Re: No more bits in vm_area_struct's vm_flags.

From: KAMEZAWA Hiroyuki
Date: Sun Sep 27 2009 - 23:07:21 EST


Hi,

On Mon, 28 Sep 2009 12:45:35 +1000
Nigel Cunningham <ncunningham@xxxxxxxxxxx> wrote:
> KAMEZAWA Hiroyuki wrote:
> > On Fri, 25 Sep 2009 18:34:56 +1000
> > Nigel Cunningham <ncunningham@xxxxxxxxxxx> wrote:
> >> Hi.
> >>
> >> KAMEZAWA Hiroyuki wrote:
> >>>> I have some code in TuxOnIce that needs a bit too (explicitly mark the
> >>>> VMA as needing to be atomically copied, for GEM objects), and am not
> >>>> sure what the canonical way to proceed is. Should a new unsigned long be
> >>>> added? The difficulty I see with that is that my flag was used in
> >>>> shmem_file_setup's flags parameter (drm_gem_object_alloc), so that
> >>>> function would need an extra parameter too..
> >>> Hmm, how about adding vma->vm_flags2 ?
> >> The difficulty there is that some functions pass these flags as arguments.
> >>
> > Ah yes. But I wonder some special flags, which is rarey used, can be moved
> > to vm_flags2...
> >
> > For example,
> >
> > #define VM_SEQ_READ 0x00008000 /* App will access data sequentially */
> > #define VM_RAND_READ 0x00010000 /* App will not benefit from clustered reads */
> > are all capsuled under
> > mm.h
> > 117 #define VM_READHINTMASK (VM_SEQ_READ | VM_RAND_READ)
> > 118 #define VM_ClearReadHint(v) (v)->vm_flags &= ~VM_READHINTMASK
> > 119 #define VM_NormalReadHint(v) (!((v)->vm_flags & VM_READHINTMASK))
> > 120 #define VM_SequentialReadHint(v) ((v)->vm_flags & VM_SEQ_READ)
> > 121 #define VM_RandomReadHint(v) ((v)->vm_flags & VM_RAND_READ)
> >
> > Or
> >
> > 105 #define VM_PFN_AT_MMAP 0x40000000 /* PFNMAP vma that is fully mapped at mmap time */
> > is only used under special situation.
> >
> > etc..
> >
> > They'll be able to be moved to other(new) flag field, IIUC.
>
CCing Fengguang.

Breif Summary of thread:
Now, vm->vm_flags has no more avialable bits. Then, I proposed Nigel
to move some flags from vm->vm_flags to other flags as vm->vm_????.
It seems readahead-hints are candidates for this......

> I'm working on a patch to do this, and am looking at is_mergeable_vma,
> which is invoked via can_vma_merge_after from vma_merge from
> madvise_behaviour (which potentially modifies these hint flags). Should
> those hints be considered in that function? (Do I need to pass the hints
> in as well and check they're equal?)

I think it should be handled.
But, yes it implies to add a new argument to several functions in mmap.c
and maybe a patch will be ugly.

How about addding this check ?

is_mergeable_vma(...)
....
if (vma->vm_hints)
return 0;

And not calling vma_merge() at madvice(ACCESS_PATTERN_HINT).

I wonder there are little chances when madice(ACCESS_PATTERN_HINT) is
given against mapped-file-vma...

>
> By the way, VM_ClearReadHint and VM_NormalReadHint are currently unused.
> madvise_behaviour manipulates the flags directly (in preparing
> potential replacement values). Not sure if something should be done
> about that.
>
> By the way #2, in response to the later message in this thread, I'm
> calling the new var vma->vm_hints, and have put it at the end of the
> struct at the moment. Is that a good place?
>
I think it seems nice. But please before entries under CONFIG.
How about just after vm_private_data ?

Regards,
-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/