Re: Potential race condition in drivers/ata/sata_mv.ko

[Pavel Andrianov]

Hi!

I have found such example:

... ->

ata_exec_internal_sg ->

    ata_qc_issue ->

        mv_qc_issue ->

            mv_clear_and_enable_port_irqs ->

                mv_enable_port_irqs ->

                    mv_set_main_irq_mask


ata_exec_internal_sg acquires spin_lock(ap->lock) and call of the last 
function mv_set_main_irq_mask is with this lock. mv_interrupt acquires 
spin_lock(host->lock) before call of the same function. I am not sure is 
it correct to add one more spin_lock or move a call of request_irq in 
ata_host_activate, thus I can not easily fix the issue.

One more question is related to ata_exec_internal_sg. In comments there 
is an information the function is called without locking. However, 
ata_exec_internal_sg calls ata_eh_release before ata_eh_acquire (lines 
1650, 1655).There is a block of code under spinlock and eh context can 
not be acquired there. The comment may be wrong and eh_context is 
acquired somewhere before, but I also can not find it. Do you know where 
is the initial acquire of eh_context in this case?


10.08.2016 06:51, Tejun Heo ÐÐÑÐÑ:
> Hello,
>
> On Fri, Aug 05, 2016 at 03:43:30PM +0300, Pavel Andrianov wrote:
> > In drivers/ata/sata_mv.ko function mv_set_main_irq_mask is called several
> > times. Twice with a spinlock, twice from init function and once without any
> > protection. The call without protection rises to several handlers from
> > ata_port_operations. The structure with the ata_port_operations is included
> > into a structure 'host' in mv_platform_probe and in mv_pci_init_one. At the
> > end of these functions ata_host operations are activated together with
> > interrupt handler. The conclusion is: interrupt handler may be executed in
> > parallel with handlers from ata_port_operations, or, more formally, it may
> > interrupt its execution.
> >
> > In mv_set_main_irq_mask and in interrupt handler mv_interrupt the interrupt
> > mask is modified, but, as I said, handlers from ata_port_operations do not
> > acquire any lock. Thus, the interrupt mask may be set incorrectly if the are
> > two conflicting modifications.
> It depends on which operations.  Most are only called from EH context
> and racing there isn't likely to cause any actual issues.  Care to
> submit a patch to fix the issue?
>
> Thanks.

--
Pavel Andrianov
Linux Verification Center, ISPRAS
web: http://linuxtesting.org
e-mail: andrianov@xxxxxxxxx