On 03/31/11 15:50, Kelly Anderson wrote:

Just to let you know. I'm using a gtx460 also. I haven't had a chance to reboot into the newly patched kernel yet. It is interesting that we both have gtx460's. I have a machine with a gtx430 that doesn't seem to exhibit the problem.

cat /etc/modprobe.d/snd_hda_intel.conf

options snd-hda-intel probe_mask=-1,0xa

OK, I debugged the problem and came up with a solution. The bottom line is that it's a problem with a signed to unsigned compare. First the fix:

--- ./sound/core/pcm_lib.c.orig 2011-03-27 12:37:20.000000000 -0600 +++ ./sound/core/pcm_lib.c 2011-03-31 19:01:35.392739127 -0600 @@ -379,17 +379,17 @@ static int snd_pcm_update_hw_ptr0(struct * Without regular period interrupts, we have to check * the elapsed time to detect xruns. */ - jdelta = jiffies - runtime->hw_ptr_jiffies; - if (jdelta < runtime->hw_ptr_buffer_jiffies / 2) + jdelta = (snd_pcm_sframes_t)(jiffies - runtime->hw_ptr_jiffies); + if (jdelta < (snd_pcm_sframes_t)runtime->hw_ptr_buffer_jiffies / 2) goto no_delta_check; hdelta = jdelta - delta * HZ / runtime->rate; - while (hdelta > runtime->hw_ptr_buffer_jiffies / 2 + 1) { + while (hdelta > (snd_pcm_sframes_t)runtime->hw_ptr_buffer_jiffies / 2 + 1) { delta += runtime->buffer_size; hw_base += runtime->buffer_size; if (hw_base >= runtime->boundary) hw_base = 0; new_hw_ptr = hw_base + pos; - hdelta -= runtime->hw_ptr_buffer_jiffies; + hdelta -= (snd_pcm_sframes_t)runtime->hw_ptr_buffer_jiffies; } goto no_delta_check; }

And here's the patch I used to debug the problem:

--- ./sound/core/pcm_lib.c.orig 2011-03-27 12:37:20.000000000 -0600 +++ ./sound/core/pcm_lib.c 2011-03-31 18:44:23.701411841 -0600 @@ -379,18 +379,31 @@ static int snd_pcm_update_hw_ptr0(struct * Without regular period interrupts, we have to check * the elapsed time to detect xruns. */ + int loopCount = 0; + BUG_ON(runtime->hw_ptr_jiffies < 0); + BUG_ON(runtime->hw_ptr_jiffies == 0); + BUG_ON(runtime->hw_ptr_jiffies > jiffies); jdelta = jiffies - runtime->hw_ptr_jiffies; if (jdelta < runtime->hw_ptr_buffer_jiffies / 2) goto no_delta_check; hdelta = jdelta - delta * HZ / runtime->rate; - while (hdelta > runtime->hw_ptr_buffer_jiffies / 2 + 1) { + printk(KERN_INFO "Before loop: jdelta=%ld, hdelta=%ld runtime->hw_ptr_buffer_jiffies=%lu\n" + , jdelta , hdelta , runtime->hw_ptr_buffer_jiffies); + while (hdelta > runtime->hw_ptr_buffer_jiffies / 2 + 1) { delta += runtime->buffer_size; hw_base += runtime->buffer_size; if (hw_base >= runtime->boundary) hw_base = 0; new_hw_ptr = hw_base + pos; - hdelta -= runtime->hw_ptr_buffer_jiffies; + hdelta -= runtime->hw_ptr_buffer_jiffies; + if (++loopCount >= 1024*1024*4) + { + printk(KERN_INFO "stopped the madness at 4 thousand\n"); + break; + } } + printk(KERN_INFO "After loop: jdelta=%ld, hdelta=%ld runtime->hw_ptr_buffer_jiffies=%lu\n" + , jdelta , hdelta , runtime->hw_ptr_buffer_jiffies); goto no_delta_check; }

The kernel output with the testing patch applied:

Mar 31 18:37:02 speedy kernel: Before loop: jdelta=352, hdelta=352 runtime->hw_ptr_buffer_jiffies=371 Mar 31 18:37:02 speedy kernel: stopped the madness at 4 million Mar 31 18:37:02 speedy kernel: After loop: jdelta=352, hdelta=-1556086432 runtime->hw_ptr_buffer_jiffies=371

Kelly Anderson wrote:

OK, I debugged the problem and came up with a solution. The bottom line is that it's a problem with a signed to unsigned compare.

Aaargh, I didn't suspect this.

First the fix:

--- ./sound/core/pcm_lib.c.orig 2011-03-27 12:37:20.000000000 -0600 +++ ./sound/core/pcm_lib.c 2011-03-31 19:01:35.392739127 -0600 @@ -379,17 +379,17 @@ static int snd_pcm_update_hw_ptr0(struct * Without regular period interrupts, we have to check * the elapsed time to detect xruns. */

•    jdelta = jiffies - runtime->hw_ptr_jiffies;
  

•    if (jdelta < runtime->hw_ptr_buffer_jiffies / 2)
  

•    jdelta = (snd_pcm_sframes_t)(jiffies - runtime->hw_ptr_jiffies);
  

•    if (jdelta < (snd_pcm_sframes_t)runtime->hw_ptr_buffer_jiffies / 2)
          goto no_delta_check;
      hdelta = jdelta - delta * HZ / runtime->rate;
  

•    while (hdelta > runtime->hw_ptr_buffer_jiffies / 2 + 1) {
  

•    while (hdelta > (snd_pcm_sframes_t)runtime->hw_ptr_buffer_jiffies / 2 + 1) {
          delta += runtime->buffer_size;
          hw_base += runtime->buffer_size;
          if (hw_base >= runtime->boundary)
              hw_base = 0;
          new_hw_ptr = hw_base + pos;
  

•        hdelta -= runtime->hw_ptr_buffer_jiffies;
  

•        hdelta -= (snd_pcm_sframes_t)runtime->hw_ptr_buffer_jiffies;
      }
      goto no_delta_check;
  }
  

Most of these casts have no effect; only the loop condition is important. Please test the patch below.

Kelly, to apply this patch I need a Signed-off-by tag from you (see Documentation/SubmittingPatches).

--8<---------------------------------------------------------------->8-- ALSA: pcm: fix infinite loop in snd_pcm_update_hw_ptr0()

When period interrupts are disabled, snd_pcm_update_hw_ptr0() compares the current time against the time estimated for the current hardware pointer to detect xruns. The somewhat fuzzy threshold in the while loop makes it possible that hdelta becomes negative; the comparison being done with unsigned types then makes the loop go through the entire 2^63 negative range, and, depending on the value, never reaching an unsigned value that is small enough to stop the loop. Doing this with interrupts disabled results in the machine locking up.

To prevent this, ensure that the loop condition uses signed types for both operands so that the comparison is correctly done.

Many thanks to Kelly Anderson for debugging this.

Reported-by: Nix nix@esperi.org.uk Reported-by: "Christopher K." c.krooss@googlemail.com Reported-by: Kelly Anderson kelly@silka.with-linux.com [cl: remove unneeded casts; use a temp variable] Signed-off-by: Clemens Ladisch clemens@ladisch.de

--- a/sound/core/pcm_lib.c +++ b/sound/core/pcm_lib.c @@ -375,6 +375,7 @@ static int snd_pcm_update_hw_ptr0(struct snd_pcm_substream *substream, }

if (runtime->no_period_wakeup) { + snd_pcm_sframes_t xrun_threshold; /* * Without regular period interrupts, we have to check * the elapsed time to detect xruns. @@ -383,7 +384,8 @@ static int snd_pcm_update_hw_ptr0(struct snd_pcm_substream *substream, if (jdelta < runtime->hw_ptr_buffer_jiffies / 2) goto no_delta_check; hdelta = jdelta - delta * HZ / runtime->rate; + xrun_threshold = runtime->hw_ptr_buffer_jiffies / 2 + 1; - while (hdelta > runtime->hw_ptr_buffer_jiffies / 2 + 1) { + while (hdelta > xrun_threshold) { delta += runtime->buffer_size; hw_base += runtime->buffer_size; if (hw_base >= runtime->boundary)