On Tue, 3 Aug 2021, Matthew Wilcox wrote:
> On Tue, Aug 03, 2021 at 04:14:38PM +0800, Huang, Ying wrote:
> > Matthew Wilcox <willy@xxxxxxxxxxxxx> writes:
> > > But I REALLY REALLY REALLY want a reproducer. Right now, I have a hard
> > > time believing this, or any of the other races can really happen.
> >
> > I think the race is only theoretical too. Firstly, swapoff is a rare
> > operations in practice; secondly, the race window is really small.
>
> So do something to provoke it. Widen the window. Put an msleep(1000)
> between *pagep = NULL and the call to get_swap_device(). That's assuming
> that the swapon/swapoff loop that I proposed doesn't work. Did you
> try it?
I've been doing that swapon/swapoff loop for years, while running kernel
builds on tmpfs going out to swap; for better or worse on baremetal not VM.
You're right that few will ever need that level of reliability; but it
has caught problems from time to time, and I do insist on fixing them.
I'm not as insistent as you on wanting a reproducer; and we all take pride
sometimes in fixing ever more inconceivable bugs. I'm not against that,
but it's easy to end up with a fix more dangerous than what it claims to
fix, rather like with random newbie cleanups.
I've never seen the swapoff race claimed by Miaohe, and don't expect to;
but he's probably right, given the current code. I just dislike adding
unnecessary complexity, and siting it in the wrong place (mm/shmem.c).
Yang, is it possible that 5.1 commit 8fd2e0b505d1 ("mm: swap: check if
swap backing device is congested or not") was actually developed and
measured on 4.1 or earlier, which still had blk_set_queue_congested()?
I cannot explain its usefulness nowadays, on congested HDD anyway:
Matthew is right that NFS and a few others may still be setting
congested flags, but they're not what that commit was proposed for.
If it is still useful, then I contend (but Huang Ying will disagree)
that the get_swap_device() and put_swap_device() should be around
8fd2e0b505d1's inode_read_congested() block in swap_cluster_readahead(),
not encroaching into mm/shmem.c.
But if that block is not useful, then it should simply be removed (later).
Hugh
