[PATCH 0/3] xfs: fix xfs_extent_busy_flush() deadlock in EFI processing

[Dave Chinner <david@fromorbit.com>]

Hi folks,

This patchset is largely a rework of the patch that Wengang posted
here:

https://lore.kernel.org/linux-xfs/20230519171829.4108-1-wen.gang.wang@oracle.com/

Review has run aground because I simply don't have the time to
explain every deep, subtle corner case on every issue that is raised
before progress can be made fixing this issue. Indeed, this is the
second attempt to get this bug fixed that has run aground like this.

Hence I've decided that it's less time and effort to just take what
we have, split it, clean it up, fixed it up, remove all the
unnecessary bits and run it through testing and push it back out for
further review.

I split the alloc flags out as a separate variable that is passed
down the stack; I renamed them all "alloc_flags" so that it's clear
that the flags being used and passed between the functions are the
XFS_ALLOC_FLAG* flag values. If we want to, it will now be trivial
to pull these back into the struct xfs_alloc_arg if we so desire -
there is no mix-and-match of args->flags and function paramter flags
to confuse the issue as was in the original patch.

The changes to xfs_extent_busy_flush() mean that it can now return
-EFSCORRUPTED or -EIO from xfs_log_force(), not just -EAGAIN to
indicate the transaction must be committed before the allocation is
retried. This has been exercised by recoveryloop testing and it
appears that the new corruption/error detection conditions do not
introduce any new failures.

Cheers,

Dave.