Am Freitag, den 15.05.2020, 15:31 -0500 schrieb Pete Zaitcev: > > and usb_poison_anchored_urbs. If you think that poisoning may help against > what the bot identified, we may try this instead: Sure. Would you send in a patch? > I'm still concerned that we didn't identify the scenario tht led to bot's > findings. So am I. Yet as far as I can tell the code of usblp is correct. Even worse, it is casting doubt on our testing framework. > The usblp->present was supposed to play a role of the poison pill, at the > driver level. The difference with poisoning the anchor is that ->present > is protected by the most outlying mutex, and therefore cannot be meaninfully > checked in URB callbacks. But the anchor's poison flag is protected by a > spinlock, so callbacks check it. But what does it matter for us? This driver > does not re-submit URBs from callbacks. It also makes resubmission impossible. In fact, considering that, we might better poison before we go for usblp->present. > So, I'm suspicious of attempts to hit at the problem in the dark and hope > for a miracle. Right. The only thing worse would be doing nothing. At the risk of repeating myself, usblp looks correct to me. Regards Oliver
