Re: [PATCH 0/4] riscv: Allow userspace to directly access perf counters

[Alexandre Ghiti]

Hi Ian,

On Tue, Apr 18, 2023 at 10:30 PM Atish Patra <atishp@xxxxxxxxxxxxxx> wrote:
>
> On Tue, Apr 18, 2023 at 11:46 PM Ian Rogers <irogers@xxxxxxxxxx> wrote:
> >
> > On Tue, Apr 18, 2023 at 9:43 AM Atish Patra <atishp@xxxxxxxxxxxxxx> wrote:
> > >
> > > On Fri, Apr 14, 2023 at 2:40 AM David Laight <David.Laight@xxxxxxxxxx> wrote:
> > > >
> > > > From: Atish Patra
> > > > > Sent: 13 April 2023 20:18
> > > > >
> > > > > On Thu, Apr 13, 2023 at 9:47 PM Alexandre Ghiti <alexghiti@xxxxxxxxxxxx> wrote:
> > > > > >
> > > > > > riscv used to allow direct access to cycle/time/instret counters,
> > > > > > bypassing the perf framework, this patchset intends to allow the user to
> > > > > > mmap any counter when accessed through perf. But we can't break the
> > > > > > existing behaviour so we introduce a sysctl perf_user_access like arm64
> > > > > > does, which defaults to the legacy mode described above.
> > > > > >
> > > > >
> > > > > It would be good provide additional direction for user space packages:
> > > > >
> > > > > The legacy behavior is supported for now in order to avoid breaking
> > > > > existing software.
> > > > > However, reading counters directly without perf interaction may
> > > > > provide incorrect values which
> > > > > the userspace software must avoid. We are hoping that the user space
> > > > > packages which
> > > > > read the cycle/instret directly, will move to the proper interface
> > > > > eventually if they actually need it.
> > > > > Most of the users are supposed to read "time" instead of "cycle" if
> > > > > they intend to read timestamps.
> > > >
> > > > If you are trying to measure the performance of short code
> > > > fragments then you need pretty much raw access directly to
> > > > the cycle/clock count register.
> > > >
> > > > I've done this on x86 to compare the actual cycle times
> > > > of different implementations of the IP checksum loop
> > > > (and compare them to the theoretical limit).
> > > > The perf framework just added far too much latency,
> > > > only directly reading the cpu registers gave anything
> > > > like reliable (and consistent) answers.
> > > >
> > >
> > > This series allows direct access to the counters once configured
> > > through the perf.
> > > Earlier the cycle/instret counters are directly exposed to the
> > > userspace without kernel/perf frameworking knowing
> > > when/which user space application is reading it. That has security implications.
> > >
> > > With this series applied, the user space application just needs to
> > > configure the event (cycle/instret) through perf syscall.
> > > Once configured, the userspace application can find out the counter
> > > information from the mmap & directly
> > > read the counter. There is no latency while reading the counters.
> > >
> > > This mechanism allows stop/clear the counters when the requesting task
> > > is not running. It also takes care of context switching
> > > which may result in invalid values as you mentioned below. This is
> > > nothing new and all other arch (x86, ARM64) allow user space
> > > counter read through the same mechanism.
> > >
> > > Here is the relevant upstream discussion:
> > > https://lore.kernel.org/lkml/Y7wLa7I2hlz3rKw%2F@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/T/
> > >
> > > ARM64:
> > > https://docs.kernel.org/arm64/perf.html?highlight=perf_user_access#perf-userspace-pmu-hardware-counter-access
> > >
> > > example usage in x86:
> > > https://github.com/andikleen/pmu-tools/blob/master/jevents/rdpmc.c
> >
> > The canonical implementation of this should be:
> > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/lib/perf/mmap.c#n400
>
> Thanks for sharing the libperf implementation.
>
> > which is updated in these patches but the tests are not:
> > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/tests/mmap-basic.c#n287
> > Which appears to be an oversight. The tests display some differences
>
> Yes. It's an oversight. We should make sure that perf mmap tests pass
> for RISC-V as well.

Yes, that's an oversight, I had a local test adapted from this one but
forgot to update it afterwards, I'll do that in the next version.

Thanks for your quick feedbacks and sorry for being late,

Alex


>
>
> > between x86 and aarch64 that have assumed userspace hardware counter
> > access, and everything else that it is assumed don't.
> >
> > Thanks,
> > Ian
> >
> > > > Clearly process switches (especially cpu migrations) cause
> > > > problems, but they are obviously invalid values and can
> > > > be ignored.
> > > >
> > > > So while a lot of uses may be 'happy' with the values the
> > > > perf framework gives, sometimes you do need to directly
> > > > read the relevant registers.
> > > >
> > > >         David
> > > >
> > > > -
> > > > Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> > > > Registration No: 1397386 (Wales)
> > >
> > >
> > >
> > > --
> > > Regards,
> > > Atish
>
>
>
> --
> Regards,
> Atish