Re: rl(4) is not ready for mpsafenet net enough? (silent reboots)

On Thu, 21 Jul 2005, Andrey Chernov wrote:

> On Thu, Jul 21, 2005 at 09:18:44AM -0600, M. Warner Losh wrote:
>> *DRIVER* concept, not a hardware one in this context).  The problem
>> with REV C and earlier is that their performance sucks because the DMA
>> engine used in the cards was LAME.  Realtec fixed this in newer
>> revisions of the chip, and the re driver is able to take advantage of
>> that (as well as support the newer gige chips).
>
> The problem is that it is on-board embedded chip I try to use. Looks 
> like I need to buy some external card like "Intel PRO/100+ PCI" and turn 
> this one off...
>
> The question remains: why rl(4) driver pretend to be mpsafenet itself? 
> Other drivers with problems, like de(4), are GIANT-locked by default.

Given that this is the first report of such a feature that I've seen, it's 
likely that it's marked as MPSAFE because it appeared to the author to be 
MPSAFE.  You'll notice that if_de is marked as requiring Giant because it 
doesn't have any locking.  Although actually, John Baldwin spent the last 
day or so fixing if_de, so presumably that will go into the tree shortly 
and that will cease to be the case.

FYI, most locking bugs in network interfaces that I've seen don't result 
in a spontaneous reboot, and that's a somewhat worrying symptom.  Is this 
something you can easily reproduce in a short period of time, or in 
particular, using a particular program or system call?  Is there any 
chance your box has firewire and you can use a firewire debugger to 
inspect memory?

Robert N M Watson