Re: [RFC] drm/exynos: g2d: fix runtime PM

[Tobias Jakobi <tjakobi@math.uni-bielefeld.de>]

Tobias Jakobi wrote:
> The commit b05984e21a7e000bf5074ace00d7a574944b2c16 broke
> operation of the G2D. After this commit the following
> happens.
> - exynos_g2d_exec_ioctl() prepares a runqueue node and
>   calls g2d_exec_runqueue()
> - g2d_exec_runqueue() calls g2d_dma_start() which gets
>   runtime PM sync
> - runtime PM calls g2d_runtime_resume()
> - g2d_runtime_resume() calls g2d_exec_runqueue()
> 
> Luckily for us this doesn't really loop, but creates a
> mutex lockup, which the kernel even predicts.
> 
> Anyway, we fix this by reintroducing dedicated sleep
> operations again, and only letting runtime PM control
> the gate clock.
> 
> This is not enough to fix the issue though.
> - We switch runtime PM to autosuspend. Currently clocks get
>   disabled, and then re-enabled again in the runqueue worker
>   when a node is completed and the next is started.
>   With the upcoming introduction of IOMMU runtime PM this
>   situations gets worse, since now also the IOMMU goes
>   through a disable/enable cycle before the next node is
>   started.
> - We consolidate all runtime PM management to the runqueue
>   worker.
> - We introduce g2d_wait_finish() which waits until the currently
>   processed runqueue node is finished.
>   If this takes too long, the engine is forcibly reset. This
>   is necessary to properly close the driver in case the engine
>   should hang with read/write access to some area of memory.
>   In this situation we can't properly unmap GEM/userptr
>   objects, since this might create a pagefault.
> - Sleep suspend issues a suspend of the runqueue and then calls
>   g2d_wait_finished(), which returns the engine in the idle state.
This should read 'g2d_wait_finish()'.



>   The current runqueue node gets completed, all other queued
>   nodes stay in the queue. There is no hardware state that
>   needs to be retained.
> - Sleep resume just pokes the runqueue worker, which, should
>   something be in queue, continues its work, or just exits.
> 
> Signed-off-by: Tobias Jakobi <tjakobi@math.uni-bielefeld.de>

<snip>

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel