> The following patch will:
>
> - avoid GFP_ATOMIC in grow_buffers(). Instead, we attempt to get
> clean buffers by sleeping and by waiting for kflushd() to flush
> dirty buffers.
[...]
I did some tests on 2.0.31-pre6 without and with your patch. Very promising.
Setup: 64 MB RAM, 130 MB swap, SCSI AIC-7880, Dual P5-200
WITHOUT: bonnie -s 800 or bonnie -s 150 & bonnie -s 800 -
lots of "can not get a free page", "can not allocate skb" (already
reported on linux-kernel)
other tasks almost blocked, completed successfully
booting with "mem=4M single", swap active
screen (open two shells)
mke2fs /dev/hbd1 (500 MB IDE partition) - successfull, but got "can not get
a free page" twice
after that:
badblocks -wv /dev/hdb1 503896
after starting writing: "CNGaFP" many times
then "bus error" and:
swap_out: physical ram 4096 KB min pages 96 KB
swap_out: free pages 44 KB async pages 0 KB
repeated badblocks command:
the same but "out of memory for screen\n out of memory for bash"
instead of "bus error", swap_out twice with free pages:
70 KB and 92 KB, all shells killed
Ctrl-Alt-Del worked properly and run shutdown scripts.
WITH your patch:
64 MB: (simultaneous bonnie - OK, speed same as without patch +-10%)
4 MB:
repeated the same, not a single error message
even bonnie -s 150 worked, aprox. 20% slower than with 64 MB RAM
only 5-10 s waiting when typing "free" or "w" in the other screen window
during "badblocks"
BTW. I did "umount /" before exiting single mode. No error, but Ctrl-Alt-Del
caused "INIT: PANIC: segmentation violation". I still had bash, but I had to
press RESET. On reboot / was clean, but other two SCSI partitions were
fsck-ed (in single mode all partitions were mounted RW). I have init 2.64.
I am now going to make some speed measurements (make -j zImage running
with/without bonnie) with/without your patch.
Greetings,
-- Tomasz Motylewski