Re: Migration issues in qemu.git

[Alex Williamson <alex.williamson@redhat.com>]

On Mon, 2010-08-02 at 12:42 +0300, Avi Kivity wrote:
> On 08/02/2010 12:06 PM, Avi Kivity wrote:
> >  I'm hitting some migration issues merging qemu.git into qemu-kvm.git:
> >
> > 1. Crash in mig_cancel test:
> >
> > (gdb) bt
> > #0  0x0000003a91c83dbb in memcpy () from /lib64/libc.so.6
> > #1  0x000000000049c2ff in qemu_get_buffer (f=0x302d870, buf=<value 
> > optimized out>, size1=4096) at /usr/include/bits/string3.h:52
> > #2  0x0000000000409464 in ram_load (f=0x302d870, opaque=<value 
> > optimized out>, version_id=4) at 
> > /build/home/tlv/akivity/qemu-kvm/arch_init.c:407
> > #3  0x000000000049cb4c in qemu_loadvm_state (f=0x302d870) at 
> > savevm.c:1708
> > #4  0x0000000000494169 in process_incoming_migration (f=<value 
> > optimized out>) at migration.c:63
> > #5  0x0000000000494517 in tcp_accept_incoming_migration (opaque=<value 
> > optimized out>) at migration-tcp.c:163
> > #6  0x000000000041b67e in main_loop_wait (nonblocking=<value optimized 
> > out>) at /build/home/tlv/akivity/qemu-kvm/vl.c:1300
> > #7  0x00000000004314e7 in kvm_main_loop () at 
> > /build/home/tlv/akivity/qemu-kvm/qemu-kvm.c:1710
> > #8  0x000000000041c67f in main_loop (argc=<value optimized out>, 
> > argv=<value optimized out>, envp=<value optimized out>)
> >     at /build/home/tlv/akivity/qemu-kvm/vl.c:1340
> > #9  main (argc=<value optimized out>, argv=<value optimized out>, 
> > envp=<value optimized out>) at /build/home/tlv/akivity/qemu-kvm/vl.c:3069
> >
> > This is on the incoming side so the test completes successfully, only 
> > leaving a core dump to fill my disks.
> 
> 
> This appears to be
> 
> > static inline void *host_from_stream_offset(QEMUFile *f,
> >                                             ram_addr_t offset,
> >                                             int flags)
> > {
> >     static RAMBlock *block = NULL;
> >     char id[256];
> >     uint8_t len;
> >
> >     if (flags & RAM_SAVE_FLAG_CONTINUE) {
> >         if (!block) {
> >             fprintf(stderr, "Ack, bad migration stream!\n");
> >             return NULL;
> >         }
> >
> >         return block->host + offset;
> >     }
> 
> with block == NULL, if my gdb-fu got a static variable in an inlined 
> function examined correctly.

If block == NULL, are you getting the fprintf?

> I don't see any special reason for block to be NULL on a cancelled 
> migration.  Though perhaps the incoming stream was terminated without us 
> noticing, and we're migrating from some random buffer and confusing the 
> code?

Yeah, I don't understand that either, block == NULL should only be an
initial state, once we've seen a block it shouldn't happen.  Does this
patch solve anything:

http://lists.nongnu.org/archive/html/qemu-devel/2010-07/msg01114.html

I could see this fixing it if the migration was re-attempted after the
cancel.

Alex