Re: linux-next: manual merge of the akpm tree with the nfs tree

[Myklebust, Trond]

On Tue, 2012-07-31 at 16:19 +0100, Mel Gorman wrote:
> On Tue, Jul 31, 2012 at 02:37:24PM +0000, Myklebust, Trond wrote:
> > On Tue, 2012-07-31 at 11:33 +0100, Mel Gorman wrote:
> > > On Tue, Jul 31, 2012 at 02:24:41PM +1000, Stephen Rothwell wrote:
> > > > Hi Andrew,
> > > > 
> > > > Today's linux-next merge of the akpm tree got a conflict in
> > > > net/sunrpc/xprtsock.c between commit 5cf02d09b50b ("nfs: skip commit in
> > > > releasepage if we're freeing memory for fs-related reasons") from the nfs
> > > > tree and commit "nfs: enable swap on NFS" from the akpm tree.
> > > > 
> > > > Just context changes?  I fixed it up (I think - see below) and can carry
> > > > the fix as necessary.
> > > 
> > > Functionally it looks fine. As you say, it all looks like context
> > > changes. Arguably code like this
> > > 
> > > current->flags &= ~PF_FSTRANS
> > > 
> > > could use tsk_restore_flags instead() even though it should never be
> > > necessary as PF_FSTRANS would not be set on function entry. However,
> > > it would set up a depedency between the patch sets that is undesirable.
> > > If both sets get merged then it might make sense as a cleanup to use
> > > tsk_restore_flags() but not until then.
> > > 
> > > Thanks Stephen.
> > > 
> > 
> > Do we really need to set both PF_FSTRANS and PF_MEMALLOC here? The
> > reason why I merged the PF_FSTRANS patch is that we have the deadlock
> > problem when allocating a new socket even before we add swap-over-nfs.
> > Adding PF_FSTRANS to disallow entry into the NFS layer by the memory
> > allocator fixes that issue.
> 
> PF_FSTRANS is to prevent recursion into NFS and is set whether swap-over-NFS
> is used or not and for all requests.
> 
> > What value does PF_MEMALLOC add? Is that in order to prevent recursion
> > into other areas of the swap code (say, if you mix swap-over-nfs with
> > ordinary swap-to-disk)?
> > 
> 
> PF_MEMALLOC is normally to prevent the page reclaim recursing into
> itself. Page reclaim can call the page allocator and that cannot re-enter
> page reclaim.
> 
> In the case of swap-over-NFS, PF_MEMALLOC is set only if the socket is
> being used for swapping. In softirq context, the allocation request is
> allowed to use PFMEMALLOC reserves to avoid deadlock.
> 
> I do not see an obvious way to collapse the two flags together.
> PF_FSTRANS should not mean the PFMEMALLOC reserves can be used and
> PFMEMALLOC is not set for all requests.

Right, but in this case, we're talking about a GFP_KERNEL allocation
that always happens in an rpciod workqueue process context, so we still
won't be able to access the PFMEMALLOC reserves if I understand you
correctly?

I understand the value of preventing the page reclaim recursing into
itself, but in this case, we're talking about a separate process that is
operating on behalf of the allocator (much like kswapd does).

Cheers
  Trond
N‹§²æìr¸›yúèšØb²X¬¶ÇvØ^–)Þ{.nÇ+‰·¥Š{±‘êçzX§¶›¡Ü}©ž²ÆzÚ&j:+v‰¨¾«‘êçzZ+€Ê+zf£¢·hšˆ§~†­†Ûiÿûàz¹®w¥¢¸?™¨è­Ú&¢)ßf”ù^jÇy§m…á@A«a¶Úÿ0¶ìh®å’i