On Tue, May 26, 2020 at 02:16:29PM -0400, Brian Foster wrote:
> On Fri, May 22, 2020 at 06:36:14PM -0700, Darrick J. Wong wrote:
> > On Fri, May 22, 2020 at 01:18:28PM -0400, Brian Foster wrote:
> > > Signed-off-by: Brian Foster <bfoster@xxxxxxxxxx>
> > > ---
> > >
> > > Darrick mentioned on IRC a few days ago that he'd seen an issue that
> > > looked similar to the problem with the rmapbt based extent swap
> > > algorithm when the associated inodes happen to bounce between extent and
> > > btree format. That problem caused repeated bmapbt block allocations and
> > > frees that exhausted the transaction block reservation across the
> > > sequence of transaction rolls. The workaround for that was to use an
> > > oversized block reservation, but that is not a generic or efficient
> > > solution.
> > >
> > > I was originally playing around with some hacks to set an optional base
> > > block reservation on the transaction that we would attempt to replenish
> > > across transaction roll sequences as the block reservation depletes, but
> > > eventually noticed that there isn't much difference between stuffing
> > > block frees in the transaction reservation counter vs. the delta counter
> > > when lazy sb accounting is enabled (which is required for v5 supers). As
> > > such, the following patch seems to address the rmapbt issue in my
> > > isolated tests.
> > >
> > > I think one tradeoff with this logic is that chains of rolling/freeing
> > > transactions would now aggregate freed space until the final transaction
> > > commits vs. as transactions roll. It's not immediately clear to me how
> > > much of an issue that is, but it sounds a bit dicey when considering
> > > things like truncates of large files. This behavior could still be tied
> > > to a transaction flag to restrict its use to situations like rmapbt
> > > swapext, however. Anyways, this is mostly untested outside of the extent
> > > swap use case so I wanted to throw this on the list as an RFC for now
> > > and see if anybody has thoughts or other ideas.
> >
> > Hmm, well, this /would/ fix the immediate problem of running out of
> > block reservation, but I wonder if there are other weird subtleties.
> > If we're nearly out of space and we're mounted with -odiscard and the
> > disk is really slow at processing discard, can we encounter weird
> > failure cases where we end up stuck waiting for the extent busy tree to
> > say that one of our pingponged blocks is ok to use again?
> >
>
> Yeah, I think something like that could happen. I don't think it should
> be a failure scenario though as the busy extent list should involve a
> log force and retry in the worst case. Either way, we could always
> mitigate risk by making this an optional accounting mode for particular
> (extent swap) transactions...
Hmmm... OTOH I wonder how many people really run fsr? Even I don't...
:)
> > In the meantime, I noticed that xfs/227 on a pmem fs (or possibly
> > anything with synchronous writes?) and reflink+rmap enabled seemed to
> > fail pretty consistently. In a hastily done and incomprehensi{ve,ble}
> > survey I noted that I couldn't make the disastrous pingpong happen if
> > there were more than ~4 blocks in the bmapbt, so maybe this would help
> > there.
> >
>
> Do you mean with this patch or with current upstream? I don't see
> xfs/227 failures on my current setups (this patch passed a weekend auto
> test run), but I'll have to retry with something synchronous...
It happens semi-frequently with current upstream, and all the time with
the atomic file swap series.
> BTW, is xfs/227 related to the problem you had mentioned on IRC? I
> wasn't quite sure what operation was involved with whatever error report
> you had. xfs/227 looks like an xfs_fsr test, so I'd have thought the
> upstream workaround would have addressed that.. (though I see some attr
> ops in there as well so perhaps this is related to the attr fork..?).
It's related, but only in the sense that the "zomg hundreds of thousands
of intents sitting around in memory" were a side effect of creating a
test that creates two files with ~50000 extents and fsr'ing them.
--D
> Brian
>
> > In unrelated news, I also tried fixing the log recovery defer ops chain
> > transactions to absorb the unused block reservations that the
> > xfs_*i_item_recover functions created, but that just made fdblocks be
> > wrong. But it didn't otherwise blow up! :P
> >
> > --D
> >
> > > Brian
> > >
> > > fs/xfs/xfs_bmap_util.c | 11 -----------
> > > fs/xfs/xfs_trans.c | 4 ++++
> > > 2 files changed, 4 insertions(+), 11 deletions(-)
> > >
> > > diff --git a/fs/xfs/xfs_bmap_util.c b/fs/xfs/xfs_bmap_util.c
> > > index f37f5cc4b19f..74b3bad6c414 100644
> > > --- a/fs/xfs/xfs_bmap_util.c
> > > +++ b/fs/xfs/xfs_bmap_util.c
> > > @@ -1628,17 +1628,6 @@ xfs_swap_extents(
> > > */
> > > resblks = XFS_SWAP_RMAP_SPACE_RES(mp, ipnext, w);
> > > resblks += XFS_SWAP_RMAP_SPACE_RES(mp, tipnext, w);
> > > -
> > > - /*
> > > - * Handle the corner case where either inode might straddle the
> > > - * btree format boundary. If so, the inode could bounce between
> > > - * btree <-> extent format on unmap -> remap cycles, freeing and
> > > - * allocating a bmapbt block each time.
> > > - */
> > > - if (ipnext == (XFS_IFORK_MAXEXT(ip, w) + 1))
> > > - resblks += XFS_IFORK_MAXEXT(ip, w);
> > > - if (tipnext == (XFS_IFORK_MAXEXT(tip, w) + 1))
> > > - resblks += XFS_IFORK_MAXEXT(tip, w);
> > > }
> > > error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
> > > if (error)
> > > diff --git a/fs/xfs/xfs_trans.c b/fs/xfs/xfs_trans.c
> > > index 28b983ff8b11..b421d27445c1 100644
> > > --- a/fs/xfs/xfs_trans.c
> > > +++ b/fs/xfs/xfs_trans.c
> > > @@ -370,6 +370,10 @@ xfs_trans_mod_sb(
> > > tp->t_blk_res_used += (uint)-delta;
> > > if (tp->t_blk_res_used > tp->t_blk_res)
> > > xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
> > > + } else if (delta > 0 &&
> > > + xfs_sb_version_haslazysbcount(&mp->m_sb)) {
> > > + tp->t_blk_res += delta;
> > > + delta = 0;
> > > }
> > > tp->t_fdblocks_delta += delta;
> > > if (xfs_sb_version_haslazysbcount(&mp->m_sb))
> > > --
> > > 2.21.1
> > >
> >
>
