Re: [PATCH] virt: acrn: fix invalid check past list iterator

[Li Fei1 <fei1.li@intel.com>]

On Fri, Apr 01, 2022 at 10:50:51AM +0200, Jakob Koschel wrote:
> 
> 
> > On 1. Apr 2022, at 09:57, Li Fei1 <fei1.li@intel.com> wrote:
> > 
> > On Fri, Apr 01, 2022 at 09:16:48AM +0200, Jakob Koschel wrote:
> >> 
> >> 
> >>> On 1. Apr 2022, at 05:57, Li Fei1 <fei1.li@intel.com> wrote:
> >>> 
> >>> On Fri, Apr 01, 2022 at 05:22:36AM +0200, Jakob Koschel wrote:
> >>>> 
> >>>>> On 1. Apr 2022, at 03:15, Li Fei1 <fei1.li@intel.com> wrote:
> >>>>> 
> >>>>> On Thu, Mar 31, 2022 at 01:20:50PM +0200, Jakob Koschel wrote:
> >>>>>> 
> >>>>>>> On 30. Mar 2022, at 09:57, Li Fei1 <fei1.li@intel.com> wrote:
> >>>>>>> 
> >>>>>>> On Sat, Mar 19, 2022 at 09:38:19PM +0100, Jakob Koschel wrote:
> >>>>>>>> The condition retry == 0 is theoretically possible even if 'client'
> >>>>>>>> does not point to a valid element because no break was hit.
> >>>>>>>> 
> >>>>>>>> To only execute the dev_warn if actually a break within the loop was
> >>>>>>>> hit, a separate variable is used that is only set if it is ensured to
> >>>>>>>> point to a valid client struct.
> >>>>>>>> 
> >>>>>>> Hi Koschel
> >>>>>>> 
> >>>>>>> Thanks for you to help us to try to improve the code. Maybe you don't get the point.
> >>>>>>> The dev_warn should only been called when has_pending = true && retry == 0
> >>>>>> 
> >>>>>> Maybe I don't understand but looking isolated at this function I could see a way to call
> >>>>>> the dev_warn() with has_pending = false && retry == 0.
> >>>>> Yes, even has_pending = true && retry == 0 at the beginning, we could not make sure
> >>>>> has_pending is true after schedule_timeout_interruptible and we even didn't check
> >>>>> there're other pending client on the ioreq_clients (because we can't make sure
> >>>>> when we dev_warn this clent is still pending). So we just use dev_warn not higher log level.
> >>>> 
> >>>> I'm sorry, I don't quite understand what you mean by that.
> >>>> Do you agree that has_pending = false && retry == 0 is possible when calling the dev_warn()
> >>>> or not?
> >>>> 
> >>> Yes, so what ? It just a hint there may have pending request.
> >> 
> >> if has_pending == false && retry == 0 when calling dev_warn() there are very clear
> >> dependencies met:
> >> 
> >> * has_pending == false means that the list_for_each_entry() macro it *not* exit early.
> >> * since list_for_each_entry() did *not* exit early, client will not hold a valid list entry
> >> * using client->name is not safe and will not point to a valid string (perhaps not even an address)
> > So you'd better to check when the client in ioreq_clients would been destroyed, right ?
> 
> I'm afraid I don't know exactly what you mean here.
> 
> More specifically:
> to check what? and I'm also not sure what you mean with "when client in ioreq_clients would been destroyed".
> 
> Actually thinking about it more the check should be
> 
> 	if (client && retry == 0)
> 
> to be correct.
The client in ioreq_clients would always been "valid" (here valid means the client struct would not
been destroyed) when this function been called. That's guaranteed by the code logic.
> 
> if has_pending == false I read the code as if no client was found that has a pending request 
> so I don't follow how:
> 
> 	"%s cannot flush pending request!\n", client->name);
> 
> can be valid since no client has a pending request.
> 
> > 
> >> 
> >> I'm *only* talking about the case where has_pending == false, in case that's not clear.
> >> 
> >> 
> >>> Even retry == 0 && has_pending = true,
> >>> When we call dev_warn, the clent may not is pending.
> >>> 
> >>> 
> >>>>>> 
> >>>>>>> 		list_for_each_entry(client, &vm->ioreq_clients, list) {
> >>>>>>> 			has_pending = has_pending_request(client);
> >>>>>>> 			if (has_pending)
> >>>>>>> 		}
> >>>>>>> 		spin_unlock_bh(&vm->ioreq_clients_lock);
> >>>>>> 
> >>>>>> imagine has_pending == false && retry == 1 here, then client will not hold a valid list entry.
> >>>>> What do you mean "client will not hold a valid list entry" ? 
> >>>> 
> >>>> Imagine a very simple example:
> >>>> 
> >>>> 	struct acrn_ioreq_client *client;
> >>>> 	list_for_each_entry(client, &vm->ioreq_clients, list) {
> >>>> 		continue;
> >>>> 	}
> >>>> 
> >>>> 	dev_warn(acrn_dev.this_device,
> >>>> 		 "%s cannot flush pending request!\n", client->name); /* NOT GOOD */
> >>>> 
> >>> If there're pending request, we would call schedule to schedule out then schedule back
> >>> to check the list from the beginning. If there's no pending client, we point to the last
> >>> client and break out the while loop.
> >>> 
> >>> The code doesn't want to find the pending client and break out the while loop and call
> >>> dev_warn. Please see the function comment.
> >>> 
> >>> 
> >>>> 
> >>>> Since there is no break for the list_for_each_entry() iterator to return early,
> >>>> client *cannot* be a valid entry of the list. In fact if you look at the list_for_each_entry()
> >>>> macro, it will be an 'bogus' pointer, pointing somewhere into 'vm'.
> >>>> Essentially before the terminating condition of the list traversal the following code is called:
> >>>> 
> >>>> 	list_entry(&vm->ioreq_clients, struct acrn_ioreq_client *, list);
> >>>> 
> >>>> resulting in a:
> >>>> 
> >>>> 	container_of(&vm->ioreq_clients, struct acrn_ioreq_client *, list);
> >>>> 
> >>>> &vm->ioreq_clients however is not contained in a struct acrn_ioreq_client, making
> >>>> this call compute an invalid pointer.
> >>>> Therefore using 'client' as in the example above (e.g. client->name) after the loop is
> >>>> not safe. Since the loop can never return early in the simple example above it will
> >>>> always break. On other cases (the one we are discussing here) there might be a chance that
> >>>> there is one code path (in theory) where the loop did not exit early and 'client'
> >>>> holds that 'invalid entry'.
> >>>> 
> >>>> This would be the case with has_pending = false && retry == 0.
> >>>> 
> >>>> I hope this makes sense.
> >>>> 
> >>>>> 
> >>>>>> 
> >>>>>>> 
> >>>>>>> 		if (has_pending)
> >>>>>>> 			schedule_timeout_interruptible(HZ / 100);
> >>>>>>> 	} while (has_pending && --retry > 0);
> >>>>>> 
> >>>>>> since has_pending && --retry > 0 is no longer true the loop stops.
> >>>>>> 
> >>>>>>> 	if (retry == 0)
> >>>>>>> 		dev_warn(acrn_dev.this_device,
> >>>>>>> 			 "%s cannot flush pending request!\n", client->name);
> >>>>>> client->name is accessed since retry == 0 now, but client is not a valid struct ending up
> >>>>>> in a type confusion.
> >>>>>> 
> >>>>>>> 
> >>>>>>> If retry > 0 and has_pending is true, we would call schedule_timeout_interruptible
> >>>>>>> to schedule out to wait all the pending I/O requests would been completed.
> >>>>>>> 
> >>>>>>> Thanks.
> >>>>>> 
> >>>>>> Again, I'm not sure if this is realistically possible. I'm trying to remove
> >>>>>> any use of the list iterator after the loop to make such potentially issues detectable
> >>>>> You may think we still in the loop (could we ?), even that we didn't write the list iterator then.
> >>>> 
> >>>> I'm not exactly sure which loop you are referring to but no, I don't think we are still in
> >>>> the do while loop.
> >>>> 
> >>>> The only thing we know after the do while loop is that: !has_pending || retry == 0
> >>>> And iff has_pending && retry == 0, then we shouldn't call the dev_warn().
> >>>> 
> >>>>>> at compile time instead of relying on certain (difficult to maintain) conditions to be met
> >>>>>> to avoid the type confusion.
> >>>>>> 
> >>>>>> Thanks,
> >>>>>> Jakob
> >>>> 
> >>>> 	Jakob
> >> 
> >> 	Jakob
>