drm/radeon: Fixup hw vblank counter/ts for new drm_update_vblank_count() (v2)
commit 4dfd6486 "drm: Use vblank timestamps to guesstimate how many vblanks were missed" introduced in Linux 4.4-rc1 makes the drm core more fragile to drivers which don't update hw vblank counters and vblank timestamps in sync with firing of the vblank irq and essentially at leading edge of vblank. This exposed a problem with radeon-kms/amdgpu-kms which do not satisfy above requirements: The vblank irq fires a few scanlines before start of vblank, but programmed pageflips complete at start of vblank and vblank timestamps update at start of vblank, whereas the hw vblank counter increments only later, at start of vsync. This leads to problems like off by one errors for vblank counter updates, vblank counters apparently going backwards or vblank timestamps apparently having time going backwards. The net result is stuttering of graphics in games, or little hangs, as well as total failure of timing sensitive applications. See bug #93147 for an example of the regression on Linux 4.4-rc: https://bugs.freedesktop.org/show_bug.cgi?id=93147 This patch tries to align all above events better from the viewpoint of the drm core / of external callers to fix the problem: 1. The apparent start of vblank is shifted a few scanlines earlier, so the vblank irq now always happens after start of this extended vblank interval and thereby drm_update_vblank_count() always samples the updated vblank count and timestamp of the new vblank interval. To achieve this, the reporting of scanout positions by radeon_get_crtc_scanoutpos() now operates as if the vblank starts radeon_crtc->lb_vblank_lead_lines before the real start of the hw vblank interval. This means that the vblank timestamps which are based on these scanout positions will now update at this earlier start of vblank. 2. The driver->get_vblank_counter() function will bump the returned vblank count as read from the hw by +1 if the query happens after the shifted earlier start of the vblank, but before the real hw increment at start of vsync, so the counter appears to increment at start of vblank in sync with the timestamp update. 3. Calls from vblank irq-context and regular non-irq calls are now treated identical, always simulating the shifted vblank start, to avoid inconsistent results for queries happening from vblank irq vs. happening from drm_vblank_enable() or vblank_disable_fn(). 4. The radeon_flip_work_func will delay mmio programming a pageflip until the start of the real vblank iff it happens to execute inside the shifted earlier start of the vblank, so pageflips now also appear to execute at start of the shifted vblank, in sync with vblank counter and timestamp updates. This to avoid some races between updates of vblank count and timestamps that are used for swap scheduling and pageflip execution which could cause pageflips to execute before the scheduled target vblank. The lb_vblank_lead_lines "fudge" value is calculated as the size of the display controllers line buffer in scanlines for the given video mode: Vblank irq's are triggered by the line buffer logic when the line buffer refill for a video frame ends, ie. when the line buffer source read position enters the hw vblank. This means that a vblank irq could fire at most as many scanlines before the current reported scanout position of the crtc timing generator as the number of scanlines the line buffer can maximally hold for a given video mode. This patch has been successfully tested on a RV730 card with DCE-3 display engine and on a evergreen card with DCE-4 display engine, in single-display and dual-display configuration, with different video modes. A similar patch is needed for amdgpu-kms to fix the same problem. Limitations: - Line buffer sizes in pixels are hard-coded on < DCE-4 to a value i just guessed to be high enough to work ok, lacking info on the true sizes atm. Fixes: fdo#93147 Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Cc: Alex Deucher <alexander.deucher@amd.com> Cc: Michel Dänzer <michel.daenzer@amd.com> Cc: Harry Wentland <Harry.Wentland@amd.com> Cc: Ville Syrjälä <ville.syrjala@linux.intel.com> (v1) Tested-by: Dave Witbrodt <dawitbro@sbcglobal.net> (v2) Refine radeon_flip_work_func() for better efficiency: In radeon_flip_work_func, replace the busy waiting udelay(5) with event lock held by a more performance and energy efficient usleep_range() until at least predicted true start of hw vblank, with some slack for scheduler happiness. Release the event lock during waits to not delay other outputs in doing their stuff, as the waiting can last up to 200 usecs in some cases. Retested on DCE-3 and DCE-4 to verify it still works nicely. (v2) Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Showing
Please register or sign in to comment