]> git.ipfire.org Git - thirdparty/xfsprogs-dev.git/commitdiff
xfs_repair: rebuild the realtime refcount btree
authorDarrick J. Wong <djwong@kernel.org>
Mon, 24 Feb 2025 18:22:06 +0000 (10:22 -0800)
committerDarrick J. Wong <djwong@kernel.org>
Tue, 25 Feb 2025 17:16:02 +0000 (09:16 -0800)
Use the collected reference count information to rebuild the btree.

Signed-off-by: "Darrick J. Wong" <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
libxfs/libxfs_api_defs.h
repair/Makefile
repair/agbtree.c
repair/phase5.c
repair/phase6.c
repair/rmap.c
repair/rmap.h
repair/rtrefcount_repair.c [new file with mode: 0644]

index b5a39856bc1e80e0292cfa97ed5958ae2b895b0e..66cbb34f05a48f7a5cecaf32e552229cda0ab1ce 100644 (file)
 #define xfs_rtgroup_get                        libxfs_rtgroup_get
 #define xfs_rtgroup_put                        libxfs_rtgroup_put
 
+#define xfs_rtrefcountbt_absolute_maxlevels    libxfs_rtrefcountbt_absolute_maxlevels
 #define xfs_rtrefcountbt_calc_reserves libxfs_rtrefcountbt_calc_reserves
+#define xfs_rtrefcountbt_calc_size             libxfs_rtrefcountbt_calc_size
+#define xfs_rtrefcountbt_commit_staged_btree   libxfs_rtrefcountbt_commit_staged_btree
+#define xfs_rtrefcountbt_create                libxfs_rtrefcountbt_create
 #define xfs_rtrefcountbt_droot_maxrecs libxfs_rtrefcountbt_droot_maxrecs
 #define xfs_rtrefcountbt_init_cursor   libxfs_rtrefcountbt_init_cursor
 #define xfs_rtrefcountbt_maxlevels_ondisk      libxfs_rtrefcountbt_maxlevels_ondisk
 #define xfs_rtrefcountbt_maxrecs       libxfs_rtrefcountbt_maxrecs
+#define xfs_rtrefcountbt_stage_cursor  libxfs_rtrefcountbt_stage_cursor
 
 #define xfs_rtrmapbt_calc_reserves     libxfs_rtrmapbt_calc_reserves
 #define xfs_rtrmapbt_calc_size         libxfs_rtrmapbt_calc_size
index 6f4ec3b3a9c4dc2000dab368647b702d3d05f252..ff5b1f5abedac01f548985d2ba605a1b5c02b0d1 100644 (file)
@@ -73,6 +73,7 @@ CFILES = \
        rcbag.c \
        rmap.c \
        rt.c \
+       rtrefcount_repair.c \
        rtrmap_repair.c \
        sb.c \
        scan.c \
index 01066130767cb67e504c9c0f347a3e81d6ba5a8c..983b645e1a35a3b77ebfa9593f2497e2411b5b6a 100644 (file)
@@ -730,7 +730,7 @@ init_refc_cursor(
 
        /* Compute how many blocks we'll need. */
        error = -libxfs_btree_bload_compute_geometry(btr->cur, &btr->bload,
-                       refcount_record_count(sc->mp, agno));
+                       refcount_record_count(sc->mp, false, agno));
        if (error)
                do_error(
 _("Unable to compute refcount btree geometry, error %d.\n"), error);
index cacaf74dda3a601000c2b68eee9f8e6494fd5bb7..4cf28d8ae1a2503212ffac0acf8b8e344a917507 100644 (file)
@@ -705,7 +705,7 @@ are_packed_btrees_needed(
         * If we don't have inode-based metadata, we can let the AG btrees
         * pack as needed; there are no global space concerns here.
         */
-       if (!xfs_has_rtrmapbt(mp))
+       if (!xfs_has_rtrmapbt(mp) && !xfs_has_rtreflink(mp))
                return false;
 
        while ((pag = xfs_perag_next(mp, pag))) {
@@ -716,8 +716,10 @@ are_packed_btrees_needed(
                fdblocks += ag_fdblocks;
        }
 
-       while ((rtg = xfs_rtgroup_next(mp, rtg)))
+       while ((rtg = xfs_rtgroup_next(mp, rtg))) {
                metadata_blocks += estimate_rtrmapbt_blocks(rtg);
+               metadata_blocks += estimate_rtrefcountbt_blocks(rtg);
+       }
 
        /*
         * If we think we'll have more metadata blocks than free space, then
index 30ea19fda9fd875d136671d31d385c744f07b485..4064a84b24509f4899f2174c19e95dddf09e09bf 100644 (file)
@@ -696,6 +696,15 @@ ensure_rtgroup_rmapbt(
                populate_rtgroup_rmapbt(rtg, est_fdblocks);
 }
 
+static void
+ensure_rtgroup_refcountbt(
+       struct xfs_rtgroup      *rtg,
+       xfs_filblks_t           est_fdblocks)
+{
+       if (ensure_rtgroup_file(rtg, XFS_RTGI_REFCOUNT))
+               populate_rtgroup_refcountbt(rtg, est_fdblocks);
+}
+
 /* Initialize a root directory. */
 static int
 init_fs_root_dir(
@@ -3405,8 +3414,10 @@ reset_rt_metadir_inodes(
         * maximally.
         */
        if (!need_packed_btrees) {
-               while ((rtg = xfs_rtgroup_next(mp, rtg)))
+               while ((rtg = xfs_rtgroup_next(mp, rtg))) {
                        metadata_blocks += estimate_rtrmapbt_blocks(rtg);
+                       metadata_blocks += estimate_rtrefcountbt_blocks(rtg);
+               }
 
                if (mp->m_sb.sb_fdblocks > metadata_blocks)
                        est_fdblocks = mp->m_sb.sb_fdblocks - metadata_blocks;
@@ -3427,6 +3438,7 @@ _("        - resetting contents of realtime bitmap and summary inodes\n"));
                ensure_rtgroup_bitmap(rtg);
                ensure_rtgroup_summary(rtg);
                ensure_rtgroup_rmapbt(rtg, est_fdblocks);
+               ensure_rtgroup_refcountbt(rtg, est_fdblocks);
        }
 }
 
index 638e5ea92278cb0ac8c6bca3b4c5868a54f142c8..97510dd875911abb098fccde155483c5dc4a1666 100644 (file)
@@ -1682,9 +1682,10 @@ _("Unable to fix reflink flag on inode %"PRIu64".\n"),
 uint64_t
 refcount_record_count(
        struct xfs_mount        *mp,
+       bool                    isrt,
        xfs_agnumber_t          agno)
 {
-       struct xfs_ag_rmap      *x = rmaps_for_group(false, agno);
+       struct xfs_ag_rmap      *x = rmaps_for_group(isrt, agno);
 
        return slab_count(x->ar_refcount_items);
 }
@@ -2081,3 +2082,22 @@ estimate_rtrmapbt_blocks(
        nr_recs = xmbuf_bytes(x->ar_xmbtp) / sizeof(struct xfs_rmap_rec);
        return libxfs_rtrmapbt_calc_size(mp, nr_recs);
 }
+
+/* Estimate the size of the ondisk rtrefcountbt from the incore data. */
+xfs_filblks_t
+estimate_rtrefcountbt_blocks(
+       struct xfs_rtgroup      *rtg)
+{
+       struct xfs_mount        *mp = rtg_mount(rtg);
+       struct xfs_ag_rmap      *x;
+
+       if (!rmap_needs_work(mp) || !xfs_has_rtreflink(mp))
+               return 0;
+
+       x = &rg_rmaps[rtg_rgno(rtg)];
+       if (!x->ar_refcount_items)
+               return 0;
+
+       return libxfs_rtrefcountbt_calc_size(mp,
+                       slab_count(x->ar_refcount_items));
+}
index 80e82a4ac4c008d75b1bd0f2de5c46b05f6b1726..1f234b8be32e721a2a350c4f70e25998933917f5 100644 (file)
@@ -39,7 +39,8 @@ extern void rmap_high_key_from_rec(struct xfs_rmap_irec *rec,
                struct xfs_rmap_irec *key);
 
 int compute_refcounts(struct xfs_mount *mp, bool isrt, xfs_agnumber_t agno);
-uint64_t refcount_record_count(struct xfs_mount *mp, xfs_agnumber_t agno);
+uint64_t refcount_record_count(struct xfs_mount *mp, bool isrt,
+               xfs_agnumber_t agno);
 int init_refcount_cursor(bool isrt, xfs_agnumber_t agno,
                struct xfs_slab_cursor **pcur);
 extern void refcount_avoid_check(struct xfs_mount *mp);
@@ -64,4 +65,8 @@ void populate_rtgroup_rmapbt(struct xfs_rtgroup *rtg,
                xfs_filblks_t est_fdblocks);
 xfs_filblks_t estimate_rtrmapbt_blocks(struct xfs_rtgroup *rtg);
 
+int populate_rtgroup_refcountbt(struct xfs_rtgroup *rtg,
+               xfs_filblks_t est_fdblocks);
+xfs_filblks_t estimate_rtrefcountbt_blocks(struct xfs_rtgroup *rtg);
+
 #endif /* RMAP_H_ */
diff --git a/repair/rtrefcount_repair.c b/repair/rtrefcount_repair.c
new file mode 100644 (file)
index 0000000..228d080
--- /dev/null
@@ -0,0 +1,257 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * Copyright (c) 2021-2025 Oracle.  All Rights Reserved.
+ * Author: Darrick J. Wong <djwong@kernel.org>
+ */
+#include <libxfs.h>
+#include "btree.h"
+#include "err_protos.h"
+#include "libxlog.h"
+#include "incore.h"
+#include "globals.h"
+#include "dinode.h"
+#include "slab.h"
+#include "rmap.h"
+#include "bulkload.h"
+
+/*
+ * Realtime Reference Count (RTREFCBT) Repair
+ * ==========================================
+ *
+ * Gather all the reference count records for the realtime device, reset the
+ * incore fork, then recreate the btree.
+ */
+struct xrep_rtrefc {
+       /* rtrefcbt slab cursor */
+       struct xfs_slab_cursor  *slab_cursor;
+
+       /* New fork. */
+       struct bulkload         new_fork_info;
+       struct xfs_btree_bload  rtrefc_bload;
+
+       struct repair_ctx       *sc;
+       struct xfs_rtgroup      *rtg;
+
+       /* Estimated free space after building all rt btrees */
+       xfs_filblks_t           est_fdblocks;
+};
+
+/* Retrieve rtrefc data for bulk load. */
+STATIC int
+xrep_rtrefc_get_records(
+       struct xfs_btree_cur            *cur,
+       unsigned int                    idx,
+       struct xfs_btree_block          *block,
+       unsigned int                    nr_wanted,
+       void                            *priv)
+{
+       struct xfs_refcount_irec        *rec;
+       struct xrep_rtrefc              *rc = priv;
+       union xfs_btree_rec             *block_rec;
+       unsigned int                    loaded;
+
+       for (loaded = 0; loaded < nr_wanted; loaded++, idx++) {
+               rec = pop_slab_cursor(rc->slab_cursor);
+               memcpy(&cur->bc_rec.rc, rec, sizeof(struct xfs_refcount_irec));
+
+               block_rec = libxfs_btree_rec_addr(cur, idx, block);
+               cur->bc_ops->init_rec_from_cur(cur, block_rec);
+       }
+
+       return loaded;
+}
+
+/* Feed one of the new btree blocks to the bulk loader. */
+STATIC int
+xrep_rtrefc_claim_block(
+       struct xfs_btree_cur    *cur,
+       union xfs_btree_ptr     *ptr,
+       void                    *priv)
+{
+       struct xrep_rtrefc      *rr = priv;
+
+       return bulkload_claim_block(cur, &rr->new_fork_info, ptr);
+}
+
+/* Figure out how much space we need to create the incore btree root block. */
+STATIC size_t
+xrep_rtrefc_iroot_size(
+       struct xfs_btree_cur    *cur,
+       unsigned int            level,
+       unsigned int            nr_this_level,
+       void                    *priv)
+{
+       return xfs_rtrefcount_broot_space_calc(cur->bc_mp, level,
+                       nr_this_level);
+}
+
+/* Reserve new btree blocks and bulk load all the rtrmap records. */
+STATIC int
+xrep_rtrefc_btree_load(
+       struct xrep_rtrefc      *rr,
+       struct xfs_btree_cur    *rtrmap_cur)
+{
+       struct repair_ctx       *sc = rr->sc;
+       int                     error;
+
+       rr->rtrefc_bload.get_records = xrep_rtrefc_get_records;
+       rr->rtrefc_bload.claim_block = xrep_rtrefc_claim_block;
+       rr->rtrefc_bload.iroot_size = xrep_rtrefc_iroot_size;
+       bulkload_estimate_inode_slack(sc->mp, &rr->rtrefc_bload,
+                       rr->est_fdblocks);
+
+       /* Compute how many blocks we'll need. */
+       error = -libxfs_btree_bload_compute_geometry(rtrmap_cur,
+                       &rr->rtrefc_bload,
+                       refcount_record_count(sc->mp, true, rtg_rgno(rr->rtg)));
+       if (error)
+               return error;
+
+       /*
+        * Guess how many blocks we're going to need to rebuild an entire
+        * rtrefcountbt from the number of extents we found, and pump up our
+        * transaction to have sufficient block reservation.
+        */
+       error = -libxfs_trans_reserve_more(sc->tp, rr->rtrefc_bload.nr_blocks,
+                       0);
+       if (error)
+               return error;
+
+       /*
+        * Reserve the space we'll need for the new btree.  Drop the cursor
+        * while we do this because that can roll the transaction and cursors
+        * can't handle that.
+        */
+       error = bulkload_alloc_file_blocks(&rr->new_fork_info,
+                       rr->rtrefc_bload.nr_blocks);
+       if (error)
+               return error;
+
+       /* Add all observed rtrmap records. */
+       error = init_refcount_cursor(true, rtg_rgno(rr->rtg), &rr->slab_cursor);
+       if (error)
+               return error;
+       error = -libxfs_btree_bload(rtrmap_cur, &rr->rtrefc_bload, rr);
+       free_slab_cursor(&rr->slab_cursor);
+       return error;
+}
+
+/* Update the inode counters. */
+STATIC int
+xrep_rtrefc_reset_counters(
+       struct xrep_rtrefc      *rr)
+{
+       struct repair_ctx       *sc = rr->sc;
+
+       /*
+        * Update the inode block counts to reflect the btree we just
+        * generated.
+        */
+       sc->ip->i_nblocks = rr->new_fork_info.ifake.if_blocks;
+       libxfs_trans_log_inode(sc->tp, sc->ip, XFS_ILOG_CORE);
+
+       /* Quotas don't exist so we're done. */
+       return 0;
+}
+
+/*
+ * Use the collected rmap information to stage a new rt refcount btree.  If
+ * this is successful we'll return with the new btree root information logged
+ * to the repair transaction but not yet committed.
+ */
+static int
+xrep_rtrefc_build_new_tree(
+       struct xrep_rtrefc      *rr)
+{
+       struct xfs_owner_info   oinfo;
+       struct xfs_btree_cur    *cur;
+       struct repair_ctx       *sc = rr->sc;
+       struct xbtree_ifakeroot *ifake = &rr->new_fork_info.ifake;
+       int                     error;
+
+       /*
+        * Prepare to construct the new fork by initializing the new btree
+        * structure and creating a fake ifork in the ifakeroot structure.
+        */
+       libxfs_rmap_ino_bmbt_owner(&oinfo, sc->ip->i_ino, XFS_DATA_FORK);
+       bulkload_init_inode(&rr->new_fork_info, sc, XFS_DATA_FORK, &oinfo);
+       cur = libxfs_rtrefcountbt_init_cursor(NULL, rr->rtg);
+       libxfs_btree_stage_ifakeroot(cur, ifake);
+
+       /*
+        * Figure out the size and format of the new fork, then fill it with
+        * all the rtrmap records we've found.  Join the inode to the
+        * transaction so that we can roll the transaction while holding the
+        * inode locked.
+        */
+       libxfs_trans_ijoin(sc->tp, sc->ip, 0);
+       ifake->if_fork->if_format = XFS_DINODE_FMT_META_BTREE;
+       error = xrep_rtrefc_btree_load(rr, cur);
+       if (error)
+               goto err_cur;
+
+       /*
+        * Install the new fork in the inode.  After this point the old mapping
+        * data are no longer accessible and the new tree is live.  We delete
+        * the cursor immediately after committing the staged root because the
+        * staged fork might be in extents format.
+        */
+       libxfs_rtrefcountbt_commit_staged_btree(cur, sc->tp);
+       libxfs_btree_del_cursor(cur, 0);
+
+       /* Reset the inode counters now that we've changed the fork. */
+       error = xrep_rtrefc_reset_counters(rr);
+       if (error)
+               goto err_newbt;
+
+       /* Dispose of any unused blocks and the accounting infomation. */
+       error = bulkload_commit(&rr->new_fork_info);
+       if (error)
+               return error;
+
+       return -libxfs_trans_roll_inode(&sc->tp, sc->ip);
+err_cur:
+       if (cur)
+               libxfs_btree_del_cursor(cur, error);
+err_newbt:
+       bulkload_cancel(&rr->new_fork_info);
+       return error;
+}
+
+/* Store the realtime reference counts in the rtrefcbt. */
+int
+populate_rtgroup_refcountbt(
+       struct xfs_rtgroup      *rtg,
+       xfs_filblks_t           est_fdblocks)
+{
+       struct xfs_mount        *mp = rtg_mount(rtg);
+       struct xfs_inode        *ip = rtg_refcount(rtg);
+       struct repair_ctx       sc = {
+               .mp             = mp,
+               .ip             = ip,
+       };
+       struct xrep_rtrefc      rr = {
+               .sc             = &sc,
+               .rtg            = rtg,
+               .est_fdblocks   = est_fdblocks,
+       };
+       int                     error;
+
+       if (!xfs_has_rtreflink(mp))
+               return 0;
+
+       error = -libxfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0,
+                       &sc.tp);
+       if (error)
+               return error;
+
+       error = xrep_rtrefc_build_new_tree(&rr);
+       if (error)
+               goto out_cancel;
+
+       return -libxfs_trans_commit(sc.tp);
+
+out_cancel:
+       libxfs_trans_cancel(sc.tp);
+       return error;
+}