On Mon, Sep 16, 2024 at 3:24 AM syzbot <syzbot+3e563d99e70973c0755c@xxxxxxxxxxxxxxxxxxxxxxxxx> wrote: > > syzbot has found a reproducer for the following issue on: > > HEAD commit: 68d4209158f4 sub: cdns3: Use predefined PCI vendor ID cons.. > git tree: https://git.kernel.org/pub/scm/linux/kernel/git/gregkh/usb.git usb-testing > console output: https://syzkaller.appspot.com/x/log.txt?x=10a96200580000 > kernel config: https://syzkaller.appspot.com/x/.config?x=cb61872d4d8c5df9 > dashboard link: https://syzkaller.appspot.com/bug?extid=3e563d99e70973c0755c > compiler: gcc (Debian 12.2.0-14) 12.2.0, GNU ld (GNU Binutils for Debian) 2.40 > syz repro: https://syzkaller.appspot.com/x/repro.syz?x=1297cc07980000 > C reproducer: https://syzkaller.appspot.com/x/repro.c?x=1217c8a9980000 Hi Chang, I saw your patch testing request for this bug [1] (thank you for looking into it!), but your patch doesn't appear correct to me, even though it did satisfy syzbot. One issue that I see with the patch is calling spin_unlock_irqrestore(&dev->lock) after kref_put(&dev->count): kref_put() might free the device struct and spin_unlock_irqrestore() will then do a UAF. I'm not sure what the correct patch would be though, as I don't understand what the issue is. It seems that dev_free() indeed gets called twice, but since it's guarded by kref_put(), this shouldn't happen AFAIU. Or at least we should get a bad refcount report. Thanks! [1] https://lore.kernel.org/all/6721d497.050a0220.35b515.0001.GAE@xxxxxxxxxx/T/#mca35ec73d63ff30d7219b6739ed0801688f15f35