<div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Wed, Jul 5, 2017 at 6:04 PM, Dave Airlie <span dir="ltr"><<a href="mailto:airlied@gmail.com" target="_blank">airlied@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">From: Dave Airlie <<a href="mailto:airlied@redhat.com">airlied@redhat.com</a>><br>
<br>
This interface will allow sync object to be used to back<br>
Vulkan fences. This API is pretty much the vulkan fence waiting<br>
API, and I've ported the code from amdgpu.<br>
<br>
v2: accept relative timeout, pass remaining time back<br>
to userspace.<br>
v3: return to absolute timeouts.<br>
v4: absolute zero = poll,<br>
    rewrite any/all code to have same operation for arrays<br>
    return -EINVAL for 0 fences.<br>
v4.1: fixup fences allocation check, use u64_to_user_ptr<br>
v5: move to sec/nsec, and use timespec64 for calcs.<br>
v6: use -ETIME and drop the out status flag. (-ETIME<br>
is suggested by ickle, I can feel a shed painting)<br>
<br>
Signed-off-by: Dave Airlie <<a href="mailto:airlied@redhat.com">airlied@redhat.com</a>><br>
---<br>
 drivers/gpu/drm/drm_internal.h |   2 +<br>
 drivers/gpu/drm/drm_ioctl.c    |   2 +<br>
 drivers/gpu/drm/drm_syncobj.c  | 142 ++++++++++++++++++++++++++++++<wbr>+++++++++++<br>
 include/uapi/drm/drm.h         |  13 ++++<br>
 4 files changed, 159 insertions(+)<br>
<br>
diff --git a/drivers/gpu/drm/drm_<wbr>internal.h b/drivers/gpu/drm/drm_<wbr>internal.h<br>
index 5cecc97..d71b50d 100644<br>
--- a/drivers/gpu/drm/drm_<wbr>internal.h<br>
+++ b/drivers/gpu/drm/drm_<wbr>internal.h<br>
@@ -157,3 +157,5 @@ int drm_syncobj_handle_to_fd_<wbr>ioctl(struct drm_device *dev, void *data,<br>
                                   struct drm_file *file_private);<br>
 int drm_syncobj_fd_to_handle_<wbr>ioctl(struct drm_device *dev, void *data,<br>
                                   struct drm_file *file_private);<br>
+int drm_syncobj_wait_ioctl(struct drm_device *dev, void *data,<br>
+                          struct drm_file *file_private);<br>
diff --git a/drivers/gpu/drm/drm_ioctl.c b/drivers/gpu/drm/drm_ioctl.c<br>
index f1e5681..385ce74 100644<br>
--- a/drivers/gpu/drm/drm_ioctl.c<br>
+++ b/drivers/gpu/drm/drm_ioctl.c<br>
@@ -657,6 +657,8 @@ static const struct drm_ioctl_desc drm_ioctls[] = {<br>
                      DRM_UNLOCKED|DRM_RENDER_ALLOW)<wbr>,<br>
        DRM_IOCTL_DEF(DRM_IOCTL_<wbr>SYNCOBJ_FD_TO_HANDLE, drm_syncobj_fd_to_handle_<wbr>ioctl,<br>
                      DRM_UNLOCKED|DRM_RENDER_ALLOW)<wbr>,<br>
+       DRM_IOCTL_DEF(DRM_IOCTL_<wbr>SYNCOBJ_WAIT, drm_syncobj_wait_ioctl,<br>
+                     DRM_UNLOCKED|DRM_RENDER_ALLOW)<wbr>,<br>
 };<br>
<br>
 #define DRM_CORE_IOCTL_COUNT   ARRAY_SIZE( drm_ioctls )<br>
diff --git a/drivers/gpu/drm/drm_syncobj.<wbr>c b/drivers/gpu/drm/drm_syncobj.<wbr>c<br>
index 89441bc..2d5a7a1 100644<br>
--- a/drivers/gpu/drm/drm_syncobj.<wbr>c<br>
+++ b/drivers/gpu/drm/drm_syncobj.<wbr>c<br>
@@ -1,5 +1,7 @@<br>
 /*<br>
  * Copyright 2017 Red Hat<br>
+ * Parts ported from amdgpu (fence wait code).<br>
+ * Copyright 2016 Advanced Micro Devices, Inc.<br>
  *<br>
  * Permission is hereby granted, free of charge, to any person obtaining a<br>
  * copy of this software and associated documentation files (the "Software"),<br>
@@ -31,6 +33,9 @@<br>
  * that contain an optional fence. The fence can be updated with a new<br>
  * fence, or be NULL.<br>
  *<br>
+ * syncobj's can be waited upon, where it will wait for the underlying<br>
+ * fence.<br>
+ *<br>
  * syncobj's can be export to fd's and back, these fd's are opaque and<br>
  * have no other use case, except passing the syncobj between processes.<br>
  *<br>
@@ -451,3 +456,140 @@ drm_syncobj_fd_to_handle_<wbr>ioctl(struct drm_device *dev, void *data,<br>
        return drm_syncobj_fd_to_handle(file_<wbr>private, args->fd,<br>
                                        &args->handle);<br>
 }<br>
+<br>
+/**<br>
+ * drm_timeout_abs_to_jiffies - calculate jiffies timeout from absolute value<br>
+ *<br>
+ * @timeout_sec: timeout sec component, 0 for poll<br>
+ * @timeout_nsec: timeout nsec component in ns, 0 for poll<br>
+ * both must be 0 for poll.<br>
+ *<br>
+ * Calculate the timeout in jiffies from an absolute time in sec/nsec.<br>
+ */<br>
+static unsigned long drm_timeout_abs_to_jiffies(<wbr>int64_t timeout_sec, uint64_t timeout_nsec)<br>
+{<br>
+       struct timespec64 abs_timeout, timeout, max_jiffy_timespec;<br>
+       unsigned long timeout_jiffies;<br>
+<br>
+       /* make 0 timeout means poll - absolute 0 doesn't seem valid */<br>
+       if (timeout_sec == 0 && timeout_nsec == 0)<br>
+               return 0;<br>
+<br>
+       abs_timeout.tv_sec = timeout_sec;<br>
+       abs_timeout.tv_nsec = timeout_nsec;<br>
+<br>
+       /* clamp timeout if it's to large */<br>
+       if (!timespec64_valid_strict(&<wbr>abs_timeout))<br>
+               return MAX_SCHEDULE_TIMEOUT - 1;<br>
+<br>
+       timeout = timespec64_sub(abs_timeout, ktime_to_timespec64(ktime_get(<wbr>)));<br>
+       if (!timespec64_valid(&timeout))<br>
+               return 0;<br>
+<br>
+       jiffies_to_timespec64(MAX_<wbr>JIFFY_OFFSET, &max_jiffy_timespec);<br>
+       if (timespec64_compare(&timeout, &max_jiffy_timespec) >= 0)<br>
+               return MAX_SCHEDULE_TIMEOUT - 1;<br>
+<br>
+       timeout_jiffies = timespec64_to_jiffies(&<wbr>timeout);<br>
+       /*  clamp timeout to avoid infinite timeout */<br>
+       if (timeout_jiffies >= MAX_SCHEDULE_TIMEOUT)<br>
+               return MAX_SCHEDULE_TIMEOUT - 1;<br>
+<br>
+       return timeout_jiffies + 1;<br>
+}<br>
+<br>
+static int drm_syncobj_wait_fences(struct drm_device *dev,<br>
+                                  struct drm_file *file_private,<br>
+                                  struct drm_syncobj_wait *wait,<br>
+                                  struct dma_fence **fences)<br>
+{<br>
+       unsigned long timeout = drm_timeout_abs_to_jiffies(<wbr>wait->timeout_sec, wait->timeout_nsec);<br>
+       int ret = 0;<br>
+       uint32_t first = ~0;<br>
+<br>
+       if (wait->flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_<wbr>ALL) {<br>
+               int i;<br>
+               for (i = 0; i < wait->count_handles; i++) {<br>
+                       ret = dma_fence_wait_timeout(fences[<wbr>i], true, timeout);<br>
+<br>
+                       if (ret < 0)<br>
+                               return ret;<br>
+                       if (ret == 0)<br>
+                               break;<br>
+                       timeout = ret;<br>
+               }<br>
+               first = 0;<br>
+       } else {<br>
+               ret = dma_fence_wait_any_timeout(<wbr>fences,<br>
+                                                wait->count_handles,<br>
+                                                true, timeout,<br>
+                                                &first);<br>
+       }<br>
+<br>
+       if (ret < 0)<br>
+               return ret;<br>
+<br>
+       wait->first_signaled = first;<br>
+       if (ret == 0)<br>
+               return -ETIME;<br>
+       return 0;<br>
+}<br>
+<br>
+int<br>
+drm_syncobj_wait_ioctl(struct drm_device *dev, void *data,<br>
+                      struct drm_file *file_private)<br>
+{<br>
+       struct drm_syncobj_wait *args = data;<br>
+       uint32_t *handles;<br>
+       struct dma_fence **fences;<br>
+       int ret = 0;<br>
+       int i;<br>
+<br>
+       if (!drm_core_check_feature(dev, DRIVER_SYNCOBJ))<br>
+               return -ENODEV;<br>
+<br>
+       if (args->flags != 0 && args->flags != DRM_SYNCOBJ_WAIT_FLAGS_WAIT_<wbr>ALL)<br>
+               return -EINVAL;<br>
+<br>
+       if (args->count_handles == 0)<br>
+               return -EINVAL;<br>
+<br>
+       /* Get the handles from userspace */<br>
+       handles = kmalloc_array(args->count_<wbr>handles, sizeof(uint32_t),<br>
+                               GFP_KERNEL);<br>
+       if (handles == NULL)<br>
+               return -ENOMEM;<br>
+<br>
+       if (copy_from_user(handles,<br>
+                          u64_to_user_ptr(args->handles)<wbr>,<br>
+                          sizeof(uint32_t) * args->count_handles)) {<br>
+               ret = -EFAULT;<br>
+               goto err_free_handles;<br>
+       }<br>
+<br>
+       fences = kcalloc(args->count_handles,<br>
+                        sizeof(struct dma_fence *), GFP_KERNEL);<br>
+       if (!fences) {<br>
+               ret = -ENOMEM;<br>
+               goto err_free_handles;<br>
+       }<br>
+<br>
+       for (i = 0; i < args->count_handles; i++) {<br>
+               ret = drm_syncobj_fence_get(file_<wbr>private, handles[i],<br>
+                                           &fences[i]);<br>
+               if (ret)<br>
+                       goto err_free_fence_array;<br>
+       }<br>
+<br>
+       ret = drm_syncobj_wait_fences(dev, file_private,<br>
+                                     args, fences);<br></blockquote><div><br></div><div>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.  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.<br></div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
+<br>
+err_free_fence_array:<br>
+       for (i = 0; i < args->count_handles; i++)<br>
+               dma_fence_put(fences[i]);<br>
+       kfree(fences);<br>
+err_free_handles:<br>
+       kfree(handles);<br>
+<br>
+       return ret;<br>
+}<br>
diff --git a/include/uapi/drm/drm.h b/include/uapi/drm/drm.h<br>
index 101593a..91746a7 100644<br>
--- a/include/uapi/drm/drm.h<br>
+++ b/include/uapi/drm/drm.h<br>
@@ -718,6 +718,18 @@ struct drm_syncobj_handle {<br>
        __u32 pad;<br>
 };<br>
<br>
+#define DRM_SYNCOBJ_WAIT_FLAGS_WAIT_<wbr>ALL (1 << 0)<br>
+struct drm_syncobj_wait {<br>
+       __u64 handles;<br>
+       /* absolute timeout */<br>
+       __s64 timeout_sec;<br>
+       __s64 timeout_nsec;<br>
+       __u32 count_handles;<br>
+       __u32 flags;<br>
+       __u32 first_signaled; /* only valid when not waiting all */<br>
+       __u32 pad;<br>
+};<br>
+<br>
 #if defined(__cplusplus)<br>
 }<br>
 #endif<br>
@@ -840,6 +852,7 @@ extern "C" {<br>
 #define DRM_IOCTL_SYNCOBJ_DESTROY      DRM_IOWR(0xC0, struct drm_syncobj_destroy)<br>
 #define DRM_IOCTL_SYNCOBJ_HANDLE_TO_FD DRM_IOWR(0xC1, struct drm_syncobj_handle)<br>
 #define DRM_IOCTL_SYNCOBJ_FD_TO_HANDLE DRM_IOWR(0xC2, struct drm_syncobj_handle)<br>
+#define DRM_IOCTL_SYNCOBJ_WAIT         DRM_IOWR(0xC3, struct drm_syncobj_wait)<br>
<br>
 /**<br>
  * Device specific ioctls should only be in their respective headers<br>
<span class="HOEnZb"><font color="#888888">--<br>
2.9.4<br>
<br>
______________________________<wbr>_________________<br>
dri-devel mailing list<br>
<a href="mailto:dri-devel@lists.freedesktop.org">dri-devel@lists.freedesktop.<wbr>org</a><br>
<a href="https://lists.freedesktop.org/mailman/listinfo/dri-devel" rel="noreferrer" target="_blank">https://lists.freedesktop.org/<wbr>mailman/listinfo/dri-devel</a><br>
</font></span></blockquote></div><br></div></div>