[PATCH] Powerpc/Watchpoint: Restore nvgprs while returning from exception

[Ravi Bangoria]

On 6/7/19 11:20 AM, Michael Ellerman wrote:
> Ravi Bangoria <ravi.bangoria at linux.ibm.com> writes:
> 
>> Powerpc hw triggers watchpoint before executing the instruction.
>> To make trigger-after-execute behavior, kernel emulates the
>> instruction. If the instruction is 'load something into non-
>> volatile register', exception handler should restore emulated
>> register state while returning back, otherwise there will be
>> register state corruption. Ex, Adding a watchpoint on a list
>> can corrput the list:
>>
>>   # cat /proc/kallsyms | grep kthread_create_list
>>   c00000000121c8b8 d kthread_create_list
>>
>> Add watchpoint on kthread_create_list->next:
>>
>>   # perf record -e mem:0xc00000000121c8c0
>>
>> Run some workload such that new kthread gets invoked. Ex, I
>> just logged out from console:
>>
>>   list_add corruption. next->prev should be prev (c000000001214e00), \
>> 	but was c00000000121c8b8. (next=c00000000121c8b8).
>>   WARNING: CPU: 59 PID: 309 at lib/list_debug.c:25 __list_add_valid+0xb4/0xc0
>>   CPU: 59 PID: 309 Comm: kworker/59:0 Kdump: loaded Not tainted 5.1.0-rc7+ #69
>>   ...
>>   NIP __list_add_valid+0xb4/0xc0
>>   LR __list_add_valid+0xb0/0xc0
>>   Call Trace:
>>   __list_add_valid+0xb0/0xc0 (unreliable)
>>   __kthread_create_on_node+0xe0/0x260
>>   kthread_create_on_node+0x34/0x50
>>   create_worker+0xe8/0x260
>>   worker_thread+0x444/0x560
>>   kthread+0x160/0x1a0
>>   ret_from_kernel_thread+0x5c/0x70
> 
> This all depends on what code the compiler generates for the list
> access.

True. list corruption is just an example. But any load instruction that uses
non-volatile register and hits a watchpoint, will result in register state
corruption.

> Can you include a disassembly of the relevant code in your
> kernel so we have an example of the bad case.

Register state from WARN_ON():

  GPR00: c00000000059a3a0 c000007ff23afb50 c000000001344e00 0000000000000075
  GPR04: 0000000000000000 0000000000000000 0000001852af8bc1 0000000000000000
  GPR08: 0000000000000001 0000000000000007 0000000000000006 00000000000004aa
  GPR12: 0000000000000000 c000007ffffeb080 c000000000137038 c000005ff62aaa00
  GPR16: 0000000000000000 0000000000000000 c000007fffbe7600 c000007fffbe7370
  GPR20: c000007fffbe7320 c000007fffbe7300 c000000001373a00 0000000000000000
  GPR24: fffffffffffffef7 c00000000012e320 c000007ff23afcb0 c000000000cb8628
  GPR28: c00000000121c8b8 c000000001214e00 c000007fef5b17e8 c000007fef5b17c0

Snippet from __kthread_create_on_node:

  c000000000136be8:       ed ff a2 3f     addis   r29,r2,-19
  c000000000136bec:       c0 7a bd eb     ld      r29,31424(r29)
          if (!__list_add_valid(new, prev, next))
  c000000000136bf0:       78 f3 c3 7f     mr      r3,r30
  c000000000136bf4:       78 e3 85 7f     mr      r5,r28
  c000000000136bf8:       78 eb a4 7f     mr      r4,r29
  c000000000136bfc:       fd 36 46 48     bl      c00000000059a2f8 <__list_add_valid+0x8>

Watchpoint hit at 0xc000000000136bec. 

  addis   r29,r2,-19
   => r29 = 0xc000000001344e00 + (-19 << 16)
   => r29 = 0xc000000001214e00

  ld      r29,31424(r29)
   => r29 = *(0xc000000001214e00 + 31424)
   => r29 = *(0xc00000000121c8c0)

0xc00000000121c8c0 is where we placed a watchpoint and thus this instruction was
emulated by emulate_step. But because handle_dabr_fault did not restore emulated
register state, r29 still contains stale value in above register state.