Re: [device property] 995fe757ec: BUG:kernel_NULL_pointer_dereference,address

From: Hans de Goede
Date: Wed Nov 17 2021 - 06:54:58 EST


Hi,

On 11/17/21 01:10, Daniel Scally wrote:
> Hi Hans, Andy
>
> On 16/11/2021 16:59, Andy Shevchenko wrote:
>> On Tue, Nov 16, 2021 at 03:55:00PM +0100, Hans de Goede wrote:
>>> On 11/16/21 08:41, kernel test robot wrote:
>>>> FYI, we noticed the following commit (built with gcc-9):
>>>>
>>>> commit: 995fe757ecaeac44e023458af64d27655f9dbf73 ("[PATCH] device property: Check fwnode->secondary when finding properties")
>>>> url: https://github.com/0day-ci/linux/commits/Daniel-Scally/device-property-Check-fwnode-secondary-when-finding-properties/20211114-044259
>>>> base: https://git.kernel.org/cgit/linux/kernel/git/gregkh/driver-core.git b5013d084e03e82ceeab4db8ae8ceeaebe76b0eb
>>>> patch link: https://lore.kernel.org/lkml/20211113204141.520924-1-djrscally@xxxxxxxxx
>>>>
>>>> in testcase: boot
>>>>
>>>> on test machine: qemu-system-i386 -enable-kvm -cpu SandyBridge -smp 2 -m 4G
>>>>
>>>> caused below changes (please refer to attached dmesg/kmsg for entire log/backtrace):
>>>>
>>>>
>>>> +---------------------------------------------+------------+------------+
>>>> | | b5013d084e | 995fe757ec |
>>>> +---------------------------------------------+------------+------------+
>>>> | boot_successes | 23 | 0 |
>>>> | boot_failures | 0 | 22 |
>>>> | BUG:kernel_NULL_pointer_dereference,address | 0 | 22 |
>>>> | Oops:#[##] | 0 | 22 |
>>>> | EIP:fwnode_property_get_reference_args | 0 | 22 |
>>>> | Kernel_panic-not_syncing:Fatal_exception | 0 | 22 |
>>>> +---------------------------------------------+------------+------------+
>>>>
>>>>
>>>> If you fix the issue, kindly add following tag
>>>> Reported-by: kernel test robot <oliver.sang@xxxxxxxxx>
>>> Ok, so this patch likely needs a v2 which changes the if to this:
>>>
>>> if (ret == -EINVAL && !IS_ERR_OR_NULL(fwnode) &&
>>> !IS_ERR_OR_NULL(fwnode->secondary))
>>> ret = fwnode_call_int_op(fwnode->secondary, get_reference_args,
>>> prop, nargs_prop, nargs, index, args);
>>>
>>>
>>> So that we check fwnode before dereferencing it, note this also changes the
>>> (ret < 0) check to (ret == -EINVAL), this makes the secondary node handling
>>> identical to fwnode_property_read_int_array() and
>>> fwnode_property_read_string_array()
>>>
>>> Danny, can you send a v2 with this change please?
>> Hmm... So, you are suggesting that we need to check it only for EINVAL and
>> ENOENT in this case the one that brings us to the NULL pointer dereference.
>> But I don't understand what's the difference here.
>
>
> Sticking point; the ACPI version of .get_reference_args() returns
> -ENOENT (converted from -EINVAL [1]) if the property you ask for doesn't
> exist against that fwnode, which unless I'm missing something means this
> won't work in our use case. This confused me for a while because we
> definitely call fwnode_property_read_int_array() in sensor driver probes
> through v4l2_fwnode_endpoint_alloc_parse(), but it turns out the ACPI
> version of _that_ operation has no matching conversion of the error
> code, so when that fails to find the property it sends back -EINVAL and
> so the form that exists in fwnode_property_read_int_array() currently
> works fine.
>
>
> We could align them all to if (ret < 0 && !IS_ERR_OR_NULL(fwnode) &&
> !IS_ERR_OR_NULL(fwnode->secondary)). This is probably my preferred
> option, because I can't really see why we'd only want to do the
> secondary check on -EINVAL anyway - but maybe I miss something here.
> Alternatively we can take Hans suggestion so they all match the existing
> code, but this means we have to handle that conversion first - I
> couldn't see from a cursory look that any of the direct callers check
> the value of the return beyond "is it 0?", but of course it could be
> done somewhere in calls to the fwnode->ops->get_reference_args()
> callback instead.
>
>
> Thoughts?

I missed that just checking for -EINVAL will not work for the ipu3 case
(I did not test) in that case I think using "ret < 0" as check instead
is probably fine for this patch.

As for modifying the existing 2 code paths, IMHO it does make sense
to try and preserve the error code (and not try the secondary fwnode)
when the error is an error other then the one indicating the property
is not there.

So keeping those as -EINVAL is probably best and maybe for the
the fwnode_find_reference instead of (ret < 0) use:
(ret == -EINVAL || ret == -ENOENT) ?

Regards,

Hans


>>>> [ 17.327851][ T7] BUG: kernel NULL pointer dereference, address: 00000000
>>>> [ 17.329758][ T7] #PF: supervisor read access in kernel mode
>>>> [ 17.331371][ T7] #PF: error_code(0x0000) - not-present page
>>>> [ 17.332992][ T7] *pde = 00000000
>>>> [ 17.334107][ T7] Oops: 0000 [#1] PREEMPT
>>>> [ 17.335310][ T7] CPU: 0 PID: 7 Comm: kworker/u2:0 Tainted: G S 5.15.0-11191-g995fe757ecae #1
>>>> [ 17.338036][ T7] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-1 04/01/2014
>>>> [ 17.340544][ T7] Workqueue: events_unbound deferred_probe_work_func
>>>> [ 17.342291][ T7] EIP: fwnode_property_get_reference_args (drivers/base/property.c:486 (discriminator 1))
>>>> [ 17.344051][ T7] Code: 8b 45 0c 50 8b 45 08 50 89 d8 89 55 f4 ff d6 83 c4 0c 89 c6 85 c0 78 55 8d 65 f8 89 f0 5b 5e 5d c3 8d 74 26 00 be fa ff ff ff <8b> 03 85 c0 74 e8 3d 00 f0 ff ff 77 e1 8b 58 04 85 db 74 37 8b 5b
>>>> All code
>>>> ========
>>>> 0: 8b 45 0c mov 0xc(%rbp),%eax
>>>> 3: 50 push %rax
>>>> 4: 8b 45 08 mov 0x8(%rbp),%eax
>>>> 7: 50 push %rax
>>>> 8: 89 d8 mov %ebx,%eax
>>>> a: 89 55 f4 mov %edx,-0xc(%rbp)
>>>> d: ff d6 callq *%rsi
>>>> f: 83 c4 0c add $0xc,%esp
>>>> 12: 89 c6 mov %eax,%esi
>>>> 14: 85 c0 test %eax,%eax
>>>> 16: 78 55 js 0x6d
>>>> 18: 8d 65 f8 lea -0x8(%rbp),%esp
>>>> 1b: 89 f0 mov %esi,%eax
>>>> 1d: 5b pop %rbx
>>>> 1e: 5e pop %rsi
>>>> 1f: 5d pop %rbp
>>>> 20: c3 retq
>>>> 21: 8d 74 26 00 lea 0x0(%rsi,%riz,1),%esi
>>>> 25: be fa ff ff ff mov $0xfffffffa,%esi
>>>> 2a:* 8b 03 mov (%rbx),%eax <-- trapping instruction
>>>> 2c: 85 c0 test %eax,%eax
>>>> 2e: 74 e8 je 0x18
>>>> 30: 3d 00 f0 ff ff cmp $0xfffff000,%eax
>>>> 35: 77 e1 ja 0x18
>>>> 37: 8b 58 04 mov 0x4(%rax),%ebx
>>>> 3a: 85 db test %ebx,%ebx
>>>> 3c: 74 37 je 0x75
>>>> 3e: 8b .byte 0x8b
>>>> 3f: 5b pop %rbx
>>>>
>>>> Code starting with the faulting instruction
>>>> ===========================================
>>>> 0: 8b 03 mov (%rbx),%eax
>>>> 2: 85 c0 test %eax,%eax
>>>> 4: 74 e8 je 0xffffffffffffffee
>>>> 6: 3d 00 f0 ff ff cmp $0xfffff000,%eax
>>>> b: 77 e1 ja 0xffffffffffffffee
>>>> d: 8b 58 04 mov 0x4(%rax),%ebx
>>>> 10: 85 db test %ebx,%ebx
>>>> 12: 74 37 je 0x4b
>>>> 14: 8b .byte 0x8b
>>>> 15: 5b pop %rbx
>>>> [ 17.350847][ T7] EAX: 00000000 EBX: 00000000 ECX: 00000000 EDX: c37cd6d8
>>>> [ 17.352783][ T7] ESI: ffffffea EDI: f5b5a400 EBP: c4cffd24 ESP: c4cffd14
>>>> [ 17.354673][ T7] DS: 007b ES: 007b FS: 0000 GS: 0000 SS: 0068 EFLAGS: 00010246
>>>> [ 17.362075][ T7] CR0: 80050033 CR2: 00000000 CR3: 04206000 CR4: 00000690
>>>> [ 17.363993][ T7] Call Trace:
>>>> [ 17.365018][ T7] fwnode_find_reference (drivers/base/property.c:514)
>>>> [ 17.366430][ T7] ? __this_cpu_preempt_check (lib/smp_processor_id.c:67)
>>>> [ 17.367825][ T7] ? lockdep_init_map_type (kernel/locking/lockdep.c:4813)
>>>> [ 17.369325][ T7] ? phylink_run_resolve+0x20/0x20
>>>> [ 17.370897][ T7] ? init_timer_key (kernel/time/timer.c:818)
>>>> [ 17.372228][ T7] fwnode_get_phy_node (drivers/net/phy/phy_device.c:2986)
>>>> [ 17.373574][ T7] phylink_fwnode_phy_connect (drivers/net/phy/phylink.c:1180 drivers/net/phy/phylink.c:1166)
>>>> [ 17.375014][ T7] phylink_of_phy_connect (drivers/net/phy/phylink.c:1152)
>>>> [ 17.376373][ T7] dsa_slave_create (net/dsa/slave.c:1889 net/dsa/slave.c:2036)
>>>> [ 17.377765][ T7] dsa_tree_setup_switches (net/dsa/dsa2.c:477 net/dsa/dsa2.c:977)
>>>> [ 17.379282][ T7] dsa_register_switch (net/dsa/dsa2.c:1065 net/dsa/dsa2.c:1565 net/dsa/dsa2.c:1579)
>>>> [ 17.380762][ T7] dsa_loop_drv_probe (drivers/net/dsa/dsa_loop.c:333)
>>>> [ 17.382137][ T7] mdio_probe (drivers/net/phy/mdio_device.c:157)
>