Re: [PATCH v7 00/10] per lruvec lru_lock for memcg

[Hugh Dickins <hughd-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org>]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 1804 bytes --]

On Fri, 10 Jan 2020, Alex Shi wrote:
> 在 2020/1/2 下午6:21, Alex Shi 写道:
> > 在 2020/1/1 上午7:05, Andrew Morton 写道:
> >> On Wed, 25 Dec 2019 17:04:16 +0800 Alex Shi <alex.shi-KPsoFbNs7GizrGE5bRqYAg@public.gmane.orgm> wrote:
> >>
> >>> This patchset move lru_lock into lruvec, give a lru_lock for each of
> >>> lruvec, thus bring a lru_lock for each of memcg per node.
> >>
> >> I see that there has been plenty of feedback on previous versions, but
> >> no acked/reviewed tags as yet.
> >>
> >> I think I'll take a pass for now, see what the audience feedback looks
> >> like ;)
> >>
> > 
> 
> Hi Johannes,
> 
> Any comments of this version? :)

I (Hugh) tried to test it on v5.5-rc5, but did not get very far at all -
perhaps because my particular interest tends towards tmpfs and swap,
and swap always made trouble for lruvec lock - one of the reasons why
our patches were more complicated than you thought necessary.

Booted a smallish kernel in mem=700M with 1.5G of swap, with intention
of running small kernel builds in tmpfs and in ext4-on-loop-on-tmpfs
(losetup was the last command started but I doubt it played much part):

mount -t tmpfs -o size=470M tmpfs /tst
cp /dev/zero /tst
losetup /dev/loop0 /tst/zero

and kernel crashed on the

VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != page->mem_cgroup, page);
kernel BUG at mm/memcontrol.c:1268!
lock_page_lruvec_irqsave
relock_page_lruvec_irqsave
pagevec_lru_move_fn
__pagevec_lru_add
lru_add_drain_cpu
lru_add_drain
swap_cluster_readahead
shmem_swapin
shmem_swapin_page
shmem_getpage_gfp
shmem_getpage
shmem_write_begin
generic_perform_write
__generic_file_write_iter
generic_file_write_iter
new_sync_write
__vfs_write
vfs_write
ksys_write
__x86_sys_write
do_syscall_64

Hugh