Re: [PATCH v3] drm/i915: Remove unsafe i915.enable_rc6

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Thu, 2017-11-02 at 07:47 -0700, Rodrigo Vivi wrote:
> On Thu, Nov 02, 2017 at 08:06:29AM +0000, Jani Nikula wrote:
> > On Wed, 01 Nov 2017, Rodrigo Vivi <rodrigo.vivi@xxxxxxxxx> wrote:
> > > On Wed, Nov 01, 2017 at 04:21:08PM +0000, Ben Widawsky wrote:
> > > > On 17-11-01 18:09:47, Joonas Lahtinen wrote:
> > > > > + Kimmo and Paul
> > > > > 
> > > > > On Wed, 2017-11-01 at 07:43 -0700, Ben Widawsky wrote:
> > > > > > On 17-11-01 14:07:28, Joonas Lahtinen wrote:
> > > > > > > On Mon, 2017-10-30 at 10:48 -0700, Rodrigo Vivi wrote:
> > > > > > > > On Mon, Oct 30, 2017 at 01:00:51PM +0000, David Weinehall wrote:
> > > > > > > > > On Fri, Oct 27, 2017 at 01:57:09PM -0700, Daniele Ceraolo Spurio wrote:
> > > > > > > > > > 
> > > > > > > > > > 
> > > > > > > > > > On 26/10/17 03:32, Chris Wilson wrote:
> > > > > > > > > > > It has been many years since the last confirmed sighting (and fix) of an
> > > > > > > > > > > RC6 related bug (usually a system hang). Remove the parameter to stop
> > > > > > > > > > > users from setting dangerous values, as they often set it during triage
> > > > > > > > > > > and end up disabling the entire runtime pm instead (the option is not a
> > > > > > > > > > > fine scalpel!).
> > > > > > > > > > > 
> > > > > > > > > > > Furthermore, it allows users to set known dangerous values which were
> > > > > > > > > > > intended for testing and not for production use. For testing, we can
> > > > > > > > > > > always patch in the required setting without having to expose ourselves
> > > > > > > > > > > to random abuse.
> > > > > > > > > > > 
> > > > > > > > > > > v2: Fixup NEEDS_WaRsDisableCoarsePowerGating fumble, and document the
> > > > > > > > > > > lack of ilk support better.
> > > > > > > > > > > v3: Clear intel_info->rc6p if we don't support rc6 itself.
> > > > > > > > > > > 
> > > > > > > > > > > Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> > > > > > > > > > > Cc: Rodrigo Vivi <rodrigo.vivi@xxxxxxxxx>
> > > > > > > > > > > Cc: Joonas Lahtinen <joonas.lahtinen@xxxxxxxxxxxxxxx>
> > > > > > > > > > > Cc: Jani Nikula <jani.nikula@xxxxxxxxx>
> > > > > > > > > > > Cc: Imre Deak <imre.deak@xxxxxxxxx>
> > > > > > > > > > > Cc: Daniel Vetter <daniel.vetter@xxxxxxxx>
> > > > > > > > > > > Acked-by: Daniel Vetter <daniel.vetter@xxxxxxxx>
> > > > > > > > > > > ---
> > > > > > > > > > 
> > > > > > > > > > I think that for execution/debug on early silicon we might still want the
> > > > > > > > > > ability to turn features like RC6 off. Maybe we can add a debug kconfig to
> > > > > > > > > > force info->has_rc6 = 0? Not a blocker to this patch but worth considering
> > > > > > > > > > IMO.
> > > > > > > > > 
> > > > > > > > > Most of the BIOSes I've seen on our RVPs have had an option to disable
> > > > > > > > > RC6.
> > > > > > > > 
> > > > > > > > BIOS option don't block our code to run and set some MMIOs.
> > > > > > > > Not sure how the GPU will behave on such cases.
> > > > > > > > 
> > > > > > > > I like the idea of removing some and keeping the parameters clean.
> > > > > > > > But there are few ones like RC6 and disable_power_wells that are very
> > > > > > > > useful on platform enabling and also when assisting others to debug issues.
> > > > > > > > 
> > > > > > > > For instance right now that we fixed RC6 on CNL someone told that
> > > > > > > > he believe seeing more hangs, so I immediately asked to boot with
> > > > > > > > i915.enable_rc6=0 to double check. It is easier and straighforward
> > > > > > > > to direct them to the unsafe param than to ask them to compile the code
> > > > > > > > with different options or to use some BIOS options that we are not sure.
> > > > > > > > 
> > > > > > > > Also on bug triage some options like this are helpful.
> > > > > > > > 
> > > > > > > > Also BIOS and compile are saved flags. So if you need to do a quick test
> > > > > > > > you have to save it, and then unsave later. Parameters are very convinient
> > > > > > > > for 1 boot only check.
> > > > > > > 
> > > > > > > It's convenient for sure, but the unsafe module parameters seems to be
> > > > > > > finding their way into way too many HOWTOs, and from there to some
> > > > > > > "productized" use-cases. Chris states that setting .enable_rc6=0 to
> > > > > > > solving an issue on publicly shipping products has been some years ago,
> > > > > > > so I don't see a need for carrying this.
> > > > > > > 
> > > > > > > We shouldn't allow the convenience of not having to change one line and
> > > > > > > recompile kernel during development to affect the end-users who are
> > > > > > > Googling how to get the best performance out of their hardware (I could
> > > > > > > mention some distro here).
> > > > > > > 
> > > > > > > This seems the like the best option as I don't think introducing kernel
> > > > > > > parameters that only exists on debug builds would be too convenient
> > > > > > > either. It'd maybe just add more confusion.
> > > > > > > 
> > > > > > > Regards, Joonas
> > > > > > 
> > > > > > I believe the ability to disable RC6 is valuable not just for debugging
> > > > > > purposes. Folks with very latency sensitive workloads are often willing to
> > > > > > forego power savings. The real problem I see is that we don't test without rc6
> > > > > > in our setup, which indeed makes it unsafe. As such, I see the other option here
> > > > > > going back to the ability to toggle rc6 after load (either module parameter, or
> > > > > > make it sysfs), and actually run some subset of our workloads with RC6. I
> > > > > > suspect people will poop on that suggestion, but I figured I'd mention.
> > > > > 
> > > > > I agree there, but by my understanding there's really no ask to support
> > > > > the feature in upstream. And the original motive from Chris to drop the
> > > > > feature is that it's unsafe as it currently is.
> > > > > 
> > > > > So unless we've got the resources to bring it back from the unsafe
> > > > > zone, I think we should drop it like this patch proposes.
> > > > > 
> > > > > Regards, Joonas
> > > > 
> > > > Yep, I agree. One other option would be to move i915_forcewake_user to sysfs and
> > > > let them use that.
> > > 
> > > Well, I won't try to block that. I just put my 2 cents that I believe it is a very
> > > useful parameter.
> > > 
> > > It wasn't that long ago the last time that we needed the flag to allow
> > > end users to have a functional machine: https://plus.google.com/+JonMasters/posts/BqWLEjenLKv.
> > > 
> > > or to debug a related issue:
> > > https://bugzilla.redhat.com/show_bug.cgi?id=1440988
> > > https://bugzilla.kernel.org/show_bug.cgi?id=116431
> > > 
> > > Although date on few seems over than 1 year. We need to consider that
> > > that was our latest new gpu... gen9.
> > > 
> > > If products are recommending the use of enable_rc6=0 I can see they
> > > adding the patch to disable that. Effect is the same and our convenience is gone.
> > > 
> > > But again, just my view here. Not a nack ;)
> > 
> > I suppose the compromise would be to make it a boolean module parameter
> > to only allow disabling rc6 on platforms where it's enabled by default,
> > but not letting you enable rc6 where it's disabled by default. I.e. only
> > support i915.enable_rc6=0 to be passed by the user.
> 
> +1. I like this approach.

Umm, it still doesn't resolve the issue that it's not being tested.

I try to be super clear; until we have resources to support that
specific code path, I'd much prefer not to have an easy kernel
parameter to set it.

Regards, Joonas
-- 
Joonas Lahtinen
Open Source Technology Center
Intel Corporation
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx




[Index of Archives]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]
  Powered by Linux