GCC 10 optimizes the scheduler code differently than its predecessors.

When CONFIG_DEBUG_SECTION_MISMATCH=y, the Makefile forces GCC not
to inline some functions (-fno-inline-functions-called-once). Before GCC
10, "no-inlined" __schedule() starts with the usual prologue:

  push %bp
  mov %sp, %bp

So the ORC unwinder simply picks stack pointer from %bp and
unwinds from __schedule() just perfectly:

  $ cat /proc/1/stack
  [<0>] ep_poll+0x3e9/0x450
  [<0>] do_epoll_wait+0xaa/0xc0
  [<0>] __x64_sys_epoll_wait+0x1a/0x20
  [<0>] do_syscall_64+0x33/0x40
  [<0>] entry_SYSCALL_64_after_hwframe+0x44/0xa9

But now, with GCC 10, there is no %bp prologue in __schedule():

  $ cat /proc/1/stack
  <nothing>

The ORC entry of the point in __schedule() is:

  sp:sp+88 bp:last_sp-48 type:call end:0

In this case, nobody subtracts sizeof "struct inactive_task_frame" in
__unwind_start(). The struct is put on the stack by __switch_to_asm() and
only then __switch_to_asm() stores %sp to task->thread.sp. But we start
unwinding from a point in __schedule() (stored in frame->ret_addr by
'call') and not in __switch_to_asm().

So for these example values in __unwind_start():

  sp=ffff94b50001fdc8 bp=ffff8e1f41d29340 ip=__schedule+0x1f0

The stack is:

  ffff94b50001fdc8: ffff8e1f41578000 # struct inactive_task_frame
  ffff94b50001fdd0: 0000000000000000
  ffff94b50001fdd8: ffff8e1f41d29340
  ffff94b50001fde0: ffff8e1f41611d40 # ...
  ffff94b50001fde8: ffffffff93c41920 # bx
  ffff94b50001fdf0: ffff8e1f41d29340 # bp
  ffff94b50001fdf8: ffffffff9376cad0 # ret_addr (and end of the struct)

0xffffffff9376cad0 is __schedule+0x1f0 (after the call to
__switch_to_asm).  Now follow those 88 bytes from the ORC entry (sp+88).
The entry is correct, __schedule() really pushes 48 bytes (8*7) + 32 bytes
via subq to store some local values (like 4U below). So to unwind, look
at the offset 88-sizeof(long) = 0x50 from here:

  ffff94b50001fe00: ffff8e1f41578618
  ffff94b50001fe08: 00000cc000000255
  ffff94b50001fe10: 0000000500000004
  ffff94b50001fe18: 7793fab6956b2d00 # NOTE (see below)
  ffff94b50001fe20: ffff8e1f41578000
  ffff94b50001fe28: ffff8e1f41578000
  ffff94b50001fe30: ffff8e1f41578000
  ffff94b50001fe38: ffff8e1f41578000
  ffff94b50001fe40: ffff94b50001fed8
  ffff94b50001fe48: ffff8e1f41577ff0
  ffff94b50001fe50: ffffffff9376cf12

Here                ^^^^^^^^^^^^^^^^ is the correct ret addr from
__schedule(). It translates to schedule+0x42 (insn after a call to
__schedule()).

BUT, unwind_next_frame() tries to take the address starting from
0xffff94b50001fdc8. That is exactly from thread.sp+88-sizeof(long) =
0xffff94b50001fdc8+88-8 = 0xffff94b50001fe18, which is garbage marked as
NOTE above. So this quits the unwinding as 7793fab6956b2d00 is obviously
not a kernel address.

There was a fix to skip 'struct inactive_task_frame' in
unwind_get_return_address_ptr in the following commit:

  187b96db ("x86/unwind/orc: Fix unwind_get_return_address_ptr() for inactive tasks")

But we need to skip the struct already in the unwinder proper. So
subtract the size (increase the stack pointer) of the structure in
__unwind_start() directly. This allows for removal of the code added by
commit 187b96db completely, as the address is now at
'(unsigned long *)state->sp - 1', the same as in the generic case.

[ mingo: Cleaned up the changelog a bit, for better readability. ]

Fixes: ee9f8fce ("x86/unwind: Add the ORC unwinder")
Bug: https://bugzilla.suse.com/show_bug.cgi?id=1176907Signed-off-by: Jiri Slaby <jslaby@suse.cz>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20201014053051.24199-1-jslaby@suse.cz