ALSA/ASoC: hda: Fix race between unsol events and unbind
cover ALSA/ASoC: hda: Fix race between unsol events and unbind
Here is another series of small fixes for HD-audio, beginning with the fix for an issue reported by Sashiko, and extended the fixes to ASoC HD-audio and a potential improvement that were spotted during the development of the first patch.
Takashi
===
Takashi Iwai (3):
ALSA: hda: Stop unsol events and jack polling before codec unbind
ASoC: codecs: hda: Stop unsol events and jack polling before codec
removal
ALSA: hda: Make hdac_device.registered a plain bool
include/sound/hdaudio.h | 4 +++- sound/hda/common/bind.c | 7 +++++++ sound/hda/common/codec.c | 4 ++-- sound/hda/core/bus.c | 21 ++++++++++++++++++++- sound/soc/codecs/hda.c | 7 +++++++ 5 files changed, 39 insertions(+), 4 deletions(-)
Signature · 1 line
2.55.0
PATCH 1/3 ALSA: hda: Stop unsol events and jack polling before codec unbind
At unbinding a codec driver, hda_codec_driver_remove() calls the codec's remove callback that releases the driver resources, followed by snd_hda_codec_cleanup_for_unbind(). Meanwhile, the unsolicited events are processed asynchronously in bus->unsol_work, and snd_hdac_bus_process_unsol_events() checks only codec->registered flag before calling the driver's unsol_event callback without the lock. Since codec->registered is cleared only in snd_hda_codec_cleanup_for_unbind(), and there is no flush of the unsol work at unbinding, the unsol event handler may run concurrently during the running remove callback, which may lead to a UAF. The same problem applies to the jack polling work, which is canceled only after the remove callback.
For addressing those races, introduce a new flag unsol_disabled to hdac_device, to be checked it in the unsol event worker, while a new helper snd_hdac_device_disable_unsol() sets this flag and flushes the pending unsol work. The helper is called at hda_codec_driver_remove() together with the cancel of jackpoll_work to assure that no asynchronous jack handling can run. The flag is cleared again at probing the codec driver, while the events are still blocked by the registered flag until the codec gets registered.
--- include/sound/hdaudio.h | 2 ++ sound/hda/common/bind.c | 7 +++++++ sound/hda/core/bus.c | 21 ++++++++++++++++++++- 3 files changed, 29 insertions(+), 1 deletion(-)
include/sound/hdaudio.h +2 -0
@@ -97,6 +97,7 @@ struct hdac_device { bool caps_overwriting:1; /* caps overwrite being in process */ bool cache_coef:1; /* cache COEF read/write too */ unsigned int registered:1; /* codec was registered */ +bool unsol_disabled;/* unsol events blocked; protected by bus->reg_lock */ };/* device/driver type used for matching */
@@ -123,6 +124,7 @@ int snd_hdac_device_init(struct hdac_device *dev, struct hdac_bus *bus, const char *name, unsigned int addr);void snd_hdac_device_exit (struct hdac_device *dev);int snd_hdac_device_register (struct hdac_device *codec); +void snd_hdac_device_disable_unsol (struct hdac_device *codec);void snd_hdac_device_unregister (struct hdac_device *codec);int snd_hdac_device_set_chip_name (struct hdac_device *codec,const char *name);int snd_hdac_codec_modalias (const struct hdac_device *hdac,char *buf,size_t size);
sound/hda/common/bind.c +7 -0
@@ -100,6 +100,9 @@ static int hda_codec_driver_probe(struct device *dev)if (WARN_ON (!codec->preset))return -EINVAL ; +/* unsol events are still blocked until registered */ + codec->core.unsol_disabled = false; + err =snd_hda_codec_set_name (codec, codec->preset->name);if (err <0 ) goto error;
@@ -160,6 +163,10 @@ static int hda_codec_driver_remove(struct device *dev)return codec->bus->core.ext_ops->hdev_detach (&codec->core); } +/* stop asynchronous jack handling before freeing driver resources */ +snd_hdac_device_disable_unsol (&codec->core); +cancel_delayed_work_sync (&codec->jackpoll_work); +snd_hda_codec_disconnect_pcms (codec);snd_hda_jack_tbl_disconnect (codec);snd_refcount_sync (&codec->pcm_ref);
sound/hda/core/bus.c +20 -1
@@ -180,7 +180,7 @@ static void snd_hdac_bus_process_unsol_events(struct work_struct *work)if (!(caddr & (1 <<4 )))/* no unsolicited event? */ continue ; codec = bus->caddr_tbl[caddr &0x0f ]; -if (!codec || !codec->registered) +if (!codec || !codec->registered || codec->unsol_disabled)continue ;spin_unlock_irq (&bus->reg_lock); drv =drv_to_hdac_driver (codec->dev.driver);
@@ -191,6 +191,25 @@ static void snd_hdac_bus_process_unsol_events(struct work_struct *work)spin_unlock_irq (&bus->reg_lock); } +/** +* snd_hdac_device_disable_unsol - block and flush unsol events for the codec +* @codec: the HDA core device +* +* Stop dispatching the unsolicited events to the given codec, and wait for +* the pending unsol event handler to finish. Called at unbinding the codec +* driver before releasing the driver resources. +*/ +void snd_hdac_device_disable_unsol (struct hdac_device *codec) +{ +struct hdac_bus *bus = codec->bus; + +spin_lock_irq (&bus->reg_lock); + codec->unsol_disabled = true; +spin_unlock_irq (&bus->reg_lock); +flush_work (&bus->unsol_work); +} +EXPORT_SYMBOL_GPL (snd_hdac_device_disable_unsol); +/** * snd_hdac_bus_add_device - Add a codec to bus * @bus: HDA core bus
Signature · 1 line
2.55.0
Re: [PATCH 1/3] ALSA: hda: Stop unsol events and jack polling before codec unbind
Quoted 25 lines · Takashi Iwai
At unbinding a codec driver,
hda_codec_driver_remove()calls the codec's remove callback that releases the driver resources, followed bysnd_hda_codec_cleanup_for_unbind(). Meanwhile, the unsolicited events are processed asynchronously inbus->unsol_work, andsnd_hdac_bus_process_unsol_events()checks onlycodec->registeredflag before calling the driver'sunsol_eventcallback without the lock. Sincecodec->registeredis cleared only insnd_hda_codec_cleanup_for_unbind(), and there is no flush of the unsol work at unbinding, the unsol event handler may run concurrently during the running remove callback, which may lead to a UAF. The same problem applies to the jack polling work, which is canceled only after the remove callback.+++
b/sound/hda/common/bind.csound/hda/common/bind.c +3
-0@@ -100,6 +100,9 @@ static int hda_codec_driver_probe(struct device *dev)if (WARN_ON (!codec->preset))return -EINVAL ; +/* unsol events are still blocked until registered */ + codec->core.unsol_disabled = false; + err =snd_hda_codec_set_name (codec, codec->preset->name);if (err <0 ) goto error;
sound/hda/common/bind.c static int hda_codec_driver_remove(struct device *dev) · line 168
Quoted 6 lines
sound/hda/common/bind.c +3
-0@@ -160,6 +163,10 @@ static int hda_codec_driver_remove(struct device *dev)return codec->bus->core.ext_ops->hdev_detach (&codec->core); } +/* stop asynchronous jack handling before freeing driver resources */ +snd_hdac_device_disable_unsol (&codec->core); +cancel_delayed_work_sync (&codec->jackpoll_work);
IMHO this is a proof that ->registered flag doesn't do its job. snd_hda_codec_cleanup_for_unbind() is called right after and it does contain cancel_delayed_work_sync(&codec->jackpoll_work). It does contain snd_hda_codec_disconnect_pcms() and snd_hda_jack_tbl_disconnect() operations too.
I'd say it is snd_hda_codec_cleanup_for_unbind() that needs an update and perhaps be responsible for calling driver->remove() too.
Quoted 4 lines
sound/hda/common/bind.c +1
-0+snd_hda_codec_disconnect_pcms (codec);snd_hda_jack_tbl_disconnect (codec);snd_refcount_sync (&codec->pcm_ref);
Re: [PATCH 1/3] ALSA: hda: Stop unsol events and jack polling before codec unbind
Quoted 30 lines · Cezary Rojewski
Quoted 28 lines · Takashi Iwai
At unbinding a codec driver,
hda_codec_driver_remove()calls the codec's remove callback that releases the driver resources, followed bysnd_hda_codec_cleanup_for_unbind(). Meanwhile, the unsolicited events are processed asynchronously inbus->unsol_work, andsnd_hdac_bus_process_unsol_events()checks onlycodec->registeredflag before calling the driver'sunsol_eventcallback without the lock. Sincecodec->registeredis cleared only insnd_hda_codec_cleanup_for_unbind(), and there is no flush of the unsol work at unbinding, the unsol event handler may run concurrently during the running remove callback, which may lead to a UAF. The same problem applies to the jack polling work, which is canceled only after the remove callback.+++
b/sound/hda/common/bind.csound/hda/common/bind.c +3
-0@@ -100,6 +100,9 @@ static int hda_codec_driver_probe(struct device *dev)if (WARN_ON (!codec->preset))return -EINVAL ; +/* unsol events are still blocked until registered */ + codec->core.unsol_disabled = false; + err =snd_hda_codec_set_name (codec, codec->preset->name);if (err <0 ) goto error;@@ -160,6 +163,10 @@ static int hda_codec_driver_remove(struct device *dev)return codec->bus->core.ext_ops->hdev_detach (&codec->core); }
Quoted 10 lines
sound/hda/common/bind.c static int hda_codec_driver_remove(struct device *dev) · line 168
Quoted 3 lines
sound/hda/common/bind.c +3
-0+/* stop asynchronous jack handling before freeing driver resources */ +snd_hdac_device_disable_unsol (&codec->core); +cancel_delayed_work_sync (&codec->jackpoll_work);IMHO this is a proof that ->registered flag doesn't do its job.
snd_hda_codec_cleanup_for_unbind()is called right after and it does containcancel_delayed_work_sync(&codec->jackpoll_work). It does containsnd_hda_codec_disconnect_pcms()andsnd_hda_jack_tbl_disconnect()operations too.
It's about the pending works, and the flag doesn't help there alone unless they are synchronized; at the time point you flip the flag, the work might be already in-flight, so the flag switch may miss the chances.
Quoted 2 lines
I'd say it is
snd_hda_codec_cleanup_for_unbind()that needs an update and perhaps be responsible for callingdriver->remove()too.
Yeah, some code reorganization would be good.
thanks,
Takashi
Re: [PATCH 1/3] ALSA: hda: Stop unsol events and jack polling before codec unbind
Quoted 31 lines · Takashi Iwai
Quoted 18 lines · Cezary Rojewski
Quoted 15 lines · Takashi Iwai
+++
b/sound/hda/common/bind.csound/hda/common/bind.c +3
-0@@ -100,6 +100,9 @@ static int hda_codec_driver_probe(struct device *dev)if (WARN_ON (!codec->preset))return -EINVAL ; +/* unsol events are still blocked until registered */ + codec->core.unsol_disabled = false; + err =snd_hda_codec_set_name (codec, codec->preset->name);if (err <0 ) goto error;@@ -160,6 +163,10 @@ static int hda_codec_driver_remove(struct device *dev)return codec->bus->core.ext_ops->hdev_detach (&codec->core); }Quoted 10 lines
sound/hda/common/bind.c static int hda_codec_driver_remove(struct device *dev) · line 168
Quoted 3 lines
sound/hda/common/bind.c +3
-0+/* stop asynchronous jack handling before freeing driver resources */ +snd_hdac_device_disable_unsol (&codec->core); +cancel_delayed_work_sync (&codec->jackpoll_work);IMHO this is a proof that ->registered flag doesn't do its job.
snd_hda_codec_cleanup_for_unbind()is called right after and it does containcancel_delayed_work_sync(&codec->jackpoll_work). It does containsnd_hda_codec_disconnect_pcms()andsnd_hda_jack_tbl_disconnect()operations too.
Quoted 4 lines
It's about the pending works, and the flag doesn't help there alone unless they are synchronized; at the time point you flip the flag, the work might be already in-flight, so the flag switch may miss the chances.
I agree, my suggestion was to make ->registered the sole owner of the subject, including the synchronization. At the same time, fixing the problem first and expanding the work later is also a valid approach.
Do you have a method for reproducing the problem easily?
Quoted 4 lines
Quoted 2 lines
I'd say it is
snd_hda_codec_cleanup_for_unbind()that needs an update and perhaps be responsible for callingdriver->remove()too.Yeah, some code reorganization would be good.
Re: [PATCH 1/3] ALSA: hda: Stop unsol events and jack polling before codec unbind
Quoted 38 lines · Cezary Rojewski
Quoted 31 lines · Takashi Iwai
Quoted 18 lines · Cezary Rojewski
Quoted 15 lines · Takashi Iwai
+++
b/sound/hda/common/bind.csound/hda/common/bind.c +3
-0@@ -100,6 +100,9 @@ static int hda_codec_driver_probe(struct device *dev)if (WARN_ON (!codec->preset))return -EINVAL ; +/* unsol events are still blocked until registered */ + codec->core.unsol_disabled = false; + err =snd_hda_codec_set_name (codec, codec->preset->name);if (err <0 ) goto error;@@ -160,6 +163,10 @@ static int hda_codec_driver_remove(struct device *dev)return codec->bus->core.ext_ops->hdev_detach (&codec->core); }Quoted 10 lines
sound/hda/common/bind.c static int hda_codec_driver_remove(struct device *dev) · line 168
Quoted 3 lines
sound/hda/common/bind.c +3
-0+/* stop asynchronous jack handling before freeing driver resources */ +snd_hdac_device_disable_unsol (&codec->core); +cancel_delayed_work_sync (&codec->jackpoll_work);IMHO this is a proof that ->registered flag doesn't do its job.
snd_hda_codec_cleanup_for_unbind()is called right after and it does containcancel_delayed_work_sync(&codec->jackpoll_work). It does containsnd_hda_codec_disconnect_pcms()andsnd_hda_jack_tbl_disconnect()operations too.Quoted 4 lines
It's about the pending works, and the flag doesn't help there alone unless they are synchronized; at the time point you flip the flag, the work might be already in-flight, so the flag switch may miss the chances.
Quoted 5 lines
I agree, my suggestion was to make ->registered the sole owner of the subject, including the synchronization. At the same time, fixing the problem first and expanding the work later is also a valid approach.
Do you have a method for reproducing the problem easily?
No, it's currently only hypothetical.
I guess it can be done with userfaltd, though. Maybe AI can help for creating a reproducer.
Quoted 4 lines
Quoted 4 lines
Quoted 2 lines
I'd say it is
snd_hda_codec_cleanup_for_unbind()that needs an update and perhaps be responsible for callingdriver->remove()too.Yeah, some code reorganization would be good.
Looking forward to seeing patches! ;)
Takashi
PATCH 2/3 ASoC: codecs: hda: Stop unsol events and jack polling before codec removal
In reply to Takashi IwaiThe ASoC HD-audio codec driver has the same race as the legacy HDA codec driver unbind: hda_codec_remove() calls the codec's remove callback that releases the driver resources, followed by snd_hda_codec_cleanup_for_unbind(), while the unsolicited event work and the jack polling work may still run concurrently and access the released resources, leading to a UAF.
Apply the same fix as done for the legacy driver: call snd_hdac_device_disable_unsol() and cancel jackpoll_work at the beginning of hda_codec_remove(), and clear the unsol_disabled flag again at hda_codec_probe() so that the events get delivered after the codec is re-registered.
--- sound/soc/codecs/hda.c | 7 +++++++ 1 file changed, 7 insertions(+)
sound/soc/codecs/hda.c +7 -0
@@ -218,6 +218,9 @@ static int hda_codec_probe(struct snd_soc_component *component) goto err; } +/* unsol events are still blocked until registered */ + codec->core.unsol_disabled = false; + ret = driver->ops->probe (codec, codec->preset);if (ret <0 ) {dev_err (&hdev->dev,"codec init failed: %d\n" , ret);
@@ -276,6 +279,10 @@ static void hda_codec_remove(struct snd_soc_component *component)/* Don't allow any more runtime suspends */ pm_runtime_forbid (&hdev->dev); +/* stop asynchronous jack handling before freeing driver resources */ +snd_hdac_device_disable_unsol (hdev); +cancel_delayed_work_sync (&codec->jackpoll_work); +hda_codec_unregister_dais (codec, component);if (driver->ops->remove)
Signature · 1 line
2.55.0
PATCH 3/3 ALSA: hda: Make hdac_device.registered a plain bool
In reply to Takashi IwaiThe registered field of hdac_device is a bitfield sharing the same word with other bitfields like lazy_cache, and those are updated in a non-atomic read-modify-write manner without any lock. Meanwhile, the registered flag is read from the unsolicited event worker concurrently, so the concurrent updates of the neighboring bitfields may race with it.
Convert it to a plain bool for avoiding such a subtle race. Also replace the assignments of 0/1 with false/true accordingly.
--- include/sound/hdaudio.h | 2 +- sound/hda/common/codec.c | 4 ++-- 2 files changed, 3 insertions(+), 3 deletions(-)
include/sound/hdaudio.h +1 -1
@@ -96,7 +96,7 @@ struct hdac_device { bool lazy_cache:1; /* don't wake up for writes */ bool caps_overwriting:1; /* caps overwrite being in process */ bool cache_coef:1; /* cache COEF read/write too */ - unsigned int registered:1; /* codec was registered */ +bool registered;/* codec was registered */ bool unsol_disabled;/* unsol events blocked; protected by bus->reg_lock */ };
sound/hda/common/codec.c +2 -2
@@ -761,7 +761,7 @@ void snd_hda_codec_cleanup_for_unbind(struct hda_codec *codec)/* pm_runtime_put() is called in snd_hdac_device_exit() */ pm_runtime_get_noresume (hda_codec_dev (codec));pm_runtime_disable (hda_codec_dev (codec)); - codec->core.registered =0 ; + codec->core.registered = false; }snd_hda_codec_disconnect_pcms (codec);
@@ -815,7 +815,7 @@ void snd_hda_codec_register(struct hda_codec *codec)pm_runtime_enable (hda_codec_dev (codec));/* it was powered up in snd_hda_codec_new(), now all done */ snd_hda_power_down (codec); - codec->core.registered =1 ; + codec->core.registered = true; } }EXPORT_SYMBOL_GPL (snd_hda_codec_register);
Signature · 1 line
2.55.0
Quoted 8 lines · Takashi Iwai
Here is another series of small fixes for HD-audio, beginning with the fix for an issue reported by Sashiko, and extended the fixes to ASoC HD-audio and a potential improvement that were spotted during the development of the first patch.
Takashi
Quoted 7 lines
===
Takashi Iwai (3):
ALSA: hda: Stop unsol events and jack polling before codec unbind
ASoC: codecs: hda: Stop unsol events and jack polling before codec
removal
ALSA: hda: Make hdac_device.registered a plain bool
Thank you for fixing the ASoC's hda-driver too. The first two patches I've reviewed thoroughly, look good. The third is something I can learn from.