Rewrite mlockall logic?

[hakavlad]

1. Use /dev/shm instead of /tmp.
   /dev/shm always uses tmpfs.

2. Use mlockall() (works with python3):

from ctypes import CDLL

def mlockall():
    """Lock all memory to prevent swapping process."""

    MCL_CURRENT = 1
    MCL_FUTURE = 2
    MCL_ONFAULT = 4

    libc = CDLL('libc.so.6', use_errno=True)

    result = libc.mlockall(
        MCL_CURRENT | MCL_FUTURE | MCL_ONFAULT
    )
    if result != 0:
        result = libc.mlockall(
            MCL_CURRENT | MCL_FUTURE
        )
        if result != 0:
            print('Cannot lock all memory')
        else:
            print('All memory locked with MCL_CURRENT | MCL_FUTURE')
    else:
        print('All memory locked with MCL_CURRENT | MCL_FUTURE | MCL_ONFAULT')

mlockall()

[tobixen]

I've moved the first one to a separate issue.

The second suggestion, I'm not sure I understand it. Rewritten slightly, this is the gist of your suggestion:

from ctypes import CDLL
CDLL('libc.so.6', use_errno=True).mlockall(3)

While this is roughly the logic as it is in the code today:

from ctypes import cdll
cdll.LoadLibrary('libc.so.6').mlockall(3)

both approaches seems to work both under python2 and python3. Both yields 0 for me when I run it as root, and -1 when I run it as an ordinary user.

[hakavlad]

mlockall(3) is not good in Fedora 29, for example: the daemon will use more then 200 MiB VmRSS with mlockall(3). We should use mlockall(MCL_CURRENT | MCL_FUTURE | MCL_ONFAULT).
Same problem: golang/go#28114.

MCL_ONFAULT appeared in Linux 4.4.

[hakavlad]

MCL_ONFAULT reduces memory usage on modern systems.

[hakavlad]

status with MCL_ONFAULT on Debian 9:

cat /proc/21763/status
Name:	thrash-protect.
Umask:	0022
State:	S (sleeping)
Tgid:	21763
Ngid:	0
Pid:	21763
PPid:	21762
TracerPid:	0
Uid:	0	0	0	0
Gid:	0	0	0	0
FDSize:	64
Groups:	0 
NStgid:	21763
NSpid:	21763
NSpgid:	21762
NSsid:	21718
VmPeak:	   39480 kB
VmSize:	   39480 kB
VmLck:	   39464 kB
VmPin:	       0 kB
VmHWM:	   12276 kB
VmRSS:	   12140 kB

Without MCL_ONFAULT:

Name:	thrash-protect.
Umask:	0022
State:	S (sleeping)
Tgid:	21793
Ngid:	0
Pid:	21793
PPid:	21792
TracerPid:	0
Uid:	0	0	0	0
Gid:	0	0	0	0
FDSize:	64
Groups:	0 
NStgid:	21793
NSpid:	21793
NSpgid:	21792
NSsid:	21718
VmPeak:	   39480 kB
VmSize:	   39480 kB
VmLck:	   39464 kB
VmPin:	       0 kB
VmHWM:	   15372 kB
VmRSS:	   15372 kB

[hakavlad]

thrash-protect's VmRSS on Fedora 29: 222 MiB! Using MCL_ONFAULT fixes the problem.

[hakavlad]

IMHO thrash-protect can be optimized to use only 10-12 MiB RSS.

[tobixen]

Is this sufficient?

[hakavlad]

Excellent, thanks.
Now VmRSS is 12.5 MiB on Fedora 29.