VM breakdown, 2.4.0 family

Since the testN series and up through ac12, I experience total loss of control when memory is nearly exhausted.

I start with 256M and eat it up with programs until there is only about 7 megs left, no swap. From that point all user processes stall and the disk begins to grind nonstop. It will continue to grind for about 25-30 minutes until it goes completely silent. No processes get killed, no VM messages are emitted.

The only recourse is the magic key. If I reboot before the disk goes silent I can cleanly kill X with sysrq-E and restart.

If I wait until it goes silent, all is lost. I have to sysrq-SUB.

David Ford Wrote:

Quoted 9 lines

Since the testN series and up through ac12, I experience total loss of control when memory is nearly exhausted.

I start with 256M and eat it up with programs until there is only about 7 megs left, no swap. From that point all user processes stall and the disk begins to grind nonstop. It will continue to grind for about 25-30 minutes until it goes completely silent. No processes get killed, no VM messages are emitted.

Quoted 4 lines

The only recourse is the magic key. If I reboot before the disk goes silent I can cleanly kill X with sysrq-E and restart.

If I wait until it goes silent, all is lost. I have to sysrq-SUB.

You might want to try:

http://bazar.conectiva.com.br/~marcelo/patches/v2.4/2.4.1pre10/bg_page_aging.patch

or

ftp://ftp.cam.org/users/tomlins/pte_aging_limit_swaps.diff

The first patch from Marcelo fixes a problem with aging the wrong pages. The second patch is sort of a 'best of Marcelo' patch. It contains the aging fix and adds conditional bg pte aging (if with activate fast than we age down...). It also has code to trottle swapouts when under preasure - it only swaps out as much as we need now.

I have fives days of uptime with it here (on test9 and test10).

Feedback Welcome,

Ed Tomlinson <[email protected]>

I have Marcelo's patch. It isn't applicable because I am purposely not enabling any swap. The problem is the system gets down to about 7 megs of buffers free and within three seconds has become functionally dead. Zero response on any user input/output device save the magic key.

The system will then grind the harddrive solid for about 25-30 minutes then everything will go silent.

The brokenness is that the OOM code never activates.