iio: health: max30102: fix NULL dereference in interrupt handler
PATCH 1/1 iio: health: max30102: fix NULL dereference in interrupt handler
The interrupt is requested in max30102_probe() and stays enabled for the lifetime of the device, but indio_dev->active_scan_mask is only valid while a buffer is enabled. When an interrupt arrives while no buffer is enabled, the handler dereferences the NULL active_scan_mask:
Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000 pc : __bitmap_weight+0x64/0x98 lr : max30102_interrupt_handler+0x48/0x160 [max30102] Call trace: __bitmap_weight+0x64/0x98 (P) max30102_interrupt_handler+0x48/0x160 [max30102] irq_thread_fn+0x28/0xa8 irq_thread+0x184/0x30c kthread+0x118/0x124 ret_from_fork+0x10/0x20
Return early when no buffer is enabled. The interrupt status register is read before returning, since reading it deasserts the chip's active-low interrupt pin. Otherwise the pin would stay asserted and no further edges would be delivered.
---
I ran into this while interfacing with the MAX30102 over I2C on a Raspberry Pi 4 running the IIO subsystem tree testing branch and learning the IIO sysfs interface for the first time. When physically rearranging INT pin wiring with no buffer enabled, the kernel oopsed. This happened because max30102_interrupt_handler() attempted to dereference active_scan_mask, which is NULL because no buffer is enabled. With this patch, the same situation no longer oopses and the buffered capture was tested to work normally afterward.
Some things I am unsure about though:
- Is IRQ_HANDLED or IRQ_NONE preferred here? I chose IRQ_HANDLED because of the status register read to deassert the INT pin, but I am not 100% confident on this choice.
- Should the regmap_read() return value be checked? I did not add a check because I do not see a useful recovery path from this I2C failure, but I can add a check in a v2 if it is better.
Thank you.
drivers/iio/health/max30102.c | 17 +++++++++++++++--
1 file changed, 15 insertions(+), 2 deletions(-)
drivers/iio/health/max30102.c +15 -2
@@ -290,10 +290,23 @@ static irqreturn_t max30102_interrupt_handler(int irq, void *private) {struct iio_dev *indio_dev = private;struct max30102_data *data =iio_priv (indio_dev); -unsigned int measurements =bitmap_weight (indio_dev->active_scan_mask, -iio_get_masklength (indio_dev)); +unsigned int measurements, val;int ret, cnt =0 ; +if (!indio_dev->active_scan_mask) { +/* +* No buffer is enabled so there is nothing to read. Read the +* status register anyway to deassert the max30102's interrupt +* pin; otherwise it would stay asserted and further edges +* would not be delivered. +*/ +regmap_read (data->regmap,MAX30102_REG_INT_STATUS , &val); +return IRQ_HANDLED ; + } + + measurements =bitmap_weight (indio_dev->active_scan_mask, +iio_get_masklength (indio_dev)); +mutex_lock (&data->lock);while (cnt || (cnt =max30102_fifo_count (data)) >0 ) {
Signature · 1 line
2.55.0
David Lechner
reviewerQuoted 1 line · Marco Chen
The interrupt is requested in
max30102_probe()and stays enabled
I've always wondered why most IIO drivers seem to do this. It seems to me that requesting the irq with IRQF_NO_AUTOEN and only enabling it when everything is set up for reading data would avoid spurious interrupts like this.
Quoted 21 lines
for the lifetime of the device, but
indio_dev->active_scan_maskis only valid while a buffer is enabled. When an interrupt arrives while no buffer is enabled, the handler dereferences the NULLactive_scan_mask:Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000 pc : __bitmap_weight+0x64/0x98 lr : max30102_interrupt_handler+0x48/0x160 [max30102] Call trace: __bitmap_weight+0x64/0x98 (P) max30102_interrupt_handler+0x48/0x160 [max30102] irq_thread_fn+0x28/0xa8 irq_thread+0x184/0x30c kthread+0x118/0x124 ret_from_fork+0x10/0x20Return early when no buffer is enabled. The interrupt status register is read before returning, since reading it deasserts the chip's active-low interrupt pin. Otherwise the pin would stay asserted and no further edges would be delivered.
Quoted 7 lines
Fixes 90579b69e94b ("iio: health: max30102: Add MAX30105 support")Signed-off-by Marco Chen <[email protected]>---
I ran into this while interfacing with the MAX30102 over I2C on a Raspberry Pi 4 running the IIO subsystem tree testing branch and learning the IIO sysfs interface for the first time. When physically rearranging INT pin wiring with no buffer enabled, the kernel oopsed.
This is one of the reasons why you are supposed to turn off power before moving wires. ;-)
Quoted 9 lines
This happened because
max30102_interrupt_handler()attempted to dereferenceactive_scan_mask, which is NULL because no buffer is enabled. With this patch, the same situation no longer oopses and the buffered capture was tested to work normally afterward.Some things I am unsure about though:
- IsIRQ_HANDLEDorIRQ_NONEpreferred here? I choseIRQ_HANDLEDbecause of the status register read to deassert the INT pin, but I am not 100% confident on this choice.
IRQ_NONE would cause the handler to just run again as soon as it exits, so it is almost never the right value to return. IRQ_HANDLED is correct.
Quoted 3 lines
- Should the
regmap_read()return value be checked? I did not add a check because I do not see a useful recovery path from this I2C failure, but I can add a check in a v2 if it is better.
There isn't anything we could do other than log the error.
Quoted 35 lines
Thank you.
drivers/iio/health/max30102.c| 17 +++++++++++++++--1 file changed, 15 insertions(+), 2 deletions(-)drivers/iio/health/max30102.c +15
-2@@ -290,10 +290,23 @@ static irqreturn_t max30102_interrupt_handler(int irq, void *private) {struct iio_dev *indio_dev = private;struct max30102_data *data =iio_priv (indio_dev); -unsigned int measurements =bitmap_weight (indio_dev->active_scan_mask, -iio_get_masklength (indio_dev)); +unsigned int measurements, val;int ret, cnt =0 ; +if (!indio_dev->active_scan_mask) { +/* +* No buffer is enabled so there is nothing to read. Read the +* status register anyway to deassert the max30102's interrupt +* pin; otherwise it would stay asserted and further edges +* would not be delivered. +*/ +regmap_read (data->regmap,MAX30102_REG_INT_STATUS , &val); +return IRQ_HANDLED ; + } + + measurements =bitmap_weight (indio_dev->active_scan_mask, +iio_get_masklength (indio_dev)); +mutex_lock (&data->lock);while (cnt || (cnt =max30102_fifo_count (data)) >0 ) {
Jonathan Cameron
maintainerQuoted 7 lines · On Sat, 1 Aug 2026 10:18:42 -0500 David Lechner <[email protected]> wrote:
Quoted 1 line · Marco Chen
The interrupt is requested in
max30102_probe()and stays enabledI've always wondered why most IIO drivers seem to do this. It seems to me that requesting the irq with
IRQF_NO_AUTOENand only enabling it when everything is set up for reading data would avoid spurious interrupts like this.
It teaches discipline. More serious from what I recall the no autoen thing used to be done via a racey bit of trying to turn it off as fast as you can.
Most of the time we should be able to write drivers that don't need to use that flag and don't touch enable_irq() at all.
Obviously that's not always true.
Quoted 38 lines
Quoted 22 lines
for the lifetime of the device, but
indio_dev->active_scan_maskis only valid while a buffer is enabled. When an interrupt arrives while no buffer is enabled, the handler dereferences the NULLactive_scan_mask:Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000 pc : __bitmap_weight+0x64/0x98 lr : max30102_interrupt_handler+0x48/0x160 [max30102] Call trace: __bitmap_weight+0x64/0x98 (P) max30102_interrupt_handler+0x48/0x160 [max30102] irq_thread_fn+0x28/0xa8 irq_thread+0x184/0x30c kthread+0x118/0x124 ret_from_fork+0x10/0x20Return early when no buffer is enabled. The interrupt status register is read before returning, since reading it deasserts the chip's active-low interrupt pin. Otherwise the pin would stay asserted and no further edges would be delivered.
Quoted 7 lines
Fixes 90579b69e94b ("iio: health: max30102: Add MAX30105 support")Signed-off-by Marco Chen <[email protected]>---
I ran into this while interfacing with the MAX30102 over I2C on a Raspberry Pi 4 running the IIO subsystem tree testing branch and learning the IIO sysfs interface for the first time. When physically rearranging INT pin wiring with no buffer enabled, the kernel oopsed.This is one of the reasons why you are supposed to turn off power before moving wires. ;-)
Quoted 5 lines
This happened because
max30102_interrupt_handler()attempted to dereferenceactive_scan_mask, which is NULL because no buffer is enabled. With this patch, the same situation no longer oopses and the buffered capture was tested to work normally afterward.
Quoted 8 lines
Quoted 4 lines
Some things I am unsure about though:
- IsIRQ_HANDLEDorIRQ_NONEpreferred here? I choseIRQ_HANDLEDbecause of the status register read to deassert the INT pin, but I am not 100% confident on this choice.
IRQ_NONEwould cause the handler to just run again as soon as it exits, so it is almost never the right value to return.IRQ_HANDLEDis correct.
This is always a fun debate. If you do return IRQ_NONE you will enter the handler again, but not for long. The spurious interrupt detection code kicks in and disables the interrupt with a nice loud splat. That may or may not be appropriate.
Quoted 45 lines
Quoted 4 lines
- Should the
regmap_read()return value be checked? I did not add a check because I do not see a useful recovery path from this I2C failure, but I can add a check in a v2 if it is better.There isn't anything we could do other than log the error.
Quoted 35 lines
Thank you.
drivers/iio/health/max30102.c| 17 +++++++++++++++--1 file changed, 15 insertions(+), 2 deletions(-)drivers/iio/health/max30102.c +15
-2@@ -290,10 +290,23 @@ static irqreturn_t max30102_interrupt_handler(int irq, void *private) {struct iio_dev *indio_dev = private;struct max30102_data *data =iio_priv (indio_dev); -unsigned int measurements =bitmap_weight (indio_dev->active_scan_mask, -iio_get_masklength (indio_dev)); +unsigned int measurements, val;int ret, cnt =0 ; +if (!indio_dev->active_scan_mask) { +/* +* No buffer is enabled so there is nothing to read. Read the +* status register anyway to deassert the max30102's interrupt +* pin; otherwise it would stay asserted and further edges +* would not be delivered. +*/ +regmap_read (data->regmap,MAX30102_REG_INT_STATUS , &val); +return IRQ_HANDLED ; + } + + measurements =bitmap_weight (indio_dev->active_scan_mask, +iio_get_masklength (indio_dev)); +mutex_lock (&data->lock);while (cnt || (cnt =max30102_fifo_count (data)) >0 ) {
Jonathan Cameron
maintainerQuoted 51 lines · On Fri, 31 Jul 2026 14:41:23 -0400 Marco Chen <[email protected]> wrote:
The interrupt is requested in
max30102_probe()and stays enabled for the lifetime of the device, butindio_dev->active_scan_maskis only valid while a buffer is enabled. When an interrupt arrives while no buffer is enabled, the handler dereferences the NULLactive_scan_mask:Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000 pc : __bitmap_weight+0x64/0x98 lr : max30102_interrupt_handler+0x48/0x160 [max30102] Call trace: __bitmap_weight+0x64/0x98 (P) max30102_interrupt_handler+0x48/0x160 [max30102] irq_thread_fn+0x28/0xa8 irq_thread+0x184/0x30c kthread+0x118/0x124 ret_from_fork+0x10/0x20Return early when no buffer is enabled. The interrupt status register is read before returning, since reading it deasserts the chip's active-low interrupt pin. Otherwise the pin would stay asserted and no further edges would be delivered.
Fixes 90579b69e94b ("iio: health: max30102: Add MAX30105 support")Signed-off-by Marco Chen <[email protected]>---
I ran into this while interfacing with the MAX30102 over I2C on a Raspberry Pi 4 running the IIO subsystem tree testing branch and learning the IIO sysfs interface for the first time. When physically rearranging INT pin wiring with no buffer enabled, the kernel oopsed. This happened becausemax30102_interrupt_handler()attempted to dereferenceactive_scan_mask, which is NULL because no buffer is enabled. With this patch, the same situation no longer oopses and the buffered capture was tested to work normally afterward.Some things I am unsure about though:
- IsIRQ_HANDLEDorIRQ_NONEpreferred here? I choseIRQ_HANDLEDbecause of the status register read to deassert the INT pin, but I am not 100% confident on this choice.
- Should theregmap_read()return value be checked? I did not add a check because I do not see a useful recovery path from this I2C failure, but I can add a check in a v2 if it is better.
Thank you.
drivers/iio/health/max30102.c| 17 +++++++++++++++--1 file changed, 15 insertions(+), 2 deletions(-)drivers/iio/health/max30102.c +0
-0
drivers/iio/health/max30102.c static irqreturn_t max30102_interrupt_handler(int irq, void *private) · line 296
Quoted 9 lines
drivers/iio/health/max30102.c +2
-2@@ -290,10 +290,23 @@ static irqreturn_t max30102_interrupt_handler(int irq, void *private) {struct iio_dev *indio_dev = private;struct max30102_data *data =iio_priv (indio_dev); -unsigned int measurements =bitmap_weight (indio_dev->active_scan_mask, -iio_get_masklength (indio_dev)); +unsigned int measurements, val;int ret, cnt =0 ; +if (!indio_dev->active_scan_mask) {
That is racy as this could be going on in parallel with the buffer being disabled so we check here but it's gone before it is accessed.
It does protect against the spurious interrupt when the device isn't really in use.
Hmm. I'm not sure what the right fix for this is or how much effort we put into protecting against this level of things going wrong.
We could do a buffer mode claim, but then we'd need to do that in every driver that touches active_scan_mask. Also for that matter that pushes to buffers as they can change at an time when buffered mode isn't in use.
drivers/iio/health/max30102.c static irqreturn_t max30102_interrupt_handler(int irq, void *private) · line 303
Quoted 7 lines
drivers/iio/health/max30102.c +7
-0+/* +* No buffer is enabled so there is nothing to read. Read the +* status register anyway to deassert the max30102's interrupt +* pin; otherwise it would stay asserted and further edges +* would not be delivered. +*/ +regmap_read (data->regmap,MAX30102_REG_INT_STATUS , &val);
That we should only do if we have any reason to believe it was set in the first
place. Did we check it? Looks like we didn't.
Which leads me to suggest alternative fix - check INT_STATUS and if nothing set it isn't our interrupt. That doesn't get into the potential buffer enabled / disabled races but should close your condition I think
- looks like it is read in the fifo part below. You may need to refactor a little to not read it twice.
Jonathan
Quoted 9 lines
drivers/iio/health/max30102.c +6
-0+return IRQ_HANDLED ; + } + + measurements =bitmap_weight (indio_dev->active_scan_mask, +iio_get_masklength (indio_dev)); +mutex_lock (&data->lock);while (cnt || (cnt =max30102_fifo_count (data)) >0 ) {
David Lechner
reviewerQuoted 11 lines · Jonathan Cameron
Quoted 7 lines · On Sat, 1 Aug 2026 10:18:42 -0500 David Lechner <[email protected]> wrote:
Quoted 1 line · Marco Chen
The interrupt is requested in
max30102_probe()and stays enabledI've always wondered why most IIO drivers seem to do this. It seems to me that requesting the irq with
IRQF_NO_AUTOENand only enabling it when everything is set up for reading data would avoid spurious interrupts like this.
Quoted 8 lines
It teaches discipline. More serious from what I recall the no autoen thing used to be done via a racey bit of trying to turn it off as fast as you can.
Most of the time we should be able to write drivers that don't need to use that flag and don't touch
enable_irq()at all.Obviously that's not always true.
Hmm... so I should go back and change [1] since it isn't strictly required there?
Jonathan Cameron
maintainerQuoted 18 lines · On Sun, 2 Aug 2026 10:21:41 -0500 David Lechner <[email protected]> wrote:
Quoted 17 lines · Jonathan Cameron
Quoted 7 lines · On Sat, 1 Aug 2026 10:18:42 -0500 David Lechner <[email protected]> wrote:
Quoted 1 line · Marco Chen
The interrupt is requested in
max30102_probe()and stays enabledI've always wondered why most IIO drivers seem to do this. It seems to me that requesting the irq with
IRQF_NO_AUTOENand only enabling it when everything is set up for reading data would avoid spurious interrupts like this.It teaches discipline. More serious from what I recall the no autoen thing used to be done via a racey bit of trying to turn it off as fast as you can.
Most of the time we should be able to write drivers that don't need to use that flag and don't touch
enable_irq()at all.
Quoted 9 lines
Quoted 2 lines
Obviously that's not always true.
Hmm... so I should go back and change [1] since it isn't strictly required there?
Ah. I was assuming that was spurious irq territory or a device where we couldn't disable it device end.
I remember a discussion a long time back with the irq maintainers (probably tglx) about the fact that they really advise drivers not to turn interrupts on or off because it can be a very expensive operation on some systems and such control belongs at the device end when possible. What a host does with a disabled irq is just too variable. The only guarantee is it won't immediately report it.
Maybe that has changed over time - I have no idea!
Anyhow, yes - I would drop that enable / disable dance if it was about spurious prevention. If we want to prevent spurious then better thing to do is to check the software interrupt status register.
Looks like for that part there is a drdy status bit. Bit annoying that you'd
have to drop to a thread to read it but we probably should do that before the complete. Maybe we could speculatively complete and check it in the main thread.
I note you do use that for polling. I'd treat the irq as a way to ensure you only need to read it once rather than not read it at all.
Jonathan
Quoted 4 lines · Jonathan Cameron
drivers/iio/health/max30102.c static irqreturn_t max30102_interrupt_handler(int irq, void *private) · line 296
Quoted 1 line
drivers/iio/health/max30102.c +1
-0+if (!indio_dev->active_scan_mask) {That is racy as this could be going on in parallel with the buffer being disabled so we check here but it's gone before it is accessed.
I understand what you mean. I will remove this active_scan_mask check.
Quoted 5 lines
Which leads me to suggest alternative fix - check
INT_STATUSand if nothing set it isn't our interrupt. That doesn't get into the potential buffer enabled / disabled races but should close your condition I think
- looks like it is read in the fifo part below. You may need to refactor a little to not read it twice.
This makes a lot of sense. For the v2 patch, I will plan to read INT_STATUS once at the top of the handler and pass that value into max30102_fifo_count() so it is not read twice. Consequently I will remove the INT_STATUS read from max30102_fifo_count(), too.
For checking INT_STATUS to see if it isn't our interrupt, instead of checking if nothing is set, I was thinking to check the MAX30102_REG_INT_STATUS_FIFO_RDY bit specifically because that is the only interrupt enabled in max30102_chip_init().
This will close the reproducer I was hitting earlier as you said. Since we only ever reach bitmap_weight() when FIFO_RDY is set, the NULL dereference is avoided. I think there is a theoretical window if the buffer is disabled between the FIFO_RDY check and bitmap_weight() call, but as you said, closing this would need the buffered mode claim. I plan to leave that out of this fix.
Quoted 2 lines · David Lechner
This is one of the reasons why you are supposed to turn off power before moving wires. ;-)
Fair point, I was pretty careless.
Quoted 3 lines
IRQ_NONEwould cause the handler to just run again as soon as it exits, so it is almost never the right value to return.IRQ_HANDLEDis correct.
Got it. I will keep IRQ_HANDLED for the v2.
Quoted 1 line
There isn't anything we could do other than log the error.
Understood. Since the INT_STATUS read at the top of the interrupt handler now determines if it is our interrupt, I will check the return value of regmap_read() and log the error on failure.
Jonathan Cameron
maintainerQuoted 19 lines · On Mon, 3 Aug 2026 00:13:33 -0400 Marco Chen <[email protected]> wrote:
Quoted 4 lines · Jonathan Cameron
drivers/iio/health/max30102.c static irqreturn_t max30102_interrupt_handler(int irq, void *private) · line 296
Quoted 1 line
drivers/iio/health/max30102.c +1
-0+if (!indio_dev->active_scan_mask) {That is racy as this could be going on in parallel with the buffer being disabled so we check here but it's gone before it is accessed.
I understand what you mean. I will remove this
active_scan_maskcheck.Quoted 5 lines
Which leads me to suggest alternative fix - check
INT_STATUSand if nothing set it isn't our interrupt. That doesn't get into the potential buffer enabled / disabled races but should close your condition I think
- looks like it is read in the fifo part below. You may need to refactor a little to not read it twice.This makes a lot of sense. For the v2 patch, I will plan to read
INT_STATUSonce at the top of the handler and pass that value intomax30102_fifo_count()so it is not read twice. Consequently I will remove theINT_STATUSread frommax30102_fifo_count(), too.
Quoted 4 lines
For checking
INT_STATUSto see if it isn't our interrupt, instead of checking if nothing is set, I was thinking to check theMAX30102_REG_INT_STATUS_FIFO_RDYbit specifically because that is the only interrupt enabled inmax30102_chip_init().
Agreed. That is a better solution than mine :)
Quoted 7 lines
This will close the reproducer I was hitting earlier as you said. Since we only ever reach
bitmap_weight()whenFIFO_RDYis set, the NULL dereference is avoided. I think there is a theoretical window if the buffer is disabled between theFIFO_RDYcheck andbitmap_weight()call, but as you said, closing this would need the buffered mode claim. I plan to leave that out of this fix.