xfs: repair the AGFL
authorDarrick J. Wong <darrick.wong@oracle.com>
Fri, 10 Aug 2018 05:43:02 +0000 (22:43 -0700)
committerDarrick J. Wong <darrick.wong@oracle.com>
Fri, 10 Aug 2018 18:44:31 +0000 (11:44 -0700)
Repair the AGFL from the rmap data.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
fs/xfs/scrub/agheader_repair.c
fs/xfs/scrub/bitmap.c
fs/xfs/scrub/bitmap.h
fs/xfs/scrub/repair.h
fs/xfs/scrub/scrub.c

index aa18049..9ce3023 100644 (file)
@@ -433,3 +433,284 @@ out_revert:
        memcpy(agf, &old_agf, sizeof(old_agf));
        return error;
 }
+
+/* AGFL */
+
+struct xrep_agfl {
+       /* Bitmap of other OWN_AG metadata blocks. */
+       struct xfs_bitmap       agmetablocks;
+
+       /* Bitmap of free space. */
+       struct xfs_bitmap       *freesp;
+
+       struct xfs_scrub        *sc;
+};
+
+/* Record all OWN_AG (free space btree) information from the rmap data. */
+STATIC int
+xrep_agfl_walk_rmap(
+       struct xfs_btree_cur    *cur,
+       struct xfs_rmap_irec    *rec,
+       void                    *priv)
+{
+       struct xrep_agfl        *ra = priv;
+       xfs_fsblock_t           fsb;
+       int                     error = 0;
+
+       if (xchk_should_terminate(ra->sc, &error))
+               return error;
+
+       /* Record all the OWN_AG blocks. */
+       if (rec->rm_owner == XFS_RMAP_OWN_AG) {
+               fsb = XFS_AGB_TO_FSB(cur->bc_mp, cur->bc_private.a.agno,
+                               rec->rm_startblock);
+               error = xfs_bitmap_set(ra->freesp, fsb, rec->rm_blockcount);
+               if (error)
+                       return error;
+       }
+
+       return xfs_bitmap_set_btcur_path(&ra->agmetablocks, cur);
+}
+
+/*
+ * Map out all the non-AGFL OWN_AG space in this AG so that we can deduce
+ * which blocks belong to the AGFL.
+ *
+ * Compute the set of old AGFL blocks by subtracting from the list of OWN_AG
+ * blocks the list of blocks owned by all other OWN_AG metadata (bnobt, cntbt,
+ * rmapbt).  These are the old AGFL blocks, so return that list and the number
+ * of blocks we're actually going to put back on the AGFL.
+ */
+STATIC int
+xrep_agfl_collect_blocks(
+       struct xfs_scrub        *sc,
+       struct xfs_buf          *agf_bp,
+       struct xfs_bitmap       *agfl_extents,
+       xfs_agblock_t           *flcount)
+{
+       struct xrep_agfl        ra;
+       struct xfs_mount        *mp = sc->mp;
+       struct xfs_btree_cur    *cur;
+       struct xfs_bitmap_range *br;
+       struct xfs_bitmap_range *n;
+       int                     error;
+
+       ra.sc = sc;
+       ra.freesp = agfl_extents;
+       xfs_bitmap_init(&ra.agmetablocks);
+
+       /* Find all space used by the free space btrees & rmapbt. */
+       cur = xfs_rmapbt_init_cursor(mp, sc->tp, agf_bp, sc->sa.agno);
+       error = xfs_rmap_query_all(cur, xrep_agfl_walk_rmap, &ra);
+       if (error)
+               goto err;
+       xfs_btree_del_cursor(cur, error);
+
+       /* Find all blocks currently being used by the bnobt. */
+       cur = xfs_allocbt_init_cursor(mp, sc->tp, agf_bp, sc->sa.agno,
+                       XFS_BTNUM_BNO);
+       error = xfs_bitmap_set_btblocks(&ra.agmetablocks, cur);
+       if (error)
+               goto err;
+       xfs_btree_del_cursor(cur, error);
+
+       /* Find all blocks currently being used by the cntbt. */
+       cur = xfs_allocbt_init_cursor(mp, sc->tp, agf_bp, sc->sa.agno,
+                       XFS_BTNUM_CNT);
+       error = xfs_bitmap_set_btblocks(&ra.agmetablocks, cur);
+       if (error)
+               goto err;
+
+       xfs_btree_del_cursor(cur, error);
+
+       /*
+        * Drop the freesp meta blocks that are in use by btrees.
+        * The remaining blocks /should/ be AGFL blocks.
+        */
+       error = xfs_bitmap_disunion(agfl_extents, &ra.agmetablocks);
+       xfs_bitmap_destroy(&ra.agmetablocks);
+       if (error)
+               return error;
+
+       /*
+        * Calculate the new AGFL size.  If we found more blocks than fit in
+        * the AGFL we'll free them later.
+        */
+       *flcount = 0;
+       for_each_xfs_bitmap_extent(br, n, agfl_extents) {
+               *flcount += br->len;
+               if (*flcount > xfs_agfl_size(mp))
+                       break;
+       }
+       if (*flcount > xfs_agfl_size(mp))
+               *flcount = xfs_agfl_size(mp);
+       return 0;
+
+err:
+       xfs_bitmap_destroy(&ra.agmetablocks);
+       xfs_btree_del_cursor(cur, error);
+       return error;
+}
+
+/* Update the AGF and reset the in-core state. */
+STATIC void
+xrep_agfl_update_agf(
+       struct xfs_scrub        *sc,
+       struct xfs_buf          *agf_bp,
+       xfs_agblock_t           flcount)
+{
+       struct xfs_agf          *agf = XFS_BUF_TO_AGF(agf_bp);
+
+       ASSERT(flcount <= xfs_agfl_size(sc->mp));
+
+       /* Trigger fdblocks recalculation */
+       xfs_force_summary_recalc(sc->mp);
+
+       /* Update the AGF counters. */
+       if (sc->sa.pag->pagf_init)
+               sc->sa.pag->pagf_flcount = flcount;
+       agf->agf_flfirst = cpu_to_be32(0);
+       agf->agf_flcount = cpu_to_be32(flcount);
+       agf->agf_fllast = cpu_to_be32(flcount - 1);
+
+       xfs_alloc_log_agf(sc->tp, agf_bp,
+                       XFS_AGF_FLFIRST | XFS_AGF_FLLAST | XFS_AGF_FLCOUNT);
+}
+
+/* Write out a totally new AGFL. */
+STATIC void
+xrep_agfl_init_header(
+       struct xfs_scrub        *sc,
+       struct xfs_buf          *agfl_bp,
+       struct xfs_bitmap       *agfl_extents,
+       xfs_agblock_t           flcount)
+{
+       struct xfs_mount        *mp = sc->mp;
+       __be32                  *agfl_bno;
+       struct xfs_bitmap_range *br;
+       struct xfs_bitmap_range *n;
+       struct xfs_agfl         *agfl;
+       xfs_agblock_t           agbno;
+       unsigned int            fl_off;
+
+       ASSERT(flcount <= xfs_agfl_size(mp));
+
+       /*
+        * Start rewriting the header by setting the bno[] array to
+        * NULLAGBLOCK, then setting AGFL header fields.
+        */
+       agfl = XFS_BUF_TO_AGFL(agfl_bp);
+       memset(agfl, 0xFF, BBTOB(agfl_bp->b_length));
+       agfl->agfl_magicnum = cpu_to_be32(XFS_AGFL_MAGIC);
+       agfl->agfl_seqno = cpu_to_be32(sc->sa.agno);
+       uuid_copy(&agfl->agfl_uuid, &mp->m_sb.sb_meta_uuid);
+
+       /*
+        * Fill the AGFL with the remaining blocks.  If agfl_extents has more
+        * blocks than fit in the AGFL, they will be freed in a subsequent
+        * step.
+        */
+       fl_off = 0;
+       agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agfl_bp);
+       for_each_xfs_bitmap_extent(br, n, agfl_extents) {
+               agbno = XFS_FSB_TO_AGBNO(mp, br->start);
+
+               trace_xrep_agfl_insert(mp, sc->sa.agno, agbno, br->len);
+
+               while (br->len > 0 && fl_off < flcount) {
+                       agfl_bno[fl_off] = cpu_to_be32(agbno);
+                       fl_off++;
+                       agbno++;
+
+                       /*
+                        * We've now used br->start by putting it in the AGFL,
+                        * so bump br so that we don't reap the block later.
+                        */
+                       br->start++;
+                       br->len--;
+               }
+
+               if (br->len)
+                       break;
+               list_del(&br->list);
+               kmem_free(br);
+       }
+
+       /* Write new AGFL to disk. */
+       xfs_trans_buf_set_type(sc->tp, agfl_bp, XFS_BLFT_AGFL_BUF);
+       xfs_trans_log_buf(sc->tp, agfl_bp, 0, BBTOB(agfl_bp->b_length) - 1);
+}
+
+/* Repair the AGFL. */
+int
+xrep_agfl(
+       struct xfs_scrub        *sc)
+{
+       struct xfs_owner_info   oinfo;
+       struct xfs_bitmap       agfl_extents;
+       struct xfs_mount        *mp = sc->mp;
+       struct xfs_buf          *agf_bp;
+       struct xfs_buf          *agfl_bp;
+       xfs_agblock_t           flcount;
+       int                     error;
+
+       /* We require the rmapbt to rebuild anything. */
+       if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
+               return -EOPNOTSUPP;
+
+       xchk_perag_get(sc->mp, &sc->sa);
+       xfs_bitmap_init(&agfl_extents);
+
+       /*
+        * Read the AGF so that we can query the rmapbt.  We hope that there's
+        * nothing wrong with the AGF, but all the AG header repair functions
+        * have this chicken-and-egg problem.
+        */
+       error = xfs_alloc_read_agf(mp, sc->tp, sc->sa.agno, 0, &agf_bp);
+       if (error)
+               return error;
+       if (!agf_bp)
+               return -ENOMEM;
+
+       /*
+        * Make sure we have the AGFL buffer, as scrub might have decided it
+        * was corrupt after xfs_alloc_read_agfl failed with -EFSCORRUPTED.
+        */
+       error = xfs_trans_read_buf(mp, sc->tp, mp->m_ddev_targp,
+                       XFS_AG_DADDR(mp, sc->sa.agno, XFS_AGFL_DADDR(mp)),
+                       XFS_FSS_TO_BB(mp, 1), 0, &agfl_bp, NULL);
+       if (error)
+               return error;
+       agfl_bp->b_ops = &xfs_agfl_buf_ops;
+
+       /* Gather all the extents we're going to put on the new AGFL. */
+       error = xrep_agfl_collect_blocks(sc, agf_bp, &agfl_extents, &flcount);
+       if (error)
+               goto err;
+
+       /*
+        * Update AGF and AGFL.  We reset the global free block counter when
+        * we adjust the AGF flcount (which can fail) so avoid updating any
+        * buffers until we know that part works.
+        */
+       xrep_agfl_update_agf(sc, agf_bp, flcount);
+       xrep_agfl_init_header(sc, agfl_bp, &agfl_extents, flcount);
+
+       /*
+        * Ok, the AGFL should be ready to go now.  Roll the transaction to
+        * make the new AGFL permanent before we start using it to return
+        * freespace overflow to the freespace btrees.
+        */
+       sc->sa.agf_bp = agf_bp;
+       sc->sa.agfl_bp = agfl_bp;
+       error = xrep_roll_ag_trans(sc);
+       if (error)
+               goto err;
+
+       /* Dump any AGFL overflow. */
+       xfs_rmap_ag_owner(&oinfo, XFS_RMAP_OWN_AG);
+       return xrep_reap_extents(sc, &agfl_extents, &oinfo, XFS_AG_RESV_AGFL);
+err:
+       xfs_bitmap_destroy(&agfl_extents);
+       return error;
+}
index c770e2d..fdadc9e 100644 (file)
@@ -9,6 +9,7 @@
 #include "xfs_format.h"
 #include "xfs_trans_resv.h"
 #include "xfs_mount.h"
+#include "xfs_btree.h"
 #include "scrub/xfs_scrub.h"
 #include "scrub/scrub.h"
 #include "scrub/common.h"
@@ -209,3 +210,94 @@ out:
 }
 #undef LEFT_ALIGNED
 #undef RIGHT_ALIGNED
+
+/*
+ * Record all btree blocks seen while iterating all records of a btree.
+ *
+ * We know that the btree query_all function starts at the left edge and walks
+ * towards the right edge of the tree.  Therefore, we know that we can walk up
+ * the btree cursor towards the root; if the pointer for a given level points
+ * to the first record/key in that block, we haven't seen this block before;
+ * and therefore we need to remember that we saw this block in the btree.
+ *
+ * So if our btree is:
+ *
+ *    4
+ *  / | \
+ * 1  2  3
+ *
+ * Pretend for this example that each leaf block has 100 btree records.  For
+ * the first btree record, we'll observe that bc_ptrs[0] == 1, so we record
+ * that we saw block 1.  Then we observe that bc_ptrs[1] == 1, so we record
+ * block 4.  The list is [1, 4].
+ *
+ * For the second btree record, we see that bc_ptrs[0] == 2, so we exit the
+ * loop.  The list remains [1, 4].
+ *
+ * For the 101st btree record, we've moved onto leaf block 2.  Now
+ * bc_ptrs[0] == 1 again, so we record that we saw block 2.  We see that
+ * bc_ptrs[1] == 2, so we exit the loop.  The list is now [1, 4, 2].
+ *
+ * For the 102nd record, bc_ptrs[0] == 2, so we continue.
+ *
+ * For the 201st record, we've moved on to leaf block 3.  bc_ptrs[0] == 1, so
+ * we add 3 to the list.  Now it is [1, 4, 2, 3].
+ *
+ * For the 300th record we just exit, with the list being [1, 4, 2, 3].
+ */
+
+/*
+ * Record all the buffers pointed to by the btree cursor.  Callers already
+ * engaged in a btree walk should call this function to capture the list of
+ * blocks going from the leaf towards the root.
+ */
+int
+xfs_bitmap_set_btcur_path(
+       struct xfs_bitmap       *bitmap,
+       struct xfs_btree_cur    *cur)
+{
+       struct xfs_buf          *bp;
+       xfs_fsblock_t           fsb;
+       int                     i;
+       int                     error;
+
+       for (i = 0; i < cur->bc_nlevels && cur->bc_ptrs[i] == 1; i++) {
+               xfs_btree_get_block(cur, i, &bp);
+               if (!bp)
+                       continue;
+               fsb = XFS_DADDR_TO_FSB(cur->bc_mp, bp->b_bn);
+               error = xfs_bitmap_set(bitmap, fsb, 1);
+               if (error)
+                       return error;
+       }
+
+       return 0;
+}
+
+/* Collect a btree's block in the bitmap. */
+STATIC int
+xfs_bitmap_collect_btblock(
+       struct xfs_btree_cur    *cur,
+       int                     level,
+       void                    *priv)
+{
+       struct xfs_bitmap       *bitmap = priv;
+       struct xfs_buf          *bp;
+       xfs_fsblock_t           fsbno;
+
+       xfs_btree_get_block(cur, level, &bp);
+       if (!bp)
+               return 0;
+
+       fsbno = XFS_DADDR_TO_FSB(cur->bc_mp, bp->b_bn);
+       return xfs_bitmap_set(bitmap, fsbno, 1);
+}
+
+/* Walk the btree and mark the bitmap wherever a btree block is found. */
+int
+xfs_bitmap_set_btblocks(
+       struct xfs_bitmap       *bitmap,
+       struct xfs_btree_cur    *cur)
+{
+       return xfs_btree_visit_blocks(cur, xfs_bitmap_collect_btblock, bitmap);
+}
index dad652e..ae8ecbc 100644 (file)
@@ -28,5 +28,9 @@ void xfs_bitmap_destroy(struct xfs_bitmap *bitmap);
 
 int xfs_bitmap_set(struct xfs_bitmap *bitmap, uint64_t start, uint64_t len);
 int xfs_bitmap_disunion(struct xfs_bitmap *bitmap, struct xfs_bitmap *sub);
+int xfs_bitmap_set_btcur_path(struct xfs_bitmap *bitmap,
+               struct xfs_btree_cur *cur);
+int xfs_bitmap_set_btblocks(struct xfs_bitmap *bitmap,
+               struct xfs_btree_cur *cur);
 
 #endif /* __XFS_SCRUB_BITMAP_H__ */
index 6f0903c..1d28336 100644 (file)
@@ -59,6 +59,7 @@ int xrep_ino_dqattach(struct xfs_scrub *sc);
 int xrep_probe(struct xfs_scrub *sc);
 int xrep_superblock(struct xfs_scrub *sc);
 int xrep_agf(struct xfs_scrub *sc);
+int xrep_agfl(struct xfs_scrub *sc);
 
 #else
 
@@ -83,6 +84,7 @@ xrep_calc_ag_resblks(
 #define xrep_probe                     xrep_notsupported
 #define xrep_superblock                        xrep_notsupported
 #define xrep_agf                       xrep_notsupported
+#define xrep_agfl                      xrep_notsupported
 
 #endif /* CONFIG_XFS_ONLINE_REPAIR */
 
index 1e8a17c..2670f4c 100644 (file)
@@ -220,7 +220,7 @@ static const struct xchk_meta_ops meta_scrub_ops[] = {
                .type   = ST_PERAG,
                .setup  = xchk_setup_fs,
                .scrub  = xchk_agfl,
-               .repair = xrep_notsupported,
+               .repair = xrep_agfl,
        },
        [XFS_SCRUB_TYPE_AGI] = {        /* agi */
                .type   = ST_PERAG,