Re: [PATCH v3] mm: simplify find_vma_prev

[KAMEZAWA Hiroyuki]

On Mon, 12 Dec 2011 10:31:57 -0500
KOSAKI Motohiro <kosaki.motohiro@xxxxxxxxx> wrote:

> (12/12/11 4:27 AM), KAMEZAWA Hiroyuki wrote:
> > On Mon, 12 Dec 2011 09:49:30 +0900
> > KAMEZAWA Hiroyuki<kamezawa.hiroyu@xxxxxxxxxxxxxx>  wrote:
> >
> >> On Fri,  9 Dec 2011 17:48:40 -0500
> >> kosaki.motohiro@xxxxxxxxx wrote:
> >>
> >>> From: KOSAKI Motohiro<kosaki.motohiro@xxxxxxxxxxxxxx>
> >>>
> >>> commit 297c5eee37 (mm: make the vma list be doubly linked) added
> >>> vm_prev member into vm_area_struct. Therefore we can simplify
> >>> find_vma_prev() by using it. Also, this change help to improve
> >>> page fault performance because it has strong locality of reference.
> >>>
> >>> Signed-off-by: KOSAKI Motohiro<kosaki.motohiro@xxxxxxxxxxxxxx>
> >>
> >> Reviewed-by: KAMEZAWA Hiroyuki<kamezawa.hiroyu@xxxxxxxxxxxxxx>
> >>
> >
> > Hmm, your work remind me of a patch I tried in past.
> > Here is a refleshed one...how do you think ?
> >
> > ==
> >  From c0261936fc01322d06425731d33f38b2021e8067 Mon Sep 17 00:00:00 2001
> > From: KAMEZAWA Hiroyuki<kamezawa.hiroyu@xxxxxxxxxxxxxx>
> > Date: Mon, 12 Dec 2011 18:31:19 +0900
> > Subject: [PATCH] per thread vma cache.
> >
> > This is a toy patch. How do you think ?
> >
> > This is a patch for per-thread mmap_cache without heavy atomic ops.
> >
> > I'm sure overhead of find_vma() is pretty small in usual application
> > and this will not show good improvement. But I think, if we need
> > to have cache of vma, it should be per thread rather than per mm.
> 
> Agreed. per-thread is better.
> 
> 
> > This patch adds thread->mmap_cache, a pointer for vm_area_struct
> > and update it appropriately. Because we have no refcnt on vm_area_struct,
> > thread->mmap_cache may be a stale pointer. This patch detects stale
> > pointer by checking
> >
> >      - thread->mmap_cache is one of SLABs in vm_area_cachep.
> >      - thread->mmap_cache->vm_mm == mm.
> >
> > vma->vm_mm will be cleared before kmem_cache_free() by this patch.
> 
> Do you mean the cache can make mishit with unrelated vma when freed vma 
> was reused?

yes.

> If so, it is most tricky part of this patch, I strongly hope you write
> a comment more.
> 

Sure.
-Kame




--
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/