[PATCH 1/2] soundwire: bus: Prevent lockdep asserts when stream has multiple buses
Give the bus_lock and msg_lock of each bus a different unique key so that it is possible to acquire the locks of multiple buses without lockdep asserting a possible deadlock.
Using mutex_init() to initialize a mutex gives all those mutexes the same lock class. Lockdep checking treats it as an error to attempt to take a mutex while already holding a mutex of the same class. This causes a lockdep assert when sdw_acquire_bus_lock() attempts to lock multiple buses, and when do_bank_switch() takes multiple msg_lock.
[ 138.697350] WARNING: possible recursive locking detected [ 138.697366] 6.3.0-test #1 Tainted: G E [ 138.697380] -------------------------------------------- [ 138.697394] play/903 is trying to acquire lock: [ 138.697409] ffff99b8c41aa8c8 (&bus->bus_lock){+.+.}-{3:3}, at: sdw_prepare_stream+0x52/0x2e0 [ 138.697443] but task is already holding lock: [ 138.697468] ffff99b8c41af8c8 (&bus->bus_lock){+.+.}-{3:3}, at: sdw_prepare_stream+0x52/0x2e0 [ 138.697493] other info that might help us debug this: [ 138.697521] Possible unsafe locking scenario:
[ 138.697540] CPU0 [ 138.697550] ---- [ 138.697559] lock(&bus->bus_lock); [ 138.697570] lock(&bus->bus_lock); [ 138.697581] *** DEADLOCK ***
Giving each mutex a unique key allows multiple to be held without triggering a lockdep assert. But note that it does not allow them to be taken in one order then a different order. If two mutexes are taken in the order A, B then they must always be taken in that order otherwise they could deadlock.
Signed-off-by: Richard Fitzgerald rf@opensource.cirrus.com --- drivers/soundwire/bus.c | 15 +++++++++++++-- include/linux/soundwire/sdw.h | 3 +++ 2 files changed, 16 insertions(+), 2 deletions(-)
diff --git a/drivers/soundwire/bus.c b/drivers/soundwire/bus.c index b44f8d0affa6..dba920ec88f6 100644 --- a/drivers/soundwire/bus.c +++ b/drivers/soundwire/bus.c @@ -69,8 +69,17 @@ int sdw_bus_master_add(struct sdw_bus *bus, struct device *parent, return -EINVAL; }
- mutex_init(&bus->msg_lock); - mutex_init(&bus->bus_lock); + /* + * Give each bus_lock and msg_lock a unique key so that lockdep won't + * trigger a deadlock warning when the locks of several buses are + * grabbed during configuration of a multi-bus stream. + */ + lockdep_register_key(&bus->msg_lock_key); + __mutex_init(&bus->msg_lock, "msg_lock", &bus->msg_lock_key); + + lockdep_register_key(&bus->bus_lock_key); + __mutex_init(&bus->bus_lock, "bus_lock", &bus->bus_lock_key); + INIT_LIST_HEAD(&bus->slaves); INIT_LIST_HEAD(&bus->m_rt_list);
@@ -181,6 +190,8 @@ void sdw_bus_master_delete(struct sdw_bus *bus) sdw_master_device_del(bus);
sdw_bus_debugfs_exit(bus); + lockdep_unregister_key(&bus->bus_lock_key); + lockdep_unregister_key(&bus->msg_lock_key); ida_free(&sdw_bus_ida, bus->id); } EXPORT_SYMBOL(sdw_bus_master_delete); diff --git a/include/linux/soundwire/sdw.h b/include/linux/soundwire/sdw.h index c076a3f879b3..f523ceabd059 100644 --- a/include/linux/soundwire/sdw.h +++ b/include/linux/soundwire/sdw.h @@ -5,6 +5,7 @@ #define __SOUNDWIRE_H
#include <linux/bug.h> +#include <linux/lockdep_types.h> #include <linux/mod_devicetable.h> #include <linux/bitfield.h>
@@ -907,7 +908,9 @@ struct sdw_bus { struct list_head slaves; DECLARE_BITMAP(assigned, SDW_MAX_DEVICES); struct mutex bus_lock; + struct lock_class_key bus_lock_key; struct mutex msg_lock; + struct lock_class_key msg_lock_key; int (*compute_params)(struct sdw_bus *bus); const struct sdw_master_ops *ops; const struct sdw_master_port_ops *port_ops;
Always add buses to the stream->master_list in a fixed order. The unique bus->id is used to order the adding of buses to the list.
This prevents lockdep asserts and possible deadlocks on streams that have multiple buses.
sdw_acquire_bus_lock() takes bus_lock in the order that buses are listed in stream->master_list. do_bank_switch() takes all the msg_lock in the same order.
To prevent a lockdep assert, and a possible real deadlock, the relative order of taking these mutexes must always be the same.
For example, if a stream takes the mutexes in the order (bus0, bus1) lockdep will assert if another stream takes them in the order (bus1, bus0).
More complex relative ordering will also assert, for example if two streams take (bus0, bus1) and (bus1, bus2), then a third stream takes (bus2, bus0).
Previously sdw_stream_add_master() simply added the given bus to the end of the list, requiring the caller to guarantee that buses are added in a fixed order. This isn't reasonable or necessary - it's an internal implementation detail that should not be exposed by the API. It doesn't really make sense when there could be multiple independent calling drivers, to say "you must add your buses in the same order as a different driver, that you don't know about, added them".
Signed-off-by: Richard Fitzgerald rf@opensource.cirrus.com --- drivers/soundwire/stream.c | 18 ++++++++++++++++-- 1 file changed, 16 insertions(+), 2 deletions(-)
diff --git a/drivers/soundwire/stream.c b/drivers/soundwire/stream.c index 93baca08a0de..d77a8a0d42c8 100644 --- a/drivers/soundwire/stream.c +++ b/drivers/soundwire/stream.c @@ -1150,7 +1150,8 @@ static struct sdw_master_runtime *sdw_master_rt_alloc(struct sdw_bus *bus, struct sdw_stream_runtime *stream) { - struct sdw_master_runtime *m_rt; + struct sdw_master_runtime *m_rt, *walk_m_rt; + struct list_head *insert_after;
m_rt = kzalloc(sizeof(*m_rt), GFP_KERNEL); if (!m_rt) @@ -1159,7 +1160,20 @@ static struct sdw_master_runtime /* Initialization of Master runtime handle */ INIT_LIST_HEAD(&m_rt->port_list); INIT_LIST_HEAD(&m_rt->slave_rt_list); - list_add_tail(&m_rt->stream_node, &stream->master_list); + + /* + * Add in order of bus id so that when taking the bus_lock + * of multiple buses they will always be taken in the same + * order to prevent a mutex deadlock. + */ + insert_after = &stream->master_list; + list_for_each_entry_reverse(walk_m_rt, &stream->master_list, stream_node) { + if (walk_m_rt->bus->id < bus->id) { + insert_after = &walk_m_rt->stream_node; + break; + } + } + list_add(&m_rt->stream_node, insert_after);
list_add_tail(&m_rt->bus_node, &bus->m_rt_list);
On 6/15/23 16:12, Richard Fitzgerald wrote:
Always add buses to the stream->master_list in a fixed order. The unique bus->id is used to order the adding of buses to the list.
This prevents lockdep asserts and possible deadlocks on streams that have multiple buses.
sdw_acquire_bus_lock() takes bus_lock in the order that buses are listed in stream->master_list. do_bank_switch() takes all the msg_lock in the same order.
To prevent a lockdep assert, and a possible real deadlock, the relative order of taking these mutexes must always be the same.
For example, if a stream takes the mutexes in the order (bus0, bus1) lockdep will assert if another stream takes them in the order (bus1, bus0).
More complex relative ordering will also assert, for example if two streams take (bus0, bus1) and (bus1, bus2), then a third stream takes (bus2, bus0).
Previously sdw_stream_add_master() simply added the given bus to the end of the list, requiring the caller to guarantee that buses are added in a fixed order. This isn't reasonable or necessary - it's an internal implementation detail that should not be exposed by the API. It doesn't really make sense when there could be multiple independent calling drivers, to say "you must add your buses in the same order as a different driver, that you don't know about, added them".
Makes sense to me. The other way to look at this is that the notion of 'stream' and dailink are virtually synonyms, and 'sdw_stream_add_master' is called from each DAI of a dailink, hence in a fixed order. But nothing really defines how dailinks include the dais, and in a hypothetical case with multiple controllers, each with multiple links, there would be an ambiguity anyways so using the ida-allocated bus->id is a good solution indeed.
Reviewed-by: Pierre-Louis Bossart pierre-louis.bossart@linux.intel.com
On 6/15/23 16:12, Richard Fitzgerald wrote:
Give the bus_lock and msg_lock of each bus a different unique key so that it is possible to acquire the locks of multiple buses without lockdep asserting a possible deadlock.
Using mutex_init() to initialize a mutex gives all those mutexes the same lock class. Lockdep checking treats it as an error to attempt to take a mutex while already holding a mutex of the same class. This causes a lockdep assert when sdw_acquire_bus_lock() attempts to lock multiple buses, and when do_bank_switch() takes multiple msg_lock.
[ 138.697350] WARNING: possible recursive locking detected [ 138.697366] 6.3.0-test #1 Tainted: G E [ 138.697380] -------------------------------------------- [ 138.697394] play/903 is trying to acquire lock: [ 138.697409] ffff99b8c41aa8c8 (&bus->bus_lock){+.+.}-{3:3}, at: sdw_prepare_stream+0x52/0x2e0 [ 138.697443] but task is already holding lock: [ 138.697468] ffff99b8c41af8c8 (&bus->bus_lock){+.+.}-{3:3}, at: sdw_prepare_stream+0x52/0x2e0 [ 138.697493] other info that might help us debug this: [ 138.697521] Possible unsafe locking scenario:
[ 138.697540] CPU0 [ 138.697550] ---- [ 138.697559] lock(&bus->bus_lock); [ 138.697570] lock(&bus->bus_lock); [ 138.697581] *** DEADLOCK ***
Giving each mutex a unique key allows multiple to be held without triggering a lockdep assert. But note that it does not allow them to be taken in one order then a different order. If two mutexes are taken in the order A, B then they must always be taken in that order otherwise they could deadlock.
Signed-off-by: Richard Fitzgerald rf@opensource.cirrus.com
Reviewed-by: Pierre-Louis Bossart pierre-louis.bossart@linux.intel.com
drivers/soundwire/bus.c | 15 +++++++++++++-- include/linux/soundwire/sdw.h | 3 +++ 2 files changed, 16 insertions(+), 2 deletions(-)
diff --git a/drivers/soundwire/bus.c b/drivers/soundwire/bus.c index b44f8d0affa6..dba920ec88f6 100644 --- a/drivers/soundwire/bus.c +++ b/drivers/soundwire/bus.c @@ -69,8 +69,17 @@ int sdw_bus_master_add(struct sdw_bus *bus, struct device *parent, return -EINVAL; }
- mutex_init(&bus->msg_lock);
- mutex_init(&bus->bus_lock);
- /*
* Give each bus_lock and msg_lock a unique key so that lockdep won't
* trigger a deadlock warning when the locks of several buses are
* grabbed during configuration of a multi-bus stream.
*/
- lockdep_register_key(&bus->msg_lock_key);
- __mutex_init(&bus->msg_lock, "msg_lock", &bus->msg_lock_key);
- lockdep_register_key(&bus->bus_lock_key);
- __mutex_init(&bus->bus_lock, "bus_lock", &bus->bus_lock_key);
- INIT_LIST_HEAD(&bus->slaves); INIT_LIST_HEAD(&bus->m_rt_list);
@@ -181,6 +190,8 @@ void sdw_bus_master_delete(struct sdw_bus *bus) sdw_master_device_del(bus);
sdw_bus_debugfs_exit(bus);
- lockdep_unregister_key(&bus->bus_lock_key);
- lockdep_unregister_key(&bus->msg_lock_key); ida_free(&sdw_bus_ida, bus->id);
} EXPORT_SYMBOL(sdw_bus_master_delete); diff --git a/include/linux/soundwire/sdw.h b/include/linux/soundwire/sdw.h index c076a3f879b3..f523ceabd059 100644 --- a/include/linux/soundwire/sdw.h +++ b/include/linux/soundwire/sdw.h @@ -5,6 +5,7 @@ #define __SOUNDWIRE_H
#include <linux/bug.h> +#include <linux/lockdep_types.h> #include <linux/mod_devicetable.h> #include <linux/bitfield.h>
@@ -907,7 +908,9 @@ struct sdw_bus { struct list_head slaves; DECLARE_BITMAP(assigned, SDW_MAX_DEVICES); struct mutex bus_lock;
- struct lock_class_key bus_lock_key; struct mutex msg_lock;
- struct lock_class_key msg_lock_key; int (*compute_params)(struct sdw_bus *bus); const struct sdw_master_ops *ops; const struct sdw_master_port_ops *port_ops;
On 15-06-23, 15:12, Richard Fitzgerald wrote:
Give the bus_lock and msg_lock of each bus a different unique key so that it is possible to acquire the locks of multiple buses without lockdep asserting a possible deadlock.
Using mutex_init() to initialize a mutex gives all those mutexes the same lock class. Lockdep checking treats it as an error to attempt to take a mutex while already holding a mutex of the same class. This causes a lockdep assert when sdw_acquire_bus_lock() attempts to lock multiple buses, and when do_bank_switch() takes multiple msg_lock.
Applied both, thanks
participants (3)
-
Pierre-Louis Bossart
-
Richard Fitzgerald
-
Vinod Koul