[Qemu-devel] [RFC DEBUG PATCH 0/3] debug patch for lookup-ptr hang

[Alex Bennée]

Hi,

These are debug patches only but represent how much I have narrowed
down the problem so far. I've included Thomas' patch to fix the
thread=single|multi option as that is currently broken upstream. So
far it seems though the problem is unrelated to multi-threading.

As discussed in the other thread I found not returning to a
tb_htable_lookup but adding it to the tb_jmp_cache made the problem go
away. I also tried various printfs but they also seemed to un-wedge
the hang I was seeing. It is not really a hang rather than a busy-spin
that will eventually given enough time unwind.

So added a new TB flag (is_magic) which if set would skip returning
the code ptr and default to exiting the loop via the epilogue and set
it for all DISAS_JUMP/DISAS_UPDATE paths that trigger
lookup_and_goto_ptr. After selectively commenting them out I found the
RET instruction is responsible for my particular fail case.

I find this confusing because BL and BLR basically do the same thing
and they seem to work fine.

I have an uneasy feeling there is some subtle black magic in the
interaction between cpu_get_tb_cpu_state(env, &pc, &cs_base, &flags),
addr and the TCGv cpu_pc but I haven't nailed it down. I'm posting
this for those that still have some Friday left in case it prompts any
thoughts.

Over to you, hopefully inspiration will strike before I return to the
fray on Monday ;-)

Cheers,

Alex Bennée (2):
  tcg-runtime: light re-factor of lookup_tb_ptr
  translate-a64: fix lookup_tb_ptr hang (DEBUG!)

Thomas Huth (1):
  vl: Fix broken thread=xxx option of the --accel parameter

 include/exec/exec-all.h    |  2 ++
 target/arm/translate-a64.c | 21 +++++++++++++++++----
 target/arm/translate.h     |  2 ++
 tcg-runtime.c              | 37 +++++++++++++++++++++----------------
 vl.c                       | 13 +++++--------
 5 files changed, 47 insertions(+), 28 deletions(-)

-- 
2.13.0