Re: [PATCH v2] xfs: use per-AG reservations for the finobt

["Darrick J. Wong" <darrick.wong@oracle.com>]

On Tue, Jan 24, 2017 at 08:50:50PM +0100, Christoph Hellwig wrote:
> On Tue, Jan 24, 2017 at 08:48:37AM -0800, Darrick J. Wong wrote:
> > I'd just print out:
> > 
> > "Per-AG reservation for AG ${agno} failed.  Filesystem may run out of space."
> 
> Sure, I can do that.
> 
> > 
> > if any of the (now three) invocations of __xfs_ag_resv_init() returns
> > ENOSPC, since the AG is more than 98% full.  Some day, a sysadmin could
> > then use the fsmap data to identify which inodes are using the most
> > space in that AG and arrange to migrate/remove/etc the data to another
> > AG.
> 
> But for all but the reflink one this can't happen as the reservations
> have been there since the beginning.  And if they fail neverless we fail
> the mount.  Or am I missing something here?

The code I wrote doesn't fail the mount (or remount) on ENOSPC so that
it's still possible to delete files off the fs even when space is low.

In theory it wasn't possible to run out, but I've been thinking that
it's better to warn noisily just in case we /do/ run out of space later,
since it's not entirely obvious when the fs goes offline because we
tried to expand a btree and hit ENOSPC.

(Also if people perform underhanded "upgrades" then they can hit this
problem, but maybe that was just me. :P)

--D