Re: Lockup with --accel tcg,thread=single

[Alex Bennée]

Peter Maydell <address@hidden> writes:

> On Mon, 30 Sep 2019 at 14:17, Doug Gale <address@hidden> wrote:
>>
>> I found a lockup in single threaded TCG, with OVMF bios, 16 CPUs.
>>
>> qemu-system-x86_64 --accel tcg,thread=single -smp cpus=16 -bios
>> /usr/share/ovmf/OVMF.fd
>>
>> Using Ubuntu 18.04 LTS, default gnome desktop. There is no guest OS,
>> there is no hard drive, just the OVMF firmware locks it up. (I
>> narrowed it down to this from a much larger repro)
>
>> Peter Maydell helped me bisect it in IRC.
>>  Works fine at commit 1e8a98b53867f61
>>  Fails at commit 9458a9a1df1a4c7
>>
>> MTTCG works fine all the way up to master.
>
> Thanks for this bug report. I've reproduced it and think
> I have figured out what is going on here.
>
> Commit 9458a9a1df1a4c719e245 is Paolo's commit that fixes the
> TCG-vs-dirty-bitmap race by having the thread which is
> doing a memory access wait for the vCPU thread to finish
> its current TB using a no-op run_on_cpu() operation.
>
> In the case of the hang the thread doing the memory access
> is the iothread, like this:
>
> #14 0x000055c150c0a98c in run_on_cpu (cpu=0x55c153801c60,
> func=0x55c150bbb542 <do_nothing>, data=...)
>     at /home/petmay01/linaro/qemu-from-laptop/qemu/cpus.c:1205
> #15 0x000055c150bbb58c in tcg_log_global_after_sync
> (listener=0x55c1538410c8) at
> /home/petmay01/linaro/qemu-from-laptop/qemu/exec.c:2963
> #16 0x000055c150c1fe18 in memory_global_after_dirty_log_sync () at
> /home/petmay01/linaro/qemu-from-laptop/qemu/memory.c:2598
> #17 0x000055c150c1e6b8 in memory_region_snapshot_and_clear_dirty
> (mr=0x55c1541e82b0, addr=0, size=1920000, client=0)
>     at /home/petmay01/linaro/qemu-from-laptop/qemu/memory.c:2106
> #18 0x000055c150c76c05 in vga_draw_graphic (s=0x55c1541e82a0, full_update=0)
>     at /home/petmay01/linaro/qemu-from-laptop/qemu/hw/display/vga.c:1661
> #19 0x000055c150c771c4 in vga_update_display (opaque=0x55c1541e82a0)
> at /home/petmay01/linaro/qemu-from-laptop/qemu/hw/display/vga.c:1785
> #20 0x000055c151052a83 in graphic_hw_update (con=0x55c1536dfaa0) at
> /home/petmay01/linaro/qemu-from-laptop/qemu/ui/console.c:268
> #21 0x000055c151091490 in gd_refresh (dcl=0x55c1549af090) at
> /home/petmay01/linaro/qemu-from-laptop/qemu/ui/gtk.c:484
> #22 0x000055c151056571 in dpy_refresh (s=0x55c1542f9d90) at
> /home/petmay01/linaro/qemu-from-laptop/qemu/ui/console.c:1629
> #23 0x000055c1510527f0 in gui_update (opaque=0x55c1542f9d90) at
> /home/petmay01/linaro/qemu-from-laptop/qemu/ui/console.c:206
> #24 0x000055c1511ee67c in timerlist_run_timers
> (timer_list=0x55c15370c280) at
> /home/petmay01/linaro/qemu-from-laptop/qemu/util/qemu-timer.c:592
> #25 0x000055c1511ee726 in qemu_clock_run_timers
> (type=QEMU_CLOCK_REALTIME) at
> /home/petmay01/linaro/qemu-from-laptop/qemu/util/qemu-timer.c:606
> #26 0x000055c1511ee9e5 in qemu_clock_run_all_timers () at
> /home/petmay01/linaro/qemu-from-laptop/qemu/util/qemu-timer.c:692
> #27 0x000055c1511ef181 in main_loop_wait (nonblocking=0) at
> /home/petmay01/linaro/qemu-from-laptop/qemu/util/main-loop.c:524
> #28 0x000055c150db03fe in main_loop () at
> /home/petmay01/linaro/qemu-from-laptop/qemu/vl.c:1806
>
> run_on_cpu() adds the do_nothing function to a cpu queued-work list.
>
> However, the single-threaded TCG runloop qemu_tcg_rr_cpu_thread_fn()
> has the basic structure:
>
>    while (1) {
>        for (each vCPU) {
>            run this vCPU for a timeslice;
>        }
>        qemu_tcg_rr_wait_io_event();
>    }
>
> and it processes cpu work queues only in qemu_tcg_rr_wait_io_event().
> This is a problem, because the thing which causes us to stop running
> one vCPU when its timeslice ends and move on to the next is the
> tcg_kick_vcpu_timer -- and the iothread will never process that timer
> and kick the vcpu because it is currently blocked in run_on_cpu() !
>
> Not sure currently what the best fix is here.

We seem to be repeating ourselves because:

  a8efa60633575a2ee4dbf807a71cb44d44b0e0f8
  Author:     Paolo Bonzini <address@hidden>
  AuthorDate: Wed Nov 14 12:36:57 2018 +0100
  cpus: run work items for all vCPUs if single-threaded

However looking at the commit I can still see we have the problem of not
advancing to the next vCPU if the kick timer (or some other event)
doesn't bring the execution to an exit. I suspect you could get this in
Linux but it's probably sufficiently busy to ensure vCPUs are always
exiting for some reason or another.

So options I can think of so far are:

1. Kick 'em all when not inter-vCPU

Something like this untested patch...

--8<---------------cut here---------------start------------->8---
1 file changed, 17 insertions(+), 5 deletions(-)
cpus.c | 22 +++++++++++++++++-----

modified   cpus.c
@@ -949,8 +949,8 @@ static inline int64_t qemu_tcg_next_kick(void)
     return qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + TCG_KICK_PERIOD;
 }

-/* Kick the currently round-robin scheduled vCPU */
-static void qemu_cpu_kick_rr_cpu(void)
+/* Kick the currently round-robin scheduled vCPU to next */
+static void qemu_cpu_kick_rr_next_cpu(void)
 {
     CPUState *cpu;
     do {
@@ -961,6 +961,16 @@ static void qemu_cpu_kick_rr_cpu(void)
     } while (cpu != atomic_mb_read(&tcg_current_rr_cpu));
 }

+/* Kick all RR vCPUs */
+static void qemu_cpu_kick_rr_cpus(void)
+{
+    CPUState *cpu;
+
+    CPU_FOREACH(cpu) {
+        cpu_exit(cpu);
+    };
+}
+
 static void do_nothing(CPUState *cpu, run_on_cpu_data unused)
 {
 }
@@ -993,7 +1003,7 @@ void qemu_timer_notify_cb(void *opaque, QEMUClockType type)
 static void kick_tcg_thread(void *opaque)
 {
     timer_mod(tcg_kick_vcpu_timer, qemu_tcg_next_kick());
-    qemu_cpu_kick_rr_cpu();
+    qemu_cpu_kick_rr_next_cpu();
 }

 static void start_tcg_kick_timer(void)
@@ -1828,9 +1838,11 @@ void qemu_cpu_kick(CPUState *cpu)
 {
     qemu_cond_broadcast(cpu->halt_cond);
     if (tcg_enabled()) {
+        if (qemu_tcg_mttcg_enabled()) {
         cpu_exit(cpu);
-        /* NOP unless doing single-thread RR */
-        qemu_cpu_kick_rr_cpu();
+        } else {
+            qemu_cpu_kick_rr_cpus();
+        }
     } else {
         if (hax_enabled()) {
             /*
--8<---------------cut here---------------end--------------->8---

2. Add handling of kicking all VCPUs to do_run_on_cpu when current_cpu
== NULL

Which would basically kick all vCPUs when events come from outside the
emulation.

3. Figure out multi-threaded icount and record/replay and drop the
special RR case.

This might take a while.

> Side note -- this use of run_on_cpu() means that we now drop
> the iothread lock within memory_region_snapshot_and_clear_dirty(),
> which we didn't before. This means that a vCPU thread can now
> get in and execute an access to the device registers of
> hw/display/vga.c, updating its state in a way I suspect that the
> device model code is not expecting... So maybe the right answer
> here should be to come up with a fix for the race that 9458a9a1
> addresses that doesn't require us to drop the iothread lock in
> memory_region_snapshot_and_clear_dirty() ? Alternatively we need
> to audit the callers and flag in the documentation that this
> function has the unexpected side effect of briefly dropping the
> iothread lock.

There was a series Emilio posted to get rid of more of the BQL in place
of per-CPU locks which IIRC also stopped some of the bouncing we do in
the *_on_cpu functions. Each time we have to do a lock shuffle to get
things moving is a bit of a red flag.

>
> thanks
> -- PMM


--
Alex Bennée