Skip to contentSkip to conversation

mm, memcg: Do not high throttle allocators based on wraparound

PATCH Applied Chris Down 3 participants 6 messages Mar 31, 2020 – Apr 15, 2020

Subsystems (3)
lore
Share & tools
lore

Chris Down

author
#0 lore raw

PATCH 1/1 mm, memcg: Do not high throttle allocators based on wraparound

From: Jakub Kicinski <[email protected]>

If a cgroup violates its memory.high constraints, we may end up unduly penalising it. For example, for the following hierarchy:

A:   max high, 20 usage
A/B: 9 high, 10 usage
A/C: max high, 10 usage

We would end up doing the following calculation below when calculating high delay for A/B:

A/B: 10 - 9 = 1...
A:   20 - PAGE_COUNTER_MAX = 21, so set max_overage to 21.

This gets worse with higher disparities in usage in the parent.

I have no idea how this disappeared from the final version of the patch, but it is certainly Not Good(tm). This wasn't obvious in testing because, for a simple cgroup hierarchy with only one child, the result is usually roughly the same. It's only in more complex hierarchies that things go really awry (although still, the effects are limited to a maximum of 2 seconds in schedule_timeout_killable at a maximum).

[[email protected]: changelog]

Fixes e26733e0d0ec ("mm, memcg: throttle allocators based on ancestral memory.high")
Cc Johannes Weiner <[email protected]>
---
 mm/memcontrol.c | 3 +++
 1 file changed, 3 insertions(+)
mm/memcontrol.c +3 -0
index eecf003b0c56..75a978307863 100644
@@ -2336,6 +2336,9 @@ static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
 		usage = page_counter_read(&memcg->memory);
 		high = READ_ONCE(memcg->high);
 
+		if (usage <= high)
+			continue;
+
 		/*
 		 * Prevent division by 0 in overage calculation by acting as if
 		 * it was a threshold of 1 page
Signature · 1 line
2.26.0
author
#1 lore raw

Andrew, this is a pretty bad one that could definitely affect memory.high users. We should probably expedite it going in.

Sorry for the trouble, especially on a -stable patch...

maintainer Acked-by on v1 1/1
#2 lore raw
In reply to Chris Down
Quoted 15 lines · Chris Down

From: Jakub Kicinski <[email protected]>

If a cgroup violates its memory.high constraints, we may end up unduly penalising it. For example, for the following hierarchy:

A:   max high, 20 usage
A/B: 9 high, 10 usage
A/C: max high, 10 usage

We would end up doing the following calculation below when calculating high delay for A/B:

A/B: 10 - 9 = 1...
A:   20 - PAGE_COUNTER_MAX = 21, so set max_overage to 21.
Quoted 8 lines

This gets worse with higher disparities in usage in the parent.

I have no idea how this disappeared from the final version of the patch, but it is certainly Not Good(tm). This wasn't obvious in testing because, for a simple cgroup hierarchy with only one child, the result is usually roughly the same. It's only in more complex hierarchies that things go really awry (although still, the effects are limited to a maximum of 2 seconds in schedule_timeout_killable at a maximum).

I find this paragraph rather confusing. This is essentially an unsigned underflow when any of the memcg up the hierarchy is below the high limit, right? There doesn't really seem anything complex in such a hierarchy.

Quoted 7 lines

[[email protected]: changelog]

Fixes e26733e0d0ec ("mm, memcg: throttle allocators based on ancestral memory.high")
Cc Johannes Weiner <[email protected]>

To the patch

Quoted 21 lines
---
 mm/memcontrol.c | 3 +++
 1 file changed, 3 insertions(+)
mm/memcontrol.c +3 -0
index eecf003b0c56..75a978307863 100644
@@ -2336,6 +2336,9 @@ static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
 		usage = page_counter_read(&memcg->memory);
 		high = READ_ONCE(memcg->high);
 
+		if (usage <= high)
+			continue;
+
 		/*
 		 * Prevent division by 0 in overage calculation by acting as if
 		 * it was a threshold of 1 page
Signature · 1 line
2.26.0
Signature · 2 lines
Michal Hocko
SUSE Labs
author
#3 lore raw
Quoted 4 lines · Michal Hocko

I find this paragraph rather confusing. This is essentially an unsigned underflow when any of the memcg up the hierarchy is below the high limit, right? There doesn't really seem anything complex in such a hierarchy.

The conditions to trigger the bug itself are easy, but having it obviously visible in tests requires a moderately complex hierarchy, since in the basic case ancestor_usage is "similar enough" to the test leaf cgroup's usage.

author
#4 lore raw
Quoted 7 lines · Chris Down
Quoted 4 lines · Michal Hocko

I find this paragraph rather confusing. This is essentially an unsigned underflow when any of the memcg up the hierarchy is below the high limit, right? There doesn't really seem anything complex in such a hierarchy.

The conditions to trigger the bug itself are easy, but having it obviously visible in tests requires a moderately complex hierarchy, since in the basic case ancestor_usage is "similar enough" to the test leaf cgroup's usage.

Here is another reason why this wasn't caught -- division usually renders the overage 0 anyway with such a large input.

With the attached patch applied before this fix, you can see that usually division results in an overage of 0, so the result is the same. Here's an example where pid 213 is a cgroup in system.slice/foo.service hitting its own memory.high, and system.slice has no memory.high configuresd:

	[root@ktst ~]# cat /sys/kernel/debug/tracing/trace
	# tracer: nop
	#
	# entries-in-buffer/entries-written: 33/33   #P:4
	#
	#                              _-----=> irqs-off
	#                             / _----=> need-resched
	#                            | / _---=> hardirq/softirq
	#                            || / _--=> preempt-depth
	#                            ||| /     delay
	#           TASK-PID   CPU#  ||||    TIMESTAMP  FUNCTION
	#              | |       |   ||||       |         |
		(bash)-213   [002] .N..    58.873988: mem_cgroup_handle_over_high: usage: 32, high: 1
		(bash)-213   [002] .N..    58.873993: mem_cgroup_handle_over_high: 1 overage before shifting (31)
		(bash)-213   [002] .N..    58.873994: mem_cgroup_handle_over_high: 1 overage after shifting (32505856)
		(bash)-213   [002] .N..    58.873995: mem_cgroup_handle_over_high: 1 overage after div (32505856)
		(bash)-213   [002] .N..    58.873996: mem_cgroup_handle_over_high: 1 cgroup new overage (32505856)
		(bash)-213   [002] .N..    58.873998: mem_cgroup_handle_over_high: usage: 18641, high: 2251799813685247
		(bash)-213   [002] .N..    58.873998: mem_cgroup_handle_over_high: 2 overage before shifting (18444492273895885010)
		(bash)-213   [002] .N..    58.873999: mem_cgroup_handle_over_high: 2 overage after shifting (19547553792)
		(bash)-213   [002] .N..    58.874000: mem_cgroup_handle_over_high: 2 overage after div (0)
		(bash)-213   [002] .N..    58.874001: mem_cgroup_handle_over_high: 2 cgroup too low (0)
		(bash)-213   [002] .N..    58.874002: mem_cgroup_handle_over_high: Used 1 from leaf to get result

From df96928bc8d482d8b26c277c4ca0b075783c7aed Mon Sep 17 00:00:00 2001
From: Chris Down <[email protected]>
Date: Tue, 31 Mar 2020 19:16:23 +0100
Subject: [PATCH] temp

---
 mm/memcontrol.c | 16 +++++++++++++++-
 1 file changed, 15 insertions(+), 1 deletion(-)
mm/memcontrol.c +15 -1
index eecf003b0c56..c33e317c3667 100644
@@ -2328,11 +2328,14 @@ static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
 {
 	unsigned long penalty_jiffies;
 	u64 max_overage = 0;
+	int i = 0, i_overage = 0;
 
 	do {
 		unsigned long usage, high;
 		u64 overage;
 
+		i++;
+
 		usage = page_counter_read(&memcg->memory);
 		high = READ_ONCE(memcg->high);
 
@@ -2342,18 +2345,29 @@ static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
 		 */
 		high = max(high, 1UL);
 
+		trace_printk("usage: %lu, high: %lu\n", usage, high);
 		overage = usage - high;
+		trace_printk("%d overage before shifting (%llu)\n", i, overage);
 		overage <<= MEMCG_DELAY_PRECISION_SHIFT;
+		trace_printk("%d overage after shifting (%llu)\n", i, overage);
 		overage = div64_u64(overage, high);
+		trace_printk("%d overage after div (%llu)\n", i, overage);
 
-		if (overage > max_overage)
+		if (overage > max_overage) {
+			trace_printk("%d cgroup new overage (%llu)\n", i, overage);
+			i_overage = i;
 			max_overage = overage;
+		} else {
+			trace_printk("%d cgroup too low (%llu)\n", i, overage);
+		}
 	} while ((memcg = parent_mem_cgroup(memcg)) &&
 		 !mem_cgroup_is_root(memcg));
 
 	if (!max_overage)
 		return 0;
 
+	trace_printk("Used %d from leaf to get result\n", i_overage);
+
 	/*
 	 * We use overage compared to memory.high to calculate the number of
 	 * jiffies to sleep (penalty_jiffies). Ideally this value should be
Signature · 1 line
2.26.0
maintainer Acked-by on v1 1/1
#5 lore raw
In reply to Chris Down
Quoted 24 lines · Chris Down

From: Jakub Kicinski <[email protected]>

If a cgroup violates its memory.high constraints, we may end up unduly penalising it. For example, for the following hierarchy:

A:   max high, 20 usage
A/B: 9 high, 10 usage
A/C: max high, 10 usage

We would end up doing the following calculation below when calculating high delay for A/B:

A/B: 10 - 9 = 1...
A:   20 - PAGE_COUNTER_MAX = 21, so set max_overage to 21.

This gets worse with higher disparities in usage in the parent.

I have no idea how this disappeared from the final version of the patch, but it is certainly Not Good(tm). This wasn't obvious in testing because, for a simple cgroup hierarchy with only one child, the result is usually roughly the same. It's only in more complex hierarchies that things go really awry (although still, the effects are limited to a maximum of 2 seconds in schedule_timeout_killable at a maximum).

Quoted 7 lines

[[email protected]: changelog]

Fixes e26733e0d0ec ("mm, memcg: throttle allocators based on ancestral memory.high")
Cc Johannes Weiner <[email protected]>

Oops.