Skip to contentSkip to conversation

Dangerous devm_request_irq() conversions

DISCUSSION Dmitry Torokhov 3 participants 6 messages Feb 22, 2013

lore
Share & tools
lore
#0 lore raw

Hi,

It looks like a whole slew of devm_request_irq() conversions just got applied to mainline and many of them are quite broken.

Consider fd5231ce336e038037b4f0190a6838bdd6e17c6d or c1879fe80c61f3be6f2ddb82509c2e7f92a484fe: the drivers udsed first to free IRQ and then unregister the corresponding device ensuring that IRQ handler, while it runs, has the device available. The mechanic conversion to devm_request_irq() reverses the order of these operations opening the race window where IRQ can reference device (or other resource) that is already gone.

It would be nice if these could be reverted and revioewed again for correctness.

In general any conversion to devm_request_irq() needs double and triple checking.

Thanks.

Signature · 1 line
Dmitry
#1 lore raw
Quoted 14 lines · Dmitry Torokhov

Hi,

It looks like a whole slew of devm_request_irq() conversions just got applied to mainline and many of them are quite broken.

Consider fd5231ce336e038037b4f0190a6838bdd6e17c6d or c1879fe80c61f3be6f2ddb82509c2e7f92a484fe: the drivers udsed first to free IRQ and then unregister the corresponding device ensuring that IRQ handler, while it runs, has the device available. The mechanic conversion to devm_request_irq() reverses the order of these operations opening the race window where IRQ can reference device (or other resource) that is already gone.

Quoted 2 lines

It would be nice if these could be reverted and revioewed again for correctness.

Um, other RTC drivers already have been using devm_request_threaded_irq() or devm_request_irq() like this, before I added these patches.

For example,
./drivers/rtc/rtc-tegra.c
./drivers/rtc/rtc-spear.c
./drivers/rtc/rtc-s3c.c
./drivers/rtc/rtc-mxc.c
./drivers/rtc/rtc-ds1553.c
./drivers/rtc/rtc-ds1511.c
./drivers/rtc/rtc-snvs.c
./drivers/rtc/rtc-imxdi.c
./drivers/rtc/rtc-tx4939.c
./drivers/rtc/rtc-mv.c
./drivers/rtc/rtc-coh901331.c
./drivers/rtc/rtc-stk17ta8.c
./drivers/rtc/rtc-lpc32xx.c
./drivers/rtc/rtc-tps65910.c
./drivers/rtc/rtc-rc5t583.c

Also, even more, some RTC drivers calls rtc_device_unregister() first, then calls free_irq() later.

For example,
./drivers/rtc/rtc-vr41xx.c
./drivers/rtc/rtc-da9052.c
./drivers/rtc/rtc-isl1208.c
./drivers/rtc/rtc-88pm860x.c
./drivers/rtc/rtc-tps6586x.c
./drivers/rtc/rtc-mpc5121.c
./drivers/rtc/rtc-m48t59.c

Please, don't argue revert without concrete reasons.

If these devm_request_threaded_irq() or devm_request_irq() make the problem, devm_free_irq() will be added later.

Quoted 8 lines

In general any conversion to devm_request_irq() needs double and triple checking.

Thanks.

--
Dmitry

#2 lore raw
Quoted 52 lines · Jingoo Han
Quoted 14 lines · Dmitry Torokhov

Hi,

It looks like a whole slew of devm_request_irq() conversions just got applied to mainline and many of them are quite broken.

Consider fd5231ce336e038037b4f0190a6838bdd6e17c6d or c1879fe80c61f3be6f2ddb82509c2e7f92a484fe: the drivers udsed first to free IRQ and then unregister the corresponding device ensuring that IRQ handler, while it runs, has the device available. The mechanic conversion to devm_request_irq() reverses the order of these operations opening the race window where IRQ can reference device (or other resource) that is already gone.

Quoted 2 lines

It would be nice if these could be reverted and revioewed again for correctness.

Um, other RTC drivers already have been using devm_request_threaded_irq() or devm_request_irq() like this, before I added these patches.

For example,
./drivers/rtc/rtc-tegra.c
./drivers/rtc/rtc-spear.c
./drivers/rtc/rtc-s3c.c
./drivers/rtc/rtc-mxc.c
./drivers/rtc/rtc-ds1553.c
./drivers/rtc/rtc-ds1511.c
./drivers/rtc/rtc-snvs.c
./drivers/rtc/rtc-imxdi.c
./drivers/rtc/rtc-tx4939.c
./drivers/rtc/rtc-mv.c
./drivers/rtc/rtc-coh901331.c
./drivers/rtc/rtc-stk17ta8.c
./drivers/rtc/rtc-lpc32xx.c
./drivers/rtc/rtc-tps65910.c
./drivers/rtc/rtc-rc5t583.c

Also, even more, some RTC drivers calls rtc_device_unregister() first, then calls free_irq() later.

For example,
./drivers/rtc/rtc-vr41xx.c
./drivers/rtc/rtc-da9052.c
./drivers/rtc/rtc-isl1208.c
./drivers/rtc/rtc-88pm860x.c
./drivers/rtc/rtc-tps6586x.c
./drivers/rtc/rtc-mpc5121.c
./drivers/rtc/rtc-m48t59.c

Quoted 1 line

Please, don't argue revert without concrete reasons.

What more concrete reason do you need? I explained to you the exact reason on the patches I noticed before and also on the 2 commits referenced above: blind conversion to devm_* changes order of operation which may be deadly with IRQs (but others, like clocks and regulators, are important too).

The fact that crap slipped in the kernel before is not the valid reason for adding more of the same crap.

Please *understand* APIs you are using before making changes.

Quoted 3 lines

If these devm_request_threaded_irq() or devm_request_irq() make the problem, devm_free_irq() will be added later.

And the point? If you use devm_request_irq() and then call devm_free_irq() manually in all paths what you achieved is waste of memory required for devm_* tracking.

Signature · 1 line
Dmitry
#3 lore raw
Quoted 67 lines · Dmitry Torokhov
Quoted 52 lines · Jingoo Han
Quoted 14 lines · Dmitry Torokhov

Hi,

It looks like a whole slew of devm_request_irq() conversions just got applied to mainline and many of them are quite broken.

Consider fd5231ce336e038037b4f0190a6838bdd6e17c6d or c1879fe80c61f3be6f2ddb82509c2e7f92a484fe: the drivers udsed first to free IRQ and then unregister the corresponding device ensuring that IRQ handler, while it runs, has the device available. The mechanic conversion to devm_request_irq() reverses the order of these operations opening the race window where IRQ can reference device (or other resource) that is already gone.

Quoted 2 lines

It would be nice if these could be reverted and revioewed again for correctness.

Um, other RTC drivers already have been using devm_request_threaded_irq() or devm_request_irq() like this, before I added these patches.

For example,
./drivers/rtc/rtc-tegra.c
./drivers/rtc/rtc-spear.c
./drivers/rtc/rtc-s3c.c
./drivers/rtc/rtc-mxc.c
./drivers/rtc/rtc-ds1553.c
./drivers/rtc/rtc-ds1511.c
./drivers/rtc/rtc-snvs.c
./drivers/rtc/rtc-imxdi.c
./drivers/rtc/rtc-tx4939.c
./drivers/rtc/rtc-mv.c
./drivers/rtc/rtc-coh901331.c
./drivers/rtc/rtc-stk17ta8.c
./drivers/rtc/rtc-lpc32xx.c
./drivers/rtc/rtc-tps65910.c
./drivers/rtc/rtc-rc5t583.c

Also, even more, some RTC drivers calls rtc_device_unregister() first, then calls free_irq() later.

For example,
./drivers/rtc/rtc-vr41xx.c
./drivers/rtc/rtc-da9052.c
./drivers/rtc/rtc-isl1208.c
./drivers/rtc/rtc-88pm860x.c
./drivers/rtc/rtc-tps6586x.c
./drivers/rtc/rtc-mpc5121.c
./drivers/rtc/rtc-m48t59.c

Quoted 1 line

Please, don't argue revert without concrete reasons.

What more concrete reason do you need? I explained to you the exact reason on the patches I noticed before and also on the 2 commits referenced above: blind conversion to devm_* changes order of operation which may be deadly with IRQs (but others, like clocks and regulators, are important too).

The fact that crap slipped in the kernel before is not the valid reason for adding more of the same crap.

Please *understand* APIs you are using before making changes.

Quoted 1 line
Quoted 6 lines
Quoted 2 lines

If these devm_request_threaded_irq() or devm_request_irq() make the problem, devm_free_irq() will be added later.

And the point? If you use devm_request_irq() and then call devm_free_irq() manually in all paths what you achieved is waste of memory required for devm_* tracking.

CC'ed Al Viro, Tejun Heo

So, is there any report that the devm_request_threaded_irq() makes the deadly problem related IRQ in such cases?

According to your comment, it seems that there is no reason to use devm_request_irq() or devm_request_threaded_irq().

Please, argue that it would be better to deprecate devm_request_irq() or devm_request_threaded_irq().

Quoted 3 lines

--
Dmitry

#4 lore raw
Quoted 75 lines · Jingoo Han
Quoted 67 lines · Dmitry Torokhov
Quoted 52 lines · Jingoo Han
Quoted 14 lines · Dmitry Torokhov

Hi,

It looks like a whole slew of devm_request_irq() conversions just got applied to mainline and many of them are quite broken.

Consider fd5231ce336e038037b4f0190a6838bdd6e17c6d or c1879fe80c61f3be6f2ddb82509c2e7f92a484fe: the drivers udsed first to free IRQ and then unregister the corresponding device ensuring that IRQ handler, while it runs, has the device available. The mechanic conversion to devm_request_irq() reverses the order of these operations opening the race window where IRQ can reference device (or other resource) that is already gone.

Quoted 2 lines

It would be nice if these could be reverted and revioewed again for correctness.

Um, other RTC drivers already have been using devm_request_threaded_irq() or devm_request_irq() like this, before I added these patches.

For example,
./drivers/rtc/rtc-tegra.c
./drivers/rtc/rtc-spear.c
./drivers/rtc/rtc-s3c.c
./drivers/rtc/rtc-mxc.c
./drivers/rtc/rtc-ds1553.c
./drivers/rtc/rtc-ds1511.c
./drivers/rtc/rtc-snvs.c
./drivers/rtc/rtc-imxdi.c
./drivers/rtc/rtc-tx4939.c
./drivers/rtc/rtc-mv.c
./drivers/rtc/rtc-coh901331.c
./drivers/rtc/rtc-stk17ta8.c
./drivers/rtc/rtc-lpc32xx.c
./drivers/rtc/rtc-tps65910.c
./drivers/rtc/rtc-rc5t583.c

Also, even more, some RTC drivers calls rtc_device_unregister() first, then calls free_irq() later.

For example,
./drivers/rtc/rtc-vr41xx.c
./drivers/rtc/rtc-da9052.c
./drivers/rtc/rtc-isl1208.c
./drivers/rtc/rtc-88pm860x.c
./drivers/rtc/rtc-tps6586x.c
./drivers/rtc/rtc-mpc5121.c
./drivers/rtc/rtc-m48t59.c

Quoted 1 line

Please, don't argue revert without concrete reasons.

What more concrete reason do you need? I explained to you the exact reason on the patches I noticed before and also on the 2 commits referenced above: blind conversion to devm_* changes order of operation which may be deadly with IRQs (but others, like clocks and regulators, are important too).

The fact that crap slipped in the kernel before is not the valid reason for adding more of the same crap.

Please *understand* APIs you are using before making changes.

Quoted 1 line
Quoted 6 lines
Quoted 2 lines

If these devm_request_threaded_irq() or devm_request_irq() make the problem, devm_free_irq() will be added later.

And the point? If you use devm_request_irq() and then call devm_free_irq() manually in all paths what you achieved is waste of memory required for devm_* tracking.

Quoted 8 lines

CC'ed Al Viro, Tejun Heo

So, is there any report that the devm_request_threaded_irq() makes the deadly problem related IRQ in such cases?

According to your comment, it seems that there is no reason to use devm_request_irq() or devm_request_threaded_irq().

devm_request_irq() or devm_request_threaded_irq() are OK if:

1. _ALL_ resources in the driver are controlled by devm_* so that order of freeing is not disturbed, or

2. You can shut off interrupts in the chip so that interrupts will not be generated even though IRQ handler is installed, or

3. Some combination of above.

Note that it is not only interrupts, other resources like clocks and regulators and pins and other objects are important too, IRQ is just the most dangerous.

Quoted 4 lines

Please, argue that it would be better to deprecate devm_request_irq() or devm_request_threaded_irq().

No, they do have their use. But probably people doing the conversion should be required to put $100 in escrow ;)

Signature · 1 line
Dmitry
#5 lore raw

Hey, guys.

Quoted 5 lines · Dmitry Torokhov
Quoted 2 lines

Please, argue that it would be better to deprecate devm_request_irq() or devm_request_threaded_irq().

No, they do have their use. But probably people doing the conversion should be required to put $100 in escrow ;)

So, what devm does is simple - it records resources being allocated using devm and releases them in the reverse order when the driver detaches from the device. If all or most resources are covered by devm, this is the right thing to do most of the time and there's nothing much to worry about.

However, singling out one resource and converting only that one to devm is likely to break something if the resource in question has release ordering dependency and irqs do definitely have ordering requirements while tearing down, so, no, one can't single out request_irq() and change only that one to devm_request_irq(). It's like mixing managed objects with manual new/delete objects. You can't do that willy-nilly and it generaly isn't a good idea as evidenced by only c++ doing it.

The recommended way to do devm conversion is doing it all the way. Look at the whole subsystem, identify resources used by various drivers (they usually aren't that diverse in a single subsystem), identify the pattern and implement helpers in the subsystem as appropriate and convert the whole thing, preferably to the point where ->remove() almost doesn't have to worry about freeing resources other than shutting down devices or doing one-off cleanups not covered by devm.

If there are resources which are frequently used by a group of drivers but devm interface doesn't currently exist, add a devm interface either as a generic or subsystem-specific helper. devm is specifically designed to allow such extensions.

So, in summary, no, please don't single out irq requests and convert only those to devm. That's bound to break stuff.

Thanks.

Signature · 1 line
tejun