| Message ID | 151935121574.21654.7158330354778549108.stgit@magnolia (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
On Thu, Feb 22, 2018 at 06:00:15PM -0800, Darrick J. Wong wrote: > From: Darrick J. Wong <darrick.wong@oracle.com> Yay, no commit message! "Fix the AGFL wrapping issues introduced with the v5 disk format and semi-corrected later in 4.5. In short, if we find an incorrectly wrapped agfl at mount time, fix the list wrapping so that we can start opreations. At unmount time, adjust all the agfls so that they don't touch the controversial last element. Skip all this fuss if the filesystem has features that didn't exist prior to the agfl fixing in 4.5." --D > > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> > --- > fs/xfs/Makefile | 1 > fs/xfs/xfs_fixups.c | 310 +++++++++++++++++++++++++++++++++++++++++++++++++++ > fs/xfs/xfs_fixups.h | 26 ++++ > fs/xfs/xfs_mount.c | 21 +++ > fs/xfs/xfs_super.c | 10 ++ > 5 files changed, 367 insertions(+), 1 deletion(-) > create mode 100644 fs/xfs/xfs_fixups.c > create mode 100644 fs/xfs/xfs_fixups.h > > > diff --git a/fs/xfs/Makefile b/fs/xfs/Makefile > index b03c77e..f88368a 100644 > --- a/fs/xfs/Makefile > +++ b/fs/xfs/Makefile > @@ -86,6 +86,7 @@ xfs-y += xfs_aops.o \ > xfs_extent_busy.o \ > xfs_file.o \ > xfs_filestream.o \ > + xfs_fixups.o \ > xfs_fsmap.o \ > xfs_fsops.o \ > xfs_globals.o \ > diff --git a/fs/xfs/xfs_fixups.c b/fs/xfs/xfs_fixups.c > new file mode 100644 > index 0000000..0cad7bb > --- /dev/null > +++ b/fs/xfs/xfs_fixups.c > @@ -0,0 +1,310 @@ > +/* > + * Copyright (C) 2018 Oracle. All Rights Reserved. > + * > + * Author: Darrick J. Wong <darrick.wong@oracle.com> > + * > + * This program is free software; you can redistribute it and/or > + * modify it under the terms of the GNU General Public License > + * as published by the Free Software Foundation; either version 2 > + * of the License, or (at your option) any later version. > + * > + * This program is distributed in the hope that it would be useful, > + * but WITHOUT ANY WARRANTY; without even the implied warranty of > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the > + * GNU General Public License for more details. > + * > + * You should have received a copy of the GNU General Public License > + * along with this program; if not, write the Free Software Foundation, > + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. > + */ > +#include "xfs.h" > +#include "xfs_fs.h" > +#include "xfs_shared.h" > +#include "xfs_format.h" > +#include "xfs_log_format.h" > +#include "xfs_trans_resv.h" > +#include "xfs_sb.h" > +#include "xfs_mount.h" > +#include "xfs_alloc.h" > +#include "xfs_trans.h" > +#include "xfs_fixups.h" > + > +/* > + * v5 AGFL padding defects > + * > + * When the v5 format was first introduced, there was a defect in the struct > + * xfs_agfl definition that resulted in XFS_AGFL_SIZE returning different > + * values depending on the compiler padding. On a fs with 512-byte sectors, > + * this meant that XFS_AGFL_SIZE was 119 on i386, but 118 on x64. Commit > + * 96f859d52bcb1 ("libxfs: pack the agfl header structure so XFS_AGFL_SIZE is > + * correct") changed the definition to disable padding the end of the > + * structure, and was accepted into Linux 4.5. Since then, the AGFL has > + * always used the larger size (e.g. 119 entries on a 512b sector fs). > + * > + * Unfortunately, pre-4.5 kernels can produce filesystems with AGFLs that wrap > + * at the smaller size, and those kernels are not prepared to handle the > + * longer size. This typically manifests itself as an AGF verifier corruption > + * error followed by a filesystem shutdown. While we encourage admins to stay > + * current with software, we would like to avoid this intermittent breakage. > + * > + * Any v5 filesystem which has a feature bit set for a feature that was > + * introduced after Linux 4.5 will not have this problem, as such kernels > + * cannot be mounted on older kernels. v4 filesystems are also unaffected. > + * > + * Therefore, we add two fixup functions -- the first runs at mount time to > + * detect a short-wrapped AGFL and fix it; the second runs at unmount, freeze, > + * or remount-ro time to move a wrapped AGFL to the beginning of the list. > + * This reduces the likelihood of a screwup to the scenario where you have (a) > + * a filesystem with no post-4.5 features (reflink, rmap), (b) the AGFL wraps, > + * (c) the filesystem goes down leaving a dirty log, and (d) the dirty > + * filesystem is mounted on an old kernel. > + */ > + > +/* > + * Decide if we need to have the agfl wrapping fixes applied. This only > + * affects v5 filesystems that do not have any features enabled that did not > + * exist when the agfl padding fix went in. > + * > + * Features already present when the fix went in were finobt, ftype, spinodes. > + * If we see something new (e.g. reflink) then don't bother. > + */ > +#define XFS_SB_FEAT_RO_COMPAT_AGFL_WRAP_ALREADY_FIXED \ > + (~(XFS_SB_FEAT_RO_COMPAT_FINOBT)) > +#define XFS_SB_FEAT_INCOMPAT_AGFL_WRAP_ALREADY_FIXED \ > + (~(XFS_SB_FEAT_INCOMPAT_FTYPE | \ > + XFS_SB_FEAT_INCOMPAT_SPINODES)) > +#define XFS_SB_FEAT_INCOMPAT_LOG_AGFL_WRAP_ALREADY_FIXED \ > + (~0) > +static inline bool xfs_sb_version_needs_agfl_wrap_fixes(struct xfs_sb *sbp) > +{ > + return xfs_sb_version_hascrc(sbp) && > + !xfs_sb_has_incompat_feature(sbp, > + XFS_SB_FEAT_INCOMPAT_AGFL_WRAP_ALREADY_FIXED) && > + !xfs_sb_has_ro_compat_feature(sbp, > + XFS_SB_FEAT_RO_COMPAT_AGFL_WRAP_ALREADY_FIXED) && > + !xfs_sb_has_incompat_log_feature(sbp, > + XFS_SB_FEAT_INCOMPAT_LOG_AGFL_WRAP_ALREADY_FIXED); > +} > + > +/* > + * Fix an AGFL wrapping that falls short of the end of the block by filling the > + * gap at the end of the block. > + */ > +STATIC int > +xfs_fixup_freelist_wrap_mount( > + struct xfs_trans *tp, > + struct xfs_buf *agfbp, > + struct xfs_perag *pag) > +{ > + struct xfs_mount *mp = tp->t_mountp; > + struct xfs_agf *agf; > + struct xfs_buf *agflbp; > + __be32 *agfl_bno; > + xfs_agnumber_t agno; > + uint32_t agfl_size; > + uint32_t flfirst; > + uint32_t fllast; > + int32_t active; > + int offset; > + int len; > + int error; > + > + if (pag->pagf_flcount == 0) > + return 0; > + > + agfl_size = xfs_agfl_size(mp); > + agf = XFS_BUF_TO_AGF(agfbp); > + agno = be32_to_cpu(agf->agf_seqno); > + flfirst = be32_to_cpu(agf->agf_flfirst); > + fllast = be32_to_cpu(agf->agf_fllast); > + > + /* Make sure we're either spot on or off by 1. */ > + active = fllast - flfirst + 1; > + if (active <= 0) > + active += agfl_size; > + if (active == pag->pagf_flcount) > + return 0; > + else if (active != pag->pagf_flcount + 1) > + return -EFSCORRUPTED; > + > + /* Would this have even passed muster on an old system? */ > + if (flfirst >= agfl_size - 1 || fllast >= agfl_size - 1 || > + pag->pagf_flcount > agfl_size - 1) > + return -EFSCORRUPTED; > + > + /* > + * Convert a 40-byte-padded agfl into a 36-byte-padded AGFL. > + * Therefore, we need to move the AGFL blocks > + * bno[flfirst..agfl_size - 2] to bno[flfirst + 1...agfl_size - 1]. > + * > + * Reusing the example above, if we had flfirst == 116, we need > + * to move bno[116] and bno[117] into bno[117] and bno[118], > + * respectively, and then increment flfirst. > + */ > + error = xfs_alloc_read_agfl(mp, tp, agno, &agflbp); > + if (error) > + return error; > + agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); > + > + len = (agfl_size - flfirst - 1) * sizeof(xfs_agblock_t); > + memmove(&agfl_bno[flfirst + 1], &agfl_bno[flfirst], len); > + offset = (char *)&agfl_bno[flfirst + 1] - (char *)agflbp->b_addr; > + be32_add_cpu(&agf->agf_flfirst, 1); > + > + xfs_trans_buf_set_type(tp, agflbp, XFS_BLFT_AGFL_BUF); > + xfs_trans_log_buf(tp, agflbp, offset, offset + len - 1); > + xfs_trans_brelse(tp, agflbp); > + agflbp = NULL; > + xfs_alloc_log_agf(tp, agfbp, XFS_AGF_FLFIRST); > + > + return 0; > +} > + > +/* > + * Fix an AGFL that touches the end of the block by moving the first or last > + * part of the list elsewhere in the AGFL so that old kernels don't trip over > + * wrapping issues. > + */ > +STATIC int > +xfs_fixup_freelist_wrap_unmount( > + struct xfs_trans *tp, > + struct xfs_buf *agfbp, > + struct xfs_perag *pag) > +{ > + struct xfs_mount *mp = tp->t_mountp; > + struct xfs_agf *agf; > + struct xfs_buf *agflbp; > + __be32 *agfl_bno; > + xfs_agnumber_t agno; > + uint32_t agfl_size; > + uint32_t flfirst; > + uint32_t fllast; > + int offset; > + int len; > + int error; > + > + agfl_size = xfs_agfl_size(mp); > + agf = XFS_BUF_TO_AGF(agfbp); > + agno = be32_to_cpu(agf->agf_seqno); > + flfirst = be32_to_cpu(agf->agf_flfirst); > + fllast = be32_to_cpu(agf->agf_fllast); > + > + /* Empty AGFL? Make sure we aren't pointing at the end. */ > + if (pag->pagf_flcount == 0) { > + if (flfirst >= agfl_size || fllast >= agfl_size) { > + agf->agf_flfirst = cpu_to_be32(1); > + agf->agf_fllast = 0; > + xfs_alloc_log_agf(tp, agfbp, > + XFS_AGF_FLFIRST | XFS_AGF_FLLAST); > + } > + return 0; > + } > + > + /* If we don't hit the end, we're done. */ > + if (flfirst < fllast && fllast != agfl_size - 1) > + return 0; > + > + /* > + * Move a start of a wrapped list towards the start of the agfl block. > + * Therefore, we need to move the AGFL blocks > + * bno[flfirst..agfl_size - 1] to bno[fllast + 1...agfl_size - flfirst]. > + * Then we reset flfirst and fllast appropriately. > + * > + * Reusing the example above, if we had flfirst == 117 and fllast == 4, > + * we need to move bno[117] and bno[118] into bno[5] and bno[6], > + * respectively, and then reset flfirst and fllast. > + * > + * If it's just the last block that touches the end, only move that. > + */ > + error = xfs_alloc_read_agfl(mp, tp, agno, &agflbp); > + if (error) > + return error; > + agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); > + > + if (fllast == agfl_size - 1) { > + /* Back the AGFL off from the end of the block. */ > + len = sizeof(xfs_agblock_t); > + agfl_bno[flfirst - 1] = agfl_bno[agfl_size - 1]; > + offset = (char *)&agfl_bno[flfirst - 1] - (char *)agflbp->b_addr; > + be32_add_cpu(&agf->agf_fllast, -1); > + be32_add_cpu(&agf->agf_flfirst, -1); > + } else { > + /* Move the first part of the AGFL towards the front. */ > + len = (agfl_size - flfirst) * sizeof(xfs_agblock_t); > + memcpy(&agfl_bno[fllast + 1], &agfl_bno[flfirst], len); > + offset = (char *)&agfl_bno[fllast + 1] - (char *)agflbp->b_addr; > + agf->agf_flfirst = 0; > + agf->agf_fllast = cpu_to_be32(pag->pagf_flcount - 1); > + } > + > + xfs_trans_buf_set_type(tp, agflbp, XFS_BLFT_AGFL_BUF); > + xfs_trans_log_buf(tp, agflbp, offset, offset + len - 1); > + xfs_trans_brelse(tp, agflbp); > + agflbp = NULL; > + xfs_alloc_log_agf(tp, agfbp, XFS_AGF_FLFIRST | XFS_AGF_FLLAST); > + > + return 0; > +} > + > +typedef int (*xfs_agf_apply_fn_t)(struct xfs_trans *tp, struct xfs_buf *agfbp, > + struct xfs_perag *pag); > + > +/* Apply something to every AGF. */ > +STATIC int > +xfs_fixup_agf_apply( > + struct xfs_mount *mp, > + xfs_agf_apply_fn_t fn) > +{ > + struct xfs_trans *tp; > + struct xfs_perag *pag; > + struct xfs_buf *agfbp; > + xfs_agnumber_t agno; > + int error; > + > + error = xfs_trans_alloc(mp, &M_RES(mp)->tr_growdata, 0, 0, 0, &tp); > + if (error) > + return error; > + > + for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) { > + error = xfs_alloc_read_agf(mp, tp, agno, 0, &agfbp); > + if (error) > + goto cancel; > + if (!agfbp) { > + error = -ENOMEM; > + goto cancel; > + } > + pag = xfs_perag_get(mp, agno); > + error = fn(tp, agfbp, pag); > + xfs_perag_put(pag); > + xfs_trans_brelse(tp, agfbp); > + if (error) > + goto cancel; > + } > + > + return xfs_trans_commit(tp); > +cancel: > + xfs_trans_cancel(tp); > + return error; > +} > + > +/* Fix AGFL wrapping so we can use the filesystem. */ > +int > +xfs_fixup_agfl_wrap_mount( > + struct xfs_mount *mp) > +{ > + if (!xfs_sb_version_needs_agfl_wrap_fixes(&mp->m_sb)) > + return 0; > + > + return xfs_fixup_agf_apply(mp, xfs_fixup_freelist_wrap_mount); > +} > + > +/* Fix AGFL wrapping so old kernels can use this filesystem. */ > +int > +xfs_fixup_agfl_wrap_unmount( > + struct xfs_mount *mp) > +{ > + if (!xfs_sb_version_needs_agfl_wrap_fixes(&mp->m_sb)) > + return 0; > + > + return xfs_fixup_agf_apply(mp, xfs_fixup_freelist_wrap_unmount); > +} > diff --git a/fs/xfs/xfs_fixups.h b/fs/xfs/xfs_fixups.h > new file mode 100644 > index 0000000..fb52a96 > --- /dev/null > +++ b/fs/xfs/xfs_fixups.h > @@ -0,0 +1,26 @@ > +/* > + * Copyright (C) 2018 Oracle. All Rights Reserved. > + * > + * Author: Darrick J. Wong <darrick.wong@oracle.com> > + * > + * This program is free software; you can redistribute it and/or > + * modify it under the terms of the GNU General Public License > + * as published by the Free Software Foundation; either version 2 > + * of the License, or (at your option) any later version. > + * > + * This program is distributed in the hope that it would be useful, > + * but WITHOUT ANY WARRANTY; without even the implied warranty of > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the > + * GNU General Public License for more details. > + * > + * You should have received a copy of the GNU General Public License > + * along with this program; if not, write the Free Software Foundation, > + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. > + */ > +#ifndef __XFS_FIXUPS_H__ > +#define __XFS_FIXUPS_H__ > + > +int xfs_fixup_agfl_wrap_mount(struct xfs_mount *mp); > +int xfs_fixup_agfl_wrap_unmount(struct xfs_mount *mp); > + > +#endif /* __XFS_FIXUPS_H__ */ > diff --git a/fs/xfs/xfs_mount.c b/fs/xfs/xfs_mount.c > index 98fd41c..eb284aa 100644 > --- a/fs/xfs/xfs_mount.c > +++ b/fs/xfs/xfs_mount.c > @@ -46,7 +46,7 @@ > #include "xfs_refcount_btree.h" > #include "xfs_reflink.h" > #include "xfs_extent_busy.h" > - > +#include "xfs_fixups.h" > > static DEFINE_MUTEX(xfs_uuid_table_mutex); > static int xfs_uuid_table_size; > @@ -875,6 +875,16 @@ xfs_mountfs( > } > > /* > + * Make sure our AGFL counters do not wrap the end of the block > + * in a troublesome manner. > + */ > + error = xfs_fixup_agfl_wrap_mount(mp); > + if (error) { > + xfs_warn(mp, "Failed to fix agfl wrapping. Run xfs_repair."); > + goto out_log_dealloc; > + } > + > + /* > * Get and sanity-check the root inode. > * Save the pointer to it in the mount structure. > */ > @@ -1128,6 +1138,15 @@ xfs_unmountfs( > xfs_qm_unmount(mp); > > /* > + * Make sure our AGFL counters do not wrap the end of the block > + * in a troublesome manner for old kernels. > + */ > + error = xfs_fixup_agfl_wrap_unmount(mp); > + if (error) > + xfs_warn(mp, "Unable to fix agfl wrapping. " > + "This may cause problems on next mount."); > + > + /* > * Unreserve any blocks we have so that when we unmount we don't account > * the reserved free space as used. This is really only necessary for > * lazy superblock counting because it trusts the incore superblock > diff --git a/fs/xfs/xfs_super.c b/fs/xfs/xfs_super.c > index 624a802..d9aa39a 100644 > --- a/fs/xfs/xfs_super.c > +++ b/fs/xfs/xfs_super.c > @@ -50,6 +50,7 @@ > #include "xfs_refcount_item.h" > #include "xfs_bmap_item.h" > #include "xfs_reflink.h" > +#include "xfs_fixups.h" > > #include <linux/namei.h> > #include <linux/dax.h> > @@ -1206,6 +1207,15 @@ xfs_quiesce_attr( > xfs_reclaim_inodes(mp, 0); > xfs_reclaim_inodes(mp, SYNC_WAIT); > > + /* > + * Make sure our AGFL counters do not wrap the end of the block > + * in a troublesome manner for old kernels. > + */ > + error = xfs_fixup_agfl_wrap_unmount(mp); > + if (error) > + xfs_warn(mp, "Unable to fix agfl wrapping. " > + "This may cause problems on next mount."); > + > /* Push the superblock and write an unmount record */ > error = xfs_log_sbcount(mp); > if (error) > > -- > To unsubscribe from this list: send the line "unsubscribe linux-xfs" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Feb 22, 2018 at 06:00:15PM -0800, Darrick J. Wong wrote: > From: Darrick J. Wong <darrick.wong@oracle.com> > > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> > --- > fs/xfs/Makefile | 1 > fs/xfs/xfs_fixups.c | 310 +++++++++++++++++++++++++++++++++++++++++++++++++++ > fs/xfs/xfs_fixups.h | 26 ++++ > fs/xfs/xfs_mount.c | 21 +++ > fs/xfs/xfs_super.c | 10 ++ > 5 files changed, 367 insertions(+), 1 deletion(-) > create mode 100644 fs/xfs/xfs_fixups.c > create mode 100644 fs/xfs/xfs_fixups.h > > > diff --git a/fs/xfs/Makefile b/fs/xfs/Makefile > index b03c77e..f88368a 100644 > --- a/fs/xfs/Makefile > +++ b/fs/xfs/Makefile > @@ -86,6 +86,7 @@ xfs-y += xfs_aops.o \ > xfs_extent_busy.o \ > xfs_file.o \ > xfs_filestream.o \ > + xfs_fixups.o \ > xfs_fsmap.o \ > xfs_fsops.o \ > xfs_globals.o \ > diff --git a/fs/xfs/xfs_fixups.c b/fs/xfs/xfs_fixups.c > new file mode 100644 > index 0000000..0cad7bb > --- /dev/null > +++ b/fs/xfs/xfs_fixups.c > @@ -0,0 +1,310 @@ > +/* > + * Copyright (C) 2018 Oracle. All Rights Reserved. > + * > + * Author: Darrick J. Wong <darrick.wong@oracle.com> > + * > + * This program is free software; you can redistribute it and/or > + * modify it under the terms of the GNU General Public License > + * as published by the Free Software Foundation; either version 2 > + * of the License, or (at your option) any later version. > + * > + * This program is distributed in the hope that it would be useful, > + * but WITHOUT ANY WARRANTY; without even the implied warranty of > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the > + * GNU General Public License for more details. > + * > + * You should have received a copy of the GNU General Public License > + * along with this program; if not, write the Free Software Foundation, > + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. > + */ > +#include "xfs.h" > +#include "xfs_fs.h" > +#include "xfs_shared.h" > +#include "xfs_format.h" > +#include "xfs_log_format.h" > +#include "xfs_trans_resv.h" > +#include "xfs_sb.h" > +#include "xfs_mount.h" > +#include "xfs_alloc.h" > +#include "xfs_trans.h" > +#include "xfs_fixups.h" > + > +/* > + * v5 AGFL padding defects > + * > + * When the v5 format was first introduced, there was a defect in the struct > + * xfs_agfl definition that resulted in XFS_AGFL_SIZE returning different > + * values depending on the compiler padding. On a fs with 512-byte sectors, > + * this meant that XFS_AGFL_SIZE was 119 on i386, but 118 on x64. Commit > + * 96f859d52bcb1 ("libxfs: pack the agfl header structure so XFS_AGFL_SIZE is > + * correct") changed the definition to disable padding the end of the > + * structure, and was accepted into Linux 4.5. Since then, the AGFL has > + * always used the larger size (e.g. 119 entries on a 512b sector fs). > + * > + * Unfortunately, pre-4.5 kernels can produce filesystems with AGFLs that wrap > + * at the smaller size, and those kernels are not prepared to handle the > + * longer size. This typically manifests itself as an AGF verifier corruption > + * error followed by a filesystem shutdown. While we encourage admins to stay > + * current with software, we would like to avoid this intermittent breakage. > + * > + * Any v5 filesystem which has a feature bit set for a feature that was > + * introduced after Linux 4.5 will not have this problem, as such kernels > + * cannot be mounted on older kernels. v4 filesystems are also unaffected. > + * > + * Therefore, we add two fixup functions -- the first runs at mount time to > + * detect a short-wrapped AGFL and fix it; the second runs at unmount, freeze, > + * or remount-ro time to move a wrapped AGFL to the beginning of the list. > + * This reduces the likelihood of a screwup to the scenario where you have (a) > + * a filesystem with no post-4.5 features (reflink, rmap), (b) the AGFL wraps, > + * (c) the filesystem goes down leaving a dirty log, and (d) the dirty > + * filesystem is mounted on an old kernel. > + */ > + > +/* > + * Decide if we need to have the agfl wrapping fixes applied. This only > + * affects v5 filesystems that do not have any features enabled that did not > + * exist when the agfl padding fix went in. > + * > + * Features already present when the fix went in were finobt, ftype, spinodes. > + * If we see something new (e.g. reflink) then don't bother. > + */ > +#define XFS_SB_FEAT_RO_COMPAT_AGFL_WRAP_ALREADY_FIXED \ > + (~(XFS_SB_FEAT_RO_COMPAT_FINOBT)) > +#define XFS_SB_FEAT_INCOMPAT_AGFL_WRAP_ALREADY_FIXED \ > + (~(XFS_SB_FEAT_INCOMPAT_FTYPE | \ > + XFS_SB_FEAT_INCOMPAT_SPINODES)) > +#define XFS_SB_FEAT_INCOMPAT_LOG_AGFL_WRAP_ALREADY_FIXED \ > + (~0) > +static inline bool xfs_sb_version_needs_agfl_wrap_fixes(struct xfs_sb *sbp) > +{ > + return xfs_sb_version_hascrc(sbp) && > + !xfs_sb_has_incompat_feature(sbp, > + XFS_SB_FEAT_INCOMPAT_AGFL_WRAP_ALREADY_FIXED) && > + !xfs_sb_has_ro_compat_feature(sbp, > + XFS_SB_FEAT_RO_COMPAT_AGFL_WRAP_ALREADY_FIXED) && > + !xfs_sb_has_incompat_log_feature(sbp, > + XFS_SB_FEAT_INCOMPAT_LOG_AGFL_WRAP_ALREADY_FIXED); > +} > + > +/* > + * Fix an AGFL wrapping that falls short of the end of the block by filling the > + * gap at the end of the block. > + */ > +STATIC int > +xfs_fixup_freelist_wrap_mount( > + struct xfs_trans *tp, > + struct xfs_buf *agfbp, > + struct xfs_perag *pag) > +{ > + struct xfs_mount *mp = tp->t_mountp; > + struct xfs_agf *agf; > + struct xfs_buf *agflbp; > + __be32 *agfl_bno; > + xfs_agnumber_t agno; > + uint32_t agfl_size; > + uint32_t flfirst; > + uint32_t fllast; > + int32_t active; > + int offset; > + int len; > + int error; > + > + if (pag->pagf_flcount == 0) > + return 0; > + > + agfl_size = xfs_agfl_size(mp); > + agf = XFS_BUF_TO_AGF(agfbp); > + agno = be32_to_cpu(agf->agf_seqno); > + flfirst = be32_to_cpu(agf->agf_flfirst); > + fllast = be32_to_cpu(agf->agf_fllast); > + > + /* Make sure we're either spot on or off by 1. */ > + active = fllast - flfirst + 1; > + if (active <= 0) > + active += agfl_size; > + if (active == pag->pagf_flcount) > + return 0; > + else if (active != pag->pagf_flcount + 1) > + return -EFSCORRUPTED; > + > + /* Would this have even passed muster on an old system? */ > + if (flfirst >= agfl_size - 1 || fllast >= agfl_size - 1 || > + pag->pagf_flcount > agfl_size - 1) > + return -EFSCORRUPTED; > + > + /* > + * Convert a 40-byte-padded agfl into a 36-byte-padded AGFL. > + * Therefore, we need to move the AGFL blocks > + * bno[flfirst..agfl_size - 2] to bno[flfirst + 1...agfl_size - 1]. > + * > + * Reusing the example above, if we had flfirst == 116, we need > + * to move bno[116] and bno[117] into bno[117] and bno[118], > + * respectively, and then increment flfirst. > + */ > + error = xfs_alloc_read_agfl(mp, tp, agno, &agflbp); > + if (error) > + return error; > + agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); > + > + len = (agfl_size - flfirst - 1) * sizeof(xfs_agblock_t); > + memmove(&agfl_bno[flfirst + 1], &agfl_bno[flfirst], len); > + offset = (char *)&agfl_bno[flfirst + 1] - (char *)agflbp->b_addr; > + be32_add_cpu(&agf->agf_flfirst, 1); > + > + xfs_trans_buf_set_type(tp, agflbp, XFS_BLFT_AGFL_BUF); > + xfs_trans_log_buf(tp, agflbp, offset, offset + len - 1); > + xfs_trans_brelse(tp, agflbp); > + agflbp = NULL; > + xfs_alloc_log_agf(tp, agfbp, XFS_AGF_FLFIRST); > + > + return 0; > +} > + > +/* > + * Fix an AGFL that touches the end of the block by moving the first or last > + * part of the list elsewhere in the AGFL so that old kernels don't trip over > + * wrapping issues. > + */ > +STATIC int > +xfs_fixup_freelist_wrap_unmount( > + struct xfs_trans *tp, > + struct xfs_buf *agfbp, > + struct xfs_perag *pag) > +{ > + struct xfs_mount *mp = tp->t_mountp; > + struct xfs_agf *agf; > + struct xfs_buf *agflbp; > + __be32 *agfl_bno; > + xfs_agnumber_t agno; > + uint32_t agfl_size; > + uint32_t flfirst; > + uint32_t fllast; > + int offset; > + int len; > + int error; > + > + agfl_size = xfs_agfl_size(mp); > + agf = XFS_BUF_TO_AGF(agfbp); > + agno = be32_to_cpu(agf->agf_seqno); > + flfirst = be32_to_cpu(agf->agf_flfirst); > + fllast = be32_to_cpu(agf->agf_fllast); > + > + /* Empty AGFL? Make sure we aren't pointing at the end. */ > + if (pag->pagf_flcount == 0) { > + if (flfirst >= agfl_size || fllast >= agfl_size) { > + agf->agf_flfirst = cpu_to_be32(1); > + agf->agf_fllast = 0; > + xfs_alloc_log_agf(tp, agfbp, > + XFS_AGF_FLFIRST | XFS_AGF_FLLAST); > + } > + return 0; > + } > + > + /* If we don't hit the end, we're done. */ > + if (flfirst < fllast && fllast != agfl_size - 1) > + return 0; > + > + /* > + * Move a start of a wrapped list towards the start of the agfl block. > + * Therefore, we need to move the AGFL blocks > + * bno[flfirst..agfl_size - 1] to bno[fllast + 1...agfl_size - flfirst]. > + * Then we reset flfirst and fllast appropriately. > + * > + * Reusing the example above, if we had flfirst == 117 and fllast == 4, > + * we need to move bno[117] and bno[118] into bno[5] and bno[6], > + * respectively, and then reset flfirst and fllast. > + * > + * If it's just the last block that touches the end, only move that. > + */ > + error = xfs_alloc_read_agfl(mp, tp, agno, &agflbp); > + if (error) > + return error; > + agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); > + > + if (fllast == agfl_size - 1) { > + /* Back the AGFL off from the end of the block. */ > + len = sizeof(xfs_agblock_t); > + agfl_bno[flfirst - 1] = agfl_bno[agfl_size - 1]; > + offset = (char *)&agfl_bno[flfirst - 1] - (char *)agflbp->b_addr; > + be32_add_cpu(&agf->agf_fllast, -1); > + be32_add_cpu(&agf->agf_flfirst, -1); > + } else { > + /* Move the first part of the AGFL towards the front. */ > + len = (agfl_size - flfirst) * sizeof(xfs_agblock_t); > + memcpy(&agfl_bno[fllast + 1], &agfl_bno[flfirst], len); > + offset = (char *)&agfl_bno[fllast + 1] - (char *)agflbp->b_addr; > + agf->agf_flfirst = 0; > + agf->agf_fllast = cpu_to_be32(pag->pagf_flcount - 1); > + } > + > + xfs_trans_buf_set_type(tp, agflbp, XFS_BLFT_AGFL_BUF); > + xfs_trans_log_buf(tp, agflbp, offset, offset + len - 1); > + xfs_trans_brelse(tp, agflbp); > + agflbp = NULL; > + xfs_alloc_log_agf(tp, agfbp, XFS_AGF_FLFIRST | XFS_AGF_FLLAST); > + > + return 0; > +} > + > +typedef int (*xfs_agf_apply_fn_t)(struct xfs_trans *tp, struct xfs_buf *agfbp, > + struct xfs_perag *pag); > + > +/* Apply something to every AGF. */ > +STATIC int > +xfs_fixup_agf_apply( > + struct xfs_mount *mp, > + xfs_agf_apply_fn_t fn) > +{ > + struct xfs_trans *tp; > + struct xfs_perag *pag; > + struct xfs_buf *agfbp; > + xfs_agnumber_t agno; > + int error; > + > + error = xfs_trans_alloc(mp, &M_RES(mp)->tr_growdata, 0, 0, 0, &tp); This can get called when we're in freeze context, so I think this needs to be: error = xfs_trans_alloc(mp, &M_RES(mp)->tr_growdata, 0, 0, XFS_TRANS_NO_WRITECOUNT, &tp); I saw xfs/119 cough up an error about locking problems and deadlock. --D > + if (error) > + return error; > + > + for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) { > + error = xfs_alloc_read_agf(mp, tp, agno, 0, &agfbp); > + if (error) > + goto cancel; > + if (!agfbp) { > + error = -ENOMEM; > + goto cancel; > + } > + pag = xfs_perag_get(mp, agno); > + error = fn(tp, agfbp, pag); > + xfs_perag_put(pag); > + xfs_trans_brelse(tp, agfbp); > + if (error) > + goto cancel; > + } > + > + return xfs_trans_commit(tp); > +cancel: > + xfs_trans_cancel(tp); > + return error; > +} > + > +/* Fix AGFL wrapping so we can use the filesystem. */ > +int > +xfs_fixup_agfl_wrap_mount( > + struct xfs_mount *mp) > +{ > + if (!xfs_sb_version_needs_agfl_wrap_fixes(&mp->m_sb)) > + return 0; > + > + return xfs_fixup_agf_apply(mp, xfs_fixup_freelist_wrap_mount); > +} > + > +/* Fix AGFL wrapping so old kernels can use this filesystem. */ > +int > +xfs_fixup_agfl_wrap_unmount( > + struct xfs_mount *mp) > +{ > + if (!xfs_sb_version_needs_agfl_wrap_fixes(&mp->m_sb)) > + return 0; > + > + return xfs_fixup_agf_apply(mp, xfs_fixup_freelist_wrap_unmount); > +} > diff --git a/fs/xfs/xfs_fixups.h b/fs/xfs/xfs_fixups.h > new file mode 100644 > index 0000000..fb52a96 > --- /dev/null > +++ b/fs/xfs/xfs_fixups.h > @@ -0,0 +1,26 @@ > +/* > + * Copyright (C) 2018 Oracle. All Rights Reserved. > + * > + * Author: Darrick J. Wong <darrick.wong@oracle.com> > + * > + * This program is free software; you can redistribute it and/or > + * modify it under the terms of the GNU General Public License > + * as published by the Free Software Foundation; either version 2 > + * of the License, or (at your option) any later version. > + * > + * This program is distributed in the hope that it would be useful, > + * but WITHOUT ANY WARRANTY; without even the implied warranty of > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the > + * GNU General Public License for more details. > + * > + * You should have received a copy of the GNU General Public License > + * along with this program; if not, write the Free Software Foundation, > + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. > + */ > +#ifndef __XFS_FIXUPS_H__ > +#define __XFS_FIXUPS_H__ > + > +int xfs_fixup_agfl_wrap_mount(struct xfs_mount *mp); > +int xfs_fixup_agfl_wrap_unmount(struct xfs_mount *mp); > + > +#endif /* __XFS_FIXUPS_H__ */ > diff --git a/fs/xfs/xfs_mount.c b/fs/xfs/xfs_mount.c > index 98fd41c..eb284aa 100644 > --- a/fs/xfs/xfs_mount.c > +++ b/fs/xfs/xfs_mount.c > @@ -46,7 +46,7 @@ > #include "xfs_refcount_btree.h" > #include "xfs_reflink.h" > #include "xfs_extent_busy.h" > - > +#include "xfs_fixups.h" > > static DEFINE_MUTEX(xfs_uuid_table_mutex); > static int xfs_uuid_table_size; > @@ -875,6 +875,16 @@ xfs_mountfs( > } > > /* > + * Make sure our AGFL counters do not wrap the end of the block > + * in a troublesome manner. > + */ > + error = xfs_fixup_agfl_wrap_mount(mp); > + if (error) { > + xfs_warn(mp, "Failed to fix agfl wrapping. Run xfs_repair."); > + goto out_log_dealloc; > + } > + > + /* > * Get and sanity-check the root inode. > * Save the pointer to it in the mount structure. > */ > @@ -1128,6 +1138,15 @@ xfs_unmountfs( > xfs_qm_unmount(mp); > > /* > + * Make sure our AGFL counters do not wrap the end of the block > + * in a troublesome manner for old kernels. > + */ > + error = xfs_fixup_agfl_wrap_unmount(mp); > + if (error) > + xfs_warn(mp, "Unable to fix agfl wrapping. " > + "This may cause problems on next mount."); > + > + /* > * Unreserve any blocks we have so that when we unmount we don't account > * the reserved free space as used. This is really only necessary for > * lazy superblock counting because it trusts the incore superblock > diff --git a/fs/xfs/xfs_super.c b/fs/xfs/xfs_super.c > index 624a802..d9aa39a 100644 > --- a/fs/xfs/xfs_super.c > +++ b/fs/xfs/xfs_super.c > @@ -50,6 +50,7 @@ > #include "xfs_refcount_item.h" > #include "xfs_bmap_item.h" > #include "xfs_reflink.h" > +#include "xfs_fixups.h" > > #include <linux/namei.h> > #include <linux/dax.h> > @@ -1206,6 +1207,15 @@ xfs_quiesce_attr( > xfs_reclaim_inodes(mp, 0); > xfs_reclaim_inodes(mp, SYNC_WAIT); > > + /* > + * Make sure our AGFL counters do not wrap the end of the block > + * in a troublesome manner for old kernels. > + */ > + error = xfs_fixup_agfl_wrap_unmount(mp); > + if (error) > + xfs_warn(mp, "Unable to fix agfl wrapping. " > + "This may cause problems on next mount."); > + > /* Push the superblock and write an unmount record */ > error = xfs_log_sbcount(mp); > if (error) > > -- > To unsubscribe from this list: send the line "unsubscribe linux-xfs" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Feb 22, 2018 at 06:00:15PM -0800, Darrick J. Wong wrote: > From: Darrick J. Wong <darrick.wong@oracle.com> > > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> > --- > fs/xfs/Makefile | 1 > fs/xfs/xfs_fixups.c | 310 +++++++++++++++++++++++++++++++++++++++++++++++++++ > fs/xfs/xfs_fixups.h | 26 ++++ > fs/xfs/xfs_mount.c | 21 +++ > fs/xfs/xfs_super.c | 10 ++ > 5 files changed, 367 insertions(+), 1 deletion(-) > create mode 100644 fs/xfs/xfs_fixups.c > create mode 100644 fs/xfs/xfs_fixups.h > > > diff --git a/fs/xfs/Makefile b/fs/xfs/Makefile > index b03c77e..f88368a 100644 > --- a/fs/xfs/Makefile > +++ b/fs/xfs/Makefile > @@ -86,6 +86,7 @@ xfs-y += xfs_aops.o \ > xfs_extent_busy.o \ > xfs_file.o \ > xfs_filestream.o \ > + xfs_fixups.o \ > xfs_fsmap.o \ > xfs_fsops.o \ > xfs_globals.o \ > diff --git a/fs/xfs/xfs_fixups.c b/fs/xfs/xfs_fixups.c > new file mode 100644 > index 0000000..0cad7bb > --- /dev/null > +++ b/fs/xfs/xfs_fixups.c > @@ -0,0 +1,310 @@ > +/* > + * Copyright (C) 2018 Oracle. All Rights Reserved. > + * > + * Author: Darrick J. Wong <darrick.wong@oracle.com> > + * > + * This program is free software; you can redistribute it and/or > + * modify it under the terms of the GNU General Public License > + * as published by the Free Software Foundation; either version 2 > + * of the License, or (at your option) any later version. > + * > + * This program is distributed in the hope that it would be useful, > + * but WITHOUT ANY WARRANTY; without even the implied warranty of > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the > + * GNU General Public License for more details. > + * > + * You should have received a copy of the GNU General Public License > + * along with this program; if not, write the Free Software Foundation, > + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. > + */ > +#include "xfs.h" > +#include "xfs_fs.h" > +#include "xfs_shared.h" > +#include "xfs_format.h" > +#include "xfs_log_format.h" > +#include "xfs_trans_resv.h" > +#include "xfs_sb.h" > +#include "xfs_mount.h" > +#include "xfs_alloc.h" > +#include "xfs_trans.h" > +#include "xfs_fixups.h" > + > +/* > + * v5 AGFL padding defects > + * > + * When the v5 format was first introduced, there was a defect in the struct > + * xfs_agfl definition that resulted in XFS_AGFL_SIZE returning different XFS_AGFL_SIZE() no longer exists as of the previous patch. It might be better to size refer to "the agfl size" generically. > + * values depending on the compiler padding. On a fs with 512-byte sectors, > + * this meant that XFS_AGFL_SIZE was 119 on i386, but 118 on x64. Commit > + * 96f859d52bcb1 ("libxfs: pack the agfl header structure so XFS_AGFL_SIZE is > + * correct") changed the definition to disable padding the end of the > + * structure, and was accepted into Linux 4.5. Since then, the AGFL has > + * always used the larger size (e.g. 119 entries on a 512b sector fs). > + * > + * Unfortunately, pre-4.5 kernels can produce filesystems with AGFLs that wrap > + * at the smaller size, and those kernels are not prepared to handle the > + * longer size. This typically manifests itself as an AGF verifier corruption > + * error followed by a filesystem shutdown. While we encourage admins to stay > + * current with software, we would like to avoid this intermittent breakage. > + * > + * Any v5 filesystem which has a feature bit set for a feature that was > + * introduced after Linux 4.5 will not have this problem, as such kernels > + * cannot be mounted on older kernels. v4 filesystems are also unaffected. > + * > + * Therefore, we add two fixup functions -- the first runs at mount time to > + * detect a short-wrapped AGFL and fix it; the second runs at unmount, freeze, > + * or remount-ro time to move a wrapped AGFL to the beginning of the list. > + * This reduces the likelihood of a screwup to the scenario where you have (a) > + * a filesystem with no post-4.5 features (reflink, rmap), (b) the AGFL wraps, > + * (c) the filesystem goes down leaving a dirty log, and (d) the dirty > + * filesystem is mounted on an old kernel. > + */ > + While the mount vs. unmount time fixups serve different purposes and have different conditions, do we really need two separate fixup implementations? E.g., if we have to unwrap the AGFL at unmount time, why not just reuse that fixup at mount time (when needed) as well? I suspect the difference would only be the length of what we can consider valid in the agfl at the time. > +/* > + * Decide if we need to have the agfl wrapping fixes applied. This only > + * affects v5 filesystems that do not have any features enabled that did not > + * exist when the agfl padding fix went in. > + * > + * Features already present when the fix went in were finobt, ftype, spinodes. > + * If we see something new (e.g. reflink) then don't bother. > + */ > +#define XFS_SB_FEAT_RO_COMPAT_AGFL_WRAP_ALREADY_FIXED \ > + (~(XFS_SB_FEAT_RO_COMPAT_FINOBT)) > +#define XFS_SB_FEAT_INCOMPAT_AGFL_WRAP_ALREADY_FIXED \ > + (~(XFS_SB_FEAT_INCOMPAT_FTYPE | \ > + XFS_SB_FEAT_INCOMPAT_SPINODES)) > +#define XFS_SB_FEAT_INCOMPAT_LOG_AGFL_WRAP_ALREADY_FIXED \ > + (~0) > +static inline bool xfs_sb_version_needs_agfl_wrap_fixes(struct xfs_sb *sbp) > +{ > + return xfs_sb_version_hascrc(sbp) && > + !xfs_sb_has_incompat_feature(sbp, > + XFS_SB_FEAT_INCOMPAT_AGFL_WRAP_ALREADY_FIXED) && > + !xfs_sb_has_ro_compat_feature(sbp, > + XFS_SB_FEAT_RO_COMPAT_AGFL_WRAP_ALREADY_FIXED) && > + !xfs_sb_has_incompat_log_feature(sbp, > + XFS_SB_FEAT_INCOMPAT_LOG_AGFL_WRAP_ALREADY_FIXED); > +} > + > +/* > + * Fix an AGFL wrapping that falls short of the end of the block by filling the > + * gap at the end of the block. > + */ > +STATIC int > +xfs_fixup_freelist_wrap_mount( > + struct xfs_trans *tp, > + struct xfs_buf *agfbp, > + struct xfs_perag *pag) > +{ > + struct xfs_mount *mp = tp->t_mountp; > + struct xfs_agf *agf; > + struct xfs_buf *agflbp; > + __be32 *agfl_bno; > + xfs_agnumber_t agno; > + uint32_t agfl_size; > + uint32_t flfirst; > + uint32_t fllast; > + int32_t active; > + int offset; > + int len; > + int error; > + > + if (pag->pagf_flcount == 0) > + return 0; > + > + agfl_size = xfs_agfl_size(mp); > + agf = XFS_BUF_TO_AGF(agfbp); > + agno = be32_to_cpu(agf->agf_seqno); > + flfirst = be32_to_cpu(agf->agf_flfirst); > + fllast = be32_to_cpu(agf->agf_fllast); > + > + /* Make sure we're either spot on or off by 1. */ > + active = fllast - flfirst + 1; > + if (active <= 0) > + active += agfl_size; > + if (active == pag->pagf_flcount) > + return 0; > + else if (active != pag->pagf_flcount + 1) > + return -EFSCORRUPTED; > + So we're not attempting to cover the case where the agfl has 1 more block than the agfl size (i.e., the case where an fs goes back to a kernel with an unpacked header)? I'm wondering whether using the unmount algorithm in both places (as noted above) would also facilitate this mechanism to work both ways (unpacked <-> packed), provided particular conditions are met. I suppose that could be considered regardless, but the less variance the better IMO. > + /* Would this have even passed muster on an old system? */ Comment doesn't really explain what is going on here..? > + if (flfirst >= agfl_size - 1 || fllast >= agfl_size - 1 || > + pag->pagf_flcount > agfl_size - 1) > + return -EFSCORRUPTED; > + I take it these are checks for whether the agfl was previously corrupted based on the unpacked size? FWIW, it might be a bit more clear to use an old/unpacked_agfl_size variable to declare intent (here and in some of the comments and whatnot that follow). > + /* > + * Convert a 40-byte-padded agfl into a 36-byte-padded AGFL. > + * Therefore, we need to move the AGFL blocks > + * bno[flfirst..agfl_size - 2] to bno[flfirst + 1...agfl_size - 1]. > + * > + * Reusing the example above, if we had flfirst == 116, we need > + * to move bno[116] and bno[117] into bno[117] and bno[118], > + * respectively, and then increment flfirst. > + */ Kind of a strange example to use given that it doesn't mention wrapping.. this only triggers if the agfl was previously wrapped, right? > + error = xfs_alloc_read_agfl(mp, tp, agno, &agflbp); > + if (error) > + return error; > + agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); > + > + len = (agfl_size - flfirst - 1) * sizeof(xfs_agblock_t); > + memmove(&agfl_bno[flfirst + 1], &agfl_bno[flfirst], len); > + offset = (char *)&agfl_bno[flfirst + 1] - (char *)agflbp->b_addr; > + be32_add_cpu(&agf->agf_flfirst, 1); > + > + xfs_trans_buf_set_type(tp, agflbp, XFS_BLFT_AGFL_BUF); > + xfs_trans_log_buf(tp, agflbp, offset, offset + len - 1); > + xfs_trans_brelse(tp, agflbp); > + agflbp = NULL; > + xfs_alloc_log_agf(tp, agfbp, XFS_AGF_FLFIRST); > + > + return 0; > +} > + > +/* > + * Fix an AGFL that touches the end of the block by moving the first or last > + * part of the list elsewhere in the AGFL so that old kernels don't trip over > + * wrapping issues. > + */ > +STATIC int > +xfs_fixup_freelist_wrap_unmount( > + struct xfs_trans *tp, > + struct xfs_buf *agfbp, > + struct xfs_perag *pag) > +{ > + struct xfs_mount *mp = tp->t_mountp; > + struct xfs_agf *agf; > + struct xfs_buf *agflbp; > + __be32 *agfl_bno; > + xfs_agnumber_t agno; > + uint32_t agfl_size; > + uint32_t flfirst; > + uint32_t fllast; > + int offset; > + int len; > + int error; > + > + agfl_size = xfs_agfl_size(mp); > + agf = XFS_BUF_TO_AGF(agfbp); > + agno = be32_to_cpu(agf->agf_seqno); > + flfirst = be32_to_cpu(agf->agf_flfirst); > + fllast = be32_to_cpu(agf->agf_fllast); > + > + /* Empty AGFL? Make sure we aren't pointing at the end. */ > + if (pag->pagf_flcount == 0) { > + if (flfirst >= agfl_size || fllast >= agfl_size) { When would either field be >= agfl_size? Isn't that where they wrap? > + agf->agf_flfirst = cpu_to_be32(1); > + agf->agf_fllast = 0; > + xfs_alloc_log_agf(tp, agfbp, > + XFS_AGF_FLFIRST | XFS_AGF_FLLAST); > + } > + return 0; > + } > + > + /* If we don't hit the end, we're done. */ > + if (flfirst < fllast && fllast != agfl_size - 1) > + return 0; > + > + /* > + * Move a start of a wrapped list towards the start of the agfl block. FWIW, this and the subsequent comments kind of gave me the impression that the agfl would be "shifted" towards the start of the block. The code doesn't do that, but rather rotates the start/head of the agfl to the tail and adjusts the pointers (changing the effective order). That seems technically Ok, but I had to grok the code and read back to understand the comment rather than the other way around. > + * Therefore, we need to move the AGFL blocks > + * bno[flfirst..agfl_size - 1] to bno[fllast + 1...agfl_size - flfirst]. > + * Then we reset flfirst and fllast appropriately. > + * > + * Reusing the example above, if we had flfirst == 117 and fllast == 4, > + * we need to move bno[117] and bno[118] into bno[5] and bno[6], > + * respectively, and then reset flfirst and fllast. > + * > + * If it's just the last block that touches the end, only move that. > + */ > + error = xfs_alloc_read_agfl(mp, tp, agno, &agflbp); > + if (error) > + return error; > + agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); > + > + if (fllast == agfl_size - 1) { > + /* Back the AGFL off from the end of the block. */ > + len = sizeof(xfs_agblock_t); > + agfl_bno[flfirst - 1] = agfl_bno[agfl_size - 1]; What if the agfl is full (flfirst == 0)? > + offset = (char *)&agfl_bno[flfirst - 1] - (char *)agflbp->b_addr; > + be32_add_cpu(&agf->agf_fllast, -1); > + be32_add_cpu(&agf->agf_flfirst, -1); > + } else { > + /* Move the first part of the AGFL towards the front. */ > + len = (agfl_size - flfirst) * sizeof(xfs_agblock_t); > + memcpy(&agfl_bno[fllast + 1], &agfl_bno[flfirst], len); > + offset = (char *)&agfl_bno[fllast + 1] - (char *)agflbp->b_addr; > + agf->agf_flfirst = 0; > + agf->agf_fllast = cpu_to_be32(pag->pagf_flcount - 1); Similar question here as above.. it looks like the copy may be a no-op, but we still reset flfirst/fllast (which may also be fine, but unnecessary). The more interesting case here is flcount == 1 and flfirst == fllast, where it looks like we'd wrongly reset first/last (and perhaps copy more than we should..?). Ugh. Remind me again why we can't just detect a mismatch, fail the mount and tell the user to repair? :P Random thought... have we considered anything less invasive/more simple for this issue than low level buffer manipulation? For example, could we address the unpacked -> packed case by just allocating a block to fill the gap at the end of the wrapped agfl? Perhaps that could even be done more discreetly in xfs_alloc_fix_freelist() when we already have a transaction set up, etc. That wouldn't help going the other direction, but maybe we could consider similar logic to free an extraneous block and then worry about protecting users who won't upgrade broken kernels separately if that really continues to be a problem. Brian > + } > + > + xfs_trans_buf_set_type(tp, agflbp, XFS_BLFT_AGFL_BUF); > + xfs_trans_log_buf(tp, agflbp, offset, offset + len - 1); > + xfs_trans_brelse(tp, agflbp); > + agflbp = NULL; > + xfs_alloc_log_agf(tp, agfbp, XFS_AGF_FLFIRST | XFS_AGF_FLLAST); > + > + return 0; > +} > + > +typedef int (*xfs_agf_apply_fn_t)(struct xfs_trans *tp, struct xfs_buf *agfbp, > + struct xfs_perag *pag); > + > +/* Apply something to every AGF. */ > +STATIC int > +xfs_fixup_agf_apply( > + struct xfs_mount *mp, > + xfs_agf_apply_fn_t fn) > +{ > + struct xfs_trans *tp; > + struct xfs_perag *pag; > + struct xfs_buf *agfbp; > + xfs_agnumber_t agno; > + int error; > + > + error = xfs_trans_alloc(mp, &M_RES(mp)->tr_growdata, 0, 0, 0, &tp); > + if (error) > + return error; > + > + for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) { > + error = xfs_alloc_read_agf(mp, tp, agno, 0, &agfbp); > + if (error) > + goto cancel; > + if (!agfbp) { > + error = -ENOMEM; > + goto cancel; > + } > + pag = xfs_perag_get(mp, agno); > + error = fn(tp, agfbp, pag); > + xfs_perag_put(pag); > + xfs_trans_brelse(tp, agfbp); > + if (error) > + goto cancel; > + } > + > + return xfs_trans_commit(tp); > +cancel: > + xfs_trans_cancel(tp); > + return error; > +} > + > +/* Fix AGFL wrapping so we can use the filesystem. */ > +int > +xfs_fixup_agfl_wrap_mount( > + struct xfs_mount *mp) > +{ > + if (!xfs_sb_version_needs_agfl_wrap_fixes(&mp->m_sb)) > + return 0; > + > + return xfs_fixup_agf_apply(mp, xfs_fixup_freelist_wrap_mount); > +} > + > +/* Fix AGFL wrapping so old kernels can use this filesystem. */ > +int > +xfs_fixup_agfl_wrap_unmount( > + struct xfs_mount *mp) > +{ > + if (!xfs_sb_version_needs_agfl_wrap_fixes(&mp->m_sb)) > + return 0; > + > + return xfs_fixup_agf_apply(mp, xfs_fixup_freelist_wrap_unmount); > +} > diff --git a/fs/xfs/xfs_fixups.h b/fs/xfs/xfs_fixups.h > new file mode 100644 > index 0000000..fb52a96 > --- /dev/null > +++ b/fs/xfs/xfs_fixups.h > @@ -0,0 +1,26 @@ > +/* > + * Copyright (C) 2018 Oracle. All Rights Reserved. > + * > + * Author: Darrick J. Wong <darrick.wong@oracle.com> > + * > + * This program is free software; you can redistribute it and/or > + * modify it under the terms of the GNU General Public License > + * as published by the Free Software Foundation; either version 2 > + * of the License, or (at your option) any later version. > + * > + * This program is distributed in the hope that it would be useful, > + * but WITHOUT ANY WARRANTY; without even the implied warranty of > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the > + * GNU General Public License for more details. > + * > + * You should have received a copy of the GNU General Public License > + * along with this program; if not, write the Free Software Foundation, > + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. > + */ > +#ifndef __XFS_FIXUPS_H__ > +#define __XFS_FIXUPS_H__ > + > +int xfs_fixup_agfl_wrap_mount(struct xfs_mount *mp); > +int xfs_fixup_agfl_wrap_unmount(struct xfs_mount *mp); > + > +#endif /* __XFS_FIXUPS_H__ */ > diff --git a/fs/xfs/xfs_mount.c b/fs/xfs/xfs_mount.c > index 98fd41c..eb284aa 100644 > --- a/fs/xfs/xfs_mount.c > +++ b/fs/xfs/xfs_mount.c > @@ -46,7 +46,7 @@ > #include "xfs_refcount_btree.h" > #include "xfs_reflink.h" > #include "xfs_extent_busy.h" > - > +#include "xfs_fixups.h" > > static DEFINE_MUTEX(xfs_uuid_table_mutex); > static int xfs_uuid_table_size; > @@ -875,6 +875,16 @@ xfs_mountfs( > } > > /* > + * Make sure our AGFL counters do not wrap the end of the block > + * in a troublesome manner. > + */ > + error = xfs_fixup_agfl_wrap_mount(mp); > + if (error) { > + xfs_warn(mp, "Failed to fix agfl wrapping. Run xfs_repair."); > + goto out_log_dealloc; > + } > + > + /* > * Get and sanity-check the root inode. > * Save the pointer to it in the mount structure. > */ > @@ -1128,6 +1138,15 @@ xfs_unmountfs( > xfs_qm_unmount(mp); > > /* > + * Make sure our AGFL counters do not wrap the end of the block > + * in a troublesome manner for old kernels. > + */ > + error = xfs_fixup_agfl_wrap_unmount(mp); > + if (error) > + xfs_warn(mp, "Unable to fix agfl wrapping. " > + "This may cause problems on next mount."); > + > + /* > * Unreserve any blocks we have so that when we unmount we don't account > * the reserved free space as used. This is really only necessary for > * lazy superblock counting because it trusts the incore superblock > diff --git a/fs/xfs/xfs_super.c b/fs/xfs/xfs_super.c > index 624a802..d9aa39a 100644 > --- a/fs/xfs/xfs_super.c > +++ b/fs/xfs/xfs_super.c > @@ -50,6 +50,7 @@ > #include "xfs_refcount_item.h" > #include "xfs_bmap_item.h" > #include "xfs_reflink.h" > +#include "xfs_fixups.h" > > #include <linux/namei.h> > #include <linux/dax.h> > @@ -1206,6 +1207,15 @@ xfs_quiesce_attr( > xfs_reclaim_inodes(mp, 0); > xfs_reclaim_inodes(mp, SYNC_WAIT); > > + /* > + * Make sure our AGFL counters do not wrap the end of the block > + * in a troublesome manner for old kernels. > + */ > + error = xfs_fixup_agfl_wrap_unmount(mp); > + if (error) > + xfs_warn(mp, "Unable to fix agfl wrapping. " > + "This may cause problems on next mount."); > + > /* Push the superblock and write an unmount record */ > error = xfs_log_sbcount(mp); > if (error) > > -- > To unsubscribe from this list: send the line "unsubscribe linux-xfs" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Tue, Feb 27, 2018 at 02:35:49PM -0500, Brian Foster wrote: > On Thu, Feb 22, 2018 at 06:00:15PM -0800, Darrick J. Wong wrote: > > From: Darrick J. Wong <darrick.wong@oracle.com> > > > > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> > > --- > > fs/xfs/Makefile | 1 > > fs/xfs/xfs_fixups.c | 310 +++++++++++++++++++++++++++++++++++++++++++++++++++ > > fs/xfs/xfs_fixups.h | 26 ++++ > > fs/xfs/xfs_mount.c | 21 +++ > > fs/xfs/xfs_super.c | 10 ++ > > 5 files changed, 367 insertions(+), 1 deletion(-) > > create mode 100644 fs/xfs/xfs_fixups.c > > create mode 100644 fs/xfs/xfs_fixups.h > > > > > > diff --git a/fs/xfs/Makefile b/fs/xfs/Makefile > > index b03c77e..f88368a 100644 > > --- a/fs/xfs/Makefile > > +++ b/fs/xfs/Makefile > > @@ -86,6 +86,7 @@ xfs-y += xfs_aops.o \ > > xfs_extent_busy.o \ > > xfs_file.o \ > > xfs_filestream.o \ > > + xfs_fixups.o \ > > xfs_fsmap.o \ > > xfs_fsops.o \ > > xfs_globals.o \ > > diff --git a/fs/xfs/xfs_fixups.c b/fs/xfs/xfs_fixups.c > > new file mode 100644 > > index 0000000..0cad7bb > > --- /dev/null > > +++ b/fs/xfs/xfs_fixups.c > > @@ -0,0 +1,310 @@ > > +/* > > + * Copyright (C) 2018 Oracle. All Rights Reserved. > > + * > > + * Author: Darrick J. Wong <darrick.wong@oracle.com> > > + * > > + * This program is free software; you can redistribute it and/or > > + * modify it under the terms of the GNU General Public License > > + * as published by the Free Software Foundation; either version 2 > > + * of the License, or (at your option) any later version. > > + * > > + * This program is distributed in the hope that it would be useful, > > + * but WITHOUT ANY WARRANTY; without even the implied warranty of > > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the > > + * GNU General Public License for more details. > > + * > > + * You should have received a copy of the GNU General Public License > > + * along with this program; if not, write the Free Software Foundation, > > + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. > > + */ > > +#include "xfs.h" > > +#include "xfs_fs.h" > > +#include "xfs_shared.h" > > +#include "xfs_format.h" > > +#include "xfs_log_format.h" > > +#include "xfs_trans_resv.h" > > +#include "xfs_sb.h" > > +#include "xfs_mount.h" > > +#include "xfs_alloc.h" > > +#include "xfs_trans.h" > > +#include "xfs_fixups.h" > > + > > +/* > > + * v5 AGFL padding defects > > + * > > + * When the v5 format was first introduced, there was a defect in the struct > > + * xfs_agfl definition that resulted in XFS_AGFL_SIZE returning different > > XFS_AGFL_SIZE() no longer exists as of the previous patch. It might be > better to size refer to "the agfl size" generically. How about: "...the agfl length calculation formula (at the time known as XFS_AGFL_SIZE)..." > > + * values depending on the compiler padding. On a fs with 512-byte sectors, > > + * this meant that XFS_AGFL_SIZE was 119 on i386, but 118 on x64. Commit > > + * 96f859d52bcb1 ("libxfs: pack the agfl header structure so XFS_AGFL_SIZE is > > + * correct") changed the definition to disable padding the end of the > > + * structure, and was accepted into Linux 4.5. Since then, the AGFL has > > + * always used the larger size (e.g. 119 entries on a 512b sector fs). > > + * > > + * Unfortunately, pre-4.5 kernels can produce filesystems with AGFLs that wrap > > + * at the smaller size, and those kernels are not prepared to handle the > > + * longer size. This typically manifests itself as an AGF verifier corruption > > + * error followed by a filesystem shutdown. While we encourage admins to stay > > + * current with software, we would like to avoid this intermittent breakage. > > + * > > + * Any v5 filesystem which has a feature bit set for a feature that was > > + * introduced after Linux 4.5 will not have this problem, as such kernels > > + * cannot be mounted on older kernels. v4 filesystems are also unaffected. > > + * > > + * Therefore, we add two fixup functions -- the first runs at mount time to > > + * detect a short-wrapped AGFL and fix it; the second runs at unmount, freeze, > > + * or remount-ro time to move a wrapped AGFL to the beginning of the list. > > + * This reduces the likelihood of a screwup to the scenario where you have (a) > > + * a filesystem with no post-4.5 features (reflink, rmap), (b) the AGFL wraps, > > + * (c) the filesystem goes down leaving a dirty log, and (d) the dirty > > + * filesystem is mounted on an old kernel. > > + */ > > + > > While the mount vs. unmount time fixups serve different purposes and > have different conditions, do we really need two separate fixup > implementations? E.g., if we have to unwrap the AGFL at unmount time, > why not just reuse that fixup at mount time (when needed) as well? I > suspect the difference would only be the length of what we can consider > valid in the agfl at the time. They're doing two different things. I will use ASCII art to demonstrate the mount-time function: Let's say the AGFL block has space for 10 items. It looks like this: 0 1 2 3 4 5 6 7 8 9 Header | | | | | | | | | | | When the fs is freshly formatted or repaired, the AGFL will look like: 0 1 2 3 4 5 6 7 8 9 Header | A | B | C | D | E | F | | | | | ^-- flfirst ^-- fllast Due to the padding problems prior to 4.5 ("4.5 agfl padding fix"), XFS_AGFL_SIZE would return 10 on 32-bit systems and 9 on 64-bit systems. Therefore, if the AGFL wrapped on a 64-bit kernel we would end up with: 0 1 2 3 4 5 6 7 8 9 Header | D | E | F | | | | A | B | C | ? | ^-- fllast ^-- flfirst Note that block 9's contents are undefined because we didn't write anything there. Now the "4.5 agfl padding fix" has corrected XFS_AGFL_SIZE to return 10 in all cases, this looks like an AGF corruption because flcount is 6 but the distance between flfirst and fllast is 7. So long as the list wraps and the mismatch is exactly one block we can fix this. The mount time fixer in this patch does this by moving the records between flfirst and xfs_agfl_size (A, B, and C) to the end of the block to close the gap: 0 1 2 3 4 5 6 7 8 9 Header | D | E | F | | | | | A | B | C | ^-- fllast ^-- flfirst If the count is off by more than 1 then the AGF is truly corrupt and we bail out. Now let's say that we run for a while and want to unmount, but our AGFL wraps like this: 0 1 2 3 4 5 6 7 8 9 Header | T | U | V | | | | | Q | R | S | ^-- fllast ^-- flfirst We don't know that the next kernel to mount this filesystem will have the "4.5 agfl padding fix" applied to it; if it does not, it will flag the AGF as corrupted because flcount is 6 but in its view the distance between flfirst and fllast (which omits bno[9]) is 5. We don't want it to choke on that, so the unmount fixer moves all the records between flfirst and the end (Q, R, and S) towards the start of the block and resets flfirst/fllast: 0 1 2 3 4 5 6 7 8 9 Header | T | U | V | Q | R | S | | | | | ^-- flfirst ^-- fllast Since the AGFL no longer wraps at all, it doesn't matter if the next kernel to mount this filesystem has the "4.5 agfl padding fix" applied; all kernels can handle this correctly. > > +/* > > + * Decide if we need to have the agfl wrapping fixes applied. This only > > + * affects v5 filesystems that do not have any features enabled that did not > > + * exist when the agfl padding fix went in. > > + * > > + * Features already present when the fix went in were finobt, ftype, spinodes. > > + * If we see something new (e.g. reflink) then don't bother. > > + */ > > +#define XFS_SB_FEAT_RO_COMPAT_AGFL_WRAP_ALREADY_FIXED \ > > + (~(XFS_SB_FEAT_RO_COMPAT_FINOBT)) > > +#define XFS_SB_FEAT_INCOMPAT_AGFL_WRAP_ALREADY_FIXED \ > > + (~(XFS_SB_FEAT_INCOMPAT_FTYPE | \ > > + XFS_SB_FEAT_INCOMPAT_SPINODES)) > > +#define XFS_SB_FEAT_INCOMPAT_LOG_AGFL_WRAP_ALREADY_FIXED \ > > + (~0) > > +static inline bool xfs_sb_version_needs_agfl_wrap_fixes(struct xfs_sb *sbp) > > +{ > > + return xfs_sb_version_hascrc(sbp) && > > + !xfs_sb_has_incompat_feature(sbp, > > + XFS_SB_FEAT_INCOMPAT_AGFL_WRAP_ALREADY_FIXED) && > > + !xfs_sb_has_ro_compat_feature(sbp, > > + XFS_SB_FEAT_RO_COMPAT_AGFL_WRAP_ALREADY_FIXED) && > > + !xfs_sb_has_incompat_log_feature(sbp, > > + XFS_SB_FEAT_INCOMPAT_LOG_AGFL_WRAP_ALREADY_FIXED); > > +} > > + > > +/* > > + * Fix an AGFL wrapping that falls short of the end of the block by filling the > > + * gap at the end of the block. > > + */ > > +STATIC int > > +xfs_fixup_freelist_wrap_mount( > > + struct xfs_trans *tp, > > + struct xfs_buf *agfbp, > > + struct xfs_perag *pag) > > +{ > > + struct xfs_mount *mp = tp->t_mountp; > > + struct xfs_agf *agf; > > + struct xfs_buf *agflbp; > > + __be32 *agfl_bno; > > + xfs_agnumber_t agno; > > + uint32_t agfl_size; > > + uint32_t flfirst; > > + uint32_t fllast; > > + int32_t active; > > + int offset; > > + int len; > > + int error; > > + > > + if (pag->pagf_flcount == 0) > > + return 0; > > + > > + agfl_size = xfs_agfl_size(mp); > > + agf = XFS_BUF_TO_AGF(agfbp); > > + agno = be32_to_cpu(agf->agf_seqno); > > + flfirst = be32_to_cpu(agf->agf_flfirst); > > + fllast = be32_to_cpu(agf->agf_fllast); > > + > > + /* Make sure we're either spot on or off by 1. */ > > + active = fllast - flfirst + 1; > > + if (active <= 0) > > + active += agfl_size; > > + if (active == pag->pagf_flcount) > > + return 0; > > + else if (active != pag->pagf_flcount + 1) > > + return -EFSCORRUPTED; > > + > > So we're not attempting to cover the case where the agfl has 1 more > block than the agfl size (i.e., the case where an fs goes back to a > kernel with an unpacked header)? We don't know how the next kernel to touch this filesystem will define XFS_AGFL_SIZE -- it could be a 4.5+ kernel (same agfl size), a 32-bit pre-4.5 kernel (same agfl size), or a 64-bit pre-4.5 kernel (small agfl size). > I'm wondering whether using the unmount algorithm in both places (as > noted above) would also facilitate this mechanism to work both ways > (unpacked <-> packed), provided particular conditions are met. I suppose > that could be considered regardless, but the less variance the better > IMO. In principle I suppose we could have a single function to handle all the rearranging, but I'd rather have two functions to handle the two cases. I'm open to refactoring the common parts out of the xfs_fixup_freelist_wrap_*mount functions. > > + /* Would this have even passed muster on an old system? */ > > Comment doesn't really explain what is going on here..? /* * If the distance between flfirst and fllast mismatches flcount by more * than 1, then there's more wrong with this agfl than just the padding * problem. Bail out completely, which will force the admin to run * xfs_repair. */ > > > + if (flfirst >= agfl_size - 1 || fllast >= agfl_size - 1 || > > + pag->pagf_flcount > agfl_size - 1) > > + return -EFSCORRUPTED; > > + > > I take it these are checks for whether the agfl was previously corrupted > based on the unpacked size? FWIW, it might be a bit more clear to use an > old/unpacked_agfl_size variable to declare intent (here and in some of > the comments and whatnot that follow). Ok. > > + /* > > + * Convert a 40-byte-padded agfl into a 36-byte-padded AGFL. > > + * Therefore, we need to move the AGFL blocks > > + * bno[flfirst..agfl_size - 2] to bno[flfirst + 1...agfl_size - 1]. > > + * > > + * Reusing the example above, if we had flfirst == 116, we need > > + * to move bno[116] and bno[117] into bno[117] and bno[118], > > + * respectively, and then increment flfirst. > > + */ > > Kind of a strange example to use given that it doesn't mention > wrapping.. this only triggers if the agfl was previously wrapped, right? Right. I used to have a check that if the list didn't wrap then we would just exit because no fixing is required... but it must've fallen out. I agree that we don't want to rely on a subtlety here to stay out of here if the agfl wasn't initially wrapped. > > + error = xfs_alloc_read_agfl(mp, tp, agno, &agflbp); > > + if (error) > > + return error; > > + agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); > > + > > + len = (agfl_size - flfirst - 1) * sizeof(xfs_agblock_t); > > + memmove(&agfl_bno[flfirst + 1], &agfl_bno[flfirst], len); > > + offset = (char *)&agfl_bno[flfirst + 1] - (char *)agflbp->b_addr; > > + be32_add_cpu(&agf->agf_flfirst, 1); > > + > > + xfs_trans_buf_set_type(tp, agflbp, XFS_BLFT_AGFL_BUF); > > + xfs_trans_log_buf(tp, agflbp, offset, offset + len - 1); > > + xfs_trans_brelse(tp, agflbp); > > + agflbp = NULL; > > + xfs_alloc_log_agf(tp, agfbp, XFS_AGF_FLFIRST); > > + > > + return 0; > > +} > > + > > +/* > > + * Fix an AGFL that touches the end of the block by moving the first or last > > + * part of the list elsewhere in the AGFL so that old kernels don't trip over > > + * wrapping issues. > > + */ > > +STATIC int > > +xfs_fixup_freelist_wrap_unmount( > > + struct xfs_trans *tp, > > + struct xfs_buf *agfbp, > > + struct xfs_perag *pag) > > +{ > > + struct xfs_mount *mp = tp->t_mountp; > > + struct xfs_agf *agf; > > + struct xfs_buf *agflbp; > > + __be32 *agfl_bno; > > + xfs_agnumber_t agno; > > + uint32_t agfl_size; > > + uint32_t flfirst; > > + uint32_t fllast; > > + int offset; > > + int len; > > + int error; > > + > > + agfl_size = xfs_agfl_size(mp); > > + agf = XFS_BUF_TO_AGF(agfbp); > > + agno = be32_to_cpu(agf->agf_seqno); > > + flfirst = be32_to_cpu(agf->agf_flfirst); > > + fllast = be32_to_cpu(agf->agf_fllast); > > + > > + /* Empty AGFL? Make sure we aren't pointing at the end. */ > > + if (pag->pagf_flcount == 0) { > > + if (flfirst >= agfl_size || fllast >= agfl_size) { > > When would either field be >= agfl_size? Isn't that where they wrap? Oops, you're right, that should be flfirst >= agfl_size - 1. Or really, old_agfl_size = agfl_size - 1 if flfirst >= old_agfl_size or fllast >= old_agfl_size: blahblahblah > > > + agf->agf_flfirst = cpu_to_be32(1); > > + agf->agf_fllast = 0; > > + xfs_alloc_log_agf(tp, agfbp, > > + XFS_AGF_FLFIRST | XFS_AGF_FLLAST); > > + } > > + return 0; > > + } > > + > > + /* If we don't hit the end, we're done. */ > > + if (flfirst < fllast && fllast != agfl_size - 1) > > + return 0; > > + > > + /* > > + * Move a start of a wrapped list towards the start of the agfl block. > > FWIW, this and the subsequent comments kind of gave me the impression > that the agfl would be "shifted" towards the start of the block. The > code doesn't do that, but rather rotates the start/head of the agfl to > the tail and adjusts the pointers (changing the effective order). That > seems technically Ok, but I had to grok the code and read back to > understand the comment rather than the other way around. Yeah. Sorry about the confusion, maybe it would be better if I took out these confusing comments and replaced it all with the ascii art above? > > + * Therefore, we need to move the AGFL blocks > > + * bno[flfirst..agfl_size - 1] to bno[fllast + 1...agfl_size - flfirst]. > > + * Then we reset flfirst and fllast appropriately. > > + * > > + * Reusing the example above, if we had flfirst == 117 and fllast == 4, > > + * we need to move bno[117] and bno[118] into bno[5] and bno[6], > > + * respectively, and then reset flfirst and fllast. > > + * > > + * If it's just the last block that touches the end, only move that. > > + */ > > + error = xfs_alloc_read_agfl(mp, tp, agno, &agflbp); > > + if (error) > > + return error; > > + agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); > > + > > + if (fllast == agfl_size - 1) { > > + /* Back the AGFL off from the end of the block. */ > > + len = sizeof(xfs_agblock_t); > > + agfl_bno[flfirst - 1] = agfl_bno[agfl_size - 1]; > > What if the agfl is full (flfirst == 0)? I don't think we ever unmount with a full agfl -- we're guaranteed at least 118 items in the AGFL, and the maximum size the agfl will ever need is 8 blocks for each of bnobt, cntbt, rmapbt. However, we could just detect a full agfl and free the end block. > > + offset = (char *)&agfl_bno[flfirst - 1] - (char *)agflbp->b_addr; > > + be32_add_cpu(&agf->agf_fllast, -1); > > + be32_add_cpu(&agf->agf_flfirst, -1); > > + } else { This needs to check for flfirst > fllast. > > + /* Move the first part of the AGFL towards the front. */ > > + len = (agfl_size - flfirst) * sizeof(xfs_agblock_t); > > + memcpy(&agfl_bno[fllast + 1], &agfl_bno[flfirst], len); > > + offset = (char *)&agfl_bno[fllast + 1] - (char *)agflbp->b_addr; > > + agf->agf_flfirst = 0; > > + agf->agf_fllast = cpu_to_be32(pag->pagf_flcount - 1); > > Similar question here as above.. it looks like the copy may be a no-op, > but we still reset flfirst/fllast (which may also be fine, but > unnecessary). > > The more interesting case here is flcount == 1 and flfirst == fllast, > where it looks like we'd wrongly reset first/last (and perhaps copy more > than we should..?). > > Ugh. Remind me again why we can't just detect a mismatch, fail the mount > and tell the user to repair? :P Not all the distros included xfs_repair in the initrd or the logic to call it (as opposed to "fsck -y") if the root fs fails to mount, which means that if this wrapping problem hits a root filesystem then the administrator will have to reboot with a recovery iso and run xfs_repair from there. That's going to generate a /lot/ of support work compared to us figuring out how to repair the filesystems automatically. > Random thought... have we considered anything less invasive/more > simple for this issue than low level buffer manipulation? For example, > could we address the unpacked -> packed case by just allocating a > block to fill the gap at the end of the wrapped agfl? The allocation could fail, in which case we'd be forced to resort to list manipulation. > Perhaps that could even be done more discreetly in > xfs_alloc_fix_freelist() when we already have a transaction set up, > etc. Possible, though this will add more logic to every hot allocation call for something that only needs fixing at mount and at unmount. I think there's a stronger argument for doing the unmount cleanup on every fix_freelist since that would eliminate the small chance that a new kernel crashes and gets recovered on an old kernel. > That wouldn't help going the other direction, but maybe we could > consider similar logic to free an extraneous block and then worry > about protecting users who won't upgrade broken kernels separately if > that really continues to be a problem. It is already a problem here. --D > Brian > > > + } > > + > > + xfs_trans_buf_set_type(tp, agflbp, XFS_BLFT_AGFL_BUF); > > + xfs_trans_log_buf(tp, agflbp, offset, offset + len - 1); > > + xfs_trans_brelse(tp, agflbp); > > + agflbp = NULL; > > + xfs_alloc_log_agf(tp, agfbp, XFS_AGF_FLFIRST | XFS_AGF_FLLAST); > > + > > + return 0; > > +} > > + > > +typedef int (*xfs_agf_apply_fn_t)(struct xfs_trans *tp, struct xfs_buf *agfbp, > > + struct xfs_perag *pag); > > + > > +/* Apply something to every AGF. */ > > +STATIC int > > +xfs_fixup_agf_apply( > > + struct xfs_mount *mp, > > + xfs_agf_apply_fn_t fn) > > +{ > > + struct xfs_trans *tp; > > + struct xfs_perag *pag; > > + struct xfs_buf *agfbp; > > + xfs_agnumber_t agno; > > + int error; > > + > > + error = xfs_trans_alloc(mp, &M_RES(mp)->tr_growdata, 0, 0, 0, &tp); > > + if (error) > > + return error; > > + > > + for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) { > > + error = xfs_alloc_read_agf(mp, tp, agno, 0, &agfbp); > > + if (error) > > + goto cancel; > > + if (!agfbp) { > > + error = -ENOMEM; > > + goto cancel; > > + } > > + pag = xfs_perag_get(mp, agno); > > + error = fn(tp, agfbp, pag); > > + xfs_perag_put(pag); > > + xfs_trans_brelse(tp, agfbp); > > + if (error) > > + goto cancel; > > + } > > + > > + return xfs_trans_commit(tp); > > +cancel: > > + xfs_trans_cancel(tp); > > + return error; > > +} > > + > > +/* Fix AGFL wrapping so we can use the filesystem. */ > > +int > > +xfs_fixup_agfl_wrap_mount( > > + struct xfs_mount *mp) > > +{ > > + if (!xfs_sb_version_needs_agfl_wrap_fixes(&mp->m_sb)) > > + return 0; > > + > > + return xfs_fixup_agf_apply(mp, xfs_fixup_freelist_wrap_mount); > > +} > > + > > +/* Fix AGFL wrapping so old kernels can use this filesystem. */ > > +int > > +xfs_fixup_agfl_wrap_unmount( > > + struct xfs_mount *mp) > > +{ > > + if (!xfs_sb_version_needs_agfl_wrap_fixes(&mp->m_sb)) > > + return 0; > > + > > + return xfs_fixup_agf_apply(mp, xfs_fixup_freelist_wrap_unmount); > > +} > > diff --git a/fs/xfs/xfs_fixups.h b/fs/xfs/xfs_fixups.h > > new file mode 100644 > > index 0000000..fb52a96 > > --- /dev/null > > +++ b/fs/xfs/xfs_fixups.h > > @@ -0,0 +1,26 @@ > > +/* > > + * Copyright (C) 2018 Oracle. All Rights Reserved. > > + * > > + * Author: Darrick J. Wong <darrick.wong@oracle.com> > > + * > > + * This program is free software; you can redistribute it and/or > > + * modify it under the terms of the GNU General Public License > > + * as published by the Free Software Foundation; either version 2 > > + * of the License, or (at your option) any later version. > > + * > > + * This program is distributed in the hope that it would be useful, > > + * but WITHOUT ANY WARRANTY; without even the implied warranty of > > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the > > + * GNU General Public License for more details. > > + * > > + * You should have received a copy of the GNU General Public License > > + * along with this program; if not, write the Free Software Foundation, > > + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. > > + */ > > +#ifndef __XFS_FIXUPS_H__ > > +#define __XFS_FIXUPS_H__ > > + > > +int xfs_fixup_agfl_wrap_mount(struct xfs_mount *mp); > > +int xfs_fixup_agfl_wrap_unmount(struct xfs_mount *mp); > > + > > +#endif /* __XFS_FIXUPS_H__ */ > > diff --git a/fs/xfs/xfs_mount.c b/fs/xfs/xfs_mount.c > > index 98fd41c..eb284aa 100644 > > --- a/fs/xfs/xfs_mount.c > > +++ b/fs/xfs/xfs_mount.c > > @@ -46,7 +46,7 @@ > > #include "xfs_refcount_btree.h" > > #include "xfs_reflink.h" > > #include "xfs_extent_busy.h" > > - > > +#include "xfs_fixups.h" > > > > static DEFINE_MUTEX(xfs_uuid_table_mutex); > > static int xfs_uuid_table_size; > > @@ -875,6 +875,16 @@ xfs_mountfs( > > } > > > > /* > > + * Make sure our AGFL counters do not wrap the end of the block > > + * in a troublesome manner. > > + */ > > + error = xfs_fixup_agfl_wrap_mount(mp); > > + if (error) { > > + xfs_warn(mp, "Failed to fix agfl wrapping. Run xfs_repair."); > > + goto out_log_dealloc; > > + } > > + > > + /* > > * Get and sanity-check the root inode. > > * Save the pointer to it in the mount structure. > > */ > > @@ -1128,6 +1138,15 @@ xfs_unmountfs( > > xfs_qm_unmount(mp); > > > > /* > > + * Make sure our AGFL counters do not wrap the end of the block > > + * in a troublesome manner for old kernels. > > + */ > > + error = xfs_fixup_agfl_wrap_unmount(mp); > > + if (error) > > + xfs_warn(mp, "Unable to fix agfl wrapping. " > > + "This may cause problems on next mount."); > > + > > + /* > > * Unreserve any blocks we have so that when we unmount we don't account > > * the reserved free space as used. This is really only necessary for > > * lazy superblock counting because it trusts the incore superblock > > diff --git a/fs/xfs/xfs_super.c b/fs/xfs/xfs_super.c > > index 624a802..d9aa39a 100644 > > --- a/fs/xfs/xfs_super.c > > +++ b/fs/xfs/xfs_super.c > > @@ -50,6 +50,7 @@ > > #include "xfs_refcount_item.h" > > #include "xfs_bmap_item.h" > > #include "xfs_reflink.h" > > +#include "xfs_fixups.h" > > > > #include <linux/namei.h> > > #include <linux/dax.h> > > @@ -1206,6 +1207,15 @@ xfs_quiesce_attr( > > xfs_reclaim_inodes(mp, 0); > > xfs_reclaim_inodes(mp, SYNC_WAIT); > > > > + /* > > + * Make sure our AGFL counters do not wrap the end of the block > > + * in a troublesome manner for old kernels. > > + */ > > + error = xfs_fixup_agfl_wrap_unmount(mp); > > + if (error) > > + xfs_warn(mp, "Unable to fix agfl wrapping. " > > + "This may cause problems on next mount."); > > + > > /* Push the superblock and write an unmount record */ > > error = xfs_log_sbcount(mp); > > if (error) > > > > -- > > To unsubscribe from this list: send the line "unsubscribe linux-xfs" in > > the body of a message to majordomo@vger.kernel.org > > More majordomo info at http://vger.kernel.org/majordomo-info.html > -- > To unsubscribe from this list: send the line "unsubscribe linux-xfs" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Tue, Feb 27, 2018 at 01:03:13PM -0800, Darrick J. Wong wrote: > On Tue, Feb 27, 2018 at 02:35:49PM -0500, Brian Foster wrote: > > On Thu, Feb 22, 2018 at 06:00:15PM -0800, Darrick J. Wong wrote: > > > From: Darrick J. Wong <darrick.wong@oracle.com> > > > > > > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> > > > --- > > > fs/xfs/Makefile | 1 > > > fs/xfs/xfs_fixups.c | 310 +++++++++++++++++++++++++++++++++++++++++++++++++++ > > > fs/xfs/xfs_fixups.h | 26 ++++ > > > fs/xfs/xfs_mount.c | 21 +++ > > > fs/xfs/xfs_super.c | 10 ++ > > > 5 files changed, 367 insertions(+), 1 deletion(-) > > > create mode 100644 fs/xfs/xfs_fixups.c > > > create mode 100644 fs/xfs/xfs_fixups.h > > > > > > > > > diff --git a/fs/xfs/Makefile b/fs/xfs/Makefile > > > index b03c77e..f88368a 100644 > > > --- a/fs/xfs/Makefile > > > +++ b/fs/xfs/Makefile > > > @@ -86,6 +86,7 @@ xfs-y += xfs_aops.o \ > > > xfs_extent_busy.o \ > > > xfs_file.o \ > > > xfs_filestream.o \ > > > + xfs_fixups.o \ > > > xfs_fsmap.o \ > > > xfs_fsops.o \ > > > xfs_globals.o \ > > > diff --git a/fs/xfs/xfs_fixups.c b/fs/xfs/xfs_fixups.c > > > new file mode 100644 > > > index 0000000..0cad7bb > > > --- /dev/null > > > +++ b/fs/xfs/xfs_fixups.c > > > @@ -0,0 +1,310 @@ > > > +/* > > > + * Copyright (C) 2018 Oracle. All Rights Reserved. > > > + * > > > + * Author: Darrick J. Wong <darrick.wong@oracle.com> > > > + * > > > + * This program is free software; you can redistribute it and/or > > > + * modify it under the terms of the GNU General Public License > > > + * as published by the Free Software Foundation; either version 2 > > > + * of the License, or (at your option) any later version. > > > + * > > > + * This program is distributed in the hope that it would be useful, > > > + * but WITHOUT ANY WARRANTY; without even the implied warranty of > > > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the > > > + * GNU General Public License for more details. > > > + * > > > + * You should have received a copy of the GNU General Public License > > > + * along with this program; if not, write the Free Software Foundation, > > > + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. > > > + */ > > > +#include "xfs.h" > > > +#include "xfs_fs.h" > > > +#include "xfs_shared.h" > > > +#include "xfs_format.h" > > > +#include "xfs_log_format.h" > > > +#include "xfs_trans_resv.h" > > > +#include "xfs_sb.h" > > > +#include "xfs_mount.h" > > > +#include "xfs_alloc.h" > > > +#include "xfs_trans.h" > > > +#include "xfs_fixups.h" > > > + > > > +/* > > > + * v5 AGFL padding defects > > > + * > > > + * When the v5 format was first introduced, there was a defect in the struct > > > + * xfs_agfl definition that resulted in XFS_AGFL_SIZE returning different > > > > XFS_AGFL_SIZE() no longer exists as of the previous patch. It might be > > better to size refer to "the agfl size" generically. > > How about: > > "...the agfl length calculation formula (at the time known as > XFS_AGFL_SIZE)..." > The formula formerly known as..? ;) Heh, sounds fine. > > > + * values depending on the compiler padding. On a fs with 512-byte sectors, > > > + * this meant that XFS_AGFL_SIZE was 119 on i386, but 118 on x64. Commit > > > + * 96f859d52bcb1 ("libxfs: pack the agfl header structure so XFS_AGFL_SIZE is > > > + * correct") changed the definition to disable padding the end of the > > > + * structure, and was accepted into Linux 4.5. Since then, the AGFL has > > > + * always used the larger size (e.g. 119 entries on a 512b sector fs). > > > + * > > > + * Unfortunately, pre-4.5 kernels can produce filesystems with AGFLs that wrap > > > + * at the smaller size, and those kernels are not prepared to handle the > > > + * longer size. This typically manifests itself as an AGF verifier corruption > > > + * error followed by a filesystem shutdown. While we encourage admins to stay > > > + * current with software, we would like to avoid this intermittent breakage. > > > + * > > > + * Any v5 filesystem which has a feature bit set for a feature that was > > > + * introduced after Linux 4.5 will not have this problem, as such kernels > > > + * cannot be mounted on older kernels. v4 filesystems are also unaffected. > > > + * > > > + * Therefore, we add two fixup functions -- the first runs at mount time to > > > + * detect a short-wrapped AGFL and fix it; the second runs at unmount, freeze, > > > + * or remount-ro time to move a wrapped AGFL to the beginning of the list. > > > + * This reduces the likelihood of a screwup to the scenario where you have (a) > > > + * a filesystem with no post-4.5 features (reflink, rmap), (b) the AGFL wraps, > > > + * (c) the filesystem goes down leaving a dirty log, and (d) the dirty > > > + * filesystem is mounted on an old kernel. > > > + */ > > > + > > > > While the mount vs. unmount time fixups serve different purposes and > > have different conditions, do we really need two separate fixup > > implementations? E.g., if we have to unwrap the AGFL at unmount time, > > why not just reuse that fixup at mount time (when needed) as well? I > > suspect the difference would only be the length of what we can consider > > valid in the agfl at the time. > > They're doing two different things. I will use ASCII art to demonstrate > the mount-time function: > Yes, the question was whether we need to do two different things. For example... > Let's say the AGFL block has space for 10 items. It looks like this: > > 0 1 2 3 4 5 6 7 8 9 > Header | | | | | | | | | | | > > > When the fs is freshly formatted or repaired, the AGFL will look like: > > 0 1 2 3 4 5 6 7 8 9 > Header | A | B | C | D | E | F | | | | | > ^-- flfirst ^-- fllast > > Due to the padding problems prior to 4.5 ("4.5 agfl padding fix"), > XFS_AGFL_SIZE would return 10 on 32-bit systems and 9 on 64-bit systems. > Therefore, if the AGFL wrapped on a 64-bit kernel we would end up with: > > 0 1 2 3 4 5 6 7 8 9 > Header | D | E | F | | | | A | B | C | ? | > ^-- fllast ^-- flfirst > > Note that block 9's contents are undefined because we didn't write > anything there. Now the "4.5 agfl padding fix" has corrected > XFS_AGFL_SIZE to return 10 in all cases, this looks like an AGF > corruption because flcount is 6 but the distance between flfirst and > fllast is 7. So long as the list wraps and the mismatch is exactly one > block we can fix this. The mount time fixer in this patch does this by > moving the records between flfirst and xfs_agfl_size (A, B, and C) to > the end of the block to close the gap: > > 0 1 2 3 4 5 6 7 8 9 > Header | D | E | F | | | | | A | B | C | > ^-- fllast ^-- flfirst > > If the count is off by more than 1 then the AGF is truly corrupt and we > bail out. > > Now let's say that we run for a while and want to unmount, but our AGFL > wraps like this: > > 0 1 2 3 4 5 6 7 8 9 > Header | T | U | V | | | | | Q | R | S | > ^-- fllast ^-- flfirst > > We don't know that the next kernel to mount this filesystem will have > the "4.5 agfl padding fix" applied to it; if it does not, it will flag > the AGF as corrupted because flcount is 6 but in its view the distance > between flfirst and fllast (which omits bno[9]) is 5. We don't want > it to choke on that, so the unmount fixer moves all the records between > flfirst and the end (Q, R, and S) towards the start of the block and > resets flfirst/fllast: > > 0 1 2 3 4 5 6 7 8 9 > Header | T | U | V | Q | R | S | | | | | > ^-- flfirst ^-- fllast > Could we use this basic algorithm at mount time as well? I know it wouldn't be _exactly_ the same operation at mount time as it is for unmount since slot 9 is a gap in the former case, but afaict the only difference from an algorithmic perspective is the length of the shift. IOW, if we were to parameterize the length of the shift and have the mount fixup call the unmount algorithm, would it not also address the problem? > Since the AGFL no longer wraps at all, it doesn't matter if the next > kernel to mount this filesystem has the "4.5 agfl padding fix" applied; > all kernels can handle this correctly. > > > > +/* > > > + * Decide if we need to have the agfl wrapping fixes applied. This only > > > + * affects v5 filesystems that do not have any features enabled that did not > > > + * exist when the agfl padding fix went in. > > > + * > > > + * Features already present when the fix went in were finobt, ftype, spinodes. > > > + * If we see something new (e.g. reflink) then don't bother. > > > + */ > > > +#define XFS_SB_FEAT_RO_COMPAT_AGFL_WRAP_ALREADY_FIXED \ > > > + (~(XFS_SB_FEAT_RO_COMPAT_FINOBT)) > > > +#define XFS_SB_FEAT_INCOMPAT_AGFL_WRAP_ALREADY_FIXED \ > > > + (~(XFS_SB_FEAT_INCOMPAT_FTYPE | \ > > > + XFS_SB_FEAT_INCOMPAT_SPINODES)) > > > +#define XFS_SB_FEAT_INCOMPAT_LOG_AGFL_WRAP_ALREADY_FIXED \ > > > + (~0) > > > +static inline bool xfs_sb_version_needs_agfl_wrap_fixes(struct xfs_sb *sbp) > > > +{ > > > + return xfs_sb_version_hascrc(sbp) && > > > + !xfs_sb_has_incompat_feature(sbp, > > > + XFS_SB_FEAT_INCOMPAT_AGFL_WRAP_ALREADY_FIXED) && > > > + !xfs_sb_has_ro_compat_feature(sbp, > > > + XFS_SB_FEAT_RO_COMPAT_AGFL_WRAP_ALREADY_FIXED) && > > > + !xfs_sb_has_incompat_log_feature(sbp, > > > + XFS_SB_FEAT_INCOMPAT_LOG_AGFL_WRAP_ALREADY_FIXED); > > > +} > > > + > > > +/* > > > + * Fix an AGFL wrapping that falls short of the end of the block by filling the > > > + * gap at the end of the block. > > > + */ > > > +STATIC int > > > +xfs_fixup_freelist_wrap_mount( > > > + struct xfs_trans *tp, > > > + struct xfs_buf *agfbp, > > > + struct xfs_perag *pag) > > > +{ > > > + struct xfs_mount *mp = tp->t_mountp; > > > + struct xfs_agf *agf; > > > + struct xfs_buf *agflbp; > > > + __be32 *agfl_bno; > > > + xfs_agnumber_t agno; > > > + uint32_t agfl_size; > > > + uint32_t flfirst; > > > + uint32_t fllast; > > > + int32_t active; > > > + int offset; > > > + int len; > > > + int error; > > > + > > > + if (pag->pagf_flcount == 0) > > > + return 0; > > > + > > > + agfl_size = xfs_agfl_size(mp); > > > + agf = XFS_BUF_TO_AGF(agfbp); > > > + agno = be32_to_cpu(agf->agf_seqno); > > > + flfirst = be32_to_cpu(agf->agf_flfirst); > > > + fllast = be32_to_cpu(agf->agf_fllast); > > > + > > > + /* Make sure we're either spot on or off by 1. */ > > > + active = fllast - flfirst + 1; > > > + if (active <= 0) > > > + active += agfl_size; > > > + if (active == pag->pagf_flcount) > > > + return 0; > > > + else if (active != pag->pagf_flcount + 1) > > > + return -EFSCORRUPTED; > > > + > > > > So we're not attempting to cover the case where the agfl has 1 more > > block than the agfl size (i.e., the case where an fs goes back to a > > kernel with an unpacked header)? > > We don't know how the next kernel to touch this filesystem will define > XFS_AGFL_SIZE -- it could be a 4.5+ kernel (same agfl size), a 32-bit > pre-4.5 kernel (same agfl size), or a 64-bit pre-4.5 kernel (small agfl > size). > I don't think I was clear.. I'm envisioning whether we could come up with a patch that would generically fix up the agfl on disk to be sane relative to the current kernel. This patch covers the case of a packed agfl kernel mounting an unpacked on-disk agfl. It would be nice if we could implement something that also handled a packed on-disk agfl to an unpacked agfl kernel (for easier backport to unpacked kernels, for e.g.). > > I'm wondering whether using the unmount algorithm in both places (as > > noted above) would also facilitate this mechanism to work both ways > > (unpacked <-> packed), provided particular conditions are met. I suppose > > that could be considered regardless, but the less variance the better > > IMO. > > In principle I suppose we could have a single function to handle all the > rearranging, but I'd rather have two functions to handle the two cases. > I'm open to refactoring the common parts out of the > xfs_fixup_freelist_wrap_*mount functions. > As noted above, I'm not really commenting on the function factoring here as much as the use of multiple algorithms (using one would certainly facilitate some mechanical refactoring). > > > + /* Would this have even passed muster on an old system? */ > > > > Comment doesn't really explain what is going on here..? > > /* > * If the distance between flfirst and fllast mismatches flcount by more > * than 1, then there's more wrong with this agfl than just the padding > * problem. Bail out completely, which will force the admin to run > * xfs_repair. > */ > *nod* > > > > > + if (flfirst >= agfl_size - 1 || fllast >= agfl_size - 1 || > > > + pag->pagf_flcount > agfl_size - 1) > > > + return -EFSCORRUPTED; > > > + > > > > I take it these are checks for whether the agfl was previously corrupted > > based on the unpacked size? FWIW, it might be a bit more clear to use an > > old/unpacked_agfl_size variable to declare intent (here and in some of > > the comments and whatnot that follow). > > Ok. > > > > + /* > > > + * Convert a 40-byte-padded agfl into a 36-byte-padded AGFL. > > > + * Therefore, we need to move the AGFL blocks > > > + * bno[flfirst..agfl_size - 2] to bno[flfirst + 1...agfl_size - 1]. > > > + * > > > + * Reusing the example above, if we had flfirst == 116, we need > > > + * to move bno[116] and bno[117] into bno[117] and bno[118], > > > + * respectively, and then increment flfirst. > > > + */ > > > > Kind of a strange example to use given that it doesn't mention > > wrapping.. this only triggers if the agfl was previously wrapped, right? > > Right. I used to have a check that if the list didn't wrap then we > would just exit because no fixing is required... but it must've fallen > out. I agree that we don't want to rely on a subtlety here to stay out > of here if the agfl wasn't initially wrapped. > > > > + error = xfs_alloc_read_agfl(mp, tp, agno, &agflbp); > > > + if (error) > > > + return error; > > > + agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); > > > + > > > + len = (agfl_size - flfirst - 1) * sizeof(xfs_agblock_t); > > > + memmove(&agfl_bno[flfirst + 1], &agfl_bno[flfirst], len); > > > + offset = (char *)&agfl_bno[flfirst + 1] - (char *)agflbp->b_addr; > > > + be32_add_cpu(&agf->agf_flfirst, 1); > > > + > > > + xfs_trans_buf_set_type(tp, agflbp, XFS_BLFT_AGFL_BUF); > > > + xfs_trans_log_buf(tp, agflbp, offset, offset + len - 1); > > > + xfs_trans_brelse(tp, agflbp); > > > + agflbp = NULL; > > > + xfs_alloc_log_agf(tp, agfbp, XFS_AGF_FLFIRST); > > > + > > > + return 0; > > > +} > > > + > > > +/* > > > + * Fix an AGFL that touches the end of the block by moving the first or last > > > + * part of the list elsewhere in the AGFL so that old kernels don't trip over > > > + * wrapping issues. > > > + */ > > > +STATIC int > > > +xfs_fixup_freelist_wrap_unmount( > > > + struct xfs_trans *tp, > > > + struct xfs_buf *agfbp, > > > + struct xfs_perag *pag) > > > +{ > > > + struct xfs_mount *mp = tp->t_mountp; > > > + struct xfs_agf *agf; > > > + struct xfs_buf *agflbp; > > > + __be32 *agfl_bno; > > > + xfs_agnumber_t agno; > > > + uint32_t agfl_size; > > > + uint32_t flfirst; > > > + uint32_t fllast; > > > + int offset; > > > + int len; > > > + int error; > > > + > > > + agfl_size = xfs_agfl_size(mp); > > > + agf = XFS_BUF_TO_AGF(agfbp); > > > + agno = be32_to_cpu(agf->agf_seqno); > > > + flfirst = be32_to_cpu(agf->agf_flfirst); > > > + fllast = be32_to_cpu(agf->agf_fllast); > > > + > > > + /* Empty AGFL? Make sure we aren't pointing at the end. */ > > > + if (pag->pagf_flcount == 0) { > > > + if (flfirst >= agfl_size || fllast >= agfl_size) { > > > > When would either field be >= agfl_size? Isn't that where they wrap? > > Oops, you're right, that should be flfirst >= agfl_size - 1. > > Or really, > > old_agfl_size = agfl_size - 1 > if flfirst >= old_agfl_size or fllast >= old_agfl_size: > blahblahblah > > > > > > + agf->agf_flfirst = cpu_to_be32(1); > > > + agf->agf_fllast = 0; > > > + xfs_alloc_log_agf(tp, agfbp, > > > + XFS_AGF_FLFIRST | XFS_AGF_FLLAST); > > > + } > > > + return 0; > > > + } > > > + > > > + /* If we don't hit the end, we're done. */ > > > + if (flfirst < fllast && fllast != agfl_size - 1) > > > + return 0; > > > + > > > + /* > > > + * Move a start of a wrapped list towards the start of the agfl block. > > > > FWIW, this and the subsequent comments kind of gave me the impression > > that the agfl would be "shifted" towards the start of the block. The > > code doesn't do that, but rather rotates the start/head of the agfl to > > the tail and adjusts the pointers (changing the effective order). That > > seems technically Ok, but I had to grok the code and read back to > > understand the comment rather than the other way around. > > Yeah. Sorry about the confusion, maybe it would be better if I took out > these confusing comments and replaced it all with the ascii art above? > I'm still hoping we can do something a bit more simple and flexible in general, but that aside, the ascii art describes the associated algorithms much better. > > > + * Therefore, we need to move the AGFL blocks > > > + * bno[flfirst..agfl_size - 1] to bno[fllast + 1...agfl_size - flfirst]. > > > + * Then we reset flfirst and fllast appropriately. > > > + * > > > + * Reusing the example above, if we had flfirst == 117 and fllast == 4, > > > + * we need to move bno[117] and bno[118] into bno[5] and bno[6], > > > + * respectively, and then reset flfirst and fllast. > > > + * > > > + * If it's just the last block that touches the end, only move that. > > > + */ > > > + error = xfs_alloc_read_agfl(mp, tp, agno, &agflbp); > > > + if (error) > > > + return error; > > > + agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); > > > + > > > + if (fllast == agfl_size - 1) { > > > + /* Back the AGFL off from the end of the block. */ > > > + len = sizeof(xfs_agblock_t); > > > + agfl_bno[flfirst - 1] = agfl_bno[agfl_size - 1]; > > > > What if the agfl is full (flfirst == 0)? > > I don't think we ever unmount with a full agfl -- we're guaranteed at > least 118 items in the AGFL, and the maximum size the agfl will ever > need is 8 blocks for each of bnobt, cntbt, rmapbt. > > However, we could just detect a full agfl and free the end block. > I suppose it's unlikely we'll ever see corner cases such as a completely full or empty agfl. That said, I think the code needs to handle those cases sanely. That could mean explicit checks and skipping the fixup for all I care, just as long as we don't have memory corruption vectors and whatnot if (when?) assumptions prove wrong. > > > + offset = (char *)&agfl_bno[flfirst - 1] - (char *)agflbp->b_addr; > > > + be32_add_cpu(&agf->agf_fllast, -1); > > > + be32_add_cpu(&agf->agf_flfirst, -1); > > > + } else { > > This needs to check for flfirst > fllast. > > > > + /* Move the first part of the AGFL towards the front. */ > > > + len = (agfl_size - flfirst) * sizeof(xfs_agblock_t); > > > + memcpy(&agfl_bno[fllast + 1], &agfl_bno[flfirst], len); > > > + offset = (char *)&agfl_bno[fllast + 1] - (char *)agflbp->b_addr; > > > + agf->agf_flfirst = 0; > > > + agf->agf_fllast = cpu_to_be32(pag->pagf_flcount - 1); > > > > Similar question here as above.. it looks like the copy may be a no-op, > > but we still reset flfirst/fllast (which may also be fine, but > > unnecessary). > > > > The more interesting case here is flcount == 1 and flfirst == fllast, > > where it looks like we'd wrongly reset first/last (and perhaps copy more > > than we should..?). > > > > Ugh. Remind me again why we can't just detect a mismatch, fail the mount > > and tell the user to repair? :P > > Not all the distros included xfs_repair in the initrd or the logic to > call it (as opposed to "fsck -y") if the root fs fails to mount, which > means that if this wrapping problem hits a root filesystem then the > administrator will have to reboot with a recovery iso and run xfs_repair > from there. That's going to generate a /lot/ of support work compared > to us figuring out how to repair the filesystems automatically. > Yeah, sorry. Good argument for the unpacked -> packed fixup. > > Random thought... have we considered anything less invasive/more > > simple for this issue than low level buffer manipulation? For example, > > could we address the unpacked -> packed case by just allocating a > > block to fill the gap at the end of the wrapped agfl? > > The allocation could fail, in which case we'd be forced to resort to > list manipulation. > Yeah.. obvious tradeoff, but how likely is that? It seems a reasonable tradeoff to me given there's still the repair path. Also note that we don't have to necessarily allocate from the btrees. (FWIW, I hacked around on this a bit and ended up with something slightly different[1].) > > Perhaps that could even be done more discreetly in > > xfs_alloc_fix_freelist() when we already have a transaction set up, > > etc. > > Possible, though this will add more logic to every hot allocation call > for something that only needs fixing at mount and at unmount. I think > there's a stronger argument for doing the unmount cleanup on every > fix_freelist since that would eliminate the small chance that a new > kernel crashes and gets recovered on an old kernel. > Eh, I don't think that's anything we couldn't address if need be. E.g., amortize the padding fixup against actual allocations, set/test a pag bit to track the padding fixup, or check it on agf read and set a "needs fixup" bit, etc. > > That wouldn't help going the other direction, but maybe we could > > consider similar logic to free an extraneous block and then worry > > about protecting users who won't upgrade broken kernels separately if > > that really continues to be a problem. > > It is already a problem here. > Then it should be able to stand alone as an independent patch (with independent commit log that explains/justifies it, etc.). Brian [1] This is hacky and brute force, but the idea was to try and do something that could work in either direction and reuse more existing code. It basically just reads the on-disk entries "manually" and reinserts them into a reinitialized (empty) agfl. Perhaps it could be refactored into something that only rotates a given number of entries or something of that nature. Anyways, just an experiment and probably broken.. int xfs_agfl_fixup( struct xfs_trans *tp, struct xfs_buf *agbp, struct xfs_perag *pag) { struct xfs_mount *mp = tp->t_mountp; struct xfs_agf *agf = XFS_BUF_TO_AGF(agbp); int agfl_size = XFS_AGFL_SIZE(mp); int ofirst, olast, osize; int nfirst, nlast; struct xfs_buf *agflbp; __be32 *agfl_bno; int active; xfs_agblock_t bno; bool wasempty = false; if (pag->pagf_flcount == 0) wasempty = true; ofirst = be32_to_cpu(agf->agf_flfirst); olast = be32_to_cpu(agf->agf_fllast); if (olast >= ofirst) active = olast - ofirst + 1; else active = agfl_size - ofirst + olast + 1; if (active == pag->pagf_flcount + 1) osize = agfl_size - 1; else if ((active == pag->pagf_flcount - 1) || ofirst == agfl_size || olast == agfl_size) osize = agfl_size + 1; else if (active == pag->pagf_flcount) osize = agfl_size; else return -EFSCORRUPTED; xfs_alloc_read_agfl(mp, tp, be32_to_cpu(agf->agf_seqno), &agflbp); agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); nlast = do_mod(olast, agfl_size); nfirst = nlast + 1; nfirst = do_mod(nfirst, agfl_size); ASSERT(nfirst != ofirst); agf->agf_flfirst = cpu_to_be32(nfirst); agf->agf_fllast = cpu_to_be32(nlast); agf->agf_flcount = 0; xfs_alloc_log_agf(tp, agbp, XFS_AGF_FLFIRST | XFS_AGF_FLLAST | XFS_AGF_FLCOUNT); pag->pagf_flcount = 0; if (wasempty) goto out; while (true) { ofirst = do_mod(ofirst, osize); bno = be32_to_cpu(agfl_bno[ofirst]); xfs_alloc_put_freelist(tp, agbp, agflbp, bno, 0); if (ofirst == olast) break; ofirst++; } out: xfs_trans_brelse(tp, agflbp); return 0; } -- To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, Feb 28, 2018 at 05:43:51PM -0500, Brian Foster wrote: > On Tue, Feb 27, 2018 at 01:03:13PM -0800, Darrick J. Wong wrote: > > On Tue, Feb 27, 2018 at 02:35:49PM -0500, Brian Foster wrote: > > > On Thu, Feb 22, 2018 at 06:00:15PM -0800, Darrick J. Wong wrote: > > > > From: Darrick J. Wong <darrick.wong@oracle.com> > > > > > > > > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> > > > > --- > > > > fs/xfs/Makefile | 1 > > > > fs/xfs/xfs_fixups.c | 310 +++++++++++++++++++++++++++++++++++++++++++++++++++ > > > > fs/xfs/xfs_fixups.h | 26 ++++ > > > > fs/xfs/xfs_mount.c | 21 +++ > > > > fs/xfs/xfs_super.c | 10 ++ > > > > 5 files changed, 367 insertions(+), 1 deletion(-) > > > > create mode 100644 fs/xfs/xfs_fixups.c > > > > create mode 100644 fs/xfs/xfs_fixups.h > > > > > > > > > > > > diff --git a/fs/xfs/Makefile b/fs/xfs/Makefile > > > > index b03c77e..f88368a 100644 > > > > --- a/fs/xfs/Makefile > > > > +++ b/fs/xfs/Makefile > > > > @@ -86,6 +86,7 @@ xfs-y += xfs_aops.o \ > > > > xfs_extent_busy.o \ > > > > xfs_file.o \ > > > > xfs_filestream.o \ > > > > + xfs_fixups.o \ > > > > xfs_fsmap.o \ > > > > xfs_fsops.o \ > > > > xfs_globals.o \ > > > > diff --git a/fs/xfs/xfs_fixups.c b/fs/xfs/xfs_fixups.c > > > > new file mode 100644 > > > > index 0000000..0cad7bb > > > > --- /dev/null > > > > +++ b/fs/xfs/xfs_fixups.c > > > > @@ -0,0 +1,310 @@ > > > > +/* > > > > + * Copyright (C) 2018 Oracle. All Rights Reserved. > > > > + * > > > > + * Author: Darrick J. Wong <darrick.wong@oracle.com> > > > > + * > > > > + * This program is free software; you can redistribute it and/or > > > > + * modify it under the terms of the GNU General Public License > > > > + * as published by the Free Software Foundation; either version 2 > > > > + * of the License, or (at your option) any later version. > > > > + * > > > > + * This program is distributed in the hope that it would be useful, > > > > + * but WITHOUT ANY WARRANTY; without even the implied warranty of > > > > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the > > > > + * GNU General Public License for more details. > > > > + * > > > > + * You should have received a copy of the GNU General Public License > > > > + * along with this program; if not, write the Free Software Foundation, > > > > + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. > > > > + */ > > > > +#include "xfs.h" > > > > +#include "xfs_fs.h" > > > > +#include "xfs_shared.h" > > > > +#include "xfs_format.h" > > > > +#include "xfs_log_format.h" > > > > +#include "xfs_trans_resv.h" > > > > +#include "xfs_sb.h" > > > > +#include "xfs_mount.h" > > > > +#include "xfs_alloc.h" > > > > +#include "xfs_trans.h" > > > > +#include "xfs_fixups.h" > > > > + > > > > +/* > > > > + * v5 AGFL padding defects > > > > + * > > > > + * When the v5 format was first introduced, there was a defect in the struct > > > > + * xfs_agfl definition that resulted in XFS_AGFL_SIZE returning different > > > > > > XFS_AGFL_SIZE() no longer exists as of the previous patch. It might be > > > better to size refer to "the agfl size" generically. > > > > How about: > > > > "...the agfl length calculation formula (at the time known as > > XFS_AGFL_SIZE)..." > > > > The formula formerly known as..? ;) Heh, sounds fine. lol. :) > > > > + * values depending on the compiler padding. On a fs with 512-byte sectors, > > > > + * this meant that XFS_AGFL_SIZE was 119 on i386, but 118 on x64. Commit > > > > + * 96f859d52bcb1 ("libxfs: pack the agfl header structure so XFS_AGFL_SIZE is > > > > + * correct") changed the definition to disable padding the end of the > > > > + * structure, and was accepted into Linux 4.5. Since then, the AGFL has > > > > + * always used the larger size (e.g. 119 entries on a 512b sector fs). > > > > + * > > > > + * Unfortunately, pre-4.5 kernels can produce filesystems with AGFLs that wrap > > > > + * at the smaller size, and those kernels are not prepared to handle the > > > > + * longer size. This typically manifests itself as an AGF verifier corruption > > > > + * error followed by a filesystem shutdown. While we encourage admins to stay > > > > + * current with software, we would like to avoid this intermittent breakage. > > > > + * > > > > + * Any v5 filesystem which has a feature bit set for a feature that was > > > > + * introduced after Linux 4.5 will not have this problem, as such kernels > > > > + * cannot be mounted on older kernels. v4 filesystems are also unaffected. > > > > + * > > > > + * Therefore, we add two fixup functions -- the first runs at mount time to > > > > + * detect a short-wrapped AGFL and fix it; the second runs at unmount, freeze, > > > > + * or remount-ro time to move a wrapped AGFL to the beginning of the list. > > > > + * This reduces the likelihood of a screwup to the scenario where you have (a) > > > > + * a filesystem with no post-4.5 features (reflink, rmap), (b) the AGFL wraps, > > > > + * (c) the filesystem goes down leaving a dirty log, and (d) the dirty > > > > + * filesystem is mounted on an old kernel. > > > > + */ > > > > + > > > > > > While the mount vs. unmount time fixups serve different purposes and > > > have different conditions, do we really need two separate fixup > > > implementations? E.g., if we have to unwrap the AGFL at unmount time, > > > why not just reuse that fixup at mount time (when needed) as well? I > > > suspect the difference would only be the length of what we can consider > > > valid in the agfl at the time. > > > > They're doing two different things. I will use ASCII art to demonstrate > > the mount-time function: > > > > Yes, the question was whether we need to do two different things. For > example... > > > Let's say the AGFL block has space for 10 items. It looks like this: > > > > 0 1 2 3 4 5 6 7 8 9 > > Header | | | | | | | | | | | > > > > > > When the fs is freshly formatted or repaired, the AGFL will look like: > > > > 0 1 2 3 4 5 6 7 8 9 > > Header | A | B | C | D | E | F | | | | | > > ^-- flfirst ^-- fllast > > > > Due to the padding problems prior to 4.5 ("4.5 agfl padding fix"), > > XFS_AGFL_SIZE would return 10 on 32-bit systems and 9 on 64-bit systems. > > Therefore, if the AGFL wrapped on a 64-bit kernel we would end up with: > > > > 0 1 2 3 4 5 6 7 8 9 > > Header | D | E | F | | | | A | B | C | ? | > > ^-- fllast ^-- flfirst > > > > Note that block 9's contents are undefined because we didn't write > > anything there. Now the "4.5 agfl padding fix" has corrected > > XFS_AGFL_SIZE to return 10 in all cases, this looks like an AGF > > corruption because flcount is 6 but the distance between flfirst and > > fllast is 7. So long as the list wraps and the mismatch is exactly one > > block we can fix this. The mount time fixer in this patch does this by > > moving the records between flfirst and xfs_agfl_size (A, B, and C) to > > the end of the block to close the gap: > > > > 0 1 2 3 4 5 6 7 8 9 > > Header | D | E | F | | | | | A | B | C | > > ^-- fllast ^-- flfirst > > > > If the count is off by more than 1 then the AGF is truly corrupt and we > > bail out. > > > > Now let's say that we run for a while and want to unmount, but our AGFL > > wraps like this: > > > > 0 1 2 3 4 5 6 7 8 9 > > Header | T | U | V | | | | | Q | R | S | > > ^-- fllast ^-- flfirst > > > > We don't know that the next kernel to mount this filesystem will have > > the "4.5 agfl padding fix" applied to it; if it does not, it will flag > > the AGF as corrupted because flcount is 6 but in its view the distance > > between flfirst and fllast (which omits bno[9]) is 5. We don't want > > it to choke on that, so the unmount fixer moves all the records between > > flfirst and the end (Q, R, and S) towards the start of the block and > > resets flfirst/fllast: > > > > 0 1 2 3 4 5 6 7 8 9 > > Header | T | U | V | Q | R | S | | | | | > > ^-- flfirst ^-- fllast > > > > Could we use this basic algorithm at mount time as well? I know it > wouldn't be _exactly_ the same operation at mount time as it is for > unmount since slot 9 is a gap in the former case, but afaict the only > difference from an algorithmic perspective is the length of the shift. > > IOW, if we were to parameterize the length of the shift and have the > mount fixup call the unmount algorithm, would it not also address the > problem? Yes, I believe that would work. I think it would be more efficient in the patch below to memmove the entries instead of put_freelist'ing them individually. If the agfl is completely full then fix_freelist ought to trim it down by at least one element. The algorithm then becomes: if mounting and flfirst < fllast: return 0 if flcount == agfl_size: assert !mounting fix_freelist() assert flcount < agfl_size if flfirst < fllast: return 0 movelen = agfl_size - flfirst if active == flcount - 1: movelen-- memmove(&agflbno[fllast + 1], &agflbno[flfirst], movelen) flfirst = 0 fllast = fllast + movelen > > Since the AGFL no longer wraps at all, it doesn't matter if the next > > kernel to mount this filesystem has the "4.5 agfl padding fix" applied; > > all kernels can handle this correctly. > > > > > > +/* > > > > + * Decide if we need to have the agfl wrapping fixes applied. This only > > > > + * affects v5 filesystems that do not have any features enabled that did not > > > > + * exist when the agfl padding fix went in. > > > > + * > > > > + * Features already present when the fix went in were finobt, ftype, spinodes. > > > > + * If we see something new (e.g. reflink) then don't bother. > > > > + */ > > > > +#define XFS_SB_FEAT_RO_COMPAT_AGFL_WRAP_ALREADY_FIXED \ > > > > + (~(XFS_SB_FEAT_RO_COMPAT_FINOBT)) > > > > +#define XFS_SB_FEAT_INCOMPAT_AGFL_WRAP_ALREADY_FIXED \ > > > > + (~(XFS_SB_FEAT_INCOMPAT_FTYPE | \ > > > > + XFS_SB_FEAT_INCOMPAT_SPINODES)) > > > > +#define XFS_SB_FEAT_INCOMPAT_LOG_AGFL_WRAP_ALREADY_FIXED \ > > > > + (~0) > > > > +static inline bool xfs_sb_version_needs_agfl_wrap_fixes(struct xfs_sb *sbp) > > > > +{ > > > > + return xfs_sb_version_hascrc(sbp) && > > > > + !xfs_sb_has_incompat_feature(sbp, > > > > + XFS_SB_FEAT_INCOMPAT_AGFL_WRAP_ALREADY_FIXED) && > > > > + !xfs_sb_has_ro_compat_feature(sbp, > > > > + XFS_SB_FEAT_RO_COMPAT_AGFL_WRAP_ALREADY_FIXED) && > > > > + !xfs_sb_has_incompat_log_feature(sbp, > > > > + XFS_SB_FEAT_INCOMPAT_LOG_AGFL_WRAP_ALREADY_FIXED); > > > > +} > > > > + > > > > +/* > > > > + * Fix an AGFL wrapping that falls short of the end of the block by filling the > > > > + * gap at the end of the block. > > > > + */ > > > > +STATIC int > > > > +xfs_fixup_freelist_wrap_mount( > > > > + struct xfs_trans *tp, > > > > + struct xfs_buf *agfbp, > > > > + struct xfs_perag *pag) > > > > +{ > > > > + struct xfs_mount *mp = tp->t_mountp; > > > > + struct xfs_agf *agf; > > > > + struct xfs_buf *agflbp; > > > > + __be32 *agfl_bno; > > > > + xfs_agnumber_t agno; > > > > + uint32_t agfl_size; > > > > + uint32_t flfirst; > > > > + uint32_t fllast; > > > > + int32_t active; > > > > + int offset; > > > > + int len; > > > > + int error; > > > > + > > > > + if (pag->pagf_flcount == 0) > > > > + return 0; > > > > + > > > > + agfl_size = xfs_agfl_size(mp); > > > > + agf = XFS_BUF_TO_AGF(agfbp); > > > > + agno = be32_to_cpu(agf->agf_seqno); > > > > + flfirst = be32_to_cpu(agf->agf_flfirst); > > > > + fllast = be32_to_cpu(agf->agf_fllast); > > > > + > > > > + /* Make sure we're either spot on or off by 1. */ > > > > + active = fllast - flfirst + 1; > > > > + if (active <= 0) > > > > + active += agfl_size; > > > > + if (active == pag->pagf_flcount) > > > > + return 0; > > > > + else if (active != pag->pagf_flcount + 1) > > > > + return -EFSCORRUPTED; > > > > + > > > > > > So we're not attempting to cover the case where the agfl has 1 more > > > block than the agfl size (i.e., the case where an fs goes back to a > > > kernel with an unpacked header)? > > > > We don't know how the next kernel to touch this filesystem will define > > XFS_AGFL_SIZE -- it could be a 4.5+ kernel (same agfl size), a 32-bit > > pre-4.5 kernel (same agfl size), or a 64-bit pre-4.5 kernel (small agfl > > size). > > > > I don't think I was clear.. I'm envisioning whether we could come up > with a patch that would generically fix up the agfl on disk to be sane > relative to the current kernel. This patch covers the case of a packed > agfl kernel mounting an unpacked on-disk agfl. It would be nice if we > could implement something that also handled a packed on-disk agfl to an > unpacked agfl kernel (for easier backport to unpacked kernels, for > e.g.). If we're going to touch an old kernel's source at all I'd rather we backport both the packing fix and this fixer-upper. > > > I'm wondering whether using the unmount algorithm in both places (as > > > noted above) would also facilitate this mechanism to work both ways > > > (unpacked <-> packed), provided particular conditions are met. I suppose > > > that could be considered regardless, but the less variance the better > > > IMO. > > > > In principle I suppose we could have a single function to handle all the > > rearranging, but I'd rather have two functions to handle the two cases. > > I'm open to refactoring the common parts out of the > > xfs_fixup_freelist_wrap_*mount functions. > > > > As noted above, I'm not really commenting on the function factoring here > as much as the use of multiple algorithms (using one would certainly > facilitate some mechanical refactoring). TBH the patch you sent is pretty close to what I was imagining. > > > > + /* Would this have even passed muster on an old system? */ > > > > > > Comment doesn't really explain what is going on here..? > > > > /* > > * If the distance between flfirst and fllast mismatches flcount by more > > * than 1, then there's more wrong with this agfl than just the padding > > * problem. Bail out completely, which will force the admin to run > > * xfs_repair. > > */ > > > > *nod* > > > > > > > > + if (flfirst >= agfl_size - 1 || fllast >= agfl_size - 1 || > > > > + pag->pagf_flcount > agfl_size - 1) > > > > + return -EFSCORRUPTED; > > > > + > > > > > > I take it these are checks for whether the agfl was previously corrupted > > > based on the unpacked size? FWIW, it might be a bit more clear to use an > > > old/unpacked_agfl_size variable to declare intent (here and in some of > > > the comments and whatnot that follow). > > > > Ok. > > > > > > + /* > > > > + * Convert a 40-byte-padded agfl into a 36-byte-padded AGFL. > > > > + * Therefore, we need to move the AGFL blocks > > > > + * bno[flfirst..agfl_size - 2] to bno[flfirst + 1...agfl_size - 1]. > > > > + * > > > > + * Reusing the example above, if we had flfirst == 116, we need > > > > + * to move bno[116] and bno[117] into bno[117] and bno[118], > > > > + * respectively, and then increment flfirst. > > > > + */ > > > > > > Kind of a strange example to use given that it doesn't mention > > > wrapping.. this only triggers if the agfl was previously wrapped, right? > > > > Right. I used to have a check that if the list didn't wrap then we > > would just exit because no fixing is required... but it must've fallen > > out. I agree that we don't want to rely on a subtlety here to stay out > > of here if the agfl wasn't initially wrapped. > > > > > > + error = xfs_alloc_read_agfl(mp, tp, agno, &agflbp); > > > > + if (error) > > > > + return error; > > > > + agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); > > > > + > > > > + len = (agfl_size - flfirst - 1) * sizeof(xfs_agblock_t); > > > > + memmove(&agfl_bno[flfirst + 1], &agfl_bno[flfirst], len); > > > > + offset = (char *)&agfl_bno[flfirst + 1] - (char *)agflbp->b_addr; > > > > + be32_add_cpu(&agf->agf_flfirst, 1); > > > > + > > > > + xfs_trans_buf_set_type(tp, agflbp, XFS_BLFT_AGFL_BUF); > > > > + xfs_trans_log_buf(tp, agflbp, offset, offset + len - 1); > > > > + xfs_trans_brelse(tp, agflbp); > > > > + agflbp = NULL; > > > > + xfs_alloc_log_agf(tp, agfbp, XFS_AGF_FLFIRST); > > > > + > > > > + return 0; > > > > +} > > > > + > > > > +/* > > > > + * Fix an AGFL that touches the end of the block by moving the first or last > > > > + * part of the list elsewhere in the AGFL so that old kernels don't trip over > > > > + * wrapping issues. > > > > + */ > > > > +STATIC int > > > > +xfs_fixup_freelist_wrap_unmount( > > > > + struct xfs_trans *tp, > > > > + struct xfs_buf *agfbp, > > > > + struct xfs_perag *pag) > > > > +{ > > > > + struct xfs_mount *mp = tp->t_mountp; > > > > + struct xfs_agf *agf; > > > > + struct xfs_buf *agflbp; > > > > + __be32 *agfl_bno; > > > > + xfs_agnumber_t agno; > > > > + uint32_t agfl_size; > > > > + uint32_t flfirst; > > > > + uint32_t fllast; > > > > + int offset; > > > > + int len; > > > > + int error; > > > > + > > > > + agfl_size = xfs_agfl_size(mp); > > > > + agf = XFS_BUF_TO_AGF(agfbp); > > > > + agno = be32_to_cpu(agf->agf_seqno); > > > > + flfirst = be32_to_cpu(agf->agf_flfirst); > > > > + fllast = be32_to_cpu(agf->agf_fllast); > > > > + > > > > + /* Empty AGFL? Make sure we aren't pointing at the end. */ > > > > + if (pag->pagf_flcount == 0) { > > > > + if (flfirst >= agfl_size || fllast >= agfl_size) { > > > > > > When would either field be >= agfl_size? Isn't that where they wrap? > > > > Oops, you're right, that should be flfirst >= agfl_size - 1. > > > > Or really, > > > > old_agfl_size = agfl_size - 1 > > if flfirst >= old_agfl_size or fllast >= old_agfl_size: > > blahblahblah > > > > > > > > > + agf->agf_flfirst = cpu_to_be32(1); > > > > + agf->agf_fllast = 0; > > > > + xfs_alloc_log_agf(tp, agfbp, > > > > + XFS_AGF_FLFIRST | XFS_AGF_FLLAST); > > > > + } > > > > + return 0; > > > > + } > > > > + > > > > + /* If we don't hit the end, we're done. */ > > > > + if (flfirst < fllast && fllast != agfl_size - 1) > > > > + return 0; > > > > + > > > > + /* > > > > + * Move a start of a wrapped list towards the start of the agfl block. > > > > > > FWIW, this and the subsequent comments kind of gave me the impression > > > that the agfl would be "shifted" towards the start of the block. The > > > code doesn't do that, but rather rotates the start/head of the agfl to > > > the tail and adjusts the pointers (changing the effective order). That > > > seems technically Ok, but I had to grok the code and read back to > > > understand the comment rather than the other way around. > > > > Yeah. Sorry about the confusion, maybe it would be better if I took out > > these confusing comments and replaced it all with the ascii art above? > > > > I'm still hoping we can do something a bit more simple and flexible in > general, but that aside, the ascii art describes the associated > algorithms much better. > > > > > + * Therefore, we need to move the AGFL blocks > > > > + * bno[flfirst..agfl_size - 1] to bno[fllast + 1...agfl_size - flfirst]. > > > > + * Then we reset flfirst and fllast appropriately. > > > > + * > > > > + * Reusing the example above, if we had flfirst == 117 and fllast == 4, > > > > + * we need to move bno[117] and bno[118] into bno[5] and bno[6], > > > > + * respectively, and then reset flfirst and fllast. > > > > + * > > > > + * If it's just the last block that touches the end, only move that. > > > > + */ > > > > + error = xfs_alloc_read_agfl(mp, tp, agno, &agflbp); > > > > + if (error) > > > > + return error; > > > > + agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); > > > > + > > > > + if (fllast == agfl_size - 1) { > > > > + /* Back the AGFL off from the end of the block. */ > > > > + len = sizeof(xfs_agblock_t); > > > > + agfl_bno[flfirst - 1] = agfl_bno[agfl_size - 1]; > > > > > > What if the agfl is full (flfirst == 0)? > > > > I don't think we ever unmount with a full agfl -- we're guaranteed at > > least 118 items in the AGFL, and the maximum size the agfl will ever > > need is 8 blocks for each of bnobt, cntbt, rmapbt. > > > > However, we could just detect a full agfl and free the end block. > > > > I suppose it's unlikely we'll ever see corner cases such as a completely > full or empty agfl. That said, I think the code needs to handle those > cases sanely. That could mean explicit checks and skipping the fixup for > all I care, just as long as we don't have memory corruption vectors and > whatnot if (when?) assumptions prove wrong. > > > > > + offset = (char *)&agfl_bno[flfirst - 1] - (char *)agflbp->b_addr; > > > > + be32_add_cpu(&agf->agf_fllast, -1); > > > > + be32_add_cpu(&agf->agf_flfirst, -1); > > > > + } else { > > > > This needs to check for flfirst > fllast. > > > > > > + /* Move the first part of the AGFL towards the front. */ > > > > + len = (agfl_size - flfirst) * sizeof(xfs_agblock_t); > > > > + memcpy(&agfl_bno[fllast + 1], &agfl_bno[flfirst], len); > > > > + offset = (char *)&agfl_bno[fllast + 1] - (char *)agflbp->b_addr; > > > > + agf->agf_flfirst = 0; > > > > + agf->agf_fllast = cpu_to_be32(pag->pagf_flcount - 1); > > > > > > Similar question here as above.. it looks like the copy may be a no-op, > > > but we still reset flfirst/fllast (which may also be fine, but > > > unnecessary). > > > > > > The more interesting case here is flcount == 1 and flfirst == fllast, > > > where it looks like we'd wrongly reset first/last (and perhaps copy more > > > than we should..?). > > > > > > Ugh. Remind me again why we can't just detect a mismatch, fail the mount > > > and tell the user to repair? :P > > > > Not all the distros included xfs_repair in the initrd or the logic to > > call it (as opposed to "fsck -y") if the root fs fails to mount, which > > means that if this wrapping problem hits a root filesystem then the > > administrator will have to reboot with a recovery iso and run xfs_repair > > from there. That's going to generate a /lot/ of support work compared > > to us figuring out how to repair the filesystems automatically. > > > > Yeah, sorry. Good argument for the unpacked -> packed fixup. > > > > Random thought... have we considered anything less invasive/more > > > simple for this issue than low level buffer manipulation? For example, > > > could we address the unpacked -> packed case by just allocating a > > > block to fill the gap at the end of the wrapped agfl? > > > > The allocation could fail, in which case we'd be forced to resort to > > list manipulation. > > > > Yeah.. obvious tradeoff, but how likely is that? It seems a reasonable > tradeoff to me given there's still the repair path. Also note that we > don't have to necessarily allocate from the btrees. > > (FWIW, I hacked around on this a bit and ended up with something > slightly different[1].) > > > > Perhaps that could even be done more discreetly in > > > xfs_alloc_fix_freelist() when we already have a transaction set up, > > > etc. > > > > Possible, though this will add more logic to every hot allocation call > > for something that only needs fixing at mount and at unmount. I think > > there's a stronger argument for doing the unmount cleanup on every > > fix_freelist since that would eliminate the small chance that a new > > kernel crashes and gets recovered on an old kernel. > > > > Eh, I don't think that's anything we couldn't address if need be. E.g., > amortize the padding fixup against actual allocations, set/test a pag > bit to track the padding fixup, or check it on agf read and set a "needs > fixup" bit, etc. > > > > That wouldn't help going the other direction, but maybe we could > > > consider similar logic to free an extraneous block and then worry > > > about protecting users who won't upgrade broken kernels separately if > > > that really continues to be a problem. > > > > It is already a problem here. > > > > Then it should be able to stand alone as an independent patch (with > independent commit log that explains/justifies it, etc.). > > Brian > > [1] This is hacky and brute force, but the idea was to try and do > something that could work in either direction and reuse more existing > code. It basically just reads the on-disk entries "manually" and > reinserts them into a reinitialized (empty) agfl. Perhaps it could be > refactored into something that only rotates a given number of entries or > something of that nature. Anyways, just an experiment and probably > broken.. I'll give this a whirl and see where we end up. --D > int > xfs_agfl_fixup( > struct xfs_trans *tp, > struct xfs_buf *agbp, > struct xfs_perag *pag) > { > struct xfs_mount *mp = tp->t_mountp; > struct xfs_agf *agf = XFS_BUF_TO_AGF(agbp); > int agfl_size = XFS_AGFL_SIZE(mp); > int ofirst, olast, osize; > int nfirst, nlast; > struct xfs_buf *agflbp; > __be32 *agfl_bno; > int active; > xfs_agblock_t bno; > bool wasempty = false; > > if (pag->pagf_flcount == 0) > wasempty = true; > > ofirst = be32_to_cpu(agf->agf_flfirst); > olast = be32_to_cpu(agf->agf_fllast); > if (olast >= ofirst) > active = olast - ofirst + 1; > else > active = agfl_size - ofirst + olast + 1; > > if (active == pag->pagf_flcount + 1) > osize = agfl_size - 1; > else if ((active == pag->pagf_flcount - 1) || > ofirst == agfl_size || olast == agfl_size) > osize = agfl_size + 1; > else if (active == pag->pagf_flcount) > osize = agfl_size; > else > return -EFSCORRUPTED; > > xfs_alloc_read_agfl(mp, tp, be32_to_cpu(agf->agf_seqno), &agflbp); > agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); > > nlast = do_mod(olast, agfl_size); > nfirst = nlast + 1; > nfirst = do_mod(nfirst, agfl_size); > ASSERT(nfirst != ofirst); > > agf->agf_flfirst = cpu_to_be32(nfirst); > agf->agf_fllast = cpu_to_be32(nlast); > agf->agf_flcount = 0; > xfs_alloc_log_agf(tp, agbp, XFS_AGF_FLFIRST | XFS_AGF_FLLAST | > XFS_AGF_FLCOUNT); > pag->pagf_flcount = 0; > > if (wasempty) > goto out; > > while (true) { > ofirst = do_mod(ofirst, osize); > bno = be32_to_cpu(agfl_bno[ofirst]); > xfs_alloc_put_freelist(tp, agbp, agflbp, bno, 0); > if (ofirst == olast) > break; > ofirst++; > } > > out: > xfs_trans_brelse(tp, agflbp); > return 0; > } > -- To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Tue, Feb 27, 2018 at 02:35:49PM -0500, Brian Foster wrote: > On Thu, Feb 22, 2018 at 06:00:15PM -0800, Darrick J. Wong wrote: > > From: Darrick J. Wong <darrick.wong@oracle.com> > > > > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> > > --- > > fs/xfs/Makefile | 1 > > fs/xfs/xfs_fixups.c | 310 +++++++++++++++++++++++++++++++++++++++++++++++++++ > > fs/xfs/xfs_fixups.h | 26 ++++ > > fs/xfs/xfs_mount.c | 21 +++ > > fs/xfs/xfs_super.c | 10 ++ > > 5 files changed, 367 insertions(+), 1 deletion(-) > > create mode 100644 fs/xfs/xfs_fixups.c > > create mode 100644 fs/xfs/xfs_fixups.h > > > > > > diff --git a/fs/xfs/Makefile b/fs/xfs/Makefile > > index b03c77e..f88368a 100644 > > --- a/fs/xfs/Makefile > > +++ b/fs/xfs/Makefile > > @@ -86,6 +86,7 @@ xfs-y += xfs_aops.o \ > > xfs_extent_busy.o \ > > xfs_file.o \ > > xfs_filestream.o \ > > + xfs_fixups.o \ > > xfs_fsmap.o \ > > xfs_fsops.o \ > > xfs_globals.o \ > > diff --git a/fs/xfs/xfs_fixups.c b/fs/xfs/xfs_fixups.c > > new file mode 100644 > > index 0000000..0cad7bb > > --- /dev/null > > +++ b/fs/xfs/xfs_fixups.c > > @@ -0,0 +1,310 @@ > > +/* > > + * Copyright (C) 2018 Oracle. All Rights Reserved. > > + * > > + * Author: Darrick J. Wong <darrick.wong@oracle.com> > > + * > > + * This program is free software; you can redistribute it and/or > > + * modify it under the terms of the GNU General Public License > > + * as published by the Free Software Foundation; either version 2 > > + * of the License, or (at your option) any later version. > > + * > > + * This program is distributed in the hope that it would be useful, > > + * but WITHOUT ANY WARRANTY; without even the implied warranty of > > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the > > + * GNU General Public License for more details. > > + * > > + * You should have received a copy of the GNU General Public License > > + * along with this program; if not, write the Free Software Foundation, > > + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. > > + */ > > +#include "xfs.h" > > +#include "xfs_fs.h" > > +#include "xfs_shared.h" > > +#include "xfs_format.h" > > +#include "xfs_log_format.h" > > +#include "xfs_trans_resv.h" > > +#include "xfs_sb.h" > > +#include "xfs_mount.h" > > +#include "xfs_alloc.h" > > +#include "xfs_trans.h" > > +#include "xfs_fixups.h" > > + > > +/* > > + * v5 AGFL padding defects > > + * > > + * When the v5 format was first introduced, there was a defect in the struct > > + * xfs_agfl definition that resulted in XFS_AGFL_SIZE returning different > > XFS_AGFL_SIZE() no longer exists as of the previous patch. It might be > better to size refer to "the agfl size" generically. > > > + * values depending on the compiler padding. On a fs with 512-byte sectors, > > + * this meant that XFS_AGFL_SIZE was 119 on i386, but 118 on x64. Commit > > + * 96f859d52bcb1 ("libxfs: pack the agfl header structure so XFS_AGFL_SIZE is > > + * correct") changed the definition to disable padding the end of the > > + * structure, and was accepted into Linux 4.5. Since then, the AGFL has > > + * always used the larger size (e.g. 119 entries on a 512b sector fs). > > + * > > + * Unfortunately, pre-4.5 kernels can produce filesystems with AGFLs that wrap > > + * at the smaller size, and those kernels are not prepared to handle the > > + * longer size. This typically manifests itself as an AGF verifier corruption > > + * error followed by a filesystem shutdown. While we encourage admins to stay > > + * current with software, we would like to avoid this intermittent breakage. > > + * > > + * Any v5 filesystem which has a feature bit set for a feature that was > > + * introduced after Linux 4.5 will not have this problem, as such kernels > > + * cannot be mounted on older kernels. v4 filesystems are also unaffected. > > + * > > + * Therefore, we add two fixup functions -- the first runs at mount time to > > + * detect a short-wrapped AGFL and fix it; the second runs at unmount, freeze, > > + * or remount-ro time to move a wrapped AGFL to the beginning of the list. > > + * This reduces the likelihood of a screwup to the scenario where you have (a) > > + * a filesystem with no post-4.5 features (reflink, rmap), (b) the AGFL wraps, > > + * (c) the filesystem goes down leaving a dirty log, and (d) the dirty > > + * filesystem is mounted on an old kernel. > > + */ > > + > > While the mount vs. unmount time fixups serve different purposes and > have different conditions, do we really need two separate fixup > implementations? E.g., if we have to unwrap the AGFL at unmount time, > why not just reuse that fixup at mount time (when needed) as well? I > suspect the difference would only be the length of what we can consider > valid in the agfl at the time. TBH I rewrote the whole thing (again...) so that we move all the list items to the front of the agfl block no matter what problem we detect. I also added more ascii art to describe the two other things that the unmount fixer has to detect (list hits the end of the block at unmount) and (empty list has pointers touching the end of the block). It seems to have passed the quick run so I'll send it out right after this. I think my email receiving is broken. --D > > > +/* > > + * Decide if we need to have the agfl wrapping fixes applied. This only > > + * affects v5 filesystems that do not have any features enabled that did not > > + * exist when the agfl padding fix went in. > > + * > > + * Features already present when the fix went in were finobt, ftype, spinodes. > > + * If we see something new (e.g. reflink) then don't bother. > > + */ > > +#define XFS_SB_FEAT_RO_COMPAT_AGFL_WRAP_ALREADY_FIXED \ > > + (~(XFS_SB_FEAT_RO_COMPAT_FINOBT)) > > +#define XFS_SB_FEAT_INCOMPAT_AGFL_WRAP_ALREADY_FIXED \ > > + (~(XFS_SB_FEAT_INCOMPAT_FTYPE | \ > > + XFS_SB_FEAT_INCOMPAT_SPINODES)) > > +#define XFS_SB_FEAT_INCOMPAT_LOG_AGFL_WRAP_ALREADY_FIXED \ > > + (~0) > > +static inline bool xfs_sb_version_needs_agfl_wrap_fixes(struct xfs_sb *sbp) > > +{ > > + return xfs_sb_version_hascrc(sbp) && > > + !xfs_sb_has_incompat_feature(sbp, > > + XFS_SB_FEAT_INCOMPAT_AGFL_WRAP_ALREADY_FIXED) && > > + !xfs_sb_has_ro_compat_feature(sbp, > > + XFS_SB_FEAT_RO_COMPAT_AGFL_WRAP_ALREADY_FIXED) && > > + !xfs_sb_has_incompat_log_feature(sbp, > > + XFS_SB_FEAT_INCOMPAT_LOG_AGFL_WRAP_ALREADY_FIXED); > > +} > > + > > +/* > > + * Fix an AGFL wrapping that falls short of the end of the block by filling the > > + * gap at the end of the block. > > + */ > > +STATIC int > > +xfs_fixup_freelist_wrap_mount( > > + struct xfs_trans *tp, > > + struct xfs_buf *agfbp, > > + struct xfs_perag *pag) > > +{ > > + struct xfs_mount *mp = tp->t_mountp; > > + struct xfs_agf *agf; > > + struct xfs_buf *agflbp; > > + __be32 *agfl_bno; > > + xfs_agnumber_t agno; > > + uint32_t agfl_size; > > + uint32_t flfirst; > > + uint32_t fllast; > > + int32_t active; > > + int offset; > > + int len; > > + int error; > > + > > + if (pag->pagf_flcount == 0) > > + return 0; > > + > > + agfl_size = xfs_agfl_size(mp); > > + agf = XFS_BUF_TO_AGF(agfbp); > > + agno = be32_to_cpu(agf->agf_seqno); > > + flfirst = be32_to_cpu(agf->agf_flfirst); > > + fllast = be32_to_cpu(agf->agf_fllast); > > + > > + /* Make sure we're either spot on or off by 1. */ > > + active = fllast - flfirst + 1; > > + if (active <= 0) > > + active += agfl_size; > > + if (active == pag->pagf_flcount) > > + return 0; > > + else if (active != pag->pagf_flcount + 1) > > + return -EFSCORRUPTED; > > + > > So we're not attempting to cover the case where the agfl has 1 more > block than the agfl size (i.e., the case where an fs goes back to a > kernel with an unpacked header)? > > I'm wondering whether using the unmount algorithm in both places (as > noted above) would also facilitate this mechanism to work both ways > (unpacked <-> packed), provided particular conditions are met. I suppose > that could be considered regardless, but the less variance the better > IMO. > > > + /* Would this have even passed muster on an old system? */ > > Comment doesn't really explain what is going on here..? > > > + if (flfirst >= agfl_size - 1 || fllast >= agfl_size - 1 || > > + pag->pagf_flcount > agfl_size - 1) > > + return -EFSCORRUPTED; > > + > > I take it these are checks for whether the agfl was previously corrupted > based on the unpacked size? FWIW, it might be a bit more clear to use an > old/unpacked_agfl_size variable to declare intent (here and in some of > the comments and whatnot that follow). > > > + /* > > + * Convert a 40-byte-padded agfl into a 36-byte-padded AGFL. > > + * Therefore, we need to move the AGFL blocks > > + * bno[flfirst..agfl_size - 2] to bno[flfirst + 1...agfl_size - 1]. > > + * > > + * Reusing the example above, if we had flfirst == 116, we need > > + * to move bno[116] and bno[117] into bno[117] and bno[118], > > + * respectively, and then increment flfirst. > > + */ > > Kind of a strange example to use given that it doesn't mention > wrapping.. this only triggers if the agfl was previously wrapped, right? > > > + error = xfs_alloc_read_agfl(mp, tp, agno, &agflbp); > > + if (error) > > + return error; > > + agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); > > + > > + len = (agfl_size - flfirst - 1) * sizeof(xfs_agblock_t); > > + memmove(&agfl_bno[flfirst + 1], &agfl_bno[flfirst], len); > > + offset = (char *)&agfl_bno[flfirst + 1] - (char *)agflbp->b_addr; > > + be32_add_cpu(&agf->agf_flfirst, 1); > > + > > + xfs_trans_buf_set_type(tp, agflbp, XFS_BLFT_AGFL_BUF); > > + xfs_trans_log_buf(tp, agflbp, offset, offset + len - 1); > > + xfs_trans_brelse(tp, agflbp); > > + agflbp = NULL; > > + xfs_alloc_log_agf(tp, agfbp, XFS_AGF_FLFIRST); > > + > > + return 0; > > +} > > + > > +/* > > + * Fix an AGFL that touches the end of the block by moving the first or last > > + * part of the list elsewhere in the AGFL so that old kernels don't trip over > > + * wrapping issues. > > + */ > > +STATIC int > > +xfs_fixup_freelist_wrap_unmount( > > + struct xfs_trans *tp, > > + struct xfs_buf *agfbp, > > + struct xfs_perag *pag) > > +{ > > + struct xfs_mount *mp = tp->t_mountp; > > + struct xfs_agf *agf; > > + struct xfs_buf *agflbp; > > + __be32 *agfl_bno; > > + xfs_agnumber_t agno; > > + uint32_t agfl_size; > > + uint32_t flfirst; > > + uint32_t fllast; > > + int offset; > > + int len; > > + int error; > > + > > + agfl_size = xfs_agfl_size(mp); > > + agf = XFS_BUF_TO_AGF(agfbp); > > + agno = be32_to_cpu(agf->agf_seqno); > > + flfirst = be32_to_cpu(agf->agf_flfirst); > > + fllast = be32_to_cpu(agf->agf_fllast); > > + > > + /* Empty AGFL? Make sure we aren't pointing at the end. */ > > + if (pag->pagf_flcount == 0) { > > + if (flfirst >= agfl_size || fllast >= agfl_size) { > > When would either field be >= agfl_size? Isn't that where they wrap? > > > + agf->agf_flfirst = cpu_to_be32(1); > > + agf->agf_fllast = 0; > > + xfs_alloc_log_agf(tp, agfbp, > > + XFS_AGF_FLFIRST | XFS_AGF_FLLAST); > > + } > > + return 0; > > + } > > + > > + /* If we don't hit the end, we're done. */ > > + if (flfirst < fllast && fllast != agfl_size - 1) > > + return 0; > > + > > + /* > > + * Move a start of a wrapped list towards the start of the agfl block. > > FWIW, this and the subsequent comments kind of gave me the impression > that the agfl would be "shifted" towards the start of the block. The > code doesn't do that, but rather rotates the start/head of the agfl to > the tail and adjusts the pointers (changing the effective order). That > seems technically Ok, but I had to grok the code and read back to > understand the comment rather than the other way around. > > > + * Therefore, we need to move the AGFL blocks > > + * bno[flfirst..agfl_size - 1] to bno[fllast + 1...agfl_size - flfirst]. > > + * Then we reset flfirst and fllast appropriately. > > + * > > + * Reusing the example above, if we had flfirst == 117 and fllast == 4, > > + * we need to move bno[117] and bno[118] into bno[5] and bno[6], > > + * respectively, and then reset flfirst and fllast. > > + * > > + * If it's just the last block that touches the end, only move that. > > + */ > > + error = xfs_alloc_read_agfl(mp, tp, agno, &agflbp); > > + if (error) > > + return error; > > + agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); > > + > > + if (fllast == agfl_size - 1) { > > + /* Back the AGFL off from the end of the block. */ > > + len = sizeof(xfs_agblock_t); > > + agfl_bno[flfirst - 1] = agfl_bno[agfl_size - 1]; > > What if the agfl is full (flfirst == 0)? > > > + offset = (char *)&agfl_bno[flfirst - 1] - (char *)agflbp->b_addr; > > + be32_add_cpu(&agf->agf_fllast, -1); > > + be32_add_cpu(&agf->agf_flfirst, -1); > > + } else { > > + /* Move the first part of the AGFL towards the front. */ > > + len = (agfl_size - flfirst) * sizeof(xfs_agblock_t); > > + memcpy(&agfl_bno[fllast + 1], &agfl_bno[flfirst], len); > > + offset = (char *)&agfl_bno[fllast + 1] - (char *)agflbp->b_addr; > > + agf->agf_flfirst = 0; > > + agf->agf_fllast = cpu_to_be32(pag->pagf_flcount - 1); > > Similar question here as above.. it looks like the copy may be a no-op, > but we still reset flfirst/fllast (which may also be fine, but > unnecessary). > > The more interesting case here is flcount == 1 and flfirst == fllast, > where it looks like we'd wrongly reset first/last (and perhaps copy more > than we should..?). > > Ugh. Remind me again why we can't just detect a mismatch, fail the mount > and tell the user to repair? :P Random thought... have we considered > anything less invasive/more simple for this issue than low level buffer > manipulation? For example, could we address the unpacked -> packed case > by just allocating a block to fill the gap at the end of the wrapped > agfl? Perhaps that could even be done more discreetly in > xfs_alloc_fix_freelist() when we already have a transaction set up, etc. > That wouldn't help going the other direction, but maybe we could > consider similar logic to free an extraneous block and then worry about > protecting users who won't upgrade broken kernels separately if that > really continues to be a problem. > > Brian > > > + } > > + > > + xfs_trans_buf_set_type(tp, agflbp, XFS_BLFT_AGFL_BUF); > > + xfs_trans_log_buf(tp, agflbp, offset, offset + len - 1); > > + xfs_trans_brelse(tp, agflbp); > > + agflbp = NULL; > > + xfs_alloc_log_agf(tp, agfbp, XFS_AGF_FLFIRST | XFS_AGF_FLLAST); > > + > > + return 0; > > +} > > + > > +typedef int (*xfs_agf_apply_fn_t)(struct xfs_trans *tp, struct xfs_buf *agfbp, > > + struct xfs_perag *pag); > > + > > +/* Apply something to every AGF. */ > > +STATIC int > > +xfs_fixup_agf_apply( > > + struct xfs_mount *mp, > > + xfs_agf_apply_fn_t fn) > > +{ > > + struct xfs_trans *tp; > > + struct xfs_perag *pag; > > + struct xfs_buf *agfbp; > > + xfs_agnumber_t agno; > > + int error; > > + > > + error = xfs_trans_alloc(mp, &M_RES(mp)->tr_growdata, 0, 0, 0, &tp); > > + if (error) > > + return error; > > + > > + for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) { > > + error = xfs_alloc_read_agf(mp, tp, agno, 0, &agfbp); > > + if (error) > > + goto cancel; > > + if (!agfbp) { > > + error = -ENOMEM; > > + goto cancel; > > + } > > + pag = xfs_perag_get(mp, agno); > > + error = fn(tp, agfbp, pag); > > + xfs_perag_put(pag); > > + xfs_trans_brelse(tp, agfbp); > > + if (error) > > + goto cancel; > > + } > > + > > + return xfs_trans_commit(tp); > > +cancel: > > + xfs_trans_cancel(tp); > > + return error; > > +} > > + > > +/* Fix AGFL wrapping so we can use the filesystem. */ > > +int > > +xfs_fixup_agfl_wrap_mount( > > + struct xfs_mount *mp) > > +{ > > + if (!xfs_sb_version_needs_agfl_wrap_fixes(&mp->m_sb)) > > + return 0; > > + > > + return xfs_fixup_agf_apply(mp, xfs_fixup_freelist_wrap_mount); > > +} > > + > > +/* Fix AGFL wrapping so old kernels can use this filesystem. */ > > +int > > +xfs_fixup_agfl_wrap_unmount( > > + struct xfs_mount *mp) > > +{ > > + if (!xfs_sb_version_needs_agfl_wrap_fixes(&mp->m_sb)) > > + return 0; > > + > > + return xfs_fixup_agf_apply(mp, xfs_fixup_freelist_wrap_unmount); > > +} > > diff --git a/fs/xfs/xfs_fixups.h b/fs/xfs/xfs_fixups.h > > new file mode 100644 > > index 0000000..fb52a96 > > --- /dev/null > > +++ b/fs/xfs/xfs_fixups.h > > @@ -0,0 +1,26 @@ > > +/* > > + * Copyright (C) 2018 Oracle. All Rights Reserved. > > + * > > + * Author: Darrick J. Wong <darrick.wong@oracle.com> > > + * > > + * This program is free software; you can redistribute it and/or > > + * modify it under the terms of the GNU General Public License > > + * as published by the Free Software Foundation; either version 2 > > + * of the License, or (at your option) any later version. > > + * > > + * This program is distributed in the hope that it would be useful, > > + * but WITHOUT ANY WARRANTY; without even the implied warranty of > > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the > > + * GNU General Public License for more details. > > + * > > + * You should have received a copy of the GNU General Public License > > + * along with this program; if not, write the Free Software Foundation, > > + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. > > + */ > > +#ifndef __XFS_FIXUPS_H__ > > +#define __XFS_FIXUPS_H__ > > + > > +int xfs_fixup_agfl_wrap_mount(struct xfs_mount *mp); > > +int xfs_fixup_agfl_wrap_unmount(struct xfs_mount *mp); > > + > > +#endif /* __XFS_FIXUPS_H__ */ > > diff --git a/fs/xfs/xfs_mount.c b/fs/xfs/xfs_mount.c > > index 98fd41c..eb284aa 100644 > > --- a/fs/xfs/xfs_mount.c > > +++ b/fs/xfs/xfs_mount.c > > @@ -46,7 +46,7 @@ > > #include "xfs_refcount_btree.h" > > #include "xfs_reflink.h" > > #include "xfs_extent_busy.h" > > - > > +#include "xfs_fixups.h" > > > > static DEFINE_MUTEX(xfs_uuid_table_mutex); > > static int xfs_uuid_table_size; > > @@ -875,6 +875,16 @@ xfs_mountfs( > > } > > > > /* > > + * Make sure our AGFL counters do not wrap the end of the block > > + * in a troublesome manner. > > + */ > > + error = xfs_fixup_agfl_wrap_mount(mp); > > + if (error) { > > + xfs_warn(mp, "Failed to fix agfl wrapping. Run xfs_repair."); > > + goto out_log_dealloc; > > + } > > + > > + /* > > * Get and sanity-check the root inode. > > * Save the pointer to it in the mount structure. > > */ > > @@ -1128,6 +1138,15 @@ xfs_unmountfs( > > xfs_qm_unmount(mp); > > > > /* > > + * Make sure our AGFL counters do not wrap the end of the block > > + * in a troublesome manner for old kernels. > > + */ > > + error = xfs_fixup_agfl_wrap_unmount(mp); > > + if (error) > > + xfs_warn(mp, "Unable to fix agfl wrapping. " > > + "This may cause problems on next mount."); > > + > > + /* > > * Unreserve any blocks we have so that when we unmount we don't account > > * the reserved free space as used. This is really only necessary for > > * lazy superblock counting because it trusts the incore superblock > > diff --git a/fs/xfs/xfs_super.c b/fs/xfs/xfs_super.c > > index 624a802..d9aa39a 100644 > > --- a/fs/xfs/xfs_super.c > > +++ b/fs/xfs/xfs_super.c > > @@ -50,6 +50,7 @@ > > #include "xfs_refcount_item.h" > > #include "xfs_bmap_item.h" > > #include "xfs_reflink.h" > > +#include "xfs_fixups.h" > > > > #include <linux/namei.h> > > #include <linux/dax.h> > > @@ -1206,6 +1207,15 @@ xfs_quiesce_attr( > > xfs_reclaim_inodes(mp, 0); > > xfs_reclaim_inodes(mp, SYNC_WAIT); > > > > + /* > > + * Make sure our AGFL counters do not wrap the end of the block > > + * in a troublesome manner for old kernels. > > + */ > > + error = xfs_fixup_agfl_wrap_unmount(mp); > > + if (error) > > + xfs_warn(mp, "Unable to fix agfl wrapping. " > > + "This may cause problems on next mount."); > > + > > /* Push the superblock and write an unmount record */ > > error = xfs_log_sbcount(mp); > > if (error) > > > > -- > > To unsubscribe from this list: send the line "unsubscribe linux-xfs" in > > the body of a message to majordomo@vger.kernel.org > > More majordomo info at http://vger.kernel.org/majordomo-info.html > -- > To unsubscribe from this list: send the line "unsubscribe linux-xfs" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, Feb 28, 2018 at 03:20:32PM -0800, Darrick J. Wong wrote: > On Wed, Feb 28, 2018 at 05:43:51PM -0500, Brian Foster wrote: > > On Tue, Feb 27, 2018 at 01:03:13PM -0800, Darrick J. Wong wrote: > > > On Tue, Feb 27, 2018 at 02:35:49PM -0500, Brian Foster wrote: > > > > On Thu, Feb 22, 2018 at 06:00:15PM -0800, Darrick J. Wong wrote: > > > > > From: Darrick J. Wong <darrick.wong@oracle.com> ... > > > > > diff --git a/fs/xfs/xfs_fixups.c b/fs/xfs/xfs_fixups.c > > > > > new file mode 100644 > > > > > index 0000000..0cad7bb > > > > > --- /dev/null > > > > > +++ b/fs/xfs/xfs_fixups.c > > > > > @@ -0,0 +1,310 @@ ... > > > Now let's say that we run for a while and want to unmount, but our AGFL > > > wraps like this: > > > > > > 0 1 2 3 4 5 6 7 8 9 > > > Header | T | U | V | | | | | Q | R | S | > > > ^-- fllast ^-- flfirst > > > > > > We don't know that the next kernel to mount this filesystem will have > > > the "4.5 agfl padding fix" applied to it; if it does not, it will flag > > > the AGF as corrupted because flcount is 6 but in its view the distance > > > between flfirst and fllast (which omits bno[9]) is 5. We don't want > > > it to choke on that, so the unmount fixer moves all the records between > > > flfirst and the end (Q, R, and S) towards the start of the block and > > > resets flfirst/fllast: > > > > > > 0 1 2 3 4 5 6 7 8 9 > > > Header | T | U | V | Q | R | S | | | | | > > > ^-- flfirst ^-- fllast > > > > > > > Could we use this basic algorithm at mount time as well? I know it > > wouldn't be _exactly_ the same operation at mount time as it is for > > unmount since slot 9 is a gap in the former case, but afaict the only > > difference from an algorithmic perspective is the length of the shift. > > > > IOW, if we were to parameterize the length of the shift and have the > > mount fixup call the unmount algorithm, would it not also address the > > problem? > > Yes, I believe that would work. I think it would be more efficient in > the patch below to memmove the entries instead of put_freelist'ing them > individually. > > If the agfl is completely full then fix_freelist ought to trim it down > by at least one element. The algorithm then becomes: > > if mounting and flfirst < fllast: > return 0 > if flcount == agfl_size: > assert !mounting > fix_freelist() Not sure I follow where the fix_freelist() came in, but I think we need to be careful here. What if flfirst == 118 and that slot is garbage? Don't we need to fix up the agfl before we can allow any traditional operations to proceed? > assert flcount < agfl_size > if flfirst < fllast: > return 0 > movelen = agfl_size - flfirst > if active == flcount - 1: > movelen-- > memmove(&agflbno[fllast + 1], &agflbno[flfirst], movelen) > flfirst = 0 > fllast = fllast + movelen > ... > > > > > > > > So we're not attempting to cover the case where the agfl has 1 more > > > > block than the agfl size (i.e., the case where an fs goes back to a > > > > kernel with an unpacked header)? > > > > > > We don't know how the next kernel to touch this filesystem will define > > > XFS_AGFL_SIZE -- it could be a 4.5+ kernel (same agfl size), a 32-bit > > > pre-4.5 kernel (same agfl size), or a 64-bit pre-4.5 kernel (small agfl > > > size). > > > > > > > I don't think I was clear.. I'm envisioning whether we could come up > > with a patch that would generically fix up the agfl on disk to be sane > > relative to the current kernel. This patch covers the case of a packed > > agfl kernel mounting an unpacked on-disk agfl. It would be nice if we > > could implement something that also handled a packed on-disk agfl to an > > unpacked agfl kernel (for easier backport to unpacked kernels, for > > e.g.). > > If we're going to touch an old kernel's source at all I'd rather we > backport both the packing fix and this fixer-upper. > Not sure I parse... the "old kernel" is essentially the rhel example where we apparently have deliberately maintained the unpacked format to avoid this incompatibility problem. If we had a patch that generically converted on-disk format (packed or unpacked) to the current kernel (packed or unpacked), we could merge that patch to upstream, stable as well as distro kernels that might not include the agfl packing fix and eliminate compatibility problems between them (even if the packing fix comes in out of order). Otherwise, we need a separate unpacked -> packed fixup for packed kernels (i.e., this patch) and a packed -> unpacked fixup for unpacked kernels and to make sure they are used in the right places. Trying to see if we could avoid this kind of dependency matrix was one of the objectives around the hack I posted previously. I'm not married to that particular implementation, but I'm much less concerned about inefficiency (even if it dictates a mount time fixup over a runtime one) in comparison to simplicity and flexibility. Perhaps we can accomplish something similarly flexible via direct buffer manipulation..? FWIW, I've appended another variant of the previous hack that is less brute force, but I think is still able to convert back and forth. The tradeoff is essentially that it no longer uses the same generic algorithm. Brian --- 8< --- static int xfs_agfl_ondisk_size( struct xfs_mount *mp, int first, int last, int flcount) { int active; int size; int agfl_size = XFS_AGFL_SIZE(mp); if (last >= first) active = last - first + 1; else active = agfl_size - first + last + 1; if (active == flcount + 1) size = agfl_size - 1; else if ((active == flcount - 1) || first == agfl_size || last == agfl_size) size = agfl_size + 1; else if (active == flcount) size = agfl_size; else return -EFSCORRUPTED; return size; } int xfs_agfl_fixup( struct xfs_trans *tp, struct xfs_buf *agbp, struct xfs_perag *pag) { struct xfs_mount *mp = tp->t_mountp; struct xfs_agf *agf = XFS_BUF_TO_AGF(agbp); int agfl_size = XFS_AGFL_SIZE(mp); int ofirst, olast, osize; int nfirst, nlast; struct xfs_buf *agflbp; __be32 *agfl_bno; xfs_agblock_t bno = -1; int tidx = -1; bool empty = false; int logflags = 0; int error; ofirst = nfirst = be32_to_cpu(agf->agf_flfirst); olast = nlast = be32_to_cpu(agf->agf_fllast); osize = xfs_agfl_ondisk_size(mp, ofirst, olast, pag->pagf_flcount); if (osize < 0) return osize; if (pag->pagf_flcount == 0) empty = true; /* sizes match, nothing to do */ if (osize == agfl_size) return 0; /* size mismatch, read the agfl.. */ error = xfs_alloc_read_agfl(mp, tp, be32_to_cpu(agf->agf_seqno), &agflbp); if (error) return error; agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); /* * If the on-disk agfl is smaller than what the kernel expects, the last * slot of the on-disk agfl is a gap with bogus data. Allocate the first * valid block from the agfl, manually place it in the gap and fix up * the count. */ if (osize < agfl_size) { ASSERT(!empty); error = xfs_alloc_get_freelist(tp, agbp, &bno, 0); if (error) goto out_relse; pag->pagf_flcount++; be32_add_cpu(&agf->agf_flcount, 1); logflags |= XFS_AGF_FLCOUNT; tidx = agfl_size - 1; goto done; } /* * Otherwise, the on-disk agfl is larger than what the current kernel * can manage. If the agfl was empty, we just fix up the first and last * pointers. If not, move the inaccessible block in the last slot to the * next valid, open slot. */ nfirst = do_mod(nfirst, agfl_size); if (empty) { nlast = (nfirst == 0 ? agfl_size - 1 : nfirst - 1); goto done; } if (nlast != agfl_size) nlast++; nlast = do_mod(nlast, agfl_size); tidx = nlast; bno = be32_to_cpu(agfl_bno[osize - 1]); done: if (nfirst != ofirst) { agf->agf_flfirst = cpu_to_be32(nfirst); logflags |= XFS_AGF_FLFIRST; } if (nlast != olast) { agf->agf_fllast = cpu_to_be32(nlast); logflags |= XFS_AGF_FLLAST; } if (bno != -1) { int startoff; agfl_bno[tidx] = cpu_to_be32(bno); xfs_trans_buf_set_type(tp, agflbp, XFS_BLFT_AGFL_BUF); startoff = (char *) &agfl_bno[tidx] - (char *) agflbp->b_addr; xfs_trans_log_buf(tp, agflbp, startoff, startoff + sizeof(xfs_agblock_t) - 1); } if (logflags) xfs_alloc_log_agf(tp, agbp, logflags); out_relse: xfs_trans_brelse(tp, agflbp); return error; } -- To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Mar 01, 2018 at 12:28:33PM -0500, Brian Foster wrote: > On Wed, Feb 28, 2018 at 03:20:32PM -0800, Darrick J. Wong wrote: > > On Wed, Feb 28, 2018 at 05:43:51PM -0500, Brian Foster wrote: > > > On Tue, Feb 27, 2018 at 01:03:13PM -0800, Darrick J. Wong wrote: > > > > On Tue, Feb 27, 2018 at 02:35:49PM -0500, Brian Foster wrote: > > > > > On Thu, Feb 22, 2018 at 06:00:15PM -0800, Darrick J. Wong wrote: > > > > > > From: Darrick J. Wong <darrick.wong@oracle.com> > ... > > > > > > diff --git a/fs/xfs/xfs_fixups.c b/fs/xfs/xfs_fixups.c > > > > > > new file mode 100644 > > > > > > index 0000000..0cad7bb > > > > > > --- /dev/null > > > > > > +++ b/fs/xfs/xfs_fixups.c > > > > > > @@ -0,0 +1,310 @@ > ... > > > > Now let's say that we run for a while and want to unmount, but our AGFL > > > > wraps like this: > > > > > > > > 0 1 2 3 4 5 6 7 8 9 > > > > Header | T | U | V | | | | | Q | R | S | > > > > ^-- fllast ^-- flfirst > > > > > > > > We don't know that the next kernel to mount this filesystem will have > > > > the "4.5 agfl padding fix" applied to it; if it does not, it will flag > > > > the AGF as corrupted because flcount is 6 but in its view the distance > > > > between flfirst and fllast (which omits bno[9]) is 5. We don't want > > > > it to choke on that, so the unmount fixer moves all the records between > > > > flfirst and the end (Q, R, and S) towards the start of the block and > > > > resets flfirst/fllast: > > > > > > > > 0 1 2 3 4 5 6 7 8 9 > > > > Header | T | U | V | Q | R | S | | | | | > > > > ^-- flfirst ^-- fllast > > > > > > > > > > Could we use this basic algorithm at mount time as well? I know it > > > wouldn't be _exactly_ the same operation at mount time as it is for > > > unmount since slot 9 is a gap in the former case, but afaict the only > > > difference from an algorithmic perspective is the length of the shift. > > > > > > IOW, if we were to parameterize the length of the shift and have the > > > mount fixup call the unmount algorithm, would it not also address the > > > problem? > > > > Yes, I believe that would work. I think it would be more efficient in > > the patch below to memmove the entries instead of put_freelist'ing them > > individually. > > > > If the agfl is completely full then fix_freelist ought to trim it down > > by at least one element. The algorithm then becomes: > > > > if mounting and flfirst < fllast: > > return 0 > > if flcount == agfl_size: > > assert !mounting > > fix_freelist() > > Not sure I follow where the fix_freelist() came in, but I think we need > to be careful here. What if flfirst == 118 and that slot is garbage? > Don't we need to fix up the agfl before we can allow any traditional > operations to proceed? This got confusing, so you might notice in the v2 patch that I separated this out... > > assert flcount < agfl_size > > if flfirst < fllast: > > return 0 > > movelen = agfl_size - flfirst > > if active == flcount - 1: > > movelen-- > > memmove(&agflbno[fllast + 1], &agflbno[flfirst], movelen) > > flfirst = 0 > > fllast = fllast + movelen > > > ... > > > > > > > > > > So we're not attempting to cover the case where the agfl has 1 more > > > > > block than the agfl size (i.e., the case where an fs goes back to a > > > > > kernel with an unpacked header)? > > > > > > > > We don't know how the next kernel to touch this filesystem will define > > > > XFS_AGFL_SIZE -- it could be a 4.5+ kernel (same agfl size), a 32-bit > > > > pre-4.5 kernel (same agfl size), or a 64-bit pre-4.5 kernel (small agfl > > > > size). > > > > > > > > > > I don't think I was clear.. I'm envisioning whether we could come up > > > with a patch that would generically fix up the agfl on disk to be sane > > > relative to the current kernel. This patch covers the case of a packed > > > agfl kernel mounting an unpacked on-disk agfl. It would be nice if we > > > could implement something that also handled a packed on-disk agfl to an > > > unpacked agfl kernel (for easier backport to unpacked kernels, for > > > e.g.). > > > > If we're going to touch an old kernel's source at all I'd rather we > > backport both the packing fix and this fixer-upper. > > > > Not sure I parse... the "old kernel" is essentially the rhel example > where we apparently have deliberately maintained the unpacked format to > avoid this incompatibility problem. If we had a patch that generically > converted on-disk format (packed or unpacked) to the current kernel > (packed or unpacked), we could merge that patch to upstream, stable as > well as distro kernels that might not include the agfl packing fix and > eliminate compatibility problems between them (even if the packing fix > comes in out of order). ...I wonder if we've gotten to the 'talking past each other' point, but let me try one more time. Going forward, I want the number of unpacked kernels to decrease as quickly as possible. I understand that distro kernel maintainers are not willing to apply the packing patch to their kernel until we come up with a smooth transition path. I don't want to support fixing agfls to be 118 units long on 64-bit unpacked kernels and 119 units long on 32-bit unpacked kernels, and I only want to support the packed kernels with their 119 unit long agfls. An AGFL that starts at 0 and ends at flcount-1 is compatible with packed and unpacked kernels, so the v2 patch I sent last night removes the delicate per-case surgery in favor of a new strategy where the mount time and unmount time helpers both look for agfl configurationss that are known to cause problems, and solves them all with the same solution: moving the agfl list towards the start of the block. The fixup patch requires that the kernel already include the packing fix. Its goals are: a) At mount time, ensure that all AGFLs which are not compatible with a packed kernel are made compatible with a packed kernel. b) At unmount time, ensure all AGFLs are compatible with packed kernels and unpacked kernels. Therefore, the first step for a kernel maintainer is to ensure that the packing fix is applied (which it will be for any 4.5+ kernel). The second step is to apply this fixer patch. --D > Otherwise, we need a separate unpacked -> packed fixup for packed > kernels (i.e., this patch) and a packed -> unpacked fixup for unpacked > kernels and to make sure they are used in the right places. Trying to > see if we could avoid this kind of dependency matrix was one of the > objectives around the hack I posted previously. I'm not married to that > particular implementation, but I'm much less concerned about > inefficiency (even if it dictates a mount time fixup over a runtime one) > in comparison to simplicity and flexibility. Perhaps we can accomplish > something similarly flexible via direct buffer manipulation..? > > FWIW, I've appended another variant of the previous hack that is less > brute force, but I think is still able to convert back and forth. The > tradeoff is essentially that it no longer uses the same generic > algorithm. > > Brian > > --- 8< --- > > static int > xfs_agfl_ondisk_size( > struct xfs_mount *mp, > int first, > int last, > int flcount) > { > int active; > int size; > int agfl_size = XFS_AGFL_SIZE(mp); > > if (last >= first) > active = last - first + 1; > else > active = agfl_size - first + last + 1; > > if (active == flcount + 1) > size = agfl_size - 1; > else if ((active == flcount - 1) || > first == agfl_size || last == agfl_size) > size = agfl_size + 1; > else if (active == flcount) > size = agfl_size; > else > return -EFSCORRUPTED; > > return size; > } > > int > xfs_agfl_fixup( > struct xfs_trans *tp, > struct xfs_buf *agbp, > struct xfs_perag *pag) > { > struct xfs_mount *mp = tp->t_mountp; > struct xfs_agf *agf = XFS_BUF_TO_AGF(agbp); > int agfl_size = XFS_AGFL_SIZE(mp); > int ofirst, olast, osize; > int nfirst, nlast; > struct xfs_buf *agflbp; > __be32 *agfl_bno; > xfs_agblock_t bno = -1; > int tidx = -1; > bool empty = false; > int logflags = 0; > int error; > > ofirst = nfirst = be32_to_cpu(agf->agf_flfirst); > olast = nlast = be32_to_cpu(agf->agf_fllast); > osize = xfs_agfl_ondisk_size(mp, ofirst, olast, pag->pagf_flcount); > if (osize < 0) > return osize; > if (pag->pagf_flcount == 0) > empty = true; > > /* sizes match, nothing to do */ > if (osize == agfl_size) > return 0; > > /* size mismatch, read the agfl.. */ > error = xfs_alloc_read_agfl(mp, tp, be32_to_cpu(agf->agf_seqno), > &agflbp); > if (error) > return error; > agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); > > /* > * If the on-disk agfl is smaller than what the kernel expects, the last > * slot of the on-disk agfl is a gap with bogus data. Allocate the first > * valid block from the agfl, manually place it in the gap and fix up > * the count. > */ > if (osize < agfl_size) { > ASSERT(!empty); > error = xfs_alloc_get_freelist(tp, agbp, &bno, 0); > if (error) > goto out_relse; > > pag->pagf_flcount++; > be32_add_cpu(&agf->agf_flcount, 1); > logflags |= XFS_AGF_FLCOUNT; > tidx = agfl_size - 1; > goto done; > } > > /* > * Otherwise, the on-disk agfl is larger than what the current kernel > * can manage. If the agfl was empty, we just fix up the first and last > * pointers. If not, move the inaccessible block in the last slot to the > * next valid, open slot. > */ > nfirst = do_mod(nfirst, agfl_size); > if (empty) { > nlast = (nfirst == 0 ? agfl_size - 1 : nfirst - 1); > goto done; > } > if (nlast != agfl_size) > nlast++; > nlast = do_mod(nlast, agfl_size); > tidx = nlast; > bno = be32_to_cpu(agfl_bno[osize - 1]); > > done: > if (nfirst != ofirst) { > agf->agf_flfirst = cpu_to_be32(nfirst); > logflags |= XFS_AGF_FLFIRST; > } > if (nlast != olast) { > agf->agf_fllast = cpu_to_be32(nlast); > logflags |= XFS_AGF_FLLAST; > } > if (bno != -1) { > int startoff; > > agfl_bno[tidx] = cpu_to_be32(bno); > xfs_trans_buf_set_type(tp, agflbp, XFS_BLFT_AGFL_BUF); > startoff = (char *) &agfl_bno[tidx] - (char *) agflbp->b_addr; > xfs_trans_log_buf(tp, agflbp, startoff, > startoff + sizeof(xfs_agblock_t) - 1); > } > if (logflags) > xfs_alloc_log_agf(tp, agbp, logflags); > > out_relse: > xfs_trans_brelse(tp, agflbp); > return error; > } > -- To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Mar 01, 2018 at 12:55:41PM -0800, Darrick J. Wong wrote: > On Thu, Mar 01, 2018 at 12:28:33PM -0500, Brian Foster wrote: > > On Wed, Feb 28, 2018 at 03:20:32PM -0800, Darrick J. Wong wrote: > > > On Wed, Feb 28, 2018 at 05:43:51PM -0500, Brian Foster wrote: > > > > On Tue, Feb 27, 2018 at 01:03:13PM -0800, Darrick J. Wong wrote: > > > > > On Tue, Feb 27, 2018 at 02:35:49PM -0500, Brian Foster wrote: > > > > > > On Thu, Feb 22, 2018 at 06:00:15PM -0800, Darrick J. Wong wrote: > > > > > > > From: Darrick J. Wong <darrick.wong@oracle.com> > > ... > > > > > > > diff --git a/fs/xfs/xfs_fixups.c b/fs/xfs/xfs_fixups.c > > > > > > > new file mode 100644 > > > > > > > index 0000000..0cad7bb > > > > > > > --- /dev/null > > > > > > > +++ b/fs/xfs/xfs_fixups.c > > > > > > > @@ -0,0 +1,310 @@ > > ... > > > > > Now let's say that we run for a while and want to unmount, but our AGFL > > > > > wraps like this: > > > > > > > > > > 0 1 2 3 4 5 6 7 8 9 > > > > > Header | T | U | V | | | | | Q | R | S | > > > > > ^-- fllast ^-- flfirst > > > > > > > > > > We don't know that the next kernel to mount this filesystem will have > > > > > the "4.5 agfl padding fix" applied to it; if it does not, it will flag > > > > > the AGF as corrupted because flcount is 6 but in its view the distance > > > > > between flfirst and fllast (which omits bno[9]) is 5. We don't want > > > > > it to choke on that, so the unmount fixer moves all the records between > > > > > flfirst and the end (Q, R, and S) towards the start of the block and > > > > > resets flfirst/fllast: > > > > > > > > > > 0 1 2 3 4 5 6 7 8 9 > > > > > Header | T | U | V | Q | R | S | | | | | > > > > > ^-- flfirst ^-- fllast > > > > > > > > > > > > > Could we use this basic algorithm at mount time as well? I know it > > > > wouldn't be _exactly_ the same operation at mount time as it is for > > > > unmount since slot 9 is a gap in the former case, but afaict the only > > > > difference from an algorithmic perspective is the length of the shift. > > > > > > > > IOW, if we were to parameterize the length of the shift and have the > > > > mount fixup call the unmount algorithm, would it not also address the > > > > problem? > > > > > > Yes, I believe that would work. I think it would be more efficient in > > > the patch below to memmove the entries instead of put_freelist'ing them > > > individually. > > > > > > If the agfl is completely full then fix_freelist ought to trim it down > > > by at least one element. The algorithm then becomes: > > > > > > if mounting and flfirst < fllast: > > > return 0 > > > if flcount == agfl_size: > > > assert !mounting > > > fix_freelist() > > > > Not sure I follow where the fix_freelist() came in, but I think we need > > to be careful here. What if flfirst == 118 and that slot is garbage? > > Don't we need to fix up the agfl before we can allow any traditional > > operations to proceed? > > This got confusing, so you might notice in the v2 patch that I separated > this out... > > > > assert flcount < agfl_size > > > if flfirst < fllast: > > > return 0 > > > movelen = agfl_size - flfirst > > > if active == flcount - 1: > > > movelen-- > > > memmove(&agflbno[fllast + 1], &agflbno[flfirst], movelen) > > > flfirst = 0 > > > fllast = fllast + movelen > > > > > ... > > > > > > > > > > > > So we're not attempting to cover the case where the agfl has 1 more > > > > > > block than the agfl size (i.e., the case where an fs goes back to a > > > > > > kernel with an unpacked header)? > > > > > > > > > > We don't know how the next kernel to touch this filesystem will define > > > > > XFS_AGFL_SIZE -- it could be a 4.5+ kernel (same agfl size), a 32-bit > > > > > pre-4.5 kernel (same agfl size), or a 64-bit pre-4.5 kernel (small agfl > > > > > size). > > > > > > > > > > > > > I don't think I was clear.. I'm envisioning whether we could come up > > > > with a patch that would generically fix up the agfl on disk to be sane > > > > relative to the current kernel. This patch covers the case of a packed > > > > agfl kernel mounting an unpacked on-disk agfl. It would be nice if we > > > > could implement something that also handled a packed on-disk agfl to an > > > > unpacked agfl kernel (for easier backport to unpacked kernels, for > > > > e.g.). > > > > > > If we're going to touch an old kernel's source at all I'd rather we > > > backport both the packing fix and this fixer-upper. > > > > > > > Not sure I parse... the "old kernel" is essentially the rhel example > > where we apparently have deliberately maintained the unpacked format to > > avoid this incompatibility problem. If we had a patch that generically > > converted on-disk format (packed or unpacked) to the current kernel > > (packed or unpacked), we could merge that patch to upstream, stable as > > well as distro kernels that might not include the agfl packing fix and > > eliminate compatibility problems between them (even if the packing fix > > comes in out of order). > > ...I wonder if we've gotten to the 'talking past each other' point, but > let me try one more time. > > Going forward, I want the number of unpacked kernels to decrease as > quickly as possible. I understand that distro kernel maintainers are > not willing to apply the packing patch to their kernel until we come up > with a smooth transition path. > I agree wrt to upstream, but note that I don't anticipate including the padding fix downstream any time soon. > I don't want to support fixing agfls to be 118 units long on 64-bit > unpacked kernels and 119 units long on 32-bit unpacked kernels, and I > only want to support the packed kernels with their 119 unit long agfls. > An AGFL that starts at 0 and ends at flcount-1 is compatible with packed > and unpacked kernels, so the v2 patch I sent last night removes the > delicate per-case surgery in favor of a new strategy where the mount > time and unmount time helpers both look for agfl configurationss that > are known to cause problems, and solves them all with the same solution: > moving the agfl list towards the start of the block. > The purpose of the patch I sent was not for upstream unpacked support going forward. Upstream has clearly moved forward with the packed format. The goal of the patch was to explore a single/generic patch that could be merged upstream/downstream and handle compatibility cleanly. IOW, I considered what I thought will be necessary to handle compatibility downstream and attempted to incorporate that with an upstream fix because then the patch is straightforward and safe to backport (upstream acceptance mostly justifies inclusion downstream). Then it is possible to create a set of downstream systems, going forward, that are explicitly 100% upstream compatible. So upstream is technically fine without a packed -> unpacked variant, to suggest otherwise wasn't really the intent. Perhaps a BUILD_BUG_ON() or some such that asserts a packed agfl would be effective to prevent blind backporting to unpacked upstream kernels (and rhel could just drop that bit) and would have documented that more clearly. > The fixup patch requires that the kernel already include the packing > fix. Its goals are: > > a) At mount time, ensure that all AGFLs which are not compatible with a > packed kernel are made compatible with a packed kernel. > > b) At unmount time, ensure all AGFLs are compatible with packed kernels > and unpacked kernels. > > Therefore, the first step for a kernel maintainer is to ensure that the > packing fix is applied (which it will be for any 4.5+ kernel). The > second step is to apply this fixer patch. > I'm pretty clear on what the patch does at this point. ;) I agree on goal A in principle. It is generally sane for newer kernels to expect weird/broken formats that could be caused by older kernels and rectify them. There is a convenience value for users upgrading from broken kernels. As noted above, the purpose of my patch was to try and further genericize the implementation to make it useful for downstream (and perhaps more simple/less code). I don't see the value in B. The approach is not 100% effective so it doesn't allow those who have to support unpacked kernels to support any use case that they couldn't before. It seems to be motivated by the existence of essentially stale[1] stable kernels, but we may have already released more[2] kernels since v4.5 that would never get the unmount fix than we had unpacked kernel releases in the first place. Those newer stale kernels can continue to break the older stale kernels regardless of the existence of B. So in other words, the only way the B variant of problems truly disappear is for those stale/unpacked kernel users to upgrade. As such, B mostly seems like a placebo, requres 2-3x more code than is necessary for A (particularly if we consider an on-demand fixup to elminate the need for a mount time scan) and as a result I'm having a hard time justifying doing anything at all here so long as these fixes remain artificially squashed together. At the end of the day it's not my decision but that's the only remaining feedback I can offer on the v2 patch (having skimmed it). Brian [1] Unless I'm mistaken, anybody still on a stale/stable unpacked upstream kernel should have a minor release update readily available that includes the padding fix. [2] AFAICT, the same logic around the stale unpacked kernels applies to future stable non-unmount-fix-capable kernels were we to take this approach. > --D > > > Otherwise, we need a separate unpacked -> packed fixup for packed > > kernels (i.e., this patch) and a packed -> unpacked fixup for unpacked > > kernels and to make sure they are used in the right places. Trying to > > see if we could avoid this kind of dependency matrix was one of the > > objectives around the hack I posted previously. I'm not married to that > > particular implementation, but I'm much less concerned about > > inefficiency (even if it dictates a mount time fixup over a runtime one) > > in comparison to simplicity and flexibility. Perhaps we can accomplish > > something similarly flexible via direct buffer manipulation..? > > > > FWIW, I've appended another variant of the previous hack that is less > > brute force, but I think is still able to convert back and forth. The > > tradeoff is essentially that it no longer uses the same generic > > algorithm. > > > > Brian > > > > --- 8< --- > > > > static int > > xfs_agfl_ondisk_size( > > struct xfs_mount *mp, > > int first, > > int last, > > int flcount) > > { > > int active; > > int size; > > int agfl_size = XFS_AGFL_SIZE(mp); > > > > if (last >= first) > > active = last - first + 1; > > else > > active = agfl_size - first + last + 1; > > > > if (active == flcount + 1) > > size = agfl_size - 1; > > else if ((active == flcount - 1) || > > first == agfl_size || last == agfl_size) > > size = agfl_size + 1; > > else if (active == flcount) > > size = agfl_size; > > else > > return -EFSCORRUPTED; > > > > return size; > > } > > > > int > > xfs_agfl_fixup( > > struct xfs_trans *tp, > > struct xfs_buf *agbp, > > struct xfs_perag *pag) > > { > > struct xfs_mount *mp = tp->t_mountp; > > struct xfs_agf *agf = XFS_BUF_TO_AGF(agbp); > > int agfl_size = XFS_AGFL_SIZE(mp); > > int ofirst, olast, osize; > > int nfirst, nlast; > > struct xfs_buf *agflbp; > > __be32 *agfl_bno; > > xfs_agblock_t bno = -1; > > int tidx = -1; > > bool empty = false; > > int logflags = 0; > > int error; > > > > ofirst = nfirst = be32_to_cpu(agf->agf_flfirst); > > olast = nlast = be32_to_cpu(agf->agf_fllast); > > osize = xfs_agfl_ondisk_size(mp, ofirst, olast, pag->pagf_flcount); > > if (osize < 0) > > return osize; > > if (pag->pagf_flcount == 0) > > empty = true; > > > > /* sizes match, nothing to do */ > > if (osize == agfl_size) > > return 0; > > > > /* size mismatch, read the agfl.. */ > > error = xfs_alloc_read_agfl(mp, tp, be32_to_cpu(agf->agf_seqno), > > &agflbp); > > if (error) > > return error; > > agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); > > > > /* > > * If the on-disk agfl is smaller than what the kernel expects, the last > > * slot of the on-disk agfl is a gap with bogus data. Allocate the first > > * valid block from the agfl, manually place it in the gap and fix up > > * the count. > > */ > > if (osize < agfl_size) { > > ASSERT(!empty); > > error = xfs_alloc_get_freelist(tp, agbp, &bno, 0); > > if (error) > > goto out_relse; > > > > pag->pagf_flcount++; > > be32_add_cpu(&agf->agf_flcount, 1); > > logflags |= XFS_AGF_FLCOUNT; > > tidx = agfl_size - 1; > > goto done; > > } > > > > /* > > * Otherwise, the on-disk agfl is larger than what the current kernel > > * can manage. If the agfl was empty, we just fix up the first and last > > * pointers. If not, move the inaccessible block in the last slot to the > > * next valid, open slot. > > */ > > nfirst = do_mod(nfirst, agfl_size); > > if (empty) { > > nlast = (nfirst == 0 ? agfl_size - 1 : nfirst - 1); > > goto done; > > } > > if (nlast != agfl_size) > > nlast++; > > nlast = do_mod(nlast, agfl_size); > > tidx = nlast; > > bno = be32_to_cpu(agfl_bno[osize - 1]); > > > > done: > > if (nfirst != ofirst) { > > agf->agf_flfirst = cpu_to_be32(nfirst); > > logflags |= XFS_AGF_FLFIRST; > > } > > if (nlast != olast) { > > agf->agf_fllast = cpu_to_be32(nlast); > > logflags |= XFS_AGF_FLLAST; > > } > > if (bno != -1) { > > int startoff; > > > > agfl_bno[tidx] = cpu_to_be32(bno); > > xfs_trans_buf_set_type(tp, agflbp, XFS_BLFT_AGFL_BUF); > > startoff = (char *) &agfl_bno[tidx] - (char *) agflbp->b_addr; > > xfs_trans_log_buf(tp, agflbp, startoff, > > startoff + sizeof(xfs_agblock_t) - 1); > > } > > if (logflags) > > xfs_alloc_log_agf(tp, agbp, logflags); > > > > out_relse: > > xfs_trans_brelse(tp, agflbp); > > return error; > > } > > > -- > To unsubscribe from this list: send the line "unsubscribe linux-xfs" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/fs/xfs/Makefile b/fs/xfs/Makefile index b03c77e..f88368a 100644 --- a/fs/xfs/Makefile +++ b/fs/xfs/Makefile @@ -86,6 +86,7 @@ xfs-y += xfs_aops.o \ xfs_extent_busy.o \ xfs_file.o \ xfs_filestream.o \ + xfs_fixups.o \ xfs_fsmap.o \ xfs_fsops.o \ xfs_globals.o \ diff --git a/fs/xfs/xfs_fixups.c b/fs/xfs/xfs_fixups.c new file mode 100644 index 0000000..0cad7bb --- /dev/null +++ b/fs/xfs/xfs_fixups.c @@ -0,0 +1,310 @@ +/* + * Copyright (C) 2018 Oracle. All Rights Reserved. + * + * Author: Darrick J. Wong <darrick.wong@oracle.com> + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation; either version 2 + * of the License, or (at your option) any later version. + * + * This program is distributed in the hope that it would be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write the Free Software Foundation, + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. + */ +#include "xfs.h" +#include "xfs_fs.h" +#include "xfs_shared.h" +#include "xfs_format.h" +#include "xfs_log_format.h" +#include "xfs_trans_resv.h" +#include "xfs_sb.h" +#include "xfs_mount.h" +#include "xfs_alloc.h" +#include "xfs_trans.h" +#include "xfs_fixups.h" + +/* + * v5 AGFL padding defects + * + * When the v5 format was first introduced, there was a defect in the struct + * xfs_agfl definition that resulted in XFS_AGFL_SIZE returning different + * values depending on the compiler padding. On a fs with 512-byte sectors, + * this meant that XFS_AGFL_SIZE was 119 on i386, but 118 on x64. Commit + * 96f859d52bcb1 ("libxfs: pack the agfl header structure so XFS_AGFL_SIZE is + * correct") changed the definition to disable padding the end of the + * structure, and was accepted into Linux 4.5. Since then, the AGFL has + * always used the larger size (e.g. 119 entries on a 512b sector fs). + * + * Unfortunately, pre-4.5 kernels can produce filesystems with AGFLs that wrap + * at the smaller size, and those kernels are not prepared to handle the + * longer size. This typically manifests itself as an AGF verifier corruption + * error followed by a filesystem shutdown. While we encourage admins to stay + * current with software, we would like to avoid this intermittent breakage. + * + * Any v5 filesystem which has a feature bit set for a feature that was + * introduced after Linux 4.5 will not have this problem, as such kernels + * cannot be mounted on older kernels. v4 filesystems are also unaffected. + * + * Therefore, we add two fixup functions -- the first runs at mount time to + * detect a short-wrapped AGFL and fix it; the second runs at unmount, freeze, + * or remount-ro time to move a wrapped AGFL to the beginning of the list. + * This reduces the likelihood of a screwup to the scenario where you have (a) + * a filesystem with no post-4.5 features (reflink, rmap), (b) the AGFL wraps, + * (c) the filesystem goes down leaving a dirty log, and (d) the dirty + * filesystem is mounted on an old kernel. + */ + +/* + * Decide if we need to have the agfl wrapping fixes applied. This only + * affects v5 filesystems that do not have any features enabled that did not + * exist when the agfl padding fix went in. + * + * Features already present when the fix went in were finobt, ftype, spinodes. + * If we see something new (e.g. reflink) then don't bother. + */ +#define XFS_SB_FEAT_RO_COMPAT_AGFL_WRAP_ALREADY_FIXED \ + (~(XFS_SB_FEAT_RO_COMPAT_FINOBT)) +#define XFS_SB_FEAT_INCOMPAT_AGFL_WRAP_ALREADY_FIXED \ + (~(XFS_SB_FEAT_INCOMPAT_FTYPE | \ + XFS_SB_FEAT_INCOMPAT_SPINODES)) +#define XFS_SB_FEAT_INCOMPAT_LOG_AGFL_WRAP_ALREADY_FIXED \ + (~0) +static inline bool xfs_sb_version_needs_agfl_wrap_fixes(struct xfs_sb *sbp) +{ + return xfs_sb_version_hascrc(sbp) && + !xfs_sb_has_incompat_feature(sbp, + XFS_SB_FEAT_INCOMPAT_AGFL_WRAP_ALREADY_FIXED) && + !xfs_sb_has_ro_compat_feature(sbp, + XFS_SB_FEAT_RO_COMPAT_AGFL_WRAP_ALREADY_FIXED) && + !xfs_sb_has_incompat_log_feature(sbp, + XFS_SB_FEAT_INCOMPAT_LOG_AGFL_WRAP_ALREADY_FIXED); +} + +/* + * Fix an AGFL wrapping that falls short of the end of the block by filling the + * gap at the end of the block. + */ +STATIC int +xfs_fixup_freelist_wrap_mount( + struct xfs_trans *tp, + struct xfs_buf *agfbp, + struct xfs_perag *pag) +{ + struct xfs_mount *mp = tp->t_mountp; + struct xfs_agf *agf; + struct xfs_buf *agflbp; + __be32 *agfl_bno; + xfs_agnumber_t agno; + uint32_t agfl_size; + uint32_t flfirst; + uint32_t fllast; + int32_t active; + int offset; + int len; + int error; + + if (pag->pagf_flcount == 0) + return 0; + + agfl_size = xfs_agfl_size(mp); + agf = XFS_BUF_TO_AGF(agfbp); + agno = be32_to_cpu(agf->agf_seqno); + flfirst = be32_to_cpu(agf->agf_flfirst); + fllast = be32_to_cpu(agf->agf_fllast); + + /* Make sure we're either spot on or off by 1. */ + active = fllast - flfirst + 1; + if (active <= 0) + active += agfl_size; + if (active == pag->pagf_flcount) + return 0; + else if (active != pag->pagf_flcount + 1) + return -EFSCORRUPTED; + + /* Would this have even passed muster on an old system? */ + if (flfirst >= agfl_size - 1 || fllast >= agfl_size - 1 || + pag->pagf_flcount > agfl_size - 1) + return -EFSCORRUPTED; + + /* + * Convert a 40-byte-padded agfl into a 36-byte-padded AGFL. + * Therefore, we need to move the AGFL blocks + * bno[flfirst..agfl_size - 2] to bno[flfirst + 1...agfl_size - 1]. + * + * Reusing the example above, if we had flfirst == 116, we need + * to move bno[116] and bno[117] into bno[117] and bno[118], + * respectively, and then increment flfirst. + */ + error = xfs_alloc_read_agfl(mp, tp, agno, &agflbp); + if (error) + return error; + agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); + + len = (agfl_size - flfirst - 1) * sizeof(xfs_agblock_t); + memmove(&agfl_bno[flfirst + 1], &agfl_bno[flfirst], len); + offset = (char *)&agfl_bno[flfirst + 1] - (char *)agflbp->b_addr; + be32_add_cpu(&agf->agf_flfirst, 1); + + xfs_trans_buf_set_type(tp, agflbp, XFS_BLFT_AGFL_BUF); + xfs_trans_log_buf(tp, agflbp, offset, offset + len - 1); + xfs_trans_brelse(tp, agflbp); + agflbp = NULL; + xfs_alloc_log_agf(tp, agfbp, XFS_AGF_FLFIRST); + + return 0; +} + +/* + * Fix an AGFL that touches the end of the block by moving the first or last + * part of the list elsewhere in the AGFL so that old kernels don't trip over + * wrapping issues. + */ +STATIC int +xfs_fixup_freelist_wrap_unmount( + struct xfs_trans *tp, + struct xfs_buf *agfbp, + struct xfs_perag *pag) +{ + struct xfs_mount *mp = tp->t_mountp; + struct xfs_agf *agf; + struct xfs_buf *agflbp; + __be32 *agfl_bno; + xfs_agnumber_t agno; + uint32_t agfl_size; + uint32_t flfirst; + uint32_t fllast; + int offset; + int len; + int error; + + agfl_size = xfs_agfl_size(mp); + agf = XFS_BUF_TO_AGF(agfbp); + agno = be32_to_cpu(agf->agf_seqno); + flfirst = be32_to_cpu(agf->agf_flfirst); + fllast = be32_to_cpu(agf->agf_fllast); + + /* Empty AGFL? Make sure we aren't pointing at the end. */ + if (pag->pagf_flcount == 0) { + if (flfirst >= agfl_size || fllast >= agfl_size) { + agf->agf_flfirst = cpu_to_be32(1); + agf->agf_fllast = 0; + xfs_alloc_log_agf(tp, agfbp, + XFS_AGF_FLFIRST | XFS_AGF_FLLAST); + } + return 0; + } + + /* If we don't hit the end, we're done. */ + if (flfirst < fllast && fllast != agfl_size - 1) + return 0; + + /* + * Move a start of a wrapped list towards the start of the agfl block. + * Therefore, we need to move the AGFL blocks + * bno[flfirst..agfl_size - 1] to bno[fllast + 1...agfl_size - flfirst]. + * Then we reset flfirst and fllast appropriately. + * + * Reusing the example above, if we had flfirst == 117 and fllast == 4, + * we need to move bno[117] and bno[118] into bno[5] and bno[6], + * respectively, and then reset flfirst and fllast. + * + * If it's just the last block that touches the end, only move that. + */ + error = xfs_alloc_read_agfl(mp, tp, agno, &agflbp); + if (error) + return error; + agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); + + if (fllast == agfl_size - 1) { + /* Back the AGFL off from the end of the block. */ + len = sizeof(xfs_agblock_t); + agfl_bno[flfirst - 1] = agfl_bno[agfl_size - 1]; + offset = (char *)&agfl_bno[flfirst - 1] - (char *)agflbp->b_addr; + be32_add_cpu(&agf->agf_fllast, -1); + be32_add_cpu(&agf->agf_flfirst, -1); + } else { + /* Move the first part of the AGFL towards the front. */ + len = (agfl_size - flfirst) * sizeof(xfs_agblock_t); + memcpy(&agfl_bno[fllast + 1], &agfl_bno[flfirst], len); + offset = (char *)&agfl_bno[fllast + 1] - (char *)agflbp->b_addr; + agf->agf_flfirst = 0; + agf->agf_fllast = cpu_to_be32(pag->pagf_flcount - 1); + } + + xfs_trans_buf_set_type(tp, agflbp, XFS_BLFT_AGFL_BUF); + xfs_trans_log_buf(tp, agflbp, offset, offset + len - 1); + xfs_trans_brelse(tp, agflbp); + agflbp = NULL; + xfs_alloc_log_agf(tp, agfbp, XFS_AGF_FLFIRST | XFS_AGF_FLLAST); + + return 0; +} + +typedef int (*xfs_agf_apply_fn_t)(struct xfs_trans *tp, struct xfs_buf *agfbp, + struct xfs_perag *pag); + +/* Apply something to every AGF. */ +STATIC int +xfs_fixup_agf_apply( + struct xfs_mount *mp, + xfs_agf_apply_fn_t fn) +{ + struct xfs_trans *tp; + struct xfs_perag *pag; + struct xfs_buf *agfbp; + xfs_agnumber_t agno; + int error; + + error = xfs_trans_alloc(mp, &M_RES(mp)->tr_growdata, 0, 0, 0, &tp); + if (error) + return error; + + for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) { + error = xfs_alloc_read_agf(mp, tp, agno, 0, &agfbp); + if (error) + goto cancel; + if (!agfbp) { + error = -ENOMEM; + goto cancel; + } + pag = xfs_perag_get(mp, agno); + error = fn(tp, agfbp, pag); + xfs_perag_put(pag); + xfs_trans_brelse(tp, agfbp); + if (error) + goto cancel; + } + + return xfs_trans_commit(tp); +cancel: + xfs_trans_cancel(tp); + return error; +} + +/* Fix AGFL wrapping so we can use the filesystem. */ +int +xfs_fixup_agfl_wrap_mount( + struct xfs_mount *mp) +{ + if (!xfs_sb_version_needs_agfl_wrap_fixes(&mp->m_sb)) + return 0; + + return xfs_fixup_agf_apply(mp, xfs_fixup_freelist_wrap_mount); +} + +/* Fix AGFL wrapping so old kernels can use this filesystem. */ +int +xfs_fixup_agfl_wrap_unmount( + struct xfs_mount *mp) +{ + if (!xfs_sb_version_needs_agfl_wrap_fixes(&mp->m_sb)) + return 0; + + return xfs_fixup_agf_apply(mp, xfs_fixup_freelist_wrap_unmount); +} diff --git a/fs/xfs/xfs_fixups.h b/fs/xfs/xfs_fixups.h new file mode 100644 index 0000000..fb52a96 --- /dev/null +++ b/fs/xfs/xfs_fixups.h @@ -0,0 +1,26 @@ +/* + * Copyright (C) 2018 Oracle. All Rights Reserved. + * + * Author: Darrick J. Wong <darrick.wong@oracle.com> + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation; either version 2 + * of the License, or (at your option) any later version. + * + * This program is distributed in the hope that it would be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write the Free Software Foundation, + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. + */ +#ifndef __XFS_FIXUPS_H__ +#define __XFS_FIXUPS_H__ + +int xfs_fixup_agfl_wrap_mount(struct xfs_mount *mp); +int xfs_fixup_agfl_wrap_unmount(struct xfs_mount *mp); + +#endif /* __XFS_FIXUPS_H__ */ diff --git a/fs/xfs/xfs_mount.c b/fs/xfs/xfs_mount.c index 98fd41c..eb284aa 100644 --- a/fs/xfs/xfs_mount.c +++ b/fs/xfs/xfs_mount.c @@ -46,7 +46,7 @@ #include "xfs_refcount_btree.h" #include "xfs_reflink.h" #include "xfs_extent_busy.h" - +#include "xfs_fixups.h" static DEFINE_MUTEX(xfs_uuid_table_mutex); static int xfs_uuid_table_size; @@ -875,6 +875,16 @@ xfs_mountfs( } /* + * Make sure our AGFL counters do not wrap the end of the block + * in a troublesome manner. + */ + error = xfs_fixup_agfl_wrap_mount(mp); + if (error) { + xfs_warn(mp, "Failed to fix agfl wrapping. Run xfs_repair."); + goto out_log_dealloc; + } + + /* * Get and sanity-check the root inode. * Save the pointer to it in the mount structure. */ @@ -1128,6 +1138,15 @@ xfs_unmountfs( xfs_qm_unmount(mp); /* + * Make sure our AGFL counters do not wrap the end of the block + * in a troublesome manner for old kernels. + */ + error = xfs_fixup_agfl_wrap_unmount(mp); + if (error) + xfs_warn(mp, "Unable to fix agfl wrapping. " + "This may cause problems on next mount."); + + /* * Unreserve any blocks we have so that when we unmount we don't account * the reserved free space as used. This is really only necessary for * lazy superblock counting because it trusts the incore superblock diff --git a/fs/xfs/xfs_super.c b/fs/xfs/xfs_super.c index 624a802..d9aa39a 100644 --- a/fs/xfs/xfs_super.c +++ b/fs/xfs/xfs_super.c @@ -50,6 +50,7 @@ #include "xfs_refcount_item.h" #include "xfs_bmap_item.h" #include "xfs_reflink.h" +#include "xfs_fixups.h" #include <linux/namei.h> #include <linux/dax.h> @@ -1206,6 +1207,15 @@ xfs_quiesce_attr( xfs_reclaim_inodes(mp, 0); xfs_reclaim_inodes(mp, SYNC_WAIT); + /* + * Make sure our AGFL counters do not wrap the end of the block + * in a troublesome manner for old kernels. + */ + error = xfs_fixup_agfl_wrap_unmount(mp); + if (error) + xfs_warn(mp, "Unable to fix agfl wrapping. " + "This may cause problems on next mount."); + /* Push the superblock and write an unmount record */ error = xfs_log_sbcount(mp); if (error)