reboot: don't use cb_data to tell if the static sys-off handler is free
PATCH 1/1 reboot: don't use cb_data to tell if the static sys-off handler is free
There is one static handler for SYS_OFF_PRIO_PLATFORM, shared by all modes, so the second platform priority registration has to fail with -EBUSY. alloc_sys_off_handler() decides that by testing cb_data.
cb_data is caller data though, and nothing says it has to be set. A registration that passes NULL for it leaves the static handler looking free, so the next one gets the same struct back while the first is still live: sys_off_cb and list get overwritten, and the same notifier_block is registered into a second chain. A notifier_block has one next pointer, so both chains end up wrong.
register_platform_power_off() always passes the power off callback as cb_data, which is why this held up when the static handler was added. Callers using register_sys_off_handler() directly have no such rule.
Test sys_off_cb instead. Every handler has one, it is what the callback path actually needs, and it is set for as long as the handler is registered.
--- kernel/reboot.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
kernel/reboot.c +1 -1
@@ -366,7 +366,7 @@ static struct sys_off_handler *alloc_sys_off_handler(int priority)*/ if (priority ==SYS_OFF_PRIO_PLATFORM ) { handler = &platform_sys_off_handler; -if (handler->cb_data) +if (handler->sys_off_cb)return ERR_PTR (-EBUSY ); }else {if (system_state >SYSTEM_RUNNING )
Signature · 1 line
2.47.3
Quoted 15 lines · On Sun, 26 Jul 2026 22:55:41 +0000 Bradley Morgan <[email protected]> wrote:
There is one static handler for
SYS_OFF_PRIO_PLATFORM, shared by all modes, so the second platform priority registration has to fail with-EBUSY.alloc_sys_off_handler()decides that by testingcb_data.
cb_datais caller data though, and nothing says it has to be set. A registration that passes NULL for it leaves the static handler looking free, so the next one gets the same struct back while the first is still live:sys_off_cband list get overwritten, and the samenotifier_blockis registered into a second chain. Anotifier_blockhas one next pointer, so both chains end up wrong.
register_platform_power_off()always passes the power off callback ascb_data, which is why this held up when the static handler was added. Callers usingregister_sys_off_handler()directly have no such rule.
Quoted 5 lines
Test
sys_off_cbinstead. Every handler has one, it is what the callback path actually needs, and it is set for as long as the handler is registered.Fixes 587b9bfe0668 ("kernel/reboot: Use static handler for register_platform_power_off()")
Please do try to hunt down the people who are best situated to review proposed changes. `git show 587b9bfe0668' ain't hard!
I assume you saw the possible pre-existing thing:
https://sashiko.dev/#/patchset/[email protected]
Quoted 16 lines
...
kernel/reboot.c +1
-1@@ -366,7 +366,7 @@ static struct sys_off_handler *alloc_sys_off_handler(int priority)*/ if (priority ==SYS_OFF_PRIO_PLATFORM ) { handler = &platform_sys_off_handler; -if (handler->cb_data) +if (handler->sys_off_cb)return ERR_PTR (-EBUSY ); }else {if (system_state >SYSTEM_RUNNING )Signature · 1 line
2.47.3