xfs: Remove inode iolock held check during allocation
[safe/jmp/linux-2.6] / fs / xfs / xfs_trans.c
index 2eb6027..237badc 100644 (file)
@@ -24,7 +24,6 @@
 #include "xfs_trans.h"
 #include "xfs_sb.h"
 #include "xfs_ag.h"
-#include "xfs_dir.h"
 #include "xfs_dir2.h"
 #include "xfs_dmapi.h"
 #include "xfs_mount.h"
@@ -33,7 +32,6 @@
 #include "xfs_bmap_btree.h"
 #include "xfs_alloc_btree.h"
 #include "xfs_ialloc_btree.h"
-#include "xfs_dir_sf.h"
 #include "xfs_dir2_sf.h"
 #include "xfs_attr_sf.h"
 #include "xfs_dinode.h"
@@ -45,6 +43,7 @@
 #include "xfs_quota.h"
 #include "xfs_trans_priv.h"
 #include "xfs_trans_space.h"
+#include "xfs_inode_item.h"
 
 
 STATIC void    xfs_trans_apply_sb_deltas(xfs_trans_t *);
@@ -55,10 +54,141 @@ STATIC void        xfs_trans_committed(xfs_trans_t *, int);
 STATIC void    xfs_trans_chunk_committed(xfs_log_item_chunk_t *, xfs_lsn_t, int);
 STATIC void    xfs_trans_free(xfs_trans_t *);
 
-kmem_zone_t            *xfs_trans_zone;
+kmem_zone_t    *xfs_trans_zone;
 
 
 /*
+ * Reservation functions here avoid a huge stack in xfs_trans_init
+ * due to register overflow from temporaries in the calculations.
+ */
+
+STATIC uint
+xfs_calc_write_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_WRITE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
+}
+
+STATIC uint
+xfs_calc_itruncate_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_ITRUNCATE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
+}
+
+STATIC uint
+xfs_calc_rename_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_RENAME_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
+}
+
+STATIC uint
+xfs_calc_link_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_LINK_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
+}
+
+STATIC uint
+xfs_calc_remove_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_REMOVE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
+}
+
+STATIC uint
+xfs_calc_symlink_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_SYMLINK_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
+}
+
+STATIC uint
+xfs_calc_create_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_CREATE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
+}
+
+STATIC uint
+xfs_calc_mkdir_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_MKDIR_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
+}
+
+STATIC uint
+xfs_calc_ifree_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_IFREE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
+}
+
+STATIC uint
+xfs_calc_ichange_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_ICHANGE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
+}
+
+STATIC uint
+xfs_calc_growdata_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_GROWDATA_LOG_RES(mp);
+}
+
+STATIC uint
+xfs_calc_growrtalloc_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_GROWRTALLOC_LOG_RES(mp);
+}
+
+STATIC uint
+xfs_calc_growrtzero_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_GROWRTZERO_LOG_RES(mp);
+}
+
+STATIC uint
+xfs_calc_growrtfree_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_GROWRTFREE_LOG_RES(mp);
+}
+
+STATIC uint
+xfs_calc_swrite_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_SWRITE_LOG_RES(mp);
+}
+
+STATIC uint
+xfs_calc_writeid_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_WRITEID_LOG_RES(mp);
+}
+
+STATIC uint
+xfs_calc_addafork_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_ADDAFORK_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
+}
+
+STATIC uint
+xfs_calc_attrinval_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_ATTRINVAL_LOG_RES(mp);
+}
+
+STATIC uint
+xfs_calc_attrset_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_ATTRSET_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
+}
+
+STATIC uint
+xfs_calc_attrrm_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_ATTRRM_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
+}
+
+STATIC uint
+xfs_calc_clear_agi_bucket_reservation(xfs_mount_t *mp)
+{
+       return XFS_CALC_CLEAR_AGI_BUCKET_LOG_RES(mp);
+}
+
+/*
  * Initialize the precomputed transaction reservation values
  * in the mount structure.
  */
@@ -69,39 +199,27 @@ xfs_trans_init(
        xfs_trans_reservations_t        *resp;
 
        resp = &(mp->m_reservations);
-       resp->tr_write =
-               (uint)(XFS_CALC_WRITE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp));
-       resp->tr_itruncate =
-               (uint)(XFS_CALC_ITRUNCATE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp));
-       resp->tr_rename =
-               (uint)(XFS_CALC_RENAME_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp));
-       resp->tr_link = (uint)XFS_CALC_LINK_LOG_RES(mp);
-       resp->tr_remove =
-               (uint)(XFS_CALC_REMOVE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp));
-       resp->tr_symlink =
-               (uint)(XFS_CALC_SYMLINK_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp));
-       resp->tr_create =
-               (uint)(XFS_CALC_CREATE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp));
-       resp->tr_mkdir =
-               (uint)(XFS_CALC_MKDIR_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp));
-       resp->tr_ifree =
-               (uint)(XFS_CALC_IFREE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp));
-       resp->tr_ichange =
-               (uint)(XFS_CALC_ICHANGE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp));
-       resp->tr_growdata = (uint)XFS_CALC_GROWDATA_LOG_RES(mp);
-       resp->tr_swrite = (uint)XFS_CALC_SWRITE_LOG_RES(mp);
-       resp->tr_writeid = (uint)XFS_CALC_WRITEID_LOG_RES(mp);
-       resp->tr_addafork =
-               (uint)(XFS_CALC_ADDAFORK_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp));
-       resp->tr_attrinval = (uint)XFS_CALC_ATTRINVAL_LOG_RES(mp);
-       resp->tr_attrset =
-               (uint)(XFS_CALC_ATTRSET_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp));
-       resp->tr_attrrm =
-               (uint)(XFS_CALC_ATTRRM_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp));
-       resp->tr_clearagi = (uint)XFS_CALC_CLEAR_AGI_BUCKET_LOG_RES(mp);
-       resp->tr_growrtalloc = (uint)XFS_CALC_GROWRTALLOC_LOG_RES(mp);
-       resp->tr_growrtzero = (uint)XFS_CALC_GROWRTZERO_LOG_RES(mp);
-       resp->tr_growrtfree = (uint)XFS_CALC_GROWRTFREE_LOG_RES(mp);
+       resp->tr_write = xfs_calc_write_reservation(mp);
+       resp->tr_itruncate = xfs_calc_itruncate_reservation(mp);
+       resp->tr_rename = xfs_calc_rename_reservation(mp);
+       resp->tr_link = xfs_calc_link_reservation(mp);
+       resp->tr_remove = xfs_calc_remove_reservation(mp);
+       resp->tr_symlink = xfs_calc_symlink_reservation(mp);
+       resp->tr_create = xfs_calc_create_reservation(mp);
+       resp->tr_mkdir = xfs_calc_mkdir_reservation(mp);
+       resp->tr_ifree = xfs_calc_ifree_reservation(mp);
+       resp->tr_ichange = xfs_calc_ichange_reservation(mp);
+       resp->tr_growdata = xfs_calc_growdata_reservation(mp);
+       resp->tr_swrite = xfs_calc_swrite_reservation(mp);
+       resp->tr_writeid = xfs_calc_writeid_reservation(mp);
+       resp->tr_addafork = xfs_calc_addafork_reservation(mp);
+       resp->tr_attrinval = xfs_calc_attrinval_reservation(mp);
+       resp->tr_attrset = xfs_calc_attrset_reservation(mp);
+       resp->tr_attrrm = xfs_calc_attrrm_reservation(mp);
+       resp->tr_clearagi = xfs_calc_clear_agi_bucket_reservation(mp);
+       resp->tr_growrtalloc = xfs_calc_growrtalloc_reservation(mp);
+       resp->tr_growrtzero = xfs_calc_growrtzero_reservation(mp);
+       resp->tr_growrtfree = xfs_calc_growrtfree_reservation(mp);
 }
 
 /*
@@ -117,35 +235,29 @@ xfs_trans_alloc(
        xfs_mount_t     *mp,
        uint            type)
 {
-       fs_check_frozen(XFS_MTOVFS(mp), SB_FREEZE_TRANS);
-       atomic_inc(&mp->m_active_trans);
-
-       return (_xfs_trans_alloc(mp, type));
-
+       xfs_wait_for_freeze(mp, SB_FREEZE_TRANS);
+       return _xfs_trans_alloc(mp, type, KM_SLEEP);
 }
 
 xfs_trans_t *
 _xfs_trans_alloc(
        xfs_mount_t     *mp,
-       uint            type)
+       uint            type,
+       uint            memflags)
 {
        xfs_trans_t     *tp;
 
-       ASSERT(xfs_trans_zone != NULL);
-       tp = kmem_zone_zalloc(xfs_trans_zone, KM_SLEEP);
+       atomic_inc(&mp->m_active_trans);
 
-       /*
-        * Initialize the transaction structure.
-        */
+       tp = kmem_zone_zalloc(xfs_trans_zone, memflags);
        tp->t_magic = XFS_TRANS_MAGIC;
        tp->t_type = type;
        tp->t_mountp = mp;
        tp->t_items_free = XFS_LIC_NUM_SLOTS;
        tp->t_busy_free = XFS_LBC_NUM_SLOTS;
-       XFS_LIC_INIT(&(tp->t_items));
+       xfs_lic_init(&(tp->t_items));
        XFS_LBC_INIT(&(tp->t_busy));
-
-       return (tp);
+       return tp;
 }
 
 /*
@@ -172,21 +284,21 @@ xfs_trans_dup(
        ntp->t_mountp = tp->t_mountp;
        ntp->t_items_free = XFS_LIC_NUM_SLOTS;
        ntp->t_busy_free = XFS_LBC_NUM_SLOTS;
-       XFS_LIC_INIT(&(ntp->t_items));
+       xfs_lic_init(&(ntp->t_items));
        XFS_LBC_INIT(&(ntp->t_busy));
 
        ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
        ASSERT(tp->t_ticket != NULL);
 
        ntp->t_flags = XFS_TRANS_PERM_LOG_RES | (tp->t_flags & XFS_TRANS_RESERVE);
-       ntp->t_ticket = tp->t_ticket;
+       ntp->t_ticket = xfs_log_ticket_get(tp->t_ticket);
        ntp->t_blk_res = tp->t_blk_res - tp->t_blk_res_used;
        tp->t_blk_res = tp->t_blk_res_used;
        ntp->t_rtx_res = tp->t_rtx_res - tp->t_rtx_res_used;
        tp->t_rtx_res = tp->t_rtx_res_used;
-       PFLAGS_DUP(&tp->t_pflags, &ntp->t_pflags);
+       ntp->t_pflags = tp->t_pflags;
 
-       XFS_TRANS_DUP_DQINFO(tp->t_mountp, tp, ntp);
+       xfs_trans_dup_dqinfo(tp, ntp);
 
        atomic_inc(&tp->t_mountp->m_active_trans);
        return ntp;
@@ -216,14 +328,11 @@ xfs_trans_reserve(
        uint            logcount)
 {
        int             log_flags;
-       int             error;
-       int     rsvd;
-
-       error = 0;
-       rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
+       int             error = 0;
+       int             rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
 
        /* Mark this thread as being in a transaction */
-        PFLAGS_SET_FSTRANS(&tp->t_pflags);
+       current_set_flags_nested(&tp->t_pflags, PF_FSTRANS);
 
        /*
         * Attempt to reserve the needed disk blocks by decrementing
@@ -232,9 +341,9 @@ xfs_trans_reserve(
         */
        if (blocks > 0) {
                error = xfs_mod_incore_sb(tp->t_mountp, XFS_SBS_FDBLOCKS,
-                                         -blocks, rsvd);
+                                         -((int64_t)blocks), rsvd);
                if (error != 0) {
-                        PFLAGS_RESTORE_FSTRANS(&tp->t_pflags);
+                       current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
                        return (XFS_ERROR(ENOSPC));
                }
                tp->t_blk_res += blocks;
@@ -273,7 +382,7 @@ xfs_trans_reserve(
         */
        if (rtextents > 0) {
                error = xfs_mod_incore_sb(tp->t_mountp, XFS_SBS_FREXTENTS,
-                                         -rtextents, rsvd);
+                                         -((int64_t)rtextents), rsvd);
                if (error) {
                        error = XFS_ERROR(ENOSPC);
                        goto undo_log;
@@ -303,13 +412,13 @@ undo_log:
 undo_blocks:
        if (blocks > 0) {
                (void) xfs_mod_incore_sb(tp->t_mountp, XFS_SBS_FDBLOCKS,
-                                        blocks, rsvd);
+                                        (int64_t)blocks, rsvd);
                tp->t_blk_res = 0;
        }
 
-        PFLAGS_RESTORE_FSTRANS(&tp->t_pflags);
+       current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
 
-       return (error);
+       return error;
 }
 
 
@@ -320,20 +429,34 @@ undo_blocks:
  *
  * Mark the transaction structure to indicate that the superblock
  * needs to be updated before committing.
+ *
+ * Because we may not be keeping track of allocated/free inodes and
+ * used filesystem blocks in the superblock, we do not mark the
+ * superblock dirty in this transaction if we modify these fields.
+ * We still need to update the transaction deltas so that they get
+ * applied to the incore superblock, but we don't want them to
+ * cause the superblock to get locked and logged if these are the
+ * only fields in the superblock that the transaction modifies.
  */
 void
 xfs_trans_mod_sb(
        xfs_trans_t     *tp,
        uint            field,
-       long            delta)
+       int64_t         delta)
 {
+       uint32_t        flags = (XFS_TRANS_DIRTY|XFS_TRANS_SB_DIRTY);
+       xfs_mount_t     *mp = tp->t_mountp;
 
        switch (field) {
        case XFS_TRANS_SB_ICOUNT:
                tp->t_icount_delta += delta;
+               if (xfs_sb_version_haslazysbcount(&mp->m_sb))
+                       flags &= ~XFS_TRANS_SB_DIRTY;
                break;
        case XFS_TRANS_SB_IFREE:
                tp->t_ifree_delta += delta;
+               if (xfs_sb_version_haslazysbcount(&mp->m_sb))
+                       flags &= ~XFS_TRANS_SB_DIRTY;
                break;
        case XFS_TRANS_SB_FDBLOCKS:
                /*
@@ -346,6 +469,8 @@ xfs_trans_mod_sb(
                        ASSERT(tp->t_blk_res_used <= tp->t_blk_res);
                }
                tp->t_fdblocks_delta += delta;
+               if (xfs_sb_version_haslazysbcount(&mp->m_sb))
+                       flags &= ~XFS_TRANS_SB_DIRTY;
                break;
        case XFS_TRANS_SB_RES_FDBLOCKS:
                /*
@@ -355,6 +480,8 @@ xfs_trans_mod_sb(
                 */
                ASSERT(delta < 0);
                tp->t_res_fdblocks_delta += delta;
+               if (xfs_sb_version_haslazysbcount(&mp->m_sb))
+                       flags &= ~XFS_TRANS_SB_DIRTY;
                break;
        case XFS_TRANS_SB_FREXTENTS:
                /*
@@ -371,7 +498,7 @@ xfs_trans_mod_sb(
        case XFS_TRANS_SB_RES_FREXTENTS:
                /*
                 * The allocation has already been applied to the
-                * in-core superblocks's counter.  This should only
+                * in-core superblock's counter.  This should only
                 * be applied to the on-disk superblock.
                 */
                ASSERT(delta < 0);
@@ -408,7 +535,7 @@ xfs_trans_mod_sb(
                return;
        }
 
-       tp->t_flags |= (XFS_TRANS_SB_DIRTY | XFS_TRANS_DIRTY);
+       tp->t_flags |= flags;
 }
 
 /*
@@ -423,7 +550,7 @@ STATIC void
 xfs_trans_apply_sb_deltas(
        xfs_trans_t     *tp)
 {
-       xfs_sb_t        *sbp;
+       xfs_dsb_t       *sbp;
        xfs_buf_t       *bp;
        int             whole = 0;
 
@@ -437,82 +564,91 @@ xfs_trans_apply_sb_deltas(
               (tp->t_ag_freeblks_delta + tp->t_ag_flist_delta +
                tp->t_ag_btree_delta));
 
-       if (tp->t_icount_delta != 0) {
-               INT_MOD(sbp->sb_icount, ARCH_CONVERT, tp->t_icount_delta);
-       }
-       if (tp->t_ifree_delta != 0) {
-               INT_MOD(sbp->sb_ifree, ARCH_CONVERT, tp->t_ifree_delta);
+       /*
+        * Only update the superblock counters if we are logging them
+        */
+       if (!xfs_sb_version_haslazysbcount(&(tp->t_mountp->m_sb))) {
+               if (tp->t_icount_delta)
+                       be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta);
+               if (tp->t_ifree_delta)
+                       be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta);
+               if (tp->t_fdblocks_delta)
+                       be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta);
+               if (tp->t_res_fdblocks_delta)
+                       be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta);
        }
 
-       if (tp->t_fdblocks_delta != 0) {
-               INT_MOD(sbp->sb_fdblocks, ARCH_CONVERT, tp->t_fdblocks_delta);
-       }
-       if (tp->t_res_fdblocks_delta != 0) {
-               INT_MOD(sbp->sb_fdblocks, ARCH_CONVERT, tp->t_res_fdblocks_delta);
-       }
+       if (tp->t_frextents_delta)
+               be64_add_cpu(&sbp->sb_frextents, tp->t_frextents_delta);
+       if (tp->t_res_frextents_delta)
+               be64_add_cpu(&sbp->sb_frextents, tp->t_res_frextents_delta);
 
-       if (tp->t_frextents_delta != 0) {
-               INT_MOD(sbp->sb_frextents, ARCH_CONVERT, tp->t_frextents_delta);
-       }
-       if (tp->t_res_frextents_delta != 0) {
-               INT_MOD(sbp->sb_frextents, ARCH_CONVERT, tp->t_res_frextents_delta);
-       }
-       if (tp->t_dblocks_delta != 0) {
-               INT_MOD(sbp->sb_dblocks, ARCH_CONVERT, tp->t_dblocks_delta);
+       if (tp->t_dblocks_delta) {
+               be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta);
                whole = 1;
        }
-       if (tp->t_agcount_delta != 0) {
-               INT_MOD(sbp->sb_agcount, ARCH_CONVERT, tp->t_agcount_delta);
+       if (tp->t_agcount_delta) {
+               be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta);
                whole = 1;
        }
-       if (tp->t_imaxpct_delta != 0) {
-               INT_MOD(sbp->sb_imax_pct, ARCH_CONVERT, tp->t_imaxpct_delta);
+       if (tp->t_imaxpct_delta) {
+               sbp->sb_imax_pct += tp->t_imaxpct_delta;
                whole = 1;
        }
-       if (tp->t_rextsize_delta != 0) {
-               INT_MOD(sbp->sb_rextsize, ARCH_CONVERT, tp->t_rextsize_delta);
+       if (tp->t_rextsize_delta) {
+               be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta);
                whole = 1;
        }
-       if (tp->t_rbmblocks_delta != 0) {
-               INT_MOD(sbp->sb_rbmblocks, ARCH_CONVERT, tp->t_rbmblocks_delta);
+       if (tp->t_rbmblocks_delta) {
+               be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta);
                whole = 1;
        }
-       if (tp->t_rblocks_delta != 0) {
-               INT_MOD(sbp->sb_rblocks, ARCH_CONVERT, tp->t_rblocks_delta);
+       if (tp->t_rblocks_delta) {
+               be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta);
                whole = 1;
        }
-       if (tp->t_rextents_delta != 0) {
-               INT_MOD(sbp->sb_rextents, ARCH_CONVERT, tp->t_rextents_delta);
+       if (tp->t_rextents_delta) {
+               be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta);
                whole = 1;
        }
-       if (tp->t_rextslog_delta != 0) {
-               INT_MOD(sbp->sb_rextslog, ARCH_CONVERT, tp->t_rextslog_delta);
+       if (tp->t_rextslog_delta) {
+               sbp->sb_rextslog += tp->t_rextslog_delta;
                whole = 1;
        }
 
        if (whole)
                /*
-                * Log the whole thing, the fields are discontiguous.
+                * Log the whole thing, the fields are noncontiguous.
                 */
-               xfs_trans_log_buf(tp, bp, 0, sizeof(xfs_sb_t) - 1);
+               xfs_trans_log_buf(tp, bp, 0, sizeof(xfs_dsb_t) - 1);
        else
                /*
                 * Since all the modifiable fields are contiguous, we
                 * can get away with this.
                 */
-               xfs_trans_log_buf(tp, bp, offsetof(xfs_sb_t, sb_icount),
-                                 offsetof(xfs_sb_t, sb_frextents) +
+               xfs_trans_log_buf(tp, bp, offsetof(xfs_dsb_t, sb_icount),
+                                 offsetof(xfs_dsb_t, sb_frextents) +
                                  sizeof(sbp->sb_frextents) - 1);
-
-       XFS_MTOVFS(tp->t_mountp)->vfs_super->s_dirt = 1;
 }
 
 /*
- * xfs_trans_unreserve_and_mod_sb() is called to release unused
- * reservations and apply superblock counter changes to the in-core
- * superblock.
+ * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations
+ * and apply superblock counter changes to the in-core superblock.  The
+ * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT
+ * applied to the in-core superblock.  The idea is that that has already been
+ * done.
  *
  * This is done efficiently with a single call to xfs_mod_incore_sb_batch().
+ * However, we have to ensure that we only modify each superblock field only
+ * once because the application of the delta values may not be atomic. That can
+ * lead to ENOSPC races occurring if we have two separate modifcations of the
+ * free space counter to put back the entire reservation and then take away
+ * what we used.
+ *
+ * If we are not logging superblock counters, then the inode allocated/free and
+ * used block counts are not updated in the on disk superblock. In this case,
+ * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we
+ * still need to update the incore superblock with the changes.
  */
 STATIC void
 xfs_trans_unreserve_and_mod_sb(
@@ -520,98 +656,100 @@ xfs_trans_unreserve_and_mod_sb(
 {
        xfs_mod_sb_t    msb[14];        /* If you add cases, add entries */
        xfs_mod_sb_t    *msbp;
+       xfs_mount_t     *mp = tp->t_mountp;
        /* REFERENCED */
        int             error;
        int             rsvd;
+       int64_t         blkdelta = 0;
+       int64_t         rtxdelta = 0;
 
        msbp = msb;
        rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
 
-       /*
-        * Release any reserved blocks.  Any that were allocated
-        * will be taken back again by fdblocks_delta below.
-        */
-       if (tp->t_blk_res > 0) {
+       /* calculate free blocks delta */
+       if (tp->t_blk_res > 0)
+               blkdelta = tp->t_blk_res;
+
+       if ((tp->t_fdblocks_delta != 0) &&
+           (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
+            (tp->t_flags & XFS_TRANS_SB_DIRTY)))
+               blkdelta += tp->t_fdblocks_delta;
+
+       if (blkdelta != 0) {
                msbp->msb_field = XFS_SBS_FDBLOCKS;
-               msbp->msb_delta = tp->t_blk_res;
+               msbp->msb_delta = blkdelta;
                msbp++;
        }
 
-       /*
-        * Release any reserved real time extents .  Any that were
-        * allocated will be taken back again by frextents_delta below.
-        */
-       if (tp->t_rtx_res > 0) {
+       /* calculate free realtime extents delta */
+       if (tp->t_rtx_res > 0)
+               rtxdelta = tp->t_rtx_res;
+
+       if ((tp->t_frextents_delta != 0) &&
+           (tp->t_flags & XFS_TRANS_SB_DIRTY))
+               rtxdelta += tp->t_frextents_delta;
+
+       if (rtxdelta != 0) {
                msbp->msb_field = XFS_SBS_FREXTENTS;
-               msbp->msb_delta = tp->t_rtx_res;
+               msbp->msb_delta = rtxdelta;
                msbp++;
        }
 
-       /*
-        * Apply any superblock modifications to the in-core version.
-        * The t_res_fdblocks_delta and t_res_frextents_delta fields are
-        * explicity NOT applied to the in-core superblock.
-        * The idea is that that has already been done.
-        */
-       if (tp->t_flags & XFS_TRANS_SB_DIRTY) {
+       /* apply remaining deltas */
+
+       if (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
+            (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
                if (tp->t_icount_delta != 0) {
                        msbp->msb_field = XFS_SBS_ICOUNT;
-                       msbp->msb_delta = (int)tp->t_icount_delta;
+                       msbp->msb_delta = tp->t_icount_delta;
                        msbp++;
                }
                if (tp->t_ifree_delta != 0) {
                        msbp->msb_field = XFS_SBS_IFREE;
-                       msbp->msb_delta = (int)tp->t_ifree_delta;
-                       msbp++;
-               }
-               if (tp->t_fdblocks_delta != 0) {
-                       msbp->msb_field = XFS_SBS_FDBLOCKS;
-                       msbp->msb_delta = (int)tp->t_fdblocks_delta;
-                       msbp++;
-               }
-               if (tp->t_frextents_delta != 0) {
-                       msbp->msb_field = XFS_SBS_FREXTENTS;
-                       msbp->msb_delta = (int)tp->t_frextents_delta;
+                       msbp->msb_delta = tp->t_ifree_delta;
                        msbp++;
                }
+       }
+
+       if (tp->t_flags & XFS_TRANS_SB_DIRTY) {
                if (tp->t_dblocks_delta != 0) {
                        msbp->msb_field = XFS_SBS_DBLOCKS;
-                       msbp->msb_delta = (int)tp->t_dblocks_delta;
+                       msbp->msb_delta = tp->t_dblocks_delta;
                        msbp++;
                }
                if (tp->t_agcount_delta != 0) {
                        msbp->msb_field = XFS_SBS_AGCOUNT;
-                       msbp->msb_delta = (int)tp->t_agcount_delta;
+                       msbp->msb_delta = tp->t_agcount_delta;
                        msbp++;
                }
                if (tp->t_imaxpct_delta != 0) {
                        msbp->msb_field = XFS_SBS_IMAX_PCT;
-                       msbp->msb_delta = (int)tp->t_imaxpct_delta;
+                       msbp->msb_delta = tp->t_imaxpct_delta;
                        msbp++;
                }
                if (tp->t_rextsize_delta != 0) {
                        msbp->msb_field = XFS_SBS_REXTSIZE;
-                       msbp->msb_delta = (int)tp->t_rextsize_delta;
+                       msbp->msb_delta = tp->t_rextsize_delta;
                        msbp++;
                }
                if (tp->t_rbmblocks_delta != 0) {
                        msbp->msb_field = XFS_SBS_RBMBLOCKS;
-                       msbp->msb_delta = (int)tp->t_rbmblocks_delta;
+                       msbp->msb_delta = tp->t_rbmblocks_delta;
                        msbp++;
                }
                if (tp->t_rblocks_delta != 0) {
                        msbp->msb_field = XFS_SBS_RBLOCKS;
-                       msbp->msb_delta = (int)tp->t_rblocks_delta;
+                       msbp->msb_delta = tp->t_rblocks_delta;
                        msbp++;
                }
                if (tp->t_rextents_delta != 0) {
                        msbp->msb_field = XFS_SBS_REXTENTS;
-                       msbp->msb_delta = (int)tp->t_rextents_delta;
+                       msbp->msb_delta = tp->t_rextents_delta;
                        msbp++;
                }
                if (tp->t_rextslog_delta != 0) {
                        msbp->msb_field = XFS_SBS_REXTSLOG;
-                       msbp->msb_delta = (int)tp->t_rextslog_delta;
+                       msbp->msb_delta = tp->t_rextslog_delta;
                        msbp++;
                }
        }
@@ -646,7 +784,6 @@ int
 _xfs_trans_commit(
        xfs_trans_t     *tp,
        uint            flags,
-       xfs_lsn_t       *commit_lsn_p,
        int             *log_flushed)
 {
        xfs_log_iovec_t         *log_vector;
@@ -693,20 +830,18 @@ shut_us_down:
                 * means is that we have some (non-persistent) quota
                 * reservations that need to be unreserved.
                 */
-               XFS_TRANS_UNRESERVE_AND_MOD_DQUOTS(mp, tp);
+               xfs_trans_unreserve_and_mod_dquots(tp);
                if (tp->t_ticket) {
                        commit_lsn = xfs_log_done(mp, tp->t_ticket,
                                                        NULL, log_flags);
                        if (commit_lsn == -1 && !shutdown)
                                shutdown = XFS_ERROR(EIO);
                }
-                PFLAGS_RESTORE_FSTRANS(&tp->t_pflags);
+               current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
                xfs_trans_free_items(tp, shutdown? XFS_TRANS_ABORT : 0);
                xfs_trans_free_busy(tp);
                xfs_trans_free(tp);
                XFS_STATS_INC(xs_trans_empty);
-               if (commit_lsn_p)
-                       *commit_lsn_p = commit_lsn;
                return (shutdown);
        }
        ASSERT(tp->t_ticket != NULL);
@@ -714,10 +849,9 @@ shut_us_down:
        /*
         * If we need to update the superblock, then do it now.
         */
-       if (tp->t_flags & XFS_TRANS_SB_DIRTY) {
+       if (tp->t_flags & XFS_TRANS_SB_DIRTY)
                xfs_trans_apply_sb_deltas(tp);
-       }
-       XFS_TRANS_APPLY_DQUOT_DELTAS(mp, tp);
+       xfs_trans_apply_dquot_deltas(tp);
 
        /*
         * Ask each log item how many log_vector entries it will
@@ -727,7 +861,7 @@ shut_us_down:
         */
        nvec = xfs_trans_count_vecs(tp);
        if (nvec == 0) {
-               xfs_force_shutdown(mp, XFS_LOG_IO_ERROR);
+               xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
                goto shut_us_down;
        } else if (nvec <= XFS_TRANS_LOGVEC_COUNT) {
                log_vector = log_vector_fast;
@@ -754,18 +888,15 @@ shut_us_down:
 
        tp->t_commit_lsn = commit_lsn;
        if (nvec > XFS_TRANS_LOGVEC_COUNT) {
-               kmem_free(log_vector, nvec * sizeof(xfs_log_iovec_t));
+               kmem_free(log_vector);
        }
 
-       if (commit_lsn_p)
-               *commit_lsn_p = commit_lsn;
-
        /*
         * If we got a log write error. Unpin the logitems that we
         * had pinned, clean up, free trans structure, and return error.
         */
        if (error || commit_lsn == -1) {
-                PFLAGS_RESTORE_FSTRANS(&tp->t_pflags);
+               current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
                xfs_trans_uncommit(tp, flags|XFS_TRANS_ABORT);
                return XFS_ERROR(EIO);
        }
@@ -807,7 +938,7 @@ shut_us_down:
        /*
         * Mark this thread as no longer being in a transaction
         */
-       PFLAGS_RESTORE_FSTRANS(&tp->t_pflags);
+       current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
 
        /*
         * Once all the items of the transaction have been copied
@@ -925,7 +1056,7 @@ xfs_trans_uncommit(
        }
 
        xfs_trans_unreserve_and_mod_sb(tp);
-       XFS_TRANS_UNRESERVE_AND_MOD_DQUOTS(tp->t_mountp, tp);
+       xfs_trans_unreserve_and_mod_dquots(tp);
 
        xfs_trans_free_items(tp, flags);
        xfs_trans_free_busy(tp);
@@ -1027,17 +1158,17 @@ xfs_trans_cancel(
         * filesystem.  This happens in paths where we detect
         * corruption and decide to give up.
         */
-       if ((tp->t_flags & XFS_TRANS_DIRTY) &&
-           !XFS_FORCED_SHUTDOWN(mp))
+       if ((tp->t_flags & XFS_TRANS_DIRTY) && !XFS_FORCED_SHUTDOWN(mp)) {
                XFS_ERROR_REPORT("xfs_trans_cancel", XFS_ERRLEVEL_LOW, mp);
-               xfs_force_shutdown(mp, XFS_CORRUPT_INCORE);
+               xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
+       }
 #ifdef DEBUG
        if (!(flags & XFS_TRANS_ABORT)) {
                licp = &(tp->t_items);
                while (licp != NULL) {
                        lidp = licp->lic_descs;
                        for (i = 0; i < licp->lic_unused; i++, lidp++) {
-                               if (XFS_LIC_ISFREE(licp, i)) {
+                               if (xfs_lic_isfree(licp, i)) {
                                        continue;
                                }
 
@@ -1050,7 +1181,7 @@ xfs_trans_cancel(
        }
 #endif
        xfs_trans_unreserve_and_mod_sb(tp);
-       XFS_TRANS_UNRESERVE_AND_MOD_DQUOTS(mp, tp);
+       xfs_trans_unreserve_and_mod_dquots(tp);
 
        if (tp->t_ticket) {
                if (flags & XFS_TRANS_RELEASE_LOG_RES) {
@@ -1063,7 +1194,7 @@ xfs_trans_cancel(
        }
 
        /* mark this thread as no longer being in a transaction */
-        PFLAGS_RESTORE_FSTRANS(&tp->t_pflags);
+       current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
 
        xfs_trans_free_items(tp, flags);
        xfs_trans_free_busy(tp);
@@ -1080,10 +1211,79 @@ xfs_trans_free(
        xfs_trans_t     *tp)
 {
        atomic_dec(&tp->t_mountp->m_active_trans);
-       XFS_TRANS_FREE_DQINFO(tp->t_mountp, tp);
+       xfs_trans_free_dqinfo(tp);
        kmem_zone_free(xfs_trans_zone, tp);
 }
 
+/*
+ * Roll from one trans in the sequence of PERMANENT transactions to
+ * the next: permanent transactions are only flushed out when
+ * committed with XFS_TRANS_RELEASE_LOG_RES, but we still want as soon
+ * as possible to let chunks of it go to the log. So we commit the
+ * chunk we've been working on and get a new transaction to continue.
+ */
+int
+xfs_trans_roll(
+       struct xfs_trans        **tpp,
+       struct xfs_inode        *dp)
+{
+       struct xfs_trans        *trans;
+       unsigned int            logres, count;
+       int                     error;
+
+       /*
+        * Ensure that the inode is always logged.
+        */
+       trans = *tpp;
+       xfs_trans_log_inode(trans, dp, XFS_ILOG_CORE);
+
+       /*
+        * Copy the critical parameters from one trans to the next.
+        */
+       logres = trans->t_log_res;
+       count = trans->t_log_count;
+       *tpp = xfs_trans_dup(trans);
+
+       /*
+        * Commit the current transaction.
+        * If this commit failed, then it'd just unlock those items that
+        * are not marked ihold. That also means that a filesystem shutdown
+        * is in progress. The caller takes the responsibility to cancel
+        * the duplicate transaction that gets returned.
+        */
+       error = xfs_trans_commit(trans, 0);
+       if (error)
+               return (error);
+
+       trans = *tpp;
+
+       /*
+        * transaction commit worked ok so we can drop the extra ticket
+        * reference that we gained in xfs_trans_dup()
+        */
+       xfs_log_ticket_put(trans->t_ticket);
+
+
+       /*
+        * Reserve space in the log for th next transaction.
+        * This also pushes items in the "AIL", the list of logged items,
+        * out to disk if they are taking up space at the tail of the log
+        * that we want to use.  This requires that either nothing be locked
+        * across this call, or that anything that is locked be logged in
+        * the prior and the next transactions.
+        */
+       error = xfs_trans_reserve(trans, 0, logres, 0,
+                                 XFS_TRANS_PERM_LOG_RES, count);
+       /*
+        *  Ensure that the inode is in the new transaction and locked.
+        */
+       if (error)
+               return error;
+
+       xfs_trans_ijoin(trans, dp, XFS_ILOCK_EXCL);
+       xfs_trans_ihold(trans, dp);
+       return 0;
+}
 
 /*
  * THIS SHOULD BE REWRITTEN TO USE xfs_trans_next_item().
@@ -1121,7 +1321,7 @@ xfs_trans_committed(
         * Special case the chunk embedded in the transaction.
         */
        licp = &(tp->t_items);
-       if (!(XFS_LIC_ARE_ALL_FREE(licp))) {
+       if (!(xfs_lic_are_all_free(licp))) {
                xfs_trans_chunk_committed(licp, tp->t_lsn, abortflag);
        }
 
@@ -1130,10 +1330,10 @@ xfs_trans_committed(
         */
        licp = licp->lic_next;
        while (licp != NULL) {
-               ASSERT(!XFS_LIC_ARE_ALL_FREE(licp));
+               ASSERT(!xfs_lic_are_all_free(licp));
                xfs_trans_chunk_committed(licp, tp->t_lsn, abortflag);
                next_licp = licp->lic_next;
-               kmem_free(licp, sizeof(xfs_log_item_chunk_t));
+               kmem_free(licp);
                licp = next_licp;
        }
 
@@ -1188,13 +1388,13 @@ xfs_trans_chunk_committed(
        xfs_log_item_desc_t     *lidp;
        xfs_log_item_t          *lip;
        xfs_lsn_t               item_lsn;
-       struct xfs_mount        *mp;
        int                     i;
-       SPLDECL(s);
 
        lidp = licp->lic_descs;
        for (i = 0; i < licp->lic_unused; i++, lidp++) {
-               if (XFS_LIC_ISFREE(licp, i)) {
+               struct xfs_ail          *ailp;
+
+               if (xfs_lic_isfree(licp, i)) {
                        continue;
                }
 
@@ -1230,19 +1430,19 @@ xfs_trans_chunk_committed(
                 * This would cause the earlier transaction to fail
                 * the test below.
                 */
-               mp = lip->li_mountp;
-               AIL_LOCK(mp,s);
+               ailp = lip->li_ailp;
+               spin_lock(&ailp->xa_lock);
                if (XFS_LSN_CMP(item_lsn, lip->li_lsn) > 0) {
                        /*
                         * This will set the item's lsn to item_lsn
                         * and update the position of the item in
                         * the AIL.
                         *
-                        * xfs_trans_update_ail() drops the AIL lock.
+                        * xfs_trans_ail_update() drops the AIL lock.
                         */
-                       xfs_trans_update_ail(mp, lip, item_lsn, s);
+                       xfs_trans_ail_update(ailp, lip, item_lsn);
                } else {
-                       AIL_UNLOCK(mp, s);
+                       spin_unlock(&ailp->xa_lock);
                }
 
                /*