[PATCH] drm/syncobj: add sync obj wait interface. (v6)

Jason Ekstrand jason at jlekstrand.net
Mon Jul 10 23:15:00 UTC 2017


On Mon, Jul 10, 2017 at 9:13 AM, Christian König <deathsimple at vodafone.de>
wrote:

> Am 10.07.2017 um 17:58 schrieb Xie, AlexBin:
>
> I understand this discussion from closes source driver terminology.
>
>
>
> If a process is killed before it sends out the signaling command, will
> some part of the GPU be in a waiting situation forever?
>
>
> Yes, exactly that's the problem here and the reason why that even
> Microsoft forbids that under windows.
>

To be clear, we are only discussing wait-before-submit for client CPU
waits.  The GPU will not be waiting, only some userspace process.  If that
process is a compositor, it should take measures to avoid getting hung up
by bad clients.  What this will *not* cause is the GPU (or kernel GPU
scheduling) to get hung up on waiting for some unsubmitted thing.


> Christian.
>
>
>
>
> Alex Bin Xie
>
> *From:* amd-gfx [mailto:amd-gfx-bounces at lists.freedesktop.org
> <amd-gfx-bounces at lists.freedesktop.org>] *On Behalf Of *Jason Ekstrand
> *Sent:* Monday, July 10, 2017 11:53 AM
> *To:* Christian König <deathsimple at vodafone.de> <deathsimple at vodafone.de>
> *Cc:* Dave Airlie <airlied at gmail.com> <airlied at gmail.com>; Maling list -
> DRI developers <dri-devel at lists.freedesktop.org>
> <dri-devel at lists.freedesktop.org>; amd-gfx mailing list
> <amd-gfx at lists.freedesktop.org> <amd-gfx at lists.freedesktop.org>
> *Subject:* Re: [PATCH] drm/syncobj: add sync obj wait interface. (v6)
>
>
>
> On Mon, Jul 10, 2017 at 8:45 AM, Christian König <deathsimple at vodafone.de>
> wrote:
>
> Am 10.07.2017 um 17:28 schrieb Jason Ekstrand:
>
> On Wed, Jul 5, 2017 at 6:04 PM, Dave Airlie <airlied at gmail.com> wrote:
>
> From: Dave Airlie <airlied at redhat.com>
>
> This interface will allow sync object to be used to back
> Vulkan fences. This API is pretty much the vulkan fence waiting
> API, and I've ported the code from amdgpu.
>
> v2: accept relative timeout, pass remaining time back
> to userspace.
> v3: return to absolute timeouts.
> v4: absolute zero = poll,
>     rewrite any/all code to have same operation for arrays
>     return -EINVAL for 0 fences.
> v4.1: fixup fences allocation check, use u64_to_user_ptr
> v5: move to sec/nsec, and use timespec64 for calcs.
> v6: use -ETIME and drop the out status flag. (-ETIME
> is suggested by ickle, I can feel a shed painting)
>
> Signed-off-by: Dave Airlie <airlied at redhat.com>
> ---
>  drivers/gpu/drm/drm_internal.h |   2 +
>  drivers/gpu/drm/drm_ioctl.c    |   2 +
>  drivers/gpu/drm/drm_syncobj.c  | 142 ++++++++++++++++++++++++++++++
> +++++++++++
>  include/uapi/drm/drm.h         |  13 ++++
>  4 files changed, 159 insertions(+)
>
> diff --git a/drivers/gpu/drm/drm_internal.h b/drivers/gpu/drm/drm_
> internal.h
> index 5cecc97..d71b50d 100644
> --- a/drivers/gpu/drm/drm_internal.h
> +++ b/drivers/gpu/drm/drm_internal.h
> @@ -157,3 +157,5 @@ int drm_syncobj_handle_to_fd_ioctl(struct drm_device
> *dev, void *data,
>                                    struct drm_file *file_private);
>  int drm_syncobj_fd_to_handle_ioctl(struct drm_device *dev, void *data,
>                                    struct drm_file *file_private);
> +int drm_syncobj_wait_ioctl(struct drm_device *dev, void *data,
> +                          struct drm_file *file_private);
> diff --git a/drivers/gpu/drm/drm_ioctl.c b/drivers/gpu/drm/drm_ioctl.c
> index f1e5681..385ce74 100644
> --- a/drivers/gpu/drm/drm_ioctl.c
> +++ b/drivers/gpu/drm/drm_ioctl.c
> @@ -657,6 +657,8 @@ static const struct drm_ioctl_desc drm_ioctls[] = {
>                       DRM_UNLOCKED|DRM_RENDER_ALLOW),
>         DRM_IOCTL_DEF(DRM_IOCTL_SYNCOBJ_FD_TO_HANDLE,
> drm_syncobj_fd_to_handle_ioctl,
>                       DRM_UNLOCKED|DRM_RENDER_ALLOW),
> +       DRM_IOCTL_DEF(DRM_IOCTL_SYNCOBJ_WAIT, drm_syncobj_wait_ioctl,
> +                     DRM_UNLOCKED|DRM_RENDER_ALLOW),
>  };
>
>  #define DRM_CORE_IOCTL_COUNT   ARRAY_SIZE( drm_ioctls )
> diff --git a/drivers/gpu/drm/drm_syncobj.c b/drivers/gpu/drm/drm_syncobj.c
> index 89441bc..2d5a7a1 100644
> --- a/drivers/gpu/drm/drm_syncobj.c
> +++ b/drivers/gpu/drm/drm_syncobj.c
> @@ -1,5 +1,7 @@
>  /*
>   * Copyright 2017 Red Hat
> + * Parts ported from amdgpu (fence wait code).
> + * Copyright 2016 Advanced Micro Devices, Inc.
>   *
>   * Permission is hereby granted, free of charge, to any person obtaining a
>   * copy of this software and associated documentation files (the
> "Software"),
> @@ -31,6 +33,9 @@
>   * that contain an optional fence. The fence can be updated with a new
>   * fence, or be NULL.
>   *
> + * syncobj's can be waited upon, where it will wait for the underlying
> + * fence.
> + *
>   * syncobj's can be export to fd's and back, these fd's are opaque and
>   * have no other use case, except passing the syncobj between processes.
>   *
> @@ -451,3 +456,140 @@ drm_syncobj_fd_to_handle_ioctl(struct drm_device
> *dev, void *data,
>         return drm_syncobj_fd_to_handle(file_private, args->fd,
>                                         &args->handle);
>  }
> +
> +/**
> + * drm_timeout_abs_to_jiffies - calculate jiffies timeout from absolute
> value
> + *
> + * @timeout_sec: timeout sec component, 0 for poll
> + * @timeout_nsec: timeout nsec component in ns, 0 for poll
> + * both must be 0 for poll.
> + *
> + * Calculate the timeout in jiffies from an absolute time in sec/nsec.
> + */
> +static unsigned long drm_timeout_abs_to_jiffies(int64_t timeout_sec,
> uint64_t timeout_nsec)
> +{
> +       struct timespec64 abs_timeout, timeout, max_jiffy_timespec;
> +       unsigned long timeout_jiffies;
> +
> +       /* make 0 timeout means poll - absolute 0 doesn't seem valid */
> +       if (timeout_sec == 0 && timeout_nsec == 0)
> +               return 0;
> +
> +       abs_timeout.tv_sec = timeout_sec;
> +       abs_timeout.tv_nsec = timeout_nsec;
> +
> +       /* clamp timeout if it's to large */
> +       if (!timespec64_valid_strict(&abs_timeout))
> +               return MAX_SCHEDULE_TIMEOUT - 1;
> +
> +       timeout = timespec64_sub(abs_timeout,
> ktime_to_timespec64(ktime_get()));
> +       if (!timespec64_valid(&timeout))
> +               return 0;
> +
> +       jiffies_to_timespec64(MAX_JIFFY_OFFSET, &max_jiffy_timespec);
> +       if (timespec64_compare(&timeout, &max_jiffy_timespec) >= 0)
> +               return MAX_SCHEDULE_TIMEOUT - 1;
> +
> +       timeout_jiffies = timespec64_to_jiffies(&timeout);
> +       /*  clamp timeout to avoid infinite timeout */
> +       if (timeout_jiffies >= MAX_SCHEDULE_TIMEOUT)
> +               return MAX_SCHEDULE_TIMEOUT - 1;
> +
> +       return timeout_jiffies + 1;
> +}
> +
> +static int drm_syncobj_wait_fences(struct drm_device *dev,
> +                                  struct drm_file *file_private,
> +                                  struct drm_syncobj_wait *wait,
> +                                  struct dma_fence **fences)
> +{
> +       unsigned long timeout = drm_timeout_abs_to_jiffies(wait->timeout_sec,
> wait->timeout_nsec);
> +       int ret = 0;
> +       uint32_t first = ~0;
> +
> +       if (wait->flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL) {
> +               int i;
> +               for (i = 0; i < wait->count_handles; i++) {
> +                       ret = dma_fence_wait_timeout(fences[i], true,
> timeout);
> +
> +                       if (ret < 0)
> +                               return ret;
> +                       if (ret == 0)
> +                               break;
> +                       timeout = ret;
> +               }
> +               first = 0;
> +       } else {
> +               ret = dma_fence_wait_any_timeout(fences,
> +                                                wait->count_handles,
> +                                                true, timeout,
> +                                                &first);
> +       }
> +
> +       if (ret < 0)
> +               return ret;
> +
> +       wait->first_signaled = first;
> +       if (ret == 0)
> +               return -ETIME;
> +       return 0;
> +}
> +
> +int
> +drm_syncobj_wait_ioctl(struct drm_device *dev, void *data,
> +                      struct drm_file *file_private)
> +{
> +       struct drm_syncobj_wait *args = data;
> +       uint32_t *handles;
> +       struct dma_fence **fences;
> +       int ret = 0;
> +       int i;
> +
> +       if (!drm_core_check_feature(dev, DRIVER_SYNCOBJ))
> +               return -ENODEV;
> +
> +       if (args->flags != 0 && args->flags != DRM_SYNCOBJ_WAIT_FLAGS_WAIT_
> ALL)
> +               return -EINVAL;
> +
> +       if (args->count_handles == 0)
> +               return -EINVAL;
> +
> +       /* Get the handles from userspace */
> +       handles = kmalloc_array(args->count_handles, sizeof(uint32_t),
> +                               GFP_KERNEL);
> +       if (handles == NULL)
> +               return -ENOMEM;
> +
> +       if (copy_from_user(handles,
> +                          u64_to_user_ptr(args->handles),
> +                          sizeof(uint32_t) * args->count_handles)) {
> +               ret = -EFAULT;
> +               goto err_free_handles;
> +       }
> +
> +       fences = kcalloc(args->count_handles,
> +                        sizeof(struct dma_fence *), GFP_KERNEL);
> +       if (!fences) {
> +               ret = -ENOMEM;
> +               goto err_free_handles;
> +       }
> +
> +       for (i = 0; i < args->count_handles; i++) {
> +               ret = drm_syncobj_fence_get(file_private, handles[i],
> +                                           &fences[i]);
> +               if (ret)
> +                       goto err_free_fence_array;
> +       }
> +
> +       ret = drm_syncobj_wait_fences(dev, file_private,
> +                                     args, fences);
>
>
>
> So, reading some CTS tests again, and I think we have a problem here.  The
> Vulkan spec allows you to wait on a fence that is in the unsignaled state.
>
>
>
> At least on the closed source driver that would be illegal as far as I
> know.
>
>
>
> Then they are doing workarounds in userspace.  There are definitely CTS
> tests for this:
>
> https://github.com/KhronosGroup/VK-GL-CTS/blob/master/external/vulkancts/
> modules/vulkan/synchronization/vktSynchronizationBasicFenceTests.cpp#L74
>
>
>
> You can't wait on a semaphore before the signal operation is send down to
> the kerel.
>
>
>
> We (Intel) deal with this today by tracking whether or not the fence has
> been submitted and using a condition variable in userspace to sort it all
> out.  If we ever want to share fences across processes (which we do), then
> this needs to be sorted in the kernel.
>
> --Jason
>
>
>
>
>
> Regards,
> Christian.
>
>
>
>
>   In theory, you could have thread A start waiting on a fence before
> thread B submits the work which triggers that fence.  This means that the
> dma_fence may not exist yet when vkWaitForFences gets called.  If we really
> want to support the full Vulkan usage, we need to somehow support missing
> dma_fences by waiting for the dma_fence to show up.  Unfortunately, I don't
> know enough about the internal kernel APIs to know what that would look
> like.
>
>
>
> +
> +err_free_fence_array:
> +       for (i = 0; i < args->count_handles; i++)
> +               dma_fence_put(fences[i]);
> +       kfree(fences);
> +err_free_handles:
> +       kfree(handles);
> +
> +       return ret;
> +}
> diff --git a/include/uapi/drm/drm.h b/include/uapi/drm/drm.h
> index 101593a..91746a7 100644
> --- a/include/uapi/drm/drm.h
> +++ b/include/uapi/drm/drm.h
> @@ -718,6 +718,18 @@ struct drm_syncobj_handle {
>         __u32 pad;
>  };
>
> +#define DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL (1 << 0)
> +struct drm_syncobj_wait {
> +       __u64 handles;
> +       /* absolute timeout */
> +       __s64 timeout_sec;
> +       __s64 timeout_nsec;
> +       __u32 count_handles;
> +       __u32 flags;
> +       __u32 first_signaled; /* only valid when not waiting all */
> +       __u32 pad;
> +};
> +
>  #if defined(__cplusplus)
>  }
>  #endif
> @@ -840,6 +852,7 @@ extern "C" {
>  #define DRM_IOCTL_SYNCOBJ_DESTROY      DRM_IOWR(0xC0, struct
> drm_syncobj_destroy)
>  #define DRM_IOCTL_SYNCOBJ_HANDLE_TO_FD DRM_IOWR(0xC1, struct
> drm_syncobj_handle)
>  #define DRM_IOCTL_SYNCOBJ_FD_TO_HANDLE DRM_IOWR(0xC2, struct
> drm_syncobj_handle)
> +#define DRM_IOCTL_SYNCOBJ_WAIT         DRM_IOWR(0xC3, struct
> drm_syncobj_wait)
>
>  /**
>   * Device specific ioctls should only be in their respective headers
> --
> 2.9.4
>
> _______________________________________________
> dri-devel mailing list
> dri-devel at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/dri-devel
>
>
>
>
>
> _______________________________________________
>
> amd-gfx mailing list
>
> amd-gfx at lists.freedesktop.org
>
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
>
>
>
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/amd-gfx/attachments/20170710/5309747f/attachment-0001.html>


More information about the amd-gfx mailing list