Re: [PATCH v4 1/3] dax: masking off __GFP_FS in fs DAX handlers

["Jiang, Dave" <dave.jiang@intel.com>]

On Sat, 2016-12-17 at 09:04 +1100, Dave Chinner wrote:
> On Fri, Dec 16, 2016 at 09:19:16AM -0700, Ross Zwisler wrote:
> > 
> > On Fri, Dec 16, 2016 at 12:07:30PM +1100, Dave Chinner wrote:
> > > 
> > > On Thu, Dec 15, 2016 at 04:40:41PM -0700, Dave Jiang wrote:
> > > > 
> > > > The caller into dax needs to clear __GFP_FS mask bit since it's
> > > > responsible for acquiring locks / transactions that blocks
> > > > __GFP_FS
> > > > allocation.  The caller will restore the original mask when dax
> > > > function
> > > > returns.
> > > 
> > > What's the allocation problem you're working around here? Can you
> > > please describe the call chain that is the problem?
> > > 
> > > > 
> > > >  	xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
> > > >  
> > > >  	if (IS_DAX(inode)) {
> > > > +		gfp_t old_gfp = vmf->gfp_mask;
> > > > +
> > > > +		vmf->gfp_mask &= ~__GFP_FS;
> > > >  		ret = dax_iomap_fault(vma, vmf,
> > > > &xfs_iomap_ops);
> > > > +		vmf->gfp_mask = old_gfp;
> > > 
> > > I really have to say that I hate code that clears and restores
> > > flags
> > > without any explanation of why the code needs to play flag
> > > tricks. I
> > > take one look at the XFS fault handling code and ask myself now
> > > "why
> > > the hell do we need to clear those flags?" Especially as the
> > > other
> > > paths into generic fault handlers /don't/ require us to do this.
> > > What does DAX do that require us to treat memory allocation
> > > contexts
> > > differently to the filemap_fault() path?
> > 
> > This was done in response to Jan Kara's concern:
> > 
> >   The gfp_mask that propagates from __do_fault() or
> > do_page_mkwrite() is fine
> >   because at that point it is correct. But once we grab filesystem
> > locks which
> >   are not reclaim safe, we should update vmf->gfp_mask we pass
> > further down
> >   into DAX code to not contain __GFP_FS (that's a bug we apparently
> > have
> >   there). And inside DAX code, we definitely are not generally safe
> > to add
> >   __GFP_FS to mapping_gfp_mask(). Maybe we'd be better off
> > propagating struct
> >   vm_fault into this function, using passed gfp_mask there and make
> > sure
> >   callers update gfp_mask as appropriate.
> > 
> > https://lkml.org/lkml/2016/10/4/37
> > 
> > IIUC I think the concern is that, for example, in
> > xfs_filemap_page_mkwrite()
> > we take a read lock on the struct inode.i_rwsem before we call
> > dax_iomap_fault().
> 
> That, my friends, is exactly the problem that mapping_gfp_mask() is
> meant to solve. This:
> 
> > 
> > > 
> > > > 
> > > > +	vmf.gfp_mask = mapping_gfp_mask(mapping) | __GFP_FS
> > > > |  __GFP_IO;
> 
> Is just so wrong it's not funny.
> 
> The whole point of mapping_gfp_mask() is to remove flags from the
> gfp_mask used to do mapping+page cache related allocations that the
> mapping->host considers dangerous when the host may be holding locks.
> This includes mapping tree allocations, and anything else required
> to set up a new entry in the mapping during IO path operations. That
> includes page fault operations...
> 
> e.g. in xfs_setup_inode():
> 
>         /*
>          * Ensure all page cache allocations are done from GFP_NOFS
> context to
>          * prevent direct reclaim recursion back into the filesystem
> and blowing
>          * stacks or deadlocking.
>          */
>         gfp_mask = mapping_gfp_mask(inode->i_mapping);
>         mapping_set_gfp_mask(inode->i_mapping, (gfp_mask &
> ~(__GFP_FS)));
> 
> i.e. XFS considers it invalid to use GFP_FS at all for mapping
> allocations in the io path, because we *know* that we hold
> filesystems locks over those allocations.
> 
> > 
> > dax_iomap_fault() then calls find_or_create_page(), etc. with the
> > vfm->gfp_mask we were given.
> 
> Yup. Precisely why we should be using mapping_gfp_mask() as it was
> intended for vmf.gfp_mask....
> 
> > 
> > I believe the concern is that if that memory allocation tries to do
> > FS
> > operations to free memory because __GFP_FS is part of the gfp mask,
> > then we
> > could end up deadlocking because we are already holding FS locks.
> 
> Which is a problem with the filesystem mapping mask setup, not a
> reason to sprinkle random gfpmask clear/set pairs around the code.
> i.e. For DAX inodes, the mapping mask should clear __GFP_FS as XFS
> does above, and the mapping_gfp_mask() should be used unadulterated
> by the DAX page fault code....


I'll drop this patch. We can address the issue separate from the
pmd_fault changes. 

> 
> Cheers,
> 
> Dave.