern of code, and... surprising (or not)... they're also not using the WRITE_SYNC variant to inform the blocklayer.
This patch modifies two key places that submit IO that then immediately will get waited on. The JBD code is slightly convoluted, but after some chasing of abstraction layers, these instances seem to really be of this pattern. There's one case in checkpoint.c which is another candidate, but I've not been able to get my head around the code enough to verify that this one really is of this pattern.
--- fs/jbd/commit.c | 5 +++-- fs/jbd/revoke.c | 3 ++- 2 files changed, 5 insertions(+), 3 deletions(-)
fs/jbd/commit.c +3 -2
@@ -15,6 +15,7 @@#include <linux/time.h> #include <linux/fs.h> +#include <linux/bio.h> #include <linux/jbd.h> #include <linux/errno.h> #include <linux/slab.h>
@@ -165,7 +166,7 @@ static void journal_do_submit_data(struct buffer_head **wbuf, int bufs)for (i =0 ; i < bufs; i++) { wbuf[i]->b_end_io = end_buffer_write_sync;/* We use-up our safety reference in submit_bh() */ -submit_bh (WRITE , wbuf[i]); +submit_bh (WRITE_SYNC , wbuf[i]); } }
@@ -622,7 +623,7 @@ start_journal_io:clear_buffer_dirty (bh);set_buffer_uptodate (bh); bh->b_end_io = journal_end_buffer_io_sync; -submit_bh (WRITE , bh); +submit_bh (WRITE_SYNC , bh); }cond_resched ();
fs/jbd/revoke.c +2 -1
@@ -67,6 +67,7 @@#include <linux/slab.h> #include <linux/list.h> #include <linux/init.h> +#include <linux/bio.h> #endif #include <linux/log2.h>
@@ -616,7 +617,7 @@ static void flush_descriptor(journal_t *journal,set_buffer_jwrite (bh);BUFFER_TRACE (bh,"write" );set_buffer_dirty (bh); - ll_rw_block(SWRITE, 1, &bh); + ll_rw_block(SWRITE_SYNC, 1, &bh); }#endif
Signature · 8 lines
1.5.5.1 -- If you want to reach me at my work email, use [email protected] For development, discussion and tips for power savings, visit http://www.lesswatts.org
Theodore Tso
Acked-byI will make the corresponding change for ext4. Is the "Properly notify block layer of sync writes" going to be pushed to mainline? I can't add SWRITE_SYNC to the ext4 tree without growing a dependency to this patch.
Quoted 6 lines
This patch modifies two key places that submit IO that then immediately will get waited on. The JBD code is slightly convoluted, but after some chasing of abstraction layers, these instances seem to really be of this pattern. There's one case in
checkpoint.cwhich is another candidate, but I've not been able to get my head around the code enough to verify that that one really is of this pattern as well, so for now I'll leave that one as is.
You mean the ll_rw_block() SWRITE in __flush_patch? Yep, seems to be of the same pattern.
- Ted
Quoted 6 lines · On Sat, 5 Jul 2008 10:13:35 -0400 Theodore Tso <[email protected]> wrote:
Acked-by "Theodore Ts'o" <[email protected]>I will make the corresponding change for ext4. Is the "Properly notify block layer of sync writes" going to be pushed to mainline? I can't add
SWRITE_SYNCto the ext4 tree without growing a dependency to this patch.
you have stale bits .. it's in Mainline for like at least 24 hours already ;=)
Signature · 3 lines
If you want to reach me at my work email, use [email protected] For development, discussion and tips for power savings, visit http://www.lesswatts.org
fs/jbd/commit.c static void journal_do_submit_data(struct buffer_head **wbuf, int bufs) · line 170
Quoted 8 lines · Arjan van de Ven
fs/jbd/commit.c +1
-1@@ -165,7 +165,7 @@ static void journal_do_submit_data(struct buffer_head **wbuf, int bufs)for (i =0 ; i < bufs; i++) { wbuf[i]->b_end_io = end_buffer_write_sync;/* We use-up our safety reference in submit_bh() */ -submit_bh (WRITE , wbuf[i]); +submit_bh (WRITE_SYNC , wbuf[i]); } }
So I started looking at this patch more closely when trying to replicate it for ext4. Don't you want to only use WRITE_SYNC() only for the very last time in the loop? Otherwise you end up unplugging the queue after each bufferhead, which wouldn't be a good thing, right?
- Ted
Quoted 15 lines · On Sun, 6 Jul 2008 00:15:02 -0400 Theodore Tso <[email protected]> wrote:
fs/jbd/commit.c static void journal_do_submit_data(struct · line 170
Quoted 8 lines · Arjan van de Ven
fs/jbd/commit.c +1
-1@@ -165,7 +165,7 @@ static void journal_do_submit_data(struct buffer_head **wbuf, int bufs) for (i = 0; i < bufs; i++) { wbuf[i]->b_end_io = end_buffer_write_sync;/* We use-up our safety reference in submit_bh() */ -submit_bh (WRITE , wbuf[i]); +submit_bh (WRITE_SYNC , wbuf[i]); } }So I started looking at this patch more closely when trying to replicate it for ext4. Don't you want to only use
WRITE_SYNC()only for the very last time in the loop? Otherwise you end up unplugging the queue after each bufferhead, which wouldn't be a good thing, right?
it's debatable. Because this submit-bh() will sometimes block... it wouldn't get IO started for that case. I think this is where I'd like Jens to say what he thinks the rules are; clearly the elevator needs to treat all of these guys as sync, but maybe the plugging should be .. flexible.
Signature · 3 lines
If you want to reach me at my work email, use [email protected] For development, discussion and tips for power savings, visit http://www.lesswatts.org
Quoted 5 lines · Arjan van de Ven
it's debatable. Because this submit-bh() will sometimes block... it wouldn't get IO started for that case. I think this is where I'd like Jens to say what he thinks the rules are; clearly the elevator needs to treat all of these guys as sync, but maybe the plugging should be .. flexible.
Stupid idea.... maybe the unplugging should be done on the wait side? Has this been considered already?
- Ted