Selectable block allocators

ZFS historically has had several space allocators that were
dynamically selectable.  While these have been retained in 
OpenZFS, only a single allocator has been statically compiled 
in. This patch compiles all allocators for OpenZFS and provides 
a module parameter to allow for manual selection between them.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ameer Hamza <ahamza@ixsystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Edmund Nadolski <edmund.nadolski@ixsystems.com>
Closes #15218

include/os/freebsd/spl/sys/mod_os.h

@@ -73,6 +73,9 @@
 #define	param_set_deadman_failmode_args(var) \
     CTLTYPE_STRING, NULL, 0, param_set_deadman_failmode, "A"
 
+#define	param_set_active_allocator_args(var) \
+    CTLTYPE_STRING, NULL, 0, param_set_active_allocator, "A"
+
 #define	param_set_deadman_synctime_args(var) \
     CTLTYPE_U64, NULL, 0, param_set_deadman_synctime, "QU"
 

include/sys/metaslab.h

@@ -39,6 +39,7 @@ extern "C" {
 
 
 typedef struct metaslab_ops {
+	const char *msop_name;
 	uint64_t (*msop_alloc)(metaslab_t *, uint64_t);
 } metaslab_ops_t;
 

include/sys/spa.h

@@ -1056,6 +1056,8 @@ extern uint64_t spa_deadman_synctime(spa_t *spa);
 extern uint64_t spa_deadman_ziotime(spa_t *spa);
 extern uint64_t spa_dirty_data(spa_t *spa);
 extern spa_autotrim_t spa_get_autotrim(spa_t *spa);
+extern int spa_get_allocator(spa_t *spa);
+extern void spa_set_allocator(spa_t *spa, const char *allocator);
 
 /* Miscellaneous support routines */
 extern void spa_load_failed(spa_t *spa, const char *fmt, ...)
@@ -1207,6 +1209,7 @@ int param_set_deadman_ziotime(ZFS_MODULE_PARAM_ARGS);
 int param_set_deadman_synctime(ZFS_MODULE_PARAM_ARGS);
 int param_set_slop_shift(ZFS_MODULE_PARAM_ARGS);
 int param_set_deadman_failmode(ZFS_MODULE_PARAM_ARGS);
+int param_set_active_allocator(ZFS_MODULE_PARAM_ARGS);
 
 #ifdef ZFS_DEBUG
 #define	dprintf_bp(bp, fmt, ...) do {				\

include/sys/spa_impl.h

@@ -263,6 +263,7 @@ struct spa {
 	 */
 	spa_alloc_t	*spa_allocs;
 	int		spa_alloc_count;
+	int		spa_active_allocator;	/* selectable allocator */
 
 	spa_aux_vdev_t	spa_spares;		/* hot spares */
 	spa_aux_vdev_t	spa_l2cache;		/* L2ARC cache devices */
@@ -467,6 +468,8 @@ extern int param_set_deadman_failmode_common(const char *val);
 extern void spa_set_deadman_synctime(hrtime_t ns);
 extern void spa_set_deadman_ziotime(hrtime_t ns);
 extern const char *spa_history_zone(void);
+extern const char *zfs_active_allocator;
+extern int param_set_active_allocator_common(const char *val);
 
 #ifdef	__cplusplus
 }

module/os/freebsd/zfs/sysctl_os.c

@@ -503,6 +503,24 @@ SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, max_idistance,
 
 /* metaslab.c */
 
+int
+param_set_active_allocator(SYSCTL_HANDLER_ARGS)
+{
+	char buf[16];
+	int rc;
+
+	if (req->newptr == NULL)
+		strlcpy(buf, zfs_active_allocator, sizeof (buf));
+
+	rc = sysctl_handle_string(oidp, buf, sizeof (buf), req);
+	if (rc || req->newptr == NULL)
+		return (rc);
+	if (strcmp(buf, zfs_active_allocator) == 0)
+		return (0);
+
+	return (param_set_active_allocator_common(buf));
+}
+
 /*
  * In pools where the log space map feature is not enabled we touch
  * multiple metaslabs (and their respective space maps) with each

module/os/linux/zfs/spa_misc_os.c

@@ -103,6 +103,18 @@ param_set_slop_shift(const char *buf, zfs_kernel_param_t *kp)
 	return (0);
 }
 
+int
+param_set_active_allocator(const char *val, zfs_kernel_param_t *kp)
+{
+	int error;
+
+	error = -param_set_active_allocator_common(val);
+	if (error == 0)
+		error = param_set_charp(val, kp);
+
+	return (error);
+}
+
 const char *
 spa_history_zone(void)
 {

module/zfs/metaslab.c

@@ -40,8 +40,6 @@
 #include <sys/zap.h>
 #include <sys/btree.h>
 
-#define	WITH_DF_BLOCK_ALLOCATOR
-
 #define	GANG_ALLOCATION(flags) \
 	((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER))
 
@@ -1622,9 +1620,6 @@ metaslab_block_find(zfs_btree_t *t, range_tree_t *rt, uint64_t start,
 	return (rs);
 }
 
-#if defined(WITH_DF_BLOCK_ALLOCATOR) || \
-    defined(WITH_CF_BLOCK_ALLOCATOR)
-
 /*
  * This is a helper function that can be used by the allocator to find a
  * suitable block to allocate. This will search the specified B-tree looking
@@ -1659,9 +1654,74 @@ metaslab_block_picker(range_tree_t *rt, uint64_t *cursor, uint64_t size,
 	*cursor = 0;
 	return (-1ULL);
 }
-#endif /* WITH_DF/CF_BLOCK_ALLOCATOR */
 
-#if defined(WITH_DF_BLOCK_ALLOCATOR)
+static uint64_t metaslab_df_alloc(metaslab_t *msp, uint64_t size);
+static uint64_t metaslab_cf_alloc(metaslab_t *msp, uint64_t size);
+static uint64_t metaslab_ndf_alloc(metaslab_t *msp, uint64_t size);
+metaslab_ops_t *metaslab_allocator(spa_t *spa);
+
+static metaslab_ops_t metaslab_allocators[] = {
+	{ "dynamic", metaslab_df_alloc },
+	{ "cursor", metaslab_cf_alloc },
+	{ "new-dynamic", metaslab_ndf_alloc },
+};
+
+static int
+spa_find_allocator_byname(const char *val)
+{
+	int a = ARRAY_SIZE(metaslab_allocators) - 1;
+	if (strcmp("new-dynamic", val) == 0)
+		return (-1); /* remove when ndf is working */
+	for (; a >= 0; a--) {
+		if (strcmp(val, metaslab_allocators[a].msop_name) == 0)
+			return (a);
+	}
+	return (-1);
+}
+
+void
+spa_set_allocator(spa_t *spa, const char *allocator)
+{
+	int a = spa_find_allocator_byname(allocator);
+	if (a < 0) a = 0;
+	spa->spa_active_allocator = a;
+	zfs_dbgmsg("spa allocator: %s\n", metaslab_allocators[a].msop_name);
+}
+
+int
+spa_get_allocator(spa_t *spa)
+{
+	return (spa->spa_active_allocator);
+}
+
+#if defined(_KERNEL)
+int
+param_set_active_allocator_common(const char *val)
+{
+	char *p;
+
+	if (val == NULL)
+		return (SET_ERROR(EINVAL));
+
+	if ((p = strchr(val, '\n')) != NULL)
+		*p = '\0';
+
+	int a = spa_find_allocator_byname(val);
+	if (a < 0)
+		return (SET_ERROR(EINVAL));
+
+	zfs_active_allocator = metaslab_allocators[a].msop_name;
+	return (0);
+}
+#endif
+
+metaslab_ops_t *
+metaslab_allocator(spa_t *spa)
+{
+	int allocator = spa_get_allocator(spa);
+	return (&metaslab_allocators[allocator]);
+}
+
 /*
  * ==========================================================================
  * Dynamic Fit (df) block allocator
@@ -1736,12 +1796,6 @@ metaslab_df_alloc(metaslab_t *msp, uint64_t size)
 	return (offset);
 }
 
-const metaslab_ops_t zfs_metaslab_ops = {
-	metaslab_df_alloc
-};
-#endif /* WITH_DF_BLOCK_ALLOCATOR */
-
-#if defined(WITH_CF_BLOCK_ALLOCATOR)
 /*
  * ==========================================================================
  * Cursor fit block allocator -
@@ -1784,12 +1838,6 @@ metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
 	return (offset);
 }
 
-const metaslab_ops_t zfs_metaslab_ops = {
-	metaslab_cf_alloc
-};
-#endif /* WITH_CF_BLOCK_ALLOCATOR */
-
-#if defined(WITH_NDF_BLOCK_ALLOCATOR)
 /*
  * ==========================================================================
  * New dynamic fit allocator -
@@ -1846,12 +1894,6 @@ metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
 	return (-1ULL);
 }
 
-const metaslab_ops_t zfs_metaslab_ops = {
-	metaslab_ndf_alloc
-};
-#endif /* WITH_NDF_BLOCK_ALLOCATOR */
-
-
 /*
  * ==========================================================================
  * Metaslabs
@@ -6232,3 +6274,9 @@ ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, try_hard_before_gang, INT,
 
 ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, find_max_tries, UINT, ZMOD_RW,
 	"Normally only consider this many of the best metaslabs in each vdev");
+
+/* BEGIN CSTYLED */
+ZFS_MODULE_PARAM_CALL(zfs, zfs_, active_allocator,
+	param_set_active_allocator, param_get_charp, ZMOD_RW,
+	"SPA active allocator");
+/* END CSTYLED */

module/zfs/spa.c

@@ -1295,24 +1295,26 @@ spa_thread(void *arg)
 }
 #endif
 
+extern metaslab_ops_t *metaslab_allocator(spa_t *spa);
+
 /*
  * Activate an uninitialized pool.
  */
 static void
 spa_activate(spa_t *spa, spa_mode_t mode)
 {
+	metaslab_ops_t *msp = metaslab_allocator(spa);
 	ASSERT(spa->spa_state == POOL_STATE_UNINITIALIZED);
 
 	spa->spa_state = POOL_STATE_ACTIVE;
 	spa->spa_mode = mode;
 	spa->spa_read_spacemaps = spa_mode_readable_spacemaps;
 
-	spa->spa_normal_class = metaslab_class_create(spa, &zfs_metaslab_ops);
-	spa->spa_log_class = metaslab_class_create(spa, &zfs_metaslab_ops);
-	spa->spa_embedded_log_class =
-	    metaslab_class_create(spa, &zfs_metaslab_ops);
-	spa->spa_special_class = metaslab_class_create(spa, &zfs_metaslab_ops);
-	spa->spa_dedup_class = metaslab_class_create(spa, &zfs_metaslab_ops);
+	spa->spa_normal_class = metaslab_class_create(spa, msp);
+	spa->spa_log_class = metaslab_class_create(spa, msp);
+	spa->spa_embedded_log_class = metaslab_class_create(spa, msp);
+	spa->spa_special_class = metaslab_class_create(spa, msp);
+	spa->spa_dedup_class = metaslab_class_create(spa, msp);
 
 	/* Try to create a covering process */
 	mutex_enter(&spa->spa_proc_lock);

module/zfs/spa_misc.c

@@ -389,6 +389,11 @@ static const uint64_t spa_min_slop = 128ULL * 1024 * 1024;
 static const uint64_t spa_max_slop = 128ULL * 1024 * 1024 * 1024;
 static const int spa_allocators = 4;
 
+/*
+ * Spa active allocator.
+ * Valid values are zfs_active_allocator=<dynamic|cursor|new-dynamic>.
+ */
+const char *zfs_active_allocator = "dynamic";
 
 void
 spa_load_failed(spa_t *spa, const char *fmt, ...)
@@ -710,6 +715,7 @@ spa_add(const char *name, nvlist_t *config, const char *altroot)
 	spa->spa_deadman_synctime = MSEC2NSEC(zfs_deadman_synctime_ms);
 	spa->spa_deadman_ziotime = MSEC2NSEC(zfs_deadman_ziotime_ms);
 	spa_set_deadman_failmode(spa, zfs_deadman_failmode);
+	spa_set_allocator(spa, zfs_active_allocator);
 
 	zfs_refcount_create(&spa->spa_refcount);
 	spa_config_lock_init(spa);