At Fri, 02 May 2008 19:12:57 +0200, Zoilo Gomez wrote:
Takashi Iwai wrote:
At Fri, 02 May 2008 17:15:13 +0200, Zoilo Gomez wrote:
Takashi Iwai wrote:
At Wed, 30 Apr 2008 18:49:52 +0200, Zoilo Gomez wrote:
Takashi Iwai wrote:
At Tue, 29 Apr 2008 15:25:57 +0200, Zoilo Gomez wrote:
>>> This is consistent with the apparent introduction of this bug in kernel >>> 2.6.19 (includes alsa-driver version 1.0.12rc1): no such problem >>> occurred until 2.6.18, but all kernels since 2.6.19 do suffer from this >>> problem. The line of code above first shows up in linux-2.6.19. >>> >>> Unfortunately, since I do not have a datasheet for the VT1617A chip set, >>> I cannot verify the exact semantics, or suggest an improvement. >>> >>> Can anyone with a datasheet please suggest a proper patch to this line >>> of code? >>> >>> >>> >>> >> The register 0x5c is the VIA specific one. I have a VT1617 (without >> A) datasheet, and it suggests that the bit corresponds to the >> "headphone amplifier temperature sensing control". And setting this >> bit means to _disable_ the temperature sensing control. This sounds >> rather the correct to set. >> >> However, I don't know whether any difference exists betweeen VIA1617 >> and 1617A although the codec id of both are identical. >> >> Andrey, any comments about your patch? >> >> >> In anyway, it'd be helpful if we can know which ac97 registers work >> and wich not. Please take /proc/asound/card0/codec97#*/ac97#*-regs >> file in both working and non-working cases to compare. Especially, >> the registers 0x5a and 0x5c look interesting. >> >> >> >> > Code containing "snd_ac97_write_cache(a97, 0x5c, 0x20)": > 0x5a = 8300 > 0x5c = 0000 > > Code with "snd_ac97_write_cache(a97, 0x5c, 0x20)" commented out: > 0x5a = 8301 > 0x5c = 0020 > > Quite a surprise to me, I would have expected exactly the opposite .....!? > > > I would, too. Could you double-check, e.g. by adding a printk?
I modified the driver code to:
int patch_vt1617a(struct snd_ac97 * ac97) { int err = 0;
/* we choose to not fail out at this point, but we tell the caller when we return */ err = patch_build_controls(ac97, &snd_ac97_controls_vt1617a[0], ARRAY_SIZE(snd_ac97_controls_vt1617a)); /* bring analog power consumption to normal by turning off the * headphone amplifier, like WinXP driver for EPIA SP */printk("before: snd_ac97_read(ac97, 0x5c) = %04x\n", snd_ac97_read(ac97, 0x5c)); snd_ac97_write_cache(ac97, 0x5c, 0x20); printk("after: snd_ac97_read(ac97, 0x5c) = %04x\n", snd_ac97_read(ac97, 0x5c)); ac97->ext_id |= AC97_EI_SPDIF; /* force the detection of spdif */ ac97->rates[AC97_RATES_SPDIF] = SNDRV_PCM_RATE_44100 | SNDRV_PCM_RATE_48000; ac97->build_ops = &patch_vt1616_ops;
return err;}
Surprise surprise ... the printk output in dmesg reads:
before: snd_ac97_read(ac97, 0x5c) = 0020 after: snd_ac97_read(ac97, 0x5c) = 0000
If I write 0x00 (i.o. 0x20) into 0x5c, then the register content remains unchanged, and returns: 0x20. If I write 0x20 into 0x5c twice, then the register content is also 0x00, so no toggling.
So it seems that the bit-value is inverted: write "1" clears the bit, whereas writing "0" sets the bit.
Interesting.
And the code is trying to set the bit (with good intentions), but the result is in fact the opposite.
What do you think?
This sounds like a hardware bug to me...
Agreed.
How about the patch below then?
Takashi
diff --git a/pci/ac97/ac97_patch.c b/pci/ac97/ac97_patch.c index 726320b..e3e4dac 100644 --- a/pci/ac97/ac97_patch.c +++ b/pci/ac97/ac97_patch.c @@ -3448,6 +3448,7 @@ static const struct snd_kcontrol_new snd_ac97_controls_vt1617a[] = { int patch_vt1617a(struct snd_ac97 * ac97) { int err = 0;
int val;
/* we choose to not fail out at this point, but we tell the caller when we return */
@@ -3458,7 +3459,10 @@ int patch_vt1617a(struct snd_ac97 * ac97) /* bring analog power consumption to normal by turning off the * headphone amplifier, like WinXP driver for EPIA SP */
- snd_ac97_write_cache(ac97, 0x5c, 0x20);
- val = snd_ac97_read(ac97, 0x5c);
- if (!(val & 0x20))
snd_ac97_write_cache(ac97, 0x5c, 0x20);This has no effect, since set/clear is apparently reversed here;
So, the bit 0x20 isn't set as default when you didn't write to 0x5c?
Yes it is; if I do not write to 0x5c after reset, then the register value reads back as 0x20.
I think you mean:
- if (!(val & 0x20))
snd_ac97_write_cache(ac97, 0x5c, 0x00);No I meant as is. See below.
But why so complicated? Why not omit the test, and write unconditionally:
- snd_ac97_write_cache(ac97, 0x5c, 0x00);
The patch was there for any good reason. It was introduced to fix some bugs on certain machines. Removing it blindly breaks it again.
Or perhaps even leave bit 5 in the reset state, which is 0x20 hence OK.
Hm, they why the first test didn't pass?
OK, now I get your point: you are assuming that the chip set on my machine can be recognized because it has bit 0x20 set after reset, and in that case we leave the value as is. However if bit 0x20 is not set, then we assume that it is a different chip set with "proper 1/0-set/clear-behaviour", and we set the bit to 0x20. Correct?
Yes, exactly.
Sorry for the confusion on my side.
No. The situation is really complicated:
- the current code is confirmed to work on some machines as is
- we don't know exactly whether the bit flip oocurs on every hardware with VT1617A (maybe not)
- we don't know exactly whether the bit is set or cleared as default
Understood.
OK, I confirm that your patch works fine on my machine, which (for the record) is a Jetway J7F2 Mini-ITX motherboard.
Thanks for checking. I applied the patch (with a bit more comment) to ALSA tree now. It'll be on 2.6.26 tree, too.
I have also cross-checked (well ... "cross-grepped") with the viaudiocombo OSS-driver from the Arena web site; it seems that the contents of register 0x5c are left unaltered after reset (hence the value is 0x20). So that explains why the Arena driver does not suffer from this problem.
The workaround was relatively new and I guess it's specific to some EPIA board and possibly not included in other drivers...
thanks,
Takashi