Mark Lord wrote:
Jeff Garzik wrote:
This has been bubbling on my brain for a while. I blathered on about
this on IRC to Tejun, but figured I might as well post it here and get
it archived.
In general, I think we should adopt a flexible or "loose" model for
acking interrupts on SFF controllers.
(a) whenever we are in bus-idle (qc == NULL), and get an interrupt, go
ahead and read Status.
(b) if we are expecting an interrupt, and receive one, check Status
(or AltStatus if DMAing).
(c) if condition "(b)" indicates busy, initiate status polling every
250ms until timeout occurs or BSY clears.
(d) if N seconds (4?) elapses without an interrupt, initiate polling.
keep a history of such "fail-over" events, and note each fail-over'd
command's eventual success via polling, success via interrupt, or
timeout. Use that history to decide to switch to 100% polling mode
(i.e. reach conclusion that interrupt delivery is broken, via
observation)
That should cover no-interrupts, lost interrupts, early interrupts,
screaming interrupts, insane devices, and of course normal operation.
The model could be summarized as "interrupt as a hint" operation.
..
The only question is, under which conditions do we return IRQ "handled=1",
and which times should we return 0 ?
Definitely when a real IRQ wakes us up and we see (qc != NULL &&
drive_ready),
essentially exactly as we currently do it.
But things might be trickier once polling is introduced, unless we also
mask
the device interrupt before initiating the polling.
Actually no, and that is a key benefit of this scheme: if we ensure
that the polling paths are resilient even in case where interrupts are
being delivered -- as we must do anyway -- then we don't have to worry
about interrupt masking, either on the interrupt controller or on the
device[1].
If we do get an interrupt, ack it ASAP. That covers normal operation
and screaming interrupts.
If we don't get an interrupt, we will notice after a spell and poll
Status to ensure progress occurs.
Note that this polling is a different sort of polling than running an
entire ATA command via a kernel thread. In this case, we're talking
about periodic Status (or AltStatus or LLD-specific-register status)
polling only.
A lot of fiddling with irq masking is getting around ugliness that I am
instead trying to eliminate altogether. A truly robust system follows
the spec WRT nIEN and other interrupt masking..... but then prepares
for the case where hw decides to send an interrupt anyway.
On SFF controllers, we should consider interrupts to be unreliable
messages delivered on a best effort basis by hardware. If we get them,
great, ack and act. If we lack them, make sure progress occurs.
Regards,
Jeff
[1] well, there -are- exceptions, such as when we are bitbanging the ATA
Data register
-
To unsubscribe from this list: send the line "unsubscribe linux-ide" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
