Re: Linux 2.4.10: printk() deadlocks

[Andrew Morton]

[e-mail] wrote:
> 
> Hi, andrew,
> 
> currently I'm porting our WORM filesystem (OWFS) from kernel
> 2.4.0 to kernel 2.4.10. I'm used to rely on printk() to be callable
> from every place in the kernel, but now I experience deadlocks
> when calling it from some places within our filesystem.
> (I'm writing a ~600 lines in 2 seconds and I enlarged the buffer size
> to 1MB, what worked perfectly in the past.)
> The call-chain is:
> ...

But it isn't deadlocked.  The EIP points at this statement

                spin_unlock_irqrestore(&logbuf_lock, flags);
-->             console_may_schedule = 0;
                release_console_sem();

in printk().  So the output is about to come out to
the console device(s).

> A diff between printk.c of the two kernels shows:
>  * Rewrote bits to get rid of console_lock
>  *      01Mar01 Andrew Morton <[e-mail]>
> 
> Please could you tell me a workaround or some possibility to get out
> of this situation? I think printk() should be an absolutely reliable
> function which never schedules.

printk() doesn't schedule.  However, userspace processes which
write to /dev/ttyX *do* schedule, and the two paths use the same
code and have to lock against each other, etc.  That's why it's all
rather unclear.

Sorry, but I think you must be doing something strange to make this
happen - can you please diagnose a little further?  Investigate
further with kdb?  Can you send me the wherewithals to reproduce
it?  Are you running SMP?

-