Re: [PATCH 1/3] xfs: pass alloc flags through to xfs_extent_busy_flush()

[Dave Chinner <david@fromorbit.com>]

On Thu, Jun 15, 2023 at 11:51:09PM +0000, Wengang Wang wrote:
> 
> 
> > On Jun 15, 2023, at 4:33 PM, Dave Chinner <david@fromorbit.com> wrote:
> > 
> > On Thu, Jun 15, 2023 at 11:09:41PM +0000, Wengang Wang wrote:
> >> When mounting the problematic metadump with the patches, I see the following reported.
> >> 
> >> For more information about troubleshooting your instance using a console connection, see the documentation: https://docs.cloud.oracle.com/en-us/iaas/Content/Compute/References/serialconsole.htm#four
> >> =================================================
> >> [   67.212496] loop: module loaded
> >> [   67.214732] loop0: detected capacity change from 0 to 629137408
> >> [   67.247542] XFS (loop0): Deprecated V4 format (crc=0) will not be supported after September 2030.
> >> [   67.249257] XFS (loop0): Mounting V4 Filesystem af755a98-5f62-421d-aa81-2db7bffd2c40
> >> [   72.241546] XFS (loop0): Starting recovery (logdev: internal)
> >> [   92.218256] XFS (loop0): Internal error ltbno + ltlen > bno at line 1957 of file fs/xfs/libxfs/xfs_alloc.c.  Caller xfs_free_ag_extent+0x3f6/0x870 [xfs]
> >> [   92.249802] CPU: 1 PID: 4201 Comm: mount Not tainted 6.4.0-rc6 #8
> > 
> > What is the test you are running? Please describe how you reproduced
> > this failure - a reproducer script would be the best thing here.
> 
> I was mounting a (copy of) V4 metadump from customer.

Is the metadump obfuscated? Can I get a copy of it via a private,
secure channel?

> > Does the test fail on a v5 filesytsem?
> 
> N/A.
> 
> > 
> >> I think that’s because that the same EFI record was going to be freed again
> >> by xfs_extent_free_finish_item() after it already got freed by xfs_efi_item_recover().

How is this happening? Where (and why) are we defering an extent we
have successfully freed into a new xefi that we create a new intent
for and then defer?

Can you post the debug output and analysis that lead you to this
observation? I certainly can't see how this can happen from looking
at the code

> >> I was trying to fix above issue in my previous patch by checking the intent
> >> log item’s lsn and avoid running iop_recover() in xlog_recover_process_intents().
> >> 
> >> Now I am thinking if we can pass a flag, say XFS_EFI_PROCESSED, from
> >> xfs_efi_item_recover() after it processed that record to the xfs_efi_log_item
> >> memory structure somehow. In xfs_extent_free_finish_item(), we skip to process
> >> that xfs_efi_log_item on seeing XFS_EFI_PROCESSED and return OK. By that
> >> we can avoid the double free.
> > 
> > I'm not really interested in speculation of the cause or the fix at
> > this point. I want to know how the problem is triggered so I can
> > work out exactly what caused it, along with why we don't have
> > coverage of this specific failure case in fstests already.
> > 
> 
> I get to know the cause by adding additional debug log along with
> my previous patch.

Can you please post that debug and analysis, rather than just a
stack trace that is completely lacking in context? Nothing can be
inferred from a stack trace, and what you are saying is occurring
does not match what the code should actually be doing. So I need to
actually look at what is happening in detail to work out where this
mismatch is coming from....

-Dave.
-- 
Dave Chinner
david@fromorbit.com