Message-ID: <20210728161737.GA24635@fuller.cnet>
Date:   Wed, 28 Jul 2021 13:17:37 -0300
From:   Marcelo Tosatti <mtosatti@...hat.com>
To:     Nitesh Lal <nilal@...hat.com>
Cc:     Frederic Weisbecker <frederic@...nel.org>,
        Nicolas Saenz Julienne <nsaenzju@...hat.com>,
        linux-kernel@...r.kernel.org, Christoph Lameter <cl@...ux.com>,
        Juri Lelli <juri.lelli@...hat.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Alex Belits <abelits@...vell.com>,
        Peter Xu <peterx@...hat.com>,
        Thomas Gleixner <tglx@...utronix.de>
Subject: Re: [patch 1/4] add basic task isolation prctl interface

On Wed, Jul 28, 2021 at 10:48:25AM -0400, Nitesh Lal wrote:
> On Wed, Jul 28, 2021 at 5:56 AM Marcelo Tosatti <mtosatti@...hat.com> wrote:
> 
> > On Wed, Jul 28, 2021 at 01:45:39AM +0200, Frederic Weisbecker wrote:
> > > On Tue, Jul 27, 2021 at 11:52:09AM -0300, Marcelo Tosatti wrote:
> > > > The meaning of isolated is specified as follows:
> > > >
> > > > Isolation features
> > > > ==================
> > > >
> > > > - prctl(PR_ISOL_GET, ISOL_SUP_FEATURES, 0, 0, 0) returns the supported
> > > > features as a return value.
> > > >
> > > > - prctl(PR_ISOL_SET, ISOL_FEATURES, bitmask, 0, 0) enables the
> > features in
> > > > the bitmask.
> > > >
> > > > - prctl(PR_ISOL_GET, ISOL_FEATURES, 0, 0, 0) returns the currently
> > > > enabled features.
> > >
> > > So what are the ISOL_FEATURES here? A mode that we enter such as flush
> > > vmstat _everytime_ we resume to userpace after (and including) this
> > prctl() ?
> >
> > ISOL_FEATURES is just the "command" type (which you can get and set).
> >
> 
> So, ISOL_FEATURES is really defining when the operations are really going
> to take place for eg. on every uret?

ISOL_F_QUIESCE_ON_URET enables quiescing on userspace return.

> > The bitmask would include ISOL_F_QUIESCE_ON_URET, so:
> >
> >
> When we talk about full/complete isolation 

https://lwn.net/Articles/816298/

Nohz and task isolation section

These features reduce interruptions on the isolated CPUs, but do not
fully eliminate them; task isolation is an attempt to finish the job by
removing all interruptions. A process that enters the isolation mode
will be able to run in user space with no interference from the kernel
or other processes.

> then does that translates to
> enabling all possible features supported by something like
> ISOL_F_QUIESCE_ON_URET?

Not necessarily.

If one controls what apps execute on the system (say the system 
is completly idle), ISOL_F_QUIESCE_ON_URET with vmstat sync should
be sufficient for complete isolation (one can read events via
rt-trace-bpf.py).

> - bitmask = ISOL_F_QUIESCE_ON_URET;
> > - prctl(PR_ISOL_SET, ISOL_FEATURES, bitmask, 0, 0) enables the features in
> > the bitmask.
> >
> > - quiesce_bitmap = prctl(PR_ISOL_GET, PR_ISOL_SUP_QUIESCE_CFG, 0, 0, 0)
> >   (1)
> >
> >   (returns the supported actions to be quiesced).
> >
> > - prctl(PR_ISOL_SET, PR_ISOL_QUIESCE_CFG, quiesce_bitmask, 0, 0) _sets_
> > the actions to be quiesced (2)
> >
> > If an application does not modify "quiesce_bitmask" between
> > points (1) and (2) above, it will enable quiescing of all
> > "features" the kernel supports.
> >
> > Application can, however, modify quiesce_bitmap to its preference.
> >
> > Flushing vmstat _everytime_ you resume to userspace is enabled only
> > _after_ prctl(PR_ISOL_ENTER, 0, 0, 0, 0) is performed (which happens
> > only when isolation is fully configured with the PR_ISOL_SET calls).
> >
> 
> Will this also happen if I disable  ISOL_F_QUIESCE_VMSTAT_SYNC from the
> quiesce_bitmask?

Yes.

> > OK, will better document that.
> >
> > > If so I'd rather call that ISOL_MODE because feature is too general.
> >
> > Well, in the first patchset, there was one "mode" implemented (but
> > it was possible to implement different modes in the future).
> >
> > This would allow for example easier integration of "full task isolation"
> > patchset type of functionality, disallowing syscalls.
> >
> >
> Makes sense to go back to the usage of ISOL_MODE.
> After this change, the ISOL_FEATURES will be replaced with something like
> PR_ISOL_MODE_NORMAL/PR_MODE_ISOL?
> 
> I think we'd like to keep that, so i'll keep the previous distinct modes
> > (but allow configuration of individual features on the bitmap).
> >
> > > >
> > > > The supported features are:
> > > >
> > > > ISOL_F_QUIESCE_ON_URET: quiesce deferred actions on return to
> > userspace.
> > > > ----------------------
> > > >
> > > > Quiescing of different actions can be performed on return to userspace.
> > > >
> > > > - prctl(PR_ISOL_GET, PR_ISOL_SUP_QUIESCE_CFG, 0, 0, 0) returns
> > > > the supported actions to be quiesced.
> > > >
> > > > - prctl(PR_ISOL_SET, PR_ISOL_QUIESCE_CFG, quiesce_bitmask, 0, 0)
> > returns
> >
> >                                                         s/returns/sets/
> >
> > > > the currently supported actions to be quiesced.
> > > >
> > > > - prctl(PR_ISOL_GET, PR_ISOL_QUIESCE_CFG, 0, 0, 0) returns
> > > > the currently enabled actions to be quiesced.
> > > >
> > > > #define ISOL_F_QUIESCE_VMSTAT_SYNC      (1<<0)
> > > > #define ISOL_F_QUIESCE_NOHZ_FULL        (1<<1)
> > > > #define ISOL_F_QUIESCE_DEFER_TLB_FLUSH  (1<<2)
> > >
> > > And then PR_ISOL_QUIESCE_CFG is a oneshot operation that applies only
> > upon
> > > return to this ctrl, right? If so perhaps this should be called just
> > > ISOL_QUIESCE or ISOL_QUIESCE_ONCE or ISOL_REQ ?
> >
> > There was no one-shot operation implemented in the first patchset. What
> > application would do to achieve that is:
> >
> > 1. Configure isolation with PR_ISOL_SET (say configure mode which
> > allows system calls, and when a system call happens, flush all deferred
> > actions on return to userspace).
> >
> > 2. prctl(PR_ISOL_ENTER, 0, 0, 0, 0) (this actually enables the flushing,
> > and tags the task_struct as isolated). Here we can transfer this
> > information
> > from per-task to per-CPU data, for example, to be able to implement
> > other features such as deferred TLB flushing.
> >
> > On return from this prctl(), deferrable actions are flushed.
> >
> > 3. latency sensitive loop, with no system calls.
> >
> > 4. some event which requires system calls is noticed:
> >    prctl(PR_ISOL_EXIT, 0, 0, 0, 0)
> >    (this would untag task_struct as isolated).
> >
> > 5. perform system calls A, B, C, D (with no flushing of vmstat,
> > for example).
> >
> > 6. jmp to 2.
> >
> > So there is a problem with this logic, which is that one would like
> > certain isolation functionality to remain enabled between points 4
> > and 6 (for example, blocking CPU hotplug or other blockable activities
> > that would cause interruptions).
> >
> > One way to achieve this would be to replace PR_ISOL_ENTER/PR_ISOL_EXIT
> > with PR_ISOL_ENABLE, which accepts a bitmask:
> >
> > 1. Configure isolation with PR_ISOL_SET (say configure mode which
> > allows system calls, and when a system call happens, flush all deferred
> > actions on return to userspace).
> >
> > 2. enabled_bitmask = ISOL_F_QUIESCE_ON_URET|ISOL_F_BLOCK_INTERRUPTORS;
> >    prctl(PR_ISOL_ENABLE, enabled_bitmask, 0, 0, 0)
> >
> > On return from this prctl(), deferrable actions are flushed.
> >
> > 3. latency sensitive loop, with no system calls.
> >
> > 4. some event which requires system calls is noticed:
> >
> >    prctl(PR_ISOL_ENABLE, ISOL_F_BLOCK_INTERRUPTORS, 0, 0, 0)
> >    (this would clear ISOL_F_QUIESCE_ON_URET, so no flushing
> >     is performed on return from system calls).
> >
> 
> FWIU we will still exit before this via prctl(PR_ISOL_EXIT, 0, 0, 0, 0)?
> Because if we are still in a latency-sensitive loop then not flushing
> while returning to the userspace can cause interruptions anyways.

No, PR_ISOL_ENABLE replaces PR_ISOL_ENTER/PR_ISOL_EXIT.

> > 5. perform system calls A, B, C, D (with no flushing of vmstat).
> >
> > 6. jmp to 2.
> >
> > ...
> >
> > On exit: prctl(PR_ISOL_ENABLE, 0, 0, 0, 0)
> >
> > IOW: the one-shot operation does not allow the application
> > to inform the kernel when the latency sensitive loop has
> > begun or has ended.
> >
> > >
> > > But that's just naming debate because otherwise that prctl layout looks
> > good
> > > to me.
> > >
> > > Thanks!
> >
> > Thank you for the input!
> >
> >
> 
> -- 
> Thanks
> Nitesh

Powered by blists - more mailing lists