Message-ID: <2049164886.596497.1579641536619.JavaMail.zimbra@efficios.com>
Date:   Tue, 21 Jan 2020 16:18:56 -0500 (EST)
From:   Mathieu Desnoyers <mathieu.desnoyers@...icios.com>
To:     Jann Horn <jannh@...gle.com>
Cc:     Peter Zijlstra <peterz@...radead.org>,
        Thomas Gleixner <tglx@...utronix.de>,
        linux-kernel <linux-kernel@...r.kernel.org>,
        Joel Fernandes <joelaf@...gle.com>,
        Ingo Molnar <mingo@...hat.com>,
        Catalin Marinas <catalin.marinas@....com>,
        Dave Watson <davejwatson@...com>,
        Will Deacon <will.deacon@....com>, shuah <shuah@...nel.org>,
        Andi Kleen <andi@...stfloor.org>,
        linux-kselftest <linux-kselftest@...r.kernel.org>,
        "H. Peter Anvin" <hpa@...or.com>, Chris Lameter <cl@...ux.com>,
        Russell King <linux@....linux.org.uk>,
        Michael Kerrisk <mtk.manpages@...il.com>,
        Paul <paulmck@...ux.vnet.ibm.com>, Paul Turner <pjt@...gle.com>,
        Boqun Feng <boqun.feng@...il.com>,
        Josh Triplett <josh@...htriplett.org>,
        rostedt <rostedt@...dmis.org>, Ben Maurer <bmaurer@...com>,
        linux-api <linux-api@...r.kernel.org>,
        Andy Lutomirski <luto@...capital.net>
Subject: Re: [RFC PATCH v1] pin_on_cpu: Introduce thread CPU pinning system
 call

----- On Jan 21, 2020, at 3:35 PM, Jann Horn jannh@...gle.com wrote:

> On Tue, Jan 21, 2020 at 8:47 PM Mathieu Desnoyers
> <mathieu.desnoyers@...icios.com> wrote:
>>
>> ----- On Jan 21, 2020, at 12:20 PM, Jann Horn jannh@...gle.com wrote:
>>
>> > On Tue, Jan 21, 2020 at 5:13 PM Mathieu Desnoyers
>> > <mathieu.desnoyers@...icios.com> wrote:
>> >> There is an important use-case which is not possible with the
>> >> "rseq" (Restartable Sequences) system call, which was left as
>> >> future work.
>> >>
>> >> That use-case is to modify user-space per-cpu data structures
>> >> belonging to specific CPUs which may be brought offline and
>> >> online again by CPU hotplug. This can be used by memory
>> >> allocators to migrate free memory pools when CPUs are brought
>> >> offline, or by ring buffer consumers to target specific per-CPU
>> >> buffers, even when CPUs are brought offline.
>> >>
>> >> A few rather complex prior attempts were made to solve this.
>> >> Those were based on in-kernel interpreters (cpu_opv, do_on_cpu).
>> >> That complexity was generally frowned upon, even by their author.
>> >>
>> >> This patch fulfills this use-case in a refreshingly simple way:
>> >> it introduces a "pin_on_cpu" system call, which allows user-space
>> >> threads to pin themselves on a specific CPU (which needs to be
>> >> present in the thread's allowed cpu mask), and then clear this
>> >> pinned state.
>> > [...]
>> >> For instance, this allows implementing this userspace library API
>> >> for incrementing a per-cpu counter for a specific cpu number
>> >> received as parameter:
>> >>
>> >> static inline __attribute__((always_inline))
>> >> int percpu_addv(intptr_t *v, intptr_t count, int cpu)
>> >> {
>> >>         int ret;
>> >>
>> >>         ret = rseq_addv(v, count, cpu);
>> >> check:
>> >>         if (rseq_unlikely(ret)) {
>> >>                 pin_on_cpu_set(cpu);
>> >>                 ret = rseq_addv(v, count, percpu_current_cpu());
>> >>                 pin_on_cpu_clear();
>> >>                 goto check;
>> >>         }
>> >>         return 0;
>> >> }
>> >
>> > What does userspace have to do if the set of allowed CPUs switches all
>> > the time? For example, on Android, if you first open Chrome and then
>> > look at its allowed CPUs, Chrome is allowed to use all CPU cores
>> > because it's running in the foreground:
>> >
>> > walleye:/ # ps -AZ | grep 'android.chrome$'
>> > u:r:untrusted_app:s0:c145,c256,c512,c768 u0_a145 7845 805 1474472
>> > 197868 SyS_epoll_wait f09c0194 S com.android.chrome
>> > walleye:/ # grep cpuset /proc/7845/cgroup; grep Cpus_allowed_list
>> > /proc/7845/status
>> > 3:cpuset:/top-app
>> > Cpus_allowed_list: 0-7
>> >
>> > But if you then switch to the home screen, the application is moved
>> > into a different cgroup, and is restricted to two CPU cores:
>> >
>> > walleye:/ # grep cpuset /proc/7845/cgroup; grep Cpus_allowed_list
>> > /proc/7845/status
>> > 3:cpuset:/background
>> > Cpus_allowed_list: 0-1
>>
>> Then at that point, pin_on_cpu() would only be allowed to pin on
>> CPUs 0 and 1.
> 
> Which means that you can't actually reliably use pin_on_cpu_set() to
> manipulate percpu data structures since you have to call it with the
> assumption that it might randomly fail at any time, right?

Only if the cpu affinity of the thread is being changed concurrently
by another thread which is a possibility in some applications, indeed.

> And then
> userspace needs to code a fallback path that somehow halts all the
> threads with thread-directed signals or something?

The example use of pin_on_cpu() did not include handling of the return
value in that case (-1, errno=EINVAL) for conciseness. But yes, the
application would have to handle this.

It's not so different from error handling which is required when using
sched_setaffinity(), which can fail with -1, errno=EINVAL in the following
scenario:

       EINVAL The  affinity bit mask mask contains no processors that are cur‐
              rently physically on the system  and  permitted  to  the  thread
              according  to  any  restrictions  that  may  be  imposed  by the
              "cpuset" mechanism described in cpuset(7).

> Especially if the task trying to use pin_on_cpu_set() isn't allowed to
> pin to the target CPU, but all the other tasks using the shared data
> structure are allowed to do that. Or if the CPU you want to pin to is
> always removed from your affinity mask immediately before
> pin_on_cpu_set() and added back immediately afterwards.

I am tempted to state that using pin_on_cpu() targeting a disallowed cpu
should be considered a programming error and handled accordingly by the
application.

> 
>> > At the same time, I also wonder whether it is a good idea to allow
>> > userspace to stay active on a CPU even after the task has been told to
>> > move to another CPU core - that's probably not exactly a big deal, but
>> > seems suboptimal to me.
>>
>> Do you mean the following scenario for a given task ?
>>
>> 1) set affinity to CPU 0,1,2
>> 2) pin on CPU 2
>> 3) set affinity to CPU 0,1
>>
>> In the patch I propose here, step (3) is forbidden by this added check in
>> __set_cpus_allowed_ptr, which is used by sched_setaffinity(2):
>>
>> +       /* Prevent removing the currently pinned CPU from the allowed cpu mask.
>> */
>> +       if (is_pinned_task(p) && !cpumask_test_cpu(p->pinned_cpu, new_mask)) {
>> +               ret = -EINVAL;
>> +               goto out;
>> +       }
> 
> How does that interact with things like cgroups? What happens when
> root tries to move a process from the cpuset:/top-app cgroup to the
> cpuset:/background cgroup? Does that also just throw an error code?

Looking at kernel/cgroup/cpuset.c use of set_cpus_allowed_ptr(), either
it:

A) silently ignores the error (3 instances),
or
B) triggers a WARN_ON_ONCE() (1 instance).

So yeah, that cgroup code seems rather optimistic that this operation
should always succeed (?!?)

I can say up front I am not entirely sure what's the best way to proceed
here. Either sched_setaffinity and cgroup can fail if trying to remove a
CPU from affinity mask while the task is pinned, or we let them succeed
and leave the task running on the pinned CPU anyway, or any better idea ?

> 
>> > I'm wondering whether it might be possible to rework this mechanism
>> > such that, instead of moving the current task onto a target CPU, it
>> > prevents all *other* threads of the current process from running on
>> > that CPU (either entirely or in user mode). That might be the easiest
>> > way to take care of issues like CPU hotplugging and changing cpusets
>> > all at once? The only potential issue I see with that approach would
>> > be that you wouldn't be able to use it for inter-process
>> > communication; and I have no idea whether this would be good or bad
>> > performance-wise.
>>
>> Firstly, inter-process communication over shared memory is one of my use-cases
>> (for lttng-ust's ring buffer).
>>
>> I'm not convinced that applying constraints on all other threads belonging to
>> the current process would be easier or faster than migrating the current thread
>> over to the target CPU. I'm unsure how those additional constraints would
>> fit with other threads already having their own cpu affinity masks (which
>> could generate an empty cpumask by adding an extra constraint).
> 
> Hm - is an empty cpumask a problem here? If one task is in the middle
> of a critical section for performing maintenance on a percpu data
> structure, isn't it a nice design property to exclude concurrent
> access from other tasks to that data structure automatically (by
> preventing those tasks by running on that CPU)? That way the
> (presumably rarely-executed) update path doesn't have to be
> rseq-reentrancy-safe.

Given we already have rseq, simply using it to protect against other
threads trying to touch the same per-cpu data seems rather more lightweight
than to try to exclude all other threads from that CPU for a possibly
unbounded amount of time.

Allowing a completely empty cpumask could effectively allow those
critical sections to prevent execution of possibly higher priority
threads on the system, which ends up being the definition of a priority
inversion, which I'd like to avoid.

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

Powered by blists - more mailing lists