Skip to contentSkip to conversation

watchdog: pika_wdt: portability and timer lifecycle fixes

PATCH In review Rosen Penev 3 participants 9 messages Jun 8 – Jun 9 Watchdog device drivers

lore
Share & tools
lore
#0 lore raw

v3 cover watchdog: pika_wdt: portability and timer lifecycle fixes

Fix two timer lifecycle bugs in the PIKA watchdog driver:

- pikawdt_release() could race with a concurrent open() due to clearing the open bit before stopping the timer.
- pikawdt_exit() unmapped FPGA memory without stopping the ping timer first, potentially executing the handler on freed memory.

Then replace powerpc-specific I/O accessors with generic ones to allow COMPILE_TEST builds on other architectures.

v3: fix sashiko reported issues
v2: fix preexisting issues

Rosen Penev (2):
watchdog: pika_wdt: fix timer lifecycle bugs
watchdog: pika_wdt: enable COMPILE_TEST builds

 drivers/watchdog/Kconfig    |  3 ++-
 drivers/watchdog/pika_wdt.c | 12 +++++++-----
 2 files changed, 9 insertions(+), 6 deletions(-)

--
2.54.0

#1 lore raw

PATCH v3 1/2 watchdog: pika_wdt: fix timer lifecycle bugs

Two timer lifecycle bugs in the PIKA watchdog driver:

1. pikawdt_release() clears the open bit before stopping the timer, allowing a concurrent pikawdt_open() to start a new timer that is then immediately deleted. Move clear_bit after the timer stop. Use timer_shutdown_sync() so the handler cannot re-arm after the call returns. Add timer_setup() in pikawdt_open() to re-initialize the timer for the next open.

2. pikawdt_exit() unmaps FPGA IO memory without first stopping the ping timer. If the timer handler fires after iounmap(), it executes pikawdt_ping() on freed memory. Use timer_shutdown_sync() which permanently prevents re-arming.

Assisted-by opencode:big-pickle
---
 drivers/watchdog/pika_wdt.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
drivers/watchdog/pika_wdt.c +3 -1
index 87b8988d2520..d567103ec5e4 100644
@@ -117,6 +117,7 @@ static int pikawdt_open(struct inode *inode, struct file *file)
 	if (test_and_set_bit(0, &pikawdt_private.open))
 		return -EBUSY;
 
+	timer_setup(&pikawdt_private.timer, pikawdt_ping, 0);
 	pikawdt_start();
 
 	return stream_open(inode, file);
@@ -129,7 +130,7 @@ static int pikawdt_release(struct inode *inode, struct file *file)
 {
 	/* stop internal ping */
 	if (!pikawdt_private.expect_close)
-		timer_delete(&pikawdt_private.timer);
+		timer_shutdown_sync(&pikawdt_private.timer);
 
 	clear_bit(0, &pikawdt_private.open);
 	pikawdt_private.expect_close = 0;
@@ -291,6 +292,7 @@ static void __exit pikawdt_exit(void)
 {
 	misc_deregister(&pikawdt_miscdev);
 
+	timer_shutdown_sync(&pikawdt_private.timer);
 	iounmap(pikawdt_private.fpga);
 }
 

--
2.54.0

#2 lore raw

Re: [PATCHv3 1/2] watchdog: pika_wdt: fix timer lifecycle bugs

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [Critical] Calling timer_setup() unconditionally in pikawdt_open() corrupts the active timer wheel if the…

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Critical] Calling timer_setup() unconditionally in pikawdt_open() corrupts the active timer wheel if the device was previously closed gracefully (magic close). --

commit cf2334cde532efba699588386b4f5a28da397535
Author: Rosen Penev <[email protected]>

watchdog: pika_wdt: fix timer lifecycle bugs

This commit updates the PIKA watchdog driver to use timer_shutdown_sync() during device release and module exit. It also adds a timer_setup() call to pikawdt_open() to re-initialize the timer for subsequent opens.

drivers/watchdog/pika_wdt.c static int pikawdt_open(struct inode *inode, struct file *file) · line 120

Quoted 9 lines
drivers/watchdog/pika_wdt.c +1 -0
index 87b8988d25206..d567103ec5e48 100644
@@ -117,6 +117,7 @@ static int pikawdt_open(struct inode *inode, struct file *file)
 	if (test_and_set_bit(0, &pikawdt_private.open))
 		return -EBUSY;
 
+	timer_setup(&pikawdt_private.timer, pikawdt_ping, 0);

[Severity: Critical]
Does this code corrupt the active timer wheel if the device was previously closed gracefully?

The watchdog driver supports a "magic close" feature where writing 'V' before closing sets expect_close. During pikawdt_release(), if expect_close is set, the internal ping timer is intentionally left running:

static int pikawdt_release(struct inode *inode, struct file *file)

{
	/* stop internal ping */
	if (!pikawdt_private.expect_close)
		timer_shutdown_sync(&pikawdt_private.timer);

	clear_bit(0, &pikawdt_private.open);
	...
}

Because the open bit is cleared, a subsequent open is allowed. When re-opened, this new timer_setup() is invoked unconditionally on the actively queued timer.

Calling timer_setup() on a pending timer overwrites its internal linked list pointers, which can lead to a panic when the timer softirq processes the wheel.

Quoted 3 lines
drivers/watchdog/pika_wdt.c +0 -0
 	pikawdt_start();
 
 	return stream_open(inode, file);

[ ... ]

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
maintainer
#3 lore raw

Re: [PATCHv3 1/2] watchdog: pika_wdt: fix timer lifecycle bugs

In reply to Rosen Penev
Quoted 8 lines · Rosen Penev

Two timer lifecycle bugs in the PIKA watchdog driver:

1. pikawdt_release() clears the open bit before stopping the timer, allowing a concurrent pikawdt_open() to start a new timer that is then immediately deleted. Move clear_bit after the timer stop. Use timer_shutdown_sync() so the handler cannot re-arm after the call returns. Add timer_setup() in pikawdt_open() to re-initialize the timer for the next open.

The clear bit is not moved, and the approach is wrong. More on that below.

Quoted 17 lines

2. pikawdt_exit() unmaps FPGA IO memory without first stopping the ping timer. If the timer handler fires after iounmap(), it executes pikawdt_ping() on freed memory. Use timer_shutdown_sync() which permanently prevents re-arming.

Assisted-by opencode:big-pickle
---
  drivers/watchdog/pika_wdt.c | 4 +++-
  1 file changed, 3 insertions(+), 1 deletion(-)
drivers/watchdog/pika_wdt.c +0 -0
index 87b8988d2520..d567103ec5e4 100644

drivers/watchdog/pika_wdt.c static int pikawdt_open(struct inode *inode, struct file *file) · line 120

Quoted 4 lines
drivers/watchdog/pika_wdt.c +1 -0
@@ -117,6 +117,7 @@ static int pikawdt_open(struct inode *inode, struct file *file)
  	if (test_and_set_bit(0, &pikawdt_private.open))
  		return -EBUSY;
 
+	timer_setup(&pikawdt_private.timer, pikawdt_ping, 0);

Calling timer_setup() in the open function is wrong because the timer may already be running.

drivers/watchdog/pika_wdt.c static int pikawdt_release(struct inode *inode, struct file *file) · line 133

Quoted 9 lines
drivers/watchdog/pika_wdt.c +1 -1
  	pikawdt_start();
 
  	return stream_open(inode, file);
@@ -129,7 +130,7 @@ static int pikawdt_release(struct inode *inode, struct file *file)
  {
  	/* stop internal ping */
  	if (!pikawdt_private.expect_close)
-		timer_delete(&pikawdt_private.timer);
+		timer_shutdown_sync(&pikawdt_private.timer);

This is wrong here. timer_shutdown_sync() is only supposed to be called if the timer will not be re-enabled. Anything else would be an API violation.

Quoted 15 lines
drivers/watchdog/pika_wdt.c +1 -1
  	clear_bit(0, &pikawdt_private.open);
  	pikawdt_private.expect_close = 0;
@@ -291,6 +292,7 @@ static void __exit pikawdt_exit(void)
  {
  	misc_deregister(&pikawdt_miscdev);
 
+	timer_shutdown_sync(&pikawdt_private.timer);
  	iounmap(pikawdt_private.fpga);
  }
 
--

2.54.0

#4 lore raw

PATCH v3 2/2 watchdog: pika_wdt: enable COMPILE_TEST builds

In reply to Rosen Penev

Replace powerpc-specific I/O accessors (in_be32/out_be32) with generic ones (ioread32be/iowrite32be) so the driver can be compile-tested on other architectures. Relax the Kconfig dependency accordingly and add the required HAS_IOMEM.

Assisted-by opencode:big-pickle
---
 drivers/watchdog/Kconfig    | 3 ++-
 drivers/watchdog/pika_wdt.c | 8 ++++----
 2 files changed, 6 insertions(+), 5 deletions(-)
drivers/watchdog/Kconfig +2 -1
index 8319c503319a..855736eaca2c 100644
@@ -2040,7 +2040,8 @@ config 8xxx_WDT
 
 config PIKA_WDT
 	tristate "PIKA FPGA Watchdog"
-	depends on WARP || (PPC64 && COMPILE_TEST)
+	depends on WARP || COMPILE_TEST
+	depends on HAS_IOMEM
 	default WARP
 	help
 	  This enables the watchdog in the PIKA FPGA. Currently used on
drivers/watchdog/pika_wdt.c +4 -4
index d567103ec5e4..e599aff8329b 100644
@@ -77,10 +77,10 @@ static inline void pikawdt_reset(void)
 	 *           seconds. Valid ranges are 1 to 15 seconds. The value can
 	 *           be modified dynamically.
 	 */
-	unsigned reset = in_be32(pikawdt_private.fpga + 0x14);
+	unsigned reset = ioread32be(pikawdt_private.fpga + 0x14);
 	/* enable with max timeout - 15 seconds */
 	reset |= (1 << 7) + (WDT_HW_TIMEOUT << 8);
-	out_be32(pikawdt_private.fpga + 0x14, reset);
+	iowrite32be(reset, pikawdt_private.fpga + 0x14);
 }
 
 /*
@@ -243,7 +243,7 @@ static int __init pikawdt_init(void)
 		return -ENOMEM;
 	}
 
-	ident.firmware_version = in_be32(pikawdt_private.fpga + 0x1c) & 0xffff;
+	ident.firmware_version = ioread32be(pikawdt_private.fpga + 0x1c) & 0xffff;
 
 	/* POST information is in the sd area. */
 	np = of_find_compatible_node(NULL, NULL, "pika,fpga-sd");
@@ -265,7 +265,7 @@ static int __init pikawdt_init(void)
 	 * Bit 31,   WDOG: Set to 1 when the last reset was caused by a watchdog
 	 *           timeout.
 	 */
-	post1 = in_be32(fpga + 0x40);
+	post1 = ioread32be(fpga + 0x40);
 	if (post1 & 0x80000000)
 		pikawdt_private.bootstatus = WDIOF_CARDRESET;
 

--
2.54.0

#5 lore raw

Re: [PATCHv3 2/2] watchdog: pika_wdt: enable COMPILE_TEST builds

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Pre-existing issues:
- [High] Missing timer_delete_sync in the module's __exit function leads to a Use-After-Free and kernel panic on module unload.
- [High] A race condition in pikawdt_release using a non-synchronous timer_delete can resurrect the ping timer and permanently defeat the watchdog on unexpected close. --

commit 6bf97a13e2f6f235845375a27efabc19c1cdb946
Author: Rosen Penev <[email protected]>

watchdog: pika_wdt: enable COMPILE_TEST builds

This commit replaces PowerPC-specific I/O accessors with generic ones in the PIKA FPGA watchdog driver. It updates the Kconfig to allow COMPILE_TEST builds on other architectures and adds the required HAS_IOMEM dependency.

Quoted 4 lines
drivers/watchdog/Kconfig +0 -0
index dc78729ba2a5d..c458f60f26696 100644

[ ... ]

drivers/watchdog/pika_wdt.c

Quoted 4 lines
drivers/watchdog/pika_wdt.c +0 -0
index 87b8988d25206..2d3c201540a9a 100644

[ ... ]

[Severity: High]
This isn't a bug introduced by this patch, but while reviewing this driver I noticed a potential race condition in the watchdog release handler:

drivers/watchdog/pika_wdt.c:pikawdt_release() {
    ...
	/* stop internal ping */
	if (!pikawdt_private.expect_close)
		timer_delete(&pikawdt_private.timer);

	clear_bit(0, &pikawdt_private.open);
    ...
}

If the device is closed without the magic character, pikawdt_release() calls timer_delete(). If pikawdt_ping() is executing concurrently on another CPU:

drivers/watchdog/pika_wdt.c:pikawdt_ping() {
    ...
	if (time_before(jiffies, pikawdt_private.next_heartbeat) ||
			(!nowayout && !pikawdt_private.open)) {
		pikawdt_reset();
		mod_timer(&pikawdt_private.timer, jiffies + WDT_TIMEOUT);
	}
    ...
}

Could timer_delete() return immediately while the concurrent timer handler is running, allowing clear_bit(0, &pikawdt_private.open) to execute before the handler checks the open bit?

If so, would the concurrent handler evaluate (!nowayout && !pikawdt_private.open) as true, re-arm itself via mod_timer(), and continue to ping the watchdog forever?

Should timer_delete_sync() be used here instead?

[Severity: High]
This is a pre-existing issue, but I also noticed a potential use-after-free during module unload:

drivers/watchdog/pika_wdt.c:pikawdt_exit() {
	misc_deregister(&pikawdt_miscdev);

	iounmap(pikawdt_private.fpga);
}

When the watchdog device is closed with the 'magic close' character, the ping timer is intentionally left running in the background. Since misc_deregister() allows the module reference count to reach 0, the module can be unloaded.

Because pikawdt_exit() doesn't call timer_delete_sync(), won't the timer continue to execute pikawdt_ping() from freed module memory and attempt to access the unmapped pikawdt_private.fpga I/O region, leading to a kernel panic?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
maintainer
#6 lore raw

Re: [PATCHv3 2/2] watchdog: pika_wdt: enable COMPILE_TEST builds

In reply to Rosen Penev
Quoted 7 lines · Rosen Penev

Replace powerpc-specific I/O accessors (in_be32/out_be32) with generic ones (ioread32be/iowrite32be) so the driver can be compile-tested on other architectures. Relax the Kconfig dependency accordingly and add the required HAS_IOMEM.

Assisted-by opencode:big-pickle

I really don't think this is worth the trouble just to be able to support COMPILE_TEST on all architectures. This is an old watchdog driver. It should not be touched except to fix bugs.

Didn't I say that before ?

Thanks,
Guenter

#7 lore raw

Re: [PATCHv3 2/2] watchdog: pika_wdt: enable COMPILE_TEST builds

Quoted 10 lines · Guenter Roeck
Quoted 7 lines · Rosen Penev

Replace powerpc-specific I/O accessors (in_be32/out_be32) with generic ones (ioread32be/iowrite32be) so the driver can be compile-tested on other architectures. Relax the Kconfig dependency accordingly and add the required HAS_IOMEM.

Assisted-by opencode:big-pickle
Quoted 5 lines

I really don't think this is worth the trouble just to be able to support COMPILE_TEST on all architectures. This is an old watchdog driver. It should not be touched except to fix bugs.

Didn't I say that before ?

Sounds like this driver should be left alone then.

Quoted 3 lines

Thanks,
Guenter

maintainer
#8 lore raw

Re: [PATCHv3 2/2] watchdog: pika_wdt: enable COMPILE_TEST builds

Quoted 11 lines · Rosen Penev
Quoted 10 lines · Guenter Roeck
Quoted 7 lines · Rosen Penev

Replace powerpc-specific I/O accessors (in_be32/out_be32) with generic ones (ioread32be/iowrite32be) so the driver can be compile-tested on other architectures. Relax the Kconfig dependency accordingly and add the required HAS_IOMEM.

Assisted-by opencode:big-pickle
Quoted 6 lines
Quoted 5 lines

I really don't think this is worth the trouble just to be able to support COMPILE_TEST on all architectures. This is an old watchdog driver. It should not be touched except to fix bugs.

Didn't I say that before ?

Sounds like this driver should be left alone then.

Very much so. The driver is for a FPGA based system from around 2007, and the company behind it went bankrupt in 2022. I would be surprised if the current Linux kernel even boots on it.

Guenter