[Intel-gfx] [PATCH 3/3] drm/i915: Disable PSMI idle messaging when stopping ring

[Mika Kuoppala]

Chris Wilson <chris at chris-wilson.co.uk> writes:

> Quoting Mika Kuoppala (2019-02-28 16:01:33)
>> Hardware cannot be in a middle of idle flow messaging
>> when we pull the plug from ringbuffer. Disable idle
>> messaging before we do so to avoid potential deadlock
>> on engine initialization and reset.
>> 
>> v2: INVALID_MMIO_REG, unconditional enable (Chris)
>> v3: comment (Chris), bspec reference
>> 
>> References: bspec/11751
>> Cc: Chris Wilson <chris at chris-wilson.co.uk>
>> Acked-by: Chris Wilson <chris at chris-wilson.co.uk>
>> Signed-off-by: Mika Kuoppala <mika.kuoppala at linux.intel.com>
>> ---
>>  drivers/gpu/drm/i915/intel_engine_cs.c | 64 +++++++++++++++++++++++++-
>>  1 file changed, 63 insertions(+), 1 deletion(-)
>> 
>> diff --git a/drivers/gpu/drm/i915/intel_engine_cs.c b/drivers/gpu/drm/i915/intel_engine_cs.c
>> index 2a94b92cfcd3..10cb6b22fe92 100644
>> --- a/drivers/gpu/drm/i915/intel_engine_cs.c
>> +++ b/drivers/gpu/drm/i915/intel_engine_cs.c
>> @@ -764,7 +764,7 @@ int intel_engine_init_common(struct intel_engine_cs *engine)
>>  
>>  /**
>>   * intel_engines_cleanup_common - cleans up the engine state created by
>> - *                                the common initiailizers.
>> + *                                the common initializers.
>>   * @engine: Engine to cleanup.
>>   *
>>   * This cleans up everything created by the common helpers.
>> @@ -865,6 +865,63 @@ void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine)
>>                       _MASKED_BIT_DISABLE(STOP_RING));
>>  }
>>  
>> +static i915_reg_t get_idle_poll_reg(const struct intel_engine_cs *engine)
>> +{
>> +       if (engine->id != RCS)
>> +               return INVALID_MMIO_REG;
>> +
>> +       if (IS_GEN(engine->i915, 9))
>> +               return GEN9_RCS_FE_FSM2;
>> +
>> +       if (IS_GEN(engine->i915, 8))
>> +               return GEN6_RCS_PWR_FSM;
>> +
>> +       return INVALID_MMIO_REG;
>> +}
>> +
>> +static void disable_idle_messaging(struct intel_engine_cs *engine)
>> +{
>> +       struct drm_i915_private *dev_priv = engine->i915;
>> +       i915_reg_t poll_reg;
>> +
>> +       poll_reg = get_idle_poll_reg(engine);
>> +       if (!i915_mmio_reg_valid(poll_reg))
>> +               return;
>> +
>> +       GEM_DEBUG_WARN_ON(I915_READ_FW(GEN6_RC_SLEEP_PSMI_CONTROL) &
>> +                         GEN6_PSMI_SLEEP_MSG_DISABLE);
>> +       /*
>> +        * Hardware must not be in middle of idle flow signalling
>> +        * when the RING_CTL is zeroed. In order to prevent this
>> +        * we disable the messaging temporarily.
>> +        */
>> +       I915_WRITE_FW(GEN6_RC_SLEEP_PSMI_CONTROL,
>> +                     _MASKED_BIT_ENABLE(GEN6_PSMI_SLEEP_MSG_DISABLE));
>> +
>> +       /*
>> +        * Poll the magic bits to ensure that engine state machine
>> +        * is in valid state for stopping the ring.
>> +        */
>> +       if (__intel_wait_for_register_fw(dev_priv, poll_reg,
>> +                                        0x7f, 0x30,
>> +                                        5000, 0,
>> +                                        NULL))
>> +               DRM_DEBUG_DRIVER("psmi idle msg poll timeout\n");
>> +}
>> +
>> +static void enable_idle_messaging(struct intel_engine_cs *engine)
>> +{
>> +       struct drm_i915_private *dev_priv = engine->i915;
>> +       i915_reg_t poll_reg;
>> +
>> +       poll_reg = get_idle_poll_reg(engine);
>> +       if (!i915_mmio_reg_valid(poll_reg))
>> +               return;
>> +
>> +       I915_WRITE_FW(GEN6_RC_SLEEP_PSMI_CONTROL,
>> +                     _MASKED_BIT_DISABLE(GEN6_PSMI_SLEEP_MSG_DISABLE));
>> +}
>> +
>>  int intel_engine_stop_ring(struct intel_engine_cs *engine)
>>  {
>>         struct drm_i915_private *dev_priv = engine->i915;
>> @@ -880,9 +937,14 @@ int intel_engine_stop_ring(struct intel_engine_cs *engine)
>>         I915_WRITE_FW(RING_TAIL(base), 0);
>>         POSTING_READ_FW(RING_TAIL(base));
>>  
>> +       /* Idle messaging needs to be off during ring disable */
>
> Ok, that's ott :)

Rest assured, Captain Obvious will scrap this before merging.

>
> The previous pair are a nice why, peeling back some of the layers of the
> HW to see what's going on. This one is just repeating
> disable_idle_messing().

There is this dream I had where hardware engineers
did review patches.

But it should be close enough of 'what's going on'

Thanks,
-Mika