Re: [RFC PATCH 2/2] ALSA: pcm: refresh capture pointers before drain

On Thu, 15 Oct 2026 01:52:03 +0200,
Joonho Lee wrote:

Hi Takashi,

Thanks, I've reworked the local version along those lines.

The helper takes snd_pcm_link_rwsem for reading and updates each running
member under its own stream lock. It uses snd_pcm_update_hw_ptr() for both
playback and capture. I call it after the existing pause-resume step so a
paused capture stream gets its final pointer update too. The pause and
drain-init actions remain outside the link rwsem.

A pointer error now follows the usual XRUN path and returns -EPIPE before
drain-init. This restores the group-wide error handling from v2 and removes
the "notify" boolean argument introduced in v3 proposal.

While checking this version, I also confirmed three problems in the first
patch: 
1) a reader already waiting for data could lose the capture tail
2) finishing a capture read could stop a linked playback stream
3) DELAYand SYNC_PTR(HWSYNC) rejected capture DRAINING. The local fixes preserve
the unread count when waking the reader, finish capture with
snd_pcm_drain_done(), and use the cached pointer for capture hwsync.

I reproduced all three failures with the helper plus the old first patch,
then checked the fixes in both atomic and nonatomic modes. The blocked
reader now receives all 64 final frames, capture EOF leaves linked playback
in DRAINING, and both position queries succeed without another pointer
callback. The 26 linked-stream checks, including paused drain and pointer
errors, and the 10 existing lifecycle/reset cases also passed under ARM64
QEMU with KASAN and lockdep, but physical real-world device testing is still pending in this time.

There is one concurrency point I'd like to check. The helper releases the
link rwsem before drain-init, and snd_pcm_link() allows another RUNNING
stream to join in that interval. The new member could enter drain-init
without being visited by the refresh loop.
My question is, should we handle concurrent linking across these two steps, or leave that outside this change's scope?

I guess we can leave this for now.  The invalid state like OPEN or
SUSPENDED is checked at snd_pcm_pre_drain_init().


thanks,

Takashi