[PATCH] intel8x0m: schedule timeout before warm reset check

[PATCH] intel8x0m: schedule timeout before warm reset check

[Paul Bolle]

At every resume a laptop I use prints this message (at KERN_ERR level):
    ALSA sound/pci/intel8x0m.c:904: AC'97 warm reset still in progress?
[0x2]

The thing to note here is that 0x2 corresponds to ICH_AC97COLD. Ie, what
seems to be happening is that the register involved indicated a warm
reset for some time (as the ICH_AC97WARM bit was set) but by the time
the warning is printed, and that same register is checked again, that
bit is already cleared and only the ICH_AC97COLD bit is still set.

It turns out a warm reset needs some time to settle, but it is currently
checked right away. The test therefore fails the first time it is done
and schedule_timeout_uninterruptible() will be called. Once we return
from that jiffies is already (far) past end_time on this laptop, so we
exit the loop, print a warning, and exit the function while the warm
reset actually succeeded.

One way to fix this is to call schedule_timeout_uninterruptible() at the
start of the loop. (This might punish the "finish cold reset" case, but
on this laptop such a cold reset apparently never happens, so I can't
say for sure.)

While we're at it drop the extra single tick from end_time, as it looks
rather silly.

Signed-off-by: Paul Bolle <pebolle@tiscali.nl>
---
The question I asked in the previous message ("snd_intel8x0m: error
message because return value is unused?") made little sense. I probably
got quite confused by my own efforts to debug this.

 sound/pci/intel8x0m.c |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/sound/pci/intel8x0m.c b/sound/pci/intel8x0m.c
index 13cec1e..9abff7b 100644
--- a/sound/pci/intel8x0m.c
+++ b/sound/pci/intel8x0m.c
@@ -894,11 +894,11 @@ static int snd_intel8x0m_ich_chip_init(struct intel8x0m *chip, int probing)
        /* finish cold or do warm reset */
        cnt |= (cnt & ICH_AC97COLD) == 0 ? ICH_AC97COLD : ICH_AC97WARM;
        iputdword(chip, ICHREG(GLOB_CNT), cnt);
-       end_time = (jiffies + (HZ / 4)) + 1;
+       end_time = jiffies + HZ / 4;
        do {
+               schedule_timeout_uninterruptible(1);
                if ((igetdword(chip, ICHREG(GLOB_CNT)) & ICH_AC97WARM) == 0)
                        goto __ok;
-               schedule_timeout_uninterruptible(1);
        } while (time_after_eq(end_time, jiffies));
        snd_printk(KERN_ERR "AC'97 warm reset still in progress? [0x%x]\n",
                   igetdword(chip, ICHREG(GLOB_CNT)));
-- 
1.7.4.1

Re: [PATCH] intel8x0m: schedule timeout before warm reset check

[Takashi Iwai]

At Sun, 27 Feb 2011 01:23:58 +0100,
Paul Bolle wrote:
> 
> At every resume a laptop I use prints this message (at KERN_ERR level):
>     ALSA sound/pci/intel8x0m.c:904: AC'97 warm reset still in progress?
> [0x2]
> 
> The thing to note here is that 0x2 corresponds to ICH_AC97COLD. Ie, what
> seems to be happening is that the register involved indicated a warm
> reset for some time (as the ICH_AC97WARM bit was set) but by the time
> the warning is printed, and that same register is checked again, that
> bit is already cleared and only the ICH_AC97COLD bit is still set.
> 
> It turns out a warm reset needs some time to settle, but it is currently
> checked right away. The test therefore fails the first time it is done
> and schedule_timeout_uninterruptible() will be called. Once we return
> from that jiffies is already (far) past end_time on this laptop, so we
> exit the loop, print a warning, and exit the function while the warm
> reset actually succeeded.
> 
> One way to fix this is to call schedule_timeout_uninterruptible() at the
> start of the loop. (This might punish the "finish cold reset" case, but
> on this laptop such a cold reset apparently never happens, so I can't
> say for sure.)
> 
> While we're at it drop the extra single tick from end_time, as it looks
> rather silly.
> 
> Signed-off-by: Paul Bolle <pebolle@tiscali.nl>
> ---
> The question I asked in the previous message ("snd_intel8x0m: error
> message because return value is unused?") made little sense. I probably
> got quite confused by my own efforts to debug this.

It'd be better to put a more proper delay time than tick,
e.g. usleep_range() or whatever.  schedule_timeout(1) can be a
different length depending on HZ.

Could you figure out?


thanks,

Takashi

> 
>  sound/pci/intel8x0m.c |    4 ++--
>  1 files changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/sound/pci/intel8x0m.c b/sound/pci/intel8x0m.c
> index 13cec1e..9abff7b 100644
> --- a/sound/pci/intel8x0m.c
> +++ b/sound/pci/intel8x0m.c
> @@ -894,11 +894,11 @@ static int snd_intel8x0m_ich_chip_init(struct intel8x0m *chip, int probing)
>         /* finish cold or do warm reset */
>         cnt |= (cnt & ICH_AC97COLD) == 0 ? ICH_AC97COLD : ICH_AC97WARM;
>         iputdword(chip, ICHREG(GLOB_CNT), cnt);
> -       end_time = (jiffies + (HZ / 4)) + 1;
> +       end_time = jiffies + HZ / 4;
>         do {
> +               schedule_timeout_uninterruptible(1);
>                 if ((igetdword(chip, ICHREG(GLOB_CNT)) & ICH_AC97WARM) == 0)
>                         goto __ok;
> -               schedule_timeout_uninterruptible(1);
>         } while (time_after_eq(end_time, jiffies));
>         snd_printk(KERN_ERR "AC'97 warm reset still in progress? [0x%x]\n",
>                    igetdword(chip, ICHREG(GLOB_CNT)));
> -- 
> 1.7.4.1
> 
> 
> 
>