ALSA: seq: Wait for one RCU grace period when deleting all ports
PATCH 1/1 ALSA: seq: Wait for one RCU grace period when deleting all ports
Since the port list was converted to RCU, snd_seq_delete_all_ports() calls synchronize_rcu() once per port under register_mutex. Closing a client with 254 ports now takes about 1.6s and every other client open or close waits for it.
Unlink all ports first, wait for a single grace period and then free them.
---
Tested in QEMU on v7.3 with lockdep and KASAN. Closing a client with 254 ports went from ~1.6s to ~0.02s, and a second client doing open/close in a loop no longer stalls. Closing the client while other clients send to and query its ports ran 100 rounds without reports.
sound/core/seq/seq_ports.c | 41 +++++++++++++++++++++++++------------- sound/core/seq/seq_ports.h | 1 + 2 files changed, 28 insertions(+), 14 deletions(-)
sound/core/seq/seq_ports.c +27 -14
@@ -248,17 +248,9 @@ static void clear_subscriber_list(struct snd_seq_client *client, }/* delete port data */ -static int port_delete (struct snd_seq_client *client, -struct snd_seq_client_port *port) +static void port_free (struct snd_seq_client *client, +struct snd_seq_client_port *port) { -/* set closing flag and wait for all port access are gone */ - port->closing =1 ; -/* the port has already been unlinked from the client's port list; -* wait for a grace period so that RCU readers still traversing the -* list can no longer take a new use_lock reference, then drain the -* outstanding references before freeing -*/ -synchronize_rcu ();snd_use_lock_sync (&port->use_lock);/* clear subscribers info */
@@ -272,6 +264,20 @@ static int port_delete(struct snd_seq_client *client,snd_BUG_ON (port->c_dest.count !=0 );kfree (port); +} + +static int port_delete (struct snd_seq_client *client, +struct snd_seq_client_port *port) +{ +/* set closing flag and wait for all port access are gone */ + port->closing =1 ; +/* the port has already been unlinked from the client's port list; +* wait for a grace period so that RCU readers still traversing the +* list can no longer take a new use_lock reference, then drain the +* outstanding references before freeing +*/ +synchronize_rcu (); +port_free (client, port);return 0 ; }
@@ -302,18 +308,25 @@ int snd_seq_delete_port(struct snd_seq_client *client, int port)int snd_seq_delete_all_ports (struct snd_seq_client *client) {struct snd_seq_client_port *port, *tmp; +LIST_HEAD (deleted); -/* unlink and delete each port; port_delete() waits for an RCU grace -* period before draining the port, so concurrent lockless readers can -* no longer take a new use_lock reference on it +/* unlink all ports first and wait for a single RCU grace period +* before draining and freeing them, so concurrent lockless readers +* can no longer take a new use_lock reference on any of them */ guard (mutex)(&client->ports_mutex);list_for_each_entry_safe (port, tmp, &client->ports_list_head, list) {list_del_rcu (&port->list); client->num_ports--; + port->closing =1 ; +list_add_tail (&port->delete_list, &deleted);snd_seq_system_client_ev_port_exit (port->addr.client, port->addr.port); -port_delete (client, port); } +if (list_empty (&deleted)) +return 0 ; +synchronize_rcu (); +list_for_each_entry_safe (port, tmp, &deleted, delete_list) +port_free (client, port);return 0 ; }
sound/core/seq/seq_ports.h +1 -0
@@ -49,6 +49,7 @@ struct snd_seq_client_port {struct module *owner;/* owner of this port */ char name[64 ];/* port name */ struct list_head list;/* port list */ +struct list_head delete_list;/* used by snd_seq_delete_all_ports() */ snd_use_lock_t use_lock;/* subscribers */
Signature · 1 line
2.55.0
Quoted 9 lines · Ali Ahmet Memiş
Since the port list was converted to RCU,
snd_seq_delete_all_ports()callssynchronize_rcu()once per port underregister_mutex. Closing a client with 254 ports now takes about 1.6s and every other client open or close waits for it.Unlink all ports first, wait for a single grace period and then free them.
Quoted 7 lines
Fixes 4c252fbc06d6 ("ALSA: seq: Use RCU for the client port list")Signed-off-by Ali Ahmet Memiş <[email protected]>---
Tested in QEMU on v7.3 with lockdep and KASAN. Closing a client with 254 ports went from ~1.6s to ~0.02s, and a second client doing open/close in a loop no longer stalls. Closing the client while other clients send to and query its ports ran 100 rounds without reports.
Applied now to for-next branch.
Sounds like a good performance improvement.
thanks,
Takashi