On Tue, 08 Sep 2026 09:29:44 +0100,
Raviteja Laggyshetty [off-list ref] wrote:
On 9/8/2026 1:04 PM, Abel Vesa wrote:
quoted
On 26-09-07 22:15:09, Marc Zyngier wrote:
quoted
On Mon, 07 Sep 2026 21:21:44 +0100,
Abel Vesa [off-list ref] wrote:
quoted
On 26-09-07 18:17:29, Marc Zyngier wrote:
quoted
Recent addition of QoS configuration to the X1 variants resulted in
systems that spontaneously reset at boot time or under load.
Given that this is both very unpleasant and hard to debug, give the
affected user the option to survive such event with a module parameter.
Passing "icc_rpmh.enable_qos=0" to the kernel command-line will skip the
offending QoS setup.
Lets not start doing this, please.
Can we figure out which platform deosn't work and then maybe we can
add a quirk in DT per platform?
We have at least 3 machines that are dysfunctional since [1] was
merged (mini-x, sp12 and devkit). Probably some more, but I'm not
going to be the one doing a survey.
What really should happen is [1] to be reversed until QC figures out a
set of parameters that actually works, instead of making this platform
even more unreliable than it already is.
I'd rather have the patch reverted than adding a module parameter which
users might not know about and distros will probably never addopt.
quoted
Just like [2] should have been reverted.
Until then, we'll have to paper over this stuff.
But a module parameter complicates things rather than just papering over.
I suggest we get both reverted if they break existing platforms.
I will share a patch to add interconnect support for Purwa.
Purwa is similar to Hamoa, but there are few port additions and deletions with
respect to PCIe, and trying to program QoS for them can result in crash.
This is unfortunately not solely about Purwa, as repeatedly pointed
out in this thread and the original one.
Hamoa (at least the machine I use for a lot of my upstream work) has
become completely unreliable since the QoS patch was merged, and
randomly resets under IO load. Not applying the QoS configuration
restores stability.
My conclusion is that merging this stuff was premature and has
evidently not seen enough testing and verification.
M.
--
Without deviation from the norm, progress is not possible.