[PATCH 6/6] md/raid5: derive the default group_thread_cnt from the hardware

[Hiroshi Nishida]

setup_conf() starts every array with worker_cnt_per_group = 0, i.e. a
single raid5d thread and no worker groups.  On a many-core host backing a
fast, wide array that single thread is the bottleneck; the multi-threading
that group_thread_cnt enables has to be turned on by hand, per array,
after every assembly.

Pick a starting value from the CPU count instead.  When the new
group_thread_cnt_default module parameter is left at -1 (the default),
raid5_default_group_thread_cnt() uses half the online CPU count spread
across the NUMA nodes, capped at 256:

	gtc = num_online_cpus() / (2 * nr_nodes)

The count is per NUMA node, matching alloc_thread_groups(), so dividing by
the node count keeps the total near half the CPUs regardless of socket
count.  It is only a ceiling: raid5_wakeup_stripe_thread() wakes workers
in proportion to the queued stripe count, so a lightly loaded array uses
far fewer than the maximum.  A lone worker is not worth its overhead over
raid5d, so a result of 1 collapses back to 0, which keeps boxes with two
or fewer CPUs single-threaded.

Measured on a 16-disk raid6 array of NVMe SSDs (a 32-vCPU host: 16 cores,
two NUMA nodes; steady state, interleaved runs): the derived count of 8
raises throughput 2.1-3.2x over the single-threaded default -- 4K random
write from ~39k to ~100k IOPS (2.6x), mixed database/OLTP/high-concurrency
2.1-2.5x, partial stripe write 3.2x.  There is no regression on smaller
machines: a 4-CPU box gets 2 workers and is faster or equal on every
workload, and a box with two or fewer CPUs gets 0 and is byte-for-byte
unchanged.

group_thread_cnt_default overrides the heuristic (0 forces the historical
single-threaded behaviour; a positive value pins a count, capped at 256),
and the existing per-array group_thread_cnt sysfs attribute continues to
override it at runtime and still accepts up to its own 8192 limit.

Signed-off-by: Hiroshi Nishida <nishidafmly@xxxxxxxxx>
---
 drivers/md/raid5.c | 45 ++++++++++++++++++++++++++++++++++++++++++---
 1 file changed, 42 insertions(+), 3 deletions(-)

diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c
index 8e1c2eba4241..6e91eb0ad575 100644
--- a/drivers/md/raid5.c
+++ b/drivers/md/raid5.c
@@ -88,6 +88,11 @@ module_param(stripe_cache_size_default, uint, 0644);
 MODULE_PARM_DESC(stripe_cache_size_default,
 		 "Initial stripe_cache_size for newly created arrays.  0 (the default) auto-sizes it from system memory: the historical 256 on small hosts, scaling up with RAM to a capped maximum on larger ones.  A non-zero value sets a fixed initial size.  Existing arrays are unaffected");
 
+static int group_thread_cnt_default = -1;
+module_param(group_thread_cnt_default, int, 0644);
+MODULE_PARM_DESC(group_thread_cnt_default,
+		 "Initial group_thread_cnt (raid5 worker threads per NUMA node) for newly created arrays.  A negative value (the default, -1) auto-sizes it from the CPU count; 0 forces the single-threaded raid5d; a positive value sets a fixed count (capped at 256).  The per-array group_thread_cnt sysfs attribute overrides this and allows larger values");
+
 static bool devices_handle_discard_safely = false;
 module_param(devices_handle_discard_safely, bool, 0644);
 MODULE_PARM_DESC(devices_handle_discard_safely,
@@ -7584,6 +7589,31 @@ static unsigned long raid5_cache_count(struct shrinker *shrink,
 #define RAID5_CACHE_DEFAULT_RAM_SHIFT	9	/* above it: ~1/512 of the extra RAM */
 #define RAID5_CACHE_DEFAULT_MAX		4096
 
+/*
+ * Default group_thread_cnt (worker_cnt_per_group) for a new array when the
+ * group_thread_cnt_default module parameter is left at -1.  The historical
+ * default is 0 -- a single raid5d thread -- which cannot keep a fast, wide
+ * array busy on a many-core host.  Derive a starting point from the CPU
+ * count: half the online CPUs divided across the NUMA nodes (this is a
+ * per-node count -- see alloc_thread_groups() -- so the total lands near
+ * half the online CPUs regardless of socket count), capped at
+ * RAID5_AUTO_GROUP_THREAD_MAX.  This is only a ceiling:
+ * raid5_wakeup_stripe_thread() wakes workers in proportion to the queued
+ * stripe count, so a lightly loaded array uses far fewer.  A lone worker is
+ * not worth its overhead, so 1 collapses back to 0.  The group_thread_cnt
+ * sysfs attribute overrides this per array.
+ */
+#define RAID5_AUTO_GROUP_THREAD_MAX	256
+
+static int raid5_default_group_thread_cnt(void)
+{
+	unsigned int gtc = num_online_cpus() / (2 * num_possible_nodes());
+
+	if (gtc > RAID5_AUTO_GROUP_THREAD_MAX)
+		gtc = RAID5_AUTO_GROUP_THREAD_MAX;
+	return gtc == 1 ? 0 : gtc;
+}
+
 static struct r5conf *setup_conf(struct mddev *mddev)
 {
 	struct r5conf *conf;
@@ -7594,6 +7624,7 @@ static struct r5conf *setup_conf(struct mddev *mddev)
 	int i;
 	int group_cnt;
 	struct r5worker_group *new_group;
+	int def_threads;
 	int ret = -ENOMEM;
 
 	if (mddev->new_level != 5
@@ -7677,10 +7708,18 @@ static struct r5conf *setup_conf(struct mddev *mddev)
 		goto abort;
 	for (i = 0; i < PENDING_IO_MAX; i++)
 		list_add(&conf->pending_data[i].sibling, &conf->free_list);
-	/* Don't enable multi-threading by default*/
-	if (!alloc_thread_groups(conf, 0, &group_cnt, &new_group)) {
+	/*
+	 * Multi-threading defaults to a hardware-derived worker count (see
+	 * raid5_default_group_thread_cnt()); group_thread_cnt_default overrides
+	 * the choice, and the group_thread_cnt sysfs attribute overrides it per
+	 * array.
+	 */
+	def_threads = group_thread_cnt_default < 0 ?
+		raid5_default_group_thread_cnt() :
+		min(group_thread_cnt_default, RAID5_AUTO_GROUP_THREAD_MAX);
+	if (!alloc_thread_groups(conf, def_threads, &group_cnt, &new_group)) {
 		conf->group_cnt = group_cnt;
-		conf->worker_cnt_per_group = 0;
+		conf->worker_cnt_per_group = def_threads;
 		conf->worker_groups = new_group;
 	} else
 		goto abort;
-- 
2.43.0