Reproducible deadlock w. alsa/maestro3 when sleeping (ACPI,) 2.6.0-test1

DISCUSSION P. Christeas 3 participants 6 messages Jul 14, 2003 – Jul 17, 2003

Share & tools

P. Christeas

author

2003-07-14 16:00 UTC #0 lore raw

I have been experiencing some fully reproducible deadlock when waking from sleep, using artsd over ALSA.
The scenario is:
I use ALSA, with the maestro3 device and everything else compiled as modules.

Quoted 1 line

From userspace, I launch artsd, which uses its native ALSA support to connect

to /dev/pcmXXXXX .
I only have a custom script, which sleeps the machine by a 'echo 1> /proc/acpi/sleep' . It does NOT stop alsa .
When the machine weaks, it has a deadlock. The suspended 'artsd' process waits for the /dev/pcmXXXX to become available (presumably; I can see 'D' on the 'ps' line), while the maestro3 module waits for the 'artsd' process to free /dev/pcmXXX .
I have tried to kill (w. 15, 9 etc.) the artsd process, rmmod the 'snd_maestro3' module (w. '-f' or '-l' options to rmmod). Nothing would happen. In the worst case (after the rmmod), the whole module mechanism would lock as well ('lsmod' would hang ).
I could not find any relevant message for this case. However, on a subsequent sleep, the system refuses to sleep because it cannot suspend 'artsd'.

ps. My scripts (including module-init-tools) are definitely messy. However, since 'artsd' or any ALSA client works in userspace and can be launched at any time, I believe this problem leads to serious trouble for the system.

Takashi Iwai

2003-07-16 11:32 UTC #1 lore raw

At Mon, 14 Jul 2003 19:17:22 +0300,

Quoted 9 lines · P. Christeas

I have been experiencing some fully reproducible deadlock when waking from sleep, using artsd over ALSA.
The scenario is:
I use ALSA, with the maestro3 device and everything else compiled as modules. From userspace, I launch artsd, which uses its native ALSA support to connect to /dev/pcmXXXXX .
I only have a custom script, which sleeps the machine by a 'echo 1> /proc/acpi/sleep' . It does NOT stop alsa .

could you check whether m3_suspend() and m3_resume() in sound/pci/maestro3.c are really called?

Takashi

P. Christeas

author

2003-07-16 14:01 UTC #2 lore raw

Quoted 12 lines · Takashi Iwai

At Mon, 14 Jul 2003 19:17:22 +0300,
Quoted 8 lines · P. Christeas
I have been experiencing some fully reproducible deadlock when waking from sleep, using artsd over ALSA.
The scenario is:
I use ALSA, with the maestro3 device and everything else compiled as modules. From userspace, I launch artsd, which uses its native ALSA support to connect to /dev/pcmXXXXX .
I only have a custom script, which sleeps the machine by a 'echo 1> /proc/acpi/sleep' . It does NOT stop alsa .

Quoted 5 lines

could you check whether m3_suspend() and m3_resume() in sound/pci/maestro3.c are really called?
Takashi

OK, I did that. I put two messages in both functions of the maestro3 driver. I suspended/resumed the machine. Both functions had been called.

This time, I did NOT have 'artsd' (i.e. the client) loaded. What happened was that the module was properly restored and I could load (and use) artsd even after the resume.
That brings me to the first assumption/question I have made: is there something wrong if we suspend two parts (one module and a userspace process), while they inter-communicate through the /dev/* interface?

static void m3_suspend(m3_t *chip)

{
	snd_card_t *card = chip->card;
	int i, index;

	snd_printk("m3 suspend");
	if (chip->suspend_mem == NULL)
		return;
	if (card->power_state == SNDRV_CTL_POWER_D3hot)
		return;
	...

static void m3_resume(m3_t *chip)

{
	snd_card_t *card = chip->card;
	int i, index;

	snd_printk("m3 resume");

	if (chip->suspend_mem == NULL)
		return;
	...

P. Christeas

author

2003-07-16 14:47 UTC #3 lore raw

Quoted 22 lines

Quoted 12 lines · Takashi Iwai
At Mon, 14 Jul 2003 19:17:22 +0300,
Quoted 8 lines · P. Christeas
I have been experiencing some fully reproducible deadlock when waking from sleep, using artsd over ALSA.
The scenario is:
I use ALSA, with the maestro3 device and everything else compiled as modules. From userspace, I launch artsd, which uses its native ALSA support to connect to /dev/pcmXXXXX .
I only have a custom script, which sleeps the machine by a 'echo 1> /proc/acpi/sleep' . It does NOT stop alsa .
Quoted 5 lines
could you check whether m3_suspend() and m3_resume() in sound/pci/maestro3.c are really called?
Takashi
OK, I did that. I put two messages in both functions of the maestro3 driver. I suspended/resumed the machine. Both functions had been called.

Quoted 8 lines

This time, I did NOT have 'artsd' (i.e. the client) loaded. What happened was that the module was properly restored and I could load (and use) artsd even after the resume.
That brings me to the first assumption/question I have made: is there something wrong if we suspend two parts (one module and a userspace process), while they inter-communicate through the /dev/* interface?

In a second test, I have also added messages at the end of these functions. They surely don't exit early indeed.

Has anybody else managed to reproduce the bug?
Does it happen with other drivers (say, PCI cards w. pcm interface)?

procedure:
(read the last step first; it is a warning)
1. load the sound module, like 'modprobe snd-maestro3'
2. load the client ("artsd" should be the one, others may eventually release
the descriptor. If you want, you could give them a try as well.)
3. Suspend, S1, NOT with an all-capable script. The script you use must not
try to bring down ALSA.
4. Resume.
5. Check the state of the "artsd" (or equivalent) process.

W. Note that if the process is waiting for I/O, you can do nothing but
reboot.

James Courtier-Dutton

2003-07-16 15:18 UTC #4 lore raw

Quoted 48 lines · P. Christeas

Quoted 24 lines
Quoted 13 lines · Takashi Iwai
At Mon, 14 Jul 2003 19:17:22 +0300,
Quoted 8 lines · P. Christeas
I have been experiencing some fully reproducible deadlock when waking from sleep, using artsd over ALSA.
The scenario is:
I use ALSA, with the maestro3 device and everything else compiled as modules. From userspace, I launch artsd, which uses its native ALSA support to connect to /dev/pcmXXXXX .
I only have a custom script, which sleeps the machine by a 'echo 1> /proc/acpi/sleep' . It does NOT stop alsa .
Quoted 5 lines
could you check whether m3_suspend() and m3_resume() in sound/pci/maestro3.c are really called?
Takashi
OK, I did that. I put two messages in both functions of the maestro3 driver. I suspended/resumed the machine. Both functions had been called.
Quoted 8 lines
This time, I did NOT have 'artsd' (i.e. the client) loaded. What happened was that the module was properly restored and I could load (and use) artsd even after the resume.
That brings me to the first assumption/question I have made: is there something wrong if we suspend two parts (one module and a userspace process), while they inter-communicate through the /dev/* interface?
In a second test, I have also added messages at the end of these functions. They surely don't exit early indeed.
Has anybody else managed to reproduce the bug?
Does it happen with other drivers (say, PCI cards w. pcm interface)?
procedure:
(read the last step first; it is a warning)
1. load the sound module, like 'modprobe snd-maestro3'
2. load the client ("artsd" should be the one, others may eventually release
the descriptor. If you want, you could give them a try as well.)
3. Suspend, S1, NOT with an all-capable script. The script you use must not
try to bring down ALSA.
4. Resume.
5. Check the state of the "artsd" (or equivalent) process.

Quoted 5 lines

W. Note that if the process is waiting for I/O, you can do nothing but
reboot.

I get a similar sort of problem, but that is with an USB sound card, and one unplugs the usb cable.
The process (in my case a media appication) is waiting for a return from a call to alsa-lib, but the return from the call never happens.
Maybe suspending a PCI sound card is similar to unplugging a USB sound card. In both cases, the device does not respond anymore, but alsa-lib fails to return an error to the application, but instead never returns.

Cheers
James

-------------------------------------------------------
This SF.net email is sponsored by: VM Ware
With VMware you can run multiple operating systems on a single machine. WITHOUT REBOOTING! Mix Linux / Windows / Novell virtual machines at the same time. Free trial click here: http://www.vmware.com/wl/offer/345/0

P. Christeas

author

2003-07-17 16:14 UTC #5 lore raw

Quoted 34 lines

Quoted 23 lines
Quoted 6 lines
This time, I did NOT have 'artsd' (i.e. the client) loaded. What happened was that the module was properly restored and I could load (and use) artsd even after the resume.
That brings me to the first assumption/question I have made: is there something wrong if we suspend two parts (one module and a userspace process), while they inter-communicate through the /dev/* interface?
In a second test, I have also added messages at the end of these functions. They surely don't exit early indeed.
Has anybody else managed to reproduce the bug?
Does it happen with other drivers (say, PCI cards w. pcm interface)?
procedure:
(read the last step first; it is a warning)
1. load the sound module, like 'modprobe snd-maestro3'
2. load the client ("artsd" should be the one, others may eventually
release the descriptor. If you want, you could give them a try as well.)
3. Suspend, S1, NOT with an all-capable script. The script you use must not try to bring down ALSA.
4. Resume.
5. Check the state of the "artsd" (or equivalent) process.
Quoted 2 lines
W. Note that if the process is waiting for I/O, you can do nothing but
reboot.
I get a similar sort of problem, but that is with an USB sound card, and one unplugs the usb cable.
The process (in my case a media appication) is waiting for a return from a call to alsa-lib, but the return from the call never happens. Maybe suspending a PCI sound card is similar to unplugging a USB sound card. In both cases, the device does not respond anymore, but alsa-lib fails to return an error to the application, but instead never returns.

Quoted 2 lines

Cheers
James

If you verify that you are unable to 'kill -9' the client application, then the problem becomes just little more serious. IMHO no such actions should cause the kernel to have 'blocked' resources.

One important observation I have is that 'esd' doesn't produce the bug. Esd uses '/dev/sound/dsp' rather than '/dev/snd/pcmXXX' (oss-emu. rather than ALSA). I don't even know what level of 'feedback' (in terms of function returns) OSS provides, in comparison to the ALSA system.