On 2023/4/28 0:45, Luck, Tony wrote:
But in the core dump case there is no return to user. The process is being
terminated by the signal that leads to this core dump. So even though you
may consider the page being accessed to be a "user" page, you can't fix
it by queueing work to run on return to user.
For coredump，the task work will be called too, see following code,
get_signal
sig_kernel_coredump
elf_core_dump
dump_user_range
_copy_from_iter // with MC-safe copy, return without panic
do_group_exit(ksig->info.si_signo);
do_exit
exit_task_work
task_work_run
kill_me_never
memory_failure
Nice. I didn't realize that the exit code path would clear any pending task_work() requests.
But it makes sense that this happens. Thanks for filling a gap in my knowledge.
Yep, we could be benefit from it to unify memory failure handling :)
-Tony
