diff --git a/config/Rules.am b/config/Rules.am
index af20ddb396..20a21e972b 100644
--- a/config/Rules.am
+++ b/config/Rules.am
@@ -7,7 +7,8 @@ AM_CFLAGS += ${NO_BOOL_COMPARE}
 AM_CFLAGS += -fno-strict-aliasing
 AM_CPPFLAGS  = -D_GNU_SOURCE -D__EXTENSIONS__ -D_REENTRANT
 AM_CPPFLAGS += -D_POSIX_PTHREAD_SEMANTICS -D_FILE_OFFSET_BITS=64
-AM_CPPFLAGS += -D_LARGEFILE64_SOURCE -DTEXT_DOMAIN=\"zfs-linux-user\"
+AM_CPPFLAGS += -D_LARGEFILE64_SOURCE -DHAVE_LARGE_STACKS=1
+AM_CPPFLAGS += -DTEXT_DOMAIN=\"zfs-linux-user\"
 AM_CPPFLAGS += -DLIBEXECDIR=\"$(libexecdir)\"
 AM_CPPFLAGS += -DRUNSTATEDIR=\"$(runstatedir)\"
 AM_CPPFLAGS += -DSBINDIR=\"$(sbindir)\"
diff --git a/config/kernel.m4 b/config/kernel.m4
index 30c9b84d22..ae4bed9a7b 100644
--- a/config/kernel.m4
+++ b/config/kernel.m4
@@ -473,9 +473,35 @@ AC_DEFUN([ZFS_AC_KERNEL_CONFIG], [
 		])
 	])
 
+	ZFS_AC_KERNEL_CONFIG_THREAD_SIZE
 	ZFS_AC_KERNEL_CONFIG_DEBUG_LOCK_ALLOC
 ])
 
+dnl #
+dnl # Check configured THREAD_SIZE
+dnl #
+dnl # The stack size will vary by architecture, but as of Linux 3.15 on x86_64
+dnl # the default thread stack size was increased to 16K from 8K.  Therefore,
+dnl # on newer kernels and some architectures stack usage optimizations can be
+dnl # conditionally applied to improve performance without negatively impacting
+dnl # stability.
+dnl #
+AC_DEFUN([ZFS_AC_KERNEL_CONFIG_THREAD_SIZE], [
+	AC_MSG_CHECKING([whether kernel was built with 16K or larger stacks])
+	ZFS_LINUX_TRY_COMPILE([
+		#include <linux/module.h>
+	],[
+		#if (THREAD_SIZE < 16384)
+		#error "THREAD_SIZE is less than 16K"
+		#endif
+	],[
+		AC_MSG_RESULT([yes])
+		AC_DEFINE(HAVE_LARGE_STACKS, 1, [kernel has large stacks])
+	],[
+		AC_MSG_RESULT([no])
+	])
+])
+
 dnl #
 dnl # Check CONFIG_DEBUG_LOCK_ALLOC
 dnl #
@@ -585,7 +611,7 @@ dnl #
 dnl # ZFS_LINUX_CONFIG
 dnl #
 AC_DEFUN([ZFS_LINUX_CONFIG],
-	[AC_MSG_CHECKING([whether Linux was built with CONFIG_$1])
+	[AC_MSG_CHECKING([whether kernel was built with CONFIG_$1])
 	ZFS_LINUX_TRY_COMPILE([
 		#include <linux/module.h>
 	],[
diff --git a/module/zfs/dmu_send.c b/module/zfs/dmu_send.c
index 6a349c6600..940454977c 100644
--- a/module/zfs/dmu_send.c
+++ b/module/zfs/dmu_send.c
@@ -69,7 +69,7 @@ typedef struct dump_bytes_io {
 } dump_bytes_io_t;
 
 static void
-dump_bytes_strategy(void *arg)
+dump_bytes_cb(void *arg)
 {
 	dump_bytes_io_t *dbi = (dump_bytes_io_t *)arg;
 	dmu_sendarg_t *dsp = dbi->dbi_dsp;
@@ -96,6 +96,9 @@ dump_bytes(dmu_sendarg_t *dsp, void *buf, int len)
 	dbi.dbi_buf = buf;
 	dbi.dbi_len = len;
 
+#if defined(HAVE_LARGE_STACKS)
+	dump_bytes_cb(&dbi);
+#else
 	/*
 	 * The vn_rdwr() call is performed in a taskq to ensure that there is
 	 * always enough stack space to write safely to the target filesystem.
@@ -103,7 +106,8 @@ dump_bytes(dmu_sendarg_t *dsp, void *buf, int len)
 	 * them and they are used in vdev_file.c for a similar purpose.
 	 */
 	spa_taskq_dispatch_sync(dmu_objset_spa(dsp->dsa_os), ZIO_TYPE_FREE,
-	    ZIO_TASKQ_ISSUE, dump_bytes_strategy, &dbi, TQ_SLEEP);
+	    ZIO_TASKQ_ISSUE, dump_bytes_cb, &dbi, TQ_SLEEP);
+#endif /* HAVE_LARGE_STACKS */
 
 	return (dsp->dsa_err);
 }
diff --git a/module/zfs/zio.c b/module/zfs/zio.c
index 794439bbe7..2bc88c52c9 100644
--- a/module/zfs/zio.c
+++ b/module/zfs/zio.c
@@ -1415,6 +1415,31 @@ zio_execute(zio_t *zio)
 	spl_fstrans_unmark(cookie);
 }
 
+/*
+ * Used to determine if in the current context the stack is sized large
+ * enough to allow zio_execute() to be called recursively.  A minimum
+ * stack size of 16K is required to avoid needing to re-dispatch the zio.
+ */
+boolean_t
+zio_execute_stack_check(zio_t *zio)
+{
+#if !defined(HAVE_LARGE_STACKS)
+	dsl_pool_t *dp = spa_get_dsl(zio->io_spa);
+
+	/* Executing in txg_sync_thread() context. */
+	if (dp && curthread == dp->dp_tx.tx_sync_thread)
+		return (B_TRUE);
+
+	/* Pool initialization outside of zio_taskq context. */
+	if (dp && spa_is_initializing(dp->dp_spa) &&
+	    !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) &&
+	    !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH))
+		return (B_TRUE);
+#endif /* HAVE_LARGE_STACKS */
+
+	return (B_FALSE);
+}
+
 __attribute__((always_inline))
 static inline void
 __zio_execute(zio_t *zio)
@@ -1424,8 +1449,6 @@ __zio_execute(zio_t *zio)
 	while (zio->io_stage < ZIO_STAGE_DONE) {
 		enum zio_stage pipeline = zio->io_pipeline;
 		enum zio_stage stage = zio->io_stage;
-		dsl_pool_t *dp;
-		boolean_t cut;
 		int rv;
 
 		ASSERT(!MUTEX_HELD(&zio->io_lock));
@@ -1438,10 +1461,6 @@ __zio_execute(zio_t *zio)
 
 		ASSERT(stage <= ZIO_STAGE_DONE);
 
-		dp = spa_get_dsl(zio->io_spa);
-		cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
-		    zio_requeue_io_start_cut_in_line : B_FALSE;
-
 		/*
 		 * If we are in interrupt context and this pipeline stage
 		 * will grab a config lock that is held across I/O,
@@ -1453,21 +1472,19 @@ __zio_execute(zio_t *zio)
 		 */
 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
+			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
+			    zio_requeue_io_start_cut_in_line : B_FALSE;
 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
 			return;
 		}
 
 		/*
-		 * If we executing in the context of the tx_sync_thread,
-		 * or we are performing pool initialization outside of a
-		 * zio_taskq[ZIO_TASKQ_ISSUE|ZIO_TASKQ_ISSUE_HIGH] context.
-		 * Then issue the zio asynchronously to minimize stack usage
-		 * for these deep call paths.
+		 * If the current context doesn't have large enough stacks
+		 * the zio must be issued asynchronously to prevent overflow.
 		 */
-		if ((dp && curthread == dp->dp_tx.tx_sync_thread) ||
-		    (dp && spa_is_initializing(dp->dp_spa) &&
-		    !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) &&
-		    !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH))) {
+		if (zio_execute_stack_check(zio)) {
+			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
+			    zio_requeue_io_start_cut_in_line : B_FALSE;
 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
 			return;
 		}