Re: Report 2 in ext4 and journal based on v5.17-rc1

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Sat, Mar 05, 2022 at 11:15:38PM +0900, Byungchul Park wrote:
> On Fri, Mar 04, 2022 at 10:26:23PM -0500, Theodore Ts'o wrote:
> > On Fri, Mar 04, 2022 at 09:42:37AM +0900, Byungchul Park wrote:
> > > 
> > > All contexts waiting for any of the events in the circular dependency
> > > chain will be definitely stuck if there is a circular dependency as I
> > > explained. So we need another wakeup source to break the circle. In
> > > ext4 code, you might have the wakeup source for breaking the circle.
> > > 
> > > What I agreed with is:
> > > 
> > >    The case that 1) the circular dependency is unevitable 2) there are
> > >    another wakeup source for breadking the circle and 3) the duration
> > >    in sleep is short enough, should be acceptable.
> > > 
> > > Sounds good?
> > 
> > These dependencies are part of every single ext4 metadata update,
> > and if there were any unnecessary sleeps, this would be a major
> > performance gap, and this is a very well studied part of ext4.
> > 
> > There are some places where we sleep, sure.  In some case
> > start_this_handle() needs to wait for a commit to complete, and the
> > commit thread might need to sleep for I/O to complete.  But the moment
> > the thing that we're waiting for is complete, we wake up all of the
> > processes on the wait queue.  But in the case where we wait for I/O
> > complete, that wakeupis coming from the device driver, when it
> > receives the the I/O completion interrupt from the hard drive.  Is
> > that considered an "external source"?  Maybe DEPT doesn't recognize
> > that this is certain to happen just as day follows the night?  (Well,
> > maybe the I/O completion interrupt might not happen if the disk drive
> > bursts into flames --- but then, you've got bigger problems. :-)
> 
> Almost all you've been blaming at Dept are totally non-sense. Based on
> what you're saying, I'm conviced that you don't understand how Dept
> works even 1%. You don't even try to understand it before blame.
> 
> You don't have to understand and support it. But I can't response to you
> if you keep saying silly things that way.

Byungchul, other than ext4 have there been any DEPT reports that other
subsystem maintainers' agree were valid usecases?

Regarding false-positives, just to note lockdep is not without its share of
false-positives. Just that (as you know), the signal-to-noise ratio should be
high for it to be useful. I've put up with lockdep's false positives just
because it occasionally saves me from catastrophe.

> > In any case, if DEPT is going to report these "circular dependencies
> > as bugs that MUST be fixed", it's going to be pure noise and I will
> > ignore all DEPT reports, and will push back on having Lockdep replaced
> 
> Dept is going to be improved so that what you are concerning about won't
> be reported.

Yeah I am looking forward to learning more about it however I was wondering
about the following: lockdep can already be used for modeling "resource
acquire/release" and "resource wait" semantics that are unrelated to locks,
like we do in mm reclaim. I am wondering why we cannot just use those existing
lockdep mechanisms for the wait/wake usecases (Assuming that we can agree
that circular dependencies on related to wait/wake is a bad thing. Or perhaps
there's a reason why Peter Zijlstra did not use lockdep for wait/wake
dependencies (such as multiple wake sources) considering he wrote a lot of
that code.

Keep kicking ass brother, you're doing great.

Thanks,

     Joel




[Index of Archives]     [Linux Ext4 Filesystem]     [Union Filesystem]     [Filesystem Testing]     [Ceph Users]     [Ecryptfs]     [NTFS 3]     [AutoFS]     [Kernel Newbies]     [Share Photos]     [Security]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux Cachefs]     [Reiser Filesystem]     [Linux RAID]     [NTFS 3]     [Samba]     [Device Mapper]     [CEPH Development]

  Powered by Linux