mlockall() not working properly in 2.6.x
Hi!
A year ago, I wrote a small mlockall()-wrapper ("noswap") to make certain programs unswappable. It used to work perfectly, until I upgraded to kernel 2.6.x (2.6.13.1 in my case, but that shouldn't matter), which made the mlockall() execute without error, but also without any effect (the "L" in the STAT column of "ps axf" which indicates locked pages is missing).
The complete program source and the strace log of
strace -v -f -s 4096 noswap sleep 9999
is here:
http://marc-jano.de/tmp/noswap-10935/
Here are the most important lines:
-----< snip >-----
/* mark process as unswappable for all times */
if (ret = mlockall (MCL_CURRENT | MCL_FUTURE))
perror ("mlockall()"); /* give up root privileges */
if (ret = setuid (getuid()) ) { perror ("setuid()"); exit (ret); };
if (ret = setgid (getgid()) ) { perror ("setgid()"); exit (ret); };
/* execute program */
if (ret = execvp (argv[1], exec_argv)) perror ("execvp()");
exit (ret);-----< snip >-----
The strace log shows that mlockall() returns with no error.
If I replace the execvp() call with a sleep(9999), then the current process (noswap) is properly locked, as can be checked by looking for a capital "L" in the ps STAT column.
So it seems that the mlockall() flags are not passed somehow when using execvp().
Best regards
Marc-Jano (not subscribed to the list, so for further questions, please Cc:)
Quoted 10 lines · Marc-Jano Knopp
Hi!
A year ago, I wrote a small mlockall()-wrapper ("noswap") to make certain programs unswappable. It used to work perfectly, until I upgraded to kernel 2.6.x (2.6.13.1 in my case, but that shouldn't matter), which made the
mlockall()execute without error, but also without any effect (the "L" in the STAT column of "ps axf" which indicates locked pages is missing).
Question is: what kernel version did you upgrade from?
Prior to 2.4.18 the kernel would allow MCL_FUTURE to propagate into child processes. But that was disabled in 2.4.18 and later. I seem to recall that we did this because inheriting MCL_FUTURE is standards-incorrect.
Quoted 11 lines · Andrew Morton
Quoted 8 lines · Marc-Jano Knopp
A year ago, I wrote a small mlockall()-wrapper ("noswap") to make certain programs unswappable. It used to work perfectly, until I upgraded to kernel 2.6.x (2.6.13.1 in my case, but that shouldn't matter), which made the
mlockall()execute without error, but also without any effect (the "L" in the STAT column of "ps axf" which indicates locked pages is missing).Question is: what kernel version did you upgrade from?
2.4.31. Just rebooted to 2.4.31 and tried again - mlockall() seems to work perfectly:
# ps axf|grep [9]99 1037 tty1 SL+ 0:00 \_ sleep 999 # uname -a Linux pc8 2.4.31 #3 Thu Sep 8 16:49:45 CEST 2005 i686 unknown #
Quoted 3 lines
Prior to 2.4.18 the kernel would allow
MCL_FUTUREto propagate into child processes. But that was disabled in 2.4.18 and later. I seem to recall that we did this because inheritingMCL_FUTUREis standards-incorrect.
Oh! So how can I make programs unswappable with kernel 2.6.x then?
Best regards
Marc-Jano
Quoted 8 lines · Marc-Jano Knopp
Quoted 7 lines · Andrew Morton
Quoted 6 lines · Marc-Jano Knopp
A year ago, I wrote a small mlockall()-wrapper ("noswap") to make certain programs unswappable. It used to work perfectly, until I upgraded to kernel 2.6.x (2.6.13.1 in my case, but that shouldn't matter), which made the
mlockall()execute without error, but also without any effect (the "L" in the STAT column of "ps axf" which indicates locked pages is missing).
Quoted 5 lines
Quoted 3 lines
Prior to 2.4.18 the kernel would allow
MCL_FUTUREto propagate into child processes. But that was disabled in 2.4.18 and later. I seem to recall that we did this because inheritingMCL_FUTUREis standards-incorrect.Oh! So how can I make programs unswappable with kernel 2.6.x then?
That would mean that you cannot just exec() another program that will also be mlockall()ed. The new program has to do that on its own...
MfG, JBG
Signature · 4 lines
Jan-Benedict Glaw [email protected] . +49-172-7608481 _ O _ "Eine Freie Meinung in einem Freien Kopf | Gegen Zensur | Gegen Krieg _ _ O für einen Freien Staat voll Freier Bürger" | im Internet! | im Irak! O O O ret = do_actions((curr | FREE_SPEECH) & ~(NEW_COPYRIGHT_LAW | DRM | TCPA));
Quoted 6 lines · Jan-Benedict Glaw
Quoted 3 lines
Quoted 1 line
that we did this because inheriting
MCL_FUTUREis standards-incorrect.Oh! So how can I make programs unswappable with kernel 2.6.x then?
That would mean that you cannot just
exec()another program that will also be mlockall()ed. The new program has to do that on its own...
mlockall MCL_FUTURE applies to this image only and the 2.6 behaviour is correct if less useful in some ways. It would be possible to add an inheriting MCL_ flag that was Linux specific but then how do you control the depth of inheritance ? If that isn't an issue it looks the easiest.
Another possibility would be pmlockall(pid, flag), but that looks even more nasty if it races an exec.
Quoted 8 lines · Alan Cox
Quoted 6 lines · Jan-Benedict Glaw
Quoted 3 lines
Quoted 1 line
that we did this because inheriting
MCL_FUTUREis standards-incorrect.Oh! So how can I make programs unswappable with kernel 2.6.x then?
That would mean that you cannot just
exec()another program that will also be mlockall()ed. The new program has to do that on its own...
Quoted 7 lines
mlockall
MCL_FUTUREapplies to this image only and the 2.6 behaviour is correct if less useful in some ways. It would be possible to add an inheritingMCL_flag that was Linux specific but then how do you control the depth of inheritance ? If that isn't an issue it looks the easiest.Another possibility would be
pmlockall(pid, flag), but that looks even more nasty if it races an exec.
How about clearing MCL_FUTURE on fork but allow exec to inherit it? That way a parent process could fork, mlockall in the child and exec a memlocked child. A regular fork,exec by a memlocked parent would not create a memlocked child.
Seems less messy than a new flag, while keeping the benefits.
Signature · 1 line
Zan Lynx <[email protected]>
Quoted 6 lines · Zan Lynx
How about clearing
MCL_FUTUREon fork but allow exec to inherit it? That way a parent process could fork, mlockall in the child and exec a memlocked child. A regular fork,exec by a memlocked parent would not create a memlocked child.Seems less messy than a new flag, while keeping the benefits.
The behaviour of MCL_FUTURE is standards defined so we don't get to change it. The behaviour of an added flag is up to us.