fold per-CPU vmstats remotely
Subsystems (12)
PATCH v5 5/12 add this_cpu_cmpxchg_local and asm-generic definitions
In reply to Marcelo TosattiGoal is to have vmstat_shepherd to transfer from per-CPU counters to global counters remotely. For this, an atomic this_cpu_cmpxchg is necessary.
Add this_cpu_cmpxchg_local_ helpers to asm-generic/percpu.h.
include/asm-generic/percpu.h +17 -0
@@ -424,6 +424,23 @@ do { \ this_cpu_generic_cmpxchg(pcp, oval, nval)#endif +#ifndef this_cpu_cmpxchg_local_1 +#define this_cpu_cmpxchg_local_1 (pcp, oval, nval) \ + this_cpu_generic_cmpxchg(pcp, oval, nval) +#endif +#ifndef this_cpu_cmpxchg_local_2 +#define this_cpu_cmpxchg_local_2 (pcp, oval, nval) \ + this_cpu_generic_cmpxchg(pcp, oval, nval) +#endif +#ifndef this_cpu_cmpxchg_local_4 +#define this_cpu_cmpxchg_local_4 (pcp, oval, nval) \ + this_cpu_generic_cmpxchg(pcp, oval, nval) +#endif +#ifndef this_cpu_cmpxchg_local_8 +#define this_cpu_cmpxchg_local_8 (pcp, oval, nval) \ + this_cpu_generic_cmpxchg(pcp, oval, nval) +#endif +#ifndef this_cpu_cmpxchg_double_1#define this_cpu_cmpxchg_double_1 (pcp1, pcp2, oval1, oval2, nval1, nval2) \ this_cpu_generic_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)
include/linux/percpu-defs.h +2 -0
@@ -513,6 +513,8 @@ do { \#define this_cpu_xchg (pcp, nval) __pcpu_size_call_return2(this_cpu_xchg_, pcp, nval)#define this_cpu_cmpxchg (pcp, oval, nval) \ __pcpu_size_call_return2(this_cpu_cmpxchg_, pcp, oval, nval) +#define this_cpu_cmpxchg_local (pcp, oval, nval) \ + __pcpu_size_call_return2(this_cpu_cmpxchg_local_, pcp, oval, nval)#define this_cpu_cmpxchg_double (pcp1, pcp2, oval1, oval2, nval1, nval2) \ __pcpu_double_call_return_bool(this_cpu_cmpxchg_double_, pcp1, pcp2, oval1, oval2, nval1, nval2)
PATCH v5 12/12 vmstat: add pcp remote node draining via cpu_vm_stats_fold
In reply to Marcelo TosattiLarge NUMA systems might have significant portions
of system memory to be trapped in pcp queues. The number of pcp is determined by the number of processors and nodes in a system. A system with 4 processors and 2 nodes has 8 pcps which is okay. But a system with 1024 processors and 512 nodes has 512k pcps with a high potential for large amount of memory being caught in them.
Enable remote node draining for the CONFIG_HAVE_CMPXCHG_LOCAL case, where vmstat_shepherd will perform the aging and draining via cpu_vm_stats_fold.
mm/vmstat.c +42 -2
@@ -928,7 +928,7 @@ static int refresh_cpu_vm_stats(bool do_* There cannot be any access by the offline cpu and therefore * synchronization is simplified. */ -void cpu_vm_stats_fold (int cpu) +void cpu_vm_stats_fold (int cpu,bool do_pagesets) {struct pglist_data *pgdat;struct zone *zone;
@@ -938,6 +938,9 @@ void cpu_vm_stats_fold(int cpu)for_each_populated_zone (zone) {struct per_cpu_zonestat *pzstats; +#ifdef CONFIG_NUMA +struct per_cpu_pages *pcp =per_cpu_ptr (zone->per_cpu_pageset, cpu); +#endif pzstats =per_cpu_ptr (zone->per_cpu_zonestats, cpu);
@@ -948,6 +951,11 @@ void cpu_vm_stats_fold(int cpu) v =xchg (&pzstats->vm_stat_diff[i],0 );atomic_long_add (v, &zone->vm_stat[i]); global_zone_diff[i] += v; +#ifdef CONFIG_NUMA +/* 3 seconds idle till flush */ +if (do_pagesets) + pcp->expire =3 ; +#endif } }#ifdef CONFIG_NUMA
@@ -959,6 +967,38 @@ void cpu_vm_stats_fold(int cpu) zone_numa_event_add(v, zone, i); } } + +if (do_pagesets) { +cond_resched (); +/* +* Deal with draining the remote pageset of a +* processor +* +* Check if there are pages remaining in this pageset +* if not then there is nothing to expire. +*/ +if (!pcp->expire || !pcp->count) +continue ; + +/* +* We never drain zones local to this processor. +*/ +if (zone_to_nid (zone) ==cpu_to_node (cpu)) { + pcp->expire =0 ; +continue ; + } + +WARN_ON (pcp->expire <0 ); +/* +* pcp->expire is only accessed from vmstat_shepherd context, +* therefore no locking is required. +*/ +if (--pcp->expire) +continue ; + +if (pcp->count) +drain_zone_pages (zone, pcp); + }#endif }
@@ -2060,7 +2100,7 @@ static int refresh_all_vm_stats(void)cpus_read_lock ();for_each_online_cpu (cpu) { -cpu_vm_stats_fold (cpu); +cpu_vm_stats_fold (cpu, true);cond_resched (); }cpus_read_unlock ();
include/linux/vmstat.h +1 -1
@@ -291,7 +291,7 @@ extern void __dec_zone_state(struct zoneextern void __dec_node_state (struct pglist_data *,enum node_stat_item );void quiet_vmstat (void ); -void cpu_vm_stats_fold (int cpu); +void cpu_vm_stats_fold (int cpu,bool do_pagesets);void refresh_zone_stat_thresholds (void );struct ctl_table ;
mm/page_alloc.c +1 -1
@@ -8629,7 +8629,7 @@ static int page_alloc_cpu_dead(unsigned* Zero the differential counters of the dead processor * so that the vm statistics are consistent. */ -cpu_vm_stats_fold (cpu); +cpu_vm_stats_fold (cpu, false);for_each_populated_zone (zone)zone_pcp_update (zone,0 );
v5 cover fold per-CPU vmstats remotely
This patch series addresses the following two problems:
1. A customer provided some evidence which indicates that the idle tick was stopped; albeit, CPU-specific vmstat counters still remained populated.
Thus one can only assume quiet_vmstat() was not invoked on return to the idle loop. If I understand correctly, I suspect this divergence might erroneously prevent a reclaim attempt by kswapd. If the number of zone specific free pages are below their per-cpu drift value then zone_page_state_snapshot() is used to compute a more accurate view of the aforementioned statistic. Thus any task blocked on the NUMA node specific pfmemalloc_wait queue will be unable to make significant progress via direct reclaim unless it is killed after being woken up by kswapd
(see throttle_direct_reclaim())
2. With a SCHED_FIFO task that busy loops on a given CPU, and kworker for that CPU at SCHED_OTHER priority, queuing work to sync per-vmstats will either cause that work to never execute, or stalld (i.e. stall daemon) boosts kworker priority which causes a latency violation
By having vmstat_shepherd flush the per-CPU counters to the global counters from remote CPUs.
This is done using cmpxchg to manipulate the counters, both CPU locally (via the account functions), and remotely (via cpu_vm_stats_fold).
Thanks to Aaron Tomlin for diagnosing issue 1 and writing the initial patch series.
v5:
- Drop "mm/vmstat: remove remote node draining" (Vlastimil Babka)
- Implement remote node draining for cpu_vm_stats_fold (Vlastimil Babka)
v4:
- Switch per-CPU vmstat counters to s32, required by RISC-V, ARC architectures
v3:
- Removed unused drain_zone_pages and changes variable (David Hildenbrand)
- Use xchg instead of cmpxchg in refresh_cpu_vm_stats (Peter Xu)
- Add drain_all_pages to vmstat_refresh to make
stats more accurate (Peter Xu)
- Improve changelog of
"mm/vmstat: switch counter modification to cmpxchg" (Peter Xu / David)
- Improve changelog of
"mm/vmstat: remove remote node draining" (David Hildenbrand)
v2:
- actually use LOCK CMPXCHG on counter mod/inc/dec functions (Christoph Lameter)
- use try_cmpxchg for cmpxchg loops
(Uros Bizjak / Matthew Wilcox)
arch/arm64/include/asm/percpu.h | 16 ++ arch/loongarch/include/asm/percpu.h | 23 +++- arch/s390/include/asm/percpu.h | 5 arch/x86/include/asm/percpu.h | 39 +++---- include/asm-generic/percpu.h | 17 +++ include/linux/mmzone.h | 3 include/linux/percpu-defs.h | 2 kernel/fork.c | 2 kernel/scs.c | 2 mm/page_alloc.c | 23 ---- mm/vmstat.c | 424 +++++++++++++++++++++++++++++++++++++++++------------------------------------ 11 files changed, 307 insertions(+), 249 deletions(-)
PATCH v5 3/12 this_cpu_cmpxchg: S390: switch this_cpu_cmpxchg to locked, add _local function
Goal is to have vmstat_shepherd to transfer from per-CPU counters to global counters remotely. For this, an atomic this_cpu_cmpxchg is necessary.
Following the kernel convention for cmpxchg/cmpxchg_local, add S390's this_cpu_cmpxchg_local.
arch/s390/include/asm/percpu.h +5 -0
@@ -148,6 +148,11 @@#define this_cpu_cmpxchg_4 (pcp, oval, nval) arch_this_cpu_cmpxchg(pcp, oval, nval)#define this_cpu_cmpxchg_8 (pcp, oval, nval) arch_this_cpu_cmpxchg(pcp, oval, nval) +#define this_cpu_cmpxchg_local_1 (pcp, oval, nval) arch_this_cpu_cmpxchg(pcp, oval, nval) +#define this_cpu_cmpxchg_local_2 (pcp, oval, nval) arch_this_cpu_cmpxchg(pcp, oval, nval) +#define this_cpu_cmpxchg_local_4 (pcp, oval, nval) arch_this_cpu_cmpxchg(pcp, oval, nval) +#define this_cpu_cmpxchg_local_8 (pcp, oval, nval) arch_this_cpu_cmpxchg(pcp, oval, nval) +#define arch_this_cpu_xchg (pcp, nval) \ ({ \ typeof(pcp) *ptr__; \
PATCH v5 2/12 this_cpu_cmpxchg: loongarch: switch this_cpu_cmpxchg to locked, add _local function
In reply to Marcelo TosattiGoal is to have vmstat_shepherd to transfer from per-CPU counters to global counters remotely. For this, an atomic this_cpu_cmpxchg is necessary.
Following the kernel convention for cmpxchg/cmpxchg_local, add this_cpu_cmpxchg_local helpers to Loongarch.
arch/loongarch/include/asm/percpu.h +19 -4
@@ -150,6 +150,16 @@ static inline unsigned long __percpu_xch }/* this_cpu_cmpxchg */ +#define _protect_cmpxchg (pcp, o, n) \ +({ \ + typeof(*raw_cpu_ptr(&(pcp))) __ret; \ + preempt_disable_notrace(); \ + __ret = cmpxchg(raw_cpu_ptr(&(pcp)), o, n); \ + preempt_enable_notrace(); \ + __ret; \ +}) + +/* this_cpu_cmpxchg_local */ #define _protect_cmpxchg_local (pcp, o, n) \ ({ \ typeof(*raw_cpu_ptr(&(pcp))) __ret; \
@@ -222,10 +232,15 @@ do { \#define this_cpu_xchg_4 (pcp, val) _percpu_xchg(pcp, val)#define this_cpu_xchg_8 (pcp, val) _percpu_xchg(pcp, val) -#define this_cpu_cmpxchg_1 (ptr, o, n) _protect_cmpxchg_local(ptr, o, n) -#define this_cpu_cmpxchg_2 (ptr, o, n) _protect_cmpxchg_local(ptr, o, n) -#define this_cpu_cmpxchg_4 (ptr, o, n) _protect_cmpxchg_local(ptr, o, n) -#define this_cpu_cmpxchg_8 (ptr, o, n) _protect_cmpxchg_local(ptr, o, n) +#define this_cpu_cmpxchg_local_1 (ptr, o, n) _protect_cmpxchg_local(ptr, o, n) +#define this_cpu_cmpxchg_local_2 (ptr, o, n) _protect_cmpxchg_local(ptr, o, n) +#define this_cpu_cmpxchg_local_4 (ptr, o, n) _protect_cmpxchg_local(ptr, o, n) +#define this_cpu_cmpxchg_local_8 (ptr, o, n) _protect_cmpxchg_local(ptr, o, n) + +#define this_cpu_cmpxchg_1 (ptr, o, n) _protect_cmpxchg(ptr, o, n) +#define this_cpu_cmpxchg_2 (ptr, o, n) _protect_cmpxchg(ptr, o, n) +#define this_cpu_cmpxchg_4 (ptr, o, n) _protect_cmpxchg(ptr, o, n) +#define this_cpu_cmpxchg_8 (ptr, o, n) _protect_cmpxchg(ptr, o, n)#include <asm-generic/percpu.h>
PATCH v5 6/12 convert this_cpu_cmpxchg users to this_cpu_cmpxchg_local
In reply to Marcelo Tosattithis_cpu_cmpxchg was modified to atomic version, which can be more costly than non-atomic version.
Switch users of this_cpu_cmpxchg to this_cpu_cmpxchg_local (which preserves pre-non-atomic this_cpu_cmpxchg behaviour).
kernel/fork.c +1 -1
@@ -203,7 +203,7 @@ static bool try_release_thread_stack_to_unsigned int i;for (i =0 ; i <NR_CACHED_STACKS ; i++) { -if (this_cpu_cmpxchg (cached_stacks[i],NULL , vm) !=NULL ) +if (this_cpu_cmpxchg_local (cached_stacks[i],NULL , vm) !=NULL )continue ;return true; }
kernel/scs.c +1 -1
@@ -83,7 +83,7 @@ void scs_free(void *s)*/ for (i =0 ; i <NR_CACHED_SCS ; i++) -if (this_cpu_cmpxchg (scs_cache[i],0 , s) ==NULL ) +if (this_cpu_cmpxchg_local (scs_cache[i],0 , s) ==NULL )return ;kasan_unpoison_vmalloc (s,SCS_SIZE ,KASAN_VMALLOC_PROT_NORMAL );
PATCH v5 8/12 vmstat: switch per-cpu vmstat counters to 32-bits
In reply to Marcelo TosattiSome architectures only provide xchg/cmpxchg in 32/64-bit quantities.
Since the next patch is about to use xchg on per-CPU vmstat counters, switch them to s32.
include/linux/mmzone.h +4 -4
@@ -689,8 +689,8 @@ struct per_cpu_pages {struct per_cpu_zonestat {#ifdef CONFIG_SMP -s8 vm_stat_diff[NR_VM_ZONE_STAT_ITEMS ]; -s8 stat_threshold; +s32 vm_stat_diff[NR_VM_ZONE_STAT_ITEMS ]; +s32 stat_threshold;#endif #ifdef CONFIG_NUMA /*
@@ -703,8 +703,8 @@ struct per_cpu_zonestat { };struct per_cpu_nodestat { -s8 stat_threshold; -s8 vm_node_stat_diff[NR_VM_NODE_STAT_ITEMS ]; +s32 stat_threshold; +s32 vm_node_stat_diff[NR_VM_NODE_STAT_ITEMS ]; };#endif /* !__GENERATING_BOUNDS.H */
mm/vmstat.c +16 -16
@@ -351,7 +351,7 @@ static inline void mod_zone_state(struct long delta, int overstep_mode) {struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats; -s8 __percpu *p = pcp->vm_stat_diff + item; +s32 __percpu *p = pcp->vm_stat_diff + item;long o, n, t, z; do {
@@ -428,7 +428,7 @@ static inline void mod_node_state(struct int delta, int overstep_mode) {struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats; -s8 __percpu *p = pcp->vm_node_stat_diff + item; +s32 __percpu *p = pcp->vm_node_stat_diff + item;long o, n, t, z;if (vmstat_item_in_bytes (item)) {
@@ -525,7 +525,7 @@ void __mod_zone_page_state(struct zone * long delta) {struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats; -s8 __percpu *p = pcp->vm_stat_diff + item; +s32 __percpu *p = pcp->vm_stat_diff + item;long x;long t;
@@ -556,7 +556,7 @@ void __mod_node_page_state(struct pglist long delta) {struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats; -s8 __percpu *p = pcp->vm_node_stat_diff + item; +s32 __percpu *p = pcp->vm_node_stat_diff + item;long x;long t;
@@ -614,8 +614,8 @@ EXPORT_SYMBOL(__mod_node_page_state);void __inc_zone_state (struct zone *zone,enum zone_stat_item item) {struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats; -s8 __percpu *p = pcp->vm_stat_diff + item; -s8 v, t; +s32 __percpu *p = pcp->vm_stat_diff + item; +s32 v, t;/* See __mod_node_page_state */ preempt_disable_nested ();
@@ -623,7 +623,7 @@ void __inc_zone_state(struct zone *zone, v =__this_cpu_inc_return (*p); t =__this_cpu_read (pcp->stat_threshold);if (unlikely (v > t)) { -s8 overstep = t >>1 ; +s32 overstep = t >>1 ;zone_page_state_add (v + overstep, zone, item);__this_cpu_write (*p, -overstep);
@@ -635,8 +635,8 @@ void __inc_zone_state(struct zone *zone,void __inc_node_state (struct pglist_data *pgdat,enum node_stat_item item) {struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats; -s8 __percpu *p = pcp->vm_node_stat_diff + item; -s8 v, t; +s32 __percpu *p = pcp->vm_node_stat_diff + item; +s32 v, t;VM_WARN_ON_ONCE (vmstat_item_in_bytes (item));
@@ -646,7 +646,7 @@ void __inc_node_state(struct pglist_data v =__this_cpu_inc_return (*p); t =__this_cpu_read (pcp->stat_threshold);if (unlikely (v > t)) { -s8 overstep = t >>1 ; +s32 overstep = t >>1 ;node_page_state_add (v + overstep, pgdat, item);__this_cpu_write (*p, -overstep);
@@ -670,8 +670,8 @@ EXPORT_SYMBOL(__inc_node_page_state);void __dec_zone_state (struct zone *zone,enum zone_stat_item item) {struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats; -s8 __percpu *p = pcp->vm_stat_diff + item; -s8 v, t; +s32 __percpu *p = pcp->vm_stat_diff + item; +s32 v, t;/* See __mod_node_page_state */ preempt_disable_nested ();
@@ -679,7 +679,7 @@ void __dec_zone_state(struct zone *zone, v =__this_cpu_dec_return (*p); t =__this_cpu_read (pcp->stat_threshold);if (unlikely (v < - t)) { -s8 overstep = t >>1 ; +s32 overstep = t >>1 ;zone_page_state_add (v - overstep, zone, item);__this_cpu_write (*p, overstep);
@@ -691,8 +691,8 @@ void __dec_zone_state(struct zone *zone,void __dec_node_state (struct pglist_data *pgdat,enum node_stat_item item) {struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats; -s8 __percpu *p = pcp->vm_node_stat_diff + item; -s8 v, t; +s32 __percpu *p = pcp->vm_node_stat_diff + item; +s32 v, t;VM_WARN_ON_ONCE (vmstat_item_in_bytes (item));
@@ -702,7 +702,7 @@ void __dec_node_state(struct pglist_data v =__this_cpu_dec_return (*p); t =__this_cpu_read (pcp->stat_threshold);if (unlikely (v < - t)) { -s8 overstep = t >>1 ; +s32 overstep = t >>1 ;node_page_state_add (v - overstep, pgdat, item);__this_cpu_write (*p, overstep);
PATCH v5 10/12 mm/vmstat: switch vmstat shepherd to flush per-CPU counters remotely
In reply to Marcelo TosattiNow that the counters are modified via cmpxchg both CPU locally (via the account functions), and remotely (via cpu_vm_stats_fold), its possible to switch vmstat_shepherd to perform the per-CPU vmstats folding remotely.
This fixes the following two problems:
1. A customer provided some evidence which indicates that the idle tick was stopped; albeit, CPU-specific vmstat counters still remained populated.
Thus one can only assume quiet_vmstat() was not invoked on return to the idle loop. If I understand correctly, I suspect this divergence might erroneously prevent a reclaim attempt by kswapd. If the number of zone specific free pages are below their per-cpu drift value then zone_page_state_snapshot() is used to compute a more accurate view of the aforementioned statistic. Thus any task blocked on the NUMA node specific pfmemalloc_wait queue will be unable to make significant progress via direct reclaim unless it is killed after being woken up by kswapd
(see throttle_direct_reclaim())
2. With a SCHED_FIFO task that busy loops on a given CPU, and kworker for that CPU at SCHED_OTHER priority, queuing work to sync per-vmstats will either cause that work to never execute, or stalld (i.e. stall daemon) boosts kworker priority which causes a latency violation
mm/vmstat.c +18 -0
@@ -2043,6 +2043,23 @@ static void vmstat_shepherd(struct work_ static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd); +#ifdef CONFIG_HAVE_CMPXCHG_LOCAL +/* Flush counters remotely if CPU uses cmpxchg to update its per-CPU counters */ +static void vmstat_shepherd (struct work_struct *w) +{ +int cpu; + +cpus_read_lock (); +for_each_online_cpu (cpu) { +cpu_vm_stats_fold (cpu); +cond_resched (); + } +cpus_read_unlock (); + +schedule_delayed_work (&shepherd, +round_jiffies_relative (sysctl_stat_interval)); +} +#else static void vmstat_shepherd (struct work_struct *w) {int cpu;
@@ -2062,6 +2079,7 @@ static void vmstat_shepherd(struct work_schedule_delayed_work (&shepherd, round_jiffies_relative(sysctl_stat_interval)); } +#endif static void __init start_shepherd_timer (void ) {
PATCH v5 7/12 mm/vmstat: switch counter modification to cmpxchg
In reply to Marcelo TosattiIn preparation for switching vmstat shepherd to flush per-CPU counters remotely, switch the __{mod,inc,dec} functions that modify the counters to use cmpxchg.
To facilitate reviewing, functions are ordered in the text file, as:
__{mod,inc,dec}_{zone,node}_page_state
#ifdef CONFIG_HAVE_CMPXCHG_LOCAL
{mod,inc,dec}_{zone,node}_page_state
#else
{mod,inc,dec}_{zone,node}_page_state
#endif
This patch defines the __ versions for the CONFIG_HAVE_CMPXCHG_LOCAL case to be their non-"__" counterparts:
#ifdef CONFIG_HAVE_CMPXCHG_LOCAL {mod,inc,dec}_{zone,node}_page_state __{mod,inc,dec}_{zone,node}_page_state = {mod,inc,dec}_{zone,node}_page_state #else
{mod,inc,dec}_{zone,node}_page_state __{mod,inc,dec}_{zone,node}_page_state
#endif
To test the performance difference, a page allocator microbenchmark:
https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/bench/page_bench01.c
with loops=1000000 was used, on Intel Core i7-11850H @ 2.50GHz.
For the single_page_alloc_free test, which does
/** Loop to measure **/
for (i = 0; i < rec->loops; i++) {
my_page = alloc_page(gfp_mask);
if (unlikely(my_page == NULL))
return 0;
__free_page(my_page);
}Unit is cycles.
Vanilla Patched Diff
115.25 117 1.4%
mm/vmstat.c +182 -143
@@ -334,6 +334,188 @@ void set_pgdat_percpu_threshold(pg_data_ } } +#ifdef CONFIG_HAVE_CMPXCHG_LOCAL +/* +* If we have cmpxchg_local support then we do not need to incur the overhead +* that comes with local_irq_save/restore if we use this_cpu_cmpxchg. +* +* mod_state() modifies the zone counter state through atomic per cpu +* operations. +* +* Overstep mode specifies how overstep should handled: +* 0 No overstepping +* 1 Overstepping half of threshold +* -1 Overstepping minus half of threshold +*/ +static inline void mod_zone_state (struct zone *zone,enum zone_stat_item item, +long delta,int overstep_mode) +{ +struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats; +s8 __percpu *p = pcp->vm_stat_diff + item; +long o, n, t, z; + +do { + z =0 ;/* overflow to zone counters */ + +/* +* The fetching of the stat_threshold is racy. We may apply +* a counter threshold to the wrong the cpu if we get +* rescheduled while executing here. However, the next +* counter update will apply the threshold again and +* therefore bring the counter under the threshold again. +* +* Most of the time the thresholds are the same anyways +* for all cpus in a zone. +*/ + t =this_cpu_read (pcp->stat_threshold); + + o =this_cpu_read (*p); + n = delta + o; + +if (abs (n) > t) { +int os = overstep_mode * (t >>1 ); + +/* Overflow must be added to zone counters */ + z = n + os; + n = -os; + } + }while (this_cpu_cmpxchg (*p, o, n) != o); + +if (z) +zone_page_state_add (z, zone, item); +} + +void mod_zone_page_state (struct zone *zone,enum zone_stat_item item, +long delta) +{ +mod_zone_state (zone, item, delta,0 ); +} +EXPORT_SYMBOL (mod_zone_page_state); + +void __mod_zone_page_state (struct zone *zone,enum zone_stat_item item, +long delta) +{ +mod_zone_state (zone, item, delta,0 ); +} +EXPORT_SYMBOL (__mod_zone_page_state); + +void inc_zone_page_state (struct page *page,enum zone_stat_item item) +{ +mod_zone_state (page_zone (page), item,1 ,1 ); +} +EXPORT_SYMBOL (inc_zone_page_state); + +void __inc_zone_page_state (struct page *page,enum zone_stat_item item) +{ +mod_zone_state (page_zone (page), item,1 ,1 ); +} +EXPORT_SYMBOL (__inc_zone_page_state); + +void dec_zone_page_state (struct page *page,enum zone_stat_item item) +{ +mod_zone_state (page_zone (page), item,-1 ,-1 ); +} +EXPORT_SYMBOL (dec_zone_page_state); + +void __dec_zone_page_state (struct page *page,enum zone_stat_item item) +{ +mod_zone_state (page_zone (page), item,-1 ,-1 ); +} +EXPORT_SYMBOL (__dec_zone_page_state); + +static inline void mod_node_state (struct pglist_data *pgdat, +enum node_stat_item item, +int delta,int overstep_mode) +{ +struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats; +s8 __percpu *p = pcp->vm_node_stat_diff + item; +long o, n, t, z; + +if (vmstat_item_in_bytes (item)) { +/* +* Only cgroups use subpage accounting right now; at +* the global level, these items still change in +* multiples of whole pages. Store them as pages +* internally to keep the per-cpu counters compact. +*/ +VM_WARN_ON_ONCE (delta & (PAGE_SIZE -1 )); + delta >>=PAGE_SHIFT ; + } + +do { + z =0 ;/* overflow to node counters */ + +/* +* The fetching of the stat_threshold is racy. We may apply +* a counter threshold to the wrong the cpu if we get +* rescheduled while executing here. However, the next +* counter update will apply the threshold again and +* therefore bring the counter under the threshold again. +* +* Most of the time the thresholds are the same anyways +* for all cpus in a node. +*/ + t =this_cpu_read (pcp->stat_threshold); + + o =this_cpu_read (*p); + n = delta + o; + +if (abs (n) > t) { +int os = overstep_mode * (t >>1 ); + +/* Overflow must be added to node counters */ + z = n + os; + n = -os; + } + }while (this_cpu_cmpxchg (*p, o, n) != o); + +if (z) +node_page_state_add (z, pgdat, item); +} + +void mod_node_page_state (struct pglist_data *pgdat,enum node_stat_item item, +long delta) +{ +mod_node_state (pgdat, item, delta,0 ); +} +EXPORT_SYMBOL (mod_node_page_state); + +void __mod_node_page_state (struct pglist_data *pgdat,enum node_stat_item item, +long delta) +{ +mod_node_state (pgdat, item, delta,0 ); +} +EXPORT_SYMBOL (__mod_node_page_state); + +void inc_node_state (struct pglist_data *pgdat,enum node_stat_item item) +{ +mod_node_state (pgdat, item,1 ,1 ); +} + +void inc_node_page_state (struct page *page,enum node_stat_item item) +{ +mod_node_state (page_pgdat (page), item,1 ,1 ); +} +EXPORT_SYMBOL (inc_node_page_state); + +void __inc_node_page_state (struct page *page,enum node_stat_item item) +{ +mod_node_state (page_pgdat (page), item,1 ,1 ); +} +EXPORT_SYMBOL (__inc_node_page_state); + +void dec_node_page_state (struct page *page,enum node_stat_item item) +{ +mod_node_state (page_pgdat (page), item,-1 ,-1 ); +} +EXPORT_SYMBOL (dec_node_page_state); + +void __dec_node_page_state (struct page *page,enum node_stat_item item) +{ +mod_node_state (page_pgdat (page), item,-1 ,-1 ); +} +EXPORT_SYMBOL (__dec_node_page_state); +#else /* * For use when we know that interrupts are disabled, * or when we know that preemption is disabled and that
@@ -541,149 +723,6 @@ void __dec_node_page_state(struct page * }EXPORT_SYMBOL (__dec_node_page_state); -#ifdef CONFIG_HAVE_CMPXCHG_LOCAL -/* -* If we have cmpxchg_local support then we do not need to incur the overhead -* that comes with local_irq_save/restore if we use this_cpu_cmpxchg. -* -* mod_state() modifies the zone counter state through atomic per cpu -* operations. -* -* Overstep mode specifies how overstep should handled: -* 0 No overstepping -* 1 Overstepping half of threshold -* -1 Overstepping minus half of threshold -*/ -static inline void mod_zone_state (struct zone *zone, -enum zone_stat_item item,long delta,int overstep_mode) -{ -struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats; -s8 __percpu *p = pcp->vm_stat_diff + item; -long o, n, t, z; - -do { - z =0 ;/* overflow to zone counters */ - -/* -* The fetching of the stat_threshold is racy. We may apply -* a counter threshold to the wrong the cpu if we get -* rescheduled while executing here. However, the next -* counter update will apply the threshold again and -* therefore bring the counter under the threshold again. -* -* Most of the time the thresholds are the same anyways -* for all cpus in a zone. -*/ - t =this_cpu_read (pcp->stat_threshold); - - o =this_cpu_read (*p); - n = delta + o; - -if (abs (n) > t) { -int os = overstep_mode * (t >>1 ) ; - -/* Overflow must be added to zone counters */ - z = n + os; - n = -os; - } - }while (this_cpu_cmpxchg (*p, o, n) != o); - -if (z) -zone_page_state_add (z, zone, item); -} - -void mod_zone_page_state (struct zone *zone,enum zone_stat_item item, -long delta) -{ -mod_zone_state (zone, item, delta,0 ); -} -EXPORT_SYMBOL (mod_zone_page_state); - -void inc_zone_page_state (struct page *page,enum zone_stat_item item) -{ -mod_zone_state (page_zone (page), item,1 ,1 ); -} -EXPORT_SYMBOL (inc_zone_page_state); - -void dec_zone_page_state (struct page *page,enum zone_stat_item item) -{ -mod_zone_state (page_zone (page), item,-1 ,-1 ); -} -EXPORT_SYMBOL (dec_zone_page_state); - -static inline void mod_node_state (struct pglist_data *pgdat, -enum node_stat_item item,int delta,int overstep_mode) -{ -struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats; -s8 __percpu *p = pcp->vm_node_stat_diff + item; -long o, n, t, z; - -if (vmstat_item_in_bytes (item)) { -/* -* Only cgroups use subpage accounting right now; at -* the global level, these items still change in -* multiples of whole pages. Store them as pages -* internally to keep the per-cpu counters compact. -*/ -VM_WARN_ON_ONCE (delta & (PAGE_SIZE -1 )); - delta >>=PAGE_SHIFT ; - } - -do { - z =0 ;/* overflow to node counters */ - -/* -* The fetching of the stat_threshold is racy. We may apply -* a counter threshold to the wrong the cpu if we get -* rescheduled while executing here. However, the next -* counter update will apply the threshold again and -* therefore bring the counter under the threshold again. -* -* Most of the time the thresholds are the same anyways -* for all cpus in a node. -*/ - t =this_cpu_read (pcp->stat_threshold); - - o =this_cpu_read (*p); - n = delta + o; - -if (abs (n) > t) { -int os = overstep_mode * (t >>1 ) ; - -/* Overflow must be added to node counters */ - z = n + os; - n = -os; - } - }while (this_cpu_cmpxchg (*p, o, n) != o); - -if (z) -node_page_state_add (z, pgdat, item); -} - -void mod_node_page_state (struct pglist_data *pgdat,enum node_stat_item item, -long delta) -{ -mod_node_state (pgdat, item, delta,0 ); -} -EXPORT_SYMBOL (mod_node_page_state); - -void inc_node_state (struct pglist_data *pgdat,enum node_stat_item item) -{ -mod_node_state (pgdat, item,1 ,1 ); -} - -void inc_node_page_state (struct page *page,enum node_stat_item item) -{ -mod_node_state (page_pgdat (page), item,1 ,1 ); -} -EXPORT_SYMBOL (inc_node_page_state); - -void dec_node_page_state (struct page *page,enum node_stat_item item) -{ -mod_node_state (page_pgdat (page), item,-1 ,-1 ); -} -EXPORT_SYMBOL (dec_node_page_state); -#else /* * Use interrupt disable to serialize counter updates */
mm/page_alloc.c +0 -3
@@ -8628,9 +8628,6 @@ static int page_alloc_cpu_dead(unsigned/* * Zero the differential counters of the dead processor * so that the vm statistics are consistent. -* -* This is only okay since the processor is dead and cannot -* race with what we are doing. */ cpu_vm_stats_fold (cpu);
PATCH v5 9/12 mm/vmstat: use xchg in cpu_vm_stats_fold
In reply to Marcelo TosattiIn preparation to switch vmstat shepherd to flush
per-CPU counters remotely, use xchg instead of a
pair of read/write instructions.
mm/vmstat.c +4 -7
@@ -924,7 +924,7 @@ static int refresh_cpu_vm_stats(bool do_ }/* -* Fold the data for an offline cpu into the global array. +* Fold the data for a cpu into the global array. * There cannot be any access by the offline cpu and therefore * synchronization is simplified. */
@@ -945,8 +945,7 @@ void cpu_vm_stats_fold(int cpu)if (pzstats->vm_stat_diff[i]) {int v; - v = pzstats->vm_stat_diff[i]; - pzstats->vm_stat_diff[i] =0 ; + v =xchg (&pzstats->vm_stat_diff[i],0 );atomic_long_add (v, &zone->vm_stat[i]); global_zone_diff[i] += v; }
@@ -956,8 +955,7 @@ void cpu_vm_stats_fold(int cpu)if (pzstats->vm_numa_event[i]) {unsigned long v; - v = pzstats->vm_numa_event[i]; - pzstats->vm_numa_event[i] =0 ; + v =xchg (&pzstats->vm_numa_event[i],0 );zone_numa_event_add (v, zone, i); } }
@@ -973,8 +971,7 @@ void cpu_vm_stats_fold(int cpu)if (p->vm_node_stat_diff[i]) {int v; - v = p->vm_node_stat_diff[i]; - p->vm_node_stat_diff[i] =0 ; + v =xchg (&p->vm_node_stat_diff[i],0 );atomic_long_add (v, &pgdat->vm_stat[i]); global_node_diff[i] += v; }
PATCH v5 1/12 this_cpu_cmpxchg: ARM64: switch this_cpu_cmpxchg to locked, add _local function
In reply to Marcelo TosattiGoal is to have vmstat_shepherd to transfer from per-CPU counters to global counters remotely. For this, an atomic this_cpu_cmpxchg is necessary.
Following the kernel convention for cmpxchg/cmpxchg_local, change ARM's this_cpu_cmpxchg_ helpers to be atomic, and add this_cpu_cmpxchg_local_ helpers which are not atomic.
arch/arm64/include/asm/percpu.h +13 -3
@@ -232,13 +232,23 @@ PERCPU_RET_OP(add, add, ldadd) _pcp_protect_return(xchg_relaxed, pcp, val)#define this_cpu_cmpxchg_1 (pcp, o, n) \ - _pcp_protect_return(cmpxchg_relaxed, pcp, o, n) + _pcp_protect_return(cmpxchg, pcp, o, n)#define this_cpu_cmpxchg_2 (pcp, o, n) \ - _pcp_protect_return(cmpxchg_relaxed, pcp, o, n) + _pcp_protect_return(cmpxchg, pcp, o, n)#define this_cpu_cmpxchg_4 (pcp, o, n) \ - _pcp_protect_return(cmpxchg_relaxed, pcp, o, n) + _pcp_protect_return(cmpxchg, pcp, o, n)#define this_cpu_cmpxchg_8 (pcp, o, n) \ + _pcp_protect_return(cmpxchg, pcp, o, n) + +#define this_cpu_cmpxchg_local_1 (pcp, o, n) \ _pcp_protect_return(cmpxchg_relaxed, pcp, o, n) +#define this_cpu_cmpxchg_local_2 (pcp, o, n) \ + _pcp_protect_return(cmpxchg_relaxed, pcp, o, n) +#define this_cpu_cmpxchg_local_4 (pcp, o, n) \ + _pcp_protect_return(cmpxchg_relaxed, pcp, o, n) +#define this_cpu_cmpxchg_local_8 (pcp, o, n) \ + _pcp_protect_return(cmpxchg_relaxed, pcp, o, n) +#ifdef __KVM_NVHE_HYPERVISOR__extern unsigned long __hyp_per_cpu_offset (unsigned int cpu);
PATCH v5 4/12 this_cpu_cmpxchg: x86: switch this_cpu_cmpxchg to locked, add _local function
In reply to Marcelo TosattiGoal is to have vmstat_shepherd to transfer from per-CPU counters to global counters remotely. For this, an atomic this_cpu_cmpxchg is necessary.
Following the kernel convention for cmpxchg/cmpxchg_local, change x86's this_cpu_cmpxchg_ helpers to be atomic. and add this_cpu_cmpxchg_local_ helpers which are not atomic.
arch/x86/include/asm/percpu.h +22 -17
@@ -197,11 +197,11 @@ do { \* cmpxchg has no such implied lock semantics as a result it is much * more efficient for cpu local operations. */ -#define percpu_cmpxchg_op (size, qual, _var, _oval, _nval) \ +#define percpu_cmpxchg_op (size, qual, _var, _oval, _nval, lockp) \ ({ \ __pcpu_type_##size pco_old__ = __pcpu_cast_##size(_oval); \ __pcpu_type_##size pco_new__ = __pcpu_cast_##size(_nval); \ - asm qual (__pcpu_op2_##size("cmpxchg", "%[nval]", \ + asm qual (__pcpu_op2_##size(lockp "cmpxchg", "%[nval]", \ __percpu_arg([var])) \ : [oval] "+a" (pco_old__), \ [var] "+m" (_var) \
@@ -279,16 +279,20 @@ do { \#define raw_cpu_add_return_1 (pcp, val) percpu_add_return_op(1, , pcp, val)#define raw_cpu_add_return_2 (pcp, val) percpu_add_return_op(2, , pcp, val)#define raw_cpu_add_return_4 (pcp, val) percpu_add_return_op(4, , pcp, val) -#define raw_cpu_cmpxchg_1 (pcp, oval, nval) percpu_cmpxchg_op(1, , pcp, oval, nval) -#define raw_cpu_cmpxchg_2 (pcp, oval, nval) percpu_cmpxchg_op(2, , pcp, oval, nval) -#define raw_cpu_cmpxchg_4 (pcp, oval, nval) percpu_cmpxchg_op(4, , pcp, oval, nval) +#define raw_cpu_cmpxchg_1 (pcp, oval, nval) percpu_cmpxchg_op(1, , pcp, oval, nval, "") +#define raw_cpu_cmpxchg_2 (pcp, oval, nval) percpu_cmpxchg_op(2, , pcp, oval, nval, "") +#define raw_cpu_cmpxchg_4 (pcp, oval, nval) percpu_cmpxchg_op(4, , pcp, oval, nval, "")#define this_cpu_add_return_1 (pcp, val) percpu_add_return_op(1, volatile, pcp, val)#define this_cpu_add_return_2 (pcp, val) percpu_add_return_op(2, volatile, pcp, val)#define this_cpu_add_return_4 (pcp, val) percpu_add_return_op(4, volatile, pcp, val) -#define this_cpu_cmpxchg_1 (pcp, oval, nval) percpu_cmpxchg_op(1, volatile, pcp, oval, nval) -#define this_cpu_cmpxchg_2 (pcp, oval, nval) percpu_cmpxchg_op(2, volatile, pcp, oval, nval) -#define this_cpu_cmpxchg_4 (pcp, oval, nval) percpu_cmpxchg_op(4, volatile, pcp, oval, nval) +#define this_cpu_cmpxchg_local_1 (pcp, oval, nval) percpu_cmpxchg_op(1, volatile, pcp, oval, nval, "") +#define this_cpu_cmpxchg_local_2 (pcp, oval, nval) percpu_cmpxchg_op(2, volatile, pcp, oval, nval, "") +#define this_cpu_cmpxchg_local_4 (pcp, oval, nval) percpu_cmpxchg_op(4, volatile, pcp, oval, nval, "") + +#define this_cpu_cmpxchg_1 (pcp, oval, nval) percpu_cmpxchg_op(1, volatile, pcp, oval, nval, LOCK_PREFIX) +#define this_cpu_cmpxchg_2 (pcp, oval, nval) percpu_cmpxchg_op(2, volatile, pcp, oval, nval, LOCK_PREFIX) +#define this_cpu_cmpxchg_4 (pcp, oval, nval) percpu_cmpxchg_op(4, volatile, pcp, oval, nval, LOCK_PREFIX)#ifdef CONFIG_X86_CMPXCHG64 #define percpu_cmpxchg8b_double (pcp1, pcp2, o1, o2, n1, n2) \
@@ -319,16 +323,17 @@ do { \#define raw_cpu_or_8 (pcp, val) percpu_to_op(8, , "or", (pcp), val)#define raw_cpu_add_return_8 (pcp, val) percpu_add_return_op(8, , pcp, val)#define raw_cpu_xchg_8 (pcp, nval) raw_percpu_xchg_op(pcp, nval) -#define raw_cpu_cmpxchg_8 (pcp, oval, nval) percpu_cmpxchg_op(8, , pcp, oval, nval) +#define raw_cpu_cmpxchg_8 (pcp, oval, nval) percpu_cmpxchg_op(8, , pcp, oval, nval, "") -#define this_cpu_read_8 (pcp) percpu_from_op(8, volatile, "mov", pcp) -#define this_cpu_write_8 (pcp, val) percpu_to_op(8, volatile, "mov", (pcp), val) -#define this_cpu_add_8 (pcp, val) percpu_add_op(8, volatile, (pcp), val) -#define this_cpu_and_8 (pcp, val) percpu_to_op(8, volatile, "and", (pcp), val) -#define this_cpu_or_8 (pcp, val) percpu_to_op(8, volatile, "or", (pcp), val) -#define this_cpu_add_return_8 (pcp, val) percpu_add_return_op(8, volatile, pcp, val) -#define this_cpu_xchg_8 (pcp, nval) percpu_xchg_op(8, volatile, pcp, nval) -#define this_cpu_cmpxchg_8 (pcp, oval, nval) percpu_cmpxchg_op(8, volatile, pcp, oval, nval) +#define this_cpu_read_8 (pcp) percpu_from_op(8, volatile, "mov", pcp) +#define this_cpu_write_8 (pcp, val) percpu_to_op(8, volatile, "mov", (pcp), val) +#define this_cpu_add_8 (pcp, val) percpu_add_op(8, volatile, (pcp), val) +#define this_cpu_and_8 (pcp, val) percpu_to_op(8, volatile, "and", (pcp), val) +#define this_cpu_or_8 (pcp, val) percpu_to_op(8, volatile, "or", (pcp), val) +#define this_cpu_add_return_8 (pcp, val) percpu_add_return_op(8, volatile, pcp, val) +#define this_cpu_xchg_8 (pcp, nval) percpu_xchg_op(8, volatile, pcp, nval) +#define this_cpu_cmpxchg_local_8 (pcp, oval, nval) percpu_cmpxchg_op(8, volatile, pcp, oval, nval, "") +#define this_cpu_cmpxchg_8 (pcp, oval, nval) percpu_cmpxchg_op(8, volatile, pcp, oval, nval, LOCK_PREFIX)/* * Pretty complex macro to generate cmpxchg16 instruction. The instruction
PATCH v5 11/12 mm/vmstat: refresh stats remotely instead of via work item
In reply to Marcelo TosattiRefresh per-CPU stats remotely, instead of queueing work items, for the stat_refresh procfs method.
This fixes sosreport hang (which uses vmstat_refresh) with spinning SCHED_FIFO process.
mm/vmstat.c +16 -2
@@ -1901,11 +1901,20 @@ static DEFINE_PER_CPU(struct delayed_worint sysctl_stat_interval__read_mostly =HZ ;#ifdef CONFIG_PROC_FS +#ifdef CONFIG_HAVE_CMPXCHG_LOCAL +static int refresh_all_vm_stats (void ); +#else static void refresh_vm_stats (struct work_struct *work) {refresh_cpu_vm_stats (true); } +static int refresh_all_vm_stats (void ) +{ +return schedule_on_each_cpu (refresh_vm_stats); +} +#endif +int vmstat_refresh (struct ctl_table *table,int write,void *buffer,size_t *lenp,loff_t *ppos) {
@@ -1925,7 +1934,7 @@ int vmstat_refresh(struct ctl_table *tab* transiently negative values, report an error here if any of * the stats is negative, so we know to go looking for imbalance. */ - err =schedule_on_each_cpu (refresh_vm_stats); + err =refresh_all_vm_stats ();if (err)return err;for (i =0 ; i <NR_VM_ZONE_STAT_ITEMS ; i++) {
@@ -2045,7 +2054,7 @@ static DECLARE_DEFERRABLE_WORK(shepherd,#ifdef CONFIG_HAVE_CMPXCHG_LOCAL /* Flush counters remotely if CPU uses cmpxchg to update its per-CPU counters */ -static void vmstat_shepherd (struct work_struct *w) +static int refresh_all_vm_stats (void ) {int cpu;
@@ -2055,7 +2064,12 @@ static void vmstat_shepherd(struct work_cond_resched (); }cpus_read_unlock (); +return 0 ; +} +static void vmstat_shepherd (struct work_struct *w) +{ +refresh_all_vm_stats ();schedule_delayed_work (&shepherd,round_jiffies_relative (sysctl_stat_interval)); }
Michal Hocko
reviewerQuoted 6 lines · Marcelo Tosatti
This patch series addresses the following two problems:
1. A customer provided some evidence which indicates that the idle tick was stopped; albeit, CPU-specific vmstat counters still remained populated.
Quoted 12 lines
Thus one can only assume
quiet_vmstat()was not invoked on return to the idle loop. If I understand correctly, I suspect this divergence might erroneously prevent a reclaim attempt by kswapd. If the number of zone specific free pages are below their per-cpu drift value thenzone_page_state_snapshot()is used to compute a more accurate view of the aforementioned statistic. Thus any task blocked on the NUMA node specificpfmemalloc_waitqueue will be unable to make significant progress via direct reclaim unless it is killed after being woken up by kswapd
(seethrottle_direct_reclaim())
I have hard time to follow the actual problem described above. Are you suggesting that a lack of pcp vmstat counters update has led to reclaim issues? What is the said "evidence"? Could you share more of the story please?
Quoted 6 lines
2. With a
SCHED_FIFOtask that busy loops on a given CPU, and kworker for that CPU atSCHED_OTHERpriority, queuing work to sync per-vmstats will either cause that work to never execute, or stalld (i.e. stall daemon) boosts kworker priority which causes a latency violation
Why is that a problem? Out-of-sync stats shouldn't cause major problems. Or can they?
Thanks!
Signature · 2 lines
Michal Hocko SUSE Labs
Quoted 20 lines · Michal Hocko
Quoted 6 lines · Marcelo Tosatti
This patch series addresses the following two problems:
1. A customer provided some evidence which indicates that the idle tick was stopped; albeit, CPU-specific vmstat counters still remained populated.
Quoted 12 lines
Thus one can only assume
quiet_vmstat()was not invoked on return to the idle loop. If I understand correctly, I suspect this divergence might erroneously prevent a reclaim attempt by kswapd. If the number of zone specific free pages are below their per-cpu drift value thenzone_page_state_snapshot()is used to compute a more accurate view of the aforementioned statistic. Thus any task blocked on the NUMA node specificpfmemalloc_waitqueue will be unable to make significant progress via direct reclaim unless it is killed after being woken up by kswapd
(seethrottle_direct_reclaim())
Quoted 4 lines
I have hard time to follow the actual problem described above. Are you suggesting that a lack of pcp vmstat counters update has led to reclaim issues? What is the said "evidence"? Could you share more of the story please?
- The process was trapped in throttle_direct_reclaim(). The function wait_event_killable() was called to wait condition allow_direct_reclaim(pgdat) for current node to be true. The allow_direct_reclaim(pgdat) examined the number of free pages on the node by zone_page_state() which just returns value in zone->vm_stat[NR_FREE_PAGES].
- On node #1, zone->vm_stat[NR_FREE_PAGES] was 0. However, the freelist on this node was not empty.
- This inconsistent of vmstat value was caused by percpu vmstat on nohz_full cpus. Every increment/decrement of vmstat is performed on percpu vmstat counter at first, then pooled diffs are cumulated to the zone's vmstat counter in timely manner. However, on nohz_full cpus (in case of this customer's system, 48 of 52 cpus) these pooled diffs were not cumulated once the cpu had no event on it so that the cpu started sleeping infinitely.
I checked percpu vmstat and found there were total 69 counts not cumulated to the zone's vmstat counter yet.
- In this situation, kswapd did not help the trapped process. In pgdat_balanced(), zone_wakermark_ok_safe() examined the number of free pages on the node by zone_page_state_snapshot() which checks pending counts on percpu vmstat.
Therefore kswapd could know there were 69 free pages correctly. Since zone->_watermark = {8, 20, 32}, kswapd did not work because 69 was greater than 32 as high watermark.
- As the result:
- The process waited allow_direct_reclaim(pgdat) to be true.
- allow_direct_reclaim() saw 0 via zone_page_state(). It woke kswapd since 0 was lower than min watermark.
- The kswapd did nothing.
- kswapd saw 69 via zone_page_state_snapshot(). It woke waiters without performing memory reclaim since 69 is greater than high watermark.
- The process woked by kswapd soon restart waiting for kswapd.
- Still allow_direct_reclaim() saw 0 via zone_page_state(). It woke kswapd since 0 was lower than min watermark.
Quoted 10 lines
Quoted 7 lines
2. With a
SCHED_FIFOtask that busy loops on a given CPU, and kworker for that CPU atSCHED_OTHERpriority, queuing work to sync per-vmstats will either cause that work to never execute, or stalld (i.e. stall daemon) boosts kworker priority which causes a latency violationWhy is that a problem? Out-of-sync stats shouldn't cause major problems. Or can they?
Consider SCHED_FIFO task that is polling the network queue (say testpmd).
do {
if (net_registers->state & DATA_AVAILABLE) {
process_data)();
}
} while (!stopped);Since this task runs at SCHED_FIFO priority, kworker won't be scheduled to run (therefore per-CPU vmstats won't be flushed to global vmstats).
Or, if testpmd runs at SCHED_OTHER, then the work item to flush per-CPU vmstats causes
testpmd -> kworker kworker: flush per-CPU vmstats kworker -> testpmd
And this might cause undesired latencies to the packets being processed by the testpmd task.
Quoted 19 lines · Marcelo Tosatti
Quoted 2 lines
Why is that a problem? Out-of-sync stats shouldn't cause major problems. Or can they?
Consider
SCHED_FIFOtask that is polling the network queue (say testpmd).do { if (net_registers->state & DATA_AVAILABLE) { process_data)(); } } while (!stopped);Since this task runs at
SCHED_FIFOpriority, kworker won't be scheduled to run (therefore per-CPU vmstats won't be flushed to global vmstats).Or, if testpmd runs at
SCHED_OTHER, then the work item to flush per-CPU vmstats causes
Quoted 6 lines
testpmd -> kworker kworker: flush per-CPU vmstats kworker -> testpmdAnd this might cause undesired latencies to the packets being processed by the testpmd task.
This problem is unrelated to the kswapd problem, but both are addressed by the patchset.
Michal Hocko
reviewerQuoted 47 lines · Marcelo Tosatti
Quoted 20 lines · Michal Hocko
Quoted 6 lines · Marcelo Tosatti
This patch series addresses the following two problems:
1. A customer provided some evidence which indicates that the idle tick was stopped; albeit, CPU-specific vmstat counters still remained populated.
Quoted 12 lines
Thus one can only assume
quiet_vmstat()was not invoked on return to the idle loop. If I understand correctly, I suspect this divergence might erroneously prevent a reclaim attempt by kswapd. If the number of zone specific free pages are below their per-cpu drift value thenzone_page_state_snapshot()is used to compute a more accurate view of the aforementioned statistic. Thus any task blocked on the NUMA node specificpfmemalloc_waitqueue will be unable to make significant progress via direct reclaim unless it is killed after being woken up by kswapd
(seethrottle_direct_reclaim())Quoted 4 lines
I have hard time to follow the actual problem described above. Are you suggesting that a lack of pcp vmstat counters update has led to reclaim issues? What is the said "evidence"? Could you share more of the story please?
- The process was trapped in
throttle_direct_reclaim(). The functionwait_event_killable()was called to wait conditionallow_direct_reclaim(pgdat)for current node to be true. Theallow_direct_reclaim(pgdat)examined the number of free pages on the node byzone_page_state()which just returns value in zone->vm_stat[NR_FREE_PAGES].- On node #1, zone->vm_stat[NR_FREE_PAGES] was 0. However, the freelist on this node was not empty.
- This inconsistent of vmstat value was caused by percpu vmstat on
nohz_fullcpus. Every increment/decrement of vmstat is performed on percpu vmstat counter at first, then pooled diffs are cumulated to the zone's vmstat counter in timely manner. However, onnohz_fullcpus (in case of this customer's system, 48 of 52 cpus) these pooled diffs were not cumulated once the cpu had no event on it so that the cpu started sleeping infinitely.
I checked percpu vmstat and found there were total 69 counts not cumulated to the zone's vmstat counter yet.
Quoted 7 lines
- In this situation, kswapd did not help the trapped process. In
pgdat_balanced(),zone_wakermark_ok_safe()examined the number of free pages on the node byzone_page_state_snapshot()which checks pending counts on percpu vmstat.
Therefore kswapd could know there were 69 free pages correctly. Sincezone->_watermark= {8, 20, 32}, kswapd did not work because 69 was greater than 32 as high watermark.
If the imprecision of allow_direct_reclaim is the underlying problem why haven't you used zone_page_state_snapshot instead?
Anyway, this is kind of information that is really helpful to have in the patch description.
[...]
Quoted 19 lines
Quoted 9 lines
Quoted 6 lines
2. With a
SCHED_FIFOtask that busy loops on a given CPU, and kworker for that CPU atSCHED_OTHERpriority, queuing work to sync per-vmstats will either cause that work to never execute, or stalld (i.e. stall daemon) boosts kworker priority which causes a latency violationWhy is that a problem? Out-of-sync stats shouldn't cause major problems. Or can they?
Consider
SCHED_FIFOtask that is polling the network queue (say testpmd).do { if (net_registers->state & DATA_AVAILABLE) { process_data)(); } } while (!stopped);
Quoted 3 lines
Since this task runs at
SCHED_FIFOpriority, kworker won't be scheduled to run (therefore per-CPU vmstats won't be flushed to global vmstats).
Yes, that is certainly possible. But my main point is that vmstat imprecision shouldn't cause functional problems. That is why we have _snapshot readers to get an exact value where it matters for consistency.
Quoted 9 lines
Or, if testpmd runs at
SCHED_OTHER, then the work item to flush per-CPU vmstats causestestpmd -> kworker kworker: flush per-CPU vmstats kworker -> testpmdAnd this might cause undesired latencies to the packets being processed by the testpmd task.
Right but can you have any latencies expectation in a situation like that?
Signature · 2 lines
Michal Hocko SUSE Labs
Quoted 56 lines · Michal Hocko
Quoted 47 lines · Marcelo Tosatti
Quoted 20 lines · Michal Hocko
Quoted 6 lines · Marcelo Tosatti
This patch series addresses the following two problems:
1. A customer provided some evidence which indicates that the idle tick was stopped; albeit, CPU-specific vmstat counters still remained populated.
Quoted 12 lines
Thus one can only assume
quiet_vmstat()was not invoked on return to the idle loop. If I understand correctly, I suspect this divergence might erroneously prevent a reclaim attempt by kswapd. If the number of zone specific free pages are below their per-cpu drift value thenzone_page_state_snapshot()is used to compute a more accurate view of the aforementioned statistic. Thus any task blocked on the NUMA node specificpfmemalloc_waitqueue will be unable to make significant progress via direct reclaim unless it is killed after being woken up by kswapd
(seethrottle_direct_reclaim())Quoted 4 lines
I have hard time to follow the actual problem described above. Are you suggesting that a lack of pcp vmstat counters update has led to reclaim issues? What is the said "evidence"? Could you share more of the story please?
- The process was trapped in
throttle_direct_reclaim(). The functionwait_event_killable()was called to wait conditionallow_direct_reclaim(pgdat)for current node to be true. Theallow_direct_reclaim(pgdat)examined the number of free pages on the node byzone_page_state()which just returns value in zone->vm_stat[NR_FREE_PAGES].- On node #1, zone->vm_stat[NR_FREE_PAGES] was 0. However, the freelist on this node was not empty.
- This inconsistent of vmstat value was caused by percpu vmstat on
nohz_fullcpus. Every increment/decrement of vmstat is performed on percpu vmstat counter at first, then pooled diffs are cumulated to the zone's vmstat counter in timely manner. However, onnohz_fullcpus (in case of this customer's system, 48 of 52 cpus) these pooled diffs were not cumulated once the cpu had no event on it so that the cpu started sleeping infinitely.
I checked percpu vmstat and found there were total 69 counts not cumulated to the zone's vmstat counter yet.Quoted 7 lines
- In this situation, kswapd did not help the trapped process. In
pgdat_balanced(),zone_wakermark_ok_safe()examined the number of free pages on the node byzone_page_state_snapshot()which checks pending counts on percpu vmstat.
Therefore kswapd could know there were 69 free pages correctly. Sincezone->_watermark= {8, 20, 32}, kswapd did not work because 69 was greater than 32 as high watermark.
Quoted 2 lines
If the imprecision of
allow_direct_reclaimis the underlying problem why haven't you usedzone_page_state_snapshotinstead?
It might have dealt with problem #1 for this particular case. However, looking at the callers of zone_page_state:
5 2227 mm/compaction.c <<compaction_suitable>>
zone_page_state(zone, NR_FREE_PAGES));
6 124 mm/highmem.c <<__nr_free_highpages>>
pages += zone_page_state(zone, NR_FREE_PAGES);
7 283 mm/page-writeback.c <<node_dirtyable_memory>>
nr_pages += zone_page_state(zone, NR_FREE_PAGES);
8 318 mm/page-writeback.c <<highmem_dirtyable_memory>>
nr_pages = zone_page_state(z, NR_FREE_PAGES);
9 321 mm/page-writeback.c <<highmem_dirtyable_memory>>
nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
10 322 mm/page-writeback.c <<highmem_dirtyable_memory>>
nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
11 3091 mm/page_alloc.c <<__rmqueue>>
zone_page_state(zone, NR_FREE_CMA_PAGES) >
12 3092 mm/page_alloc.c <<__rmqueue>>
zone_page_state(zone, NR_FREE_PAGES) / 2) {The suggested patchset fixes the problem of where due to nohz_full, the delayed timer for vmstat_work can be armed but not executed (which means the per-cpu counters can be out of sync for as long as the cpu is in idle while in nohz_full mode).
You probably do not want to convert all callers of zone_page_state into zone_page_state_snapshot (as a justification for the proposed patchset).
Quoted 2 lines
Anyway, this is kind of information that is really helpful to have in the patch description.
Agree: resending a new version with updated commit.
Quoted 20 lines
[...]
Quoted 19 lines
Quoted 9 lines
Quoted 6 lines
2. With a
SCHED_FIFOtask that busy loops on a given CPU, and kworker for that CPU atSCHED_OTHERpriority, queuing work to sync per-vmstats will either cause that work to never execute, or stalld (i.e. stall daemon) boosts kworker priority which causes a latency violationWhy is that a problem? Out-of-sync stats shouldn't cause major problems. Or can they?
Consider
SCHED_FIFOtask that is polling the network queue (say testpmd).do { if (net_registers->state & DATA_AVAILABLE) { process_data)(); } } while (!stopped);
Quoted 8 lines
Quoted 3 lines
Since this task runs at
SCHED_FIFOpriority, kworker won't be scheduled to run (therefore per-CPU vmstats won't be flushed to global vmstats).Yes, that is certainly possible. But my main point is that vmstat imprecision shouldn't cause functional problems. That is why we have
_snapshotreaders to get an exact value where it matters for consistency.
Understood. Perhaps allow_direct_reclaim should use zone_page_state_snapshot, as otherwise it is only precise at sysctl_stat_interval intervals?
Quoted 8 lines
Quoted 7 lines
Or, if testpmd runs at
SCHED_OTHER, then the work item to flush per-CPU vmstats causestestpmd -> kworker kworker: flush per-CPU vmstats kworker -> testpmd
Quoted 5 lines
And this might cause undesired latencies to the packets being processed by the testpmd task.
Right but can you have any latencies expectation in a situation like that?
Not sure i understand what you mean. Example:
https://www.gabrieleara.it/assets/documents/papers/conferences/2021-ieee-nfv-sdn.pdf
In general, UDPDK exhibits a much lower
latency than the in-kernel UDP stack used through the POSIX API (e.g., a 69 % reduction from 95 µs down to 29 µs), thanks to its ability to bypass the kernel entirely, which in turn outperforms the in-kernel TCP stack as available through the POSIX API, as expected.
...
Alternatively, application processes can use UDPDK with the non-blocking API calls (using the O_NONBLOCK flag) and perform some other action while waiting for packets to be ready to be sent/received to/from the UDPDK Process, instead of performing continuous busy-loops on packet queues. However, in this case the cost of a single CPU fully busy due to the UDPDK Process itself is anyway unavoidab
Michal Hocko
reviewerQuoted 80 lines · Marcelo Tosatti
Quoted 56 lines · Michal Hocko
Quoted 47 lines · Marcelo Tosatti
Quoted 20 lines · Michal Hocko
Quoted 6 lines · Marcelo Tosatti
This patch series addresses the following two problems:
1. A customer provided some evidence which indicates that the idle tick was stopped; albeit, CPU-specific vmstat counters still remained populated.
Quoted 12 lines
Thus one can only assume
quiet_vmstat()was not invoked on return to the idle loop. If I understand correctly, I suspect this divergence might erroneously prevent a reclaim attempt by kswapd. If the number of zone specific free pages are below their per-cpu drift value thenzone_page_state_snapshot()is used to compute a more accurate view of the aforementioned statistic. Thus any task blocked on the NUMA node specificpfmemalloc_waitqueue will be unable to make significant progress via direct reclaim unless it is killed after being woken up by kswapd
(seethrottle_direct_reclaim())Quoted 4 lines
I have hard time to follow the actual problem described above. Are you suggesting that a lack of pcp vmstat counters update has led to reclaim issues? What is the said "evidence"? Could you share more of the story please?
- The process was trapped in
throttle_direct_reclaim(). The functionwait_event_killable()was called to wait conditionallow_direct_reclaim(pgdat)for current node to be true. Theallow_direct_reclaim(pgdat)examined the number of free pages on the node byzone_page_state()which just returns value in zone->vm_stat[NR_FREE_PAGES].- On node #1, zone->vm_stat[NR_FREE_PAGES] was 0. However, the freelist on this node was not empty.
- This inconsistent of vmstat value was caused by percpu vmstat on
nohz_fullcpus. Every increment/decrement of vmstat is performed on percpu vmstat counter at first, then pooled diffs are cumulated to the zone's vmstat counter in timely manner. However, onnohz_fullcpus (in case of this customer's system, 48 of 52 cpus) these pooled diffs were not cumulated once the cpu had no event on it so that the cpu started sleeping infinitely.
I checked percpu vmstat and found there were total 69 counts not cumulated to the zone's vmstat counter yet.Quoted 7 lines
- In this situation, kswapd did not help the trapped process. In
pgdat_balanced(),zone_wakermark_ok_safe()examined the number of free pages on the node byzone_page_state_snapshot()which checks pending counts on percpu vmstat.
Therefore kswapd could know there were 69 free pages correctly. Sincezone->_watermark= {8, 20, 32}, kswapd did not work because 69 was greater than 32 as high watermark.Quoted 2 lines
If the imprecision of
allow_direct_reclaimis the underlying problem why haven't you usedzone_page_state_snapshotinstead?It might have dealt with problem #1 for this particular case. However, looking at the callers of
zone_page_state:5 2227 mm/compaction.c <<compaction_suitable>> zone_page_state(zone, NR_FREE_PAGES)); 6 124 mm/highmem.c <<__nr_free_highpages>> pages += zone_page_state(zone, NR_FREE_PAGES); 7 283 mm/page-writeback.c <<node_dirtyable_memory>> nr_pages += zone_page_state(zone, NR_FREE_PAGES); 8 318 mm/page-writeback.c <<highmem_dirtyable_memory>> nr_pages = zone_page_state(z, NR_FREE_PAGES); 9 321 mm/page-writeback.c <<highmem_dirtyable_memory>> nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE); 10 322 mm/page-writeback.c <<highmem_dirtyable_memory>> nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE); 11 3091 mm/page_alloc.c <<__rmqueue>> zone_page_state(zone, NR_FREE_CMA_PAGES) > 12 3092 mm/page_alloc.c <<__rmqueue>> zone_page_state(zone, NR_FREE_PAGES) / 2) {
Quoted 8 lines
The suggested patchset fixes the problem of where due to
nohz_full, the delayed timer forvmstat_workcan be armed but not executed (which means the per-cpu counters can be out of sync for as long as the cpu is in idle while innohz_fullmode).You probably do not want to convert all callers of
zone_page_stateintozone_page_state_snapshot(as a justification for the proposed patchset).
Yes, I do not really think we want or even need to convert all of them. But it seems that your direct reclaim throttling example really requires that. The thing with the remote flushing is that it would suffer from a similar imprecations as the flushing could be deferred and under certain conditions really starved. So it is definitely worth fixing the issue you are seeing without such a complex scheme.
Quoted 4 lines
Quoted 2 lines
Anyway, this is kind of information that is really helpful to have in the patch description.
Agree: resending a new version with updated commit.
I would really recommend trying out the simple fix and see if it changes the behavior.
Quoted 24 lines
Quoted 24 lines
[...]
Quoted 22 lines
Quoted 9 lines
Quoted 6 lines
2. With a
SCHED_FIFOtask that busy loops on a given CPU, and kworker for that CPU atSCHED_OTHERpriority, queuing work to sync per-vmstats will either cause that work to never execute, or stalld (i.e. stall daemon) boosts kworker priority which causes a latency violationWhy is that a problem? Out-of-sync stats shouldn't cause major problems. Or can they?
Consider
SCHED_FIFOtask that is polling the network queue (say testpmd).do { if (net_registers->state & DATA_AVAILABLE) { process_data)(); } } while (!stopped);Since this task runs at
SCHED_FIFOpriority, kworker won't be scheduled to run (therefore per-CPU vmstats won't be flushed to global vmstats).
Quoted 8 lines
Quoted 4 lines
Yes, that is certainly possible. But my main point is that vmstat imprecision shouldn't cause functional problems. That is why we have
_snapshotreaders to get an exact value where it matters for consistency.Understood. Perhaps
allow_direct_reclaimshould usezone_page_state_snapshot, as otherwise it is only precise atsysctl_stat_intervalintervals?
or even much less than that. The flusher uses WQ infrastructure and even when a WQ_MEM_RECLAIM one is used this doesn't mean that all workers could be jammed.
Quoted 17 lines
Quoted 7 lines
Quoted 6 lines
Or, if testpmd runs at
SCHED_OTHER, then the work item to flush per-CPU vmstats causestestpmd -> kworker kworker: flush per-CPU vmstats kworker -> testpmdQuoted 5 lines
And this might cause undesired latencies to the packets being processed by the testpmd task.
Right but can you have any latencies expectation in a situation like that?
Not sure i understand what you mean. Example:
https://www.gabrieleara.it/assets/documents/papers/conferences/2021-ieee-nfv-sdn.pdf
Quoted 14 lines
In general, UDPDK exhibits a much lower
latency than the in-kernel UDP stack used through the POSIX API (e.g., a 69 % reduction from 95 µs down to 29 µs), thanks to its ability to bypass the kernel entirely, which in turn outperforms the in-kernel TCP stack as available through the POSIX API, as expected.
...
Alternatively, application processes can use UDPDK with the non-blocking API calls (using theO_NONBLOCKflag) and perform some other action while waiting for packets to be ready to be sent/received to/from the UDPDK Process, instead of performing continuous busy-loops on packet queues. However, in this case the cost of a single CPU fully busy due to the UDPDK Process itself is anyway unavoidab
If the userspace workload avoids the kernel completely then it is quite unlikely that there is any pcp work to be flushed for in-kernel counters.
That being said, I am nor saying remote flushing is not useful. I just think that the issue you are reporting here could be fixed by a much simpler fix that doesn't change the way how the flushing is performed. Such a large rework should be justified by performance numbers. It should be also explained how do we end up doing a lot of work on isolated cpus or a pure user space workload.
Signature · 2 lines
Michal Hocko SUSE Labs
Quoted 86 lines · Michal Hocko
Quoted 85 lines · Marcelo Tosatti
Quoted 56 lines · Michal Hocko
Quoted 47 lines · Marcelo Tosatti
Quoted 20 lines · Michal Hocko
Quoted 6 lines · Marcelo Tosatti
This patch series addresses the following two problems:
1. A customer provided some evidence which indicates that the idle tick was stopped; albeit, CPU-specific vmstat counters still remained populated.
Quoted 12 lines
Thus one can only assume
quiet_vmstat()was not invoked on return to the idle loop. If I understand correctly, I suspect this divergence might erroneously prevent a reclaim attempt by kswapd. If the number of zone specific free pages are below their per-cpu drift value thenzone_page_state_snapshot()is used to compute a more accurate view of the aforementioned statistic. Thus any task blocked on the NUMA node specificpfmemalloc_waitqueue will be unable to make significant progress via direct reclaim unless it is killed after being woken up by kswapd
(seethrottle_direct_reclaim())Quoted 4 lines
I have hard time to follow the actual problem described above. Are you suggesting that a lack of pcp vmstat counters update has led to reclaim issues? What is the said "evidence"? Could you share more of the story please?
- The process was trapped in
throttle_direct_reclaim(). The functionwait_event_killable()was called to wait conditionallow_direct_reclaim(pgdat)for current node to be true. Theallow_direct_reclaim(pgdat)examined the number of free pages on the node byzone_page_state()which just returns value in zone->vm_stat[NR_FREE_PAGES].- On node #1, zone->vm_stat[NR_FREE_PAGES] was 0. However, the freelist on this node was not empty.
- This inconsistent of vmstat value was caused by percpu vmstat on
nohz_fullcpus. Every increment/decrement of vmstat is performed on percpu vmstat counter at first, then pooled diffs are cumulated to the zone's vmstat counter in timely manner. However, onnohz_fullcpus (in case of this customer's system, 48 of 52 cpus) these pooled diffs were not cumulated once the cpu had no event on it so that the cpu started sleeping infinitely.
I checked percpu vmstat and found there were total 69 counts not cumulated to the zone's vmstat counter yet.Quoted 7 lines
- In this situation, kswapd did not help the trapped process. In
pgdat_balanced(),zone_wakermark_ok_safe()examined the number of free pages on the node byzone_page_state_snapshot()which checks pending counts on percpu vmstat.
Therefore kswapd could know there were 69 free pages correctly. Sincezone->_watermark= {8, 20, 32}, kswapd did not work because 69 was greater than 32 as high watermark.Quoted 2 lines
If the imprecision of
allow_direct_reclaimis the underlying problem why haven't you usedzone_page_state_snapshotinstead?It might have dealt with problem #1 for this particular case. However, looking at the callers of
zone_page_state:5 2227 mm/compaction.c <<compaction_suitable>> zone_page_state(zone, NR_FREE_PAGES)); 6 124 mm/highmem.c <<__nr_free_highpages>> pages += zone_page_state(zone, NR_FREE_PAGES); 7 283 mm/page-writeback.c <<node_dirtyable_memory>> nr_pages += zone_page_state(zone, NR_FREE_PAGES); 8 318 mm/page-writeback.c <<highmem_dirtyable_memory>> nr_pages = zone_page_state(z, NR_FREE_PAGES); 9 321 mm/page-writeback.c <<highmem_dirtyable_memory>> nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE); 10 322 mm/page-writeback.c <<highmem_dirtyable_memory>> nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE); 11 3091 mm/page_alloc.c <<__rmqueue>> zone_page_state(zone, NR_FREE_CMA_PAGES) > 12 3092 mm/page_alloc.c <<__rmqueue>> zone_page_state(zone, NR_FREE_PAGES) / 2) {The suggested patchset fixes the problem of where due to
nohz_full, the delayed timer forvmstat_workcan be armed but not executed (which means the per-cpu counters can be out of sync for as long as the cpu is in idle while innohz_fullmode).
Quoted 5 lines
Quoted 3 lines
You probably do not want to convert all callers of
zone_page_stateintozone_page_state_snapshot(as a justification for the proposed patchset).Yes, I do not really think we want or even need to convert all of them.
OK.
Quoted 7 lines
But it seems that your direct reclaim throttling example really requires that. The thing with the remote flushing is that it would suffer from a similar imprecations as the flushing could be deferred and under certain conditions really starved.
So it is definitely worth fixing the
issue you are seeing without such a complex scheme.
The scheme is necessary for other reasons.
Quoted 77 lines
Quoted 4 lines
Quoted 2 lines
Anyway, this is kind of information that is really helpful to have in the patch description.
Agree: resending a new version with updated commit.
I would really recommend trying out the simple fix and see if it changes the behavior.
Quoted 24 lines
Quoted 24 lines
[...]
Quoted 22 lines
Quoted 9 lines
Quoted 6 lines
2. With a
SCHED_FIFOtask that busy loops on a given CPU, and kworker for that CPU atSCHED_OTHERpriority, queuing work to sync per-vmstats will either cause that work to never execute, or stalld (i.e. stall daemon) boosts kworker priority which causes a latency violationWhy is that a problem? Out-of-sync stats shouldn't cause major problems. Or can they?
Consider
SCHED_FIFOtask that is polling the network queue (say testpmd).do { if (net_registers->state & DATA_AVAILABLE) { process_data)(); } } while (!stopped);Since this task runs at
SCHED_FIFOpriority, kworker won't be scheduled to run (therefore per-CPU vmstats won't be flushed to global vmstats).Quoted 8 lines
Quoted 4 lines
Yes, that is certainly possible. But my main point is that vmstat imprecision shouldn't cause functional problems. That is why we have
_snapshotreaders to get an exact value where it matters for consistency.Understood. Perhaps
allow_direct_reclaimshould usezone_page_state_snapshot, as otherwise it is only precise atsysctl_stat_intervalintervals?or even much less than that. The flusher uses WQ infrastructure and even when a
WQ_MEM_RECLAIMone is used this doesn't mean that all workers could be jammed.Quoted 17 lines
Quoted 7 lines
Quoted 6 lines
Or, if testpmd runs at
SCHED_OTHER, then the work item to flush per-CPU vmstats causestestpmd -> kworker kworker: flush per-CPU vmstats kworker -> testpmdQuoted 5 lines
And this might cause undesired latencies to the packets being processed by the testpmd task.
Right but can you have any latencies expectation in a situation like that?
Not sure i understand what you mean. Example:
https://www.gabrieleara.it/assets/documents/papers/conferences/2021-ieee-nfv-sdn.pdf
Quoted 14 lines
In general, UDPDK exhibits a much lower
latency than the in-kernel UDP stack used through the POSIX API (e.g., a 69 % reduction from 95 µs down to 29 µs), thanks to its ability to bypass the kernel entirely, which in turn outperforms the in-kernel TCP stack as available through the POSIX API, as expected.
...
Alternatively, application processes can use UDPDK with the non-blocking API calls (using theO_NONBLOCKflag) and perform some other action while waiting for packets to be ready to be sent/received to/from the UDPDK Process, instead of performing continuous busy-loops on packet queues. However, in this case the cost of a single CPU fully busy due to the UDPDK Process itself is anyway unavoidab
Quoted 3 lines
If the userspace workload avoids the kernel completely then it is quite unlikely that there is any pcp work to be flushed for in-kernel counters.
This particular workload avoids the kernel. Others (were latency is still a concern) don't.
Quoted 5 lines
That being said, I am nor saying remote flushing is not useful.
I just think that the issue you are reporting here could be fixed by a much simpler fix that doesn't change the way how the flushing is performed.
OK. Must change flushing anyway, but fixing allow_direct_reclaim to use zone_page_state_snapshot won't hurt.
Quoted 1 line
Such a large rework should be justified by performance numbers.
OK.
Quoted 2 lines
It should be also explained how do we end up doing a lot of work on isolated cpus or a pure user space workload.
Again, pure user space workload is one example where latency matters, in response to the "can you have any latencies expectation in a situation like that?" question.
Will resend -v8 with allow_direct_reclaim fix.