Re: [PATCH] Linux Kernel Markers 0.5 for Linux 2.6.17 (with probe management)

[Ingo Molnar]

* Mathieu Desnoyers <compudj@xxxxxxxxxxxxxxxxxx> wrote:

> * Ingo Molnar (mingo@xxxxxxx) wrote:
> > 
> > * Mathieu Desnoyers <compudj@xxxxxxxxxxxxxxxxxx> wrote:
> > 
> > > > > Then you lose the ability to trace in-kernel minor page faults.
> > > > 
> > > > that's wrong, minor pagefaults go through __handle_mm_fault() just as 
> > > > much.
> > > > 
> > > 
> > > Hi Ingo,
> > > 
> > > On a 2.6.17 kernel tree :
> > 
> > > It seems like a shortcut path that will never call __handle_mm_fault. 
> > > This path is precisely used to handle vmalloc faults.
> > 
> > yes, but you said "minor fault", not "vmalloc fault".
> > 
> > minor faults are the things that happen when a task does read-after-COW 
> > or read-mmap-ed-pagecache-page, and they very much go through 
> > __handle_mm_fault().
> > 
> > vmalloc faults are extremely rare, x86-specific and they are a pure 
> > kernel-internal matter. (I'd never want to trace them, especially if it 
> > pushes tracepoints into every architecture's page fault handler. I 
> > implemented the initial version of them IIRC, but my memory fails 
> > precisely why. I think it was 4:4 related, but i'm unsure.)
> > 
> > (i now realize that above you said "in-kernel minor faults" - under that 
> > you meant vmalloc faults?)
> > 
> 
> Yes, sorry, my mistake. This kind of fault is not as infrequent as you 
> may think, as every newly allocated vmalloc region will cause vmalloc 
> faults on every processes on the system that are trying to access 
> them. I agree that it should not be a standard event people would be 
> interested in.

most of the vmalloc area that is allocated on a typical system are 
modules - and they get loaded on bootup and rarely unloaded. Even for 
other vmalloc-ed areas like netfilter, the activation of them is during 
bootup. So from that point on the number of vmalloc faults is quite low. 
(zero on most systems) If you still want to trace it i'd suggest a 
separate type of event for it.

(meanwhile i remember why i implemented vmalloc faults to begin with: 
during vmalloc() we used to have a for_each_process() over all 
kernel-pagetables of tasks to fix up their pagetables. This caused both 
high latencies and overhead back in the days when we still were frequent 
vmalloc()ers.)

	Ingo
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/