VM in 2.2.17pre9 : a success ?

[Willy Tarreau]

Hi Andrea, Rik and Marcello

I've been stressing the VM in 2.2.17pre9 with lots of malloc and mmap002 for
about 15 minutes now, and it seems solid.

Under 2.2.17pre7, I got a killed mmap002 once, and sometimes a complete
freeze (at least, SysRQ worked, and there were zillions of messages about
try_to_free_pages). But now, the first thing I can notice is that the system
keeps responsive during high swapping activity. Moreover, malloc correctly
terminates with a NULL and mmap002 terminates normally. Not any kill till now.
The disk activity in the mmaped files and the swap seems more regular than
before.
The hard disk sounds just as if it was running windoze, but I still can log into
the system. For info, this machine has 128 MB RAM and 136 MB swap. It's about
1 minute now that the swap is completely full (0 byte free), but still works.

Aie, it seems I wrote this mail too quickly  :-(
it agreed to work during about 1 minute without any RAM nor swap free, but
suddenly stuck. I've just caught the last messages which complain about
do_try_to_free_pages failing for malloc (a process which does while(malloc)).
Anyway, they're old since this one ended correcly a few minutes ago. But
nothing about mmap002.

Perhaps we should agree to kill some tasks when there are no ram nor free swap.
What I wonder, however, is why does the mmap consume so much memory. This will
seem like a naive question to you, but wouldn't it be possible, in case of OOM,
to shrink down the mmaped areas to a few pages only ?

I will retest this after applying some of Andrea's patches and will keep you
informed ASAP.

Willy