Skip to contentSkip to conversation

intel8x0: Improve performance in virtual environment

PATCH Jaroslav Kysela 2 participants 4 messages Oct 20, 2011 – Oct 24, 2011 3 messages not in the archive

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

#1 lore raw
Quoted 7 lines · Konstantin Ozerkov

This patch intended to improve performance in virtualized environments like Parallels Desktop or VirtualBox/QEMU (virtual ICH/AC97 audio).

I/O access is very time-expensive operation in virtual world: VCPU can be rescheduled and in the worst case we got more than 10ms delay on each I/O access.

Quoted 8 lines

In the original code normal loop exit rule
(old_civ == current_civ && old_picb == current_picb) was never satisfied, because old_picb was never the same as current_picb due delay inspired by reading current_civ. As a result loop ended by timeout and we got 10x more I/O operations.

To prevent glitch on buffer swapping the rule (old_civ == current_civ) is enough.

I don't think so. This code was added because we got some strange situations (values) with the real intel8x0 hardware. A better way might be to identify the virtualisation (maybe using special PCI IDs) and skip these extra checks in the driver for these devices.

NAK from me.

Jaroslav

Signature · 3 lines
Jaroslav Kysela <[email protected]>
Linux Kernel Sound Maintainer
ALSA Project; Red Hat, Inc.
#2 lore raw
In reply to a message not in the archive

At Thu, 20 Oct 2011 19:02:50 +0400,

Quoted 8 lines · Konstantin Ozerkov

This patch intended to improve performance in virtualized environments like Parallels Desktop or VirtualBox/QEMU (virtual ICH/AC97 audio).

I/O access is very time-expensive operation in virtual world: VCPU can be rescheduled and in the worst case we got more than 10ms delay on each I/O access.

Quoted 5 lines

In the original code normal loop exit rule
(old_civ == current_civ && old_picb == current_picb) was never satisfied, because old_picb was never the same as current_picb due delay inspired by reading current_civ. As a result loop ended by timeout and we got 10x more I/O operations.

You mean in the virtual environment, right? On the real machines, this condition can be satisfied normally because the read is much faster than the sample rate. This check is needed to avoid the broken behavior around the boundary.

Quoted 2 lines

To prevent glitch on buffer swapping the rule (old_civ == current_civ) is enough.

Not for real (thus less logical) machines, unfortunately. Both checks were needed in the real tests.

I don't mind to implement some bypass if the driver knows that it's running in a VM. But, removing blindly would cause definitely a regression.

thanks,

Takashi

Quoted 32 lines

Experimental data from Prallels Desktop 7, RHEL6 guest (I/O ops per second):

Original code:

In Port    Counter         Callback
   f014      41550         fffff00000179d00 ac97_bm_read_civ+0x000
   f018      41387         fffff0000017a580 ac97_bm_read_picb+0x000

With patch:

In Port    Counter         Callback
   f014       4090         fffff00000179d00 ac97_bm_read_civ+0x000
   f018       1964         fffff0000017a580 ac97_bm_read_picb+0x000
pci/intel8x0.c +1 -2
index 6a5b387..bad4a65 100644
@@ -1065,8 +1065,7 @@ static snd_pcm_uframes_t snd_intel8x0_pcm_pointer(struct snd_pcm_substream *subs
 			udelay(10);
 			continue;
 		}
-		if (civ == igetbyte(chip, ichdev->reg_offset + ICH_REG_OFF_CIV) &&
-		    ptr1 == igetword(chip, ichdev->reg_offset + ichdev->roff_picb))
+		if (civ == igetbyte(chip, ichdev->reg_offset + ICH_REG_OFF_CIV))
 			break;
 	} while (timeout--);
 	ptr = ichdev->last_pos;
Signature · 1 line
1.7.7

A message that is not in the archive (sent off-list or not yet received): [email protected]

#4 lore raw

At Thu, 20 Oct 2011 21:32:38 +0400,

Quoted 7 lines · Konstantin Ozerkov

Thank you for fast response and comments.

I will rewrite this patch according you suggestions:
* add VM environment detection logic (PD only, I don't have clear vision how to that for VirtualBox/QEMU) and module parameter (for manual control)
* make separated VM-optimized variant of snd_intel8x0_pcm_pointer()

Sounds reasonable.

Quoted 22 lines

Initially this patch is result of research and profiling audio emulation quality in Parallels Desktop.

Quoted 19 lines · Takashi Iwai

At Thu, 20 Oct 2011 19:02:50 +0400,

Quoted 7 lines · Konstantin Ozerkov

This patch intended to improve performance in virtualized environments like Parallels Desktop or VirtualBox/QEMU (virtual ICH/AC97 audio).

I/O access is very time-expensive operation in virtual world: VCPU can be rescheduled and in the worst case we got more than 10ms delay on each I/O access.

Quoted 5 lines

In the original code normal loop exit rule
(old_civ == current_civ&& old_picb == current_picb) was never satisfied, because old_picb was never the same as current_picb due delay inspired by reading current_civ. As a result loop ended by timeout and we got 10x more I/O operations.

You mean in the virtual environment, right? On the real machines, this condition can be satisfied normally because the read is much faster than the sample rate. This check is needed to avoid the broken behavior around the boundary.

Quoted 6 lines
Quoted 4 lines
Quoted 2 lines

To prevent glitch on buffer swapping the rule (old_civ == current_civ) is enough.

Not for real (thus less logical) machines, unfortunately. Both checks were needed in the real tests.

Can you say which hardware glitches this way and and which conditions? My simplified rule works well in many Windows AC97 drivers which I debug/analyze.

Some hardware showed the unstable CIV and PICB values when it goes crossing the boundary. Both registers should be updated simultaneously, but in reality, you may hit at a bad timing where only one of them is updated. And the order of updates of CIV/PICB isn't defined. Actually, in the earlier version, we tested only CIV, but it turned out that it's didn't suffice. So, we had to check both values.

thanks,

Takashi

A message that is not in the archive (sent off-list or not yet received): [email protected]

#6 lore raw

Re: Is locking problem in snd_intel8x0_pcm_pointer() ?

At Mon, 24 Oct 2011 15:48:44 +0400,

Quoted 5 lines · Konstantin Ozerkov

Seems that locking is broken: snd_intel8x0_pcm_pointer() uses spin_lock()/spin_unlock() semantic for lock which is shared in ISR ( snd_intel8x0_update). As result we got possible deadlock inside ISR on UP system. And needed special kludge inside snd_intel8x0_pcm_pointer() to prevent moving backward (really this is race with snd_intel8x0_update).

Hm, could you elaborate how can it a problem? snd_intel8x0_update() releases the lock before calling snd_pcm_period_update() that calls the pointer callback.

Takashi