On 18/12/13 14:07, Alexander Graf wrote:
[...]
How does it encourage a vendor to properly implement their firmware if there's
a workaround?
Alex
Hi Alex,
In short, by enabling the users to create the demand. Yes, like any workaround there's potential
for abuse, but having *something* that makes it work is the difference between "I want
virtualisation"[1] and "Dear vendor, I've tried virtualisation on your chip/board and
it's great, but it tells me I need new firmware, where do I get that?"
Having the specs tell them what to do clearly isn't sufficient, so let's give
the integrators and consumers incentive to shout at them too. The sooner proper
support is commonplace and we can deprecate clock-frequency hacks altogether,
the better.
Oh, I'm all for hacks. But please don't fall under the illusion that this will push
vendors to fix their firmware. It will have the opposite effect - vendors will just point
to the workaround and say "but it works" :).
If vendors already aren't bothering to support functionality available
in their flagship hardware, workarounds hardly make that worse, and are
a win for the user. If it can drive adoption enough to get vendors to
see the value in at least fixing future products, that's only good.
Either way, this hack is basically required because you can't program CNTFRQ because it's
controlled by secure firmware, right? So the host os already needs to know about this and
probably does have a "clock-frequency" value in its device tree entry already
to know how fast its clock ticks.
In some cases, yes. In others they don't explicitly use the arch timer
at all thus have no frequency set anywhere. In the case of the board I
have on my desk, it took hacking the non-secure part of the bootloader,
writing a shim to throw away the securely-booted non-hyp cpu0 and fire
up a secondary, and hacking a timer node into the host DT to even get as
far as having an issue with kvmtool.
Couldn't we search for that host entry and automatically pass it into the guest
if it's there? That way the whole thing becomes seamless and even less of an
issue.
In theory that would be an ideal solution, yes. In practice it means
either making KVM dependent on PROC_DEVICETREE (yuck) or cooking up some
kernel interface to expose the system timer frequency to userspace
(double yuck). Not just "global solution to local problem", but "global
solution to local
problem-that-shouldn't-even-exist-and-you-want-to-go-away-as-soon-as-possible-without-leaving-a-legacy".
Besides, that would probably just reinforce the equally wrong behaviour
of putting the frequency in the host DT instead of fixing the firmware ;)
Robin.
Alex
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
