Re: [PATCH v2 06/11] don't put symlink bodies in pagecache into highmem

[Al Viro]

On Thu, Jan 14, 2016 at 04:58:48PM +0100, Tomeu Vizoso wrote:
> > Could you add
> >         printk(KERN_ERR "i_data = %p, i_mapping = %p, flags: %lx\n",
> >                 &inode->i_data,
> >                 inode->i_mapping,
> >                 (unsigned long)inode->i_data.flags);
> > right before the return from nfs_get_link() and see what it prints?
> 
> Here it is:
> 
> [  170.136956] i_data = ed9c1b04, i_mapping = ed9c1b04, flags: 24200c0
> [  170.144567] i_data = ed9de784, i_mapping = ed9de784, flags: 24200c0
> [  170.151457] i_data = ed9dec84, i_mapping = ed9dec84, flags: 24200c0
> [  170.158358] i_data = ed9c3b84, i_mapping = ed9c3b84, flags: 24200c0
> [  170.165253] i_data = ed9d4204, i_mapping = ed9d4204, flags: 24200c0
> [  170.172131] i_data = ed9df184, i_mapping = ed9df184, flags: 24200c0
> [  170.188804] i_data = eddbce84, i_mapping = eddbce84, flags: 24200c0
> [  170.196158] i_data = ec904984, i_mapping = ec904984, flags: 24200c0
> [  170.205133] i_data = ec906784, i_mapping = ec906784, flags: 24200c0

Aha.  So ->i_data vs. ->i_mapping is irrelevant (as it ought to be here)
and inode_nohighmem() should've acted on the address_space we are hitting
here.  What do we have in flags...  ___GFP_IO | ___GFP_FS | ___GFP_HARDWALL |
___GFP_DIRECT_RECLAIM | ___GFP_KSWAPD_RECLAIM.  IOW, normal GFP_USER, no
__GFP_HIGHMEM in sight.

So either we have a highmem page somehow ending up in i_data before we
set the flags, or __page_cache_alloc() done by read_cache_page() returns
us a highmem page on GFP_USER | __GFP_COLD (or I'm misreading the things
completely)...

Could you slap
	printk(KERN_ERR "inode: %p, pages: %ld\n",
		inode, inode->i_data.nrpages);
before that read_cache_page() in nfs_get_link() and
	printk(KERN_ERR "page_address: %p\n", page_address(page));
right before the return?