[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
flashdisk real-time behavior
- Date: Tue, 29 Jan 2008 13:16:42 -0800
- From: strauman at slac.stanford.edu (Till Straumann)
- Subject: flashdisk real-time behavior
With the new flashdisk support there is an opportunity
for me to reiterate my rant at libblock/bdbuf.
I remind you that bdbuf disables task preemption to
protect internal data structures. Most notably,
the ioctl(BLKIO_REQUEST) is executed non-preemptably.
It seems to me (I have briefly looked at the code but didn't do
an extensive analysis or run-time tests) that the flash write and erase
are essentially busy-waiting. This would mean that a task
doing I/O to a flash disk may prevent higher-priority tasks
from being scheduled for intolerably long periods.
E.g., with a block size of 512b and a programming time of ~5us/byte
task-preemption is disabled for ~2.5ms.
(I don't even want to think about foreground-erasing...)
At the time (see this thread:
my concerns were answered by asserting that ioctl(BLKIO_REQUEST)
is a very short operation. This may apply to the ATA driver but not
to the flash disk driver. Also, it is not obvious to the disk driver
that the BLKIO_REQUEST ioctl is executed with preemption disabled.
Could someone familiar with the flashdisk comment, please?
Anyways, IMHO libblock/bdbuf should be converted to using mutex protection
(I have started some work in that direction).