David Miller wrote:
From: Jason Wang <jason77.wang@xxxxxxxxx>
Date: Thu, 23 Dec 2010 18:43:13 +0800
When kernel enters irqhanlder, it will check the Rx interrupt status
bit, if Rx status is set but can't call napi_schedule(), it will do
nothing and directly return form irqhandler. This situation is prone
to be produced when we repeatly call irqhandler through netpoll
interface(i.e kgdboe connecting).
This is a potential risk for those level triggered platforms(i.e
ti_omap3evm), because if we don't handle Rx int and just return from
irqhandler, the irq pin will be keeping asserted, the level triggered
platforms will have no chance to jump out from the Rx irq. The whole
system will hung into the irq subsystem.
To solve it, we add a disable/re-enable Rx int operation for this
situation, this operation can de-assert interrupt pin for this time
and will leave the received data and status in the FIFO for later
interrupts to handle.
Signed-off-by: Jason Wang <jason77.wang@xxxxxxxxx>
Hi David,
Thanks for your comments.
You absolutely cannot do this.
You now can race with the NAPI completion code turning the RX
interrupts back on, and you'll leave the chip with RX interrupts
disabled.
I think my modification almost have the same execution path as the
original design and don't produce the race condition with NAPI threads.
The original design is:
If (can call napi_schedule) {
disable rx int;
call napi_schedule()
} else {
keep rx int enabled;
return;
}
my modification is:
disable rx int;
if (can call napi_schedule) {
call napi_schedule();
} else {
re-enable rx int; // this will de-assert interrupt pin for this time
return;
}
So my modification is: if we can call napi_schedule(), we will disable
the rx int until the NAPI thread re-enable it. if we can't call
napi_schedule(), the rx int will keep enabled. This logic is almost same
as the original design. I can't figure out why original design is safe
while my modification is risky.
Add more info:
the Freescale imx31pdk, imx51pdk and ti_omap3evm boards all use this
driver, before apply this modification, their kgdboe connecting is not
stable. After applied this patch, their nfs root is as good as before
and their kgdboe connecting is stable.
Thanks,
Jason.
You must solve your level triggered interrupt some other way, every
NAPI based device must manage the interrupt disabling carefully and
only when the napi POLL is successfully scheduled in order to avoid
races.
And especially you must not make a crazy hack like this for obscure
things like kgdboe.
--
To unsubscribe from this list: send the line "unsubscribe linux-omap" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
