On Tue, Jun 23, 2020 at 05:55:57PM +0100, Mark Rutland wrote:
On Tue, Jun 23, 2020 at 04:59:01PM +0100, Will Deacon wrote:
On Thu, Jun 18, 2020 at 01:29:29PM -0400, Paul Gortmaker wrote:
BUG: sleeping function called from invalid context at kernel/locking/rtmutex.c:975 in_atomic(): 1, irqs_disabled(): 0, pid: 35658, name: gdbtest Preemption disabled at: [<ffff000010081578>] do_debug_exception+0x38/0x1a4 Call trace: dump_backtrace+0x0/0x138 show_stack+0x24/0x30 dump_stack+0x94/0xbc ___might_sleep+0x13c/0x168 rt_spin_lock+0x40/0x80 do_force_sig_info+0x30/0xe0 force_sig_fault+0x64/0x90 arm64_force_sig_fault+0x50/0x80 send_user_sigtrap+0x50/0x80 brk_handler+0x98/0xc8 do_debug_exception+0x70/0x1a4 el0_dbg+0x18/0x20
The reproducer was basically an automated gdb test that set a breakpoint on a simple "hello world" program and then quit gdb once the breakpoint was hit - i.e. "(gdb) A debugging session is active. Quit anyway? "
Hmm, the debug exception handler path was definitely written with the expectation that preemption is disabled, so this is unfortunate. For exceptions from kernelspace, we need to keep that guarantee as we implement things like BUG() using this path. For exceptions from userspace, it's plausible that we could re-enable preemption, but then we should also re-enable interrupts and debug exceptions too because we don't context-switch pstate in switch_to() and we would end up with holes in our kernel debug coverage (and these might be fatal if e.g. single step doesn't work in a kprobe OOL buffer). However, that then means that any common code when handling user and kernel debug exceptions needs to be re-entrant, which it probably isn't at the moment (I haven't checked).
I'm pretty certain existing code is not reentrant, and regardless it's going to be a mess to reason about this generally if we have to undo our strict exception nesting rules.
Are these rules written down somewhere? I'll need to update them if we get this working for preempt-rt (and we should try to do that).
I reckon we need to treat this like an NMI instead -- is that plausible?
I don't think so. It's very much a synchronous exception, and delivering a signal to the exceptional context doesn't feel like an NMI to me. There's also a fair amount of code that can run in debug context (hw_breakpoint, kprobes, uprobes, kasan) which might not be happy to suddenly be in an NMI-like environment. Furthermore, the masking rules are different depending on what triggers the exception.
One of the things I've started looking at is ripping out our dodgy hw_breakpoint code so that kernel debug exceptions are easier to reason about. Specifically, I think we end up with something like:
- On taking a non-debug exception from EL0, unmask D as soon as we can.
- On taking a debug exception from EL0, unmask {D,I} and invoke user handlers. I think this always means SIGTRAP, apart from uprobes. This will mean making those paths preemptible, as I don't think they are right now (e.g. traversing the callback hooks uses an RCU-protected list).
- On taking a non-debug, non-fatal synchronous exception from EL1, unmask D as soon as we can (i.e. we step into these exceptions). Fatal exceptions can obviously leave D masked.
- On taking an interrupt from EL1, stash MDSCR_EL1.SS in a pcpu variable and clear the register bit if it was set. Then unmask only D and leave I set. On return from the exception, set D and restore MDSCR_EL1.SS. If we decide to reschedule, unmask D (i.e. we only step into interrupts if we need a reschedule. Alternatively, we could skip the reschedule if we were stepping.)
- On taking a debug exception from EL1, leave {D,I} set. Watchpoints on uaccess are silently stepped over.
Thoughts? We could probably simplify this if we could state that stepping an instruction in kernel space could only ever be interrupted by an interrupt. That's probably true for kprobes, but relying on it feels like it might bite us later on.
Will