[PATCH 00/11 v3] xfs: inode reclaim vs the world

[Dave Chinner]

Hi folks,

The inode reclaim patchset has grown somewhat larger with all the
fixes I've accrued over the past couple of days for issues that have
been reported. I've just re-ordered the patchset to have all the bug
fixes that need to go into 4.6-rc up front, followed by other bug
fixes, followed by all the cleanups I wrote as I went along fixing
things. I want to get this out before LSFMM, so I'm posting this
before I've done a full verification that I've got the patches
correct, and they are still running through xfstests right now.

Patch 1 is new - it fixes a regression introduced in 4.6-rc1 and is
caused by clearing the vfs inode i_mode in ->evict_inode, when there
is code in evict() that requires the mode to be intact until
->destroy_inode.

Patch 2 is new - if fixes the inode log item leak that left a dirty
log item on the AIL after the inode had been reclaimed, resulting in
unmountable filesystems and a couple of use-after-free vectors.

Patch 3 is the original fix for the xfs_iflush_cluster lookup
validity checks that were incorect

Patch 4 is the original fix for avoiding flushing stale inodes.

Patch 5 is new, and comes from Alex @ Zadara to avoid having to
reallocate memory when tearing down the inode extent lists. This is
a necessary pre-requisite for patch 6.

Patch 6 is the original patch that pushed all the inode memory
freeing into the RCU callback, so that xfs_iflush_cluster didn't
reference freed memory when racing with reclaim.

Patch 7 is the original patch that made reclaim races eaasier to
detect.

Patch 8 is the original patch that drops out of xfs-iflush_cluster
on the first inode beyond the end of the current cluster.

Patch 9 is the original patch that renames the variables in
xfs_iflush_cluster().

Patch 10 and 11 are new patches that simplify the inode reclaim
tagging interfaces to remove dependencies on the struct xfs_inode
and the inode number where they are not actually required.

-Dave.