[Cluster-devel] [GFS2 PATCH] GFS2: Insert some cond_resched statements into recovery

[Cluster-devel] [GFS2 PATCH] GFS2: Insert some cond_resched statements into recovery

[Bob Peterson]

Hi,

During recovery, GFS2 can be a CPU hog while it replays its journals.
This can prevent realtime processes like corosync from getting their
proper time slices. This patch inserts some cond_resched statements
into the recovery loops, and also while it searches for the proper
starting location in the journal.

Signed-off-by: Bob Peterson <rpeterso@redhat.com>
---
 fs/gfs2/lops.c     | 3 +++
 fs/gfs2/recovery.c | 2 ++
 2 files changed, 5 insertions(+)

diff --git a/fs/gfs2/lops.c b/fs/gfs2/lops.c
index b1f9144..aa7507c 100644
--- a/fs/gfs2/lops.c
+++ b/fs/gfs2/lops.c
@@ -566,6 +566,7 @@ static int buf_lo_scan_elements(struct gfs2_jdesc *jd, unsigned int start,
 			break;
 
 		jd->jd_replayed_blocks++;
+		cond_resched();
 	}
 
 	return error;
@@ -722,6 +723,7 @@ static int revoke_lo_scan_elements(struct gfs2_jdesc *jd, unsigned int start,
 		brelse(bh);
 		offset = sizeof(struct gfs2_meta_header);
 		first = 0;
+		cond_resched();
 	}
 
 	return 0;
@@ -802,6 +804,7 @@ static int databuf_lo_scan_elements(struct gfs2_jdesc *jd, unsigned int start,
 		brelse(bh_ip);
 
 		jd->jd_replayed_blocks++;
+		cond_resched();
 	}
 
 	return error;
diff --git a/fs/gfs2/recovery.c b/fs/gfs2/recovery.c
index 113b609..b5873f7 100644
--- a/fs/gfs2/recovery.c
+++ b/fs/gfs2/recovery.c
@@ -287,6 +287,7 @@ int gfs2_find_jhead(struct gfs2_jdesc *jd, struct gfs2_log_header_host *head)
 			blk_1 = blk_m;
 		else
 			blk_2 = blk_m;
+		cond_resched();
 	}
 
 	error = jhead_scan(jd, &lh_1);
@@ -602,6 +603,7 @@ int gfs2_recover_journal(struct gfs2_jdesc *jd, bool wait)
 	rv = queue_work(gfs_recovery_wq, &jd->jd_work);
 	BUG_ON(!rv);
 
+	cond_resched();
 	if (wait)
 		wait_on_bit(&jd->jd_flags, JDF_RECOVERY,
 			    TASK_UNINTERRUPTIBLE);

[Cluster-devel] [GFS2 PATCH] GFS2: Insert some cond_resched statements into recovery

[Steven Whitehouse]

Hi,

That generally looks ok, but not sure why we would need to have a 
cond_resched() in gfs2_recover_journal() since that only schedules work 
on a work queue. Even if it is doing that for several different 
journals, that should not take very long. The other cond_resched() calls 
all make sense to me though,

Steve.


On 19/04/17 19:50, Bob Peterson wrote:
> Hi,
>
> During recovery, GFS2 can be a CPU hog while it replays its journals.
> This can prevent realtime processes like corosync from getting their
> proper time slices. This patch inserts some cond_resched statements
> into the recovery loops, and also while it searches for the proper
> starting location in the journal.
>
> Signed-off-by: Bob Peterson <rpeterso@redhat.com>
> ---
>   fs/gfs2/lops.c     | 3 +++
>   fs/gfs2/recovery.c | 2 ++
>   2 files changed, 5 insertions(+)
>
> diff --git a/fs/gfs2/lops.c b/fs/gfs2/lops.c
> index b1f9144..aa7507c 100644
> --- a/fs/gfs2/lops.c
> +++ b/fs/gfs2/lops.c
> @@ -566,6 +566,7 @@ static int buf_lo_scan_elements(struct gfs2_jdesc *jd, unsigned int start,
>   			break;
>   
>   		jd->jd_replayed_blocks++;
> +		cond_resched();
>   	}
>   
>   	return error;
> @@ -722,6 +723,7 @@ static int revoke_lo_scan_elements(struct gfs2_jdesc *jd, unsigned int start,
>   		brelse(bh);
>   		offset = sizeof(struct gfs2_meta_header);
>   		first = 0;
> +		cond_resched();
>   	}
>   
>   	return 0;
> @@ -802,6 +804,7 @@ static int databuf_lo_scan_elements(struct gfs2_jdesc *jd, unsigned int start,
>   		brelse(bh_ip);
>   
>   		jd->jd_replayed_blocks++;
> +		cond_resched();
>   	}
>   
>   	return error;
> diff --git a/fs/gfs2/recovery.c b/fs/gfs2/recovery.c
> index 113b609..b5873f7 100644
> --- a/fs/gfs2/recovery.c
> +++ b/fs/gfs2/recovery.c
> @@ -287,6 +287,7 @@ int gfs2_find_jhead(struct gfs2_jdesc *jd, struct gfs2_log_header_host *head)
>   			blk_1 = blk_m;
>   		else
>   			blk_2 = blk_m;
> +		cond_resched();
>   	}
>   
>   	error = jhead_scan(jd, &lh_1);
> @@ -602,6 +603,7 @@ int gfs2_recover_journal(struct gfs2_jdesc *jd, bool wait)
>   	rv = queue_work(gfs_recovery_wq, &jd->jd_work);
>   	BUG_ON(!rv);
>   
> +	cond_resched();
>   	if (wait)
>   		wait_on_bit(&jd->jd_flags, JDF_RECOVERY,
>   			    TASK_UNINTERRUPTIBLE);
>