On Mon, May 2, 2016 at 3:51 PM, Chen Gang <chengang@xxxxxxxxxxxxxxxx> wrote: > On 5/2/16 20:42, Alexander Potapenko wrote: >> On Mon, May 2, 2016 at 2:27 PM, Chen Gang <chengang@xxxxxxxxxxxxxxxx> wrote: >>> On 5/2/16 19:21, Dmitry Vyukov wrote: >>>> >>>> Signed counter looks good to me. >>> >>> Oh, sorry, it seems a little mess (originally, I need let the 2 patches >>> in one patch set). >>> >>> If what Alexander's idea is OK (if I did not misunderstand), I guess, >>> unsigned counter is still necessary. >> I don't think it's critical for us to use an unsigned counter. >> If we increment the counter in kasan_disable_current() and decrement >> it in kasan_enable_current(), as Dmitry suggested, we'll be naturally >> using only positive integers for the counter. >> If the counter drops below zero, or exceeds a certain number (say, >> 20), we can immediately issue a warning. >> > > OK, thanks. > > And for "kasan_depth == 1", I guess, its meaning is related with > kasan_depth[++|--] in kasan_[en|dis]able_current(): Assuming you are talking about the assignment of 1 to kasan_depth in /include/linux/init_task.h, it's somewhat counterintuitive. I think we just need to replace it with kasan_disable_current(), and add a corresponding kasan_enable_current() to the end of kasan_init. > - If kasan_depth++ in kasan_enable_current() with preventing overflow/ > underflow, it means "we always want to disable KASAN, if CONFIG_KASAN > is not under arm64 or x86_64". > > - If kasan_depth-- in kasan_enable_current() with preventing overflow/ > underflow, it means "we can enable KASAN if CONFIG_KASAN, but firstly > we disable it, if it is not under arm64 or x86_64". > > For me, I don't know which one is correct (or my whole 'guess' is > incorrect). Could any members provide your ideas? > >>>> We can both issue a WARNING and prevent the actual overflow/underflow. >>>> But I don't think that there is any sane way to handle the bug (other >>>> than just fixing the unmatched disable/enable). If we ignore an >>>> excessive disable, then we can end up with ignores enabled >>>> permanently. If we ignore an excessive enable, then we can end up with >>>> ignores enabled when they should not be enabled. The main point here >>>> is to bark loudly, so that the unmatched annotations are noticed and >>>> fixed. >>>> >>> >>> How about BUG_ON()? >> As noted by Dmitry in an offline discussion, we shouldn't bail out as >> long as it's possible to proceed, otherwise the kernel may become very >> hard to debug. >> A mismatching annotation isn't a case in which we can't proceed with >> the execution. > > OK, thanks. > > I guess, we are agree with each other: "We can both issue a WARNING and > prevent the actual overflow/underflow.". No, I am not sure think that we need to prevent the overflow. As I showed before, this may result in kasan_depth being off even in the case kasan_enable_current()/kasan_disable_current() are used consistently. > Thanks. > -- > Chen Gang (陈刚) > > Managing Natural Environments is the Duty of Human Beings. -- Alexander Potapenko Software Engineer Google Germany GmbH Erika-Mann-Straße, 33 80636 München Geschäftsführer: Matthew Scott Sucherman, Paul Terence Manicle Registergericht und -nummer: Hamburg, HRB 86891 Sitz der Gesellschaft: Hamburg -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href