[PATCH] arm64: Fix early single-stepping

[Will Deacon]

On Mon, Oct 26, 2020 at 06:29:09PM +0100, Jean-Philippe Brucker wrote:
> To use debug features such as single-step, the OS lock must be unlocked
> in the debug registers. Currently this is done in postcore_initcall
> which is now too late.
> 
> Commit 36dadef23fcc ("kprobes: Init kprobes in early_initcall") enabled
> using kprobes from early_initcall, when OS lock is still locked. So when
> kprobe attempts to single-step a patched instruction, instead of
> trapping, execution continues until it throws an undef exception:
> 
> [    0.064233] Kprobe smoke test: started
> [    0.151133] ------------[ cut here ]------------
> [    0.151458] kernel BUG at arch/arm64/kernel/traps.c:406!
> [    0.151812] Internal error: Oops - BUG: 0 [#1] PREEMPT SMP
>               ...
> [    0.162689] Call trace:
> [    0.163014]  do_undefinstr+0x1d4/0x1f4
> [    0.163336]  el1_sync_handler+0xbc/0x140
> [    0.163839]  el1_sync+0x80/0x100
> [    0.164154]  0xffffffc01001d004
> [    0.164527]  init_kprobes+0x13c/0x154
> [    0.164968]  do_one_initcall+0x54/0x2e0
> [    0.165322]  kernel_init_freeable+0xf4/0x258
> [    0.165783]  kernel_init+0x20/0x12c
> [    0.166117]  ret_from_fork+0x10/0x30
> [    0.166595] Code: 97ffff53 a9425bf5 17ffff9b f9001bf7 (d4210000)
> [    0.167084] ---[ end trace 36778fdf576e9a79 ]---
> 
> To fix this, unlock the OS lock as early as possible. Do it in
> traps_init() for CPU0, since KGDB wants to use single-step from that
> point on according to commit b322c65f8ca3 ("arm64: Call
> debug_traps_init() from trap_init() to help early kgdb").
> For secondary CPUs, setup the CPU hotplug handler at early_initcall.

Hmm, does this mean we end up setting MDSCR_EL1.KDE before we've reset the
hardware breakpoint/watchpoint registers? Why do we need kprobes so early?

Will