Re: Unified tracing buffer

Re: Unified tracing buffer

[Mathieu Desnoyers]

* Peter Zijlstra (a.p.zijlstra@chello.nl) wrote:
> On Tue, 2008-09-23 at 00:25 -0500, Tom Zanussi wrote:
> 
> > - get rid of anything having to do with padding, nobody needs it and its
> > only affect has been to horribly distort and complicate a lot of the
> > code
> > - get rid of sub-buffers, they just cause confusion
> > - get rid of mmap, nobody uses it
> > - no sub-buffers and no mmap support means we can get rid of most of the
> > callbacks, and a lot of API confusion along with them
> > - add relay flags - they probably should have been used from the
> > beginning and options made explicit instead of being shoehorned into the
> > callback functions.
> 
>  - get rid of the vmap buffers as they cause tlb pressure and eat up
> precious vspace on 32 bit platforms.
> 

Although I agree on the basic idea, namely to use a sane amount of TLB
entries for tracing, I disagree on the way proposed to reach this goal.
Such memory management concerns belong to the mm field and should not be
done "oh so cleverly" by a buffer management infrastructure in the back
of the kernel memory management infrastructure.

I think we should instead try to figure out what is currently missing in
the kernel vmap mechanism (probably the ability to vmap from large 4MB
pages after boot), and fix _that_ instead (if possible), which would not
only benefit to tracing, but also to module support.

Also, I would like to keep a contiguous address mapping within buffers
so we could keep the buffer read/write code as simple as possible,
leveraging the existing CPU MM unit.

I added Christoph Lameter to the CC list, he always comes with clever
ideas. :)

Mathieu



-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: Unified tracing buffer

[Christoph Lameter]

Mathieu Desnoyers wrote:
> 
> I think we should instead try to figure out what is currently missing in
> the kernel vmap mechanism (probably the ability to vmap from large 4MB
> pages after boot), and fix _that_ instead (if possible), which would not
> only benefit to tracing, but also to module support.

With some custom code one can vmap 2MB pages on x86. See the VMEMMAP support
in the x86 arch. The code in mm/sparse-vmemmap.c could be abstracted for a
general 2MB mapping API to reduce TLB pressure for the buffers. If there are
concerns about fragmentation then one could fallback to 4kb TLBs. See the
virtualizable compound page patchset which does something similar.

> I added Christoph Lameter to the CC list, he always comes with clever
> ideas. :)

Oh mostly we are just recycling the old ideas.

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: Unified tracing buffer

[Linus Torvalds]

On Tue, 23 Sep 2008, Christoph Lameter wrote:

> Mathieu Desnoyers wrote:
> > 
> > I think we should instead try to figure out what is currently missing in
> > the kernel vmap mechanism (probably the ability to vmap from large 4MB
> > pages after boot), and fix _that_ instead (if possible), which would not
> > only benefit to tracing, but also to module support.

No. Don't go there. Piece of absolute shit.

The problem with VMAP is that it's _limited_. We don't have reasonable 
virtual address space holes for x86-32.

The other is that physically contiguos buffers are hard to come by. 
Certainly not an acceptable solution.

The third is that if you have multiple buffers, you need to look them up 
in software anyway, so the whole notion of mis-using the TLB to avoid a 
software lookup is TOTAL CRAP.

Don't do virtual mapping. IT IS BROKEN. IT IS A TOTAL AND UTTER PIECE OF 
SHIT.

I will absolutely not take any general-purpse tracing code if I'm aware of 
it mis-using the TLB to play games.

		Linus

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>