Re: Crash with CXL + TCG on 8.2: Was Re: qemu cxl memory expander shows numa_node -1

["Alex Bennée" <alex.bennee@linaro.org>]

Peter Maydell <peter.maydell@linaro.org> writes:

> On Thu, 1 Feb 2024 at 14:01, Jonathan Cameron
> <Jonathan.Cameron@huawei.com> wrote:
>> > Can you run QEMU under gdb and give the backtrace when it stops
>> > on the abort() ? That will probably have a helpful clue. I
>> > suspect something is failing to pass a valid retaddr in
>> > when it calls a load/store function.
>
>> [Switching to Thread 0x7ffff56ff6c0 (LWP 21916)]
>> __pthread_kill_implementation (no_tid=0, signo=6, threadid=<optimized out>) at ./nptl/pthread_kill.c:44
>> 44      ./nptl/pthread_kill.c: No such file or directory.
>> (gdb) bt
>> #0  __pthread_kill_implementation (no_tid=0, signo=6, threadid=<optimized out>) at ./nptl/pthread_kill.c:44
>> #1  __pthread_kill_internal (signo=6, threadid=<optimized out>) at ./nptl/pthread_kill.c:78
>> #2  __GI___pthread_kill (threadid=<optimized out>, signo=signo@entry=6) at ./nptl/pthread_kill.c:89
>> #3  0x00007ffff77c43b6 in __GI_raise (sig=sig@entry=6) at ../sysdeps/posix/raise.c:26
>> #4  0x00007ffff77aa87c in __GI_abort () at ./stdlib/abort.c:79
>> #5  0x0000555555c0d4ce in cpu_abort
>>     (cpu=cpu@entry=0x555556fd9000, fmt=fmt@entry=0x555555fe3378 "cpu_io_recompile: could not find TB for pc=%p")
>>     at ../../cpu-target.c:359
>> #6  0x0000555555c59435 in cpu_io_recompile (cpu=cpu@entry=0x555556fd9000, retaddr=retaddr@entry=0) at ../../accel/tcg/translate-all.c:611
>> #7  0x0000555555c5c956 in io_prepare
>>     (retaddr=0, addr=19595792376, attrs=..., xlat=<optimized out>, cpu=0x555556fd9000, out_offset=<synthetic pointer>)
>>     at ../../accel/tcg/cputlb.c:1339
<snip>
>> #21 tb_htable_lookup (cpu=<optimized out>, pc=pc@entry=18446744072116178925, cs_base=0, flags=415285936, cflags=4278353920)
>>     at ../../accel/tcg/cpu-exec.c:231
>> #22 0x0000555555c50c08 in tb_lookup
>>     (cpu=cpu@entry=0x555556fd9000, pc=pc@entry=18446744072116178925, cs_base=cs_base@entry=0, flags=<optimized out>, cflags=<optimized out>) at ../../accel/tcg/cpu-exec.c:267
>> #23 0x0000555555c51e23 in helper_lookup_tb_ptr (env=0x555556fdb7c0) at ../../accel/tcg/cpu-exec.c:423
>> #24 0x00007fffa9076ead in code_gen_buffer ()
>> #25 0x0000555555c50fab in cpu_tb_exec (cpu=cpu@entry=0x555556fd9000, itb=<optimized out>, tb_exit=tb_exit@entry=0x7ffff56fe708)
>>     at ../../accel/tcg/cpu-exec.c:458
>> #26 0x0000555555c51492 in cpu_loop_exec_tb
>>     (tb_exit=0x7ffff56fe708, last_tb=<synthetic pointer>, pc=18446744072116179169, tb=<optimized out>, cpu=0x555556fd9000)
>>     at ../../accel/tcg/cpu-exec.c:920
>> #27 cpu_exec_loop (cpu=cpu@entry=0x555556fd9000, sc=sc@entry=0x7ffff56fe7a0) at ../../accel/tcg/cpu-exec.c:1041
>> #28 0x0000555555c51d11 in cpu_exec_setjmp (cpu=cpu@entry=0x555556fd9000, sc=sc@entry=0x7ffff56fe7a0) at ../../accel/tcg/cpu-exec.c:1058
>> #29 0x0000555555c523b4 in cpu_exec (cpu=cpu@entry=0x555556fd9000) at ../../accel/tcg/cpu-exec.c:1084
>> #30 0x0000555555c74053 in tcg_cpus_exec (cpu=cpu@entry=0x555556fd9000) at ../../accel/tcg/tcg-accel-ops.c:76
>> #31 0x0000555555c741a0 in mttcg_cpu_thread_fn (arg=arg@entry=0x555556fd9000) at ../../accel/tcg/tcg-accel-ops-mttcg.c:95
>> #32 0x0000555555dfb580 in qemu_thread_start (args=0x55555703c3e0) at ../../util/qemu-thread-posix.c:541
>> #33 0x00007ffff78176ba in start_thread (arg=<optimized out>) at ./nptl/pthread_create.c:444
>> #34 0x00007ffff78a60d0 in clone3 () at ../sysdeps/unix/sysv/linux/x86_64/clone3.S:81
>
> So, that looks like:
>  * we call cpu_tb_exec(), which executes some generated code
>  * that generated code calls the lookup_tb_ptr helper to see
>    if we have a generated TB already for the address we're going
>    to execute next
>  * lookup_tb_ptr probes the TLB to see if we know the host RAM
>    address for the guest address
>  * this results in a TLB walk for an instruction fetch
>  * the page table descriptor load is to IO memory
>  * io_prepare assumes it needs to do a TLB recompile, because
>    can_do_io is clear
>
> I am not surprised that the corner case of "the guest put its
> page tables in an MMIO device" has not yet come up :-)
>
> I'm really not sure how the icount handling should interact
> with that...

Its not just icount - we need to handle it for all modes now. That said
seeing as we are at the end of a block shouldn't can_do_io be set?

Does:

modified   accel/tcg/translator.c
@@ -201,6 +201,8 @@ void translator_loop(CPUState *cpu, TranslationBlock *tb, int *max_insns,
         }
     }
 
+    set_can_do_io(db, true);
+
     /* Emit code to exit the TB, as indicated by db->is_jmp.  */
     ops->tb_stop(db, cpu);
     gen_tb_end(tb, cflags, icount_start_insn, db->num_insns);

do the trick?

>
> -- PMM

-- 
Alex Bennée
Virtualisation Tech Lead @ Linaro