Re: [Qemu-devel] MTTCG sync-up call today?

[Alex Bennée]

Paolo Bonzini <address@hidden> writes:

> On 12/01/2016 16:11, Alex Bennée wrote:
>> > Sorry for the late answer, I find some time to take a look at it.
>> >
>> > Seems you were right I fixed the exit issue and it seems it was one of
>> > the problem.
>> > I think we must double check how we use cpu->exit_request as Paolo
>> > removed SIG_IPI to exit the CPU.
>> >
>> > I found one additional issue and it seems booting well right now.
>>
>> The other thing that needs cleaning up is the tcg_current_cpu and
>> current_cpu. I suspect the former should go and the restrictions on the
>> later be loosend so the TLS current_cpu is available to deferred tasks.
>
> Yes, you can make TLS current_cpu always non-NULL for multi-threaded TCG.
>
> tcg_current_cpu definitely should go, it doesn't make sense if you have
> multiple threads.
>
>> The thing I'm currently looking at is what happens when something like a
>> virtio completes in a non-CPU thread.
>
> It should just work.  It will cause a cpu_interrupt under the BQL, and
> that sets cpu->interrupt_request.  The code that modifies
> cpu->interrupt_request in the VCPU thread also runs under the BQL.

Hmm I'm seeing a virtio co-routine kick of an unmap:

qemu-system-arm: /home/alex/lsrc/qemu/qemu.git/translate-all.c:1303: 
tb_invalidate_phys_range: Assertion `have_tb_lock' failed.

Program received signal SIGABRT, Aborted.
0x00007ffff0ae9cc9 in __GI_raise (address@hidden) at 
../nptl/sysdeps/unix/sysv/linux/raise.c:56
56      ../nptl/sysdeps/unix/sysv/linux/raise.c: No such file or directory.
(gdb) bt
#0  0x00007ffff0ae9cc9 in __GI_raise (address@hidden) at 
../nptl/sysdeps/unix/sysv/linux/raise.c:56
#1  0x00007ffff0aed0d8 in __GI_abort () at abort.c:89
#2  0x00007ffff0ae2b86 in __assert_fail_base (fmt=0x7ffff0c33830 "%s%s%s:%u: 
%s%sAssertion `%s' failed.\n%n",
    address@hidden "have_tb_lock", address@hidden 
"/home/alex/lsrc/qemu/qemu.git/translate-all.c", address@hidden,
    address@hidden <__PRETTY_FUNCTION__.33460> "tb_invalidate_phys_range") at 
assert.c:92
#3  0x00007ffff0ae2c32 in __GI___assert_fail (address@hidden "have_tb_lock",
    address@hidden "/home/alex/lsrc/qemu/qemu.git/translate-all.c", 
address@hidden,
    address@hidden <__PRETTY_FUNCTION__.33460> "tb_invalidate_phys_range") at 
assert.c:101
#4  0x00005555556e5b06 in tb_invalidate_phys_range (address@hidden, 
address@hidden) at /home/alex/lsrc/qemu/qemu.git/translate-all.c:1303
#5  0x00005555556dbe42 in invalidate_and_set_dirty (address@hidden, addr=0, 
address@hidden) at /home/alex/lsrc/qemu/qemu.git/exec.c:2420
#6  0x00005555556e1890 in address_space_unmap (address@hidden 
<address_space_memory>, buffer=<optimised out>, len=<optimised out>,
    address@hidden, address@hidden) at /home/alex/lsrc/qemu/qemu.git/exec.c:2933
#7  0x00005555556e19bf in cpu_physical_memory_unmap (buffer=<optimised out>, 
len=<optimised out>, address@hidden, address@hidden)
    at /home/alex/lsrc/qemu/qemu.git/exec.c:2962
#8  0x000055555578219c in virtqueue_unmap_sg (address@hidden, address@hidden, 
vq=0x555556e6f020)
    at /home/alex/lsrc/qemu/qemu.git/hw/virtio/virtio.c:257
#9  0x0000555555782ac0 in virtqueue_fill (address@hidden, address@hidden, 
len=4097, address@hidden)
    at /home/alex/lsrc/qemu/qemu.git/hw/virtio/virtio.c:282
#10 0x0000555555782ccf in virtqueue_push (vq=0x555556e6f020, address@hidden, 
len=<optimised out>)
    at /home/alex/lsrc/qemu/qemu.git/hw/virtio/virtio.c:308
#11 0x000055555573451a in virtio_blk_complete_request (req=0x7ffe782c7ce0, 
status=<optimised out>) at 
/home/alex/lsrc/qemu/qemu.git/hw/block/virtio-blk.c:58
#12 0x0000555555734a13 in virtio_blk_req_complete (status=0 '\000', 
req=0x7ffe782c7ce0) at /home/alex/lsrc/qemu/qemu.git/hw/block/virtio-blk.c:64
#13 virtio_blk_rw_complete (opaque=<optimised out>, ret=0) at 
/home/alex/lsrc/qemu/qemu.git/hw/block/virtio-blk.c:122
---Type <return> to continue, or q <return> to quit---
#14 0x0000555555a2d822 in bdrv_co_complete (acb=0x7ffe780189c0) at 
block/io.c:2122
#15 0x0000555555a87a7a in coroutine_trampoline (i0=<optimised out>, 
i1=<optimised out>) at util/coroutine-ucontext.c:80
#16 0x00007ffff0afc8b0 in ?? () from /lib/x86_64-linux-gnu/libc.so.6
#17 0x00007fff8f5aa6e0 in ?? ()
#18 0x0000000000000000 in ?? ()

I guess the tb_lock could just be grabbed but there is stuff in that
path that assumes current_cpu is valid so I thought the thing to do was
defer the operation until a "real" vCPU can deal with it.



>
> Paolo


--
Alex Bennée