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 listamd-gfx at lists.freedesktop.orghttps://lists.freedesktop.org/mailman/listinfo/amd-gfx
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/amd-gfx/attachments/20170710/68bd85c1/attachment-0001.html>
