genirq: warn about IRQF_SHARED|IRQF_DISABLED at the right place
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): [email protected]
Back to LKML and Cc'ed a few more folks
Quoted 20 lines · Andrew Morton
Quoted 7 lines · On Sat, 13 Mar 2010 00:32:28 +0100 (CET) Thomas Gleixner <[email protected]>…
Quoted 5 lines · On Thu, 11 Mar 2010, [email protected] wrote:
From: Uwe Kleine-K?nig <[email protected]>
For shared irqs
IRQF_DISABLEDis only guaranteed for the first handler. So only warn starting at the second registration.Quoted 9 lines
Quoted 5 lines
The warning is moved to
__setup_irqhaving the additional benefit of catching actions registered usingsetup_irqnot onlyregister_irq.This doesn't fix the cases where setup order is wrong but it should report the broken cases more reliably.
Please drop this. It's just useless shuffling off code for no real benefit.
Quoted 6 lines
I think that warning on the second-registered IRQ was probably a mistake. Because registering just the first
IRQF_DISABLEDhandler only works by luck, and is vlunerable to someone else adding a handler later.But extending the warning so that it also covers
setup_irq()seems sensible enough.
Well, this is a crappy concept anyway and it would be way more worthwhile to figure out why we need IRQF_DISABLED at all.
An irq handler should just run always with irqs disabled. The flag is just there to deal with crappy drivers which take hundreds of microseconds or more in their interrupt service routines.
hint: IDE, USB ...
I'm simply refusing to apply patches which just shuffle code around to warn about obscure use cases instead of tackling the real problem of long running irq handlers which prevent us to get rid of IRQF_DISABLED and just run all interrupt handlers with irqs disabled. Getting rid of that would also simplify stuff which needs to be aware of reentrancy now.
Thanks,
tglx
Quoted 5 lines · On Sat, 13 Mar 2010 01:59:50 +0100 (CET) Thomas Gleixner <[email protected]>…
I'm simply refusing to apply patches which just shuffle code around to warn about obscure use cases instead of tackling the real problem of long running irq handlers which prevent us to get rid of
IRQF_DISABLEDand just run all interrupt handlers with irqs disabled.
I'm having trouble agreeing with that aim, really.
We can tweak and tune until we're blue in the face, but the system's IRQ latency will always be worse if handlers always run with interrupts disabled.
Plus this approach introduces dependencies between unrelated drivers (and between _each_ driver and the overall system) which simply don't exist if those drivers are being nicer to the rest of the system.
Quoted 3 lines · Andrew Morton
We can tweak and tune until we're blue in the face, but the system's IRQ latency will always be worse if handlers always run with interrupts disabled.
Are you sure?
If good, fast irq handler A runs with interrupts enabled, and good, fast irq handler B runs (interrupting A), A's latency goes _up_ not down.
If B happens first, then you get B's latency going up, and A's latency going down.
It's not really clear what happens to average latency. But if you know that some handlers require lower latency than others (e.g. serial ports with small FIFOs), it makes sense to prioritise them.
-- Jamie
Quoted 5 lines · On Mon, 15 Mar 2010 02:52:49 +0000 Jamie Lokier <[email protected]> wrote:
Quoted 3 lines · Andrew Morton
We can tweak and tune until we're blue in the face, but the system's IRQ latency will always be worse if handlers always run with interrupts disabled.
Quoted 7 lines
Are you sure?
If good, fast irq handler A runs with interrupts enabled, and good, fast irq handler B runs (interrupting A), A's latency goes _up_ not down.
If B happens first, then you get B's latency going up, and A's latency going down.
If the hardware doesn't prioritise these interrupts then the programmer can do so crudely, by running the critical one with interrupts disabled.
Hello,
Quoted 39 lines · Thomas Gleixner
Back to LKML and Cc'ed a few more folks
Quoted 20 lines · Andrew Morton
Quoted 7 lines · On Sat, 13 Mar 2010 00:32:28 +0100 (CET) Thomas Gleixner <[email protected]>…
Quoted 5 lines · On Thu, 11 Mar 2010, [email protected] wrote:
From: Uwe Kleine-K?nig <[email protected]>
For shared irqs
IRQF_DISABLEDis only guaranteed for the first handler. So only warn starting at the second registration.Quoted 9 lines
Quoted 5 lines
The warning is moved to
__setup_irqhaving the additional benefit of catching actions registered usingsetup_irqnot onlyregister_irq.This doesn't fix the cases where setup order is wrong but it should report the broken cases more reliably.
Please drop this. It's just useless shuffling off code for no real benefit.
Quoted 6 lines
I think that warning on the second-registered IRQ was probably a mistake. Because registering just the first
IRQF_DISABLEDhandler only works by luck, and is vlunerable to someone else adding a handler later.But extending the warning so that it also covers
setup_irq()seems sensible enough.Well, this is a crappy concept anyway and it would be way more worthwhile to figure out why we need
IRQF_DISABLEDat all.An irq handler should just run always with irqs disabled. The flag is just there to deal with crappy drivers which take hundreds of microseconds or more in their interrupt service routines.
Quoted 8 lines
hint: IDE, USB ...
I'm simply refusing to apply patches which just shuffle code around to warn about obscure use cases instead of tackling the real problem of long running irq handlers which prevent us to get rid of
IRQF_DISABLEDand just run all interrupt handlers with irqs disabled. Getting rid of that would also simplify stuff which needs to be aware of reentrancy now.
The intention for my patch is a problem that have occured at least twice in the past on the linux-arm-kernel mailing list. On at91 the timer irq is shared with some peripherals and we got reports that on these machines the timer core BUGged because irq were not disabled. Up to now the reporters didn't help to debug that to the end. So I'd consider it nice to make it easy for the next person running into that problem now that we know that there is a problem in that area.
I still think the patch is worthwhile as I think we're not getting rid of IRQF_DISABLED in the next few release cycles, are we?
Best regards
Uwe
Signature · 2 lines
Pengutronix e.K. | Uwe Kleine-König | Industrial Linux Solutions | http://www.pengutronix.de/ |