Hi Christoph,
While testing Jens' loop-4 patch (and not being able to find
any way to lock it up), I stumbled onto a strange behavior.
I set up an interleaved swap with one swap partition, and one
swapfile in a loopback mounted reiserfs - populated tmpfs with
a kernel tree and did hefty make -j kernel builds to generate
much I/O. Afterward (bored), I figured I'd bounce the data in
tmpfs back and forth between swap containers with swapoff. It
took much longer than I expected.
This is likely only something a bored tester would do, but it
seems odd enough to report.. repeated without loop.
[root]:# pwd
/var/tmp/linux-2.4.2-pre1
[root]:# df
Filesystem 1k-blocks Used Available Use% Mounted on
/dev/hda6 4157873 3734124 213701 95% /
shm 499092 0 499092 0% /dev/shm
/dev/hda7 256592 155715 87625 64% /var
tmpfs 622228 123136 499092 20% /var/tmp
/dev/hda8 10445318 6450096 3669907 64% /usr/local
/dev/hda1 4088532 2286968 1801564 56% /dos_c
[root]:# mount
/dev/hda6 on / type ext2 (rw)
proc on /proc type proc (rw)
shm on /dev/shm type shm (rw)
/dev/hda7 on /var type ext2 (rw)
tmpfs on /var/tmp type tmpfs (rw)
/dev/hda8 on /usr/local type ext2 (rw)
/dev/hda1 on /dos_c type vfat (rw,noexec)
devpts on /dev/pts type devpts (rw,gid=5,mode=0620)
[root]:# swapon -s
Filename Type Size Used Priority
/dev/hda2 partition 265064 127356 2
/dev/hda5 partition 265032 0 2
[root]:# time swapoff /dev/hda2
real 7m17.555s
user 0m0.010s
sys 7m3.040s
[root]:#
Hi Mike,
On Thu, 8 Feb 2001, Mike Galbraith wrote:
> Hi Christoph,
>
> While testing Jens' loop-4 patch (and not being able to find
> any way to lock it up), I stumbled onto a strange behavior.
>
> I set up an interleaved swap with one swap partition, and one
> swapfile in a loopback mounted reiserfs - populated tmpfs with
> a kernel tree and did hefty make -j kernel builds to generate
> much I/O. Afterward (bored), I figured I'd bounce the data in
> tmpfs back and forth between swap containers with swapoff. It
> took much longer than I expected.
Oh, the swapoff handling in Linux is much less then suboptimal. So I
would expect that.
To explain: For every single page on swap we scan all processes vmas
and all shm/tmpfs objects swap tables :-( And we have to hold some
locks for that...
Greetings
Christoph