Re: [PATCH] vfs: check i_count under lock in evict_inodes
On Mon, Jul 11, 2016 at 06:31:57PM -0700, David Chen wrote: > Hi Al, > > I'm not sure about the in-tree fs, but in zfsonlinux, it would offload > iput to a thread, so this would happen there. And it would wait for > the thread in put_super(), so that part is not a problem... *shrug* I hadn't looked (and won't look) at zfs glue, but I'd suggest trying something along the line of stopping that thread in the beginning of your ->kill_sb() (having told the sucker to stop offloading, of course) and only then calling generic_shutdown_super()...
Re: [PATCH] vfs: check i_count under lock in evict_inodes
On Mon, Jul 11, 2016 at 06:31:57PM -0700, David Chen wrote: > Hi Al, > > I'm not sure about the in-tree fs, but in zfsonlinux, it would offload > iput to a thread, so this would happen there. And it would wait for > the thread in put_super(), so that part is not a problem... And why exactly is your use of a broken and undistributable out of tree module our problem?
Re: [PATCH] vfs: check i_count under lock in evict_inodes
Hi Al, I'm not sure about the in-tree fs, but in zfsonlinux, it would offload iput to a thread, so this would happen there. And it would wait for the thread in put_super(), so that part is not a problem... Thanks 2016-07-11 17:46 GMT-07:00 Al Viro : > On Mon, Jul 11, 2016 at 05:15:04PM -0700, Chunwei Chen wrote: >> We need to check i_count again with i_lock held, because iput might re-add >> i_count when lazytime is on. Without this check, we could end up with >> double-free or use-after-free. > > Details, please. Ideally - with a reproducer. Who is calling that iput() > at that point of generic_shutdown_super() (has to be another thread) and > just what will happen if the same iput() is delayed until *after* > evict_inodes(), all the way into ->put_super(). At which point there's > no promise whatsoever that the data structures used by ->evict_inode() > hadn't been already freed... >
Re: [PATCH] vfs: check i_count under lock in evict_inodes
On Mon, Jul 11, 2016 at 05:15:04PM -0700, Chunwei Chen wrote: > We need to check i_count again with i_lock held, because iput might re-add > i_count when lazytime is on. Without this check, we could end up with > double-free or use-after-free. Details, please. Ideally - with a reproducer. Who is calling that iput() at that point of generic_shutdown_super() (has to be another thread) and just what will happen if the same iput() is delayed until *after* evict_inodes(), all the way into ->put_super(). At which point there's no promise whatsoever that the data structures used by ->evict_inode() hadn't been already freed...
[PATCH] vfs: check i_count under lock in evict_inodes
We need to check i_count again with i_lock held, because iput might re-add i_count when lazytime is on. Without this check, we could end up with double-free or use-after-free. Cc: Alexander Viro Cc: linux-fsde...@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: sta...@vger.kernel.org Signed-off-by: Chunwei Chen --- fs/inode.c | 7 ++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/fs/inode.c b/fs/inode.c index 4ccbc21..10bb020 100644 --- a/fs/inode.c +++ b/fs/inode.c @@ -607,7 +607,12 @@ again: continue; spin_lock(&inode->i_lock); - if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) { + /* +* check i_count again with lock, because iput might re-add +* it when lazytime is on. +*/ + if (atomic_read(&inode->i_count) || + (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE))) { spin_unlock(&inode->i_lock); continue; } -- 2.7.4