Illumos 5056 - ZFS deadlock on db_mtx and dn_holds

5056 ZFS deadlock on db_mtx and dn_holds
Author: Justin Gibbs <[email protected]>
Reviewed by: Will Andrews <[email protected]>
Reviewed by: Matt Ahrens <[email protected]>
Reviewed by: George Wilson <[email protected]>
Approved by: Dan McDonald <[email protected]>

References:
  https://www.illumos.org/issues/5056
  https://github.com/illumos/illumos-gate/commit/bc9014e

Porting Notes:

sa_handle_get_from_db():
  - the original patch includes an otherwise unmentioned fix for a
    possible usage of an uninitialised variable

dmu_objset_open_impl():
  - Under Illumos list_link_init() is the same as filling a list_node_t
    with NULLs, so they don't notice if they miss doing list_link_init()
    on a zero'd containing structure (e.g. allocated with kmem_zalloc as
    here). Under Linux, not so much: an uninitialised list_node_t goes
    "Boom!" some time later when it's used or destroyed.

dmu_objset_evict_dbufs():
  - reduce stack usage using kmem_alloc()

Ported-by: Chris Dunlop <[email protected]>
Signed-off-by: Brian Behlendorf <[email protected]>

1 files changed, 28 insertions, 10 deletions

diff --git a/module/zfs/dsl_dir.c b/module/zfs/dsl_dir.c
index 0a0a0fa7c..e85cb81ae 100644
--- a/module/zfs/dsl_dir.c
+++ b/module/zfs/dsl_dir.c
@@ -23,6 +23,7 @@
  * Copyright (c) 2013 by Delphix. All rights reserved.
  * Copyright (c) 2013 Martin Matuska. All rights reserved.
  * Copyright (c) 2014 Joyent, Inc. All rights reserved.
+ * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
  */
 
 #include <sys/dmu.h>
@@ -126,14 +127,15 @@ extern inline dsl_dir_phys_t *dsl_dir_phys(dsl_dir_t *dd);
 
 static uint64_t dsl_dir_space_towrite(dsl_dir_t *dd);
 
-/* ARGSUSED */
 static void
-dsl_dir_evict(dmu_buf_t *db, void *arg)
+dsl_dir_evict(void *dbu)
 {
-	dsl_dir_t *dd = arg;
+	dsl_dir_t *dd = dbu;
 	int t;
 	ASSERTV(dsl_pool_t *dp = dd->dd_pool);
 
+	dd->dd_dbuf = NULL;
+
 	for (t = 0; t < TXG_SIZE; t++) {
 		ASSERT(!txg_list_member(&dp->dp_dirty_dirs, dd, t));
 		ASSERT(dd->dd_tempreserved[t] == 0);
@@ -141,9 +143,9 @@ dsl_dir_evict(dmu_buf_t *db, void *arg)
 	}
 
 	if (dd->dd_parent)
-		dsl_dir_rele(dd->dd_parent, dd);
+		dsl_dir_async_rele(dd->dd_parent, dd);
 
-	spa_close(dd->dd_pool->dp_spa, dd);
+	spa_async_close(dd->dd_pool->dp_spa, dd);
 
 	/*
 	 * The props callback list should have been cleaned up by
@@ -239,8 +241,9 @@ dsl_dir_hold_obj(dsl_pool_t *dp, uint64_t ddobj,
 			dmu_buf_rele(origin_bonus, FTAG);
 		}
 
-		winner = dmu_buf_set_user_ie(dbuf, dd, dsl_dir_evict);
-		if (winner) {
+		dmu_buf_init_user(&dd->dd_dbu, dsl_dir_evict, &dd->dd_dbuf);
+		winner = dmu_buf_set_user_ie(dbuf, &dd->dd_dbu);
+		if (winner != NULL) {
 			if (dd->dd_parent)
 				dsl_dir_rele(dd->dd_parent, dd);
 			mutex_destroy(&dd->dd_lock);
@@ -284,6 +287,21 @@ dsl_dir_rele(dsl_dir_t *dd, void *tag)
 	dmu_buf_rele(dd->dd_dbuf, tag);
 }
 
+/*
+ * Remove a reference to the given dsl dir that is being asynchronously
+ * released.  Async releases occur from a taskq performing eviction of
+ * dsl datasets and dirs.  This process is identical to a normal release
+ * with the exception of using the async API for releasing the reference on
+ * the spa.
+ */
+void
+dsl_dir_async_rele(dsl_dir_t *dd, void *tag)
+{
+	dprintf_dd(dd, "%s\n", "");
+	spa_async_close(dd->dd_pool->dp_spa, tag);
+	dmu_buf_rele(dd->dd_dbuf, tag);
+}
+
 /* buf must be long enough (MAXNAMELEN + strlen(MOS_DIR_NAME) + 1 should do) */
 void
 dsl_dir_name(dsl_dir_t *dd, char *buf)
@@ -417,7 +435,7 @@ dsl_dir_hold(dsl_pool_t *dp, const char *name, void *tag,
 	}
 
 	while (next != NULL) {
-		dsl_dir_t *child_ds;
+		dsl_dir_t *child_dd;
 		err = getcomponent(next, buf, &nextnext);
 		if (err != 0)
 			break;
@@ -436,11 +454,11 @@ dsl_dir_hold(dsl_pool_t *dp, const char *name, void *tag,
 			break;
 		}
 
-		err = dsl_dir_hold_obj(dp, ddobj, buf, tag, &child_ds);
+		err = dsl_dir_hold_obj(dp, ddobj, buf, tag, &child_dd);
 		if (err != 0)
 			break;
 		dsl_dir_rele(dd, tag);
-		dd = child_ds;
+		dd = child_dd;
 		next = nextnext;
 	}