Commit a5637186 authored by Darrick J. Wong's avatar Darrick J. Wong
Browse files

xfs: dispatch metadata scrub subcommands 



Create structures needed to hold scrubbing context and dispatch incoming
commands to the individual scrubbers.

Signed-off-by: default avatarDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: default avatarDave Chinner <dchinner@redhat.com>
parent 36fd6e86

fs/xfs/scrub/scrub.c

+195 −1
Original line number Diff line number Diff line
@@ -44,11 +44,205 @@ 
#include "scrub/scrub.h"

#include "scrub/trace.h"



/*

 * Online Scrub and Repair

 *

 * Traditionally, XFS (the kernel driver) did not know how to check or

 * repair on-disk data structures.  That task was left to the xfs_check

 * and xfs_repair tools, both of which require taking the filesystem

 * offline for a thorough but time consuming examination.  Online

 * scrub & repair, on the other hand, enables us to check the metadata

 * for obvious errors while carefully stepping around the filesystem's

 * ongoing operations, locking rules, etc.

 *

 * Given that most XFS metadata consist of records stored in a btree,

 * most of the checking functions iterate the btree blocks themselves

 * looking for irregularities.  When a record block is encountered, each

 * record can be checked for obviously bad values.  Record values can

 * also be cross-referenced against other btrees to look for potential

 * misunderstandings between pieces of metadata.

 *

 * It is expected that the checkers responsible for per-AG metadata

 * structures will lock the AG headers (AGI, AGF, AGFL), iterate the

 * metadata structure, and perform any relevant cross-referencing before

 * unlocking the AG and returning the results to userspace.  These

 * scrubbers must not keep an AG locked for too long to avoid tying up

 * the block and inode allocators.

 *

 * Block maps and b-trees rooted in an inode present a special challenge

 * because they can involve extents from any AG.  The general scrubber

 * structure of lock -> check -> xref -> unlock still holds, but AG

 * locking order rules /must/ be obeyed to avoid deadlocks.  The

 * ordering rule, of course, is that we must lock in increasing AG

 * order.  Helper functions are provided to track which AG headers we've

 * already locked.  If we detect an imminent locking order violation, we

 * can signal a potential deadlock, in which case the scrubber can jump

 * out to the top level, lock all the AGs in order, and retry the scrub.

 *

 * For file data (directories, extended attributes, symlinks) scrub, we

 * can simply lock the inode and walk the data.  For btree data

 * (directories and attributes) we follow the same btree-scrubbing

 * strategy outlined previously to check the records.

 *

 * We use a bit of trickery with transactions to avoid buffer deadlocks

 * if there is a cycle in the metadata.  The basic problem is that

 * travelling down a btree involves locking the current buffer at each

 * tree level.  If a pointer should somehow point back to a buffer that

 * we've already examined, we will deadlock due to the second buffer

 * locking attempt.  Note however that grabbing a buffer in transaction

 * context links the locked buffer to the transaction.  If we try to

 * re-grab the buffer in the context of the same transaction, we avoid

 * the second lock attempt and continue.  Between the verifier and the

 * scrubber, something will notice that something is amiss and report

 * the corruption.  Therefore, each scrubber will allocate an empty

 * transaction, attach buffers to it, and cancel the transaction at the

 * end of the scrub run.  Cancelling a non-dirty transaction simply

 * unlocks the buffers.

 *

 * There are four pieces of data that scrub can communicate to

 * userspace.  The first is the error code (errno), which can be used to

 * communicate operational errors in performing the scrub.  There are

 * also three flags that can be set in the scrub context.  If the data

 * structure itself is corrupt, the CORRUPT flag will be set.  If

 * the metadata is correct but otherwise suboptimal, the PREEN flag

 * will be set.

 */



/* Scrub setup and teardown */



/* Free all the resources and finish the transactions. */

STATIC int

xfs_scrub_teardown(

	struct xfs_scrub_context	*sc,

	int				error)

{

	if (sc->tp) {

		xfs_trans_cancel(sc->tp);

		sc->tp = NULL;

	}

	return error;

}



/* Scrubbing dispatch. */



static const struct xfs_scrub_meta_ops meta_scrub_ops[] = {

};



/* This isn't a stable feature, warn once per day. */

static inline void

xfs_scrub_experimental_warning(

	struct xfs_mount	*mp)

{

	static struct ratelimit_state scrub_warning = RATELIMIT_STATE_INIT(

			"xfs_scrub_warning", 86400 * HZ, 1);

	ratelimit_set_flags(&scrub_warning, RATELIMIT_MSG_ON_RELEASE);



	if (__ratelimit(&scrub_warning))

		xfs_alert(mp,

"EXPERIMENTAL online scrub feature in use. Use at your own risk!");

}



/* Dispatch metadata scrubbing. */

int

xfs_scrub_metadata(

	struct xfs_inode		*ip,

	struct xfs_scrub_metadata	*sm)

{

	return -EOPNOTSUPP;

	struct xfs_scrub_context	sc;

	struct xfs_mount		*mp = ip->i_mount;

	const struct xfs_scrub_meta_ops	*ops;

	bool				try_harder = false;

	int				error = 0;



	trace_xfs_scrub_start(ip, sm, error);



	/* Forbidden if we are shut down or mounted norecovery. */

	error = -ESHUTDOWN;

	if (XFS_FORCED_SHUTDOWN(mp))

		goto out;

	error = -ENOTRECOVERABLE;

	if (mp->m_flags & XFS_MOUNT_NORECOVERY)

		goto out;



	/* Check our inputs. */

	error = -EINVAL;

	sm->sm_flags &= ~XFS_SCRUB_FLAGS_OUT;

	if (sm->sm_flags & ~XFS_SCRUB_FLAGS_IN)

		goto out;

	if (memchr_inv(sm->sm_reserved, 0, sizeof(sm->sm_reserved)))

		goto out;



	/* Do we know about this type of metadata? */

	error = -ENOENT;

	if (sm->sm_type >= XFS_SCRUB_TYPE_NR)

		goto out;

	ops = &meta_scrub_ops[sm->sm_type];

	if (ops->scrub == NULL)

		goto out;



	/*

	 * We won't scrub any filesystem that doesn't have the ability

	 * to record unwritten extents.  The option was made default in

	 * 2003, removed from mkfs in 2007, and cannot be disabled in

	 * v5, so if we find a filesystem without this flag it's either

	 * really old or totally unsupported.  Avoid it either way.

	 * We also don't support v1-v3 filesystems, which aren't

	 * mountable.

	 */

	error = -EOPNOTSUPP;

	if (!xfs_sb_version_hasextflgbit(&mp->m_sb))

		goto out;



	/* Does this fs even support this type of metadata? */

	error = -ENOENT;

	if (ops->has && !ops->has(&mp->m_sb))

		goto out;



	/* We don't know how to repair anything yet. */

	error = -EOPNOTSUPP;

	if (sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR)

		goto out;



	xfs_scrub_experimental_warning(mp);



retry_op:

	/* Set up for the operation. */

	memset(&sc, 0, sizeof(sc));

	sc.mp = ip->i_mount;

	sc.sm = sm;

	sc.ops = ops;

	sc.try_harder = try_harder;

	error = sc.ops->setup(&sc, ip);

	if (error)

		goto out_teardown;



	/* Scrub for errors. */

	error = sc.ops->scrub(&sc);

	if (!try_harder && error == -EDEADLOCK) {

		/*

		 * Scrubbers return -EDEADLOCK to mean 'try harder'.

		 * Tear down everything we hold, then set up again with

		 * preparation for worst-case scenarios.

		 */

		error = xfs_scrub_teardown(&sc, 0);

		if (error)

			goto out;

		try_harder = true;

		goto retry_op;

	} else if (error)

		goto out_teardown;



	if (sc.sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |

			       XFS_SCRUB_OFLAG_XCORRUPT))

		xfs_alert_ratelimited(mp, "Corruption detected during scrub.");



out_teardown:

	error = xfs_scrub_teardown(&sc, error);

out:

	trace_xfs_scrub_done(ip, sm, error);

	if (error == -EFSCORRUPTED || error == -EFSBADCRC) {

		sm->sm_flags |= XFS_SCRUB_OFLAG_CORRUPT;

		error = 0;

	}

	return error;

}

fs/xfs/scrub/scrub.h

+24 −0
Original line number Diff line number Diff line
@@ -20,6 +20,30 @@ 
#ifndef __XFS_SCRUB_SCRUB_H__

#define __XFS_SCRUB_SCRUB_H__



struct xfs_scrub_context;



struct xfs_scrub_meta_ops {

	/* Acquire whatever resources are needed for the operation. */

	int		(*setup)(struct xfs_scrub_context *,

				 struct xfs_inode *);



	/* Examine metadata for errors. */

	int		(*scrub)(struct xfs_scrub_context *);



	/* Decide if we even have this piece of metadata. */

	bool		(*has)(struct xfs_sb *);

};



struct xfs_scrub_context {

	/* General scrub state. */

	struct xfs_mount		*mp;

	struct xfs_scrub_metadata	*sm;

	const struct xfs_scrub_meta_ops	*ops;

	struct xfs_trans		*tp;

	struct xfs_inode		*ip;

	bool				try_harder;

};



/* Metadata scrubbers */



#endif	/* __XFS_SCRUB_SCRUB_H__ */

fs/xfs/scrub/trace.h

+43 −0
Original line number Diff line number Diff line
@@ -25,6 +25,49 @@ 


#include <linux/tracepoint.h>



DECLARE_EVENT_CLASS(xfs_scrub_class,

	TP_PROTO(struct xfs_inode *ip, struct xfs_scrub_metadata *sm,

		 int error),

	TP_ARGS(ip, sm, error),

	TP_STRUCT__entry(

		__field(dev_t, dev)

		__field(xfs_ino_t, ino)

		__field(unsigned int, type)

		__field(xfs_agnumber_t, agno)

		__field(xfs_ino_t, inum)

		__field(unsigned int, gen)

		__field(unsigned int, flags)

		__field(int, error)

	),

	TP_fast_assign(

		__entry->dev = ip->i_mount->m_super->s_dev;

		__entry->ino = ip->i_ino;

		__entry->type = sm->sm_type;

		__entry->agno = sm->sm_agno;

		__entry->inum = sm->sm_ino;

		__entry->gen = sm->sm_gen;

		__entry->flags = sm->sm_flags;

		__entry->error = error;

	),

	TP_printk("dev %d:%d ino %llu type %u agno %u inum %llu gen %u flags 0x%x error %d",

		  MAJOR(__entry->dev), MINOR(__entry->dev),

		  __entry->ino,

		  __entry->type,

		  __entry->agno,

		  __entry->inum,

		  __entry->gen,

		  __entry->flags,

		  __entry->error)

)

#define DEFINE_SCRUB_EVENT(name) \

DEFINE_EVENT(xfs_scrub_class, name, \

	TP_PROTO(struct xfs_inode *ip, struct xfs_scrub_metadata *sm, \

		 int error), \

	TP_ARGS(ip, sm, error))



DEFINE_SCRUB_EVENT(xfs_scrub_start);

DEFINE_SCRUB_EVENT(xfs_scrub_done);



#endif /* _TRACE_XFS_SCRUB_TRACE_H */



#undef TRACE_INCLUDE_PATH