From mboxrd@z Thu Jan 1 00:00:00 1970 From: bugtrack@alsa-project.org Subject: [ALSA - driver 0000951]: Fix bugs with early IRQs mishandling, missing silence fill, and wrong timing calculations Date: Wed, 2 Mar 2005 09:22:08 +0100 Message-ID: <44b367cff09421cff02416c95128cd60@bugtrack.alsa-project.org> Mime-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: 7bit Received: from bugtrack.alsa-project.org (gate.perex.cz [82.113.61.162]) by alsa.alsa-project.org (ALSA's E-mail Delivery System) with ESMTP id 681981E9 for ; Wed, 2 Mar 2005 09:22:08 +0100 (MET) Sender: alsa-devel-admin@lists.sourceforge.net Errors-To: alsa-devel-admin@lists.sourceforge.net List-Unsubscribe: , List-Post: List-Help: List-Subscribe: , List-Archive: To: alsa-devel@alsa-project.org List-Id: alsa-devel@alsa-project.org A NOTE has been added to this issue. ====================================================================== ====================================================================== Reported By: charles_levert Assigned To: ====================================================================== Project: ALSA - driver Issue ID: 951 Category: CORE - pcm Reproducibility: always Severity: major Priority: normal Status: new Distribution: Kernel Version: 2.4.22 ====================================================================== Date Submitted: 02-28-2005 21:26 CET Last Modified: 03-02-2005 09:22 CET ====================================================================== Summary: Fix bugs with early IRQs mishandling, missing silence fill, and wrong timing calculations Description: A patch is forthcoming. Here is the text from the patch preamble. This patch is against alsa-driver-1.0.8. It covers: alsa-kernel/core/pcm_lib.c alsa-kernel/include/pcm.h alsa-kernel/core/pcm_native.c alsa-kernel/core/oss/pcm_oss.c The main changes are in the first listed file. An original bug description was filed under "[ALSA - driver] PCI - cmipci": since the problems were first observed on this hardware. The patch fixes the following problems: -- Properly handle early IRQs in snd_pcm_update_hw_ptr_interrupt() in pcm_lib.c. This does occur, at least with CMI8738MC6 hardware. This bug had a major impact that caused the DMAs to run for too long and irritating noise to be heard at the end of playback. -- Properly compute estimated time of next interrupt in snd_pcm_update_hw_ptr_interrupt() in pcm_lib.c; this is no longer done with a "... % runtime->period_size" kludge (which didn't even attempt to round to nearest instead of merely truncating). -- Test whether hw_avail is big enough to handle the next runtime->period_size in snd_pcm_update_hw_ptr_post() in pcm_lib.c, otherwise call snd_pcm_tick_prepare() to tentatively set up a timer if needed (to be able to take earlier action in that case). -- Centralize all folding of values with regard to runtime->boundary in the new snd_pcm_next_buf() function in pcm_lib.c. All relevant values are now systematically in coherent synchronization without resorting to the ugly and error-prone "value %= runtime->boundary" kludge all over the place; unobvious wrapping conditions testing is no longer needed in these places. A safe zone worth one runtime->buffer_size is now kept on the left side of the operating interval, just like it was the case for the right side; this is necessary for proper simplified operation. -- Direct manipulations of values with respect to runtime->boundary needed to be brought in line with new proper practices (and thus simplified) in pcm.h, pcm_lib.c, pcm_native.c, and pcm_oss.c. -- Isolate the second part of snd_pcm_playback_silence() in the new snd_pcm_playback_silence_fill() function in pcm_lib.c. This makes the code more readable since the task this performs should not be confused with what is done in the remaining first part of snd_pcm_playback_silence(). The isolated code is unchanged. -- Rewrite the remaining first part of snd_pcm_playback_silence() to properly handle all situations, including computing minimum silence insertion even when "runtime->silence_size == 0". This is needed for proper operation of the aplay program, for one, and is much more forgiving on the user; insertion of silence samples in user space was always useless and a wrong solution anyway (but user space specification of kernel space silence insertion is useful and still very much supported). The code is now much more straightforward to read. Updates of runtime->control->appl_ptr (due to the user writing samples) and updates to runtime->status->hw_ptr (due to the audio hardware consuming samples) are now _both_ properly accounted for in all cases; the new code makes this much more obvious. -- All calls to snd_pcm_playback_silence() in pcm_lib.c and pcm_native.c must no longer be under the "runtime->silence_size > 0" condition. -- Properly perform rounded-up integer division in snd_pcm_system_tick_set() in pcm_lib.c. This only had a minor impact. -- Properly compute distance to next interrupt in snd_pcm_tick_prepare() in pcm_lib.c; again, avoid using any "... % runtime->period_size" kludge. -- The usage of runtime->rate in snd_pcm_tick_prepare() in pcm_lib.c was incorrect, regarless of other changes. The impact was that the timer was systematically set to 1 single tick when it needed to be set to typically hundreds of ticks. -- The usage of runtime->sleep_min in snd_pcm_tick_prepare() in pcm_lib.c was incorrect, regarless of other changes. The impact was that the number of ticks for the timer was not properly adjusted up to what sleep_min is documented to mean, when needed. ====================================================================== Relationships ID Summary ---------------------------------------------------------------------- related to 0000946 Early playback DMA IRQ with CMI8738MC6 ... ====================================================================== ---------------------------------------------------------------------- charles_levert - 03-01-05 11:15 ---------------------------------------------------------------------- The snd_pcm_tick_prepare() change could be on its own. That's one chunk with 2 old lines replaced by 1 new line. The runtime->rate and runtime->sleep_min in snd_pcm_tick_prepare() could be together on their own. That's one chunk with 3 old lines replaced by 3 new lines. Everything else, however, forms a coherent ensemble. For example, snd_pcm_playback_silence() now relies on snd_pcm_update_hw_ptr_interrupt() having already updated runtime->hw_ptr_interrupt and done the snd_pcm_next_buf() type of update. Splitting this in a fashion that would leave the code in a still-coherent state in-between each individual patch would mean writing intermediary code that does not exist at this point. It is far from obvious that what would so be made to appear as intermediary points would bring any more insight into the single main issue. It could present things in a worst way, actually. I've been burned doing this type of work for nothing in the past. Please provide me with a preliminary evaluation of the patch and some rough directions about what you would like to appear as intermediary points (and in which order of individual patch application), if any is justified, before I go any further. The patch is actually not that big. I just was very verbose in its description. ---------------------------------------------------------------------- Clemens Ladisch - 03-02-05 09:22 ---------------------------------------------------------------------- As far as I can see there are several changes to fix bugs, and other changes that are enhancements not intended to change the outwards behaviour of ALSA. Please make one patch for each bug. (Each _logical_ change should be one patch, see section 1.3 in SubmittingPatches.) There are several problems with this version of the patch: * snd_pcm_update_hw_ptr_interrupt() doesn't compute the time of the next interrupt; runtime->hw_ptr_interrupt is the beginning of the last (current) period. It is not sufficient do add one period_size because some PCM interrupts might have been lost, therefore the new value must be derived from the actual hardware pointer. (The same applies to the dist calculation in snd_pcm_tick_prepare().) * Xrun debugging was dropped. Why? * The documented behaviour for frame pointers is to wrap around to zero when the boundary is reached. Safe zones can only be implemented if the pointer values exported to user space and OSS emulation still show backwards-compatible behaviour. Issue History Date Modified Username Field Change ====================================================================== 02-28-05 21:26 charles_levert New Issue 02-28-05 21:26 charles_levert Kernel Version => 2.4.22 02-28-05 21:29 charles_levert File Added: early-irqs.patch 03-01-05 09:36 perex Relationship added related to 0000946 03-01-05 09:40 perex Note Added: 0003755 03-01-05 11:15 charles_levert Note Added: 0003757 03-02-05 09:22 Clemens LadischNote Added: 0003761 ====================================================================== ------------------------------------------------------- SF email is sponsored by - The IT Product Guide Read honest & candid reviews on hundreds of IT Products from real users. Discover which products truly live up to the hype. Start reading now. http://ads.osdn.com/?ad_id=6595&alloc_id=14396&op=click