Conserve stack in dsl_scan_visitbp()

The recursive call chain dsl_scan_visitbp() -> dsl_scan_recurse() ->
dsl_scan_visitdnode() -> dsl_scan_visitbp has been observed to consume
considerable stack resulting in a stack overflow (>8k).  The cleanest
way I see to fix this with minimal impact to the existing flow of
code, and with the fewest performance concerns, is to always inline
dsl_scan_recurse() and dsl_scan_visitdnode().  While this will increase
the function size of dsl_scan_visitbp(), by 4660 bytes, it also reduces
the stack requirements by removing the function call overhead.

Issue #174

diff --git a/module/zfs/dsl_scan.c b/module/zfs/dsl_scan.c
index 107c81b962aa862059fe8147999d175e639a246e..421ddd0418418532c78a919bd71ec3cd95c5d868 100644 (file)
--- a/module/zfs/dsl_scan.c
+++ b/module/zfs/dsl_scan.c
@@ -648,7 +648,7 @@ dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
  * Return nonzero on i/o error.
  * Return new buf to write out in *bufp.
  */
-static int
+__attribute__((always_inline)) static int
 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
     dnode_phys_t *dnp, const blkptr_t *bp,
     const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
@@ -754,7 +754,7 @@ dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
        return (0);
 }
 
-static void
+__attribute__((always_inline)) static void
 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
     dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
     uint64_t object, dmu_tx_t *tx)