watchdog: omap: Do not read the counter while the watchdog is stopped
Subsystems (2)
PATCH 1/1 watchdog: omap: Do not read the counter while the watchdog is stopped
omap_wdt only holds a runtime PM reference between omap_wdt_start() and omap_wdt_stop(), so the module is idle and its clock gated whenever the watchdog is not running. omap_wdt_get_timeleft() reads the WCRR register unconditionally, but the watchdog core calls ->get_timeleft() from the sysfs timeleft attribute without checking whether the watchdog is active.
With CONFIG_WATCHDOG_SYSFS=y the attribute is world readable, so any unprivileged user can trigger an external abort on a stopped watchdog. On a BeagleBone Black (AM335x), with panic_on_oops set:
$ id uid=1000(tester) gid=1000(tester) groups=1000(tester) $ cat /sys/class/watchdog/watchdog0/timeleft Unhandled fault: external abort on non-linefetch (0x1028) at 0xf9e35028 Internal error: : 1028 [#1] SMP ARM CPU: 0 UID: 1000 PID: 239 Comm: cat Not tainted 7.4.0-rc2-00716-g23c62304dde2-dirty #10 VOLUNTARY Hardware name: Generic AM33XX (Flattened Device Tree) PC is at omap_wdt_get_timeleft+0x4/0x14 LR is at watchdog_get_timeleft+0x24/0x130 Call trace: omap_wdt_get_timeleft from watchdog_get_timeleft+0x24/0x130 watchdog_get_timeleft from timeleft_show+0x3c/0x80 timeleft_show from dev_attr_show+0x18/0x4c dev_attr_show from sysfs_kf_seq_show+0xa0/0x138 sysfs_kf_seq_show from seq_read_iter+0x194/0x4c0 seq_read_iter from copy_splice_read+0x168/0x2d8 copy_splice_read from splice_direct_to_actor+0xb4/0x2b4 splice_direct_to_actor from do_splice_direct_actor+0x84/0xcc do_splice_direct_actor from do_splice_direct+0x24/0x30 do_splice_direct from do_sendfile+0x2fc/0x3a4 do_sendfile from sys_sendfile64+0x10c/0x128 sys_sendfile64 from ret_fast_syscall+0x0/0x54 Kernel panic - not syncing: Fatal exception
Only read the counter while the watchdog is running, under wdev->lock so the check cannot race with start() and stop(), and report 0 time left otherwise.
Tested on a BeagleBone Black: with the watchdog stopped, timeleft now reads 0 for an unprivileged user; with the watchdog started through /dev/watchdog it counts down as before.
--- drivers/watchdog/omap_wdt.c | 10 +++++++--- 1 file changed, 7 insertions(+), 3 deletions(-)
drivers/watchdog/omap_wdt.c +7 -3
@@ -206,10 +206,14 @@ static unsigned int omap_wdt_get_timeleft(struct watchdog_device *wdog) {struct omap_wdt_dev *wdev =to_omap_wdt_dev (wdog);void __iomem *base = wdev->base; -u32 value; +unsigned int timeleft =0 ; - value =readl_relaxed (base +OMAP_WATCHDOG_CRR ); -return GET_WCCR_SECS (value); +mutex_lock (&wdev->lock); +if (wdev->omap_wdt_users) + timeleft =GET_WCCR_SECS (readl_relaxed (base +OMAP_WATCHDOG_CRR )); +mutex_unlock (&wdev->lock); + +return timeleft; }static const struct watchdog_info omap_wdt_info = {
Signature · 1 line
2.25.1
Guenter Roeck
maintainerQuoted 8 lines · Nhat Tran
omap_wdtonly holds a runtime PM reference betweenomap_wdt_start()andomap_wdt_stop(), so the module is idle and its clock gated whenever the watchdog is not running.omap_wdt_get_timeleft()reads the WCRR register unconditionally, but the watchdog core calls ->get_timeleft() from the sysfs timeleft attribute without checking whether the watchdog is active.With
CONFIG_WATCHDOG_SYSFS=ythe attribute is world readable, so any
The watchdog can also be stopped while the watchdog device is open, so the problem is a bit broader than that.
Quoted 3 lines
unprivileged user can trigger an external abort on a stopped watchdog. On a BeagleBone Black (AM335x), with
panic_on_oopsset:
Seems to me the proper generic fix should be in watchdog_get_timeleft(), which should return either return the configured timeout or maybe 0 (I am open to suggestions) if the watchdog is not running.
Thanks,
Guenter