On 12/15/22 10:07 PM, Rodrigo Vivi wrote:
On Wed, Dec 14, 2022 at 11:37:19PM +0100, Andi Shyti wrote:
Hi Rodrigo,
On Tue, Dec 13, 2022 at 01:18:48PM +0000, Vivi, Rodrigo wrote:
On Tue, 2022-12-13 at 00:08 +0100, Andi Shyti wrote:
Hi Rodrigo,
On Mon, Dec 12, 2022 at 11:55:10AM -0500, Rodrigo Vivi wrote:
On Mon, Dec 12, 2022 at 05:13:38PM +0100, Andi Shyti wrote:
From: Chris Wilson chris@chris-wilson.co.uk
After applying an engine reset, on some platforms like Jasperlake, we occasionally detect that the engine state is not cleared until shortly after the resume. As we try to resume the engine with volatile internal state, the first request fails with a spurious CS event (it looks like it reports a lite-restore to the hung context, instead of the expected idle->active context switch).
Signed-off-by: Chris Wilson hris@chris-wilson.co.uk
There's a typo in the signature email I'm afraid...
oh yes, I forgot the 'C' :)
you forgot? who signed it off?
Chris, but as I was copy/pasting SoB's I might have unintentionally removed the 'c'.
Other than that, have we checked the possibility of using the driver-initiated-flr bit instead of this second loop? That should be the right way to guarantee everything is cleared on gen11+...
maybe I am misinterpreting it, but is FLR the same as resetting hardware domains individually?
No, it is bigger than that... almost the PCI FLR with some exceptions:
https://lists.freedesktop.org/archives/intel-gfx/2022-December/313956.html
yes, exactly... I would use FLR feedback if I was performing an FLR reset. But here I'm not doing that, here I'm simply gating off some power domains. It happens that those power domains turn on and off engines making them reset.
is this issue only seeing when this reset is called from the sanitize functions at probe and resumes? Or from any kind of gt reset?
I don't remember seeing any reference link to the bug in the patch, hence I'm assuming this is happening in any kind of gt reset that ends up in this function.
FLR doesn't have anything to do here, also because if you want to reset a single engine you go through this function, instead of resetting the whole GPU with whatever is annexed.
yeap. That might be to extreme depending on the case. But all that I asked was if we were considering this option since this is the recommended way of reseting our engines nowadays.
This patch is not fixing the "reset" concept of i915, but it's fixing a missing feedback that happens in one single platform when trying to gate on/off a domain.
But it is changing the reset concept and timeouts for all the reset cases in all the platforms.
Maybe I am completely off track, but I don't see connection with FLR here.
The point is that if a reset is needed, for any reason, the recommended way for Jasperlake, and any other newer platforms, is to use the FLR rather than the engine reset. But we are using the engine reset, and now twice, rather then attempt the recommended way.
(besides FLR might not be present in all the platforms)
This issue is also not present in all the platforms and you are still increasing the loops and delay for all the platforms.
Thanks a lot for your inputs,
have we looked to the Jasperlake workarounds to see if we are missing anything there that could help us in this case instead of this extreme approach of randomly increasing timeouts and attempts for all the platforms?
Hi, Rodrigo JSL_WA (Bspec: 33451) doesn't seem to have a WA similar to this issue. (Please correct me if I can't find it.)
Andi
How am I supposed to use driver_initiated_flr() in this context?
Some drivers are not even using this gt reset anymore and going directly with the driver-initiated FLR in case that guc reset failed.
I believe we can still keep the gt reset in our case as we currently have, but instead of keep retrying it until it succeeds we probably should go to the next level and do the driver initiated FLR when the GT reset failed.
Thanks, Andi
Cc: stable@vger.kernel.org Cc: Mika Kuoppala mika.kuoppala@linux.intel.com Signed-off-by: Andi Shyti andi.shyti@linux.intel.com
drivers/gpu/drm/i915/gt/intel_reset.c | 34 ++++++++++++++++++++++----- 1 file changed, 28 insertions(+), 6 deletions(-)
diff --git a/drivers/gpu/drm/i915/gt/intel_reset.c b/drivers/gpu/drm/i915/gt/intel_reset.c index ffde89c5835a4..88dfc0c5316ff 100644 --- a/drivers/gpu/drm/i915/gt/intel_reset.c +++ b/drivers/gpu/drm/i915/gt/intel_reset.c @@ -268,6 +268,7 @@ static int ilk_do_reset(struct intel_gt *gt, intel_engine_mask_t engine_mask, static int gen6_hw_domain_reset(struct intel_gt *gt, u32 hw_domain_mask) { struct intel_uncore *uncore = gt->uncore; + int loops = 2; int err; /* @@ -275,18 +276,39 @@ static int gen6_hw_domain_reset(struct intel_gt *gt, u32 hw_domain_mask) * for fifo space for the write or forcewake the chip for * the read */ - intel_uncore_write_fw(uncore, GEN6_GDRST, hw_domain_mask); + do { + intel_uncore_write_fw(uncore, GEN6_GDRST, hw_domain_mask); - /* Wait for the device to ack the reset requests */ - err = __intel_wait_for_register_fw(uncore, - GEN6_GDRST, hw_domain_mask, 0, - 500, 0, - NULL); + /* + * Wait for the device to ack the reset requests. + * + * On some platforms, e.g. Jasperlake, we see see that the + * engine register state is not cleared until shortly after + * GDRST reports completion, causing a failure as we try + * to immediately resume while the internal state is still + * in flux. If we immediately repeat the reset, the second + * reset appears to serialise with the first, and since + * it is a no-op, the registers should retain their reset + * value. However, there is still a concern that upon + * leaving the second reset, the internal engine state + * is still in flux and not ready for resuming. + */ + err = __intel_wait_for_register_fw(uncore, GEN6_GDRST,
hw_domain_mask, 0, + 2000, 0, + NULL);
Andi, fast_timeout_us is increased from 500 to 2000, and if it fails, it tries to reset it once more. How was this value of 2000 calculated?
+ } while (err == 0 && --loops); if (err) GT_TRACE(gt, "Wait for 0x%08x engines reset failed\n", hw_domain_mask);
Did GT_TRACE report an error in a situation where the problem was reported?
+ /* + * As we have observed that the engine state is still volatile + * after GDRST is acked, impose a small delay to let everything settle. + */ + udelay(50);
udelay(50) affects all platforms that can call gen6_hw_domain_reset(), is that intended?
Br,
G.G.
return err; } -- 2.38.1