[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-devel] [RFC PATCH v2 2/3] utils: Add cpuinfo helper to fetch /
Re: [Qemu-devel] [RFC PATCH v2 2/3] utils: Add cpuinfo helper to fetch /proc/cpuinfo
Tue, 10 May 2016 14:06:35 +0100
Cc'ing Ramana to get some input from the toolchain side.
On Tue, May 10, 2016 at 11:24:04AM +0100, Will Deacon wrote:
> On Mon, May 09, 2016 at 01:44:11PM +0000, Catalin Marinas wrote:
> > On Mon, May 09, 2016 at 12:21:08PM +0100, Peter Maydell wrote:
> > > On 9 May 2016 at 11:59, Suzuki K Poulose <address@hidden> wrote:
> > > > Well, we have been waiting for a use case, like this, before we merge
> > > > the series.
> > >
> > > This isn't a great strategy for moving people away from things
> > > you'd like them to avoid like parsing /proc/cpuinfo, because typically
> > > userspace app writers are not very interested in coding to facilities
> > > which don't exist yet, and will prefer to make do with what's actually
> > > present in the kernel today... You need to provide the improved API,
> > > and then it needs to get out into kernel versions in distros and
> > > otherwise, and only then are you likely to get app developers who
> > > will start to say "this is useful".
> > The problem is that the way kernel people think the API may be improved
> > does not always match the use-cases required by app writers. One example
> > here is exposing MIDR via MRS emulation, we know there are problems with
> > big.LITTLE and the only clear answer I got so far is that we ignore such
> > configurations. We don't even have a way to tell user space that this is
> > a heterogeneous CPU configuration, unless we add another HWCAP bit
> > specifically for this (or the opposite: HWCAP_HOMOGENEOUS_CPUS).
> Personally, I think we should expose big.LITTLE as-is to userspace. That
> is, if you execute an mrs instruction you'll get whichever core the
> emulation happens to run on. This might even be useful to things like
> pinned threadpools w/ userspace schedulers sitting on top.
That's the point I try to make. We "think" there may be use-cases but
there are no concrete examples yet. IIRC, the only request for mrs
handling came from the tools guys for the ifunc support. However, they
don't seem to have a solution for big.LITTLE either and they may simply
ignore this feature. OTOH, we have to maintain it in the kernel on the
long run because it became ABI (that said, I would be fine if this was
complemented by another HWCAP bit for heterogeneous systems).
The CPU feature registers wouldn't be affected by the big.LITTLE
configurations as we provide a sanitised version anyway. But, again, do
we have concrete use-cases?
> > That said, I'm perfectly fine with exposing:
> > /sys/devices/system/cpu/cpu$ID/identification/
> > \- midr
> > \- revidr
> > I had the wrong impression that we already merged this part but Suzuki
> > just pointed out to me that it's not.
> Yes, there are use-cases for this interface as well. I don't think it's
> a choice between one or the other and I firmly believe we need both (the
> sysfs and mrs code).
At least for this one we have a clear use-case: JVM and errata
> > I think our 4.7-rc1 tree is pretty much frozen to new features now,
> > though the sysfs patch is relatively small (I'll let Will comment):
> The merge window opens in less than a week, so it's fixes only atm.
We have more time to debate ;)