On 07/11/2016 10:42, Stefan Hajnoczi wrote:
Let's try to isolate the cause of this crash:
Are you able to switch -netdev user to -netdev tap so we can rule out
the slirp user network stack as the source of memory corruption?
Let me try to set that up. Using packer.io, I will have to start a VM by
hand, and then use the 'null' builder to ssh to the existing VM (whereas
normally packer fires up the qemu process by itself)
Alternatively could you re-run with virtio-blk instead of virtio-scsi to
see if that eliminates crashes?
This is what I got after changing to virtio:
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
Core was generated by `/usr/local/bin/qemu-system-x86_64 -netdev
user,id=user.0,hostfwd=tcp::2521-:22'.
Program terminated with signal SIGABRT, Aborted.
#0 0x00007fa76d645428 in __GI_raise (sig=sig@entry=6) at
../sysdeps/unix/sysv/linux/raise.c:54
54 ../sysdeps/unix/sysv/linux/raise.c: No such file or directory.
[Current thread is 1 (Thread 0x7fa76f065a80 (LWP 18155))]
(gdb) bt
#0 0x00007fa76d645428 in __GI_raise (sig=sig@entry=6) at
../sysdeps/unix/sysv/linux/raise.c:54
#1 0x00007fa76d64702a in __GI_abort () at abort.c:89
#2 0x00007fa76d63dbd7 in __assert_fail_base (fmt=<optimised out>,
assertion=assertion@entry=0x5629cea98cd5 "mr != NULL",
file=file@entry=0x5629cea7a884 "/home/nsrc/qemu-2.7.0/exec.c",
line=line@entry=2967,
function=function@entry=0x5629cea7af00 <__PRETTY_FUNCTION__.42881>
"address_space_unmap")
at assert.c:92
#3 0x00007fa76d63dc82 in __GI___assert_fail (
assertion=assertion@entry=0x5629cea98cd5 "mr != NULL",
file=file@entry=0x5629cea7a884 "/home/nsrc/qemu-2.7.0/exec.c",
line=line@entry=2967,
function=function@entry=0x5629cea7af00 <__PRETTY_FUNCTION__.42881>
"address_space_unmap")
at assert.c:101
#4 0x00005629ce6c0ffe in address_space_unmap (as=<optimised out>,
buffer=<optimised out>,
len=<optimised out>, is_write=1, access_len=4096) at
/home/nsrc/qemu-2.7.0/exec.c:2967
#5 0x00005629ce743beb in virtqueue_unmap_sg
(elem=elem@entry=0x5629d29d5290, len=len@entry=61441,
vq=0x5629d13186b0) at /home/nsrc/qemu-2.7.0/hw/virtio/virtio.c:254
#6 0x00005629ce744422 in virtqueue_fill (vq=vq@entry=0x5629d13186b0,
elem=elem@entry=0x5629d29d5290, len=61441, idx=idx@entry=0)
at /home/nsrc/qemu-2.7.0/hw/virtio/virtio.c:282
#7 0x00005629ce7445db in virtqueue_push (vq=0x5629d13186b0,
elem=elem@entry=0x5629d29d5290,
len=<optimised out>) at /home/nsrc/qemu-2.7.0/hw/virtio/virtio.c:308
#8 0x00005629ce71894d in virtio_blk_req_complete
(req=req@entry=0x5629d29d5290,
status=status@entry=0 '\000') at
/home/nsrc/qemu-2.7.0/hw/block/virtio-blk.c:58
#9 0x00005629ce718b59 in virtio_blk_rw_complete (opaque=<optimised
out>, ret=0)
at /home/nsrc/qemu-2.7.0/hw/block/virtio-blk.c:121
#10 0x00005629ce98025a in blk_aio_complete (acb=0x5629d298f370)
at /home/nsrc/qemu-2.7.0/block/block-backend.c:923
#11 0x00005629ce9efaea in coroutine_trampoline (i0=<optimised out>,
i1=<optimised out>)
at /home/nsrc/qemu-2.7.0/util/coroutine-ucontext.c:78
#12 0x00007fa76d65a5d0 in ?? () from /lib/x86_64-linux-gnu/libc.so.6
#13 0x00007ffee3d75a20 in ?? ()
#14 0x2d2d2d2d2d2d2d2d in ?? ()
---Type <return> to continue, or q <return> to quit---
#15 0x00000000000000f0 in ?? ()
#16 0x0000000000000000 in ?? ()
Aside: I see "virtqueue_unmap_sg" in the backtrace. Is this correct even
for a non-SCSI virtio?
The command line was something like this (captured by running packer
another time, so the ports and filenames are not exactly the same)
/usr/local/bin/qemu-system-x86_64 -m 4G -vnc [::]:59 -machine
type=pc,accel=kvm -netdev user,id=user.0,hostfwd=tcp::2879-:22 -boot c
-smp 8,sockets=1,cores=4,threads=2 -name vtp-nmm-201611071057.qcow2
-device virtio-net,netdev=user.0 -drive
file=output-qemu-vtp-nmm/vtp-nmm-201611071057.qcow2,if=virtio,cache=writeback,discard=ignore,format=qcow2
The core dumps are likely to contain more clues. If you are comfortable
with gdb and debugging C code you could dump the memory surround where
the junk value (mr) was loaded from. Perhaps there is a hint about who
zeroed the memory. In the first core dump you could start with:
(gdb) up 6 # go to the dma_blk_unmap() stack frame
(gdb) p *(DMAAIOCB *)0x560909ceca90
(gdb) p *((DMAAIOCB *)0x560909ceca90).sg
(gdb) up 6
#6 dma_blk_unmap (dbs=dbs@entry=0x560909ceca90) at
/home/nsrc/qemu-2.7.0/dma-helpers.c:102
102 dma_memory_unmap(dbs->sg->as, dbs->iov.iov[i].iov_base,
(gdb) p *(DMAAIOCB *)0x560909ceca90
$1 = {common = {aiocb_info = 0x560907c15690 <dma_aiocb_info>, bs = 0x0,
cb = 0x56090767e250 <scsi_dma_complete>, opaque = 0x560909c2b8e0,
refcnt = 1},
ctx = 0x5609087d82a0, acb = 0x0, sg = 0x560909af7430, offset =
4302675968,
dir = DMA_DIRECTION_FROM_DEVICE, sg_cur_index = 126, sg_cur_byte = 0,
iov = {
iov = 0x560909c6e960, niov = 126, nalloc = 126, size = 1048576}, bh
= 0x0,
io_func = 0x56090767d110 <scsi_dma_readv>, io_func_opaque =
0x560909c2b8e0}
(gdb) p *((DMAAIOCB *)0x560909ceca90).sg
$2 = {sg = 0x560909fab1e0, nsg = 126, nalloc = 143, size = 1048576, dev
= 0x5609087e5630,
as = 0x560907e20480 <address_space_memory>}
(gdb)
I'm comfortable with C, but don't really know what I'm looking for, nor
what the data structures represent :-)
(gdb) p dbs->iov.niov
$3 = 126
(gdb) p i
$4 = 125
...so it appears it was in the last iteration of the loop.
(gdb) print dbs->sg->as
$5 = (AddressSpace *) 0x560907e20480 <address_space_memory>
(gdb) print dbs->iov.iov[i].iov_base
$6 = (void *) 0x7f354099e000
(gdb) print dbs->iov.iov[i].iov_len
$7 = 8192
(gdb) print dbs->dir
$8 = DMA_DIRECTION_FROM_DEVICE
Unfortunately, much has been inlined:
(gdb) frame 4
#4 0x000056090749dffe in address_space_unmap (as=<optimised out>,
buffer=<optimised out>,
len=<optimised out>, is_write=1, access_len=8192) at
/home/nsrc/qemu-2.7.0/exec.c:2967
2967 assert(mr != NULL);
(gdb) print mr
$9 = (MemoryRegion *) 0x0
(gdb) print buffer
$10 = <optimised out>
(gdb)
Regards,
Brian.
