vfs: don't softlockup when evicting inodes
PATCH v2 1/1 vfs: don't softlockup when evicting inodes
If I run an fs_mark job that creates millions of empty files and then immediately unmount the file system I will get a softlockup during unmount. This box has ~140gb of RAM so we never hit sufficient memory pressure to evict enough inodes during the runtime of the benchmark, which means I see around 80 million inodes being evicted at unmount time. With this patch my box no longer softlocks up. Thanks,
---V1->V2:
-Still occasionally saw a softlockup in evict_inodes so add a cond_resched_lock
to that case as well.
fs/inode.c | 2 ++ 1 file changed, 2 insertions(+)
fs/inode.c +2 -0
@@ -581,6 +581,7 @@ static void dispose_list(struct list_head *head)list_del_init (&inode->i_lru);evict (inode); +cond_resched (); } }
@@ -613,6 +614,7 @@ void evict_inodes(struct super_block *sb)inode_lru_list_del (inode);spin_unlock (&inode->i_lock);list_add (&inode->i_lru, &dispose); +cond_resched_lock (&inode_sb_list_lock); }spin_unlock (&inode_sb_list_lock);
Signature · 1 line
1.9.3
Quoted 29 lines · On Fri, 19 Dec 2014 10:34:40 -0500 Josef Bacik <[email protected]> wrote:
If I run an
fs_markjob that creates millions of empty files and then immediately unmount the file system I will get a softlockup during unmount. This box has ~140gb of RAM so we never hit sufficient memory pressure to evict enough inodes during the runtime of the benchmark, which means I see around 80 million inodes being evicted at unmount time. With this patch my box no longer softlocks up. Thanks,Signed-off-by Josef Bacik <[email protected]>---
V1->V2:(quoted diff) +0
-1-Still occasionally saw a softlockup in evict_inodes so add a cond_resched_lock to that case as well. fs/inode.c | 2 ++ 1 file changed, 2 insertions(+)fs/inode.c +1
-0@@ -581,6 +581,7 @@ static void dispose_list(struct list_head *head)list_del_init (&inode->i_lru);evict (inode); +cond_resched (); } }
fs/inode.c void evict_inodes(struct super_block *sb) · line 620
Quoted 7 lines
fs/inode.c +1
-0@@ -613,6 +614,7 @@ void evict_inodes(struct super_block *sb)inode_lru_list_del (inode);spin_unlock (&inode->i_lock);list_add (&inode->i_lru, &dispose); +cond_resched_lock (&inode_sb_list_lock); }spin_unlock (&inode_sb_list_lock);
Is that safe? What guarantees that the next entry in the list (that is, the one already prefetched by list_for_each_entry_safe) will still be there once you drop and reacquire the lock?
Signature · 1 line
Jeff Layton <[email protected]>
I had convinced myself it was fine but if we have a final iput of a deleted inode it could screw us, I'll fix this to restart the loop if we resched. Thanks,
Josef
Jeff Layton <[email protected]> wrote:
Quoted 29 lines · On Fri, 19 Dec 2014 10:34:40 -0500 Josef Bacik <[email protected]> wrote:
If I run an
fs_markjob that creates millions of empty files and then immediately unmount the file system I will get a softlockup during unmount. This box has ~140gb of RAM so we never hit sufficient memory pressure to evict enough inodes during the runtime of the benchmark, which means I see around 80 million inodes being evicted at unmount time. With this patch my box no longer softlocks up. Thanks,Signed-off-by Josef Bacik <[email protected]>---
V1->V2:(quoted diff) +0
-1-Still occasionally saw a softlockup in evict_inodes so add a cond_resched_lock to that case as well. fs/inode.c | 2 ++ 1 file changed, 2 insertions(+)fs/inode.c +1
-0@@ -581,6 +581,7 @@ static void dispose_list(struct list_head *head)list_del_init (&inode->i_lru);evict (inode); +cond_resched (); } }
fs/inode.c void evict_inodes(struct super_block *sb) · line 620
Quoted 7 lines
fs/inode.c +1
-0@@ -613,6 +614,7 @@ void evict_inodes(struct super_block *sb)inode_lru_list_del (inode);spin_unlock (&inode->i_lock);list_add (&inode->i_lru, &dispose); +cond_resched_lock (&inode_sb_list_lock); }spin_unlock (&inode_sb_list_lock);
Is that safe? What guarantees that the next entry in the list (that is, the one already prefetched by list_for_each_entry_safe) will still be there once you drop and reacquire the lock?
--
Jeff Layton <[email protected]>