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

[Tomeu Vizoso]

On 14 January 2016 at 18:13, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> On Thu, Jan 14, 2016 at 05:57:42PM +0100, Tomeu Vizoso wrote:
>> Here it is:
>>
>> [  170.715356] inode: ec8c30b0, pages: 1
>> [  170.719014] page_address:   (null)
>>
>> https://lava.collabora.co.uk/scheduler/job/127698/log_file
>
>         Lovely...  And that looks like the first time that inode hits
> nfs_get_link().  Ho-hum...
>
>         Could you add WARN_ON(inode->i_mapping.nrpages) in inode_nohighmem()
> and see if that triggers?  It really shouldn't (we hit it after iget5_locked()

Indeed :(

https://lava.collabora.co.uk/scheduler/job/127782/log_file

diff --git a/fs/inode.c b/fs/inode.c
index 1e6dd388ba7f..b957744376e8 100644
--- a/fs/inode.c
+++ b/fs/inode.c
@@ -2031,6 +2031,7 @@ EXPORT_SYMBOL(inode_set_flags);

 void inode_nohighmem(struct inode *inode)
 {
+       WARN_ON(inode->i_mapping->nrpages);
        mapping_set_gfp_mask(inode->i_mapping, GFP_USER);
 }
 EXPORT_SYMBOL(inode_nohighmem);

I have ran out of time for today, but tomorrow can do more tests if needed.

Regards,

Tomeu

> in nfs_fhget() has returned us a new inode, and there shouldn't be a chance
> for any pages to get in there between struct inode allocation and that
> point), but then neither should highmem pages be added to address_space
> without __GFP_HIGHMEM in ->flags...