Re: Memory hotplug softlock issue

[Baoquan He]

On 11/15/18 at 08:30am, Michal Hocko wrote:
> On Thu 15-11-18 13:10:34, Baoquan He wrote:
> > On 11/14/18 at 04:00pm, Michal Hocko wrote:
> > > On Wed 14-11-18 22:52:50, Baoquan He wrote:
> > > > On 11/14/18 at 10:01am, Michal Hocko wrote:
> > > > > I have seen an issue when the migration cannot make a forward progress
> > > > > because of a glibc page with a reference count bumping up and down. Most
> > > > > probable explanation is the faultaround code. I am working on this and
> > > > > will post a patch soon. In any case the migration should converge and if
> > > > > it doesn't do then there is a bug lurking somewhere.
> > > > > 
> > > > > Failing on ENOMEM is a questionable thing. I haven't seen that happening
> > > > > wildly but if it is a case then I wouldn't be opposed.
> > > > 
> > > > Applied your debugging patches, it helps a lot to printing message.
> > > > 
> > > > Below is the dmesg log about the migrating failure. It can't pass
> > > > migrate_pages() and loop forever.
> > > > 
> > > > [  +0.083841] migrating pfn 10fff7d0 failed 
> > > > [  +0.000005] page:ffffea043ffdf400 count:208 mapcount:201 mapping:ffff888dff4bdda8 index:0x2
> > > > [  +0.012689] xfs_address_space_operations [xfs] 
> > > > [  +0.000030] name:"stress" 
> > > > [  +0.004556] flags: 0x5fffffc0000004(uptodate)
> > > > [  +0.007339] raw: 005fffffc0000004 ffffc900000e3d80 ffffc900000e3d80 ffff888dff4bdda8
> > > > [  +0.009488] raw: 0000000000000002 0000000000000000 000000cb000000c8 ffff888e7353d000
> > > > [  +0.007726] page->mem_cgroup:ffff888e7353d000
> > > > [  +0.084538] migrating pfn 10fff7d0 failed 
> > > > [  +0.000006] page:ffffea043ffdf400 count:210 mapcount:201 mapping:ffff888dff4bdda8 index:0x2
> > > > [  +0.012798] xfs_address_space_operations [xfs] 
> > > > [  +0.000034] name:"stress" 
> > > > [  +0.004524] flags: 0x5fffffc0000004(uptodate)
> > > > [  +0.007068] raw: 005fffffc0000004 ffffc900000e3d80 ffffc900000e3d80 ffff888dff4bdda8
> > > > [  +0.009359] raw: 0000000000000002 0000000000000000 000000cb000000c8 ffff888e7353d000
> > > > [  +0.007728] page->mem_cgroup:ffff888e7353d000
> > > 
> > > I wouldn't be surprised if this was a similar/same issue I've been
> > > chasing recently. Could you try to disable faultaround to see if that
> > > helps. It seems that it helped in my particular case but I am still
> > > waiting for the final good-to-go to post the patch as I do not own the
> > > workload which triggered that issue.
> > 
> > Tried, still stuck in last block sometime. Usually after several times
> > of hotplug/unplug. If stop stress program, the last block will be
> > offlined immediately.
> 
> Is the pattern still the same? I mean failing over few pages with
> reference count jumping up and down between attempts?

->count jumping up and down, mapcount stays the same value.

> 
> > [root@ ~]# cat /sys/kernel/debug/fault_around_bytes 
> > 4096
> 
> Can you make it 0?

I executed 'echo 0 > fault_around_bytes', value less than one page size
will round up to one page.

/*
 * fault_around_bytes must be rounded down to the nearest page order as it's
 * what do_fault_around() expects to see.
 */
static int fault_around_bytes_set(void *data, u64 val)
{
        if (val / PAGE_SIZE > PTRS_PER_PTE)
                return -EINVAL;
        if (val > PAGE_SIZE)
                fault_around_bytes = rounddown_pow_of_two(val);
        else
                fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
        return 0;
}