[Intel-gfx] [PATCH 02/11] drm/i915/fence: Avoid del_timer_sync() from inside a timer
Chris Wilson
chris at chris-wilson.co.uk
Tue Sep 19 09:24:55 UTC 2017
Quoting Chris Wilson (2017-09-11 09:41:26)
> A fence may be signaled from any context, including from inside a timer.
> One example is timer_i915_sw_fence_wake() which is used to provide a
> safety-net when waiting on an external fence. If the external fence is
> not signaled within a timely fashion, we signal our fence on its behalf,
> and so we then may process subsequent fences in the chain from within
> that timer context.
>
> Given that dma_i915_sw_fence_cb() may be from inside a timer, we cannot
> then use del_timer_sync() as that requires the timer lock for itself. To
> circumvent this, while trying to keep the signal propagation as low
> latency as possible, move the completion into a worker and use a bit of
> atomic switheroo to serialise the timer-callback and the dma-callback.
>
> Testcase: igt/gem_eio/in-flight-external
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
Ping? It's ugly but it's the best I have to offer to escape the lock
inversion whilst trying to avoid adding any delay to signaling.
-Chris
> ---
> drivers/gpu/drm/i915/Kconfig | 1 +
> drivers/gpu/drm/i915/i915_sw_fence.c | 27 +++++++++++++++++++++------
> 2 files changed, 22 insertions(+), 6 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/Kconfig b/drivers/gpu/drm/i915/Kconfig
> index e9e64e8e9765..dfd95889f4b7 100644
> --- a/drivers/gpu/drm/i915/Kconfig
> +++ b/drivers/gpu/drm/i915/Kconfig
> @@ -12,6 +12,7 @@ config DRM_I915
> select DRM_PANEL
> select DRM_MIPI_DSI
> select RELAY
> + select IRQ_WORK
> # i915 depends on ACPI_VIDEO when ACPI is enabled
> # but for select to work, need to select ACPI_VIDEO's dependencies, ick
> select BACKLIGHT_LCD_SUPPORT if ACPI
> diff --git a/drivers/gpu/drm/i915/i915_sw_fence.c b/drivers/gpu/drm/i915/i915_sw_fence.c
> index f29540f922af..808ea4d5b962 100644
> --- a/drivers/gpu/drm/i915/i915_sw_fence.c
> +++ b/drivers/gpu/drm/i915/i915_sw_fence.c
> @@ -9,6 +9,7 @@
>
> #include <linux/slab.h>
> #include <linux/dma-fence.h>
> +#include <linux/irq_work.h>
> #include <linux/reservation.h>
>
> #include "i915_sw_fence.h"
> @@ -356,31 +357,44 @@ struct i915_sw_dma_fence_cb {
> struct i915_sw_fence *fence;
> struct dma_fence *dma;
> struct timer_list timer;
> + struct irq_work work;
> };
>
> static void timer_i915_sw_fence_wake(unsigned long data)
> {
> struct i915_sw_dma_fence_cb *cb = (struct i915_sw_dma_fence_cb *)data;
> + struct i915_sw_fence *fence;
> +
> + fence = xchg(&cb->fence, NULL);
> + if (!fence)
> + return;
>
> pr_warn("asynchronous wait on fence %s:%s:%x timed out\n",
> cb->dma->ops->get_driver_name(cb->dma),
> cb->dma->ops->get_timeline_name(cb->dma),
> cb->dma->seqno);
> - dma_fence_put(cb->dma);
> - cb->dma = NULL;
>
> - i915_sw_fence_complete(cb->fence);
> - cb->timer.function = NULL;
> + i915_sw_fence_complete(fence);
> }
>
> static void dma_i915_sw_fence_wake(struct dma_fence *dma,
> struct dma_fence_cb *data)
> {
> struct i915_sw_dma_fence_cb *cb = container_of(data, typeof(*cb), base);
> + struct i915_sw_fence *fence;
> +
> + fence = xchg(&cb->fence, NULL);
> + if (fence)
> + i915_sw_fence_complete(fence);
> +
> + irq_work_queue(&cb->work);
> +}
> +
> +static void irq_i915_sw_fence_work(struct irq_work *wrk)
> +{
> + struct i915_sw_dma_fence_cb *cb = container_of(wrk, typeof(*cb), work);
>
> del_timer_sync(&cb->timer);
> - if (cb->timer.function)
> - i915_sw_fence_complete(cb->fence);
> dma_fence_put(cb->dma);
>
> kfree(cb);
> @@ -414,6 +428,7 @@ int i915_sw_fence_await_dma_fence(struct i915_sw_fence *fence,
> __setup_timer(&cb->timer,
> timer_i915_sw_fence_wake, (unsigned long)cb,
> TIMER_IRQSAFE);
> + init_irq_work(&cb->work, irq_i915_sw_fence_work);
> if (timeout) {
> cb->dma = dma_fence_get(dma);
> mod_timer(&cb->timer, round_jiffies_up(jiffies + timeout));
> --
> 2.14.1
>
More information about the Intel-gfx
mailing list