Re: [lkp-robot] [btrfs] beeeccca9b: WARNING:at_mm/util.c:#kvmalloc_node

[Omar Sandoval <osandov@osandov.com>]

On Wed, May 31, 2017 at 11:19:05AM +0200, Michal Hocko wrote:
> On Wed 31-05-17 02:12:02, Omar Sandoval wrote:
> > On Wed, May 31, 2017 at 08:51:28AM +0200, Michal Hocko wrote:
> > > On Wed 31-05-17 14:30:33, kernel test robot wrote:
> > > > 
> > > > FYI, we noticed the following commit:
> > > > 
> > > > commit: beeeccca9bebcec386cc31c250cff8a06cf27034 ("btrfs: Use kvzalloc instead of kzalloc/vmalloc in alloc_bitmap")
> > > > https://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git master
> > > 
> > > I have intentionally skipped alloc_bitmap because it relies on GFP_NOFS.
> > > This doesn't work properly when falling back to vmalloc and that is what
> > > the warning reported here says. I believe the right approach is to check
> > > whether the GFP_NOFS is _really_ needed and document why if yes.
> > > Otherwise drop the NOFS part in one patch with the explanation and
> > > convert it to kvmalloc in a separate patch.
> > 
> > Unfortunately we really do need GFP_NOFS here, the free space tree is
> > modified while we are committing a fs transaction, sometimes in the
> > critical section when we block new operations from joining the
> > transaction.
> 
> OK, please document this.
> 
> > Looking at the comment in kvmalloc_node():
> > 
> > 	/*
> > 	 * vmalloc uses GFP_KERNEL for some internal allocations (e.g page tables)
> > 	 * so the given set of flags has to be compatible.
> > 	 */
> > 	WARN_ON_ONCE((flags & GFP_KERNEL) != GFP_KERNEL);
> > 
> > has alloc_bitmap() always been broken by virtue of calling vmalloc()
> > with GFP_NOFS?
> 
> yes. vmalloc is simply not GFP_NOFS safe as it performs GFP_KERNEL
> hardcoded allocations. The way out of this is to use
> memalloc_nofs_{save,restore} around kvmalloc call.

Ok, thanks. Something like this untested patch instead of the one in
linux-next?

diff --git a/fs/btrfs/free-space-tree.c b/fs/btrfs/free-space-tree.c
index fc0bd8406758..5abd3cd71144 100644
--- a/fs/btrfs/free-space-tree.c
+++ b/fs/btrfs/free-space-tree.c
@@ -153,21 +153,18 @@ static inline u32 free_space_bitmap_size(u64 size, u32 sectorsize)
 
 static u8 *alloc_bitmap(u32 bitmap_size)
 {
-	void *mem;
+	u8 *ret;
+	unsigned int nofs_flag;
 
 	/*
-	 * The allocation size varies, observed numbers were < 4K up to 16K.
-	 * Using vmalloc unconditionally would be too heavy, we'll try
-	 * contiguous allocations first.
+	 * GFP_NOFS doesn't work with kvmalloc(), but we really can't recurse
+	 * into the filesystem as the free space bitmap can be modified in the
+	 * critical section of a transaction commit.
 	 */
-	if  (bitmap_size <= PAGE_SIZE)
-		return kzalloc(bitmap_size, GFP_NOFS);
-
-	mem = kzalloc(bitmap_size, GFP_NOFS | __GFP_NOWARN);
-	if (mem)
-		return mem;
-
-	return __vmalloc(bitmap_size, GFP_NOFS | __GFP_ZERO, PAGE_KERNEL);
+	nofs_flag = memalloc_nofs_save();
+	ret = kvmalloc(bitmap_size, GFP_KERNEL);
+	memalloc_nofs_restore(nofs_flag);
+	return ret;
 }
 
 int convert_free_space_to_bitmaps(struct btrfs_trans_handle *trans,

Dave, would you prefer to replace the patch we have now or do an
incremental patch on top of it?

Michal, is there some reason we can't have kvmalloc() with
!(gfp & __GFP_FS) just do the memalloc_nofs dance internally?