lkml.org 
[lkml]   [2022]   [Dec]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [PATCH v2] [mm-unstable] mm: Fix memcg reclaim on memory tiered systems
Date
Mina Almasry <almasrymina@google.com> writes:

> commit 3f1509c57b1b ("Revert "mm/vmscan: never demote for memcg
> reclaim"") enabled demotion in memcg reclaim, which is the right thing
> to do, but introduced a regression in the behavior of
> try_to_free_mem_cgroup_pages().
>
> The callers of try_to_free_mem_cgroup_pages() expect it to attempt to
> reclaim - not demote - nr_pages from the cgroup. I.e. the memory usage
> of the cgroup should reduce by nr_pages. The callers expect
> try_to_free_mem_cgroup_pages() to also return the number of pages
> reclaimed, not demoted.
>
> However, try_to_free_mem_cgroup_pages() actually unconditionally counts
> demoted pages as reclaimed pages. So in practice when it is called it will
> often demote nr_pages and return the number of demoted pages to the caller.
> Demoted pages don't lower the memcg usage as the caller requested.
>
> I suspect various things work suboptimally on memory systems or don't
> work at all due to this:
>
> - memory.high enforcement likely doesn't work (it just demotes nr_pages
> instead of lowering the memcg usage by nr_pages).
> - try_charge_memcg() will keep retrying the charge while
> try_to_free_mem_cgroup_pages() is just demoting pages and not actually
> making any room for the charge.
> - memory.reclaim has a wonky interface. It advertises to the user it
> reclaims the provided amount but it will actually demote that amount.
>
> There may be more effects to this issue.
>
> To fix these issues I propose shrink_folio_list() to only count pages
> demoted from inside of sc->nodemask to outside of sc->nodemask as
> 'reclaimed'.
>
> For callers such as reclaim_high() or try_charge_memcg() that set
> sc->nodemask to NULL, try_to_free_mem_cgroup_pages() will try to
> actually reclaim nr_pages and return the number of pages reclaimed. No
> demoted pages would count towards the nr_pages requirement.
>
> For callers such as memory_reclaim() that set sc->nodemask,
> try_to_free_mem_cgroup_pages() will free nr_pages from that nodemask
> with either demotion or reclaim.
>
> Tested this change using memory.reclaim interface. With this change,
>
> echo "1m" > memory.reclaim
>
> Will cause freeing of 1m of memory from the cgroup regardless of the
> demotions happening inside.
>
> echo "1m nodes=0" > memory.reclaim
>
> Will cause freeing of 1m of node 0 by demotion if a demotion target is
> available, and by reclaim if no demotion target is available.
>
> Signed-off-by: Mina Almasry <almasrymina@google.com>
>
> ---
>
> This is developed on top of mm-unstable largely to test with memory.reclaim
> nodes= arg and ensure the fix is compatible with that.
>
> v2:
> - Shortened the commit message a bit.
> - Fixed issue when demotion falls back to other allowed target nodes returned by
> node_get_allowed_targets() as Wei suggested.
>
> Cc: weixugc@google.com
> ---
> include/linux/memory-tiers.h | 7 +++++--
> mm/memory-tiers.c | 10 +++++++++-
> mm/vmscan.c | 20 +++++++++++++++++---
> 3 files changed, 31 insertions(+), 6 deletions(-)
>
> diff --git a/include/linux/memory-tiers.h b/include/linux/memory-tiers.h
> index fc9647b1b4f9..f3f359760fd0 100644
> --- a/include/linux/memory-tiers.h
> +++ b/include/linux/memory-tiers.h
> @@ -38,7 +38,8 @@ void init_node_memory_type(int node, struct memory_dev_type *default_type);
> void clear_node_memory_type(int node, struct memory_dev_type *memtype);
> #ifdef CONFIG_MIGRATION
> int next_demotion_node(int node);
> -void node_get_allowed_targets(pg_data_t *pgdat, nodemask_t *targets);
> +void node_get_allowed_targets(pg_data_t *pgdat, nodemask_t *targets,
> + nodemask_t *demote_from_targets);
> bool node_is_toptier(int node);
> #else
> static inline int next_demotion_node(int node)
> @@ -46,7 +47,9 @@ static inline int next_demotion_node(int node)
> return NUMA_NO_NODE;
> }
>
> -static inline void node_get_allowed_targets(pg_data_t *pgdat, nodemask_t *targets)
> +static inline void node_get_allowed_targets(pg_data_t *pgdat,
> + nodemask_t *targets,
> + nodemask_t *demote_from_targets)
> {
> *targets = NODE_MASK_NONE;
> }
> diff --git a/mm/memory-tiers.c b/mm/memory-tiers.c
> index c734658c6242..7f8f0b5de2b3 100644
> --- a/mm/memory-tiers.c
> +++ b/mm/memory-tiers.c
> @@ -264,7 +264,8 @@ bool node_is_toptier(int node)
> return toptier;
> }
>
> -void node_get_allowed_targets(pg_data_t *pgdat, nodemask_t *targets)
> +void node_get_allowed_targets(pg_data_t *pgdat, nodemask_t *targets,
> + nodemask_t *demote_from_targets)
> {
> struct memory_tier *memtier;
>
> @@ -280,6 +281,13 @@ void node_get_allowed_targets(pg_data_t *pgdat, nodemask_t *targets)
> else
> *targets = NODE_MASK_NONE;
> rcu_read_unlock();
> +
> + /*
> + * Exclude the demote_from_targets from the allowed targets if we're
> + * trying to demote from a specific set of nodes.
> + */
> + if (demote_from_targets)
> + nodes_andnot(*targets, *targets, *demote_from_targets);
> }

Will this cause demotion to not work when we have memory policy like
MPOL_BIND with nodemask including demotion targets?


>
> /**
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index 2b42ac9ad755..97ca0445b5dc 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -1590,7 +1590,8 @@ static struct page *alloc_demote_page(struct page *page, unsigned long private)
> * Folios which are not demoted are left on @demote_folios.
> */
> static unsigned int demote_folio_list(struct list_head *demote_folios,
> - struct pglist_data *pgdat)
> + struct pglist_data *pgdat,
> + nodemask_t *demote_from_nodemask)
> {
> int target_nid = next_demotion_node(pgdat->node_id);
> unsigned int nr_succeeded;
> @@ -1614,7 +1615,7 @@ static unsigned int demote_folio_list(struct list_head *demote_folios,
> if (target_nid == NUMA_NO_NODE)
> return 0;
>
> - node_get_allowed_targets(pgdat, &allowed_mask);
> + node_get_allowed_targets(pgdat, &allowed_mask, demote_from_nodemask);
>
> /* Demotion ignores all cpuset and mempolicy settings */
> migrate_pages(demote_folios, alloc_demote_page, NULL,
> @@ -1653,6 +1654,7 @@ static unsigned int shrink_folio_list(struct list_head *folio_list,
> LIST_HEAD(free_folios);
> LIST_HEAD(demote_folios);
> unsigned int nr_reclaimed = 0;
> + unsigned int nr_demoted = 0;
> unsigned int pgactivate = 0;
> bool do_demote_pass;
> struct swap_iocb *plug = NULL;
> @@ -2085,7 +2087,19 @@ static unsigned int shrink_folio_list(struct list_head *folio_list,
> /* 'folio_list' is always empty here */
>
> /* Migrate folios selected for demotion */
> - nr_reclaimed += demote_folio_list(&demote_folios, pgdat);
> + nr_demoted = demote_folio_list(&demote_folios, pgdat, sc->nodemask);
> +
> + /*
> + * Only count demoted folios as reclaimed if the caller has requested
> + * demotion from a specific nodemask. In this case pages inside the
> + * noedmask have been demoted to outside the nodemask and we can count
> + * these pages as reclaimed. If no nodemask is passed, then the caller
> + * is requesting reclaim from all memory, which should not count
> + * demoted pages.
> + */
> + if (sc->nodemask)
> + nr_reclaimed += nr_demoted;
> +
> /* Folios that could not be demoted are still in @demote_folios */
> if (!list_empty(&demote_folios)) {
> /* Folios which weren't demoted go back on @folio_list */
> --
> 2.39.0.rc0.267.gcb52ba06e7-goog

\
 
 \ /
  Last update: 2022-12-07 09:08    [W:0.067 / U:0.468 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site