Part 3 · 1 chapters · ~8 min

Memory Management

Physical pages, zones and the buddy allocator, slab allocators for kernel objects, virtual memory areas and page tables, page faults (minor and major), copy-on-write, transparent huge pages, reclaim with LRU lists and kswapd, overcommit, the OOM killer, cgroup memory limits, and reading /proc/meminfo.

Pages, faults, reclaim and the OOM killer

code

grep -E 'MemTotal|MemFree|MemAvailable|Cached|Dirty|AnonPages' /proc/meminfo
#  MemFree is small on a healthy server: free memory becomes page cache. MemAvailable is the honest number.
cat /proc/$(pgrep -o node)/status | grep -E 'VmRSS|VmSwap'
cat /proc/$(pgrep -o node)/smaps_rollup          # Rss, Pss, Anonymous, shared vs private
/usr/bin/time -v node app.js 2>&1 | grep -E 'Maximum resident|page faults'   # major vs minor faults
cat /proc/sys/vm/overcommit_memory               # 0 heuristic, 1 always, 2 strict
dmesg | grep -i 'out of memory'                  # OOM killer decisions, with the victim's oom_score

allocator	job
buddy allocator	hands out physical pages in power-of-two blocks
slab (SLUB)	caches of same-sized kernel objects (inodes, dentries, sk_buffs): /proc/slabinfo
page cache	file data in memory, reclaimable
reclaim (kswapd, direct reclaim)	frees clean cache, writes back dirty pages, swaps anonymous pages

Containers: a cgroup memory limit (memory.max) triggers reclaim and then an OOM kill inside that cgroup only, which is why a pod is "OOMKilled" while the node has free memory. Page cache from the container's own files counts toward its limit.

A PAGE FAULT, STEP BY STEP

virtual memory made real on first touch

swipe the figure sideways, or tap expand for full screen

1/4

mappings first

mmap and malloc create virtual memory areas (VMAs) but no physical pages. Memory is committed lazily, on first touch.

VMAs without pageslazy allocation