On Mon, Sep 14, 2026 at 1:10 PM Breno Leitao <[email protected]> wrote: > > A target running with oops_only=1 receives nothing when the kernel > oopses, which is the one thing the option exists for. Only a panic > still gets through. > > netconsole_write() bails out unless oops_in_progress is set, and that > flag is already gone by the time netconsole runs. netconsole is > CON_NBCON_ATOMIC_UNSAFE, so console_is_usable() keeps it out of the > emergency flush an oops prints in, and the records only leave the box > once the printer thread runs. > > die() wakes that thread from oops_exit(), by which point > bust_spinlocks(0) has cleared oops_in_progress again. TAINT_DIE is set > in between, so test that as well, behind a helper. > > With oops_only=1 on a netdevsim target, an lkdtm EXCEPTION produces 44 > lines of oops. The receiver gets none of them before this change and > all of them after it, while ordinary messages stay suppressed. > > The taint never clears, so from the first oops on, an oops_only target > sends everything instead of going quiet again. Sending the aftermath of > a crash beats sending nothing. > > Fixes: 7eab73b18630 ("netconsole: convert to NBCON console infrastructure") > Signed-off-by: Breno Leitao <[email protected]>
Reviewed-by: Gustavo Luiz Duarte <[email protected]>
