summaryrefslogtreecommitdiff
path: root/sound/hda
diff options
context:
space:
mode:
authorThomas Gleixner <tglx@linutronix.de>2021-11-25 01:40:01 +0300
committerTakashi Iwai <tiwai@suse.de>2021-11-29 19:42:31 +0300
commit6dd21ad81bf96478db3403b1bbe251c0612d0431 (patch)
treeced310850cdc2a442a73a6354443b09eabd63080 /sound/hda
parent7c72665c5667d4566e594ea362c50a6007b405fb (diff)
downloadlinux-6dd21ad81bf96478db3403b1bbe251c0612d0431.tar.xz
ALSA: hda: Make proper use of timecounter
HDA uses a timecounter to read a hardware clock running at 24 MHz. The conversion factor is set with a mult value of 125 and a shift value of 0, which is not converting the hardware clock to nanoseconds, it is converting to 1/3 nanoseconds because the conversion factor from 24Mhz to nanoseconds is 125/3. The usage sites divide the "nanoseconds" value returned by timecounter_read() by 3 to get a real nanoseconds value. There is a lengthy comment in azx_timecounter_init() explaining this choice. That comment makes blatantly wrong assumptions about how timecounters work and what can overflow. The comment says: * Applying the 1/3 factor as part of the multiplication * requires at least 20 bits for a decent precision, however * overflows occur after about 4 hours or less, not a option. timecounters operate on time deltas between two readouts of a clock and use the mult/shift pair to calculate a precise nanoseconds value: delta_nsec = (delta_clock * mult) >> shift; The fractional part is also taken into account and preserved to prevent accumulated rounding errors. For details see cyclecounter_cyc2ns(). The mult/shift pair has to be chosen so that the multiplication of the maximum expected delta value does not result in a 64bit overflow. As the counter wraps around on 32bit, the maximum observable delta between two reads is (1 << 32) - 1 which is about 178.9 seconds. That in turn means the maximum multiplication factor which fits into an u32 will not cause a 64bit overflow ever because it's guaranteed that: ((1 << 32) - 1) ^ 2 < (1 << 64) The resulting correct multiplication factor is 2796202667 and the shift value is 26, i.e. 26 bit precision. The overflow of the multiplication would happen exactly at a clock readout delta of 6597069765 which is way after the wrap around of the hardware clock at around 274.8 seconds which is off from the claimed 4 hours by more than an order of magnitude. If the counter ever wraps around the last read value then the calculation is off by the number of wrap arounds times 178.9 seconds because the overflow cannot be observed. Use clocks_calc_mult_shift(), which calculates the most accurate mult/shift pair based on the given clock frequency, and remove the bogus comment along with the divisions at the readout sites. Fixes: 5d890f591d15 ("ALSA: hda: support for wallclock timestamps") Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Pierre-Louis Bossart <pierre-louis.bossart@linux.intel.com> Link: https://lore.kernel.org/r/871r35kwji.ffs@tglx Signed-off-by: Takashi Iwai <tiwai@suse.de>
Diffstat (limited to 'sound/hda')
-rw-r--r--sound/hda/hdac_stream.c14
1 files changed, 4 insertions, 10 deletions
diff --git a/sound/hda/hdac_stream.c b/sound/hda/hdac_stream.c
index 9867555883c3..aa7955fdf68a 100644
--- a/sound/hda/hdac_stream.c
+++ b/sound/hda/hdac_stream.c
@@ -534,17 +534,11 @@ static void azx_timecounter_init(struct hdac_stream *azx_dev,
cc->mask = CLOCKSOURCE_MASK(32);
/*
- * Converting from 24 MHz to ns means applying a 125/3 factor.
- * To avoid any saturation issues in intermediate operations,
- * the 125 factor is applied first. The division is applied
- * last after reading the timecounter value.
- * Applying the 1/3 factor as part of the multiplication
- * requires at least 20 bits for a decent precision, however
- * overflows occur after about 4 hours or less, not a option.
+ * Calculate the optimal mult/shift values. The counter wraps
+ * around after ~178.9 seconds.
*/
-
- cc->mult = 125; /* saturation after 195 years */
- cc->shift = 0;
+ clocks_calc_mult_shift(&cc->mult, &cc->shift, 24000000,
+ NSEC_PER_SEC, 178);
nsec = 0; /* audio time is elapsed time since trigger */
timecounter_init(tc, cc, nsec);