From: Jouni Malinen (jkmaline_at_cc.hut.fi)
Date: 2002-07-24 22:06:24 UTC
At least some of the problems causing total hang on SMP hosts has now
been fixed. I was unable to hang my test setup with yesterday's version.
There were some further bugs that caused card resets, but I think these
are now fixed. I have been unable to crash either the card or the host
computer when flooding the system with few stations and performing
management operations at the same time. I would appreciate if those who
have seen crashes or kernel hangs on SMP platform with the older
versions, could test this new version (i.e., the latest CVS snapshot
from http://hostap.epitest.fi/).
I'm not 100% sure what was causing the hang, but most probably one reason was in AP code sending management frames without using the same serialization mechanism as the data frames. Other reason could well have been in complex code used with command completion events only for transmit commands.
I changed the management frames to use netdevice queue just like the data packets. This will guarantee that the frames will be subject to same restrictions than other packets (i.e., they do not perform BAP transfers or transmit commands within a data frame transmit). This made prism2_send_mgmt() function a much cleaner since it can now use the code of prism2_tx() for TX and host encryption. It should also be noted, that from now on, the net device has to be UP also for management frames (i.e., AP will see auth/assoc messages even with device down, but it will be able to reply to them only after the device has been set up).
I did not get any TX timeouts during flood tests, but noticed some problems with manually forced card resets. It's possible that the driver is not able to reset the card in all situations, but at least this did not hang the host computer. Anyway, this is not much of a problem, if the card does not end up in state requiring resetting.
-- Jouni Malinen PGP id EFC895FA