xfs: fix fdblocks accounting w/ RMAPBT per-AG reservation

In __xfs_ag_resv_init we incorrectly calculate the amount by which to
decrease fdblocks when reserving blocks for the rmapbt.  Because rmapbt
allocations do not decrease fdblocks, we must decrease fdblocks by the
entire size of the requested reservation in order to achieve our goal of
always having enough free blocks to satisfy an rmapbt expansion.

This is in contrast to the refcountbt/finobt, which /do/ subtract from
fdblocks whenever they allocate a block.  For this allocation type we
preserve the existing behavior where we decrease fdblocks only by the
requested reservation minus the size of the existing tree.

This fixes the problem where the available block counts reported by
statfs change across a remount if there had been an rmapbt size change
since mount time.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>

diff --git a/fs/xfs/libxfs/xfs_ag_resv.c b/fs/xfs/libxfs/xfs_ag_resv.c
index 84db76e0e3e3c58ae7d25b38a46a1ce2b4d5cae4..fecd187fcf2c3cd69bd79954ccc272737cb2ce2b 100644 (file)
--- a/fs/xfs/libxfs/xfs_ag_resv.c
+++ b/fs/xfs/libxfs/xfs_ag_resv.c
@@ -157,6 +157,7 @@ __xfs_ag_resv_free(
        error = xfs_mod_fdblocks(pag->pag_mount, oldresv, true);
        resv->ar_reserved = 0;
        resv->ar_asked = 0;
+       resv->ar_orig_reserved = 0;
 
        if (error)
                trace_xfs_ag_resv_free_error(pag->pag_mount, pag->pag_agno,
@@ -189,13 +190,34 @@ __xfs_ag_resv_init(
        struct xfs_mount                *mp = pag->pag_mount;
        struct xfs_ag_resv              *resv;
        int                             error;
-       xfs_extlen_t                    reserved;
+       xfs_extlen_t                    hidden_space;
 
        if (used > ask)
                ask = used;
-       reserved = ask - used;
 
-       error = xfs_mod_fdblocks(mp, -(int64_t)reserved, true);
+       switch (type) {
+       case XFS_AG_RESV_RMAPBT:
+               /*
+                * Space taken by the rmapbt is not subtracted from fdblocks
+                * because the rmapbt lives in the free space.  Here we must
+                * subtract the entire reservation from fdblocks so that we
+                * always have blocks available for rmapbt expansion.
+                */
+               hidden_space = ask;
+               break;
+       case XFS_AG_RESV_METADATA:
+               /*
+                * Space taken by all other metadata btrees are accounted
+                * on-disk as used space.  We therefore only hide the space
+                * that is reserved but not used by the trees.
+                */
+               hidden_space = ask - used;
+               break;
+       default:
+               ASSERT(0);
+               return -EINVAL;
+       }
+       error = xfs_mod_fdblocks(mp, -(int64_t)hidden_space, true);
        if (error) {
                trace_xfs_ag_resv_init_error(pag->pag_mount, pag->pag_agno,
                                error, _RET_IP_);
@@ -216,7 +238,8 @@ __xfs_ag_resv_init(
 
        resv = xfs_perag_resv(pag, type);
        resv->ar_asked = ask;
-       resv->ar_reserved = resv->ar_orig_reserved = reserved;
+       resv->ar_orig_reserved = hidden_space;
+       resv->ar_reserved = ask - used;
 
        trace_xfs_ag_resv_init(pag, type, ask);
        return 0;