Skip to contentSkip to conversation

Kernel 5.1.15 stuck in compaction

DISCUSSION Max Kellermann 1 participant 2 messages Jul 8, 2019

lore
Share & tools
lore
#0 lore raw

Hi,

one of our web servers got repeatedly stuck in the memory compaction code; two PHP processes have been busy at 100% inside memory compaction after a page fault:

100.00% 0.00% php-cgi7.0 [kernel.vmlinux] [k] page_fault

            |
            ---page_fault
               __do_page_fault
               handle_mm_fault
               __handle_mm_fault
               do_huge_pmd_anonymous_page
               __alloc_pages_nodemask
               __alloc_pages_slowpath
               __alloc_pages_direct_compact
               try_to_compact_pages
               compact_zone_order
               compact_zone
               |          
               |--61.30%--isolate_migratepages_block
               |          |          
               |          |--20.44%--node_page_state
               |          |          
               |          |--5.88%--compact_unlock_should_abort.isra.33
               |          |          
               |           --3.28%--_cond_resched
               |                     |          
               |                      --2.19%--rcu_all_qs
               |          
                --3.37%--pageblock_skip_persistent

ftrace:

           <...>-962300 [033] .... 236536.493919: node_page_state <-isolate_migratepages_block
           <...>-962300 [033] .... 236536.493919: node_page_state <-isolate_migratepages_block
           <...>-962300 [033] .... 236536.493919: node_page_state <-isolate_migratepages_block
           <...>-962300 [033] .... 236536.493919: _cond_resched <-isolate_migratepages_block
           <...>-962300 [033] .... 236536.493919: rcu_all_qs <-_cond_resched
           <...>-962300 [033] .... 236536.493919: compact_unlock_should_abort.isra.33 <-isolate_migratepages_block
           <...>-962300 [033] .... 236536.493919: pageblock_skip_persistent <-compact_zone
           <...>-962300 [033] .... 236536.493919: isolate_migratepages_block <-compact_zone
           <...>-962300 [033] .... 236536.493919: node_page_state <-isolate_migratepages_block
           <...>-962300 [033] .... 236536.493919: node_page_state <-isolate_migratepages_block
           <...>-962300 [033] .... 236536.493919: node_page_state <-isolate_migratepages_block
           <...>-962300 [033] .... 236536.493919: node_page_state <-isolate_migratepages_block
           <...>-962300 [033] .... 236536.493920: node_page_state <-isolate_migratepages_block
           <...>-962300 [033] .... 236536.493920: node_page_state <-isolate_migratepages_block
           <...>-962300 [033] .... 236536.493920: _cond_resched <-isolate_migratepages_block
           <...>-962300 [033] .... 236536.493920: rcu_all_qs <-_cond_resched
           <...>-962300 [033] .... 236536.493920: compact_unlock_should_abort.isra.33 <-isolate_migratepages_block
           <...>-962300 [033] .... 236536.493920: pageblock_skip_persistent <-compact_zone
           <...>-962300 [033] .... 236536.493920: isolate_migratepages_block <-compact_zone
           <...>-962300 [033] .... 236536.493920: node_page_state <-isolate_migratepages_block

Nothing useful in /proc/PID/{stack,wchan,syscall}.

slabinfo/kmalloc-{16,32} are going through the roof (~ 15 GB each), and this memleak-lookalike triggering the oomkiller all the time is what drew our attention to this server.

Right now, the server is still stuck, and I can attempt to collect more information on request.

Max

#1 lore raw
Quoted 3 lines · Max Kellermann

one of our web servers got repeatedly stuck in the memory compaction code; two PHP processes have been busy at 100% inside memory compaction after a page fault:

This trace maybe helpful as well; the first PHP process:

275.846 compaction:mm_compaction_isolate_migratepages:range=(0x8a48e0 ~ 0x8a48e0) nr_scanned=0 nr_taken=0
LOST 8 events!
275.894 compaction:mm_compaction_isolate_migratepages:range=(0x8a48e0 ~ 0x8a48e0) nr_scanned=0 nr_taken=0
LOST 8 events!
275.942 compaction:mm_compaction_isolate_migratepages:range=(0x8a48e0 ~ 0x8a48e0) nr_scanned=0 nr_taken=0
LOST 8 events!
275.989 compaction:mm_compaction_isolate_migratepages:range=(0x8a48e0 ~ 0x8a48e0) nr_scanned=0 nr_taken=0

This is the other PHP process:

188.501 compaction:mm_compaction_isolate_migratepages:range=(0x169f40 ~ 0x169f40) nr_scanned=0 nr_taken=0
LOST 16 events!
188.600 compaction:mm_compaction_isolate_migratepages:range=(0x169f40 ~ 0x169f40) nr_scanned=0 nr_taken=0
LOST 5 events!
188.643 compaction:mm_compaction_isolate_migratepages:range=(0x169f40 ~ 0x169f40) nr_scanned=0 nr_taken=0
LOST 17 events!
188.742 compaction:mm_compaction_isolate_migratepages:range=(0x169f40 ~ 0x169f40) nr_scanned=0 nr_taken=0

No pages are being scanned at all, start and end are the same.

However, since my perf report contains calls to compact_unlock_should_abort(), this means that the loop in isolate_migratepages_block() is not getting skipped completely, therefore the loop is just exiting too early.