Hi Andrew,
Sorry for replying a little late, it's holiday here yesterday.
On Tue, Apr 04, 2023 at 12:26:00PM -0700, Andrew Morton wrote:
> On Tue, 4 Apr 2023 23:47:16 +0800 Rongwei Wang <rongwei.wang@xxxxxxxxxxxxxxxxx> wrote:
>
> > The si->lock must be held when deleting the si from
> > the available list.
> >
> > ...
> >
> > --- a/mm/swapfile.c
> > +++ b/mm/swapfile.c
> > @@ -679,6 +679,7 @@ static void __del_from_avail_list(struct swap_info_struct *p)
> > {
> > int nid;
> >
> > + assert_spin_locked(&p->lock);
> > for_each_node(nid)
> > plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
> > }
> > @@ -2434,8 +2435,8 @@ SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
> > spin_unlock(&swap_lock);
> > goto out_dput;
> > }
> > - del_from_avail_list(p);
> > spin_lock(&p->lock);
> > + del_from_avail_list(p);
> > if (p->prio < 0) {
> > struct swap_info_struct *si = p;
> > int nid;
>
> So we have
>
> swap_avail_lock
> swap_info_struct.lock
> swap_cluster_info.lock
>
> Is the ranking of these three clearly documented somewhere?
>
I see some comments in swapfile.c mentioned something related, e.g.
above the definition of swap_avail_heads, the comment mentioned
swap_lock has to be taken before si->lock and swap_avail_lock can be
taken after si->lock is held, but I'm not aware of a place documenting
these things. Documenting these things is useful information I think,
let me see if I can come up with something later.
>
> Did you test this with lockdep fully enabled?
>
>
> I'm thinking that Aaron's a2468cc9bfdff ("swap: choose swap device
> according to numa node") is the appropriate Fixes: target - do you
> agree?
It doesn't appear to be the case. For one thing, the problematic code
that removes the swap device from the avail list without acquiring
si->lock was there before my commit and my commit didn't change that
behaviour. For another, I wanted to see if the problem is still there
without my commit(just to make sure).
I followed Rongwei's description and used stress-ng/swap test together
with some test progs that does memory allocation then MADVISE(pageout)
in a loop to reproduce this problem and I can also see the warning like
below using Linus' master branch as of today, I believe this is the
problem Rongwei described:
[ 1914.518786] ------------[ cut here ]------------
[ 1914.519049] swap_info 9 in list but !SWP_WRITEOK
[ 1914.519274] WARNING: CPU: 14 PID: 14307 at mm/swapfile.c:1085 get_swap_pages+0x3b3/0x440
[ 1914.519660] Modules linked in:
[ 1914.519811] CPU: 14 PID: 14307 Comm: swap Tainted: G W 6.3.0-rc5-00032-g99ddf2254feb #5
[ 1914.520238] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.16.1-2.fc36 04/01/2014
[ 1914.520641] RIP: 0010:get_swap_pages+0x3b3/0x440
[ 1914.520860] Code: 48 8b 4c 24 30 48 c1 e0 3a 4c 09 e0 48 89 01 e8 43 79 96 00 e9 b2 fd ff ff 41 0f be 77 48 48 c7 c78
[ 1914.521709] RSP: 0018:ffffc9000ba0f838 EFLAGS: 00010282
[ 1914.521950] RAX: 0000000000000000 RBX: ffff888154411400 RCX: 0000000000000000
[ 1914.522273] RDX: 0000000000000004 RSI: ffffffff824035cb RDI: 0000000000000001
[ 1914.522601] RBP: ffff888100d95f68 R08: 0000000000000001 R09: 0000000000000003
[ 1914.522926] R10: ffffffff82a7a420 R11: ffffffff82a7a420 R12: 0000000000000350
[ 1914.523249] R13: ffff888100d95da8 R14: ffff888100d95f50 R15: ffff888100d95c00
[ 1914.523576] FS: 00007f23abea2600(0000) GS:ffff88823b600000(0000) knlGS:0000000000000000
[ 1914.523942] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 1914.524206] CR2: 00007f23abbff000 CR3: 0000000104b86004 CR4: 0000000000770ee0
[ 1914.524534] PKRU: 55555554
[ 1914.524661] Call Trace:
[ 1914.524782] <TASK>
[ 1914.524889] folio_alloc_swap+0xde/0x230
[ 1914.525076] add_to_swap+0x36/0xb0
[ 1914.525242] shrink_folio_list+0x9ab/0xef0
[ 1914.525445] reclaim_folio_list+0x70/0x130
[ 1914.525644] reclaim_pages+0x9c/0x1c0
[ 1914.525819] madvise_cold_or_pageout_pte_range+0x79f/0xc80
[ 1914.526073] walk_pgd_range+0x4d8/0x940
[ 1914.526255] ? mt_find+0x15b/0x490
[ 1914.526426] __walk_page_range+0x211/0x230
[ 1914.526619] walk_page_range+0x17a/0x1e0
[ 1914.526807] madvise_pageout+0xef/0x250
And when I reverted my commit on the same branch(needs some manual edits),
the problem is still there.
Another thing is, I noticed Rongwei mentioned "This problem exists in
versions after stable 5.10.y." in the changelog while my commit entered
mainline in v4.14.
So either this problem is always there, i.e. earlier than my commit; or
this problem is indeed only there after v5.10, then it should be something
else that triggered it. My qemu refuses to boot v4.14 kernel so I can
not verify the former yet.
