On 09/14/2016 03:14 PM, Linus Torvalds wrote:
Ugh, I detest this patch.My gut feeling is that a driver (even a fairly core one like the random code) should not have to know these kinds of details like "schedule_work() needs system_wq to have been initialized". I'm wondering if we couldn't just initialize "system_wq" earlier. Right now init_workqueues() is an "early_initcall()", so it's at the same priority as a number of other random early initcalls. My gut feeling is that it should be initialized even earlier, with the scheduler. Because dammit, we use "schedule_work()" as if it was a pretty core scheduler thing, and having to have some odd knowledge of system_wq initialization details in the rest of the kernel sounds really really wrong. I don't think the random code is at all special in maybe wanting to schedule some work despite being an "early" initcall. Adding Tejun to the cc, and quoting the whole email. Tejun, comments? Linus
My patch does not really fix the boot problem as detailed in my follow-up email. It serves mostly to jump start the discussion on the problem that I saw. The schedule_work() call was issued as part of interrupt handling that seems to be started pretty early in the boot process before the early_initcall. I guess it is possible to move the initialization earlier, but I am not sure where will be a good place.
Cheers, Longman
