Lustre discuss - Mar 2007 - PAGE_SIZE usage in lustre and a debian fix.

Okay PAGE_SIZE and friends isn''t available for some arches in the
kernel, in fact the only architectures that PAGE_SIZE and friends are
available for code that isn''t compiled with __KERNEL__ are i386, m32r,
ppc, s390, sh64, x86_64, so after trying to compile the debian
packages (which build the userspace components separate from the
kernel components) on ia64 it failed.

Attached is a patch to ''fix'' the problem the first part of the
patch
is what should be happening. The second part of the patch is a hack to
get the source to build for debian built ia64 kernels.
>From comments on the LKML PAGE_SIZE shouldn''t ever be used by
userspace and should be using getpagesize() and perform some functions
to get the other values (PAGE_MASK and PAGE_SHIFT).

Hopefully this helps,
- David Brown
-------------- next part --------------
A non-text attachment was scrubbed...
Name: ia64-page-size-fixes.patch
Type: text/x-patch
Size: 1084 bytes
Desc: not available
Url :
http://mail.clusterfs.com/pipermail/lustre-discuss/attachments/20070305/5d193ddf/ia64-page-size-fixes.bin

On Mar 05, 2007  17:40 -0800, David Brown wrote:
> Okay PAGE_SIZE and friends isn''t available for some arches in the
> kernel, in fact the only architectures that PAGE_SIZE and friends are
> available for code that isn''t compiled with __KERNEL__ are i386,
m32r,
> ppc, s390, sh64, x86_64, so after trying to compile the debian
> packages (which build the userspace components separate from the
> kernel components) on ia64 it failed.
> 
> Attached is a patch to ''fix'' the problem the first part
of the patch
> is what should be happening. The second part of the patch is a hack to
> get the source to build for debian built ia64 kernels.
> 
> >From comments on the LKML PAGE_SIZE shouldn''t ever be used by
> userspace and should be using getpagesize() and perform some functions
> to get the other values (PAGE_MASK and PAGE_SHIFT).
I''m not sure that liblustre has ever been tested on ia64, though of
course
it is entirely possible that it works OK.  The value of PAGE_SIZE used by
liblustre is pretty arbitrary I think, a large part of it is just so that
the same code can be used by liblustre and linux.

As for your patch, I haven''t looked at it closely, but I''m
confused why
you would change PAGE_SIZE -> getpagesize(), and then later make a
#define PAGE_SIZE?  Also, there is no guarantee that the ia64 PAGE_SIZE
is 2^14, though whether liblustre really cares is another matter.

Cheers, Andreas
--
Andreas Dilger
Principal Software Engineer
Cluster File Systems, Inc.

> I''m not sure that liblustre has ever been tested on ia64, though
of course
> it is entirely possible that it works OK.  The value of PAGE_SIZE used by
> liblustre is pretty arbitrary I think, a large part of it is just so that
> the same code can be used by liblustre and linux.
>
> As for your patch, I haven''t looked at it closely, but
I''m confused why
> you would change PAGE_SIZE -> getpagesize(), and then later make a
> #define PAGE_SIZE?  Also, there is no guarantee that the ia64 PAGE_SIZE
> is 2^14, though whether liblustre really cares is another matter.
I''ve been having an on going conversation on lkml about it and they
really say, "use getpagesize()" but lustre uses PAGE_MASK PAGE_SHIFT
and all the other defines in your userspace code as well. So the first
part of the patch is what should be happening and the second part is
fast hack to get it to build with a ''safe'' value for the page
size.
Really the correct fix would be to obtain the value of PAGE_SIZE via
getpagesize() then create some functions to get PAGE_MASK and
PAGE_SHIFT from the value returned by getpagesize(). At least that''s
what I''d do...

Thanks,
- David Brown

On Mar 08, 2007  13:34 -0800, David Brown wrote:
> I''ve been having an on going conversation on lkml about it and
they
> really say, "use getpagesize()" but lustre uses PAGE_MASK
PAGE_SHIFT
> and all the other defines in your userspace code as well. So the first
> part of the patch is what should be happening and the second part is
> fast hack to get it to build with a ''safe'' value for the
page size.
> Really the correct fix would be to obtain the value of PAGE_SIZE via
> getpagesize() then create some functions to get PAGE_MASK and
> PAGE_SHIFT from the value returned by getpagesize(). At least
that''s
> what I''d do...
My point is that because liblustre is really a user-space application,
it doesn''t care so much what value is used for PAGE_SIZE.  I suspect
(though I haven''t really done an in-depth analysis) that you could
set PAGE_SIZE to some arbitrary power-of-two multiple of 4096 and not
have any problems at all.

The PAGE_SIZE define IS used in the code that is shared between liblustre
and the linux client filesystem code (llite) but those are used in
completely disjoint settings.  The only thing that matters is that the
liblustre build environment supply SOME sensible value for PAGE_SIZE &
friends so that the shared code will compile.

Cheers, Andreas
--
Andreas Dilger
Principal Software Engineer
Cluster File Systems, Inc.