Re: Warning at mm/truncate.c:740

[Vegard Nossum]

On 10 August 2015 at 11:43, Jan Kara <jack@xxxxxxx> wrote:
> On Sun 09-08-15 10:49:33, Sasha Levin wrote:
>> I saw the following warning while fuzzing with trinity:
>>
>> [385644.689209] WARNING: CPU: 1 PID: 23536 at mm/truncate.c:740 pagecache_isize_extended+0x124/0x180()
>> [385644.691780] Modules linked in:
>> [385644.692695] CPU: 1 PID: 23536 Comm: trinity-c242 Not tainted 4.2.0-rc5-next-20150806-sasha-00040-g1b47b00-dirty #2417
>> [385644.695636]  ffffffffb21300e0 ffff8800ba3cfc40 ffffffffb1e89dfc 0000000000000000
>> [385644.708128]  ffff8800ba3cfc80 ffffffffa8325106 ffffffffa869fdd4 ffff88006bbe1f10
>> [385644.710046]  0000000000001007 ffff88006bbe1f60 ffff88006bbe1f10 ffff8803daa965a0
>> [385644.722774] Call Trace:
>> [385644.723591] dump_stack (lib/dump_stack.c:52)
>> [385644.725180] warn_slowpath_common (kernel/panic.c:448)
>> [385644.728983] warn_slowpath_null (kernel/panic.c:482)
>> [385644.730679] pagecache_isize_extended (mm/truncate.c:740 (discriminator 1))
>> [385644.732630] truncate_setsize (mm/truncate.c:710)
>> [385644.734469] v9fs_vfs_setattr_dotl (fs/9p/v9fs_vfs.h:81 fs/9p/vfs_inode_dotl.c:593)
>> [385644.753009] notify_change (fs/attr.c:270)
>> [385644.754303] do_truncate (fs/open.c:64)
>> [385644.759181] do_sys_ftruncate.constprop.5 (fs/open.c:193)
>> [385644.760669] SyS_ftruncate (fs/open.c:201)
>> [385644.761818] entry_SYSCALL_64_fastpath (arch/x86/entry/entry_64.S:186)
>>
>> But I'm not really sure how that happens... truncate_setsize() changes the inode
>> size before calling pagecache_isize_extended():
>>
>>       i_size_write(inode, newsize);
>>       if (newsize > oldsize)
>>               pagecache_isize_extended(inode, oldsize, newsize);
>>       truncate_pagecache(inode, newsize);
>>
>> And notify_change() is verifying that i_mutex is held:
>>
>>       WARN_ON_ONCE(!mutex_is_locked(&inode->i_mutex));
>>
>> So it doesn't look like a race either.
>
> Well, looking at the code it can be a race which is specific to 9p
> filesystem. It seems to me that 9p can update i_size from
> v9fs_refresh_inode_dotl(). That can be called v9fs_lookup_revalidate()
> without holding i_mutex. Now I'm not sure d_revalidate() can really race
> with truncate on the same inode (whether there isn't something else
> protecting this). Al should know better...

FWIW I'm still hitting this on latest linus/master (4.9.0-rc7+).


Vegard