On 2020/04/24 0:54, Sergey Senozhatsky wrote:
> Hmm. Any chance that those NO_CONSOLE messages overflow logbuf so then
> we start losing pending YES_CONSOLE messages?
Of course there will be such possibility if printk() user is doing something
close to
while (1) {
cond_resched();
printk("hello\n");
}
which is not avoidable unless "we allocate huge logbuf" and/or "add a flow control
on the caller side". My proposal was to keep printk() synchronous by offloading
dump_tasks() to a deferrable context (i.e. a workqueue context) and adding a flow
control so that "logbuf will not overflow", and Michal Hocko is still refusing on
adding a flow control to dump_tasks() side. The possibility of overflowing logbuf
will remain no matter how printk() changes (i.e. even asynchronous printk cannot
solve).
NO_CONSOLE is not only a way for suppress printing to consoles but also a way for
reducing possibility of overflowing logbuf, by hoping that userspace syslog daemon
can read logbuf quickly enough to make free space for future NO_CONSOLE/YES_CONSOLE
messages.
