[RFC PATCH v1] powerpc: Enable KFENCE for PPC32

Wed Mar 3 21:38:58 AEDT 2021

Le 02/03/2021 à 12:39, Marco Elver a écrit :
> On Tue, 2 Mar 2021 at 12:21, Christophe Leroy
> <christophe.leroy at csgroup.eu> wrote:
> [...]
>>>> Booting with 'no_hash_pointers" I get the following. Does it helps ?
>>>>
>>>> [   16.837198] ==================================================================
>>>> [   16.848521] BUG: KFENCE: invalid read in finish_task_switch.isra.0+0x54/0x23c
>>>> [   16.848521]
>>>> [   16.857158] Invalid read at 0xdf98800a:
>>>> [   16.861004]  finish_task_switch.isra.0+0x54/0x23c
>>>> [   16.865731]  kunit_try_run_case+0x5c/0xd0
>>>> [   16.869780]  kunit_generic_run_threadfn_adapter+0x24/0x30
>>>> [   16.875199]  kthread+0x15c/0x174
>>>> [   16.878460]  ret_from_kernel_thread+0x14/0x1c
>>>> [   16.882847]
>>>> [   16.884351] CPU: 0 PID: 111 Comm: kunit_try_catch Tainted: G    B
>>>> 5.12.0-rc1-s3k-dev-01534-g4f14ae75edf0-dirty #4674
>>>> [   16.895908] NIP:  c016eb8c LR: c02f50dc CTR: c016eb38
>>>> [   16.900963] REGS: e2449d90 TRAP: 0301   Tainted: G    B
>>>> (5.12.0-rc1-s3k-dev-01534-g4f14ae75edf0-dirty)
>>>> [   16.911386] MSR:  00009032 <EE,ME,IR,DR,RI>  CR: 22000004  XER: 00000000
>>>> [   16.918153] DAR: df98800a DSISR: 20000000
>>>> [   16.918153] GPR00: c02f50dc e2449e50 c1140d00 e100dd24 c084b13c 00000008 c084b32b c016eb38
>>>> [   16.918153] GPR08: c0850000 df988000 c0d10000 e2449eb0 22000288
>>>> [   16.936695] NIP [c016eb8c] test_invalid_access+0x54/0x108
>>>> [   16.942125] LR [c02f50dc] kunit_try_run_case+0x5c/0xd0
>>>> [   16.947292] Call Trace:
>>>> [   16.949746] [e2449e50] [c005a5ec] finish_task_switch.isra.0+0x54/0x23c (unreliable)
>>>
>>> The "(unreliable)" might be a clue that it's related to ppc32 stack
>>> unwinding. Any ppc expert know what this is about?
>>>
>>>> [   16.957443] [e2449eb0] [c02f50dc] kunit_try_run_case+0x5c/0xd0
>>>> [   16.963319] [e2449ed0] [c02f63ec] kunit_generic_run_threadfn_adapter+0x24/0x30
>>>> [   16.970574] [e2449ef0] [c004e710] kthread+0x15c/0x174
>>>> [   16.975670] [e2449f30] [c001317c] ret_from_kernel_thread+0x14/0x1c
>>>> [   16.981896] Instruction dump:
>>>> [   16.984879] 8129d608 38e7eb38 81020280 911f004c 39000000 995f0024 907f0028 90ff001c
>>>> [   16.992710] 3949000a 915f0020 3d40c0d1 3d00c085 <8929000a> 3908adb0 812a4b98 3d40c02f
>>>> [   17.000711] ==================================================================
>>>> [   17.008223]     # test_invalid_access: EXPECTATION FAILED at mm/kfence/kfence_test.c:636
>>>> [   17.008223]     Expected report_matches(&expect) to be true, but is false
>>>> [   17.023243]     not ok 21 - test_invalid_access
>>>
>>> On a fault in test_invalid_access, KFENCE prints the stack trace based
>>> on the information in pt_regs. So we do not think there's anything we
>>> can do to improve stack printing pe-se.
>>
>> stack printing, probably not. Would be good anyway to mark the last level [unreliable] as the ppc does.
> 
> We use stack_trace_save_regs() + stack_trace_print().
> 
>> IIUC, on ppc the address in the stack frame of the caller is written by the caller. In most tests,
>> there is some function call being done before the fault, for instance
>> test_kmalloc_aligned_oob_read() does a call to kunit_do_assertion which populates the address of the
>> call in the stack. However this is fragile.
> 
> Interesting, this might explain it.
> 
>> This works for function calls because in order to call a subfunction, a function has to set up a
>> stack frame in order to same the value in the Link Register, which contains the address of the
>> function's parent and that will be clobbered by the sub-function call.
>>
>> However, it cannot be done by exceptions, because exceptions can happen in a function that has no
>> stack frame (because that function has no need to call a subfunction and doesn't need to same
>> anything on the stack). If the exception handler was writting the caller's address in the stack
>> frame, it would in fact write it in the parent's frame, leading to a mess.
>>
>> But in fact the information is in pt_regs, it is in regs->nip so KFENCE should be able to use that
>> instead of the stack.
> 
> Perhaps stack_trace_save_regs() needs fixing for ppc32? Although that
> seems to use arch_stack_walk().
> 
>>> What's confusing is that it's only this test, and none of the others.
>>> Given that, it might be code-gen related, which results in some subtle
>>> issue with stack unwinding. There are a few things to try, if you feel
>>> like it:
>>>
>>> -- Change the unwinder, if it's possible for ppc32.
>>
>> I don't think it is possible.
>>
>>>
>>> -- Add code to test_invalid_access(), to get the compiler to emit
>>> different code. E.g. add a bunch (unnecessary) function calls, or add
>>> barriers, etc.
>>
>> The following does the trick
>>
>> diff --git a/mm/kfence/kfence_test.c b/mm/kfence/kfence_test.c
>> index 4acf4251ee04..22550676cd1f 100644
>> --- a/mm/kfence/kfence_test.c
>> +++ b/mm/kfence/kfence_test.c
>> @@ -631,8 +631,11 @@ static void test_invalid_access(struct kunit *test)
>>                  .addr = &__kfence_pool[10],
>>                  .is_write = false,
>>          };
>> +       char *buf;
>>
>> +       buf = test_alloc(test, 4, GFP_KERNEL, ALLOCATE_RIGHT);
>>          READ_ONCE(__kfence_pool[10]);
>> +       test_free(buf);
>>          KUNIT_EXPECT_TRUE(test, report_matches(&expect));
>>    }
>>
>>
>> But as I said above, this is fragile. If for some reason one day test_alloc() gets inlined, it may
>> not work anymore.
> 
> Yeah, obviously that's hack, but interesting nevertheless.
> 
> Based on what you say above, however, it seems that
> stack_trace_save_regs()/arch_stack_walk() don't exactly do what they
> should? Can they be fixed for ppc32?

Can we really consider they don't do what they should ?

I have the feeling that excepting entry[0] of the stack trace to match the instruction pointer is 
not a valid expectation. That's probably correct on architectures that always have a stack frame for 
any function, but for powerpc who can have frameless functions, we can't expect that I think.

I have proposed a change to KFENCE in another response to this mail thread, could it be the solution ?

Thanks
Christophe