Re: MSI interrupts and disable_irq

[Jeff Garzik <jgarzik@pobox.com>]

Ayaz Abdulla wrote:
> I am trying to track down a forcedeth driver issue described by bug 9047 
> in bugzilla (2.6.23-rc7-git1 forcedeth w/ MCP55 oops under heavy load). 
> I added a patch to synchronize the timer handlers so that one handler 
> doesn't accidently enable the IRQ while another timer handler is running 
> (see attachment 'Add timer lock' in bug report) and for other processing 
> protection.
> 
> However, the system still had an Oops. So I added a lock around the 
> nv_rx_process_optimized() and the Oops has not happened (see attachment 
> 'New patch for locking' in bug report). This would imply a 
> synchronization issue. However, the only callers of that function are 
> the IRQ handler and the timer handlers (in non-NAPI case). The timer 
> handlers  use disable_irq so that the IRQ handler does not contend with 
> them. It looks as if disable_irq is not working properly.
> 
> This issue repros only with MSI interrupt and not legacy INTx 
> interrupts. Any ideas?

(added linux-kernel to CC, since I think it's more of a general kernel 
issue)

To be brutally frank, I always thought this disable_irq() mess was a 
hack both ugly and fragile.  This disable_irq() work that appeared in a 
couple net drivers was correct at the time, so I didn't feel I had the 
justification to reject it, but it still gave me a bad feeling.

I think the scenario you outline is an illustration of the approach's 
fragility:  disable_irq() is a heavy hammer that originated with INTx, 
and it relies on a chip-specific disable method (kernel/irq/manage.c) 
that practically guarantees behavior will vary across MSI/INTx/etc.

Practices like forcedeth's unique locking work for a time, but it should 
be a warning sign any time you stray from the normal spin_lock_irqsave() 
method of synchronization.

Based on your report, it is certainly possible that there is a problem 
with MSI's desc->chip->disable() method...  but I would actually 
recommend working around the problem by making the forcedeth locking 
more standardized by removing all those disable_irq() hacks.

Using spinlocks like other net drivers (note: avoid NETIF_F_LLTX 
drivers) has a high probability of both fixing your current problem, and 
giving forcedeth a more stable foundation for the long term.  In my 
humble opinion :)

	Jeff