[lkml]   [2021]   [Jul]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [patch 1/4] add basic task isolation prctl interface
On Wed, 2021-07-28 at 10:16 -0300, Marcelo Tosatti wrote:
> > For example, let's say we introduce ISOL_F_QUIESCE_DEFER_TLB_FLUSH, this will
> > defer relatively short IPIs on isolated CPUs in exchange for a longer flush
> > whenever we enter the kernel (syscall, IRQs, NMI, etc...).
> Why the flush has to be longer when you enter the kernel?

What I had in mind was cost of rapid partial flushes (IPIs) vs full flushes on
entry, although I haven't really measured anything so the extra latency cost
might as well be zero.

> ISOL_F_QUIESCE_DEFER_TLB_FLUSH might collapse multiple IPIs
> into a single IPI, so the behaviour might be beneficial
> for "standard" types of application as well.
> > A latency sensitive
> > application might be OK with the former but not with the latter.
> Two alternatives:
> 1) The pattern above, where particular subsystems that might interrupt
> the kernel are enabled automatically if the kernel supports it.
> Pros:
> Applications which implement this only need to be changed once,
> and can benefit from new kernel features.
> Applications can disable particular features if they turn
> out to be problematic.
> Cons:
> New features might break applications.
> 2) Force applications to enable each new feature individually.
> Pros: Won't cause regressions, kernel behaviour is explicitly
> controlled by userspace.
> Cons: Apps won't benefit from new features automatically.
> ---
> It seems to me 1) is preferred. Can also add a sysfs control to
> have a "default_isolation_feature" flag, which can be changed
> by a sysadmin in case a new feature is undesired.
> Thoughts?

I'd still take option 2. Nitesh has a very good point, latency requirements are
hit or miss. What's the benefit of enabling new features on an already valid
application vs the potential regression?

That said I see value in providing means for users that want all
features/modes, but it should be an through an explicit action on their part.

Nicolás Sáenz

 \ /
  Last update: 2021-07-28 19:09    [W:0.131 / U:0.140 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site