| Message ID | AM7PR03MB6660AC1E45A4490D61118BCA8FDE9@AM7PR03MB6660.eurprd03.prod.outlook.com |
|---|---|
| State | Accepted |
| Commit | f440c422b70ce76f225a34ebf168215a432e8d88 |
| Headers | show |
| Series | [FFmpeg-devel,1/3] libswscale/options: Add parent_log_context_offset to AVClass | expand |
| Context | Check | Description |
|---|---|---|
| andriy/make_x86 | success | Make finished |
| andriy/make_fate_x86 | success | Make fate finished |
| andriy/make_ppc | success | Make finished |
| andriy/make_fate_ppc | success | Make fate finished |
On Sat, Sep 18, 2021 at 05:43:25AM +0200, Andreas Rheinhardt wrote: > In this case the current code tries to warn once; to do so, it uses > ordinary static ints to store whether the warning has already been > emitted. This is both a data race (and therefore undefined behaviour) > as well as a race condition, because it is really possible for multiple > threads to be the one thread to emit the warning. This is actually > common since the introduction of the new multithreaded scaling API. > > This commit fixes this by using atomic integers for the state; > furthermore, these are not static anymore, but rather contained > in the user-facing SwsContext (i.e. the parent SwsContext in case > of slice-threading). > > Given that these atomic variables are not intended for synchronization > at all (but only for atomicity, i.e. only to output the warning once), > the atomic operations use memory_order_relaxed. > > This affected nv12, nv21, yuv420, yuv420p10, yuv422, yuv422p10 and > yuv444 filter-overlay FATE-tests. > > Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com> > --- > Of course, one could also use static atomic integers. should be ok thx [...]
diff --git a/libswscale/swscale.c b/libswscale/swscale.c index b9c9647fcb..040172752f 100644 --- a/libswscale/swscale.c +++ b/libswscale/swscale.c @@ -312,12 +312,12 @@ static int swscale(SwsContext *c, const uint8_t *src[], if (dstStride[0]&15 || dstStride[1]&15 || dstStride[2]&15 || dstStride[3]&15) { - static int warnedAlready = 0; // FIXME maybe move this into the context - if (flags & SWS_PRINT_INFO && !warnedAlready) { + SwsContext *const ctx = c->parent ? c->parent : c; + if (flags & SWS_PRINT_INFO && + !atomic_exchange_explicit(&ctx->stride_unaligned_warned, 1, memory_order_relaxed)) { av_log(c, AV_LOG_WARNING, "Warning: dstStride is not aligned!\n" " ->cannot do aligned memory accesses anymore\n"); - warnedAlready = 1; } } @@ -326,11 +326,11 @@ static int swscale(SwsContext *c, const uint8_t *src[], || dstStride[0]&15 || dstStride[1]&15 || dstStride[2]&15 || dstStride[3]&15 || srcStride[0]&15 || srcStride[1]&15 || srcStride[2]&15 || srcStride[3]&15 ) { - static int warnedAlready=0; + SwsContext *const ctx = c->parent ? c->parent : c; int cpu_flags = av_get_cpu_flags(); - if (HAVE_MMXEXT && (cpu_flags & AV_CPU_FLAG_SSE2) && !warnedAlready){ + if (HAVE_MMXEXT && (cpu_flags & AV_CPU_FLAG_SSE2) && + !atomic_exchange_explicit(&ctx->stride_unaligned_warned,1, memory_order_relaxed)) { av_log(c, AV_LOG_WARNING, "Warning: data is not aligned! This can lead to a speed loss\n"); - warnedAlready=1; } } diff --git a/libswscale/swscale_internal.h b/libswscale/swscale_internal.h index 0d60dd2e6f..e6e7b934b6 100644 --- a/libswscale/swscale_internal.h +++ b/libswscale/swscale_internal.h @@ -21,6 +21,8 @@ #ifndef SWSCALE_SWSCALE_INTERNAL_H #define SWSCALE_SWSCALE_INTERNAL_H +#include <stdatomic.h> + #include "config.h" #include "version.h" @@ -672,6 +674,8 @@ typedef struct SwsContext { unsigned int xyz_scratch_allocated; unsigned int dst_slice_align; + atomic_int stride_unaligned_warned; + atomic_int data_unaligned_warned; } SwsContext; //FIXME check init (where 0) diff --git a/libswscale/utils.c b/libswscale/utils.c index 84a29c4dc7..fcfba971c6 100644 --- a/libswscale/utils.c +++ b/libswscale/utils.c @@ -1112,6 +1112,8 @@ SwsContext *sws_alloc_context(void) if (c) { c->av_class = &ff_sws_context_class; av_opt_set_defaults(c); + atomic_init(&c->stride_unaligned_warned, 0); + atomic_init(&c->data_unaligned_warned, 0); } return c;
In this case the current code tries to warn once; to do so, it uses ordinary static ints to store whether the warning has already been emitted. This is both a data race (and therefore undefined behaviour) as well as a race condition, because it is really possible for multiple threads to be the one thread to emit the warning. This is actually common since the introduction of the new multithreaded scaling API. This commit fixes this by using atomic integers for the state; furthermore, these are not static anymore, but rather contained in the user-facing SwsContext (i.e. the parent SwsContext in case of slice-threading). Given that these atomic variables are not intended for synchronization at all (but only for atomicity, i.e. only to output the warning once), the atomic operations use memory_order_relaxed. This affected nv12, nv21, yuv420, yuv420p10, yuv422, yuv422p10 and yuv444 filter-overlay FATE-tests. Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com> --- Of course, one could also use static atomic integers. libswscale/swscale.c | 12 ++++++------ libswscale/swscale_internal.h | 4 ++++ libswscale/utils.c | 2 ++ 3 files changed, 12 insertions(+), 6 deletions(-)