This series attempts to address shortages in today's approach for memcg
stats flushing, namely occasionally stale or expensive stat reads. The
series does so by changing the threshold that we use to decide whether
to trigger a flush to be per memcg instead of global (patch 3), and then
changing flushing to be per memcg (i.e. subtree flushes) instead of
global (patch 5).
Patch 3 & 5 are the core of the series, and they include more details
and testing results. The rest are either cleanups or prep work.
This series replaces the "memcg: more sophisticated stats flushing"
series [1], which also replaces another series, in a long list of
attempts to improve memcg stats flushing. It is not a new version of
the same patchset as it is a completely different approach. This is
based on collected feedback from discussions on lkml in all previous
attempts. Hopefully, this is the final attempt.
There was a reported regression in v2 [2] for will-it-scale::fallocate
benchmark. I believe this regression should not affect production
workloads. This specific benchmark is allocating and freeing memory
(using fallocate/ftruncate) at a rate that is much faster to make actual
use of the memory. Testing this series on 100+ machines running
production workloads did not show any practical regressions in page
fault latency or allocation latency, but it showed great improvements in
stats read time. I do not have numbers about the exact improvements for
this series, but combined with another optimization for cgroup v1 [3] we
see 5-10x improvements. A significant chunk of that is coming from the
cgroup v1 optimization, but this series also made an improvement as
reported by Domenico [4].
v3 -> v4:
- Rebased on top of mm-unstable + "workload-specific and memory
pressure-driven zswap writeback" series to fix conflicts [5].
v3: https://lore.kernel.org/all/[email protected]/
[1]https://lore.kernel.org/lkml/[email protected]/
[2]https://lore.kernel.org/lkml/[email protected]/
[3]https://lore.kernel.org/lkml/[email protected]/
[4]https://lore.kernel.org/lkml/CAFYChMv_kv_KXOMRkrmTN-7MrfgBHMcK3YXv0dPYEL7nK77e2A@mail.gmail.com/
[5]https://lore.kernel.org/all/[email protected]/
Yosry Ahmed (5):
mm: memcg: change flush_next_time to flush_last_time
mm: memcg: move vmstats structs definition above flushing code
mm: memcg: make stats flushing threshold per-memcg
mm: workingset: move the stats flush into workingset_test_recent()
mm: memcg: restore subtree stats flushing
include/linux/memcontrol.h | 8 +-
mm/memcontrol.c | 272 +++++++++++++++++++++----------------
mm/vmscan.c | 2 +-
mm/workingset.c | 42 ++++--
4 files changed, 188 insertions(+), 136 deletions(-)
--
2.43.0.rc1.413.gea7ed67945-goog
The workingset code flushes the stats in workingset_refault() to get
accurate stats of the eviction memcg. In preparation for more scoped
flushed and passing the eviction memcg to the flush call, move the call
to workingset_test_recent() where we have a pointer to the eviction
memcg.
The flush call is sleepable, and cannot be made in an rcu read section.
Hence, minimize the rcu read section by also moving it into
workingset_test_recent(). Furthermore, instead of holding the rcu read
lock throughout workingset_test_recent(), only hold it briefly to get a
ref on the eviction memcg. This allows us to make the flush call after
we get the eviction memcg.
As for workingset_refault(), nothing else there appears to be protected
by rcu. The memcg of the faulted folio (which is not necessarily the
same as the eviction memcg) is protected by the folio lock, which is
held from all callsites. Add a VM_BUG_ON() to make sure this doesn't
change from under us.
No functional change intended.
Signed-off-by: Yosry Ahmed <[email protected]>
Tested-by: Domenico Cerasuolo <[email protected]>
---
mm/workingset.c | 36 ++++++++++++++++++++++++------------
1 file changed, 24 insertions(+), 12 deletions(-)
diff --git a/mm/workingset.c b/mm/workingset.c
index c17d45c6f29b0..dce41577a49d2 100644
--- a/mm/workingset.c
+++ b/mm/workingset.c
@@ -425,8 +425,16 @@ bool workingset_test_recent(void *shadow, bool file, bool *workingset)
struct pglist_data *pgdat;
unsigned long eviction;
- if (lru_gen_enabled())
- return lru_gen_test_recent(shadow, file, &eviction_lruvec, &eviction, workingset);
+ rcu_read_lock();
+
+ if (lru_gen_enabled()) {
+ bool recent = lru_gen_test_recent(shadow, file,
+ &eviction_lruvec, &eviction, workingset);
+
+ rcu_read_unlock();
+ return recent;
+ }
+
unpack_shadow(shadow, &memcgid, &pgdat, &eviction, workingset);
eviction <<= bucket_order;
@@ -448,8 +456,16 @@ bool workingset_test_recent(void *shadow, bool file, bool *workingset)
* configurations instead.
*/
eviction_memcg = mem_cgroup_from_id(memcgid);
- if (!mem_cgroup_disabled() && !eviction_memcg)
+ if (!mem_cgroup_disabled() &&
+ (!eviction_memcg || !mem_cgroup_tryget(eviction_memcg))) {
+ rcu_read_unlock();
return false;
+ }
+
+ rcu_read_unlock();
+
+ /* Flush stats (and potentially sleep) outside the RCU read section */
+ mem_cgroup_flush_stats_ratelimited();
eviction_lruvec = mem_cgroup_lruvec(eviction_memcg, pgdat);
refault = atomic_long_read(&eviction_lruvec->nonresident_age);
@@ -493,6 +509,7 @@ bool workingset_test_recent(void *shadow, bool file, bool *workingset)
}
}
+ mem_cgroup_put(eviction_memcg);
return refault_distance <= workingset_size;
}
@@ -519,19 +536,16 @@ void workingset_refault(struct folio *folio, void *shadow)
return;
}
- /* Flush stats (and potentially sleep) before holding RCU read lock */
- mem_cgroup_flush_stats_ratelimited();
-
- rcu_read_lock();
-
/*
* The activation decision for this folio is made at the level
* where the eviction occurred, as that is where the LRU order
* during folio reclaim is being determined.
*
* However, the cgroup that will own the folio is the one that
- * is actually experiencing the refault event.
+ * is actually experiencing the refault event. Make sure the folio is
+ * locked to guarantee folio_memcg() stability throughout.
*/
+ VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
nr = folio_nr_pages(folio);
memcg = folio_memcg(folio);
pgdat = folio_pgdat(folio);
@@ -540,7 +554,7 @@ void workingset_refault(struct folio *folio, void *shadow)
mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + file, nr);
if (!workingset_test_recent(shadow, file, &workingset))
- goto out;
+ return;
folio_set_active(folio);
workingset_age_nonresident(lruvec, nr);
@@ -556,8 +570,6 @@ void workingset_refault(struct folio *folio, void *shadow)
lru_note_cost_refault(folio);
mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + file, nr);
}
-out:
- rcu_read_unlock();
}
/**
--
2.43.0.rc1.413.gea7ed67945-goog
flush_next_time is an inaccurate name. It's not the next time that
periodic flushing will happen, it's rather the next time that
ratelimited flushing can happen if the periodic flusher is late.
Simplify its semantics by just storing the timestamp of the last flush
instead, flush_last_time. Move the 2*FLUSH_TIME addition to
mem_cgroup_flush_stats_ratelimited(), and add a comment explaining it.
This way, all the ratelimiting semantics live in one place.
No functional change intended.
Signed-off-by: Yosry Ahmed <[email protected]>
Tested-by: Domenico Cerasuolo <[email protected]>
Acked-by: Shakeel Butt <[email protected]>
Acked-by: Chris Li <[email protected]> (Google)
---
mm/memcontrol.c | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index f88c8fd036897..61435bd037cb4 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -593,7 +593,7 @@ static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork);
static DEFINE_PER_CPU(unsigned int, stats_updates);
static atomic_t stats_flush_ongoing = ATOMIC_INIT(0);
static atomic_t stats_flush_threshold = ATOMIC_INIT(0);
-static u64 flush_next_time;
+static u64 flush_last_time;
#define FLUSH_TIME (2UL*HZ)
@@ -653,7 +653,7 @@ static void do_flush_stats(void)
atomic_xchg(&stats_flush_ongoing, 1))
return;
- WRITE_ONCE(flush_next_time, jiffies_64 + 2*FLUSH_TIME);
+ WRITE_ONCE(flush_last_time, jiffies_64);
cgroup_rstat_flush(root_mem_cgroup->css.cgroup);
@@ -669,7 +669,8 @@ void mem_cgroup_flush_stats(void)
void mem_cgroup_flush_stats_ratelimited(void)
{
- if (time_after64(jiffies_64, READ_ONCE(flush_next_time)))
+ /* Only flush if the periodic flusher is one full cycle late */
+ if (time_after64(jiffies_64, READ_ONCE(flush_last_time) + 2*FLUSH_TIME))
mem_cgroup_flush_stats();
}
--
2.43.0.rc1.413.gea7ed67945-goog
A global counter for the magnitude of memcg stats update is maintained
on the memcg side to avoid invoking rstat flushes when the pending
updates are not significant. This avoids unnecessary flushes, which are
not very cheap even if there isn't a lot of stats to flush. It also
avoids unnecessary lock contention on the underlying global rstat lock.
Make this threshold per-memcg. The scheme is followed where percpu (now
also per-memcg) counters are incremented in the update path, and only
propagated to per-memcg atomics when they exceed a certain threshold.
This provides two benefits:
(a) On large machines with a lot of memcgs, the global threshold can be
reached relatively fast, so guarding the underlying lock becomes less
effective. Making the threshold per-memcg avoids this.
(b) Having a global threshold makes it hard to do subtree flushes, as we
cannot reset the global counter except for a full flush. Per-memcg
counters removes this as a blocker from doing subtree flushes, which
helps avoid unnecessary work when the stats of a small subtree are
needed.
Nothing is free, of course. This comes at a cost:
(a) A new per-cpu counter per memcg, consuming NR_CPUS * NR_MEMCGS * 4
bytes. The extra memory usage is insigificant.
(b) More work on the update side, although in the common case it will
only be percpu counter updates. The amount of work scales with the
number of ancestors (i.e. tree depth). This is not a new concept, adding
a cgroup to the rstat tree involves a parent loop, so is charging.
Testing results below show no significant regressions.
(c) The error margin in the stats for the system as a whole increases
from NR_CPUS * MEMCG_CHARGE_BATCH to NR_CPUS * MEMCG_CHARGE_BATCH *
NR_MEMCGS. This is probably fine because we have a similar per-memcg
error in charges coming from percpu stocks, and we have a periodic
flusher that makes sure we always flush all the stats every 2s anyway.
This patch was tested to make sure no significant regressions are
introduced on the update path as follows. The following benchmarks were
ran in a cgroup that is 2 levels deep (/sys/fs/cgroup/a/b/):
(1) Running 22 instances of netperf on a 44 cpu machine with
hyperthreading disabled. All instances are run in a level 2 cgroup, as
well as netserver:
# netserver -6
# netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
Averaging 20 runs, the numbers are as follows:
Base: 40198.0 mbps
Patched: 38629.7 mbps (-3.9%)
The regression is minimal, especially for 22 instances in the same
cgroup sharing all ancestors (so updating the same atomics).
(2) will-it-scale page_fault tests. These tests (specifically
per_process_ops in page_fault3 test) detected a 25.9% regression before
for a change in the stats update path [1]. These are the
numbers from 10 runs (+ is good) on a machine with 256 cpus:
LABEL | MEAN | MEDIAN | STDDEV |
------------------------------+-------------+-------------+-------------
page_fault1_per_process_ops | | | |
(A) base | 270249.164 | 265437.000 | 13451.836 |
(B) patched | 261368.709 | 255725.000 | 13394.767 |
| -3.29% | -3.66% | |
page_fault1_per_thread_ops | | | |
(A) base | 242111.345 | 239737.000 | 10026.031 |
(B) patched | 237057.109 | 235305.000 | 9769.687 |
| -2.09% | -1.85% | |
page_fault1_scalability | | |
(A) base | 0.034387 | 0.035168 | 0.0018283 |
(B) patched | 0.033988 | 0.034573 | 0.0018056 |
| -1.16% | -1.69% | |
page_fault2_per_process_ops | | |
(A) base | 203561.836 | 203301.000 | 2550.764 |
(B) patched | 197195.945 | 197746.000 | 2264.263 |
| -3.13% | -2.73% | |
page_fault2_per_thread_ops | | |
(A) base | 171046.473 | 170776.000 | 1509.679 |
(B) patched | 166626.327 | 166406.000 | 768.753 |
| -2.58% | -2.56% | |
page_fault2_scalability | | |
(A) base | 0.054026 | 0.053821 | 0.00062121 |
(B) patched | 0.053329 | 0.05306 | 0.00048394 |
| -1.29% | -1.41% | |
page_fault3_per_process_ops | | |
(A) base | 1295807.782 | 1297550.000 | 5907.585 |
(B) patched | 1275579.873 | 1273359.000 | 8759.160 |
| -1.56% | -1.86% | |
page_fault3_per_thread_ops | | |
(A) base | 391234.164 | 390860.000 | 1760.720 |
(B) patched | 377231.273 | 376369.000 | 1874.971 |
| -3.58% | -3.71% | |
page_fault3_scalability | | |
(A) base | 0.60369 | 0.60072 | 0.0083029 |
(B) patched | 0.61733 | 0.61544 | 0.009855 |
| +2.26% | +2.45% | |
All regressions seem to be minimal, and within the normal variance for
the benchmark. The fix for [1] assumes that 3% is noise -- and there
were no further practical complaints), so hopefully this means that such
variations in these microbenchmarks do not reflect on practical
workloads.
(3) I also ran stress-ng in a nested cgroup and did not observe any
obvious regressions.
[1]https://lore.kernel.org/all/20190520063534.GB19312@shao2-debian/
Suggested-by: Johannes Weiner <[email protected]>
Signed-off-by: Yosry Ahmed <[email protected]>
Tested-by: Domenico Cerasuolo <[email protected]>
---
mm/memcontrol.c | 50 +++++++++++++++++++++++++++++++++----------------
1 file changed, 34 insertions(+), 16 deletions(-)
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index cf05b97c1e824..93b483b379aa1 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -631,6 +631,9 @@ struct memcg_vmstats_percpu {
/* Cgroup1: threshold notifications & softlimit tree updates */
unsigned long nr_page_events;
unsigned long targets[MEM_CGROUP_NTARGETS];
+
+ /* Stats updates since the last flush */
+ unsigned int stats_updates;
};
struct memcg_vmstats {
@@ -645,6 +648,9 @@ struct memcg_vmstats {
/* Pending child counts during tree propagation */
long state_pending[MEMCG_NR_STAT];
unsigned long events_pending[NR_MEMCG_EVENTS];
+
+ /* Stats updates since the last flush */
+ atomic64_t stats_updates;
};
/*
@@ -664,9 +670,7 @@ struct memcg_vmstats {
*/
static void flush_memcg_stats_dwork(struct work_struct *w);
static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork);
-static DEFINE_PER_CPU(unsigned int, stats_updates);
static atomic_t stats_flush_ongoing = ATOMIC_INIT(0);
-static atomic_t stats_flush_threshold = ATOMIC_INIT(0);
static u64 flush_last_time;
#define FLUSH_TIME (2UL*HZ)
@@ -693,26 +697,37 @@ static void memcg_stats_unlock(void)
preempt_enable_nested();
}
+
+static bool memcg_should_flush_stats(struct mem_cgroup *memcg)
+{
+ return atomic64_read(&memcg->vmstats->stats_updates) >
+ MEMCG_CHARGE_BATCH * num_online_cpus();
+}
+
static inline void memcg_rstat_updated(struct mem_cgroup *memcg, int val)
{
+ int cpu = smp_processor_id();
unsigned int x;
if (!val)
return;
- cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
+ cgroup_rstat_updated(memcg->css.cgroup, cpu);
+
+ for (; memcg; memcg = parent_mem_cgroup(memcg)) {
+ x = __this_cpu_add_return(memcg->vmstats_percpu->stats_updates,
+ abs(val));
+
+ if (x < MEMCG_CHARGE_BATCH)
+ continue;
- x = __this_cpu_add_return(stats_updates, abs(val));
- if (x > MEMCG_CHARGE_BATCH) {
/*
- * If stats_flush_threshold exceeds the threshold
- * (>num_online_cpus()), cgroup stats update will be triggered
- * in __mem_cgroup_flush_stats(). Increasing this var further
- * is redundant and simply adds overhead in atomic update.
+ * If @memcg is already flush-able, increasing stats_updates is
+ * redundant. Avoid the overhead of the atomic update.
*/
- if (atomic_read(&stats_flush_threshold) <= num_online_cpus())
- atomic_add(x / MEMCG_CHARGE_BATCH, &stats_flush_threshold);
- __this_cpu_write(stats_updates, 0);
+ if (!memcg_should_flush_stats(memcg))
+ atomic64_add(x, &memcg->vmstats->stats_updates);
+ __this_cpu_write(memcg->vmstats_percpu->stats_updates, 0);
}
}
@@ -731,13 +746,12 @@ static void do_flush_stats(void)
cgroup_rstat_flush(root_mem_cgroup->css.cgroup);
- atomic_set(&stats_flush_threshold, 0);
atomic_set(&stats_flush_ongoing, 0);
}
void mem_cgroup_flush_stats(void)
{
- if (atomic_read(&stats_flush_threshold) > num_online_cpus())
+ if (memcg_should_flush_stats(root_mem_cgroup))
do_flush_stats();
}
@@ -751,8 +765,8 @@ void mem_cgroup_flush_stats_ratelimited(void)
static void flush_memcg_stats_dwork(struct work_struct *w)
{
/*
- * Always flush here so that flushing in latency-sensitive paths is
- * as cheap as possible.
+ * Deliberately ignore memcg_should_flush_stats() here so that flushing
+ * in latency-sensitive paths is as cheap as possible.
*/
do_flush_stats();
queue_delayed_work(system_unbound_wq, &stats_flush_dwork, FLUSH_TIME);
@@ -5809,6 +5823,10 @@ static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
}
}
}
+ statc->stats_updates = 0;
+ /* We are in a per-cpu loop here, only do the atomic write once */
+ if (atomic64_read(&memcg->vmstats->stats_updates))
+ atomic64_set(&memcg->vmstats->stats_updates, 0);
}
#ifdef CONFIG_MMU
--
2.43.0.rc1.413.gea7ed67945-goog
Stats flushing for memcg currently follows the following rules:
- Always flush the entire memcg hierarchy (i.e. flush the root).
- Only one flusher is allowed at a time. If someone else tries to flush
concurrently, they skip and return immediately.
- A periodic flusher flushes all the stats every 2 seconds.
The reason this approach is followed is because all flushes are
serialized by a global rstat spinlock. On the memcg side, flushing is
invoked from userspace reads as well as in-kernel flushers (e.g.
reclaim, refault, etc). This approach aims to avoid serializing all
flushers on the global lock, which can cause a significant performance
hit under high concurrency.
This approach has the following problems:
- Occasionally a userspace read of the stats of a non-root cgroup will
be too expensive as it has to flush the entire hierarchy [1].
- Sometimes the stats accuracy are compromised if there is an ongoing
flush, and we skip and return before the subtree of interest is
actually flushed, yielding stale stats (by up to 2s due to periodic
flushing). This is more visible when reading stats from userspace,
but can also affect in-kernel flushers.
The latter problem is particulary a concern when userspace reads stats
after an event occurs, but gets stats from before the event. Examples:
- When memory usage / pressure spikes, a userspace OOM handler may look
at the stats of different memcgs to select a victim based on various
heuristics (e.g. how much private memory will be freed by killing
this). Reading stale stats from before the usage spike in this case
may cause a wrongful OOM kill.
- A proactive reclaimer may read the stats after writing to
memory.reclaim to measure the success of the reclaim operation. Stale
stats from before reclaim may give a false negative.
- Reading the stats of a parent and a child memcg may be inconsistent
(child larger than parent), if the flush doesn't happen when the
parent is read, but happens when the child is read.
As for in-kernel flushers, they will occasionally get stale stats. No
regressions are currently known from this, but if there are regressions,
they would be very difficult to debug and link to the source of the
problem.
This patch aims to fix these problems by restoring subtree flushing,
and removing the unified/coalesced flushing logic that skips flushing if
there is an ongoing flush. This change would introduce a significant
regression with global stats flushing thresholds. With per-memcg stats
flushing thresholds, this seems to perform really well. The thresholds
protect the underlying lock from unnecessary contention.
Add a mutex to protect the underlying rstat lock from excessive memcg
flushing. The thresholds are re-checked after the mutex is grabbed to
make sure that a concurrent flush did not already get the subtree we are
trying to flush. A call to cgroup_rstat_flush() is not cheap, even if
there are no pending updates.
This patch was tested in two ways to ensure the latency of flushing is
up to bar, on a machine with 384 cpus:
- A synthetic test with 5000 concurrent workers in 500 cgroups doing
allocations and reclaim, as well as 1000 readers for memory.stat
(variation of [2]). No regressions were noticed in the total runtime.
Note that significant regressions in this test are observed with
global stats thresholds, but not with per-memcg thresholds.
- A synthetic stress test for concurrently reading memcg stats while
memory allocation/freeing workers are running in the background,
provided by Wei Xu [3]. With 250k threads reading the stats every
100ms in 50k cgroups, 99.9% of reads take <= 50us. Less than 0.01%
of reads take more than 1ms, and no reads take more than 100ms.
[1] https://lore.kernel.org/lkml/CABWYdi0c6__rh-K7dcM_pkf9BJdTRtAU08M43KO9ME4-dsgfoQ@mail.gmail.com/
[2] https://lore.kernel.org/lkml/CAJD7tka13M-zVZTyQJYL1iUAYvuQ1fcHbCjcOBZcz6POYTV-4g@mail.gmail.com/
[3] https://lore.kernel.org/lkml/CAAPL-u9D2b=iF5Lf_cRnKxUfkiEe0AMDTu6yhrUAzX0b6a6rDg@mail.gmail.com/
Signed-off-by: Yosry Ahmed <[email protected]>
Tested-by: Domenico Cerasuolo <[email protected]>
---
include/linux/memcontrol.h | 8 ++--
mm/memcontrol.c | 75 +++++++++++++++++++++++---------------
mm/vmscan.c | 2 +-
mm/workingset.c | 10 +++--
4 files changed, 58 insertions(+), 37 deletions(-)
diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
index a568f70a26774..8673140683e6e 100644
--- a/include/linux/memcontrol.h
+++ b/include/linux/memcontrol.h
@@ -1050,8 +1050,8 @@ static inline unsigned long lruvec_page_state_local(struct lruvec *lruvec,
return x;
}
-void mem_cgroup_flush_stats(void);
-void mem_cgroup_flush_stats_ratelimited(void);
+void mem_cgroup_flush_stats(struct mem_cgroup *memcg);
+void mem_cgroup_flush_stats_ratelimited(struct mem_cgroup *memcg);
void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
int val);
@@ -1566,11 +1566,11 @@ static inline unsigned long lruvec_page_state_local(struct lruvec *lruvec,
return node_page_state(lruvec_pgdat(lruvec), idx);
}
-static inline void mem_cgroup_flush_stats(void)
+static inline void mem_cgroup_flush_stats(struct mem_cgroup *memcg)
{
}
-static inline void mem_cgroup_flush_stats_ratelimited(void)
+static inline void mem_cgroup_flush_stats_ratelimited(struct mem_cgroup *memcg)
{
}
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 93b483b379aa1..5d300318bf18a 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -670,7 +670,6 @@ struct memcg_vmstats {
*/
static void flush_memcg_stats_dwork(struct work_struct *w);
static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork);
-static atomic_t stats_flush_ongoing = ATOMIC_INIT(0);
static u64 flush_last_time;
#define FLUSH_TIME (2UL*HZ)
@@ -731,35 +730,47 @@ static inline void memcg_rstat_updated(struct mem_cgroup *memcg, int val)
}
}
-static void do_flush_stats(void)
+static void do_flush_stats(struct mem_cgroup *memcg)
{
- /*
- * We always flush the entire tree, so concurrent flushers can just
- * skip. This avoids a thundering herd problem on the rstat global lock
- * from memcg flushers (e.g. reclaim, refault, etc).
- */
- if (atomic_read(&stats_flush_ongoing) ||
- atomic_xchg(&stats_flush_ongoing, 1))
- return;
-
- WRITE_ONCE(flush_last_time, jiffies_64);
-
- cgroup_rstat_flush(root_mem_cgroup->css.cgroup);
+ if (mem_cgroup_is_root(memcg))
+ WRITE_ONCE(flush_last_time, jiffies_64);
- atomic_set(&stats_flush_ongoing, 0);
+ cgroup_rstat_flush(memcg->css.cgroup);
}
-void mem_cgroup_flush_stats(void)
+/*
+ * mem_cgroup_flush_stats - flush the stats of a memory cgroup subtree
+ * @memcg: root of the subtree to flush
+ *
+ * Flushing is serialized by the underlying global rstat lock. There is also a
+ * minimum amount of work to be done even if there are no stat updates to flush.
+ * Hence, we only flush the stats if the updates delta exceeds a threshold. This
+ * avoids unnecessary work and contention on the underlying lock.
+ */
+void mem_cgroup_flush_stats(struct mem_cgroup *memcg)
{
- if (memcg_should_flush_stats(root_mem_cgroup))
- do_flush_stats();
+ static DEFINE_MUTEX(memcg_stats_flush_mutex);
+
+ if (mem_cgroup_disabled())
+ return;
+
+ if (!memcg)
+ memcg = root_mem_cgroup;
+
+ if (memcg_should_flush_stats(memcg)) {
+ mutex_lock(&memcg_stats_flush_mutex);
+ /* Check again after locking, another flush may have occurred */
+ if (memcg_should_flush_stats(memcg))
+ do_flush_stats(memcg);
+ mutex_unlock(&memcg_stats_flush_mutex);
+ }
}
-void mem_cgroup_flush_stats_ratelimited(void)
+void mem_cgroup_flush_stats_ratelimited(struct mem_cgroup *memcg)
{
/* Only flush if the periodic flusher is one full cycle late */
if (time_after64(jiffies_64, READ_ONCE(flush_last_time) + 2*FLUSH_TIME))
- mem_cgroup_flush_stats();
+ mem_cgroup_flush_stats(memcg);
}
static void flush_memcg_stats_dwork(struct work_struct *w)
@@ -768,7 +779,7 @@ static void flush_memcg_stats_dwork(struct work_struct *w)
* Deliberately ignore memcg_should_flush_stats() here so that flushing
* in latency-sensitive paths is as cheap as possible.
*/
- do_flush_stats();
+ do_flush_stats(root_mem_cgroup);
queue_delayed_work(system_unbound_wq, &stats_flush_dwork, FLUSH_TIME);
}
@@ -1664,7 +1675,7 @@ static void memcg_stat_format(struct mem_cgroup *memcg, struct seq_buf *s)
*
* Current memory state:
*/
- mem_cgroup_flush_stats();
+ mem_cgroup_flush_stats(memcg);
for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
u64 size;
@@ -4214,7 +4225,7 @@ static int memcg_numa_stat_show(struct seq_file *m, void *v)
int nid;
struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
- mem_cgroup_flush_stats();
+ mem_cgroup_flush_stats(memcg);
for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
seq_printf(m, "%s=%lu", stat->name,
@@ -4295,7 +4306,7 @@ static void memcg1_stat_format(struct mem_cgroup *memcg, struct seq_buf *s)
BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
- mem_cgroup_flush_stats();
+ mem_cgroup_flush_stats(memcg);
for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
unsigned long nr;
@@ -4791,7 +4802,7 @@ void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
struct mem_cgroup *parent;
- mem_cgroup_flush_stats();
+ mem_cgroup_flush_stats(memcg);
*pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
*pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
@@ -6886,7 +6897,7 @@ static int memory_numa_stat_show(struct seq_file *m, void *v)
int i;
struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
- mem_cgroup_flush_stats();
+ mem_cgroup_flush_stats(memcg);
for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
int nid;
@@ -8125,7 +8136,11 @@ bool obj_cgroup_may_zswap(struct obj_cgroup *objcg)
break;
}
- cgroup_rstat_flush(memcg->css.cgroup);
+ /*
+ * mem_cgroup_flush_stats() ignores small changes. Use
+ * do_flush_stats() directly to get accurate stats for charging.
+ */
+ do_flush_stats(memcg);
pages = memcg_page_state(memcg, MEMCG_ZSWAP_B) / PAGE_SIZE;
if (pages < max)
continue;
@@ -8190,8 +8205,10 @@ void obj_cgroup_uncharge_zswap(struct obj_cgroup *objcg, size_t size)
static u64 zswap_current_read(struct cgroup_subsys_state *css,
struct cftype *cft)
{
- cgroup_rstat_flush(css->cgroup);
- return memcg_page_state(mem_cgroup_from_css(css), MEMCG_ZSWAP_B);
+ struct mem_cgroup *memcg = mem_cgroup_from_css(css);
+
+ mem_cgroup_flush_stats(memcg);
+ return memcg_page_state(memcg, MEMCG_ZSWAP_B);
}
static int zswap_max_show(struct seq_file *m, void *v)
diff --git a/mm/vmscan.c b/mm/vmscan.c
index d8c3338fee0fb..0b8a0107d58d8 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -2250,7 +2250,7 @@ static void prepare_scan_control(pg_data_t *pgdat, struct scan_control *sc)
* Flush the memory cgroup stats, so that we read accurate per-memcg
* lruvec stats for heuristics.
*/
- mem_cgroup_flush_stats();
+ mem_cgroup_flush_stats(sc->target_mem_cgroup);
/*
* Determine the scan balance between anon and file LRUs.
diff --git a/mm/workingset.c b/mm/workingset.c
index dce41577a49d2..7d3dacab8451a 100644
--- a/mm/workingset.c
+++ b/mm/workingset.c
@@ -464,8 +464,12 @@ bool workingset_test_recent(void *shadow, bool file, bool *workingset)
rcu_read_unlock();
- /* Flush stats (and potentially sleep) outside the RCU read section */
- mem_cgroup_flush_stats_ratelimited();
+ /*
+ * Flush stats (and potentially sleep) outside the RCU read section.
+ * XXX: With per-memcg flushing and thresholding, is ratelimiting
+ * still needed here?
+ */
+ mem_cgroup_flush_stats_ratelimited(eviction_memcg);
eviction_lruvec = mem_cgroup_lruvec(eviction_memcg, pgdat);
refault = atomic_long_read(&eviction_lruvec->nonresident_age);
@@ -676,7 +680,7 @@ static unsigned long count_shadow_nodes(struct shrinker *shrinker,
struct lruvec *lruvec;
int i;
- mem_cgroup_flush_stats();
+ mem_cgroup_flush_stats(sc->memcg);
lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
for (pages = 0, i = 0; i < NR_LRU_LISTS; i++)
pages += lruvec_page_state_local(lruvec,
--
2.43.0.rc1.413.gea7ed67945-goog
The following patch will make use of those structs in the flushing code,
so move their definitions (and a few other dependencies) a little bit up
to reduce the diff noise in the following patch.
No functional change intended.
Signed-off-by: Yosry Ahmed <[email protected]>
Tested-by: Domenico Cerasuolo <[email protected]>
Acked-by: Shakeel Butt <[email protected]>
---
mm/memcontrol.c | 148 ++++++++++++++++++++++++------------------------
1 file changed, 74 insertions(+), 74 deletions(-)
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 61435bd037cb4..cf05b97c1e824 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -573,6 +573,80 @@ mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
return mz;
}
+/* Subset of vm_event_item to report for memcg event stats */
+static const unsigned int memcg_vm_event_stat[] = {
+ PGPGIN,
+ PGPGOUT,
+ PGSCAN_KSWAPD,
+ PGSCAN_DIRECT,
+ PGSCAN_KHUGEPAGED,
+ PGSTEAL_KSWAPD,
+ PGSTEAL_DIRECT,
+ PGSTEAL_KHUGEPAGED,
+ PGFAULT,
+ PGMAJFAULT,
+ PGREFILL,
+ PGACTIVATE,
+ PGDEACTIVATE,
+ PGLAZYFREE,
+ PGLAZYFREED,
+#if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
+ ZSWPIN,
+ ZSWPOUT,
+ ZSWP_WB,
+#endif
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+ THP_FAULT_ALLOC,
+ THP_COLLAPSE_ALLOC,
+ THP_SWPOUT,
+ THP_SWPOUT_FALLBACK,
+#endif
+};
+
+#define NR_MEMCG_EVENTS ARRAY_SIZE(memcg_vm_event_stat)
+static int mem_cgroup_events_index[NR_VM_EVENT_ITEMS] __read_mostly;
+
+static void init_memcg_events(void)
+{
+ int i;
+
+ for (i = 0; i < NR_MEMCG_EVENTS; ++i)
+ mem_cgroup_events_index[memcg_vm_event_stat[i]] = i + 1;
+}
+
+static inline int memcg_events_index(enum vm_event_item idx)
+{
+ return mem_cgroup_events_index[idx] - 1;
+}
+
+struct memcg_vmstats_percpu {
+ /* Local (CPU and cgroup) page state & events */
+ long state[MEMCG_NR_STAT];
+ unsigned long events[NR_MEMCG_EVENTS];
+
+ /* Delta calculation for lockless upward propagation */
+ long state_prev[MEMCG_NR_STAT];
+ unsigned long events_prev[NR_MEMCG_EVENTS];
+
+ /* Cgroup1: threshold notifications & softlimit tree updates */
+ unsigned long nr_page_events;
+ unsigned long targets[MEM_CGROUP_NTARGETS];
+};
+
+struct memcg_vmstats {
+ /* Aggregated (CPU and subtree) page state & events */
+ long state[MEMCG_NR_STAT];
+ unsigned long events[NR_MEMCG_EVENTS];
+
+ /* Non-hierarchical (CPU aggregated) page state & events */
+ long state_local[MEMCG_NR_STAT];
+ unsigned long events_local[NR_MEMCG_EVENTS];
+
+ /* Pending child counts during tree propagation */
+ long state_pending[MEMCG_NR_STAT];
+ unsigned long events_pending[NR_MEMCG_EVENTS];
+};
+
/*
* memcg and lruvec stats flushing
*
@@ -684,80 +758,6 @@ static void flush_memcg_stats_dwork(struct work_struct *w)
queue_delayed_work(system_unbound_wq, &stats_flush_dwork, FLUSH_TIME);
}
-/* Subset of vm_event_item to report for memcg event stats */
-static const unsigned int memcg_vm_event_stat[] = {
- PGPGIN,
- PGPGOUT,
- PGSCAN_KSWAPD,
- PGSCAN_DIRECT,
- PGSCAN_KHUGEPAGED,
- PGSTEAL_KSWAPD,
- PGSTEAL_DIRECT,
- PGSTEAL_KHUGEPAGED,
- PGFAULT,
- PGMAJFAULT,
- PGREFILL,
- PGACTIVATE,
- PGDEACTIVATE,
- PGLAZYFREE,
- PGLAZYFREED,
-#if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
- ZSWPIN,
- ZSWPOUT,
- ZSWP_WB,
-#endif
-#ifdef CONFIG_TRANSPARENT_HUGEPAGE
- THP_FAULT_ALLOC,
- THP_COLLAPSE_ALLOC,
- THP_SWPOUT,
- THP_SWPOUT_FALLBACK,
-#endif
-};
-
-#define NR_MEMCG_EVENTS ARRAY_SIZE(memcg_vm_event_stat)
-static int mem_cgroup_events_index[NR_VM_EVENT_ITEMS] __read_mostly;
-
-static void init_memcg_events(void)
-{
- int i;
-
- for (i = 0; i < NR_MEMCG_EVENTS; ++i)
- mem_cgroup_events_index[memcg_vm_event_stat[i]] = i + 1;
-}
-
-static inline int memcg_events_index(enum vm_event_item idx)
-{
- return mem_cgroup_events_index[idx] - 1;
-}
-
-struct memcg_vmstats_percpu {
- /* Local (CPU and cgroup) page state & events */
- long state[MEMCG_NR_STAT];
- unsigned long events[NR_MEMCG_EVENTS];
-
- /* Delta calculation for lockless upward propagation */
- long state_prev[MEMCG_NR_STAT];
- unsigned long events_prev[NR_MEMCG_EVENTS];
-
- /* Cgroup1: threshold notifications & softlimit tree updates */
- unsigned long nr_page_events;
- unsigned long targets[MEM_CGROUP_NTARGETS];
-};
-
-struct memcg_vmstats {
- /* Aggregated (CPU and subtree) page state & events */
- long state[MEMCG_NR_STAT];
- unsigned long events[NR_MEMCG_EVENTS];
-
- /* Non-hierarchical (CPU aggregated) page state & events */
- long state_local[MEMCG_NR_STAT];
- unsigned long events_local[NR_MEMCG_EVENTS];
-
- /* Pending child counts during tree propagation */
- long state_pending[MEMCG_NR_STAT];
- unsigned long events_pending[NR_MEMCG_EVENTS];
-};
-
unsigned long memcg_page_state(struct mem_cgroup *memcg, int idx)
{
long x = READ_ONCE(memcg->vmstats->state[idx]);
--
2.43.0.rc1.413.gea7ed67945-goog
On Wed, Nov 29, 2023 at 03:21:53AM +0000, Yosry Ahmed wrote:
> Stats flushing for memcg currently follows the following rules:
> - Always flush the entire memcg hierarchy (i.e. flush the root).
> - Only one flusher is allowed at a time. If someone else tries to flush
> concurrently, they skip and return immediately.
> - A periodic flusher flushes all the stats every 2 seconds.
>
> The reason this approach is followed is because all flushes are
> serialized by a global rstat spinlock. On the memcg side, flushing is
> invoked from userspace reads as well as in-kernel flushers (e.g.
> reclaim, refault, etc). This approach aims to avoid serializing all
> flushers on the global lock, which can cause a significant performance
> hit under high concurrency.
>
> This approach has the following problems:
> - Occasionally a userspace read of the stats of a non-root cgroup will
> be too expensive as it has to flush the entire hierarchy [1].
> - Sometimes the stats accuracy are compromised if there is an ongoing
> flush, and we skip and return before the subtree of interest is
> actually flushed, yielding stale stats (by up to 2s due to periodic
> flushing). This is more visible when reading stats from userspace,
> but can also affect in-kernel flushers.
>
> The latter problem is particulary a concern when userspace reads stats
> after an event occurs, but gets stats from before the event. Examples:
> - When memory usage / pressure spikes, a userspace OOM handler may look
> at the stats of different memcgs to select a victim based on various
> heuristics (e.g. how much private memory will be freed by killing
> this). Reading stale stats from before the usage spike in this case
> may cause a wrongful OOM kill.
> - A proactive reclaimer may read the stats after writing to
> memory.reclaim to measure the success of the reclaim operation. Stale
> stats from before reclaim may give a false negative.
> - Reading the stats of a parent and a child memcg may be inconsistent
> (child larger than parent), if the flush doesn't happen when the
> parent is read, but happens when the child is read.
>
> As for in-kernel flushers, they will occasionally get stale stats. No
> regressions are currently known from this, but if there are regressions,
> they would be very difficult to debug and link to the source of the
> problem.
>
> This patch aims to fix these problems by restoring subtree flushing,
> and removing the unified/coalesced flushing logic that skips flushing if
> there is an ongoing flush. This change would introduce a significant
> regression with global stats flushing thresholds. With per-memcg stats
> flushing thresholds, this seems to perform really well. The thresholds
> protect the underlying lock from unnecessary contention.
>
> Add a mutex to protect the underlying rstat lock from excessive memcg
> flushing. The thresholds are re-checked after the mutex is grabbed to
> make sure that a concurrent flush did not already get the subtree we are
> trying to flush. A call to cgroup_rstat_flush() is not cheap, even if
> there are no pending updates.
>
> This patch was tested in two ways to ensure the latency of flushing is
> up to bar, on a machine with 384 cpus:
> - A synthetic test with 5000 concurrent workers in 500 cgroups doing
> allocations and reclaim, as well as 1000 readers for memory.stat
> (variation of [2]). No regressions were noticed in the total runtime.
> Note that significant regressions in this test are observed with
> global stats thresholds, but not with per-memcg thresholds.
>
> - A synthetic stress test for concurrently reading memcg stats while
> memory allocation/freeing workers are running in the background,
> provided by Wei Xu [3]. With 250k threads reading the stats every
> 100ms in 50k cgroups, 99.9% of reads take <= 50us. Less than 0.01%
> of reads take more than 1ms, and no reads take more than 100ms.
>
> [1] https://lore.kernel.org/lkml/CABWYdi0c6__rh-K7dcM_pkf9BJdTRtAU08M43KO9ME4-dsgfoQ@mail.gmail.com/
> [2] https://lore.kernel.org/lkml/CAJD7tka13M-zVZTyQJYL1iUAYvuQ1fcHbCjcOBZcz6POYTV-4g@mail.gmail.com/
> [3] https://lore.kernel.org/lkml/CAAPL-u9D2b=iF5Lf_cRnKxUfkiEe0AMDTu6yhrUAzX0b6a6rDg@mail.gmail.com/
>
> Signed-off-by: Yosry Ahmed <[email protected]>
> Tested-by: Domenico Cerasuolo <[email protected]>
> ---
> include/linux/memcontrol.h | 8 ++--
> mm/memcontrol.c | 75 +++++++++++++++++++++++---------------
> mm/vmscan.c | 2 +-
> mm/workingset.c | 10 +++--
> 4 files changed, 58 insertions(+), 37 deletions(-)
>
> diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
> index a568f70a26774..8673140683e6e 100644
> --- a/include/linux/memcontrol.h
> +++ b/include/linux/memcontrol.h
> @@ -1050,8 +1050,8 @@ static inline unsigned long lruvec_page_state_local(struct lruvec *lruvec,
> return x;
> }
>
> -void mem_cgroup_flush_stats(void);
> -void mem_cgroup_flush_stats_ratelimited(void);
> +void mem_cgroup_flush_stats(struct mem_cgroup *memcg);
> +void mem_cgroup_flush_stats_ratelimited(struct mem_cgroup *memcg);
>
> void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
> int val);
> @@ -1566,11 +1566,11 @@ static inline unsigned long lruvec_page_state_local(struct lruvec *lruvec,
> return node_page_state(lruvec_pgdat(lruvec), idx);
> }
>
> -static inline void mem_cgroup_flush_stats(void)
> +static inline void mem_cgroup_flush_stats(struct mem_cgroup *memcg)
> {
> }
>
> -static inline void mem_cgroup_flush_stats_ratelimited(void)
> +static inline void mem_cgroup_flush_stats_ratelimited(struct mem_cgroup *memcg)
> {
> }
>
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index 93b483b379aa1..5d300318bf18a 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -670,7 +670,6 @@ struct memcg_vmstats {
> */
> static void flush_memcg_stats_dwork(struct work_struct *w);
> static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork);
> -static atomic_t stats_flush_ongoing = ATOMIC_INIT(0);
> static u64 flush_last_time;
>
> #define FLUSH_TIME (2UL*HZ)
> @@ -731,35 +730,47 @@ static inline void memcg_rstat_updated(struct mem_cgroup *memcg, int val)
> }
> }
>
> -static void do_flush_stats(void)
> +static void do_flush_stats(struct mem_cgroup *memcg)
> {
> - /*
> - * We always flush the entire tree, so concurrent flushers can just
> - * skip. This avoids a thundering herd problem on the rstat global lock
> - * from memcg flushers (e.g. reclaim, refault, etc).
> - */
> - if (atomic_read(&stats_flush_ongoing) ||
> - atomic_xchg(&stats_flush_ongoing, 1))
> - return;
> -
> - WRITE_ONCE(flush_last_time, jiffies_64);
> -
> - cgroup_rstat_flush(root_mem_cgroup->css.cgroup);
> + if (mem_cgroup_is_root(memcg))
> + WRITE_ONCE(flush_last_time, jiffies_64);
>
> - atomic_set(&stats_flush_ongoing, 0);
> + cgroup_rstat_flush(memcg->css.cgroup);
> }
>
> -void mem_cgroup_flush_stats(void)
> +/*
> + * mem_cgroup_flush_stats - flush the stats of a memory cgroup subtree
> + * @memcg: root of the subtree to flush
> + *
> + * Flushing is serialized by the underlying global rstat lock. There is also a
> + * minimum amount of work to be done even if there are no stat updates to flush.
> + * Hence, we only flush the stats if the updates delta exceeds a threshold. This
> + * avoids unnecessary work and contention on the underlying lock.
> + */
What is global rstat lock?
> +void mem_cgroup_flush_stats(struct mem_cgroup *memcg)
> {
> - if (memcg_should_flush_stats(root_mem_cgroup))
> - do_flush_stats();
> + static DEFINE_MUTEX(memcg_stats_flush_mutex);
> +
> + if (mem_cgroup_disabled())
> + return;
> +
> + if (!memcg)
> + memcg = root_mem_cgroup;
> +
> + if (memcg_should_flush_stats(memcg)) {
> + mutex_lock(&memcg_stats_flush_mutex);
> + /* Check again after locking, another flush may have occurred */
> + if (memcg_should_flush_stats(memcg))
> + do_flush_stats(memcg);
> + mutex_unlock(&memcg_stats_flush_mutex);
> + }
> }
>
> -void mem_cgroup_flush_stats_ratelimited(void)
> +void mem_cgroup_flush_stats_ratelimited(struct mem_cgroup *memcg)
> {
> /* Only flush if the periodic flusher is one full cycle late */
> if (time_after64(jiffies_64, READ_ONCE(flush_last_time) + 2*FLUSH_TIME))
> - mem_cgroup_flush_stats();
> + mem_cgroup_flush_stats(memcg);
> }
>
> static void flush_memcg_stats_dwork(struct work_struct *w)
> @@ -768,7 +779,7 @@ static void flush_memcg_stats_dwork(struct work_struct *w)
> * Deliberately ignore memcg_should_flush_stats() here so that flushing
> * in latency-sensitive paths is as cheap as possible.
> */
> - do_flush_stats();
> + do_flush_stats(root_mem_cgroup);
> queue_delayed_work(system_unbound_wq, &stats_flush_dwork, FLUSH_TIME);
> }
>
> @@ -1664,7 +1675,7 @@ static void memcg_stat_format(struct mem_cgroup *memcg, struct seq_buf *s)
> *
> * Current memory state:
> */
> - mem_cgroup_flush_stats();
> + mem_cgroup_flush_stats(memcg);
>
> for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
> u64 size;
> @@ -4214,7 +4225,7 @@ static int memcg_numa_stat_show(struct seq_file *m, void *v)
> int nid;
> struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
>
> - mem_cgroup_flush_stats();
> + mem_cgroup_flush_stats(memcg);
>
> for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
> seq_printf(m, "%s=%lu", stat->name,
> @@ -4295,7 +4306,7 @@ static void memcg1_stat_format(struct mem_cgroup *memcg, struct seq_buf *s)
>
> BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
>
> - mem_cgroup_flush_stats();
> + mem_cgroup_flush_stats(memcg);
>
> for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
> unsigned long nr;
> @@ -4791,7 +4802,7 @@ void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
> struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
> struct mem_cgroup *parent;
>
> - mem_cgroup_flush_stats();
> + mem_cgroup_flush_stats(memcg);
>
> *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
> *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
> @@ -6886,7 +6897,7 @@ static int memory_numa_stat_show(struct seq_file *m, void *v)
> int i;
> struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
>
> - mem_cgroup_flush_stats();
> + mem_cgroup_flush_stats(memcg);
>
> for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
> int nid;
> @@ -8125,7 +8136,11 @@ bool obj_cgroup_may_zswap(struct obj_cgroup *objcg)
> break;
> }
>
> - cgroup_rstat_flush(memcg->css.cgroup);
> + /*
> + * mem_cgroup_flush_stats() ignores small changes. Use
> + * do_flush_stats() directly to get accurate stats for charging.
> + */
> + do_flush_stats(memcg);
> pages = memcg_page_state(memcg, MEMCG_ZSWAP_B) / PAGE_SIZE;
> if (pages < max)
> continue;
> @@ -8190,8 +8205,10 @@ void obj_cgroup_uncharge_zswap(struct obj_cgroup *objcg, size_t size)
> static u64 zswap_current_read(struct cgroup_subsys_state *css,
> struct cftype *cft)
> {
> - cgroup_rstat_flush(css->cgroup);
> - return memcg_page_state(mem_cgroup_from_css(css), MEMCG_ZSWAP_B);
> + struct mem_cgroup *memcg = mem_cgroup_from_css(css);
> +
> + mem_cgroup_flush_stats(memcg);
> + return memcg_page_state(memcg, MEMCG_ZSWAP_B);
> }
>
> static int zswap_max_show(struct seq_file *m, void *v)
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index d8c3338fee0fb..0b8a0107d58d8 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -2250,7 +2250,7 @@ static void prepare_scan_control(pg_data_t *pgdat, struct scan_control *sc)
> * Flush the memory cgroup stats, so that we read accurate per-memcg
> * lruvec stats for heuristics.
> */
> - mem_cgroup_flush_stats();
> + mem_cgroup_flush_stats(sc->target_mem_cgroup);
>
> /*
> * Determine the scan balance between anon and file LRUs.
> diff --git a/mm/workingset.c b/mm/workingset.c
> index dce41577a49d2..7d3dacab8451a 100644
> --- a/mm/workingset.c
> +++ b/mm/workingset.c
> @@ -464,8 +464,12 @@ bool workingset_test_recent(void *shadow, bool file, bool *workingset)
>
> rcu_read_unlock();
>
> - /* Flush stats (and potentially sleep) outside the RCU read section */
> - mem_cgroup_flush_stats_ratelimited();
> + /*
> + * Flush stats (and potentially sleep) outside the RCU read section.
> + * XXX: With per-memcg flushing and thresholding, is ratelimiting
> + * still needed here?
> + */
> + mem_cgroup_flush_stats_ratelimited(eviction_memcg);
What if flushing is not rate-limited (e.g. above line is commented)?
>
> eviction_lruvec = mem_cgroup_lruvec(eviction_memcg, pgdat);
> refault = atomic_long_read(&eviction_lruvec->nonresident_age);
> @@ -676,7 +680,7 @@ static unsigned long count_shadow_nodes(struct shrinker *shrinker,
> struct lruvec *lruvec;
> int i;
>
> - mem_cgroup_flush_stats();
> + mem_cgroup_flush_stats(sc->memcg);
> lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
> for (pages = 0, i = 0; i < NR_LRU_LISTS; i++)
> pages += lruvec_page_state_local(lruvec,
Confused...
--
An old man doll... just what I always wanted! - Clara
On 12/1/23 20:57, Bagas Sanjaya wrote:
>> -void mem_cgroup_flush_stats(void)
>> +/*
>> + * mem_cgroup_flush_stats - flush the stats of a memory cgroup subtree
>> + * @memcg: root of the subtree to flush
>> + *
>> + * Flushing is serialized by the underlying global rstat lock. There is also a
>> + * minimum amount of work to be done even if there are no stat updates to flush.
>> + * Hence, we only flush the stats if the updates delta exceeds a threshold. This
>> + * avoids unnecessary work and contention on the underlying lock.
>> + */
> What is global rstat lock?
It is the cgroup_rstat_lock in kernel/cgroup/rstat.c.
Cheers,
Longman
On Wed, Nov 29, 2023 at 03:21:48AM +0000, Yosry Ahmed wrote:
> This series attempts to address shortages in today's approach for memcg
> stats flushing, namely occasionally stale or expensive stat reads. The
> series does so by changing the threshold that we use to decide whether
> to trigger a flush to be per memcg instead of global (patch 3), and then
> changing flushing to be per memcg (i.e. subtree flushes) instead of
> global (patch 5).
>
> Patch 3 & 5 are the core of the series, and they include more details
> and testing results. The rest are either cleanups or prep work.
>
> This series replaces the "memcg: more sophisticated stats flushing"
> series [1], which also replaces another series, in a long list of
> attempts to improve memcg stats flushing. It is not a new version of
> the same patchset as it is a completely different approach. This is
> based on collected feedback from discussions on lkml in all previous
> attempts. Hopefully, this is the final attempt.
>
> There was a reported regression in v2 [2] for will-it-scale::fallocate
> benchmark. I believe this regression should not affect production
> workloads. This specific benchmark is allocating and freeing memory
> (using fallocate/ftruncate) at a rate that is much faster to make actual
> use of the memory. Testing this series on 100+ machines running
> production workloads did not show any practical regressions in page
> fault latency or allocation latency, but it showed great improvements in
> stats read time. I do not have numbers about the exact improvements for
> this series, but combined with another optimization for cgroup v1 [3] we
> see 5-10x improvements. A significant chunk of that is coming from the
> cgroup v1 optimization, but this series also made an improvement as
> reported by Domenico [4].
>
> v3 -> v4:
> - Rebased on top of mm-unstable + "workload-specific and memory
> pressure-driven zswap writeback" series to fix conflicts [5].
>
> v3: https://lore.kernel.org/all/[email protected]/
>
> [1]https://lore.kernel.org/lkml/[email protected]/
> [2]https://lore.kernel.org/lkml/[email protected]/
> [3]https://lore.kernel.org/lkml/[email protected]/
> [4]https://lore.kernel.org/lkml/CAFYChMv_kv_KXOMRkrmTN-7MrfgBHMcK3YXv0dPYEL7nK77e2A@mail.gmail.com/
> [5]https://lore.kernel.org/all/[email protected]/
>
> Yosry Ahmed (5):
> mm: memcg: change flush_next_time to flush_last_time
> mm: memcg: move vmstats structs definition above flushing code
> mm: memcg: make stats flushing threshold per-memcg
> mm: workingset: move the stats flush into workingset_test_recent()
> mm: memcg: restore subtree stats flushing
>
> include/linux/memcontrol.h | 8 +-
> mm/memcontrol.c | 272 +++++++++++++++++++++----------------
> mm/vmscan.c | 2 +-
> mm/workingset.c | 42 ++++--
> 4 files changed, 188 insertions(+), 136 deletions(-)
>
No regressions when booting the kernel with this series applied.
Tested-by: Bagas Sanjaya <[email protected]>
--
An old man doll... just what I always wanted! - Clara
On 12/2/23 09:56, Waiman Long wrote:
>
> On 12/1/23 20:57, Bagas Sanjaya wrote:
>>> -void mem_cgroup_flush_stats(void)
>>> +/*
>>> + * mem_cgroup_flush_stats - flush the stats of a memory cgroup subtree
>>> + * @memcg: root of the subtree to flush
>>> + *
>>> + * Flushing is serialized by the underlying global rstat lock. There is also a
>>> + * minimum amount of work to be done even if there are no stat updates to flush.
>>> + * Hence, we only flush the stats if the updates delta exceeds a threshold. This
>>> + * avoids unnecessary work and contention on the underlying lock.
>>> + */
>> What is global rstat lock?
>
> It is the cgroup_rstat_lock in kernel/cgroup/rstat.c.
>
OK, I see that. Thanks!
--
An old man doll... just what I always wanted! - Clara
On Wed, Nov 29, 2023 at 03:21:51AM +0000, Yosry Ahmed wrote:
> A global counter for the magnitude of memcg stats update is maintained
> on the memcg side to avoid invoking rstat flushes when the pending
> updates are not significant. This avoids unnecessary flushes, which are
> not very cheap even if there isn't a lot of stats to flush. It also
> avoids unnecessary lock contention on the underlying global rstat lock.
>
> Make this threshold per-memcg. The scheme is followed where percpu (now
> also per-memcg) counters are incremented in the update path, and only
> propagated to per-memcg atomics when they exceed a certain threshold.
>
> This provides two benefits:
> (a) On large machines with a lot of memcgs, the global threshold can be
> reached relatively fast, so guarding the underlying lock becomes less
> effective. Making the threshold per-memcg avoids this.
>
> (b) Having a global threshold makes it hard to do subtree flushes, as we
> cannot reset the global counter except for a full flush. Per-memcg
> counters removes this as a blocker from doing subtree flushes, which
> helps avoid unnecessary work when the stats of a small subtree are
> needed.
>
> Nothing is free, of course. This comes at a cost:
> (a) A new per-cpu counter per memcg, consuming NR_CPUS * NR_MEMCGS * 4
> bytes. The extra memory usage is insigificant.
>
> (b) More work on the update side, although in the common case it will
> only be percpu counter updates. The amount of work scales with the
> number of ancestors (i.e. tree depth). This is not a new concept, adding
> a cgroup to the rstat tree involves a parent loop, so is charging.
> Testing results below show no significant regressions.
>
> (c) The error margin in the stats for the system as a whole increases
> from NR_CPUS * MEMCG_CHARGE_BATCH to NR_CPUS * MEMCG_CHARGE_BATCH *
> NR_MEMCGS. This is probably fine because we have a similar per-memcg
> error in charges coming from percpu stocks, and we have a periodic
> flusher that makes sure we always flush all the stats every 2s anyway.
>
> This patch was tested to make sure no significant regressions are
> introduced on the update path as follows. The following benchmarks were
> ran in a cgroup that is 2 levels deep (/sys/fs/cgroup/a/b/):
>
> (1) Running 22 instances of netperf on a 44 cpu machine with
> hyperthreading disabled. All instances are run in a level 2 cgroup, as
> well as netserver:
> # netserver -6
> # netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
>
> Averaging 20 runs, the numbers are as follows:
> Base: 40198.0 mbps
> Patched: 38629.7 mbps (-3.9%)
>
> The regression is minimal, especially for 22 instances in the same
> cgroup sharing all ancestors (so updating the same atomics).
>
> (2) will-it-scale page_fault tests. These tests (specifically
> per_process_ops in page_fault3 test) detected a 25.9% regression before
> for a change in the stats update path [1]. These are the
> numbers from 10 runs (+ is good) on a machine with 256 cpus:
>
> LABEL | MEAN | MEDIAN | STDDEV |
> ------------------------------+-------------+-------------+-------------
> page_fault1_per_process_ops | | | |
> (A) base | 270249.164 | 265437.000 | 13451.836 |
> (B) patched | 261368.709 | 255725.000 | 13394.767 |
> | -3.29% | -3.66% | |
> page_fault1_per_thread_ops | | | |
> (A) base | 242111.345 | 239737.000 | 10026.031 |
> (B) patched | 237057.109 | 235305.000 | 9769.687 |
> | -2.09% | -1.85% | |
> page_fault1_scalability | | |
> (A) base | 0.034387 | 0.035168 | 0.0018283 |
> (B) patched | 0.033988 | 0.034573 | 0.0018056 |
> | -1.16% | -1.69% | |
> page_fault2_per_process_ops | | |
> (A) base | 203561.836 | 203301.000 | 2550.764 |
> (B) patched | 197195.945 | 197746.000 | 2264.263 |
> | -3.13% | -2.73% | |
> page_fault2_per_thread_ops | | |
> (A) base | 171046.473 | 170776.000 | 1509.679 |
> (B) patched | 166626.327 | 166406.000 | 768.753 |
> | -2.58% | -2.56% | |
> page_fault2_scalability | | |
> (A) base | 0.054026 | 0.053821 | 0.00062121 |
> (B) patched | 0.053329 | 0.05306 | 0.00048394 |
> | -1.29% | -1.41% | |
> page_fault3_per_process_ops | | |
> (A) base | 1295807.782 | 1297550.000 | 5907.585 |
> (B) patched | 1275579.873 | 1273359.000 | 8759.160 |
> | -1.56% | -1.86% | |
> page_fault3_per_thread_ops | | |
> (A) base | 391234.164 | 390860.000 | 1760.720 |
> (B) patched | 377231.273 | 376369.000 | 1874.971 |
> | -3.58% | -3.71% | |
> page_fault3_scalability | | |
> (A) base | 0.60369 | 0.60072 | 0.0083029 |
> (B) patched | 0.61733 | 0.61544 | 0.009855 |
> | +2.26% | +2.45% | |
>
> All regressions seem to be minimal, and within the normal variance for
> the benchmark. The fix for [1] assumes that 3% is noise -- and there
> were no further practical complaints), so hopefully this means that such
> variations in these microbenchmarks do not reflect on practical
> workloads.
>
> (3) I also ran stress-ng in a nested cgroup and did not observe any
> obvious regressions.
>
> [1]https://lore.kernel.org/all/20190520063534.GB19312@shao2-debian/
>
> Suggested-by: Johannes Weiner <[email protected]>
> Signed-off-by: Yosry Ahmed <[email protected]>
> Tested-by: Domenico Cerasuolo <[email protected]>
Acked-by: Shakeel Butt <[email protected]>
On Wed, Nov 29, 2023 at 03:21:52AM +0000, Yosry Ahmed wrote:
> The workingset code flushes the stats in workingset_refault() to get
> accurate stats of the eviction memcg. In preparation for more scoped
> flushed and passing the eviction memcg to the flush call, move the call
> to workingset_test_recent() where we have a pointer to the eviction
> memcg.
>
> The flush call is sleepable, and cannot be made in an rcu read section.
> Hence, minimize the rcu read section by also moving it into
> workingset_test_recent(). Furthermore, instead of holding the rcu read
> lock throughout workingset_test_recent(), only hold it briefly to get a
> ref on the eviction memcg. This allows us to make the flush call after
> we get the eviction memcg.
>
> As for workingset_refault(), nothing else there appears to be protected
> by rcu. The memcg of the faulted folio (which is not necessarily the
> same as the eviction memcg) is protected by the folio lock, which is
> held from all callsites. Add a VM_BUG_ON() to make sure this doesn't
> change from under us.
>
> No functional change intended.
>
> Signed-off-by: Yosry Ahmed <[email protected]>
> Tested-by: Domenico Cerasuolo <[email protected]>
Acked-by: Shakeel Butt <[email protected]>
On Wed, Nov 29, 2023 at 03:21:53AM +0000, Yosry Ahmed wrote:
[...]
> +void mem_cgroup_flush_stats(struct mem_cgroup *memcg)
> {
> - if (memcg_should_flush_stats(root_mem_cgroup))
> - do_flush_stats();
> + static DEFINE_MUTEX(memcg_stats_flush_mutex);
> +
> + if (mem_cgroup_disabled())
> + return;
> +
> + if (!memcg)
> + memcg = root_mem_cgroup;
> +
> + if (memcg_should_flush_stats(memcg)) {
> + mutex_lock(&memcg_stats_flush_mutex);
What's the point of this mutex now? What is it providing? I understand
we can not try_lock here due to targeted flushing. Why not just let the
global rstat serialize the flushes? Actually this mutex can cause
latency hiccups as the mutex owner can get resched during flush and then
no one can flush for a potentially long time.
> + /* Check again after locking, another flush may have occurred */
> + if (memcg_should_flush_stats(memcg))
> + do_flush_stats(memcg);
> + mutex_unlock(&memcg_stats_flush_mutex);
> + }
> }
[..]
> > diff --git a/mm/workingset.c b/mm/workingset.c
> > index dce41577a49d2..7d3dacab8451a 100644
> > --- a/mm/workingset.c
> > +++ b/mm/workingset.c
> > @@ -464,8 +464,12 @@ bool workingset_test_recent(void *shadow, bool file, bool *workingset)
> >
> > rcu_read_unlock();
> >
> > - /* Flush stats (and potentially sleep) outside the RCU read section */
> > - mem_cgroup_flush_stats_ratelimited();
> > + /*
> > + * Flush stats (and potentially sleep) outside the RCU read section.
> > + * XXX: With per-memcg flushing and thresholding, is ratelimiting
> > + * still needed here?
> > + */
> > + mem_cgroup_flush_stats_ratelimited(eviction_memcg);
>
> What if flushing is not rate-limited (e.g. above line is commented)?
>
Hmm I think I might be misunderstanding the question. The call to
mem_cgroup_flush_stats_ratelimited() does not ratelimit other
flushers, it is rather a flush call that is itself ratelimited. IOW,
it may or may not flush based on when was the last time someone else
flushed.
This was introduced because flushing in the fault path was expensive
in some cases, so we wanted to avoid flushing if someone else recently
did a flush, as we don't expect a lot of pending changes in this case.
However, that was when flushing was always on the root level. Now that
we are flushing on the memcg level, it may no longer be needed as:
- The flush is more scoped, there should be less work to do.
- There is a per-memcg threshold now such that we only flush when
there are pending updates in this memcg.
This is why I added a comment that the ratelimited flush here may no
longer be needed. I didn't want to investigate this as part of this
series, especially that I do not have a reproducer for the fault
latency introduced by the flush before ratelimiting. Hence, I am
leaving the comment such that people know that this ratelimiting may
no longer be needed with this patch.
> >
> > eviction_lruvec = mem_cgroup_lruvec(eviction_memcg, pgdat);
> > refault = atomic_long_read(&eviction_lruvec->nonresident_age);
> > @@ -676,7 +680,7 @@ static unsigned long count_shadow_nodes(struct shrinker *shrinker,
> > struct lruvec *lruvec;
> > int i;
> >
> > - mem_cgroup_flush_stats();
> > + mem_cgroup_flush_stats(sc->memcg);
> > lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
> > for (pages = 0, i = 0; i < NR_LRU_LISTS; i++)
> > pages += lruvec_page_state_local(lruvec,
>
> Confused...
Which part is confusing? The call to mem_cgroup_flush_stats() now
receives a memcg argument as flushing is scoped to that memcg only to
avoid doing unnecessary work to flush other memcgs with global
flushing.
On Sat, Dec 2, 2023 at 12:31 AM Shakeel Butt <[email protected]> wrote:
>
> On Wed, Nov 29, 2023 at 03:21:53AM +0000, Yosry Ahmed wrote:
> [...]
> > +void mem_cgroup_flush_stats(struct mem_cgroup *memcg)
> > {
> > - if (memcg_should_flush_stats(root_mem_cgroup))
> > - do_flush_stats();
> > + static DEFINE_MUTEX(memcg_stats_flush_mutex);
> > +
> > + if (mem_cgroup_disabled())
> > + return;
> > +
> > + if (!memcg)
> > + memcg = root_mem_cgroup;
> > +
> > + if (memcg_should_flush_stats(memcg)) {
> > + mutex_lock(&memcg_stats_flush_mutex);
>
> What's the point of this mutex now? What is it providing? I understand
> we can not try_lock here due to targeted flushing. Why not just let the
> global rstat serialize the flushes? Actually this mutex can cause
> latency hiccups as the mutex owner can get resched during flush and then
> no one can flush for a potentially long time.
I was hoping this was clear from the commit message and code comments,
but apparently I was wrong, sorry. Let me give more context.
In previous versions and/or series, the mutex was only used with
flushes from userspace to guard in-kernel flushers against high
contention from userspace. Later on, I kept the mutex for all memcg
flushers for the following reasons:
(a) Allow waiters to sleep:
Unlike other flushers, the memcg flushing path can see a lot of
concurrency. The mutex avoids having a lot of CPUs spinning (e.g.
concurrent reclaimers) by allowing waiters to sleep.
(b) Check the threshold under lock but before calling cgroup_rstat_flush():
The calls to cgroup_rstat_flush() are not very cheap even if there's
nothing to flush, as we still need to iterate all CPUs. If flushers
contend directly on the rstat lock, overlapping flushes will
unnecessarily do the percpu iteration once they hold the lock. With
the mutex, they will check the threshold again once they hold the
mutex.
(c) Protect non-memcg flushers from contention from memcg flushers.
This is not as strong of an argument as protecting in-kernel flushers
from userspace flushers.
There has been discussions before about changing the rstat lock itself
to be a mutex, which would resolve (a), but there are concerns about
priority inversions if a low priority task holds the mutex and gets
preempted, as well as the amount of time the rstat lock holder keeps
the lock for:
https://lore.kernel.org/lkml/[email protected]/
I agree about possible hiccups due to the inner lock being dropped
while the mutex is held. Running a synthetic test with high
concurrency between reclaimers (in-kernel flushers) and stats readers
show no material performance difference with or without the mutex.
Maybe things cancel out, or don't really matter in practice.
I would prefer to keep the current code as I think (a) and (b) could
cause problems in the future, and the current form of the code (with
the mutex) has already seen mileage with production workloads.
On Mon, Dec 4, 2023 at 12:12 PM Yosry Ahmed <[email protected]> wrote:
>
> On Sat, Dec 2, 2023 at 12:31 AM Shakeel Butt <[email protected]> wrote:
> >
> > On Wed, Nov 29, 2023 at 03:21:53AM +0000, Yosry Ahmed wrote:
> > [...]
> > > +void mem_cgroup_flush_stats(struct mem_cgroup *memcg)
> > > {
> > > - if (memcg_should_flush_stats(root_mem_cgroup))
> > > - do_flush_stats();
> > > + static DEFINE_MUTEX(memcg_stats_flush_mutex);
> > > +
> > > + if (mem_cgroup_disabled())
> > > + return;
> > > +
> > > + if (!memcg)
> > > + memcg = root_mem_cgroup;
> > > +
> > > + if (memcg_should_flush_stats(memcg)) {
> > > + mutex_lock(&memcg_stats_flush_mutex);
> >
> > What's the point of this mutex now? What is it providing? I understand
> > we can not try_lock here due to targeted flushing. Why not just let the
> > global rstat serialize the flushes? Actually this mutex can cause
> > latency hiccups as the mutex owner can get resched during flush and then
> > no one can flush for a potentially long time.
>
> I was hoping this was clear from the commit message and code comments,
> but apparently I was wrong, sorry. Let me give more context.
>
> In previous versions and/or series, the mutex was only used with
> flushes from userspace to guard in-kernel flushers against high
> contention from userspace. Later on, I kept the mutex for all memcg
> flushers for the following reasons:
>
> (a) Allow waiters to sleep:
> Unlike other flushers, the memcg flushing path can see a lot of
> concurrency. The mutex avoids having a lot of CPUs spinning (e.g.
> concurrent reclaimers) by allowing waiters to sleep.
>
> (b) Check the threshold under lock but before calling cgroup_rstat_flush():
> The calls to cgroup_rstat_flush() are not very cheap even if there's
> nothing to flush, as we still need to iterate all CPUs. If flushers
> contend directly on the rstat lock, overlapping flushes will
> unnecessarily do the percpu iteration once they hold the lock. With
> the mutex, they will check the threshold again once they hold the
> mutex.
>
> (c) Protect non-memcg flushers from contention from memcg flushers.
> This is not as strong of an argument as protecting in-kernel flushers
> from userspace flushers.
>
> There has been discussions before about changing the rstat lock itself
> to be a mutex, which would resolve (a), but there are concerns about
> priority inversions if a low priority task holds the mutex and gets
> preempted, as well as the amount of time the rstat lock holder keeps
> the lock for:
> https://lore.kernel.org/lkml/[email protected]/
>
> I agree about possible hiccups due to the inner lock being dropped
> while the mutex is held. Running a synthetic test with high
> concurrency between reclaimers (in-kernel flushers) and stats readers
> show no material performance difference with or without the mutex.
> Maybe things cancel out, or don't really matter in practice.
>
> I would prefer to keep the current code as I think (a) and (b) could
> cause problems in the future, and the current form of the code (with
> the mutex) has already seen mileage with production workloads.
Correction: The priority inversion is possible on the memcg side due
to the mutex in this patch. Also, for point (a), the spinners will
eventually sleep once they hold the lock and hit the first CPU
boundary -- because of the lock dropping and cond_resched(). So
eventually, all spinners should be able to sleep, although it will be
a while until they do. With the mutex, they all sleep from the
beginning. Point (b) still holds though.
I am slightly inclined to keep the mutex but I can send a small fixlet
to remove it if others think otherwise.
Shakeel, Wei, any preferences?
On Mon, Dec 4, 2023 at 1:38 PM Yosry Ahmed <[email protected]> wrote:
>
> On Mon, Dec 4, 2023 at 12:12 PM Yosry Ahmed <[email protected]> wrote:
> >
> > On Sat, Dec 2, 2023 at 12:31 AM Shakeel Butt <[email protected]> wrote:
> > >
> > > On Wed, Nov 29, 2023 at 03:21:53AM +0000, Yosry Ahmed wrote:
> > > [...]
> > > > +void mem_cgroup_flush_stats(struct mem_cgroup *memcg)
> > > > {
> > > > - if (memcg_should_flush_stats(root_mem_cgroup))
> > > > - do_flush_stats();
> > > > + static DEFINE_MUTEX(memcg_stats_flush_mutex);
> > > > +
> > > > + if (mem_cgroup_disabled())
> > > > + return;
> > > > +
> > > > + if (!memcg)
> > > > + memcg = root_mem_cgroup;
> > > > +
> > > > + if (memcg_should_flush_stats(memcg)) {
> > > > + mutex_lock(&memcg_stats_flush_mutex);
> > >
> > > What's the point of this mutex now? What is it providing? I understand
> > > we can not try_lock here due to targeted flushing. Why not just let the
> > > global rstat serialize the flushes? Actually this mutex can cause
> > > latency hiccups as the mutex owner can get resched during flush and then
> > > no one can flush for a potentially long time.
> >
> > I was hoping this was clear from the commit message and code comments,
> > but apparently I was wrong, sorry. Let me give more context.
> >
> > In previous versions and/or series, the mutex was only used with
> > flushes from userspace to guard in-kernel flushers against high
> > contention from userspace. Later on, I kept the mutex for all memcg
> > flushers for the following reasons:
> >
> > (a) Allow waiters to sleep:
> > Unlike other flushers, the memcg flushing path can see a lot of
> > concurrency. The mutex avoids having a lot of CPUs spinning (e.g.
> > concurrent reclaimers) by allowing waiters to sleep.
> >
> > (b) Check the threshold under lock but before calling cgroup_rstat_flush():
> > The calls to cgroup_rstat_flush() are not very cheap even if there's
> > nothing to flush, as we still need to iterate all CPUs. If flushers
> > contend directly on the rstat lock, overlapping flushes will
> > unnecessarily do the percpu iteration once they hold the lock. With
> > the mutex, they will check the threshold again once they hold the
> > mutex.
> >
> > (c) Protect non-memcg flushers from contention from memcg flushers.
> > This is not as strong of an argument as protecting in-kernel flushers
> > from userspace flushers.
> >
> > There has been discussions before about changing the rstat lock itself
> > to be a mutex, which would resolve (a), but there are concerns about
> > priority inversions if a low priority task holds the mutex and gets
> > preempted, as well as the amount of time the rstat lock holder keeps
> > the lock for:
> > https://lore.kernel.org/lkml/[email protected]/
> >
> > I agree about possible hiccups due to the inner lock being dropped
> > while the mutex is held. Running a synthetic test with high
> > concurrency between reclaimers (in-kernel flushers) and stats readers
> > show no material performance difference with or without the mutex.
> > Maybe things cancel out, or don't really matter in practice.
> >
> > I would prefer to keep the current code as I think (a) and (b) could
> > cause problems in the future, and the current form of the code (with
> > the mutex) has already seen mileage with production workloads.
>
> Correction: The priority inversion is possible on the memcg side due
> to the mutex in this patch. Also, for point (a), the spinners will
> eventually sleep once they hold the lock and hit the first CPU
> boundary -- because of the lock dropping and cond_resched(). So
> eventually, all spinners should be able to sleep, although it will be
> a while until they do. With the mutex, they all sleep from the
> beginning. Point (b) still holds though.
>
> I am slightly inclined to keep the mutex but I can send a small fixlet
> to remove it if others think otherwise.
>
> Shakeel, Wei, any preferences?
My preference is to avoid the issue we know we see in production alot
i.e. priority inversion.
In future if you see issues with spinning then you can come up with
the lockless flush mechanism at that time.
On Mon, Dec 4, 2023 at 3:31 PM Shakeel Butt <[email protected]> wrote:
>
> On Mon, Dec 4, 2023 at 1:38 PM Yosry Ahmed <[email protected]> wrote:
> >
> > On Mon, Dec 4, 2023 at 12:12 PM Yosry Ahmed <[email protected]> wrote:
> > >
> > > On Sat, Dec 2, 2023 at 12:31 AM Shakeel Butt <[email protected]> wrote:
> > > >
> > > > On Wed, Nov 29, 2023 at 03:21:53AM +0000, Yosry Ahmed wrote:
> > > > [...]
> > > > > +void mem_cgroup_flush_stats(struct mem_cgroup *memcg)
> > > > > {
> > > > > - if (memcg_should_flush_stats(root_mem_cgroup))
> > > > > - do_flush_stats();
> > > > > + static DEFINE_MUTEX(memcg_stats_flush_mutex);
> > > > > +
> > > > > + if (mem_cgroup_disabled())
> > > > > + return;
> > > > > +
> > > > > + if (!memcg)
> > > > > + memcg = root_mem_cgroup;
> > > > > +
> > > > > + if (memcg_should_flush_stats(memcg)) {
> > > > > + mutex_lock(&memcg_stats_flush_mutex);
> > > >
> > > > What's the point of this mutex now? What is it providing? I understand
> > > > we can not try_lock here due to targeted flushing. Why not just let the
> > > > global rstat serialize the flushes? Actually this mutex can cause
> > > > latency hiccups as the mutex owner can get resched during flush and then
> > > > no one can flush for a potentially long time.
> > >
> > > I was hoping this was clear from the commit message and code comments,
> > > but apparently I was wrong, sorry. Let me give more context.
> > >
> > > In previous versions and/or series, the mutex was only used with
> > > flushes from userspace to guard in-kernel flushers against high
> > > contention from userspace. Later on, I kept the mutex for all memcg
> > > flushers for the following reasons:
> > >
> > > (a) Allow waiters to sleep:
> > > Unlike other flushers, the memcg flushing path can see a lot of
> > > concurrency. The mutex avoids having a lot of CPUs spinning (e.g.
> > > concurrent reclaimers) by allowing waiters to sleep.
> > >
> > > (b) Check the threshold under lock but before calling cgroup_rstat_flush():
> > > The calls to cgroup_rstat_flush() are not very cheap even if there's
> > > nothing to flush, as we still need to iterate all CPUs. If flushers
> > > contend directly on the rstat lock, overlapping flushes will
> > > unnecessarily do the percpu iteration once they hold the lock. With
> > > the mutex, they will check the threshold again once they hold the
> > > mutex.
> > >
> > > (c) Protect non-memcg flushers from contention from memcg flushers.
> > > This is not as strong of an argument as protecting in-kernel flushers
> > > from userspace flushers.
> > >
> > > There has been discussions before about changing the rstat lock itself
> > > to be a mutex, which would resolve (a), but there are concerns about
> > > priority inversions if a low priority task holds the mutex and gets
> > > preempted, as well as the amount of time the rstat lock holder keeps
> > > the lock for:
> > > https://lore.kernel.org/lkml/[email protected]/
> > >
> > > I agree about possible hiccups due to the inner lock being dropped
> > > while the mutex is held. Running a synthetic test with high
> > > concurrency between reclaimers (in-kernel flushers) and stats readers
> > > show no material performance difference with or without the mutex.
> > > Maybe things cancel out, or don't really matter in practice.
> > >
> > > I would prefer to keep the current code as I think (a) and (b) could
> > > cause problems in the future, and the current form of the code (with
> > > the mutex) has already seen mileage with production workloads.
> >
> > Correction: The priority inversion is possible on the memcg side due
> > to the mutex in this patch. Also, for point (a), the spinners will
> > eventually sleep once they hold the lock and hit the first CPU
> > boundary -- because of the lock dropping and cond_resched(). So
> > eventually, all spinners should be able to sleep, although it will be
> > a while until they do. With the mutex, they all sleep from the
> > beginning. Point (b) still holds though.
> >
> > I am slightly inclined to keep the mutex but I can send a small fixlet
> > to remove it if others think otherwise.
> >
> > Shakeel, Wei, any preferences?
>
> My preference is to avoid the issue we know we see in production alot
> i.e. priority inversion.
>
> In future if you see issues with spinning then you can come up with
> the lockless flush mechanism at that time.
Given that the synthetic high concurrency test doesn't show material
performance difference between the mutex and non-mutex versions, I
agree that the mutex can be taken out from this patch set (one less
global mutex to worry about).
On Mon, Dec 4, 2023 at 3:46 PM Wei Xu <[email protected]> wrote:
>
> On Mon, Dec 4, 2023 at 3:31 PM Shakeel Butt <[email protected]> wrote:
> >
> > On Mon, Dec 4, 2023 at 1:38 PM Yosry Ahmed <[email protected]> wrote:
> > >
> > > On Mon, Dec 4, 2023 at 12:12 PM Yosry Ahmed <[email protected]> wrote:
> > > >
> > > > On Sat, Dec 2, 2023 at 12:31 AM Shakeel Butt <[email protected]> wrote:
> > > > >
> > > > > On Wed, Nov 29, 2023 at 03:21:53AM +0000, Yosry Ahmed wrote:
> > > > > [...]
> > > > > > +void mem_cgroup_flush_stats(struct mem_cgroup *memcg)
> > > > > > {
> > > > > > - if (memcg_should_flush_stats(root_mem_cgroup))
> > > > > > - do_flush_stats();
> > > > > > + static DEFINE_MUTEX(memcg_stats_flush_mutex);
> > > > > > +
> > > > > > + if (mem_cgroup_disabled())
> > > > > > + return;
> > > > > > +
> > > > > > + if (!memcg)
> > > > > > + memcg = root_mem_cgroup;
> > > > > > +
> > > > > > + if (memcg_should_flush_stats(memcg)) {
> > > > > > + mutex_lock(&memcg_stats_flush_mutex);
> > > > >
> > > > > What's the point of this mutex now? What is it providing? I understand
> > > > > we can not try_lock here due to targeted flushing. Why not just let the
> > > > > global rstat serialize the flushes? Actually this mutex can cause
> > > > > latency hiccups as the mutex owner can get resched during flush and then
> > > > > no one can flush for a potentially long time.
> > > >
> > > > I was hoping this was clear from the commit message and code comments,
> > > > but apparently I was wrong, sorry. Let me give more context.
> > > >
> > > > In previous versions and/or series, the mutex was only used with
> > > > flushes from userspace to guard in-kernel flushers against high
> > > > contention from userspace. Later on, I kept the mutex for all memcg
> > > > flushers for the following reasons:
> > > >
> > > > (a) Allow waiters to sleep:
> > > > Unlike other flushers, the memcg flushing path can see a lot of
> > > > concurrency. The mutex avoids having a lot of CPUs spinning (e.g.
> > > > concurrent reclaimers) by allowing waiters to sleep.
> > > >
> > > > (b) Check the threshold under lock but before calling cgroup_rstat_flush():
> > > > The calls to cgroup_rstat_flush() are not very cheap even if there's
> > > > nothing to flush, as we still need to iterate all CPUs. If flushers
> > > > contend directly on the rstat lock, overlapping flushes will
> > > > unnecessarily do the percpu iteration once they hold the lock. With
> > > > the mutex, they will check the threshold again once they hold the
> > > > mutex.
> > > >
> > > > (c) Protect non-memcg flushers from contention from memcg flushers.
> > > > This is not as strong of an argument as protecting in-kernel flushers
> > > > from userspace flushers.
> > > >
> > > > There has been discussions before about changing the rstat lock itself
> > > > to be a mutex, which would resolve (a), but there are concerns about
> > > > priority inversions if a low priority task holds the mutex and gets
> > > > preempted, as well as the amount of time the rstat lock holder keeps
> > > > the lock for:
> > > > https://lore.kernel.org/lkml/[email protected]/
> > > >
> > > > I agree about possible hiccups due to the inner lock being dropped
> > > > while the mutex is held. Running a synthetic test with high
> > > > concurrency between reclaimers (in-kernel flushers) and stats readers
> > > > show no material performance difference with or without the mutex.
> > > > Maybe things cancel out, or don't really matter in practice.
> > > >
> > > > I would prefer to keep the current code as I think (a) and (b) could
> > > > cause problems in the future, and the current form of the code (with
> > > > the mutex) has already seen mileage with production workloads.
> > >
> > > Correction: The priority inversion is possible on the memcg side due
> > > to the mutex in this patch. Also, for point (a), the spinners will
> > > eventually sleep once they hold the lock and hit the first CPU
> > > boundary -- because of the lock dropping and cond_resched(). So
> > > eventually, all spinners should be able to sleep, although it will be
> > > a while until they do. With the mutex, they all sleep from the
> > > beginning. Point (b) still holds though.
> > >
> > > I am slightly inclined to keep the mutex but I can send a small fixlet
> > > to remove it if others think otherwise.
> > >
> > > Shakeel, Wei, any preferences?
> >
> > My preference is to avoid the issue we know we see in production alot
> > i.e. priority inversion.
> >
> > In future if you see issues with spinning then you can come up with
> > the lockless flush mechanism at that time.
>
> Given that the synthetic high concurrency test doesn't show material
> performance difference between the mutex and non-mutex versions, I
> agree that the mutex can be taken out from this patch set (one less
> global mutex to worry about).
Thanks Wei and Shakeel for your input.
Andrew, could you please squash in the fixlet below and remove the
paragraph starting with "Add a mutex to.." from the commit message?
From 19af26e01f93cbf0806d75a234b78e48c1ce9d80 Mon Sep 17 00:00:00 2001
From: Yosry Ahmed <[email protected]>
Date: Mon, 4 Dec 2023 23:43:29 +0000
Subject: [PATCH] mm: memcg: remove stats flushing mutex
The mutex was intended to make the waiters sleep instead of spin, and
such that we can check the update thresholds again after acquiring the
mutex. However, the mutex has a risk of priority inversion, especially
since the underlying rstat lock can de dropped while the mutex is held.
Synthetic testing with high concurrency of flushers shows no
regressions without the mutex, so remove it.
Suggested-by: Shakeel Butt <[email protected]>
Signed-off-by: Yosry Ahmed <[email protected]>
---
mm/memcontrol.c | 11 ++---------
1 file changed, 2 insertions(+), 9 deletions(-)
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 5d300318bf18a..0563625767349 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -749,21 +749,14 @@ static void do_flush_stats(struct mem_cgroup *memcg)
*/
void mem_cgroup_flush_stats(struct mem_cgroup *memcg)
{
- static DEFINE_MUTEX(memcg_stats_flush_mutex);
-
if (mem_cgroup_disabled())
return;
if (!memcg)
memcg = root_mem_cgroup;
- if (memcg_should_flush_stats(memcg)) {
- mutex_lock(&memcg_stats_flush_mutex);
- /* Check again after locking, another flush may have occurred */
- if (memcg_should_flush_stats(memcg))
- do_flush_stats(memcg);
- mutex_unlock(&memcg_stats_flush_mutex);
- }
+ if (memcg_should_flush_stats(memcg))
+ do_flush_stats(memcg);
}
void mem_cgroup_flush_stats_ratelimited(struct mem_cgroup *memcg)
--
2.43.0.rc2.451.g8631bc7472-goog
On Mon, Dec 04, 2023 at 03:49:01PM -0800, Yosry Ahmed wrote:
[...]
>
> From 19af26e01f93cbf0806d75a234b78e48c1ce9d80 Mon Sep 17 00:00:00 2001
> From: Yosry Ahmed <[email protected]>
> Date: Mon, 4 Dec 2023 23:43:29 +0000
> Subject: [PATCH] mm: memcg: remove stats flushing mutex
>
> The mutex was intended to make the waiters sleep instead of spin, and
> such that we can check the update thresholds again after acquiring the
> mutex. However, the mutex has a risk of priority inversion, especially
> since the underlying rstat lock can de dropped while the mutex is held.
>
> Synthetic testing with high concurrency of flushers shows no
> regressions without the mutex, so remove it.
>
> Suggested-by: Shakeel Butt <[email protected]>
> Signed-off-by: Yosry Ahmed <[email protected]>
Acked-by: Shakeel Butt <[email protected]>
On Mon, 4 Dec 2023 23:58:56 +0000 Shakeel Butt <[email protected]> wrote:
> On Mon, Dec 04, 2023 at 03:49:01PM -0800, Yosry Ahmed wrote:
> [...]
> >
> > From 19af26e01f93cbf0806d75a234b78e48c1ce9d80 Mon Sep 17 00:00:00 2001
> > From: Yosry Ahmed <[email protected]>
> > Date: Mon, 4 Dec 2023 23:43:29 +0000
> > Subject: [PATCH] mm: memcg: remove stats flushing mutex
> >
> > The mutex was intended to make the waiters sleep instead of spin, and
> > such that we can check the update thresholds again after acquiring the
> > mutex. However, the mutex has a risk of priority inversion, especially
> > since the underlying rstat lock can de dropped while the mutex is held.
> >
> > Synthetic testing with high concurrency of flushers shows no
> > regressions without the mutex, so remove it.
> >
> > Suggested-by: Shakeel Butt <[email protected]>
> > Signed-off-by: Yosry Ahmed <[email protected]>
>
> Acked-by: Shakeel Butt <[email protected]>
>
I'd like to move this series into mm-stable soon. Are we all OK with that?
On Tue, Dec 12, 2023 at 10:43 AM Andrew Morton
<[email protected]> wrote:
>
> On Mon, 4 Dec 2023 23:58:56 +0000 Shakeel Butt <[email protected]> wrote:
>
> > On Mon, Dec 04, 2023 at 03:49:01PM -0800, Yosry Ahmed wrote:
> > [...]
> > >
> > > From 19af26e01f93cbf0806d75a234b78e48c1ce9d80 Mon Sep 17 00:00:00 2001
> > > From: Yosry Ahmed <[email protected]>
> > > Date: Mon, 4 Dec 2023 23:43:29 +0000
> > > Subject: [PATCH] mm: memcg: remove stats flushing mutex
> > >
> > > The mutex was intended to make the waiters sleep instead of spin, and
> > > such that we can check the update thresholds again after acquiring the
> > > mutex. However, the mutex has a risk of priority inversion, especially
> > > since the underlying rstat lock can de dropped while the mutex is held.
> > >
> > > Synthetic testing with high concurrency of flushers shows no
> > > regressions without the mutex, so remove it.
> > >
> > > Suggested-by: Shakeel Butt <[email protected]>
> > > Signed-off-by: Yosry Ahmed <[email protected]>
> >
> > Acked-by: Shakeel Butt <[email protected]>
> >
>
> I'd like to move this series into mm-stable soon. Are we all OK with that?
OK from me.
On Tue, Dec 12, 2023 at 10:43 AM Andrew Morton
<[email protected]> wrote:
>
> On Mon, 4 Dec 2023 23:58:56 +0000 Shakeel Butt <[email protected]> wrote:
>
> > On Mon, Dec 04, 2023 at 03:49:01PM -0800, Yosry Ahmed wrote:
> > [...]
> > >
> > > From 19af26e01f93cbf0806d75a234b78e48c1ce9d80 Mon Sep 17 00:00:00 2001
> > > From: Yosry Ahmed <[email protected]>
> > > Date: Mon, 4 Dec 2023 23:43:29 +0000
> > > Subject: [PATCH] mm: memcg: remove stats flushing mutex
> > >
> > > The mutex was intended to make the waiters sleep instead of spin, and
> > > such that we can check the update thresholds again after acquiring the
> > > mutex. However, the mutex has a risk of priority inversion, especially
> > > since the underlying rstat lock can de dropped while the mutex is held.
> > >
> > > Synthetic testing with high concurrency of flushers shows no
> > > regressions without the mutex, so remove it.
> > >
> > > Suggested-by: Shakeel Butt <[email protected]>
> > > Signed-off-by: Yosry Ahmed <[email protected]>
> >
> > Acked-by: Shakeel Butt <[email protected]>
> >
>
> I'd like to move this series into mm-stable soon. Are we all OK with that?
Looking forward to that :)