]> git.ipfire.org Git - thirdparty/kernel/stable.git/commit
powerpc/watchpoint: Restore NV GPRs while returning from exception
authorRavi Bangoria <ravi.bangoria@linux.ibm.com>
Thu, 13 Jun 2019 03:30:14 +0000 (09:00 +0530)
committerMichael Ellerman <mpe@ellerman.id.au>
Wed, 19 Jun 2019 10:05:08 +0000 (20:05 +1000)
commitf474c28fbcbe42faca4eb415172c07d76adcb819
treed952f846aabad191612c4978f03568be5f6f06bf
parent1b7de1df997a57e3e144bdb31ceaee9c16d6d284
powerpc/watchpoint: Restore NV GPRs while returning from exception

powerpc hardware 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. eg, 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->prev:

  # perf record -e mem:0xc00000000121c8c0

Run some workload such that new kthread gets invoked. eg, 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

List corruption happened because it uses 'load into non-volatile
register' instruction:

Snippet from __kthread_create_on_node:

  c000000000136be8:     addis   r29,r2,-19
  c000000000136bec:     ld      r29,31424(r29)
        if (!__list_add_valid(new, prev, next))
  c000000000136bf0:     mr      r3,r30
  c000000000136bf4:     mr      r5,r28
  c000000000136bf8:     mr      r4,r29
  c000000000136bfc:     bl      c00000000059a2f8 <__list_add_valid+0x8>

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

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.

Fixes: 5aae8a5370802 ("powerpc, hw_breakpoints: Implement hw_breakpoints for 64-bit server processors")
Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Cc: stable@vger.kernel.org # 2.6.36+
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
arch/powerpc/kernel/exceptions-64s.S