Re: [patch 2/2] bootmem: Node-setup agnostic free_bootmem()

[Johannes Weiner <hannes@saeurebad.de>]

Hi,

"Yinghai Lu" <yhlu.kernel@gmail.com> writes:

> On Tue, Apr 15, 2008 at 4:51 AM, Johannes Weiner <hannes@saeurebad.de> wrote:
>> Hi,
>>
>>
>>
>>  Andrew Morton <akpm@linux-foundation.org> writes:
>>
>>  > On Tue, 15 Apr 2008 00:28:34 -0700 "Yinghai Lu" <yhlu.kernel@gmail.com> wrote:
>>  >
>>  >> On Tue, Apr 15, 2008 at 12:15 AM, Andrew Morton
>>  >> <akpm@linux-foundation.org> wrote:
>>  >> >
>>  >> > On Tue, 15 Apr 2008 00:04:03 -0700 "Yinghai Lu" <yhlu.kernel@gmail.com> wrote:
>>  >> >
>>  >> >  > On Mon, Apr 14, 2008 at 11:23 PM, Andrew Morton
>>  >> >  > <akpm@linux-foundation.org> wrote:
>>  >> >  > >
>>  >> >  > > On Sun, 13 Apr 2008 18:56:57 +0200 Andi Kleen <andi@firstfloor.org> wrote:
>>  >> >  > >
>>  >> >  > >  > Johannes Weiner <hannes@saeurebad.de> writes:
>>  >> >  > >  >
>>  >> >  > >  > > Make free_bootmem() look up the node holding the specified address
>>  >> >  > >  > > range which lets it work transparently on single-node and multi-node
>>  >> >  > >  > > configurations.
>>  >> >  > >  >
>>  >> >  > >  > Acked-by: Andi Kleen <andi@firstfloor.org>
>>  >> >  > >  >
>>  >> >  > >  > This is far better than the original change it replaces and which
>>  >> >  > >  > I also objected to in review.
>>  >> >  > >  >
>>  >> >  > >
>>  >> >  > >  So...  do we think these two patches are sufficiently safe and important for
>>  >> >  > >  2.6.25?
>>  >> >  >
>>  >> >  > the patch is wrong
>>  >> >  >
>>  >> >
>>  >> >  The last I saw was this:
>>  >> >
>>  >> >
>>  >> >  On Sun, 13 Apr 2008 12:57:22 +0200 Johannes Weiner <hannes@saeurebad.de> wrote:
>>  >> >
>>  >> >  > Hi,
>>  >> >  >
>>  >> >  > "Yinghai Lu" <yhlu.kernel@gmail.com> writes:
>>  >> >  >
>>  >> >  > > On Sat, Apr 12, 2008 at 3:33 PM, Johannes Weiner <hannes@saeurebad.de> wrote:
>>  >> >  > > ...
>>  >> >
>>  >> > > >
>>  >> >  > > could have chance that bootmem with reserved_early that is crossing
>>  >> >  > > the nodes.
>>  >> >  >
>>  >> >  > Upstream reserve_bootmem_core() would BUG() on a caller trying to cross
>>  >> >  > nodes, so I don't see where this chance could come from.
>>  >> >
>>  >> >  Is that what you're referring to?
>>  >> >
>>  >> >  Was Johannes observation incorrect?  If so, why?
>>  >>
>>  >> my patch with free_bootmem will make sure free_bootmem_core only free
>>  >> bootmem in the bdata scope.
>>  >> so free_bootmem can handle the cross_node bootmem that is done by
>>  >> reserve_early ( done in another patch, is dropped by you because took
>>  >> Jonannes).
>>  >>
>>  >> in setup_arch for x86_64 there is one free_bootmem that is used when
>>  >> ramdisk is falled out of ram map. that could be crossed by bootloader
>>  >> and kexec, and kernel or second kernel is memmap=NN@SS to execlue some
>>  >> memory.
>>  >>
>>  >> anyway that is extrem case, but my patch could handle that.
>>
>>  Has this case ever occured?  If this could become real, I have no
>>  objections to implement a way to handle it (why would I?), but until now
>>  you just said that in some time in the future, this could be useful.
>>
>>
>>  >>
>>  >> I wonder if any regression caused by my previous patch? maybe on other platform?
>>  >>
>>  >
>>  > Not that I'm aware of.
>>
>>  It papers over buggy usage of free_bootmem().  If its arguments are
>>  bogus, it will just return again where it BUG()ed out before.  The pages
>>  might be never marked free and therefor never reach the buddy allocator.
>>
>>
>>  > I restored mm-make-reserve_bootmem-can-crossed-the-nodes.patch.  Johannes,
>>  > can you please check 2.6.28-rc8-mm2, see if it looks OK?
>>
>>  I object to the way it is implemented.  If it is really needed, that
>>  should be done properly:
>>
>>         - remove the double loop over the area on the likely succeeding
>>           path and unroll the reserving on the unlikely path as it was
>>           done before.  Better to punish exceptional branches than
>>           the working paths.
>>         - make reserve_bootmem_core be strict with its arguments.  If
>>           you want to iterate over the bdata list, you should not just
>>           throw every item at the _core functions and let them work it
>>           out for themselves.  The correct parameters should be
>>           calculated in advance and then passed to a strict
>>           _bootmem_core() function that BUG()s on failure.
>>
>>  But still, Yinghai, you never brought in practical reasons for this
>>  whole thing.  You talked about extreme and theoretical cases and I don't
>>  think that this justifies breaking API or pessimizing code at all.
>
> free_bootmem(ramdisk_image, ramdisk_size) is sitting in setup_arch of
> x86_64. or make that panic directly.
>
> what i needed is: free_bootmem can free bootmem cross the nodes.
>
> on numa
> alloc_bootmem always return blocks on same nodes. but some via
> reserve_early and then to bootmem via early_res_to_bootmem could be
> crossing nodes.
>
> BTW, can you look at patches in -mm about make reserve_bootmem cross
> the nodes?

Yep, already looked at them.  My patches were initially against Linus'
tree which does not allow bootmem to act across node boundaries yet.

Regarding node-crossing, what do you think about my idea in
http://lkml.org/lkml/2008/4/15/139?  That way we could preserve the core
functions and keep them clean.  The design could of course be applied to
the other node-crossing functions too.

	Hannes