Re: [PATCH 2/9] xfs: report ag header corruption errors to the health tracking system

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

On Thu, Nov 21, 2019 at 08:26:03AM -0500, Brian Foster wrote:
> On Wed, Nov 20, 2019 at 08:43:23AM -0800, Darrick J. Wong wrote:
> > On Wed, Nov 20, 2019 at 09:20:47AM -0500, Brian Foster wrote:
> > > On Thu, Nov 14, 2019 at 10:19:26AM -0800, Darrick J. Wong wrote:
> > > > From: Darrick J. Wong <darrick.wong@oracle.com>
> > > > 
> > > > Whenever we encounter a corrupt AG header, we should report that to the
> > > > health monitoring system for later reporting.
> > > > 
> > > > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
> > > > ---
> > > >  fs/xfs/libxfs/xfs_alloc.c    |    6 ++++++
> > > >  fs/xfs/libxfs/xfs_health.h   |    6 ++++++
> > > >  fs/xfs/libxfs/xfs_ialloc.c   |    3 +++
> > > >  fs/xfs/libxfs/xfs_refcount.c |    5 ++++-
> > > >  fs/xfs/libxfs/xfs_rmap.c     |    5 ++++-
> > > >  fs/xfs/libxfs/xfs_sb.c       |    2 ++
> > > >  fs/xfs/xfs_health.c          |   17 +++++++++++++++++
> > > >  fs/xfs/xfs_inode.c           |    9 +++++++++
> > > >  8 files changed, 51 insertions(+), 2 deletions(-)
> > > > 
> > > > 
> > > > diff --git a/fs/xfs/libxfs/xfs_alloc.c b/fs/xfs/libxfs/xfs_alloc.c
> > > > index c284e10af491..e75e3ae6c912 100644
> > > > --- a/fs/xfs/libxfs/xfs_alloc.c
> > > > +++ b/fs/xfs/libxfs/xfs_alloc.c
> > > > @@ -26,6 +26,7 @@
> > > >  #include "xfs_log.h"
> > > >  #include "xfs_ag_resv.h"
> > > >  #include "xfs_bmap.h"
> > > > +#include "xfs_health.h"
> > > >  
> > > >  extern kmem_zone_t	*xfs_bmap_free_item_zone;
> > > >  
> > > > @@ -699,6 +700,8 @@ xfs_alloc_read_agfl(
> > > >  			mp, tp, mp->m_ddev_targp,
> > > >  			XFS_AG_DADDR(mp, agno, XFS_AGFL_DADDR(mp)),
> > > >  			XFS_FSS_TO_BB(mp, 1), 0, &bp, &xfs_agfl_buf_ops);
> > > > +	if (xfs_metadata_is_sick(error))
> > > > +		xfs_agno_mark_sick(mp, agno, XFS_SICK_AG_AGFL);
> > > 
> > > Any reason we couldn't do some of these in verifiers? I'm assuming we'd
> > > still need calls in various external corruption checks, but at least we
> > > wouldn't add a requirement to check all future buffer reads, etc.
> > 
> > I thought about that.  It would be wonderful if C had a syntactically
> > slick method to package a function + execution scope and pass that
> > through other functions to be called later. :)
> > 
> > For the per-AG stuff it wouldn't be hard to make the verifier functions
> > derive the AG number and call xfs_agno_mark_sick directly in the
> > verifier.  For per-inode metadata, we'd have to find a way to pass the
> > struct xfs_inode pointer to the verifier, which means that we'd have to
> > add that to struct xfs_buf.
> > 
> > xfs_buf is ~384 bytes so maybe adding another pointer for read context
> > wouldn't be terrible?  That would add a fair amount of ugly special
> > casing in the btree code to decide if we have an inode to pass through,
> > though it would solve the problem of the bmbt verifier not being able to
> > check the owner field in the btree block header.
> > 
> > OTOH that's 8 bytes of overhead that we can never get rid of even though
> > we only really need it the first time the buffer gets read in from disk.
> > 
> > Thoughts?
> > 
> 
> That doesn't seem too unreasonable, but I guess I'd have to think about
> it some more. Maybe it's worth defining a private pointer in the buffer
> that callers can use to pass specific context to verifiers for health
> processing. I suppose such a field could also be conditionally defined
> on scrub enabled kernels (at least initially), so the overhead would be
> opt-in.

Looking further into this, what if we could did something like the
following:

struct xfs_buf_verify {
	const struct xfs_buf_ops	*ops;
	struct xfs_inode		*ip;
	unsigned int			sick_flags;
	/* whatever else */
};

...then we change the _read_buf and _trans_read_buf functions to take as
the final argument a (struct xfs_buf_verify *).  In the xfs_buf_reverify
cases, we can pass this context straight through to the ->read_verify
function.

To handle the !DONE case where the buffer read completion can happen
asynchronously, we change the b_ops field definition to:

	union {
		struct xfs_buf_ops	*b_ops;
		struct xfs_buf_verify	*b_vctx;
	};

Next we define a new XBF_HAVE_VERIFY_CTX flag that means b_vctx is
active and not ops.  xfs_buf_read_map can set the flag and b_vctx for
any synchronous (!XBF_ASYNC) read because we know the caller will be
asleep waiting for b_iowait and therefore cannot kill the verifier
context structure.  Once we get to xfs_buf_ioend we can set b_ops, drop
the XBF_H_V_C flag, and call ->verify_read.

Now we actually /can/ pass the inode pointer into the verifier, along
with pretty much anything else we can think of.

Does that sound reasonable?  Or totally heinous? :)

> Anyways, I think for this series it might be reasonable to push things
> down into verifiers opportunistically where we can do so without any
> core mechanism changes. We can follow up with changes to do the rest if
> we can come up with something elegant.

Ok.  I think I will try to implement such a beast for 5.6 and then put
this series after it.

> > > >  	if (error)
> > > >  		return error;
> > > >  	xfs_buf_set_ref(bp, XFS_AGFL_REF);
> > > > @@ -722,6 +725,7 @@ xfs_alloc_update_counters(
> > > >  	if (unlikely(be32_to_cpu(agf->agf_freeblks) >
> > > >  		     be32_to_cpu(agf->agf_length))) {
> > > >  		xfs_buf_corruption_error(agbp);
> > > > +		xfs_ag_mark_sick(pag, XFS_SICK_AG_AGF);
> > > >  		return -EFSCORRUPTED;
> > > >  	}
> > > >  
> > > > @@ -2952,6 +2956,8 @@ xfs_read_agf(
> > > >  			mp, tp, mp->m_ddev_targp,
> > > >  			XFS_AG_DADDR(mp, agno, XFS_AGF_DADDR(mp)),
> > > >  			XFS_FSS_TO_BB(mp, 1), flags, bpp, &xfs_agf_buf_ops);
> > > > +	if (xfs_metadata_is_sick(error))
> > > > +		xfs_agno_mark_sick(mp, agno, XFS_SICK_AG_AGF);
> > > >  	if (error)
> > > >  		return error;
> > > >  	if (!*bpp)
> > > > diff --git a/fs/xfs/libxfs/xfs_health.h b/fs/xfs/libxfs/xfs_health.h
> > > > index 3657a9cb8490..ce8954a10c66 100644
> > > > --- a/fs/xfs/libxfs/xfs_health.h
> > > > +++ b/fs/xfs/libxfs/xfs_health.h
> > > > @@ -123,6 +123,8 @@ void xfs_rt_mark_healthy(struct xfs_mount *mp, unsigned int mask);
> > > >  void xfs_rt_measure_sickness(struct xfs_mount *mp, unsigned int *sick,
> > > >  		unsigned int *checked);
> > > >  
> > > > +void xfs_agno_mark_sick(struct xfs_mount *mp, xfs_agnumber_t agno,
> > > > +		unsigned int mask);
> > > >  void xfs_ag_mark_sick(struct xfs_perag *pag, unsigned int mask);
> > > >  void xfs_ag_mark_checked(struct xfs_perag *pag, unsigned int mask);
> > > >  void xfs_ag_mark_healthy(struct xfs_perag *pag, unsigned int mask);
> > > > @@ -203,4 +205,8 @@ void xfs_fsop_geom_health(struct xfs_mount *mp, struct xfs_fsop_geom *geo);
> > > >  void xfs_ag_geom_health(struct xfs_perag *pag, struct xfs_ag_geometry *ageo);
> > > >  void xfs_bulkstat_health(struct xfs_inode *ip, struct xfs_bulkstat *bs);
> > > >  
> > > > +#define xfs_metadata_is_sick(error) \
> > > > +	(unlikely((error) == -EFSCORRUPTED || (error) == -EIO || \
> > > > +		  (error) == -EFSBADCRC))
> > > 
> > > Why is -EIO considered sick? My understanding is that once something is
> > > marked sick, scrub is the only way to clear that state. -EIO can be
> > > transient, so afaict that means we could mark a persistent in-core state
> > > based on a transient/resolved issue.
> > 
> > I think it sounds reasonable that if the fs hits a metadata IO error
> > then the administrator should scrub that data structure to make sure
> > it's ok, and if so, clear the sick state.
> > 
> 
> I'm not totally convinced... I thought we had configurations where I/O
> errors can be reasonably expected and recovered from. For example,
> consider the thin provisioning + infinite metadata writeback error retry
> mechanism. IIRC, the whole purpose of that was to facilitate the use
> case where the thin pool runs out of space, but the admin wants some
> window of time to expand and keep the filesystem alive.

Aha, I just realized that it's not clear from the macro definition that
I was only intending it to be called from the read path.

Though I guess there's always the possibility that the PFY trips over
the PCIE cable in the datacenter and XFS hits an EIO, but the disk will
be fine a moment later when he shoves it back in.  The disk media is
fine, and by that point either we returned read error to userspace or
the transaction got cancelled and it's too late to do anything anyway.

I'll drop the EIO check for now and we'll see if I get around to
revisiting it.

> I don't necessarily think it's a bad thing to suggest a scrub any time
> errors have occurred, but for something like the above where an
> environment may have been thoroughly tested and verified through that
> particular error->expand sequence, it seems that flagging bits as sick
> might be unnecessarily ominous.

<shrug> Yeah, (sick && !checked) is a weird passive-aggressive state
like that.

> > Though I realized just now that if scrub isn't enabled then it's an
> > unfixable dead end so the EIO check should be gated on
> > CONFIG_XFS_ONLINE_SCRUB=y.
> > 
> 
> Yeah, that was my initial concern..
> 
> > > Along similar lines, what's the expected behavior in the event of any of
> > > these errors for a kernel that might not support
> > > CONFIG_XFS_ONLINE_[SCRUB|REPAIR]? Just set the states that are never
> > > used for anything? If so, that seems Ok I suppose.. but it's a little
> > > awkward if we'd see the tracepoints and such associated with the state
> > > changes.
> > 
> > Even if scrub is disabled, the kernel will still set the sick state, and
> > later the administrator can query the filesystem with xfs_spaceman to
> > observe that sick state.
> > 
> 
> Ok, so it's intended to be a valid health state independent of scrub.
> That seems reasonable in principle and can always be used to indicate
> offline repair is necessary too.

Yes.

> > In the future, I will also use the per-AG sick states to steer
> > allocations away from known problematic AGs to try to avoid
> > unexpected shutdown in the middle of a transaction.
> > 
> 
> Hmm.. I'm a little curious about how much we should steer away from
> traditional behavior on kernels that might not support scrub. I suppose
> I could see arguments for going either way, but this is getting a bit
> ahead of this patch anyways. ;)

Yeah.  I /do/ have prototype patches buried in my dev tree but they are
too ugly not to let all the magic smoke out.  What really happens is
that when we hit a corruption error, we mark the AG as offline.  Then
the sysadmin can run xfs_scrub to fix it (which would set th AG back
online) or I guess we could have a spaceman -x command to force it back
online.

I always build in /some/ kind of manual override somewhere... :)

--D

> Brian
> 
> > --D
> > 
> > > 
> > > Brian
> > > 
> > > > +
> > > >  #endif	/* __XFS_HEALTH_H__ */
> > > > diff --git a/fs/xfs/libxfs/xfs_ialloc.c b/fs/xfs/libxfs/xfs_ialloc.c
> > > > index 988cde7744e6..c401512a4350 100644
> > > > --- a/fs/xfs/libxfs/xfs_ialloc.c
> > > > +++ b/fs/xfs/libxfs/xfs_ialloc.c
> > > > @@ -27,6 +27,7 @@
> > > >  #include "xfs_trace.h"
> > > >  #include "xfs_log.h"
> > > >  #include "xfs_rmap.h"
> > > > +#include "xfs_health.h"
> > > >  
> > > >  /*
> > > >   * Lookup a record by ino in the btree given by cur.
> > > > @@ -2635,6 +2636,8 @@ xfs_read_agi(
> > > >  	error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
> > > >  			XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
> > > >  			XFS_FSS_TO_BB(mp, 1), 0, bpp, &xfs_agi_buf_ops);
> > > > +	if (xfs_metadata_is_sick(error))
> > > > +		xfs_agno_mark_sick(mp, agno, XFS_SICK_AG_AGI);
> > > >  	if (error)
> > > >  		return error;
> > > >  	if (tp)
> > > > diff --git a/fs/xfs/libxfs/xfs_refcount.c b/fs/xfs/libxfs/xfs_refcount.c
> > > > index d7d702ee4d1a..25c87834e42a 100644
> > > > --- a/fs/xfs/libxfs/xfs_refcount.c
> > > > +++ b/fs/xfs/libxfs/xfs_refcount.c
> > > > @@ -22,6 +22,7 @@
> > > >  #include "xfs_bit.h"
> > > >  #include "xfs_refcount.h"
> > > >  #include "xfs_rmap.h"
> > > > +#include "xfs_health.h"
> > > >  
> > > >  /* Allowable refcount adjustment amounts. */
> > > >  enum xfs_refc_adjust_op {
> > > > @@ -1177,8 +1178,10 @@ xfs_refcount_finish_one(
> > > >  				XFS_ALLOC_FLAG_FREEING, &agbp);
> > > >  		if (error)
> > > >  			return error;
> > > > -		if (XFS_IS_CORRUPT(tp->t_mountp, !agbp))
> > > > +		if (XFS_IS_CORRUPT(tp->t_mountp, !agbp)) {
> > > > +			xfs_agno_mark_sick(tp->t_mountp, agno, XFS_SICK_AG_AGF);
> > > >  			return -EFSCORRUPTED;
> > > > +		}
> > > >  
> > > >  		rcur = xfs_refcountbt_init_cursor(mp, tp, agbp, agno);
> > > >  		if (!rcur) {
> > > > diff --git a/fs/xfs/libxfs/xfs_rmap.c b/fs/xfs/libxfs/xfs_rmap.c
> > > > index ff9412f113c4..a54a3c129cce 100644
> > > > --- a/fs/xfs/libxfs/xfs_rmap.c
> > > > +++ b/fs/xfs/libxfs/xfs_rmap.c
> > > > @@ -21,6 +21,7 @@
> > > >  #include "xfs_errortag.h"
> > > >  #include "xfs_error.h"
> > > >  #include "xfs_inode.h"
> > > > +#include "xfs_health.h"
> > > >  
> > > >  /*
> > > >   * Lookup the first record less than or equal to [bno, len, owner, offset]
> > > > @@ -2400,8 +2401,10 @@ xfs_rmap_finish_one(
> > > >  		error = xfs_free_extent_fix_freelist(tp, agno, &agbp);
> > > >  		if (error)
> > > >  			return error;
> > > > -		if (XFS_IS_CORRUPT(tp->t_mountp, !agbp))
> > > > +		if (XFS_IS_CORRUPT(tp->t_mountp, !agbp)) {
> > > > +			xfs_agno_mark_sick(tp->t_mountp, agno, XFS_SICK_AG_AGF);
> > > >  			return -EFSCORRUPTED;
> > > > +		}
> > > >  
> > > >  		rcur = xfs_rmapbt_init_cursor(mp, tp, agbp, agno);
> > > >  		if (!rcur) {
> > > > diff --git a/fs/xfs/libxfs/xfs_sb.c b/fs/xfs/libxfs/xfs_sb.c
> > > > index 0ac69751fe85..4a923545465d 100644
> > > > --- a/fs/xfs/libxfs/xfs_sb.c
> > > > +++ b/fs/xfs/libxfs/xfs_sb.c
> > > > @@ -1169,6 +1169,8 @@ xfs_sb_read_secondary(
> > > >  	error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
> > > >  			XFS_AG_DADDR(mp, agno, XFS_SB_BLOCK(mp)),
> > > >  			XFS_FSS_TO_BB(mp, 1), 0, &bp, &xfs_sb_buf_ops);
> > > > +	if (xfs_metadata_is_sick(error))
> > > > +		xfs_agno_mark_sick(mp, agno, XFS_SICK_AG_SB);
> > > >  	if (error)
> > > >  		return error;
> > > >  	xfs_buf_set_ref(bp, XFS_SSB_REF);
> > > > diff --git a/fs/xfs/xfs_health.c b/fs/xfs/xfs_health.c
> > > > index 860dc70c99e7..36c32b108b39 100644
> > > > --- a/fs/xfs/xfs_health.c
> > > > +++ b/fs/xfs/xfs_health.c
> > > > @@ -200,6 +200,23 @@ xfs_rt_measure_sickness(
> > > >  	spin_unlock(&mp->m_sb_lock);
> > > >  }
> > > >  
> > > > +/* Mark unhealthy per-ag metadata given a raw AG number. */
> > > > +void
> > > > +xfs_agno_mark_sick(
> > > > +	struct xfs_mount	*mp,
> > > > +	xfs_agnumber_t		agno,
> > > > +	unsigned int		mask)
> > > > +{
> > > > +	struct xfs_perag	*pag = xfs_perag_get(mp, agno);
> > > > +
> > > > +	/* per-ag structure not set up yet? */
> > > > +	if (!pag)
> > > > +		return;
> > > > +
> > > > +	xfs_ag_mark_sick(pag, mask);
> > > > +	xfs_perag_put(pag);
> > > > +}
> > > > +
> > > >  /* Mark unhealthy per-ag metadata. */
> > > >  void
> > > >  xfs_ag_mark_sick(
> > > > diff --git a/fs/xfs/xfs_inode.c b/fs/xfs/xfs_inode.c
> > > > index 401da197f012..a2812cea748d 100644
> > > > --- a/fs/xfs/xfs_inode.c
> > > > +++ b/fs/xfs/xfs_inode.c
> > > > @@ -35,6 +35,7 @@
> > > >  #include "xfs_log.h"
> > > >  #include "xfs_bmap_btree.h"
> > > >  #include "xfs_reflink.h"
> > > > +#include "xfs_health.h"
> > > >  
> > > >  kmem_zone_t *xfs_inode_zone;
> > > >  
> > > > @@ -787,6 +788,8 @@ xfs_ialloc(
> > > >  	 */
> > > >  	if ((pip && ino == pip->i_ino) || !xfs_verify_dir_ino(mp, ino)) {
> > > >  		xfs_alert(mp, "Allocated a known in-use inode 0x%llx!", ino);
> > > > +		xfs_agno_mark_sick(mp, XFS_INO_TO_AGNO(mp, ino),
> > > > +				XFS_SICK_AG_INOBT);
> > > >  		return -EFSCORRUPTED;
> > > >  	}
> > > >  
> > > > @@ -2137,6 +2140,7 @@ xfs_iunlink_update_bucket(
> > > >  	 */
> > > >  	if (old_value == new_agino) {
> > > >  		xfs_buf_corruption_error(agibp);
> > > > +		xfs_agno_mark_sick(tp->t_mountp, agno, XFS_SICK_AG_AGI);
> > > >  		return -EFSCORRUPTED;
> > > >  	}
> > > >  
> > > > @@ -2203,6 +2207,7 @@ xfs_iunlink_update_inode(
> > > >  	if (!xfs_verify_agino_or_null(mp, agno, old_value)) {
> > > >  		xfs_inode_verifier_error(ip, -EFSCORRUPTED, __func__, dip,
> > > >  				sizeof(*dip), __this_address);
> > > > +		xfs_inode_mark_sick(ip, XFS_SICK_INO_CORE);
> > > >  		error = -EFSCORRUPTED;
> > > >  		goto out;
> > > >  	}
> > > > @@ -2217,6 +2222,7 @@ xfs_iunlink_update_inode(
> > > >  		if (next_agino != NULLAGINO) {
> > > >  			xfs_inode_verifier_error(ip, -EFSCORRUPTED, __func__,
> > > >  					dip, sizeof(*dip), __this_address);
> > > > +			xfs_inode_mark_sick(ip, XFS_SICK_INO_CORE);
> > > >  			error = -EFSCORRUPTED;
> > > >  		}
> > > >  		goto out;
> > > > @@ -2271,6 +2277,7 @@ xfs_iunlink(
> > > >  	if (next_agino == agino ||
> > > >  	    !xfs_verify_agino_or_null(mp, agno, next_agino)) {
> > > >  		xfs_buf_corruption_error(agibp);
> > > > +		xfs_agno_mark_sick(mp, agno, XFS_SICK_AG_AGI);
> > > >  		return -EFSCORRUPTED;
> > > >  	}
> > > >  
> > > > @@ -2408,6 +2415,7 @@ xfs_iunlink_map_prev(
> > > >  			XFS_CORRUPTION_ERROR(__func__,
> > > >  					XFS_ERRLEVEL_LOW, mp,
> > > >  					*dipp, sizeof(**dipp));
> > > > +			xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
> > > >  			error = -EFSCORRUPTED;
> > > >  			return error;
> > > >  		}
> > > > @@ -2454,6 +2462,7 @@ xfs_iunlink_remove(
> > > >  	if (!xfs_verify_agino(mp, agno, head_agino)) {
> > > >  		XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
> > > >  				agi, sizeof(*agi));
> > > > +		xfs_agno_mark_sick(mp, agno, XFS_SICK_AG_AGI);
> > > >  		return -EFSCORRUPTED;
> > > >  	}
> > > >  
> > > > 
> > > 
> > 
>