random AIO reads (do_generic_mapping_read)

["Michael R. Hines" <mhines@cs.binghamton.edu>]

[-- Attachment #1: Type: text/plain, Size: 1514 bytes --]

Howdy (please CC replies directly),

I have a very specific question regarding the behavior of
do_generic_mapping_read().

Here's my 2.6.20 setup: I have a stable block device driver that has
random-access style latencies (around 80 microseconds. zero seek-time).

From userland: I have a simple AIO io_submit/io_suspend
program that repeatedly fires off *hundreds random AIO reads*

Here's what I would like to happen:

Since my driver does not seek (i.e. I/O not helped by mm/readahead.c)
I would like my AIOs to be *overlapped* when sent to my driver.
Currently, do_generic_mapping_read() will block on every read,
one-by-one. Instead I want the bio's to be fired off to my driver as
fast as possible without calling lock_page(). Rather the read should
return and allow io_submit_one to fire off more reads without waiting
for the userspace-copy - it would instead happen later in the future.

Furthermore: I *do not* want to avoid the page cache (which temporarily
solves the problem). I still want the support of the page-cache, I just
instead need those I/Os to not block or sleep while waiting for a single
I/O to complete.

Obviously if this were a disk, such a requirement would be very stupid
(and useless). But without it, my low-latency driver is under-utilized.

Is this possible right now?

-- 
/**************************************/
Michael R. Hines
Grad Student, SUNY Binghamton
http://www.cs.binghamton.edu/~mhines
/ *************************************/


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 189 bytes --]