| Message ID | 1561655524-89276-2-git-send-email-yang.shi@linux.alibaba.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [1/2,-mm] mm: account lazy free pages separately | expand |
And how about this one? On 6/27/19 10:12 AM, Yang Shi wrote: > Available memory is one of the most important metrics for memory > pressure. Currently, lazy free pages are not accounted into available > memory, but they are reclaimable actually, like reclaimable slabs. > > Accounting lazy free pages into available memory should reflect the real > memory pressure status, and also would help administrators and/or other > high level scheduling tools make better decision. > > The /proc/meminfo would show more available memory with test which > creates ~1GB deferred split THP. > > Before: > MemAvailable: 43544272 kB > ... > AnonHugePages: 10240 kB > ShmemHugePages: 0 kB > ShmemPmdMapped: 0 kB > LazyFreePages: 1046528 kB > > After: > MemAvailable: 44415124 kB > ... > AnonHugePages: 6144 kB > ShmemHugePages: 0 kB > ShmemPmdMapped: 0 kB > LazyFreePages: 1046528 kB > > MADV_FREE pages are not accounted for NR_LAZYFREE since they have been > put on inactive file LRU and accounted into available memory. > Accounting here would double account them. > > Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com> > --- > mm/page_alloc.c | 5 +++++ > 1 file changed, 5 insertions(+) > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c > index cab50e8..58ceca5 100644 > --- a/mm/page_alloc.c > +++ b/mm/page_alloc.c > @@ -5005,6 +5005,7 @@ long si_mem_available(void) > unsigned long wmark_low = 0; > unsigned long pages[NR_LRU_LISTS]; > unsigned long reclaimable; > + unsigned long lazyfree; > struct zone *zone; > int lru; > > @@ -5038,6 +5039,10 @@ long si_mem_available(void) > global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE); > available += reclaimable - min(reclaimable / 2, wmark_low); > > + /* Lazyfree pages are reclaimable when memory pressure is hit */ > + lazyfree = global_node_page_state(NR_LAZYFREE); > + available += lazyfree - min(lazyfree / 2, wmark_low); > + > if (available < 0) > available = 0; > return available;
diff --git a/mm/page_alloc.c b/mm/page_alloc.c index cab50e8..58ceca5 100644 --- a/mm/page_alloc.c +++ b/mm/page_alloc.c @@ -5005,6 +5005,7 @@ long si_mem_available(void) unsigned long wmark_low = 0; unsigned long pages[NR_LRU_LISTS]; unsigned long reclaimable; + unsigned long lazyfree; struct zone *zone; int lru; @@ -5038,6 +5039,10 @@ long si_mem_available(void) global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE); available += reclaimable - min(reclaimable / 2, wmark_low); + /* Lazyfree pages are reclaimable when memory pressure is hit */ + lazyfree = global_node_page_state(NR_LAZYFREE); + available += lazyfree - min(lazyfree / 2, wmark_low); + if (available < 0) available = 0; return available;
Available memory is one of the most important metrics for memory pressure. Currently, lazy free pages are not accounted into available memory, but they are reclaimable actually, like reclaimable slabs. Accounting lazy free pages into available memory should reflect the real memory pressure status, and also would help administrators and/or other high level scheduling tools make better decision. The /proc/meminfo would show more available memory with test which creates ~1GB deferred split THP. Before: MemAvailable: 43544272 kB ... AnonHugePages: 10240 kB ShmemHugePages: 0 kB ShmemPmdMapped: 0 kB LazyFreePages: 1046528 kB After: MemAvailable: 44415124 kB ... AnonHugePages: 6144 kB ShmemHugePages: 0 kB ShmemPmdMapped: 0 kB LazyFreePages: 1046528 kB MADV_FREE pages are not accounted for NR_LAZYFREE since they have been put on inactive file LRU and accounted into available memory. Accounting here would double account them. Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com> --- mm/page_alloc.c | 5 +++++ 1 file changed, 5 insertions(+)