disable pcplists during page isolation
Subsystems (5)
cover disable pcplists during page isolation
As per the discussions [1] [2] this is an attempt to implement David's suggestion that page isolation should disable pcplists to avoid races. This is done without extra checks in fast paths, as I mentioned should be possible in the discussion, and explained in patch 5. Patches 1-4 are preparatory cleanups.
Note this is untested RFC for now. Based on v5.9-rc4 plus Pavel's patch [2] (slated as a quick fix for mainline+stable).
[1] https://lore.kernel.org/linux-mm/[email protected]/
[2] https://lore.kernel.org/linux-mm/[email protected]/
Vlastimil Babka (5):
mm, page_alloc: clean up pageset high and batch update mm, page_alloc: calculate pageset high and batch once per zone mm, page_alloc(): remove setup_pageset()
mm, page_alloc: cache pageset high and batch in struct zone mm, page_alloc: disable pcplists during page isolation
include/linux/gfp.h | 1 + include/linux/mmzone.h | 2 + mm/internal.h | 4 ++ mm/memory_hotplug.c | 24 +++---- mm/page_alloc.c | 138 ++++++++++++++++++++++------------------- mm/page_isolation.c | 45 +++++++++++--- 6 files changed, 127 insertions(+), 87 deletions(-)
Signature · 1 line
2.28.0
RFC 1/5 mm, page_alloc: clean up pageset high and batch update
The updates to pcplists' high and batch valued are handled by multiple functions that make the calculations hard to follow. Consolidate everything to pageset_set_high_and_batch() and remove pageset_set_batch() and pageset_set_high() wrappers.
The only special case using one of the removed wrappers was:build_all_zonelists_init()
setup_pageset()
pageset_set_batch()
which was hardcoding batch as 0, so we can just open-code a call to pageset_update() with constant parameters instead.
No functional change.
--- mm/page_alloc.c | 51 +++++++++++++++++++------------------------------ 1 file changed, 20 insertions(+), 31 deletions(-)
mm/page_alloc.c +20 -31
@@ -5834,7 +5834,7 @@ static void build_zonelists(pg_data_t *pgdat)* not check if the processor is online before following the pageset pointer. * Other parts of the kernel may not check if the zone is available. */ -static void setup_pageset (struct per_cpu_pageset *p,unsigned long batch); +static void setup_pageset (struct per_cpu_pageset *p); static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset); static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
@@ -5902,7 +5902,7 @@ build_all_zonelists_init(void)* (a chicken-egg dilemma). */ for_each_possible_cpu (cpu) -setup_pageset (&per_cpu (boot_pageset, cpu),0 ); +setup_pageset (&per_cpu (boot_pageset, cpu));mminit_verify_zonelist ();cpuset_init_current_mems_allowed ();
@@ -6218,12 +6218,6 @@ static void pageset_update(struct per_cpu_pages *pcp, unsigned long high, pcp->batch = batch; } -/* a companion to pageset_set_high() */ -static void pageset_set_batch (struct per_cpu_pageset *p,unsigned long batch) -{ -pageset_update (&p->pcp,6 * batch,max (1UL ,1 * batch)); -} -static void pageset_init (struct per_cpu_pageset *p) {struct per_cpu_pages *pcp;
@@ -6236,35 +6230,30 @@ static void pageset_init(struct per_cpu_pageset *p)INIT_LIST_HEAD (&pcp->lists[migratetype]); } -static void setup_pageset (struct per_cpu_pageset *p,unsigned long batch) +static void setup_pageset (struct per_cpu_pageset *p) {pageset_init (p); -pageset_set_batch (p, batch); -} - -/* -* pageset_set_high() sets the high water mark for hot per_cpu_pagelist -* to the value high for the pageset p. -*/ -static void pageset_set_high (struct per_cpu_pageset *p, -unsigned long high) -{ -unsigned long batch =max (1UL , high /4 ); -if ((high /4 ) > (PAGE_SHIFT *8 )) - batch =PAGE_SHIFT *8 ; - -pageset_update (&p->pcp, high, batch); +pageset_update (&p->pcp,0 ,1 ); }static void pageset_set_high_and_batch (struct zone *zone, -struct per_cpu_pageset *pcp) +struct per_cpu_pageset *p) { -if (percpu_pagelist_fraction) -pageset_set_high (pcp, - (zone_managed_pages (zone) / - percpu_pagelist_fraction)); -else -pageset_set_batch (pcp,zone_batchsize (zone)); +unsigned long new_high; +unsigned long new_batch; +int fraction =READ_ONCE (percpu_pagelist_fraction); + +if (fraction) { + new_high =zone_managed_pages (zone) / fraction; + new_batch =max (1UL , new_high /4 ); +if ((new_high /4 ) > (PAGE_SHIFT *8 )) + new_batch =PAGE_SHIFT *8 ; + }else { + new_batch =zone_batchsize (zone); + new_high =6 * new_batch; + new_batch =max (1UL ,1 * new_batch); + } +pageset_update (&p->pcp, new_high, new_batch); }static void __meminit zone_pageset_init (struct zone *zone,int cpu)
Signature · 1 line
2.28.0
Re: [RFC 1/5] mm, page_alloc: clean up pageset high and batch update
mm/page_alloc.c build_all_zonelists_init(void) · line 5905
Quoted 5 lines · Vlastimil Babka
Signed-off-by Vlastimil Babka <[email protected]>mm/page_alloc.c +1
-1for_each_possible_cpu (cpu) -setup_pageset (&per_cpu (boot_pageset, cpu),0 ); +setup_pageset (&per_cpu (boot_pageset, cpu));
This is not really anything important but I realized we have like 7 functions messing with pcp lists, and everytime I try to follow them my head spins.
Since setup_pageset is only being called here, could we replace it by the pageset_init and pageset_update?
(As I said, not important and probably a matter of taste. I just think that having so many mini functions around is not always cool,
e.g: setup_zone_pageset->zone_pageset_init)
Quoted 4 lines
mm/page_alloc.c +0
-4-/* -* pageset_set_high() sets the high water mark for hot per_cpu_pagelist -* to the value high for the pageset p. -*/
mm/page_alloc.c static void pageset_init(struct per_cpu_pageset *p) · line 6237
Quoted 10 lines
mm/page_alloc.c +1
-8-static void pageset_set_high (struct per_cpu_pageset *p, -unsigned long high) -{ -unsigned long batch =max (1UL , high /4 ); -if ((high /4 ) > (PAGE_SHIFT *8 )) - batch =PAGE_SHIFT *8 ; - -pageset_update (&p->pcp, high, batch); +pageset_update (&p->pcp,0 ,1 ); }
Could we restore the comment we had in pageset_set_high, and maybe update it to match this new function? I think it would be useful.
Quoted 4 lines
static void pageset_set_high_and_batch(struct zone *zone,
mm/page_alloc.c +1
-1- struct per_cpu_pageset *pcp) + struct per_cpu_pageset *p)
mm/page_alloc.c static void pageset_init(struct per_cpu_pageset *p) · line 6244
Quoted 10 lines
mm/page_alloc.c +3
-6{ -if (percpu_pagelist_fraction) -pageset_set_high (pcp, - (zone_managed_pages (zone) / - percpu_pagelist_fraction)); -else -pageset_set_batch (pcp,zone_batchsize (zone)); +unsigned long new_high; +unsigned long new_batch; +int fraction =READ_ONCE (percpu_pagelist_fraction);
Why the READ_ONCE? In case there is a parallel update so things to get messed up?
as I said, I'd appreciate a comment in pageset_set_high_and_batch to be restored and updated, otherwise:
Thanks
Signature · 2 lines
Oscar Salvador SUSE L3
Oscar Salvador
maintainerRe: [RFC 1/5] mm, page_alloc: clean up pageset high and batch update
Quoted 10 lines · Oscar Salvador
mm/page_alloc.c build_all_zonelists_init(void) · line 5905
Quoted 5 lines · Vlastimil Babka
Signed-off-by Vlastimil Babka <[email protected]>mm/page_alloc.c +1
-1for_each_possible_cpu (cpu) -setup_pageset (&per_cpu (boot_pageset, cpu),0 ); +setup_pageset (&per_cpu (boot_pageset, cpu));This is not really anything important but I realized we have like 7 functions messing with pcp lists, and everytime I try to follow them my head spins.
Quoted 6 lines
Since
setup_pagesetis only being called here, could we replace it by thepageset_initandpageset_update?(As I said, not important and probably a matter of taste. I just think that having so many mini functions around is not always cool,
e.g:setup_zone_pageset->zone_pageset_init)
Sorry, I did not see that you just did that in Patch#3, bleh.
Signature · 2 lines
Oscar Salvador SUSE L3
Re: [RFC 1/5] mm, page_alloc: clean up pageset high and batch update
In reply to Oscar SalvadorQuoted 10 lines · Oscar Salvador
mm/page_alloc.c build_all_zonelists_init(void) · line 5905
Quoted 5 lines · Vlastimil Babka
Signed-off-by Vlastimil Babka <[email protected]>mm/page_alloc.c +1
-1for_each_possible_cpu (cpu) -setup_pageset (&per_cpu (boot_pageset, cpu),0 ); +setup_pageset (&per_cpu (boot_pageset, cpu));This is not really anything important but I realized we have like 7 functions messing with pcp lists, and everytime I try to follow them my head spins.
Quoted 2 lines
Since
setup_pagesetis only being called here, could we replace it by thepageset_initandpageset_update?
Had the same thought, so +1.
Quoted 9 lines
Quoted 9 lines
mm/page_alloc.c +0
-9-/* -* pageset_set_high() sets the high water mark for hot per_cpu_pagelist -* to the value high for the pageset p. -*/ -static void pageset_set_high (struct per_cpu_pageset *p, -unsigned long high) -{ -unsigned long batch =max (1UL , high /4 ); -if ((high /4 ) > (PAGE_SHIFT *8 ))
Quoted 8 lines
mm/page_alloc.c static void pageset_init(struct per_cpu_pageset *p) · line 6237
Quoted 5 lines
mm/page_alloc.c +1
-3- batch =PAGE_SHIFT *8 ; - -pageset_update (&p->pcp, high, batch); +pageset_update (&p->pcp,0 ,1 ); }Could we restore the comment we had in
pageset_set_high, and maybe update it to match this new function? I think it would be useful.
At least I didn't really understand what "pageset_set_high() sets the high water mark for hot per_cpu_pagelist to the value high for the pageset p." was trying to tell me.
I think the only valuable information is the "hot", meaning it is in use and we have to be careful when updating, right?
Quoted 17 lines
Quoted 4 lines
static void pageset_set_high_and_batch(struct zone *zone,
mm/page_alloc.c +1
-1- struct per_cpu_pageset *pcp) + struct per_cpu_pageset *p)mm/page_alloc.c static void pageset_init(struct per_cpu_pageset *p) · line 6244
Quoted 10 lines
mm/page_alloc.c +3
-6{ -if (percpu_pagelist_fraction) -pageset_set_high (pcp, - (zone_managed_pages (zone) / - percpu_pagelist_fraction)); -else -pageset_set_batch (pcp,zone_batchsize (zone)); +unsigned long new_high; +unsigned long new_batch; +int fraction =READ_ONCE (percpu_pagelist_fraction);Why the
READ_ONCE? In case there is a parallel update so things to get messed up?
Agreed, this is an actual change in the code. If this is a fix, separate patch?
Apart from that, looks much better to me!
Signature · 3 lines
Thanks, David / dhildenb
Re: [RFC 1/5] mm, page_alloc: clean up pageset high and batch update
In reply to Oscar SalvadorQuoted 11 lines · Oscar Salvador
Quoted 9 lines · Vlastimil Babka
mm/page_alloc.c +0
-9-/* -* pageset_set_high() sets the high water mark for hot per_cpu_pagelist -* to the value high for the pageset p. -*/ -static void pageset_set_high (struct per_cpu_pageset *p, -unsigned long high) -{ -unsigned long batch =max (1UL , high /4 ); -if ((high /4 ) > (PAGE_SHIFT *8 ))
Quoted 8 lines
mm/page_alloc.c static void pageset_init(struct per_cpu_pageset *p) · line 6237
Quoted 5 lines
mm/page_alloc.c +1
-3- batch =PAGE_SHIFT *8 ; - -pageset_update (&p->pcp, high, batch); +pageset_update (&p->pcp,0 ,1 ); }Could we restore the comment we had in
pageset_set_high, and maybe update it to match this new function? I think it would be useful.
Same as David, I didn't find the comment useful at all. But I can try writing a better one instead.
Quoted 17 lines
Quoted 4 lines
static void pageset_set_high_and_batch(struct zone *zone,
mm/page_alloc.c +1
-1- struct per_cpu_pageset *pcp) + struct per_cpu_pageset *p)mm/page_alloc.c static void pageset_init(struct per_cpu_pageset *p) · line 6244
Quoted 10 lines
mm/page_alloc.c +3
-6{ -if (percpu_pagelist_fraction) -pageset_set_high (pcp, - (zone_managed_pages (zone) / - percpu_pagelist_fraction)); -else -pageset_set_batch (pcp,zone_batchsize (zone)); +unsigned long new_high; +unsigned long new_batch; +int fraction =READ_ONCE (percpu_pagelist_fraction);Why the
READ_ONCE? In case there is a parallel update so things to get messed up?
Yeah, I think online of a new zone can race with sysctl update to percpu_pagelist_fraction, because online doesn't take pcp_batch_high_lock... until patch 5. But you're right this should be separate.
Quoted 4 lines
as I said, I'd appreciate a comment in
pageset_set_high_and_batchto be restored and updated, otherwise:Reviewed-by Oscar Salvador <[email protected]>
Thanks!
Quoted 2 lines
Thanks
RFC 2/5 mm, page_alloc: calculate pageset high and batch once per zone
In reply to Vlastimil BabkaWe currently call pageset_set_high_and_batch() for each possible cpu, which repeats the same calculations of high and batch values.
Instead call it once per zone, and it applies the calculated values to all per-cpu pagesets of the zone.
This also allows removing zone_pageset_init() and __zone_pcp_update() wrappers.
No functional change.
--- mm/page_alloc.c | 40 +++++++++++++++++----------------------- 1 file changed, 17 insertions(+), 23 deletions(-)
mm/page_alloc.c +17 -23
@@ -6236,12 +6236,13 @@ static void setup_pageset(struct per_cpu_pageset *p)pageset_update (&p->pcp,0 ,1 ); } -static void pageset_set_high_and_batch (struct zone *zone, -struct per_cpu_pageset *p) +static void zone_set_pageset_high_and_batch (struct zone *zone) {unsigned long new_high;unsigned long new_batch;int fraction =READ_ONCE (percpu_pagelist_fraction); +int cpu; +struct per_cpu_pageset *p;if (fraction) { new_high =zone_managed_pages (zone) / fraction;
@@ -6253,23 +6254,25 @@ static void pageset_set_high_and_batch(struct zone *zone, new_high =6 * new_batch; new_batch =max (1UL ,1 * new_batch); } -pageset_update (&p->pcp, new_high, new_batch); -} - -static void __meminit zone_pageset_init (struct zone *zone,int cpu) -{ -struct per_cpu_pageset *pcp =per_cpu_ptr (zone->pageset, cpu); -pageset_init (pcp); -pageset_set_high_and_batch (zone, pcp); +for_each_possible_cpu (cpu) { + p =per_cpu_ptr (zone->pageset, cpu); +pageset_update (&p->pcp, new_high, new_batch); + } }void __meminit setup_zone_pageset (struct zone *zone) {int cpu; +struct per_cpu_pageset *p; + zone->pageset = alloc_percpu(struct per_cpu_pageset); -for_each_possible_cpu (cpu) -zone_pageset_init (zone, cpu); +for_each_possible_cpu (cpu) { + p =per_cpu_ptr (zone->pageset, cpu); +pageset_init (p); + } + +zone_set_pageset_high_and_batch (zone); }/*
@@ -8002,15 +8005,6 @@ int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,return 0 ; } -static void __zone_pcp_update (struct zone *zone) -{ -unsigned int cpu; - -for_each_possible_cpu (cpu) -pageset_set_high_and_batch (zone, -per_cpu_ptr (zone->pageset, cpu)); -} -/* * percpu_pagelist_fraction - changes the pcp->high for each zone on each * cpu. It is the fraction of total pages in each zone that a hot per cpu
@@ -8043,7 +8037,7 @@ int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write, goto out;for_each_populated_zone (zone) -__zone_pcp_update (zone); +zone_set_pageset_high_and_batch (zone); out:mutex_unlock (&pcp_batch_high_lock);return ret;
@@ -8659,7 +8653,7 @@ EXPORT_SYMBOL(free_contig_range);void __meminit zone_pcp_update (struct zone *zone) {mutex_lock (&pcp_batch_high_lock); -__zone_pcp_update (zone); +zone_set_pageset_high_and_batch (zone);mutex_unlock (&pcp_batch_high_lock); }
Signature · 1 line
2.28.0
Re: [RFC 2/5] mm, page_alloc: calculate pageset high and batch once per zone
Quoted 11 lines · Vlastimil Babka
We currently call
pageset_set_high_and_batch()for each possible cpu, which repeats the same calculations of high and batch values.Instead call it once per zone, and it applies the calculated values to all per-cpu pagesets of the zone.
This also allows removing
zone_pageset_init()and__zone_pcp_update()wrappers.No functional change.
Signed-off-by Vlastimil Babka <[email protected]>
I like this, it simplifies the things.
Signature · 2 lines
Oscar Salvador SUSE L3
Re: [RFC 2/5] mm, page_alloc: calculate pageset high and batch once per zone
In reply to Vlastimil BabkaQuoted 22 lines · Vlastimil Babka
We currently call
pageset_set_high_and_batch()for each possible cpu, which repeats the same calculations of high and batch values.Instead call it once per zone, and it applies the calculated values to all per-cpu pagesets of the zone.
This also allows removing
zone_pageset_init()and__zone_pcp_update()wrappers.No functional change.
Signed-off-by Vlastimil Babka <[email protected]>--- mm/page_alloc.c | 40 +++++++++++++++++----------------------- 1 file changed, 17 insertions(+), 23 deletions(-)mm/page_alloc.c +0
-0@@ -6236,12 +6236,13 @@ static void setup_pageset(struct per_cpu_pageset *p)pageset_update (&p->pcp,0 ,1 ); }
mm/page_alloc.c static void setup_pageset(struct per_cpu_pageset *p) · line 6245
Quoted 10 lines
mm/page_alloc.c +3
-2-static void pageset_set_high_and_batch (struct zone *zone, -struct per_cpu_pageset *p) +static void zone_set_pageset_high_and_batch (struct zone *zone) {unsigned long new_high;unsigned long new_batch;int fraction =READ_ONCE (percpu_pagelist_fraction); +int cpu; +struct per_cpu_pageset *p;
I would use reverse Christmas tree layout. (same applies to patch #1, you could put both unsigned long variables into a single line)
[...]
LGTM
Signature · 3 lines
Thanks, David / dhildenb
RFC 3/5 mm, page_alloc(): remove setup_pageset()
In reply to Vlastimil BabkaWe initialize boot-time pagesets with setup_pageset(), which sets high and batch values that effectively disable pcplists.
We can remove this wrapper if we just set these values for all pagesets in pageset_init(). Non-boot pagesets then subsequently update them to specific values.
--- mm/page_alloc.c | 13 +++++++------ 1 file changed, 7 insertions(+), 6 deletions(-)
mm/page_alloc.c +7 -6
@@ -5902,7 +5902,7 @@ build_all_zonelists_init(void)* (a chicken-egg dilemma). */ for_each_possible_cpu (cpu) -setup_pageset (&per_cpu (boot_pageset, cpu)); +pageset_init (&per_cpu (boot_pageset, cpu));mminit_verify_zonelist ();cpuset_init_current_mems_allowed ();
@@ -6228,12 +6228,13 @@ static void pageset_init(struct per_cpu_pageset *p) pcp = &p->pcp;for (migratetype =0 ; migratetype <MIGRATE_PCPTYPES ; migratetype++)INIT_LIST_HEAD (&pcp->lists[migratetype]); -} -static void setup_pageset (struct per_cpu_pageset *p) -{ -pageset_init (p); -pageset_update (&p->pcp,0 ,1 ); +/* +* Set batch and high values safe for a boot pageset. Proper pageset's +* initialization will update them. +*/ + pcp->high =0 ; + pcp->batch =1 ; }static void zone_set_pageset_high_and_batch (struct zone *zone)
Signature · 1 line
2.28.0
Re: [RFC 3/5] mm, page_alloc(): remove setup_pageset()
Quoted 8 lines · Vlastimil Babka
We initialize boot-time pagesets with
setup_pageset(), which sets high and batch values that effectively disable pcplists.We can remove this wrapper if we just set these values for all pagesets in
pageset_init(). Non-boot pagesets then subsequently update them to specific values.Signed-off-by Vlastimil Babka <[email protected]>
Just one question below:
mm/page_alloc.c static void pageset_init(struct per_cpu_pageset *p) · line 6237
Quoted 10 lines
mm/page_alloc.c +6
-4-static void setup_pageset (struct per_cpu_pageset *p) -{ -pageset_init (p); -pageset_update (&p->pcp,0 ,1 ); +/* +* Set batch and high values safe for a boot pageset. Proper pageset's +* initialization will update them. +*/ + pcp->high =0 ; + pcp->batch =1 ;
pageset_update was manipulating these values with barriers in between. I guess we do not care here because we are not really updating but initializing them, right?
Signature · 2 lines
Oscar Salvador SUSE L3
Oscar Salvador
maintainerRe: [RFC 3/5] mm, page_alloc(): remove setup_pageset()
Quoted 11 lines · Oscar Salvador
Quoted 8 lines · Vlastimil Babka
We initialize boot-time pagesets with
setup_pageset(), which sets high and batch values that effectively disable pcplists.We can remove this wrapper if we just set these values for all pagesets in
pageset_init(). Non-boot pagesets then subsequently update them to specific values.Signed-off-by Vlastimil Babka <[email protected]>Reviewed-by Oscar Salvador <[email protected]>
You would need to squash here the remove of setup_pageset's declaration. And then remove it from patch#4.
Signature · 2 lines
Oscar Salvador SUSE L3
Re: [RFC 3/5] mm, page_alloc(): remove setup_pageset()
In reply to Oscar SalvadorQuoted 11 lines · Oscar Salvador
Quoted 8 lines · Vlastimil Babka
We initialize boot-time pagesets with
setup_pageset(), which sets high and batch values that effectively disable pcplists.We can remove this wrapper if we just set these values for all pagesets in
pageset_init(). Non-boot pagesets then subsequently update them to specific values.Signed-off-by Vlastimil Babka <[email protected]>Reviewed-by Oscar Salvador <[email protected]>
Thanks!
Quoted 16 lines
Just one question below:
mm/page_alloc.c static void pageset_init(struct per_cpu_pageset *p) · line 6237
Quoted 10 lines
mm/page_alloc.c +6
-4-static void setup_pageset (struct per_cpu_pageset *p) -{ -pageset_init (p); -pageset_update (&p->pcp,0 ,1 ); +/* +* Set batch and high values safe for a boot pageset. Proper pageset's +* initialization will update them. +*/ + pcp->high =0 ; + pcp->batch =1 ;
pageset_updatewas manipulating these values with barriers in between. I guess we do not care here because we are not really updating but initializing them, right?
Sure. We just initialized all the list heads, so there can be no concurrent access at this point. But I'll mention it in the comment.
Re: [RFC 3/5] mm, page_alloc(): remove setup_pageset()
In reply to Vlastimil BabkaQuoted 7 lines · Vlastimil Babka
We initialize boot-time pagesets with
setup_pageset(), which sets high and batch values that effectively disable pcplists.We can remove this wrapper if we just set these values for all pagesets in
pageset_init(). Non-boot pagesets then subsequently update them to specific values.
subject: s/page_alloc()/page_alloc/ to make it look consistent
Quoted 28 lines
Signed-off-by Vlastimil Babka <[email protected]>--- mm/page_alloc.c | 13 +++++++------ 1 file changed, 7 insertions(+), 6 deletions(-)mm/page_alloc.c +1
-5@@ -5902,7 +5902,7 @@ build_all_zonelists_init(void)* (a chicken-egg dilemma). */ for_each_possible_cpu (cpu) -setup_pageset (&per_cpu (boot_pageset, cpu)); +pageset_init (&per_cpu (boot_pageset, cpu));mminit_verify_zonelist ();cpuset_init_current_mems_allowed ();@@ -6228,12 +6228,13 @@ static void pageset_init(struct per_cpu_pageset *p) pcp = &p->pcp;for (migratetype =0 ; migratetype <MIGRATE_PCPTYPES ; migratetype++)INIT_LIST_HEAD (&pcp->lists[migratetype]); -} -static void setup_pageset (struct per_cpu_pageset *p) -{ -pageset_init (p);
mm/page_alloc.c static void pageset_init(struct per_cpu_pageset *p) · line 6240
Quoted 10 lines
mm/page_alloc.c +6
-1-pageset_update (&p->pcp,0 ,1 ); +/* +* Set batch and high values safe for a boot pageset. Proper pageset's +* initialization will update them. +*/ + pcp->high =0 ; + pcp->batch =1 ; }static void zone_set_pageset_high_and_batch (struct zone *zone)
Had the same thought about barriers being useless when the pageset is not hot.
LGTM
Signature · 3 lines
Thanks, David / dhildenb
RFC 4/5 mm, page_alloc: cache pageset high and batch in struct zone
In reply to Vlastimil BabkaAll per-cpu pagesets for a zone use the same high and batch values, that are duplicated there just for performance (locality) reasons. This patch adds the same variables also to struct zone as 'central' ones.
This will be useful later for making possible to disable pcplists temporarily by setting high value to 0, while remembering the values for restoring them later. But we can also immediately benefit from not updating pagesets of all possible cpus in case the newly recalculated values (after sysctl change or memory online/offline) are actually unchanged from the previous ones.
--- include/linux/mmzone.h | 2 ++ mm/page_alloc.c | 18 +++++++++++++----- 2 files changed, 15 insertions(+), 5 deletions(-)
include/linux/mmzone.h +2 -0
@@ -431,6 +431,8 @@ struct zone {#endif struct pglist_data *zone_pgdat;struct per_cpu_pageset __percpu *pageset; +int pageset_high; +int pageset_batch;#ifndef CONFIG_SPARSEMEM /*
mm/page_alloc.c +13 -5
@@ -5834,7 +5834,7 @@ static void build_zonelists(pg_data_t *pgdat)* not check if the processor is online before following the pageset pointer. * Other parts of the kernel may not check if the zone is available. */ -static void setup_pageset (struct per_cpu_pageset *p); +static void pageset_init (struct per_cpu_pageset *p); static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset); static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
@@ -6237,7 +6237,7 @@ static void pageset_init(struct per_cpu_pageset *p) pcp->batch =1 ; } -static void zone_set_pageset_high_and_batch (struct zone *zone) +static void zone_set_pageset_high_and_batch (struct zone *zone,bool force_update) {unsigned long new_high;unsigned long new_batch;
@@ -6256,6 +6256,14 @@ static void zone_set_pageset_high_and_batch(struct zone *zone) new_batch =max (1UL ,1 * new_batch); } +if (zone->pageset_high != new_high || + zone->pageset_batch != new_batch) { + zone->pageset_high = new_high; + zone->pageset_batch = new_batch; + }else if (!force_update) { +return ; + } +for_each_possible_cpu (cpu) { p =per_cpu_ptr (zone->pageset, cpu);pageset_update (&p->pcp, new_high, new_batch);
@@ -6273,7 +6281,7 @@ void __meminit setup_zone_pageset(struct zone *zone)pageset_init (p); } -zone_set_pageset_high_and_batch (zone); +zone_set_pageset_high_and_batch (zone, true); }/*
@@ -8038,7 +8046,7 @@ int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write, goto out;for_each_populated_zone (zone) -zone_set_pageset_high_and_batch (zone); +zone_set_pageset_high_and_batch (zone, false); out:mutex_unlock (&pcp_batch_high_lock);return ret;
@@ -8654,7 +8662,7 @@ EXPORT_SYMBOL(free_contig_range);void __meminit zone_pcp_update (struct zone *zone) {mutex_lock (&pcp_batch_high_lock); -zone_set_pageset_high_and_batch (zone); +zone_set_pageset_high_and_batch (zone, false);mutex_unlock (&pcp_batch_high_lock); }
Signature · 1 line
2.28.0
Oscar Salvador
maintainerRe: [RFC 4/5] mm, page_alloc: cache pageset high and batch in struct zone
On Mon, Sep 07, 2020 at 06:36:27PM +0200, Vlastimil Babka wrote:
*/
mm/page_alloc.c static void build_zonelists(pg_data_t *pgdat) · line 5837
Quoted 2 lines
mm/page_alloc.c +1
-1-static void setup_pageset (struct per_cpu_pageset *p); +static void pageset_init (struct per_cpu_pageset *p);
this belongs to the respective patches
Quoted 6 lines
mm/page_alloc.c +1
-1-static void zone_set_pageset_high_and_batch (struct zone *zone) +static void zone_set_pageset_high_and_batch (struct zone *zone,bool force_update) {unsigned long new_high;unsigned long new_batch;
mm/page_alloc.c static void zone_set_pageset_high_and_batch(struct zone *zone) · line 6265
Quoted 10 lines
mm/page_alloc.c +7
-0@@ -6256,6 +6256,14 @@ static void zone_set_pageset_high_and_batch(struct zone *zone) new_batch =max (1UL ,1 * new_batch); } +if (zone->pageset_high != new_high || + zone->pageset_batch != new_batch) { + zone->pageset_high = new_high; + zone->pageset_batch = new_batch; + }else if (!force_update) { +return ; + }
I am probably missimg something obvious, so sorry, but why do we need force_update here?
AFAICS, we only want to call pageset_update() in case zone->pageset_high/batch and the new computed high/batch differs, so if everything is equal, why do we want to call it anyways?
Signature · 2 lines
Oscar Salvador SUSE L3
Re: [RFC 4/5] mm, page_alloc: cache pageset high and batch in struct zone
Quoted 6 lines · Oscar Salvador
On Mon, Sep 07, 2020 at 06:36:27PM +0200, Vlastimil Babka wrote:
*/mm/page_alloc.c static void build_zonelists(pg_data_t *pgdat) · line 5837
Quoted 2 lines
mm/page_alloc.c +1
-1-static void setup_pageset (struct per_cpu_pageset *p); +static void pageset_init (struct per_cpu_pageset *p);this belongs to the respective patches
Right, thanks.
Quoted 9 lines
Quoted 9 lines
mm/page_alloc.c +1
-1-static void zone_set_pageset_high_and_batch (struct zone *zone) +static void zone_set_pageset_high_and_batch (struct zone *zone,bool force_update) {unsigned long new_high;unsigned long new_batch;@@ -6256,6 +6256,14 @@ static void zone_set_pageset_high_and_batch(struct zone *zone) new_batch =max (1UL ,1 * new_batch); }
Quoted 13 lines
mm/page_alloc.c static void zone_set_pageset_high_and_batch(struct zone *zone) · line 6265
Quoted 7 lines
mm/page_alloc.c +7
-0+if (zone->pageset_high != new_high || + zone->pageset_batch != new_batch) { + zone->pageset_high = new_high; + zone->pageset_batch = new_batch; + }else if (!force_update) { +return ; + }I am probably missimg something obvious, so sorry, but why do we need
force_updatehere?
AFAICS, we only want to callpageset_update()in case zone->pageset_high/batch and the new computed high/batch differs, so if everything is equal, why do we want to call it anyways?
My reasoning is that initially we don't have guarantee that zone->pageset_high/batch matches the respective pcp->high/batch. So we could detect no change in the zone values and return, but leave the pcp value incoherent. But now I think it could be achieved also in a simpler way, so I'll try.
RFC 5/5 mm, page_alloc: disable pcplists during page isolation
In reply to Vlastimil BabkaPage isolation can race with process freeing pages to pcplists in a way that a page from isolated pageblock can end up on pcplist. This can be fixed by repeated draining of pcplists, as done by patch "mm/memory_hotplug: drain per-cpu pages again during memory offline" in [1].
David and Michal would prefer that this race was closed in a way that callers of page isolation don't need to care about drain. David suggested disabling pcplists usage completely during page isolation, instead of repeatedly draining them.
To achieve this without adding special cases in alloc/free fastpath, we can use the same 'trick' as boot pagesets - when pcp->high is 0, any pcplist addition will be immediately flushed.
The race can thus be closed by setting pcp->high to 0 and draining pcplists once in start_isolate_page_range(). The draining will serialize after processes that already disabled interrupts and read the old value of pcp->high in free_unref_page_commit(), and processes that have not yet disabled interrupts, will observe pcp->high == 0 when they are rescheduled, and skip pcplists. This guarantees no stray pages on pcplists in zones where isolation happens.
We can use the variable zone->nr_isolate_pageblock (protected by zone->lock) to detect transitions from 0 to 1 (to change pcp->high to 0 and issue drain) and from 1 to 0 (to restore original pcp->high and batch values cached in struct zone). We have to avoid external updates to high and batch by taking pcp_batch_high_lock. To allow multiple isolations in parallel, change this lock from mutex to rwsem.
For callers that pair start_isolate_page_range() with undo_isolated_page_range() properly, this is transparent. Currently that's alloc_contig_range(). __offline_pages() doesn't call undo_isolated_page_range() in the succes case, so it has to be carful to handle restoring pcp->high and batch and unlocking pcp_batch_high_lock.
This commit also changes drain_all_pages() to not trust reading pcp->count during drain for page isolation - I believe that could be racy and lead to missing some cpu's to drain. If others agree, this can be separated and potentially backported.
[1] https://lore.kernel.org/linux-mm/[email protected]/
--- include/linux/gfp.h | 1 + mm/internal.h | 4 ++++ mm/memory_hotplug.c | 24 ++++++++----------- mm/page_alloc.c | 58 +++++++++++++++++++++++++++++---------------- mm/page_isolation.c | 45 ++++++++++++++++++++++++++++------- 5 files changed, 89 insertions(+), 43 deletions(-)
include/linux/gfp.h +1 -0
@@ -592,6 +592,7 @@ extern void page_frag_free(void *addr);void page_alloc_init (void );void drain_zone_pages (struct zone *zone,struct per_cpu_pages *pcp); +void __drain_all_pages (struct zone *zone,bool page_isolation);void drain_all_pages (struct zone *zone);void drain_local_pages (struct zone *zone);
mm/internal.h +4 -0
@@ -7,6 +7,7 @@#ifndef __MM_INTERNAL_H#define __MM_INTERNAL_H +#include <linux/rwsem.h> #include <linux/fs.h> #include <linux/mm.h> #include <linux/pagemap.h>
@@ -201,8 +202,11 @@ extern void post_alloc_hook(struct page *page, unsigned int order, gfp_t gfp_flags);extern int user_min_free_kbytes; +extern struct rw_semaphore pcp_batch_high_lock;extern void zone_pcp_update (struct zone *zone);extern void zone_pcp_reset (struct zone *zone); +extern void zone_update_pageset_high_and_batch (struct zone *zone, +unsigned long high,unsigned long batch);#if definedCONFIG_COMPACTION || definedCONFIG_CMA
mm/memory_hotplug.c +10 -14
@@ -1485,6 +1485,7 @@ static int __ref __offline_pages(unsigned long start_pfn,struct zone *zone;struct memory_notify arg;char *reason; +bool unisolated_last = false;mem_hotplug_begin ();
@@ -1575,20 +1576,6 @@ static int __ref __offline_pages(unsigned long start_pfn,/* check again */ ret =walk_system_ram_range (start_pfn, end_pfn - start_pfn,NULL , check_pages_isolated_cb); -/* -* per-cpu pages are drained in start_isolate_page_range, but if -* there are still pages that are not free, make sure that we -* drain again, because when we isolated range we might -* have raced with another thread that was adding pages to pcp -* list. -* -* Forward progress should be still guaranteed because -* pages on the pcp list can only belong to MOVABLE_ZONE -* because has_unmovable_pages explicitly checks for -* PageBuddy on freed pages on other zones. -*/ -if (ret) -drain_all_pages (zone); }while (ret);/* Ok, all of our target is isolated.
@@ -1602,8 +1589,17 @@ static int __ref __offline_pages(unsigned long start_pfn,* pageblocks zone counter here. */ spin_lock_irqsave (&zone->lock, flags); +if (nr_isolate_pageblock && nr_isolate_pageblock == + zone->nr_isolate_pageblock) + unisolated_last = true; zone->nr_isolate_pageblock -= nr_isolate_pageblock;spin_unlock_irqrestore (&zone->lock, flags); +if (unisolated_last) { +zone_update_pageset_high_and_batch (zone, zone->pageset_high, + zone->pageset_batch); + } +/* pairs with start_isolate_page_range() */ +up_read (&pcp_batch_high_lock);/* removal success */ adjust_managed_page_count (pfn_to_page (start_pfn), -offlined_pages);
mm/page_alloc.c +38 -20
118 lines not on this page · raw message
mm/page_isolation.c +36 -9
103 lines not on this page · raw message
Signature · 1 line
2.28.0
Re: [RFC 5/5] mm, page_alloc: disable pcplists during page isolation
Here's a version that will apply on top of next-20200908. The first 4 patches need no change.
----8<----
From 8febc17272b8e8b378e2e5ea5e76b2616f029c5b Mon Sep 17 00:00:00 2001
From: Vlastimil Babka <[email protected]>
Date: Mon, 7 Sep 2020 17:20:39 +0200
Subject: [PATCH] mm, page_alloc: disable pcplists during page isolation
Page isolation can race with process freeing pages to pcplists in a way that a page from isolated pageblock can end up on pcplist. This can be fixed by repeated draining of pcplists, as done by patch "mm/memory_hotplug: drain per-cpu pages again during memory offline" in [1].
David and Michal would prefer that this race was closed in a way that callers of page isolation don't need to care about drain. David suggested disabling pcplists usage completely during page isolation, instead of repeatedly draining them.
To achieve this without adding special cases in alloc/free fastpath, we can use the same 'trick' as boot pagesets - when pcp->high is 0, any pcplist addition will be immediately flushed.
The race can thus be closed by setting pcp->high to 0 and draining pcplists once in start_isolate_page_range(). The draining will serialize after processes that already disabled interrupts and read the old value of pcp->high in free_unref_page_commit(), and processes that have not yet disabled interrupts, will observe pcp->high == 0 when they are rescheduled, and skip pcplists. This guarantees no stray pages on pcplists in zones where isolation happens.
We can use the variable zone->nr_isolate_pageblock (protected by zone->lock) to detect transitions from 0 to 1 (to change pcp->high to 0 and issue drain) and from 1 to 0 (to restore original pcp->high and batch values cached in struct zone). We have to avoid external updates to high and batch by taking pcp_batch_high_lock. To allow multiple isolations in parallel, change this lock from mutex to rwsem.
For callers that pair start_isolate_page_range() with undo_isolated_page_range() properly, this is transparent. Currently that's alloc_contig_range(). __offline_pages() doesn't call undo_isolated_page_range() in the succes case, so it has to be carful to handle restoring pcp->high and batch and unlocking pcp_batch_high_lock.
This commit also changes drain_all_pages() to not trust reading pcp->count during drain for page isolation - I believe that could be racy and lead to missing some cpu's to drain. If others agree, this can be separated and potentially backported.
[1] https://lore.kernel.org/linux-mm/[email protected]/
--- include/linux/gfp.h | 1 + mm/internal.h | 4 +++ mm/memory_hotplug.c | 55 ++++++++++++++++++++++++++++------------- mm/page_alloc.c | 60 +++++++++++++++++++++++++++++---------------- mm/page_isolation.c | 45 ++++++++++++++++++++++++++++------ 5 files changed, 119 insertions(+), 46 deletions(-)
include/linux/gfp.h +1 -0
@@ -592,6 +592,7 @@ extern void page_frag_free(void *addr);void page_alloc_init (void );void drain_zone_pages (struct zone *zone,struct per_cpu_pages *pcp); +void __drain_all_pages (struct zone *zone,bool page_isolation);void drain_all_pages (struct zone *zone);void drain_local_pages (struct zone *zone);
mm/internal.h +4 -0
@@ -7,6 +7,7 @@#ifndef __MM_INTERNAL_H#define __MM_INTERNAL_H +#include <linux/rwsem.h> #include <linux/fs.h> #include <linux/mm.h> #include <linux/pagemap.h>
@@ -196,8 +197,11 @@ extern void post_alloc_hook(struct page *page, unsigned int order, gfp_t gfp_flags);extern int user_min_free_kbytes; +extern struct rw_semaphore pcp_batch_high_lock;extern void zone_pcp_update (struct zone *zone);extern void zone_pcp_reset (struct zone *zone); +extern void zone_update_pageset_high_and_batch (struct zone *zone, +unsigned long high,unsigned long batch);#if definedCONFIG_COMPACTION || definedCONFIG_CMA
mm/memory_hotplug.c +38 -17
105 lines not on this page · raw message
mm/page_alloc.c +39 -21
126 lines not on this page · raw message
mm/page_isolation.c +37 -8
101 lines not on this page · raw message
Signature · 1 line
2.28.0
Michal Hocko
reviewerRe: [RFC 5/5] mm, page_alloc: disable pcplists during page isolation
Quoted 36 lines · Vlastimil Babka
Here's a version that will apply on top of next-20200908. The first 4 patches need no change.
----8<----
Quoted 1 line
From
8febc17272b8e8b378e2e5ea5e76b2616f029c5bMon Sep 17 00:00:00 2001From: Vlastimil Babka <[email protected]>
Date: Mon, 7 Sep 2020 17:20:39 +0200
Subject: [PATCH] mm,page_alloc: disable pcplists during page isolationPage isolation can race with process freeing pages to pcplists in a way that a page from isolated pageblock can end up on pcplist. This can be fixed by repeated draining of pcplists, as done by patch "
mm/memory_hotplug: drain per-cpu pages again during memory offline" in [1].David and Michal would prefer that this race was closed in a way that callers of page isolation don't need to care about drain. David suggested disabling pcplists usage completely during page isolation, instead of repeatedly draining them.
To achieve this without adding special cases in alloc/free fastpath, we can use the same 'trick' as boot pagesets - when
pcp->highis 0, any pcplist addition will be immediately flushed.The race can thus be closed by setting
pcp->highto 0 and draining pcplists once instart_isolate_page_range(). The draining will serialize after processes that already disabled interrupts and read the old value ofpcp->highinfree_unref_page_commit(), and processes that have not yet disabled interrupts, will observepcp->high== 0 when they are rescheduled, and skip pcplists. This guarantees no stray pages on pcplists in zones where isolation happens.We can use the variable
zone->nr_isolate_pageblock(protected byzone->lock) to detect transitions from 0 to 1 (to changepcp->highto 0 and issue drain) and from 1 to 0 (to restore originalpcp->highand batch values cached in struct zone). We have to avoid external updates to high and batch by takingpcp_batch_high_lock. To allow multiple isolations in parallel, change this lock from mutex to rwsem.
Quoted 5 lines
For callers that pair
start_isolate_page_range()withundo_isolated_page_range()properly, this is transparent. Currently that'salloc_contig_range().__offline_pages()doesn't callundo_isolated_page_range()in the succes case, so it has to be carful to handle restoringpcp->highand batch and unlockingpcp_batch_high_lock.
I was hoping that it would be possible to have this completely hidden inside start_isolate_page_range code path. If we need some sort of disable_pcp_free/enable_pcp_free then it seems like a better fit to have an explicit API for that (the naming would be obviously different because we do not want to call out pcp free lists). I strongly suspect that only the memory hotplug really cares for this hard guanrantee. alloc_contig_range simply goes with EBUSY.
Quoted 9 lines
This commit also changes
drain_all_pages()to not trust readingpcp->countduring drain for page isolation - I believe that could be racy and lead to missing some cpu's to drain. If others agree, this can be separated and potentially backported.[1] https://lore.kernel.org/linux-mm/[email protected]/
Suggested-by David Hildenbrand <[email protected]>Suggested-by Michal Hocko <[email protected]>Signed-off-by Vlastimil Babka <[email protected]>
Quoted 7 lines
--- include/linux/gfp.h | 1 + mm/internal.h | 4 +++ mm/memory_hotplug.c | 55 ++++++++++++++++++++++++++++------------- mm/page_alloc.c | 60 +++++++++++++++++++++++++++++---------------- mm/page_isolation.c | 45 ++++++++++++++++++++++++++++------ 5 files changed, 119 insertions(+), 46 deletions(-)
This has turned out much larger than I would expect.
Signature · 2 lines
Michal Hocko SUSE Labs
Re: [RFC 5/5] mm, page_alloc: disable pcplists during page isolation
Quoted 43 lines · Michal Hocko
Quoted 36 lines · Vlastimil Babka
Here's a version that will apply on top of next-20200908. The first 4 patches need no change.
----8<----
Quoted 1 line
From
8febc17272b8e8b378e2e5ea5e76b2616f029c5bMon Sep 17 00:00:00 2001From: Vlastimil Babka <[email protected]>
Date: Mon, 7 Sep 2020 17:20:39 +0200
Subject: [PATCH] mm,page_alloc: disable pcplists during page isolationPage isolation can race with process freeing pages to pcplists in a way that a page from isolated pageblock can end up on pcplist. This can be fixed by repeated draining of pcplists, as done by patch "
mm/memory_hotplug: drain per-cpu pages again during memory offline" in [1].David and Michal would prefer that this race was closed in a way that callers of page isolation don't need to care about drain. David suggested disabling pcplists usage completely during page isolation, instead of repeatedly draining them.
To achieve this without adding special cases in alloc/free fastpath, we can use the same 'trick' as boot pagesets - when
pcp->highis 0, any pcplist addition will be immediately flushed.The race can thus be closed by setting
pcp->highto 0 and draining pcplists once instart_isolate_page_range(). The draining will serialize after processes that already disabled interrupts and read the old value ofpcp->highinfree_unref_page_commit(), and processes that have not yet disabled interrupts, will observepcp->high== 0 when they are rescheduled, and skip pcplists. This guarantees no stray pages on pcplists in zones where isolation happens.We can use the variable
zone->nr_isolate_pageblock(protected byzone->lock) to detect transitions from 0 to 1 (to changepcp->highto 0 and issue drain) and from 1 to 0 (to restore originalpcp->highand batch values cached in struct zone). We have to avoid external updates to high and batch by takingpcp_batch_high_lock. To allow multiple isolations in parallel, change this lock from mutex to rwsem.Quoted 5 lines
For callers that pair
start_isolate_page_range()withundo_isolated_page_range()properly, this is transparent. Currently that'salloc_contig_range().__offline_pages()doesn't callundo_isolated_page_range()in the succes case, so it has to be carful to handle restoringpcp->highand batch and unlockingpcp_batch_high_lock.
Quoted 7 lines
I was hoping that it would be possible to have this completely hidden inside
start_isolate_page_rangecode path. If we need some sort of disable_pcp_free/enable_pcp_free then it seems like a better fit to have an explicit API for that (the naming would be obviously different because we do not want to call out pcp free lists). I strongly suspect that only the memory hotplug really cares for this hard guanrantee.alloc_contig_rangesimply goes with EBUSY.
There will be different alloc_contig_range() demands in the future: try fast (e.g., loads of small CMA allocations) vs. try hard (e.g., virtio-mem). We can add ways to specify that.
Signature · 3 lines
Thanks, David / dhildenb
Re: [RFC 5/5] mm, page_alloc: disable pcplists during page isolation
Quoted 52 lines · David Hildenbrand
Quoted 43 lines · Michal Hocko
Quoted 36 lines · Vlastimil Babka
Here's a version that will apply on top of next-20200908. The first 4 patches need no change.
----8<----
Quoted 1 line
From
8febc17272b8e8b378e2e5ea5e76b2616f029c5bMon Sep 17 00:00:00 2001From: Vlastimil Babka <[email protected]>
Date: Mon, 7 Sep 2020 17:20:39 +0200
Subject: [PATCH] mm,page_alloc: disable pcplists during page isolationPage isolation can race with process freeing pages to pcplists in a way that a page from isolated pageblock can end up on pcplist. This can be fixed by repeated draining of pcplists, as done by patch "
mm/memory_hotplug: drain per-cpu pages again during memory offline" in [1].David and Michal would prefer that this race was closed in a way that callers of page isolation don't need to care about drain. David suggested disabling pcplists usage completely during page isolation, instead of repeatedly draining them.
To achieve this without adding special cases in alloc/free fastpath, we can use the same 'trick' as boot pagesets - when
pcp->highis 0, any pcplist addition will be immediately flushed.The race can thus be closed by setting
pcp->highto 0 and draining pcplists once instart_isolate_page_range(). The draining will serialize after processes that already disabled interrupts and read the old value ofpcp->highinfree_unref_page_commit(), and processes that have not yet disabled interrupts, will observepcp->high== 0 when they are rescheduled, and skip pcplists. This guarantees no stray pages on pcplists in zones where isolation happens.We can use the variable
zone->nr_isolate_pageblock(protected byzone->lock) to detect transitions from 0 to 1 (to changepcp->highto 0 and issue drain) and from 1 to 0 (to restore originalpcp->highand batch values cached in struct zone). We have to avoid external updates to high and batch by takingpcp_batch_high_lock. To allow multiple isolations in parallel, change this lock from mutex to rwsem.Quoted 5 lines
For callers that pair
start_isolate_page_range()withundo_isolated_page_range()properly, this is transparent. Currently that'salloc_contig_range().__offline_pages()doesn't callundo_isolated_page_range()in the succes case, so it has to be carful to handle restoringpcp->highand batch and unlockingpcp_batch_high_lock.Quoted 7 lines
I was hoping that it would be possible to have this completely hidden inside
start_isolate_page_rangecode path. If we need some sort of disable_pcp_free/enable_pcp_free then it seems like a better fit to have an explicit API for that (the naming would be obviously different because we do not want to call out pcp free lists). I strongly suspect that only the memory hotplug really cares for this hard guanrantee.alloc_contig_rangesimply goes with EBUSY.
Quoted 4 lines
There will be different
alloc_contig_range()demands in the future: try fast (e.g., loads of small CMA allocations) vs. try hard (e.g., virtio-mem). We can add ways to specify that.
A reference to a related discussion regarding the "try fast" use case in CMA for the future:
https://lkml.kernel.org/r/[email protected]
Signature · 3 lines
Thanks, David / dhildenb
Re: [RFC 5/5] mm, page_alloc: disable pcplists during page isolation
In reply to Michal HockoQuoted 37 lines · Michal Hocko
Quoted 36 lines · Vlastimil Babka
Here's a version that will apply on top of next-20200908. The first 4 patches need no change.
----8<----
Quoted 1 line
From
8febc17272b8e8b378e2e5ea5e76b2616f029c5bMon Sep 17 00:00:00 2001From: Vlastimil Babka <[email protected]>
Date: Mon, 7 Sep 2020 17:20:39 +0200
Subject: [PATCH] mm,page_alloc: disable pcplists during page isolationPage isolation can race with process freeing pages to pcplists in a way that a page from isolated pageblock can end up on pcplist. This can be fixed by repeated draining of pcplists, as done by patch "
mm/memory_hotplug: drain per-cpu pages again during memory offline" in [1].David and Michal would prefer that this race was closed in a way that callers of page isolation don't need to care about drain. David suggested disabling pcplists usage completely during page isolation, instead of repeatedly draining them.
To achieve this without adding special cases in alloc/free fastpath, we can use the same 'trick' as boot pagesets - when
pcp->highis 0, any pcplist addition will be immediately flushed.The race can thus be closed by setting
pcp->highto 0 and draining pcplists once instart_isolate_page_range(). The draining will serialize after processes that already disabled interrupts and read the old value ofpcp->highinfree_unref_page_commit(), and processes that have not yet disabled interrupts, will observepcp->high== 0 when they are rescheduled, and skip pcplists. This guarantees no stray pages on pcplists in zones where isolation happens.We can use the variable
zone->nr_isolate_pageblock(protected byzone->lock) to detect transitions from 0 to 1 (to changepcp->highto 0 and issue drain) and from 1 to 0 (to restore originalpcp->highand batch values cached in struct zone). We have to avoid external updates to high and batch by takingpcp_batch_high_lock. To allow multiple isolations in parallel, change this lock from mutex to rwsem.
Quoted 8 lines
Quoted 5 lines
For callers that pair
start_isolate_page_range()withundo_isolated_page_range()properly, this is transparent. Currently that'salloc_contig_range().__offline_pages()doesn't callundo_isolated_page_range()in the succes case, so it has to be carful to handle restoringpcp->highand batch and unlockingpcp_batch_high_lock.I was hoping that it would be possible to have this completely hidden inside
start_isolate_page_rangecode path.
I hoped so too, but we can't know the moment when all processes that were in the critical part of freeing pages to pcplists have moved on (they might have been rescheduled).
We could change free_unref_page() to disable IRQs sooner, before free_unref_page_prepare(), or at least the get_pfnblock_migratetype() part. Then after the single drain, we should be safe, AFAICS?
RT guys might not be happy though, but it's much simpler than this patch. I still like some of the cleanups in 1-4 though tbh :)
Quoted 26 lines
If we need some sort of
disable_pcp_free/enable_pcp_free then it seems like a better fit to have an explicit API for that (the naming would be obviously different because we do not want to call out pcp free lists). I strongly suspect that only the memory hotplug really cares for this hard guanrantee.alloc_contig_rangesimply goes with EBUSY.Quoted 9 lines
This commit also changes
drain_all_pages()to not trust readingpcp->countduring drain for page isolation - I believe that could be racy and lead to missing some cpu's to drain. If others agree, this can be separated and potentially backported.[1] https://lore.kernel.org/linux-mm/[email protected]/
Suggested-by David Hildenbrand <[email protected]>Suggested-by Michal Hocko <[email protected]>Signed-off-by Vlastimil Babka <[email protected]>Quoted 7 lines
--- include/linux/gfp.h | 1 + mm/internal.h | 4 +++ mm/memory_hotplug.c | 55 ++++++++++++++++++++++++++++------------- mm/page_alloc.c | 60 +++++++++++++++++++++++++++++---------------- mm/page_isolation.c | 45 ++++++++++++++++++++++++++++------ 5 files changed, 119 insertions(+), 46 deletions(-)This has turned out much larger than I would expect.
Re: [RFC 5/5] mm, page_alloc: disable pcplists during page isolation
Quoted 47 lines · Vlastimil Babka
Quoted 37 lines · Michal Hocko
Quoted 36 lines · Vlastimil Babka
Here's a version that will apply on top of next-20200908. The first 4 patches need no change.
----8<----
Quoted 1 line
From
8febc17272b8e8b378e2e5ea5e76b2616f029c5bMon Sep 17 00:00:00 2001From: Vlastimil Babka <[email protected]>
Date: Mon, 7 Sep 2020 17:20:39 +0200
Subject: [PATCH] mm,page_alloc: disable pcplists during page isolationPage isolation can race with process freeing pages to pcplists in a way that a page from isolated pageblock can end up on pcplist. This can be fixed by repeated draining of pcplists, as done by patch "
mm/memory_hotplug: drain per-cpu pages again during memory offline" in [1].David and Michal would prefer that this race was closed in a way that callers of page isolation don't need to care about drain. David suggested disabling pcplists usage completely during page isolation, instead of repeatedly draining them.
To achieve this without adding special cases in alloc/free fastpath, we can use the same 'trick' as boot pagesets - when
pcp->highis 0, any pcplist addition will be immediately flushed.The race can thus be closed by setting
pcp->highto 0 and draining pcplists once instart_isolate_page_range(). The draining will serialize after processes that already disabled interrupts and read the old value ofpcp->highinfree_unref_page_commit(), and processes that have not yet disabled interrupts, will observepcp->high== 0 when they are rescheduled, and skip pcplists. This guarantees no stray pages on pcplists in zones where isolation happens.We can use the variable
zone->nr_isolate_pageblock(protected byzone->lock) to detect transitions from 0 to 1 (to changepcp->highto 0 and issue drain) and from 1 to 0 (to restore originalpcp->highand batch values cached in struct zone). We have to avoid external updates to high and batch by takingpcp_batch_high_lock. To allow multiple isolations in parallel, change this lock from mutex to rwsem.Quoted 8 lines
Quoted 5 lines
For callers that pair
start_isolate_page_range()withundo_isolated_page_range()properly, this is transparent. Currently that'salloc_contig_range().__offline_pages()doesn't callundo_isolated_page_range()in the succes case, so it has to be carful to handle restoringpcp->highand batch and unlockingpcp_batch_high_lock.I was hoping that it would be possible to have this completely hidden inside
start_isolate_page_rangecode path.
Quoted 6 lines
I hoped so too, but we can't know the moment when all processes that were in the critical part of freeing pages to pcplists have moved on (they might have been rescheduled).
We could changefree_unref_page()to disable IRQs sooner, beforefree_unref_page_prepare(), or at least theget_pfnblock_migratetype()part. Then after the single drain, we should be safe, AFAICS?
At least moving it before getting the migratetype should not be that severe?
Quoted 2 lines
RT guys might not be happy though, but it's much simpler than this patch. I still like some of the cleanups in 1-4 though tbh :)
It would certainly make this patch much simpler. Do you have a prototype lying around?
Signature · 3 lines
Thanks, David / dhildenb
Re: [RFC 5/5] mm, page_alloc: disable pcplists during page isolation
Quoted 20 lines · David Hildenbrand
Quoted 12 lines · Vlastimil Babka
Quoted 10 lines · Michal Hocko
Quoted 6 lines · Vlastimil Babka
Here's a version that will apply on top of next-20200908. The first 4 patches need no change. For callers that pair
start_isolate_page_range()withundo_isolated_page_range()properly, this is transparent. Currently that'salloc_contig_range().__offline_pages()doesn't callundo_isolated_page_range()in the succes case, so it has to be carful to handle restoringpcp->highand batch and unlockingpcp_batch_high_lock.I was hoping that it would be possible to have this completely hidden inside
start_isolate_page_rangecode path.Quoted 6 lines
I hoped so too, but we can't know the moment when all processes that were in the critical part of freeing pages to pcplists have moved on (they might have been rescheduled).
We could changefree_unref_page()to disable IRQs sooner, beforefree_unref_page_prepare(), or at least theget_pfnblock_migratetype()part. Then after the single drain, we should be safe, AFAICS?
Quoted 7 lines
At least moving it before getting the migratetype should not be that severe?
Quoted 2 lines
RT guys might not be happy though, but it's much simpler than this patch. I still like some of the cleanups in 1-4 though tbh :)
It would certainly make this patch much simpler. Do you have a prototype lying around?
Below is the critical part, while writing it I realized that there's also free_unref_page_list() where it's more ugly.
So free_unref_page() simply moves the "determine migratetype" part under disabled interrupts. For free_unref_page_list() we optimistically determine them without disabling interrupts, and with disabled interrupts we check if zone has isolated pageblocks and thus we should not trust that value anymore. That's same pattern as free_pcppages_bulk uses.
But unfortunately it means an extra page_zone() plus a test for each page on the list :/
----8<----
mm/page_alloc.c +20 -7
@@ -3103,18 +3103,21 @@ void mark_free_pages(struct zone *zone) }#endif /* CONFIG_PM */ -static bool free_unref_page_prepare (struct page *page,unsigned long pfn) +static bool free_unref_page_prepare (struct page *page) { -int migratetype; -if (!free_pcp_prepare (page))return false; - migratetype =get_pfnblock_migratetype (page, pfn); -set_pcppage_migratetype (page, migratetype);return true; } +static inline void free_unref_page_set_migratetype (struct page *page,unsigned long pfn) +{ +int migratetype =get_pfnblock_migratetype (page, pfn); + +set_pcppage_migratetype (page, migratetype); +} +static void free_unref_page_commit (struct page *page,unsigned long pfn) {struct zone *zone =page_zone (page);
@@ -3156,10 +3159,17 @@ void free_unref_page(struct page *page)unsigned long flags;unsigned long pfn =page_to_pfn (page); -if (!free_unref_page_prepare (page, pfn)) +if (!free_unref_page_prepare (page))return ; +/* +* by disabling interrupts before reading pageblock's migratetype, +* we can guarantee that after changing a pageblock to MIGRATE_ISOLATE +* and drain_all_pages(), there's nobody who would read the old +* migratetype and put a page from isoalted pageblock to pcplists +*/ local_irq_save (flags); +free_unref_page_set_migratetype (page, pfn);free_unref_page_commit (page, pfn);local_irq_restore (flags); }
@@ -3176,9 +3186,10 @@ void free_unref_page_list(struct list_head *list)/* Prepare pages for freeing */ list_for_each_entry_safe (page, next, list, lru) { pfn =page_to_pfn (page); -if (!free_unref_page_prepare (page, pfn)) +if (!free_unref_page_prepare (page))list_del (&page->lru);set_page_private (page, pfn); +free_unref_page_set_migratetype (page, pfn); }local_irq_save (flags);
@@ -3187,6 +3198,8 @@ void free_unref_page_list(struct list_head *list)set_page_private (page,0 );trace_mm_page_free_batched (page); +if (has_isolate_pageblock (page_zone (page))) +free_unref_page_set_migratetype (page, pfn);free_unref_page_commit (page, pfn);/*
Re: [RFC 5/5] mm, page_alloc: disable pcplists during page isolation
Quoted 32 lines · Vlastimil Babka
Quoted 20 lines · David Hildenbrand
Quoted 12 lines · Vlastimil Babka
Quoted 10 lines · Michal Hocko
Quoted 6 lines · Vlastimil Babka
Here's a version that will apply on top of next-20200908. The first 4 patches need no change. For callers that pair
start_isolate_page_range()withundo_isolated_page_range()properly, this is transparent. Currently that'salloc_contig_range().__offline_pages()doesn't callundo_isolated_page_range()in the succes case, so it has to be carful to handle restoringpcp->highand batch and unlockingpcp_batch_high_lock.I was hoping that it would be possible to have this completely hidden inside
start_isolate_page_rangecode path.Quoted 6 lines
I hoped so too, but we can't know the moment when all processes that were in the critical part of freeing pages to pcplists have moved on (they might have been rescheduled).
We could changefree_unref_page()to disable IRQs sooner, beforefree_unref_page_prepare(), or at least theget_pfnblock_migratetype()part. Then after the single drain, we should be safe, AFAICS?Quoted 7 lines
At least moving it before getting the migratetype should not be that severe?
Quoted 2 lines
RT guys might not be happy though, but it's much simpler than this patch. I still like some of the cleanups in 1-4 though tbh :)
It would certainly make this patch much simpler. Do you have a prototype lying around?
Below is the critical part, while writing it I realized that there's also
free_unref_page_list()where it's more ugly.
Quoted 8 lines
So
free_unref_page()simply moves the "determine migratetype" part under disabled interrupts. Forfree_unref_page_list()we optimistically determine them without disabling interrupts, and with disabled interrupts we check if zone has isolated pageblocks and thus we should not trust that value anymore. That's same pattern asfree_pcppages_bulkuses.But unfortunately it means an extra
page_zone()plus a test for each page on the list :/
That function is mostly called from mm/vmscan.c AFAIKS.
The fist thing free_unref_page_commit() does is doing page_zone(). That's easy to optimize if needed.
Quoted 65 lines
----8<----
mm/page_alloc.c +18
-7@@ -3103,18 +3103,21 @@ void mark_free_pages(struct zone *zone) }#endif /* CONFIG_PM */ -static bool free_unref_page_prepare (struct page *page,unsigned long pfn) +static bool free_unref_page_prepare (struct page *page) { -int migratetype; -if (!free_pcp_prepare (page))return false; - migratetype =get_pfnblock_migratetype (page, pfn); -set_pcppage_migratetype (page, migratetype);return true; } +static inline void free_unref_page_set_migratetype (struct page *page,unsigned long pfn) +{ +int migratetype =get_pfnblock_migratetype (page, pfn); + +set_pcppage_migratetype (page, migratetype); +} +static void free_unref_page_commit (struct page *page,unsigned long pfn) {struct zone *zone =page_zone (page);@@ -3156,10 +3159,17 @@ void free_unref_page(struct page *page)unsigned long flags;unsigned long pfn =page_to_pfn (page); -if (!free_unref_page_prepare (page, pfn)) +if (!free_unref_page_prepare (page))return ; +/* +* by disabling interrupts before reading pageblock's migratetype, +* we can guarantee that after changing a pageblock to MIGRATE_ISOLATE +* and drain_all_pages(), there's nobody who would read the old +* migratetype and put a page from isoalted pageblock to pcplists +*/ local_irq_save (flags); +free_unref_page_set_migratetype (page, pfn);free_unref_page_commit (page, pfn);local_irq_restore (flags); }@@ -3176,9 +3186,10 @@ void free_unref_page_list(struct list_head *list)/* Prepare pages for freeing */ list_for_each_entry_safe (page, next, list, lru) { pfn =page_to_pfn (page); -if (!free_unref_page_prepare (page, pfn)) +if (!free_unref_page_prepare (page))list_del (&page->lru);set_page_private (page, pfn); +free_unref_page_set_migratetype (page, pfn); }local_irq_save (flags);
mm/page_alloc.c void free_unref_page_list(struct list_head *list) · line 3204
Quoted 7 lines
mm/page_alloc.c +2
-0@@ -3187,6 +3198,8 @@ void free_unref_page_list(struct list_head *list)set_page_private (page,0 );trace_mm_page_free_batched (page); +if (has_isolate_pageblock (page_zone (page))) +free_unref_page_set_migratetype (page, pfn);free_unref_page_commit (page, pfn);
FWIW: You can safely add unlikely() for the has_isolate_pageblock(), as for most other users of that function I was able to spot.
Signature · 3 lines
Thanks, David / dhildenb
Re: [RFC 5/5] mm, page_alloc: disable pcplists during page isolation
In reply to Michal HockoQuoted 4 lines
because we do not want to call out pcp free lists). I strongly suspect that only the memory hotplug really cares for this hard guanrantee.
alloc_contig_rangesimply goes with EBUSY.
You are right, the strong requirement is for hotplug case only, but if PCP disable/enable functionality is properly implemented, and we are still draining the lists in alloc_contig_range case, I do not see a reason not to disable/enable PCP during CMA as well just to reduce number of spurious errors.
Thanks,
Pasha
Quoted 18 lines · Vlastimil Babka
As per the discussions [1] [2] this is an attempt to implement David's suggestion that page isolation should disable pcplists to avoid races. This is done without extra checks in fast paths, as I mentioned should be possible in the discussion, and explained in patch 5. Patches 1-4 are preparatory cleanups.
Note this is untested RFC for now. Based on v5.9-rc4 plus Pavel's patch [2] (slated as a quick fix for mainline+stable).
[1] https://lore.kernel.org/linux-mm/[email protected]/
[2] https://lore.kernel.org/linux-mm/[email protected]/Vlastimil Babka (5):
mm,page_alloc: clean up pageset high and batch update mm,page_alloc: calculate pageset high and batch once per zone mm,page_alloc(): removesetup_pageset()
mm,page_alloc: cache pageset high and batch in struct zone mm,page_alloc: disable pcplists during page isolation
Quoted 8 lines
include/linux/gfp.h | 1 + include/linux/mmzone.h | 2 + mm/internal.h | 4 ++ mm/memory_hotplug.c | 24 +++---- mm/page_alloc.c | 138 ++++++++++++++++++++++------------------- mm/page_isolation.c | 45 +++++++++++--- 6 files changed, 127 insertions(+), 87 deletions(-)
Thanks for looking into this! Just a heads-up that -mm and -next contain some changes to memory hotplug code, whereby new pageblocks start out in MIGRATE_ISOLATE when onlining, until we're done with the heavy lifting. Might require some tweaks, similar to when isolating pageblocks.
Will dive into this in the following days. What's you're general perception of performance aspects?
Signature · 3 lines
Thanks, David / dhildenb
Quoted 28 lines · David Hildenbrand
Quoted 18 lines · Vlastimil Babka
As per the discussions [1] [2] this is an attempt to implement David's suggestion that page isolation should disable pcplists to avoid races. This is done without extra checks in fast paths, as I mentioned should be possible in the discussion, and explained in patch 5. Patches 1-4 are preparatory cleanups.
Note this is untested RFC for now. Based on v5.9-rc4 plus Pavel's patch [2] (slated as a quick fix for mainline+stable).
[1] https://lore.kernel.org/linux-mm/[email protected]/
[2] https://lore.kernel.org/linux-mm/[email protected]/Vlastimil Babka (5):
mm,page_alloc: clean up pageset high and batch update mm,page_alloc: calculate pageset high and batch once per zone mm,page_alloc(): removesetup_pageset()
mm,page_alloc: cache pageset high and batch in struct zone mm,page_alloc: disable pcplists during page isolationQuoted 8 lines
include/linux/gfp.h | 1 + include/linux/mmzone.h | 2 + mm/internal.h | 4 ++ mm/memory_hotplug.c | 24 +++---- mm/page_alloc.c | 138 ++++++++++++++++++++++------------------- mm/page_isolation.c | 45 +++++++++++--- 6 files changed, 127 insertions(+), 87 deletions(-)
Quoted 4 lines
Thanks for looking into this! Just a heads-up that -mm and -next contain some changes to memory hotplug code, whereby new pageblocks start out in
MIGRATE_ISOLATEwhen onlining, until we're done with the heavy lifting. Might require some tweaks, similar to when isolating pageblocks.
Thanks for the heads-up. I've posted updated patch 5/5 for the -next as a reply to the first one. It was a bit tricky to order everything correctly in online_pages(), hopefully I avoided any deadlock.
Quoted 2 lines
Will dive into this in the following days. What's you're general perception of performance aspects?
Thanks! I expect no performance change while no isolation is in progress, as there are no new tests added in alloc/free paths. During page isolation there's a single drain instead of once-per-pageblock, which is a benefit. But the pcplists are effectively disabled for the whole of online_pages(), offline_pages() or alloc_contig_range(), which will affect parallel page allocator users. It depends on how long these operations take and how heavy the parallel usage is, so I have no good answers. Might be similar to the current periodic drain.
[email protected]
maintainerQuoted 8 lines · Vlastimil Babka
Thanks! I expect no performance change while no isolation is in progress, as there are no new tests added in alloc/free paths. During page isolation there's a single drain instead of once-per-pageblock, which is a benefit. But the pcplists are effectively disabled for the whole of
online_pages(),offline_pages()oralloc_contig_range(), which will affect parallel page allocator users. It depends on how long these operations take and how heavy the parallel usage is, so I have no good answers. Might be similar to the current periodic drain.
I have seen some systems taking quite some time when offlining sections due to the migration of
the respective pages not being that smooth and having do_migrate_range to do some spins.
But to be fair, online_pages and offline_pages are not routines that get called that often, and we would be safe to assume that memory-hotplug operations are not constantly happening, but are rather one-offs operations.
I am not sure about Xen and HV, IIRC Xen was using online_pages and offline_pages routines to do the ballooning?
I will dive in this in the following days, thanks for the work Vlastimil
Quoted 25 lines · On 09.09.20 13:27, [email protected] wrote:
Quoted 6 lines · Vlastimil Babka
Thanks! I expect no performance change while no isolation is in progress, as
there are no new tests added in alloc/free paths. During page isolation there's
a single drain instead of once-per-pageblock, which is a benefit. But theQuoted 8 lines
pcplists are effectively disabled for the whole of
online_pages(),offline_pages()oralloc_contig_range(), which will affect parallel page
allocator users. It depends on how long these operations take and how heavy the
parallel usage is, so I have no good answers. Might be similar to the current
periodic drain.I have seen some systems taking quite some time when offlining sections due to the migration of
the respective pages not being that smooth and havingdo_migrate_rangeto do some spins.
But to be fair,online_pagesandoffline_pagesare not routines that get called that often, and we would be safe to assume that memory-hotplug operations are not constantly happening, but are rather one-offs operations.
Quoted 2 lines
I am not sure about Xen and HV, IIRC Xen was using
online_pagesandoffline_pagesroutines to do the ballooning?
No, they only add more memory blocks and online them. Memory hot(un)plug is an expensive operation already, I don't think that's an issue.
Signature · 3 lines
Thanks, David / dhildenb