[Intel-gfx] [PATCH igt 1/3] benchmark/gem_busy: Compare polling with syncobj_wait

[Chris Wilson]

Quoting Tvrtko Ursulin (2017-09-28 10:16:58)
> 
> On 28/09/2017 10:07, Chris Wilson wrote:
> > Quoting Tvrtko Ursulin (2017-09-28 07:53:56)
> >>
> >> On 25/09/2017 21:26, Chris Wilson wrote:
> >>> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> >>> ---
> >>>    benchmarks/gem_busy.c | 73 ++++++++++++++++++++++++++++++++++++++++++++++++++-
> >>>    1 file changed, 72 insertions(+), 1 deletion(-)
> >>>
> >>> diff --git a/benchmarks/gem_busy.c b/benchmarks/gem_busy.c
> >>> index f050454b..9649ea02 100644
> >>> --- a/benchmarks/gem_busy.c
> >>> +++ b/benchmarks/gem_busy.c
> >>> @@ -58,6 +58,15 @@
> >>>    #define DMABUF 0x4
> >>>    #define WAIT 0x8
> >>>    #define SYNC 0x10
> >>> +#define SYNCOBJ 0x20
> >>> +
> >>> +#define LOCAL_I915_EXEC_FENCE_ARRAY (1 << 19)
> >>> +struct local_gem_exec_fence {
> >>> +     uint32_t handle;
> >>> +     uint32_t flags;
> >>> +#define LOCAL_EXEC_FENCE_WAIT (1 << 0)
> >>> +#define LOCAL_EXEC_FENCE_SIGNAL (1 << 1)
> >>> +};
> >>>    
> >>>    static void gem_busy(int fd, uint32_t handle)
> >>>    {
> >>> @@ -109,11 +118,54 @@ static int sync_merge(int fd1, int fd2)
> >>>        return data.fence;
> >>>    }
> >>>    
> >>> +static uint32_t __syncobj_create(int fd)
> >>> +{
> >>> +     struct local_syncobj_create {
> >>> +             uint32_t handle, flags;
> >>> +     } arg;
> >>> +#define LOCAL_IOCTL_SYNCOBJ_CREATE        DRM_IOWR(0xBF, struct local_syncobj_create)
> >>> +
> >>> +     memset(&arg, 0, sizeof(arg));
> >>> +     ioctl(fd, LOCAL_IOCTL_SYNCOBJ_CREATE, &arg);
> >>> +
> >>> +     return arg.handle;
> >>> +}
> >>> +
> >>> +static uint32_t syncobj_create(int fd)
> >>> +{
> >>> +     uint32_t ret;
> >>> +
> >>> +     igt_assert_neq((ret = __syncobj_create(fd)), 0);
> >>> +
> >>> +     return ret;
> >>> +}
> >>> +
> >>> +#define LOCAL_SYNCOBJ_WAIT_FLAGS_WAIT_ALL (1 << 0)
> >>> +#define LOCAL_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT (1 << 1)
> >>> +struct local_syncobj_wait {
> >>> +       __u64 handles;
> >>> +       /* absolute timeout */
> >>> +       __s64 timeout_nsec;
> >>> +       __u32 count_handles;
> >>> +       __u32 flags;
> >>> +       __u32 first_signaled; /* only valid when not waiting all */
> >>> +       __u32 pad;
> >>> +};
> >>> +#define LOCAL_IOCTL_SYNCOBJ_WAIT     DRM_IOWR(0xC3, struct local_syncobj_wait)
> >>> +static int __syncobj_wait(int fd, struct local_syncobj_wait *args)
> >>> +{
> >>> +     int err = 0;
> >>> +     if (drmIoctl(fd, LOCAL_IOCTL_SYNCOBJ_WAIT, args))
> >>> +             err = -errno;
> >>> +     return err;
> >>> +}
> >>> +
> >>>    static int loop(unsigned ring, int reps, int ncpus, unsigned flags)
> >>>    {
> >>>        struct drm_i915_gem_execbuffer2 execbuf;
> >>>        struct drm_i915_gem_exec_object2 obj[2];
> >>>        struct drm_i915_gem_relocation_entry reloc[2];
> >>> +     struct local_gem_exec_fence syncobj;
> >>>        unsigned engines[16];
> >>>        unsigned nengine;
> >>>        uint32_t *batch;
> >>> @@ -126,6 +178,11 @@ static int loop(unsigned ring, int reps, int ncpus, unsigned flags)
> >>>        fd = drm_open_driver(DRIVER_INTEL);
> >>>        gen = intel_gen(intel_get_drm_devid(fd));
> >>>    
> >>> +     if (flags & SYNCOBJ) {
> >>> +             syncobj.handle = syncobj_create(fd);
> >>> +             syncobj.flags = LOCAL_EXEC_FENCE_SIGNAL;
> >>> +     }
> >>> +
> >>>        memset(obj, 0, sizeof(obj));
> >>>        obj[0].handle = gem_create(fd, 4096);
> >>>        if (flags & WRITE)
> >>> @@ -144,6 +201,8 @@ static int loop(unsigned ring, int reps, int ncpus, unsigned flags)
> >>>        execbuf.buffer_count = 2;
> >>>        execbuf.flags |= LOCAL_I915_EXEC_HANDLE_LUT;
> >>>        execbuf.flags |= LOCAL_I915_EXEC_NO_RELOC;
> >>> +     if (flags & SYNCOBJ)
> >>> +             execbuf.flags |= LOCAL_I915_EXEC_FENCE_ARRAY;
> >>
> >> According to the comment in i915_drm.h, when this is specified, syncobj
> >> should be also passed in in cliprects_ptr but that's not happening?
> > 
> > You want -b support as well! :)
> 
> I just failed to figure out where is the connection between syncobj 
> (local var) and execbuf. Flag is set, so what happens next? Execbuf 
> fails since cliprects_ptr is not set?

Your observation is correct. The code never set the syncobj to be
signaled. I was just saying that in the manner of polling on the idle
syncobj, it should be ok, though strictly it will be reporting that
there is no fence attached rather than idle. I need to add code to
support the '-b' mode of polling on a busy fence.
-Chris