Skip to content

Demo: How Much Page Table a Process Has

The kernel reports the size of a process's page tables as VmPTE. Compare it with what a flat table would take.

Goal

Make the multi-level page table concrete with one number: a process's tables are tens of kilobytes, not hundreds of gigabytes, because only the branches in use exist.

Run

grep -E 'VmSize|VmRSS|VmPTE' /proc/$$/status
grep -E 'VmSize|VmRSS|VmPTE' /proc/$(pgrep -f chrom | tail -1)/status

Results and explanations

$ grep -E 'VmSize|VmRSS|VmPTE' /proc/$$/status
VmSize:    10588 kB
VmRSS:      4392 kB
VmPTE:        60 kB
$ grep -E 'VmSize|VmRSS|VmPTE' /proc/$(pgrep -f chrom | tail -1)/status
VmSize: 1522658872 kB
VmRSS:   1284720 kB
VmPTE:      7812 kB

The arithmetic first. A page table with one entry per page of the whole 128 TiB user address space, laid out flat so that the page number can index it directly, has 2^35 entries of 8 bytes: 256 GB, per process, before the process has touched anything. Even a table that only covered the shell's own 10.3 MB would be 21 KB, but a flat table cannot "only cover" the zones, because the stack is at 0x7ffd..., the libraries at 0x7f... and the program at 0x5..., and the array has to span the distance between them.

What the kernel reports is 60 KB for the shell. The table is a tree of four levels on x86-64, each level a 4 KB page of 512 entries; a level-4 table points to level-3 tables, those to level-2, those to level-1 tables whose entries point at frames. A branch with no mapped page below it is simply absent, so the shell pays for one root, a few tables on the path to each of its clusters of zones, and the leaf tables for the pages it has actually mapped: fifteen 4 KB pages of tables in all.

The browser has 1.5 TB of address space and 1.2 GB resident, and its tables take 7.8 MB. A flat table for its address space alone would be 3 GB. The tree makes the size of the table follow the size of what is mapped, not the size of what is reserved, which is the only reason a process can reserve a terabyte for nothing.

Going further

  • Run the allocate-10TB demo from lecture 04 and read its VmPTE after the 10 TB reservation: unchanged, because nothing was touched and no leaf table was needed. Then touch one byte per gigabyte and watch it grow by a few pages per touch.
  • grep -E 'PageTables' /proc/meminfo is the machine-wide total.
  • On this CPU grep -c la57 /proc/cpuinfo says whether a fifth level is available, for 57-bit addresses; the kernel uses it only when asked for an address above 128 TiB.

References

  • man 5 proc (/proc/[pid]/status: VmPTE; /proc/meminfo: PageTables)
  • Intel SDM volume 3, chapter 4, or the AMD APM volume 2, chapter 5, for the exact layout of the four levels