This is an attempt to fix the softlock on the delayed_work workqueue. As
stated in 0002 patch:
Function ceph_check_delayed_caps() is called from the mdsc->delayed_work
workqueue and it can be kept looping for quite some time if caps keep being
added back to the mdsc->cap_delay_list. This may result in the watchdog
tainting the kernel with the softlockup flag.
v2 of this fix modifies the approach by time-bounding the loop in this
function, so that any caps added to the list *after* the loop starts will
be postponed to the next wq run.
An extra change in 0001 (suggested by Jeff) allows scheduling runs for
periods smaller than the default (5 secs) period. This way,
delayed_work() can have the next run scheduled for the next list element
ci->i_hold_caps_max instead of 5 secs.
This patchset should fix the issue reported here [1], although a quick
search for "ceph_check_delayed_caps" in the tracker returns a few more
bugs, possibly duplicates.
[1] https://tracker.ceph.com/issues/46284
Luis Henriques (2):
ceph: allow schedule_delayed() callers to set delay for workqueue
ceph: reduce contention in ceph_check_delayed_caps()
fs/ceph/caps.c | 17 ++++++++++++++++-
fs/ceph/mds_client.c | 24 +++++++++++++++---------
fs/ceph/super.h | 2 +-
3 files changed, 32 insertions(+), 11 deletions(-)
Function ceph_check_delayed_caps() is called from the mdsc->delayed_work
workqueue and it can be kept looping for quite some time if caps keep
being added back to the mdsc->cap_delay_list. This may result in the
watchdog tainting the kernel with the softlockup flag.
This patch breaks this loop if the caps have been recently (i.e. during
the loop execution). Any new caps added to the list will be handled in
the next run.
Cc: [email protected]
Link: https://tracker.ceph.com/issues/46284
Signed-off-by: Luis Henriques <[email protected]>
---
fs/ceph/caps.c | 17 ++++++++++++++++-
fs/ceph/mds_client.c | 7 ++++---
fs/ceph/super.h | 2 +-
3 files changed, 21 insertions(+), 5 deletions(-)
diff --git a/fs/ceph/caps.c b/fs/ceph/caps.c
index a5e93b185515..c79b8dff25d7 100644
--- a/fs/ceph/caps.c
+++ b/fs/ceph/caps.c
@@ -4224,11 +4224,19 @@ void ceph_handle_caps(struct ceph_mds_session *session,
/*
* Delayed work handler to process end of delayed cap release LRU list.
+ *
+ * If new caps are added to the list while processing it, these won't get
+ * processed in this run. In this case, the ci->i_hold_caps_max will be
+ * returned so that the work can be scheduled accordingly.
*/
-void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
+unsigned long ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
{
struct inode *inode;
struct ceph_inode_info *ci;
+ struct ceph_mount_options *opt = mdsc->fsc->mount_options;
+ unsigned long delay_max = opt->caps_wanted_delay_max * HZ;
+ unsigned long loop_start = jiffies;
+ unsigned long delay = 0;
dout("check_delayed_caps\n");
spin_lock(&mdsc->cap_delay_lock);
@@ -4236,6 +4244,11 @@ void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
ci = list_first_entry(&mdsc->cap_delay_list,
struct ceph_inode_info,
i_cap_delay_list);
+ if (time_before(loop_start, ci->i_hold_caps_max - delay_max)) {
+ dout("%s caps added recently. Exiting loop", __func__);
+ delay = ci->i_hold_caps_max;
+ break;
+ }
if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
time_before(jiffies, ci->i_hold_caps_max))
break;
@@ -4252,6 +4265,8 @@ void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
}
}
spin_unlock(&mdsc->cap_delay_lock);
+
+ return delay;
}
/*
diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
index 08c76bf57fb1..5f1bd7f9dce1 100644
--- a/fs/ceph/mds_client.c
+++ b/fs/ceph/mds_client.c
@@ -4518,11 +4518,12 @@ static void schedule_delayed(struct ceph_mds_client *mdsc, unsigned long delay)
static void delayed_work(struct work_struct *work)
{
- int i;
struct ceph_mds_client *mdsc =
container_of(work, struct ceph_mds_client, delayed_work.work);
+ unsigned long delay;
int renew_interval;
int renew_caps;
+ int i;
dout("mdsc delayed_work\n");
@@ -4562,7 +4563,7 @@ static void delayed_work(struct work_struct *work)
}
mutex_unlock(&mdsc->mutex);
- ceph_check_delayed_caps(mdsc);
+ delay = ceph_check_delayed_caps(mdsc);
ceph_queue_cap_reclaim_work(mdsc);
@@ -4570,7 +4571,7 @@ static void delayed_work(struct work_struct *work)
maybe_recover_session(mdsc);
- schedule_delayed(mdsc, 0);
+ schedule_delayed(mdsc, delay);
}
int ceph_mdsc_init(struct ceph_fs_client *fsc)
diff --git a/fs/ceph/super.h b/fs/ceph/super.h
index 839e6b0239ee..3b5207c82767 100644
--- a/fs/ceph/super.h
+++ b/fs/ceph/super.h
@@ -1170,7 +1170,7 @@ extern void ceph_flush_snaps(struct ceph_inode_info *ci,
extern bool __ceph_should_report_size(struct ceph_inode_info *ci);
extern void ceph_check_caps(struct ceph_inode_info *ci, int flags,
struct ceph_mds_session *session);
-extern void ceph_check_delayed_caps(struct ceph_mds_client *mdsc);
+extern unsigned long ceph_check_delayed_caps(struct ceph_mds_client *mdsc);
extern void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc);
extern int ceph_drop_caps_for_unlink(struct inode *inode);
extern int ceph_encode_inode_release(void **p, struct inode *inode,
Allow schedule_delayed() callers to explicitly set the delay value instead
of defaulting to a 5 secs value.
Signed-off-by: Luis Henriques <[email protected]>
---
fs/ceph/mds_client.c | 19 ++++++++++++-------
1 file changed, 12 insertions(+), 7 deletions(-)
diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
index e5af591d3bd4..08c76bf57fb1 100644
--- a/fs/ceph/mds_client.c
+++ b/fs/ceph/mds_client.c
@@ -4502,13 +4502,18 @@ void inc_session_sequence(struct ceph_mds_session *s)
}
/*
- * delayed work -- periodically trim expired leases, renew caps with mds
+ * delayed work -- periodically trim expired leases, renew caps with mds. If
+ * the @delay parameter is set to 0 or if it's more than 5 secs, the default
+ * workqueue delay value of 5 secs will be used.
*/
-static void schedule_delayed(struct ceph_mds_client *mdsc)
+static void schedule_delayed(struct ceph_mds_client *mdsc, unsigned long delay)
{
- int delay = 5;
- unsigned hz = round_jiffies_relative(HZ * delay);
- schedule_delayed_work(&mdsc->delayed_work, hz);
+ unsigned long max_delay = round_jiffies_relative(HZ * 5);
+
+ /* 5 secs default delay */
+ if (!delay || (delay > max_delay))
+ delay = max_delay;
+ schedule_delayed_work(&mdsc->delayed_work, delay);
}
static void delayed_work(struct work_struct *work)
@@ -4565,7 +4570,7 @@ static void delayed_work(struct work_struct *work)
maybe_recover_session(mdsc);
- schedule_delayed(mdsc);
+ schedule_delayed(mdsc, 0);
}
int ceph_mdsc_init(struct ceph_fs_client *fsc)
@@ -5042,7 +5047,7 @@ void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
mdsc->mdsmap->m_epoch);
mutex_unlock(&mdsc->mutex);
- schedule_delayed(mdsc);
+ schedule_delayed(mdsc, 0);
return;
bad_unlock:
On Tue, Jun 29, 2021 at 12:14:37PM +0200, Ilya Dryomov wrote:
> On Tue, Jun 29, 2021 at 11:47 AM Luis Henriques <[email protected]> wrote:
> >
> > Allow schedule_delayed() callers to explicitly set the delay value instead
> > of defaulting to a 5 secs value.
> >
> > Signed-off-by: Luis Henriques <[email protected]>
> > ---
> > fs/ceph/mds_client.c | 19 ++++++++++++-------
> > 1 file changed, 12 insertions(+), 7 deletions(-)
> >
> > diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
> > index e5af591d3bd4..08c76bf57fb1 100644
> > --- a/fs/ceph/mds_client.c
> > +++ b/fs/ceph/mds_client.c
> > @@ -4502,13 +4502,18 @@ void inc_session_sequence(struct ceph_mds_session *s)
> > }
> >
> > /*
> > - * delayed work -- periodically trim expired leases, renew caps with mds
> > + * delayed work -- periodically trim expired leases, renew caps with mds. If
> > + * the @delay parameter is set to 0 or if it's more than 5 secs, the default
> > + * workqueue delay value of 5 secs will be used.
> > */
> > -static void schedule_delayed(struct ceph_mds_client *mdsc)
> > +static void schedule_delayed(struct ceph_mds_client *mdsc, unsigned long delay)
> > {
> > - int delay = 5;
> > - unsigned hz = round_jiffies_relative(HZ * delay);
> > - schedule_delayed_work(&mdsc->delayed_work, hz);
> > + unsigned long max_delay = round_jiffies_relative(HZ * 5);
> > +
> > + /* 5 secs default delay */
> > + if (!delay || (delay > max_delay))
> > + delay = max_delay;
> > + schedule_delayed_work(&mdsc->delayed_work, delay);
>
> Hi Luis,
>
> Is there a reason to not round the non-default delay? Does it need to
> be precise?
Ah, right. No, there's not reason for not rounding it too. So, I could
have something like this instead:
unsigned long max_delay = HZ * 5;
/* 5 secs default delay */
if (!delay || (delay > max_delay))
delay = max_delay;
schedule_delayed_work(&mdsc->delayed_work, round_jiffies_relative(delay));
I'll make sure v3 will include this change. Thanks Ilya.
Cheers,
--
Lu?s
On Tue, Jun 29, 2021 at 11:47 AM Luis Henriques <[email protected]> wrote:
>
> Allow schedule_delayed() callers to explicitly set the delay value instead
> of defaulting to a 5 secs value.
>
> Signed-off-by: Luis Henriques <[email protected]>
> ---
> fs/ceph/mds_client.c | 19 ++++++++++++-------
> 1 file changed, 12 insertions(+), 7 deletions(-)
>
> diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
> index e5af591d3bd4..08c76bf57fb1 100644
> --- a/fs/ceph/mds_client.c
> +++ b/fs/ceph/mds_client.c
> @@ -4502,13 +4502,18 @@ void inc_session_sequence(struct ceph_mds_session *s)
> }
>
> /*
> - * delayed work -- periodically trim expired leases, renew caps with mds
> + * delayed work -- periodically trim expired leases, renew caps with mds. If
> + * the @delay parameter is set to 0 or if it's more than 5 secs, the default
> + * workqueue delay value of 5 secs will be used.
> */
> -static void schedule_delayed(struct ceph_mds_client *mdsc)
> +static void schedule_delayed(struct ceph_mds_client *mdsc, unsigned long delay)
> {
> - int delay = 5;
> - unsigned hz = round_jiffies_relative(HZ * delay);
> - schedule_delayed_work(&mdsc->delayed_work, hz);
> + unsigned long max_delay = round_jiffies_relative(HZ * 5);
> +
> + /* 5 secs default delay */
> + if (!delay || (delay > max_delay))
> + delay = max_delay;
> + schedule_delayed_work(&mdsc->delayed_work, delay);
Hi Luis,
Is there a reason to not round the non-default delay? Does it need to
be precise?
Thanks,
Ilya