[PATCH STABLE 4.4-5.3] mm, vmstat: drop zone->lock in /proc/pagetypeinfo

[PATCH STABLE 4.4-5.3] mm, vmstat: drop zone->lock in /proc/pagetypeinfo

[Stephen Brennan]

Commit 93b3a674485f6a4b8ffff85d1682d5e8b7c51560 upstream

Commit 93b3a674485f ("mm,vmstat: reduce zone->lock holding time by
/proc/pagetypeinfo") upstream caps the number of iterations over each
free_list at 100,000, and also drops the zone->lock in between each
migrate type. Capping the iteration count alters the file contents in
some cases, which means this approach may not be suitable for stable
backports.

However, dropping zone->lock in between migrate types (and, as a result,
page orders) will not change the /proc/pagetypeinfo file contents. It
can significantly reduce the length of time spent with IRQs disabled,
which can prevent missed interrupts or soft lockups which we have
observed on systems with particularly large memory.

Thus, this commit is a modified version of the upstream one which only
drops the lock in between migrate types.

Fixes: 467c996c1e19 ("Print out statistics in relation to fragmentation avoidance to /proc/pagetypeinfo")
Signed-off-by: Stephen Brennan <stephen.s.brennan@oracle.com>
Reviewed-by: Aruna Ramakrishna <aruna.ramakrishna@oracle.com>
Reviewed-by: Khalid Aziz <khalid.aziz@oracle.com>
---

mm/vmstat.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/mm/vmstat.c b/mm/vmstat.c
index 13b74c4314a7e..663069cf7724a 100644
--- a/mm/vmstat.c
+++ b/mm/vmstat.c
@@ -1316,6 +1316,9 @@ static void pagetypeinfo_showfree_print(struct seq_file *m,
 			list_for_each(curr, &area->free_list[mtype])
 				freecount++;
 			seq_printf(m, "%6lu ", freecount);
+			spin_unlock_irq(&zone->lock);
+			cond_resched();
+			spin_lock_irq(&zone->lock);
 		}
 		seq_putc(m, '\n');
 	}
-- 
2.27.0

Re: [PATCH STABLE 4.4-5.3] mm, vmstat: drop zone->lock in /proc/pagetypeinfo

[Greg KH]

On Wed, May 26, 2021 at 10:46:13AM -0700, Stephen Brennan wrote:
> Commit 93b3a674485f6a4b8ffff85d1682d5e8b7c51560 upstream
> 
> Commit 93b3a674485f ("mm,vmstat: reduce zone->lock holding time by
> /proc/pagetypeinfo") upstream caps the number of iterations over each
> free_list at 100,000, and also drops the zone->lock in between each
> migrate type. Capping the iteration count alters the file contents in
> some cases, which means this approach may not be suitable for stable
> backports.
> 
> However, dropping zone->lock in between migrate types (and, as a result,
> page orders) will not change the /proc/pagetypeinfo file contents. It
> can significantly reduce the length of time spent with IRQs disabled,
> which can prevent missed interrupts or soft lockups which we have
> observed on systems with particularly large memory.
> 
> Thus, this commit is a modified version of the upstream one which only
> drops the lock in between migrate types.

Now queued up, thanks.

greg k-h