Message-ID: <20210728114548.GA293265@lothringen>
Date:   Wed, 28 Jul 2021 13:45:48 +0200
From:   Frederic Weisbecker <frederic@...nel.org>
To:     Marcelo Tosatti <mtosatti@...hat.com>
Cc:     nsaenzju@...hat.com, linux-kernel@...r.kernel.org,
        Nitesh Lal <nilal@...hat.com>,
        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 06:37:07AM -0300, Marcelo Tosatti 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).
> 
> The bitmask would include ISOL_F_QUIESCE_ON_URET, so:
> 
> - bitmask = ISOL_F_QUIESCE_ON_URET;
> - prctl(PR_ISOL_SET, ISOL_FEATURES, bitmask, 0, 0) enables the features in
> the bitmask.

But does it quiesce once or for every further uret?

> 
> - 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.

I don't get the difference between ISOL_FEATURES and PR_ISOL_QUIESCE_CFG.

> 
> 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).
> OK, will better document that.

Yes please, I'm completely confused :o)

> 
> > 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.
> 
> 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).

And I also don't see how such modes differ from configuration of individual
features on the bitmap.

> > > - 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).

So PR_ISOL_ENABLE is a way to perform action when some sort of kernel entry
happens. Then we take actions when that happens (signal, warn, etc...).

I guess we'll need to define what kind of kernel entry, and what kind of
response need to happen. Ok that's a whole issue of its own that we'll need
to handle seperately.

Thanks.

Powered by blists - more mailing lists