Slow SD card and USB HDD I/O for block sizes bigger than a couple of bytes

[Michael Sparmann]

I'm observing weird I/O behavior with natty's 2.6.38-1209-omap4 kernel running on a pandaboard:
When reading from a USB HDD, the relationship between read block sizes and throughput looks like this: http://paste.ubuntu.com/604616/
(the buffer cache has been flushed with hdparm -f between every measurement)
This HDD gets about 28MB/s on windows, and about 32MB/s without this weird foldback on my natty x86 desktop: http://paste.ubuntu.com/604619/

The same (but to a lesser extent) happens with slower USB storage devices as well: http://paste.ubuntu.com/604613/ (this thumbdrive gets 5MB/s on windows).
This also appears to affect SD card access, but also to a lesser extent: http://paste.ubuntu.com/604605/

As the maximum throughput (which is still way too slow) is reached at around 128 bytes block size, which is way below sector size, and as this happens for SD card access as well, this suggests that the problem is coming from some more generic code than the MMC or USB controller driver.

ProblemType: Bug
DistroRelease: Ubuntu 11.04
Package: linux-image-2.6.38-1209-omap4 2.6.38-1209.13
ProcVersionSignature: Ubuntu 2.6.38-1209.13-omap4 2.6.38.4
Uname: Linux 2.6.38-1209-omap4 armv7l
Architecture: armel
Date: Tue May 24 00:11:32 2011
ProcEnviron:
 LANGUAGE=de_DE:en
 LANG=de_DE.UTF-8
 SHELL=/bin/bash
SourcePackage: linux-ti-omap4
UpgradeStatus: No upgrade log present (probably fresh install)

[Jani Monoses (jani)]

Brian, does the ping workaround described here affect your measurements?
https://bugs.launchpad.net/linux-linaro/+bug/709245