Microbenchmarking the page table isolation patches in Linux 4.4.13

Running on Xeon W3580 CPUs (first generation i7, 3.33 GHz base). Kernel 4.14.13. One machine has Page Table Isolation compiled in and disabled on the commandline (“nopti”).

I will put the benchmarks for this on github this weekend, I promise!

No PTI:

     216.6 nsec/call       17.5 user      199.0 sys: getrusage
     190.0 nsec/call       21.1 user      168.9 sys: read
     127.4 nsec/call       23.9 user      103.4 sys: read1bdevzero
     783.0 nsec/call       25.0 user      758.0 sys: read8kdevzero     
    

PTI:

     542.1 nsec/call      172.1 user      369.9 sys: getrusage
     600.5 nsec/call      145.7 user      454.7 sys: read
     536.8 nsec/call      166.5 user      370.3 sys: read1bdevzero
    1194.8 nsec/call      157.8 user     1037.0 sys: read8kdevzero

All things page table changes from userland are looking better than that, but not good. Might be a problem for garbage collectors using them.

No PTI:

Size of mapped area for protect/segv tests: 8192 MB, 262144 cards
no touch by default
    1403.5 nsec/call       63.6 user     1339.9 sys: mprotect_randloc
     209.5 nsec/call       33.5 user      176.0 sys: mprotect_sameloc
    3691.7 nsec/call       44.3 user     3647.5 sys: mmap_no_touch
  268322.4 usec/call 27394000.0 user 240927000.0 sys 262144.000 hpf 0.000 spf: mmap_and_touch
cpu  2598 0 4013 23730505 3580 0 290 0 0 0
    4618.2 nsec/call      304.6 user     4313.3 sys 0.938 hpf 0.000 spf: trigger_segfaults
  (101538 segfaults, 108267 writes, watch collision ratio)
cpu  2601 0 4061 23730864 3580 0 290 0 0 0
     792.9 nsec/call       85.6 user      707.3 sys: madvise_randloc
    1386.6 nsec/call       80.4 user     1306.1 sys: mprotect_randloc
with touch by default
      12.0 usec/call     1270.5 user    10778.9 sys 8.435 hpf 0.000 spf: mprotect_randloc
     634.9 nsec/call       48.2 user      586.6 sys: mprotect_sameloc
    3949.0 nsec/call       94.7 user     3854.4 sys: mmap_no_touch
  268316.5 usec/call 23921000.0 user 244398500.0 sys 262144.000 hpf 0.000 spf: mmap_and_touch
cpu  2638 0 4457 23733898 3580 0 290 0 0 0

PTI:

Size of mapped area for protect/segv tests: 8192 MB, 262144 cards
no touch by default
    1876.7 nsec/call      304.2 user     1572.6 sys: mprotect_randloc
     612.3 nsec/call      237.6 user      374.7 sys: mprotect_sameloc
    5099.0 nsec/call      316.4 user     4782.8 sys: mmap_no_touch
  341625.5 usec/call 55585000.0 user 286046000.0 sys 262144.000 hpf 0.000 spf: mmap_and_touch
cpu  8148 0 9147 23718982 1028 0 394 0 0 0
    5680.5 nsec/call      942.9 user     4737.6 sys 0.944 hpf 0.000 spf: trigger_segfaults
  (83074 segfaults, 88021 writes, watch collision ratio)
cpu  8157 0 9193 23719362 1028 0 394 0 0 0
    1191.0 nsec/call      307.2 user      883.7 sys: madvise_randloc
    1867.1 nsec/call      292.7 user     1574.4 sys: mprotect_randloc
with touch by default
      11.4 usec/call     1548.2 user     9834.8 sys 5.968 hpf 0.000 spf: mprotect_randloc
    1018.3 nsec/call      230.2 user      788.2 sys: mprotect_sameloc
    5449.8 nsec/call      334.9 user     5108.6 sys: mmap_no_touch
  341527.5 usec/call 61519000.0 user 280016500.0 sys 262144.000 hpf 0.000 spf: mmap_and_touch
cpu  8231 0 9577 23722557 1028 0 394 0 0 0

Full logs:

• https://www.cons.org/pti/pti_off.log
• https://www.cons.org/pti/pti_on.log
• https://www.cons.org/pti/freebsd.log

Code coming.