[lustre-devel] [PATCH 0/6] dcache/namei fixes for lustre

[James Simmons]

> >> This series is a revised version of two patches I sent
> >> previously (one of which was sadly broken).
> >> That patch has been broken into multiple parts for easy
> >> review.  The other is included unchanged as the last of
> >> this series.
> >> 
> >> I was drawn to look at this code due to the tests on
> >> DCACHE_DISCONNECTED which are often wrong, and it turns out
> >> they are used wrongly in lustre too.  Fixing one led to some
> >> clean-up.  Fixing the other is straight forward.
> >> 
> >> A particular change here from the previous posting is
> >> the first patch which tests for DCACHE_PAR_LOOKUP in ll_dcompare().
> >> Without this patch, two threads can be looking up the same
> >> name in a given directory in parallel.  This parallelism lead
> >> to my concerns about needing improved locking in ll_splice_alias().
> >> Instead of improving the locking, I now avoid the need for it
> >> by fixing ll_dcompare.
> >> 
> >> This code passes basic "smoke tests".
> >> 
> >> Note that the cast to "struct dentry *" in the first patch is because
> >> we have a "const struct dentry *" but d_in_lookup() requires a
> >> pointer to a non-const structure.  I'll send a separate patch to
> >> change d_in_lookup().
> >
> > To let you know this patch has been under going testing and we have a
> > ticket open to track the progess:
> >
> > https://jira.hpdd.intel.com/browse/LU-9868
> >
> > Your patch did reveal that a piece of a fix landed earlier is missing :-(
> > So currently the client can oops. I will send the fix shortly but this
> > work will have to rebased after. As soon as we can get some cycles we will
> > figure out what is going on. Thanks for helping out.
> 
> Hi,
>  what happened about this?  I had a look around the ticket and couldn't
>  find anything about an oops.  If there is still a problem I'd be very
>  happy to help work out what it is - but I don't know where to look.

The oops is specific to the in kernel client. Some where along the way the
calls to ll_d_init() were removed from ll_splice_alias(). It was unnoticed
until your patch came along. I do have a fix that I will be pushing to 
the next staging tree very shortly.

I have been testing the patch series and for me I don't see any issue. Our 
test suite is reporting failures with this patch which I'm attempting to 
figure out how to reproduce locally on my test system. Once I have a 
reproducer I can send it to you.