Re: [2.4 patch] Port of adutux driver from 2.6 kernel to 2.4.

[Pete Zaitcev <zaitcev@redhat.com>]

On Fri, 19 Oct 2007 20:40:35 +0300, Vitaliy Ivanov <vitalivanov@gmail.com> wrote:

Hi, Vitaly, I added you on cc: for the 2.6 cleanup. Please double-check
what I'm doing there and use it for your 2.4 version. I hope my intentions
get clearer with an example. Now, about the specific question:

> Static lock minor_table_mutex is used for minor table structure.
> And dev->sem for dev manipulations and that's why for open_count.
> If you will simply browse /drivers/usb dir for 2.4 you will see that
> such approach is widely used there.
> What's not right?

The fundamental reason why you cannot lock a free-able structure with
an in-structure lock is this. Imagine thread A locks in order to process
a disconnect. Thread B wants to open and waits for the lock. Notice that
the struct is not open, so thread A frees it. At this point, thread B
is using a freed memory.

The solution is to lock the instance struct dev with dev->mtx, except
for the open count, which is locked by a static lock (I'm ignoring
interrupts here, which cannot use mutexes).

I'm sorry to say, you're quite right: a number of drivers in 2.4 got
it wrong, and some (like adutux) carried it through 2.6.23.

-- Pete