Abysmal I/O scheduling with dm-crypt

[martin f krafft <madduck@madduck.net>]

[-- Attachment #1: Type: text/plain, Size: 2048 bytes --]

Dear list,

for years I have been plagued by the following problem, and it is
almost out of last resort that I am turning to you. I have searched
the web over the months. There is talk of dirty_ratios, of caches,
and of write throttling, but I have not been able to find a way to
fix the problem.

I am using encrypted filesystems (dm-crypt) and the 3.0.0 kernel.
Underneath might be a RAID1 or a fast SSD. On top is usually LVM
with a few LVs holding the system.

Whenever an I/O-intensive task starts, such as:

  - tar -c or -x
  - dd
  - rsync
  - notmuch
  - …

the system becomes unusable for several seconds at a time, at least
once or twice per minute. What seems to happen is that Vim or the
Shell or Firefox or whatever else completely blocks, waiting for
I/O, but Linux is not satisfying those I/O requests because it's
busy servicing the aforementioned I/O-intensive task. As a result,
Vim or the Shell or Firefox or whatever do not update, forcing me to
wait for them to get an I/O service slot.

People not running dm-crypt seem unable to reproduce this problem,
making me think that it must be due to dm-crypt, and I wouldn't find
it hard to imagine, because dm-crypt basically shields the
I/O-scheduler of the kernel, doesn't it? Worse, it probably doesn't
make any effort of scheduling I/O itself. Note: I know very little
about the internals, so please correct me if I am wrong.

I am wondering if this is the case, and if so, what could be done
about it.

Do you have any ideas and/or concrete suggestions?

PS: http://bugs.debian.org/635768 was the latest incarnation of my
    attempts to solve this problem (I accidentally wrote notmuch
    when I mean to write awesome in that report…).

-- 
martin | http://madduck.net/ | http://two.sentenc.es/
 
"gilmour's guitar sounds good
 whether you've got a bottle of cider in your hand
 or a keyboard and a mouse."
                                                -- prof. bruce maxwell
 
spamtraps: madduck.bogus@madduck.net

[-- Attachment #2: Digital signature (see http://martin-krafft.net/gpg/sig-policy/999bbcc4/current) --]
[-- Type: application/pgp-signature, Size: 1124 bytes --]