[Intel-gfx] NPE in i915_gemfs_init

[Chris Wilson]

Quoting Heiner Kallweit (2019-07-14 13:51:45)
> On 14.07.2019 14:46, Chris Wilson wrote:
> > Quoting Heiner Kallweit (2019-07-14 13:39:15)
> >> On 14.07.2019 14:34, Chris Wilson wrote:
> >>> Quoting Heiner Kallweit (2019-07-13 12:12:56)
> >>>> I debugged a little bit and remount_fs isn't set in sb->s_op.
> >>>> The following at least avoids the NPE, not sure whether it's the correct fix.
> >>>
> >>> I take it you don't have CONFIG_TMPFS set?
> >>>
> >> This option is set:
> >>
> >> [root at zotac linux-next]# grep TMPFS .config
> >> CONFIG_DEVTMPFS=y
> >> CONFIG_DEVTMPFS_MOUNT=y
> >> CONFIG_TMPFS=y
> >> CONFIG_TMPFS_POSIX_ACL=y
> >> CONFIG_TMPFS_XATTR=y
> > 
> > Now that's weird, as
> > 
> > static const struct super_operations shmem_ops = {
> >         .alloc_inode    = shmem_alloc_inode,
> >         .free_inode     = shmem_free_in_core_inode,
> >         .destroy_inode  = shmem_destroy_inode,
> > #ifdef CONFIG_TMPFS
> >         .statfs         = shmem_statfs,
> >         .remount_fs     = shmem_remount_fs,
> >         .show_options   = shmem_show_options,
> > #endif
> >         .evict_inode    = shmem_evict_inode,
> >         .drop_inode     = generic_delete_inode,
> >         .put_super      = shmem_put_super,
> > #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
> >         .nr_cached_objects      = shmem_unused_huge_count,
> >         .free_cached_objects    = shmem_unused_huge_scan,
> > #endif
> > };
> > 
> > the only way it should be unset was if !CONFIG_TMPFS.
> > 
> > Hmm, we even select TMPFS so it should never be absent.
> > 
> > So the question is what mount did we get if it was not a shmemfs one?
> > -Chris
> > 
> I think that's the reason:
> 144df3b288c4 ("vfs: Convert ramfs, shmem, tmpfs, devtmpfs, rootfs to use the new mount API")
> 
> @@ -3736,7 +3849,6 @@ static const struct super_operations shmem_ops = {
>         .destroy_inode  = shmem_destroy_inode,
>  #ifdef CONFIG_TMPFS
>         .statfs         = shmem_statfs,
> -       .remount_fs     = shmem_remount_fs,
>         .show_options   = shmem_show_options,
>  #endif
>         .evict_inode    = shmem_evict_inode,
> 
> Most likely this "new mount API" impacts your code.

In that scheme, it looks like there's a "reconfigure" op to replace
remount_fs. Not sure how we hook that up yet, it appears to take the
fs_context for mounting, so presumably there's a new kern_mount() API as
well.

Or we can keep the crude hack to adjust sb_info->huge directly :|
-Chris