Re: mm: mkfs.ext4 invoked oom-killer on i386 - pagecache_get_page

[Hugh Dickins]

On Thu, 21 May 2020, Johannes Weiner wrote:
> On Thu, May 21, 2020 at 01:06:28PM -0700, Hugh Dickins wrote:
> > On Thu, 21 May 2020, Johannes Weiner wrote:
> > > do_memsw_account() used to be automatically false when the cgroup
> > > controller was disabled. Now that it's replaced by
> > > cgroup_memory_noswap, for which this isn't true, make the
> > > mem_cgroup_disabled() checks explicit in the swap control API.
> > > 
> > > [hannes@xxxxxxxxxxx: use mem_cgroup_disabled() in all API functions]
> > > Reported-by: Naresh Kamboju <naresh.kamboju@xxxxxxxxxx>
> > > Debugged-by: Hugh Dickins <hughd@xxxxxxxxxx>
> > > Debugged-by: Michal Hocko <mhocko@xxxxxxxxxx>
> > > Signed-off-by: Johannes Weiner <hannes@xxxxxxxxxxx>
> > > ---
> > >  mm/memcontrol.c | 47 +++++++++++++++++++++++++++++++++++++++++------
> > >  1 file changed, 41 insertions(+), 6 deletions(-)
> > 
> > I'm certainly not against a mem_cgroup_disabled() check in the only
> > place that's been observed to need it, as a fixup to merge into your
> > original patch; but this seems rather an over-reaction - and I'm a
> > little surprised that setting mem_cgroup_disabled() doesn't just
> > force cgroup_memory_noswap, saving repetitious checks elsewhere
> > (perhaps there's a difficulty in that, I haven't looked).
> 
> Fair enough, I changed it to set the flag at initialization time if
> mem_cgroup_disabled(). I was never a fan of the old flags, where it
> was never clear what was commandline, and what was internal runtime
> state - do_swap_account? really_do_swap_account? But I think it's
> straight-forward in this case now.
> 
> > Historically, I think we've added mem_cgroup_disabled() checks
> > (accessing a cacheline we'd rather avoid) where they're necessary,
> > rather than at every "interface".
> 
> To me that always seemed like bugs waiting to happen. Like this one!
> 
> It's a jump label nowadays, so I've been liberal with these to avoid
> subtle bugs.
> 
> > And you seem to be in a very "goto out" mood today - we all have
> > our "goto out" days, alternating with our "return 0" days :)
> 
> :-)
> 
> But I agree, best to keep this fixup self-contained and defer anything
> else to separate cleanup patches.
> 
> How about the below? It survives a swaptest with cgroup_disable=memory
> for me.

I like this version *a lot*, thank you. I got worried for a bit by
the "#define cgroup_memory_noswap 1" when #ifndef CONFIG_MEMCG_SWAP,
but now realize that fits perfectly.

> 
> Hugh, I started with your patch, which is why I kept you as the
> author, but as the patch now (and arguably the previous one) is
> sufficiently different, I dropped that now. I hope that's okay.

Absolutely okay, these are yours: I was a little uncomfortable to
see me on the From line before, but it also seemed just too petty
to insist that my name be removed.

(By the way, off-topic for this particular issue, but advance warning
that I hope to post a couple of patches to __read_swap_cache_async()
before the end of the day, first being fixup to some of your mods -
I suspect you got it working well enough, and intended to come back
to check a few details later, but never quite got around to that.)

> 
> ---
> From d9e7ed15d1c9248a3fd99e35e82437549154dac7 Mon Sep 17 00:00:00 2001
> From: Johannes Weiner <hannes@xxxxxxxxxxx>
> Date: Thu, 21 May 2020 17:44:25 -0400
> Subject: [PATCH] mm: memcontrol: prepare swap controller setup for integration
>  fix
> 
> Fix crash with cgroup_disable=memory:
> 
> > > > > + mkfs -t ext4 /dev/disk/by-id/ata-TOSHIBA_MG04ACA100N_Y8NRK0BPF6XF
> > > > > mke2fs 1.43.8 (1-Jan-2018)
> > > > > Creating filesystem with 244190646 4k blocks and 61054976 inodes
> > > > > Filesystem UUID: 3bb1a285-2cb4-44b4-b6e8-62548f3ac620
> > > > > Superblock backups stored on blocks:
> > > > > 32768, 98304, 163840, 229376, 294912, 819200, 884736, 1605632, 2654208,
> > > > > 4096000, 7962624, 11239424, 20480000, 23887872, 71663616, 78675968,
> > > > > 102400000, 214990848
> > > > > Allocating group tables:    0/7453                           done
> > > > > Writing inode tables:    0/7453                           done
> > > > > Creating journal (262144 blocks): [   35.502102] BUG: kernel NULL
> > > > > pointer dereference, address: 000000c8
> > > > > [   35.508372] #PF: supervisor read access in kernel mode
> > > > > [   35.513506] #PF: error_code(0x0000) - not-present page
> > > > > [   35.518638] *pde = 00000000
> > > > > [   35.521514] Oops: 0000 [#1] SMP
> > > > > [   35.524652] CPU: 0 PID: 145 Comm: kswapd0 Not tainted
> > > > > 5.7.0-rc6-next-20200519+ #1
> > > > > [   35.532121] Hardware name: Supermicro SYS-5019S-ML/X11SSH-F, BIOS
> > > > > 2.2 05/23/2018
> > > > > [   35.539507] EIP: mem_cgroup_get_nr_swap_pages+0x28/0x60
> 
> Swap accounting used to be implied-disabled when the cgroup controller
> was disabled. Restore that for the new cgroup_memory_noswap, so that
> we bail out of this function instead of dereferencing a NULL memcg.
> 
> Reported-by: Naresh Kamboju <naresh.kamboju@xxxxxxxxxx>
> Debugged-by: Hugh Dickins <hughd@xxxxxxxxxx>
> Debugged-by: Michal Hocko <mhocko@xxxxxxxxxx>
> Signed-off-by: Johannes Weiner <hannes@xxxxxxxxxxx>

Acked-by: Hugh Dickins <hughd@xxxxxxxxxx>

> ---
>  mm/memcontrol.c | 6 +++++-
>  1 file changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index 3e000a316b59..e3b785d6e771 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -7075,7 +7075,11 @@ static struct cftype memsw_files[] = {
>  
>  static int __init mem_cgroup_swap_init(void)
>  {
> -	if (mem_cgroup_disabled() || cgroup_memory_noswap)
> +	/* No memory control -> no swap control */
> +	if (mem_cgroup_disabled())
> +		cgroup_memory_noswap = true;
> +
> +	if (cgroup_memory_noswap)
>  		return 0;
>  
>  	WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
> -- 
> 2.26.2