[PATCH V1] fuse: Remove __GFP_FS flag to avoid allocator recursing

[PATCH V1] fuse: Remove __GFP_FS flag to avoid allocator recursing

[Pradeep P V K]

Found a deadlock between kswapd, writeback thread and fuse process
Here are the sequence of events with callstacks on the deadlock.

process#1		process#2		process#3
__switch_to+0x150	__switch_to+0x150	try_to_free_pages
__schedule+0x984	__schedule+0x984
					memalloc_noreclaim_restore
schedule+0x70		schedule+0x70		__perform_reclaim
bit_wait+0x14		__fuse_request_send+0x154
					__alloc_pages_direct_reclaim
__wait_on_bit+0x70	fuse_simple_request+0x174
inode_wait_for_writeback+0xa0
						__alloc_pages_slowpath
			fuse_flush_times+0x10c
evict+0xa4		fuse_write_inode+0x5c	__alloc_pages_nodemask
iput+0x248		__writeback_single_inode+0x3d4
dentry_unlink_inode+0xd8			__alloc_pages_node
			writeback_sb_inodes+0x4a0
__dentry_kill+0x160	__writeback_inodes_wb+0xac
shrink_dentry_list+0x170			alloc_pages_node
			wb_writeback+0x26c	fuse_copy_fill
prune_dcache_sb+0x54	wb_workfn+0x2c0		fuse_copy_one
super_cache_scan+0x114	process_one_work+0x278	fuse_read_single_forget
do_shrink_slab+0x24c	worker_thread+0x26c	fuse_read_forget
shrink_slab+0xa8	kthread+0x118		fuse_dev_do_read
shrink_node+0x118				fuse_dev_splice_read
kswapd+0x92c					do_splice_to
						do_splice

Process#1(kswapd) held an inode lock and initaited a writeback to free
the pages, as the inode superblock is fuse, process#2 forms a fuse
request. Process#3 (Fuse daemon threads) while serving process#2 request,
it requires memory(pages) and as the system is already running in low
memory it ends up in calling try_to_ free_pages(), which might now call
kswapd again, which is already stuck with an inode lock held. Thus forms
a deadlock.

So, remove __GFP_FS flag to avoid allocator recursing into the
filesystem that might already held locks.

Signed-off-by: Pradeep P V K <ppvk@codeaurora.org>
---
 fs/fuse/dev.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/fs/fuse/dev.c b/fs/fuse/dev.c
index 02b3c36..2859024 100644
--- a/fs/fuse/dev.c
+++ b/fs/fuse/dev.c
@@ -708,7 +708,7 @@ static int fuse_copy_fill(struct fuse_copy_state *cs)
 			if (cs->nr_segs >= cs->pipe->max_usage)
 				return -EIO;
 
-			page = alloc_page(GFP_HIGHUSER);
+			page = alloc_page(GFP_NOFS | __GFP_HARDWALL | __GFP_HIGHMEM);
 			if (!page)
 				return -ENOMEM;
 
-- 
2.7.4

Re: [PATCH V1] fuse: Remove __GFP_FS flag to avoid allocator recursing

[Matthew Wilcox]

On Mon, Sep 14, 2020 at 07:26:15PM +0530, Pradeep P V K wrote:
> Process#1(kswapd) held an inode lock and initaited a writeback to free
> the pages, as the inode superblock is fuse, process#2 forms a fuse
> request. Process#3 (Fuse daemon threads) while serving process#2 request,
> it requires memory(pages) and as the system is already running in low
> memory it ends up in calling try_to_ free_pages(), which might now call
> kswapd again, which is already stuck with an inode lock held. Thus forms
> a deadlock.
> 
> So, remove __GFP_FS flag to avoid allocator recursing into the
> filesystem that might already held locks.

This is the wrong way to fix the problem.  The fuse daemon threads should
have called memalloc_nofs_save() as this prevents them from inadvertently
tripping over other places where they forgot to use GFP_NOFS (or have
no way to pass a GFP_NOFS flags argument).

Re: [PATCH V1] fuse: Remove __GFP_FS flag to avoid allocator recursing

[ppvk]

On 2020-09-14 19:37, Matthew Wilcox wrote:
> On Mon, Sep 14, 2020 at 07:26:15PM +0530, Pradeep P V K wrote:
>> Process#1(kswapd) held an inode lock and initaited a writeback to free
>> the pages, as the inode superblock is fuse, process#2 forms a fuse
>> request. Process#3 (Fuse daemon threads) while serving process#2 
>> request,
>> it requires memory(pages) and as the system is already running in low
>> memory it ends up in calling try_to_ free_pages(), which might now 
>> call
>> kswapd again, which is already stuck with an inode lock held. Thus 
>> forms
>> a deadlock.
>> 
>> So, remove __GFP_FS flag to avoid allocator recursing into the
>> filesystem that might already held locks.
> 
> This is the wrong way to fix the problem.  The fuse daemon threads 
> should
> have called memalloc_nofs_save() as this prevents them from 
> inadvertently
> tripping over other places where they forgot to use GFP_NOFS (or have
> no way to pass a GFP_NOFS flags argument).

Thanks Matthew for pointing this. I will address this in my next patch 
set.

Thanks and Regards,
Pradeep