[FFmpeg-devel] [PATCH 31/92] Vulkan patchset part 1 - common code changes

[Anton Khirnov]

Quoting Lynne (2023-04-19 16:39:28)
> Apr 17, 2023, 11:25 by anton at khirnov.net:
> 
> > Quoting Lynne (2023-03-14 07:33:43)
> >
> >> From 6e2dfc44e50798264eb16bc2dcabfdbf2fbac2d7 Mon Sep 17 00:00:00 2001
> >> From: Lynne <dev at lynne.ee>
> >> Date: Thu, 15 Dec 2022 01:06:52 +0100
> >> Subject: [PATCH 24/92] hwconfig: add a new HWACCEL_CAP_THREAD_SAFE for
> >>  threadsafe hwaccels
> >>
> >> Vulkan is fully threadsafe and stateless, so we can benefit from this.
> >> ---
> >>  libavcodec/hwconfig.h      | 1 +
> >>  libavcodec/pthread_frame.c | 2 +-
> >>  2 files changed, 2 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/libavcodec/hwconfig.h b/libavcodec/hwconfig.h
> >> index 721424912c..e6b78f0160 100644
> >> --- a/libavcodec/hwconfig.h
> >> +++ b/libavcodec/hwconfig.h
> >> @@ -24,6 +24,7 @@
> >>  
> >>  
> >>  #define HWACCEL_CAP_ASYNC_SAFE      (1 << 0)
> >> +#define HWACCEL_CAP_THREAD_SAFE     (1 << 1)
> >>  
> >>  
> >>  typedef struct AVCodecHWConfigInternal {
> >> diff --git a/libavcodec/pthread_frame.c b/libavcodec/pthread_frame.c
> >> index 74864e19c5..c096287233 100644
> >> --- a/libavcodec/pthread_frame.c
> >> +++ b/libavcodec/pthread_frame.c
> >> @@ -204,7 +204,7 @@ static attribute_align_arg void *frame_worker_thread(void *arg)
> >>  
> >>  /* if the previous thread uses hwaccel then we take the lock to ensure
> >>  * the threads don't run concurrently */
> >> -        if (avctx->hwaccel) {
> >> +        if (avctx->hwaccel && !(avctx->hwaccel->caps_internal & HWACCEL_CAP_THREAD_SAFE)) {
> >>
> >
> > A major problem here is that hwaccel_priv_data exists in only a single
> > instance shared across all the worker threads. If you want frame-threaded
> > hwaccels, you'll need to add some mechanism for synchronizing it between
> > threads.
> >
> > I really wonder how does your code survive all the races involved.
> >
> 
> It works because hwaccel_priv_data is read-only. No need to synchronize
> anything, the contexts Vulkan creates are defined as driver-synchronized,
> and all the info a frame needs is provided when decoding it, rather than
> kept as a state.
> So I don't see a problem here.

>From a quick glance, at least vk_av1_free_frame_priv() modifies this
context, which seems highly unsafe.

This really seems like a disaster waiting to happen, there should be
some stronger guarantees that nobody will modify this context when they
shouldn't.

-- 
Anton Khirnov