Re: Memory hotplug softlock issue

[Michal Hocko]

On Mon 19-11-18 13:51:21, Michal Hocko wrote:
> On Mon 19-11-18 13:40:33, Michal Hocko wrote:
> > On Mon 19-11-18 18:52:02, Baoquan He wrote:
> > [...]
> > 
> > There are few stacks directly in the offline path but those should be
> > OK.
> > The real culprit seems to be the swap in code
> > 
> > > [  +1.734416] CPU: 255 PID: 5558 Comm: stress Tainted: G             L    4.20.0-rc2+ #7
> > > [  +0.007927] Hardware name:  9008/IT91SMUB, BIOS BLXSV512 03/22/2018
> > > [  +0.006297] Call Trace:
> > > [  +0.002537]  dump_stack+0x46/0x60
> > > [  +0.003386]  __migration_entry_wait.cold.65+0x5/0x14
> > > [  +0.005043]  do_swap_page+0x84e/0x960
> > > [  +0.003727]  ? arch_tlb_finish_mmu+0x29/0xc0
> > > [  +0.006412]  __handle_mm_fault+0x933/0x1330
> > > [  +0.004265]  handle_mm_fault+0xc4/0x250
> > > [  +0.003915]  __do_page_fault+0x2b7/0x510
> > > [  +0.003990]  do_page_fault+0x2c/0x110
> > > [  +0.003729]  ? page_fault+0x8/0x30
> > > [  +0.003462]  page_fault+0x1e/0x30
> > 
> > There are many traces to this path. We are 
> > 	/*
> > 	 * Once page cache replacement of page migration started, page_count
> > 	 * *must* be zero. And, we don't want to call wait_on_page_locked()
> > 	 * against a page without get_page().
> > 	 * So, we use get_page_unless_zero(), here. Even failed, page fault
> > 	 * will occur again.
> > 	 */
> > 	if (!get_page_unless_zero(page))
> > 		goto out;
> > 	pte_unmap_unlock(ptep, ptl);
> > 	wait_on_page_locked(page);
> > 
> > taking a reference to the page under the migration. I have to think
> > about this much more but I suspec this is just calling for a problem.
> > 
> > Cc migration experts. For you background information. We are seeing
> > memory offline not being able to converge because few heavily used pages
> > fail to migrate away - e.g. http://lkml.kernel.org/r/20181116012433.GU2653@MiWiFi-R3L-srv
> > A debugging page to dump stack for these pages http://lkml.kernel.org/r/20181116091409.GD14706@xxxxxxxxxxxxxx
> > shows that references are taken from the swap in code (above). How are
> > we supposed to converge when the swapin code waits for the migration to
> > finish with the reference count elevated?
> 
> Just to clarify. This is not only about swapin obviously. Any caller of
> __migration_entry_wait is affected the same way AFAICS.

In other words. Why cannot we do the following?

diff --git a/mm/migrate.c b/mm/migrate.c
index f7e4bfdc13b7..7ccab29bcf9a 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -324,19 +324,9 @@ void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
 		goto out;
 
 	page = migration_entry_to_page(entry);
-
-	/*
-	 * Once page cache replacement of page migration started, page_count
-	 * *must* be zero. And, we don't want to call wait_on_page_locked()
-	 * against a page without get_page().
-	 * So, we use get_page_unless_zero(), here. Even failed, page fault
-	 * will occur again.
-	 */
-	if (!get_page_unless_zero(page))
-		goto out;
 	pte_unmap_unlock(ptep, ptl);
-	wait_on_page_locked(page);
-	put_page(page);
+	page_lock(page);
+	page_unlock(page);
 	return;
 out:
 	pte_unmap_unlock(ptep, ptl);
-- 
Michal Hocko
SUSE Labs