active_mm and mm

active_mm and mm

[Hai Huang]

In struct task_struct, what's the difference between active_mm and mm?  I
vaguely remembers it's used for reducing cache overhead during context
switch, is this right or I'm totally off.  Thanks

-
Hai

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/

Re: active_mm and mm

[Stephen C. Tweedie]

Hi,

On Mon, Aug 19, 2002 at 10:09:50PM -0400, Hai Huang wrote:
> In struct task_struct, what's the difference between active_mm and mm?  I
> vaguely remembers it's used for reducing cache overhead during context
> switch, is this right

Yep.  Many context switches don't require us to switch to the mm of
the newly running process.  All processes share exactly the same
kernel address space, so as long as we are only accessing kernel
memory and not per-process memory, we don't need to do the mm switch.

So, for operations such as waiting on an IO event, a process might get
woken up, check some kernel space data structures, and go back to
sleep, all in side a system call and never touching user space.  It's
a waste to switch to the process's mm just for that --- we'd end up
throwing out the tlb cache of the old process for nothing.

So, Linux has a "LAZY_TLB" mode which tasks such as the idle task
(which never touch user space) all have set, and which tasks can enter
if they are spinning in kernel space for a while.  When we switch to a
LAZY_TLB task, we don't get a new mm, so the new task's active_mm
is set to whatever the old task's active_mm was.  For non-LAZY_TLB
running tasks, active_mm and mm should be the same.

--Stephen
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/

Re: active_mm and mm

[Hai Huang]

Ok, I see why we're differentiating between mm and active_mm, but is this
actually giving us a lot of benefits considering the number of context switches
that would actually take advantage of this feature is probably small
(well, it depends on the workload).  Also, is the tlb flush operation that
expensive?

-
Hai

> Hi,
>
> On Mon, Aug 19, 2002 at 10:09:50PM -0400, Hai Huang wrote:
> > In struct task_struct, what's the difference between active_mm and mm?  I
> > vaguely remembers it's used for reducing cache overhead during context
> > switch, is this right
>
> Yep.  Many context switches don't require us to switch to the mm of
> the newly running process.  All processes share exactly the same
> kernel address space, so as long as we are only accessing kernel
> memory and not per-process memory, we don't need to do the mm switch.
>
> So, for operations such as waiting on an IO event, a process might get
> woken up, check some kernel space data structures, and go back to
> sleep, all in side a system call and never touching user space.  It's
> a waste to switch to the process's mm just for that --- we'd end up
> throwing out the tlb cache of the old process for nothing.
>
> So, Linux has a "LAZY_TLB" mode which tasks such as the idle task
> (which never touch user space) all have set, and which tasks can enter
> if they are spinning in kernel space for a while.  When we switch to a
> LAZY_TLB task, we don't get a new mm, so the new task's active_mm
> is set to whatever the old task's active_mm was.  For non-LAZY_TLB
> running tasks, active_mm and mm should be the same.
>
> --Stephen
>

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/

Re: active_mm and mm

[Stephen C. Tweedie]

Hi,

On Tue, Aug 20, 2002 at 10:55:04AM -0400, Hai Huang wrote:
> Ok, I see why we're differentiating between mm and active_mm, but is this
> actually giving us a lot of benefits considering the number of context switches
> that would actually take advantage of this feature is probably small
> (well, it depends on the workload).

It's actually enormous.  There are a lot of kernel daemons that do
background IO, for example.  Those are often waking up after an IO
completes, doing a tiny amount of work to submit new IO, then sleeping
again.  Even more significant in many workloads is the idle task.

> Also, is the tlb flush operation that
> expensive?

Yes.  Modern cpus are _way_ faster than main memory, and they rely
utterly on the cache architecture to keep them busy.  Doing a tlb
flush forces the CPU to go back to main memory up to 2 times for every
single address translation that follows until the tlb is full again.
That's an enormous cost, especially on rapidly-switching workloads.

--Stephen
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/