Hi Juan,
What are the semantics of 'qemu_peek_buffer'?
- is it supposed to guarantee (if there are no errors) that
it will read 'size' bytes? (i.e. it should block)
There are currently two users of it:
* qemu_read_buffer which spins filling it's buffer up
with repeated calls to qemu_peek_buffer
* vmstate_subsection_load that returns if the size read
doesn't match what it was expecting
I can't see how both of them can be right.
The problem I'm seeing is that in my world I've got a
qemu_peek_buffer of 8 bytes, and with a repeated virt-test
local tcp migration it's failing about 1 in 8 times;
here is some debug:
19:51:15 INFO | [qemu output] qemu_peek_buffer refill case (pre); size=8
offset=0 index=32764 pending=4 buf_index=32764 buf_size=32768 pos=23302795
19:51:15 INFO | [qemu output] qemu_fill_buffer got 1
19:51:15 INFO | [qemu output] qemu_peek_buffer refill case (post); size=8
offset=0 index=0 pending=5 buf_index=0 buf_size=5 pos=23302796
19:51:15 INFO | [qemu output] qemu_peek_buffer (size>pending); size=8 offset=0
index=0 pending=5 buf_index=0 buf_size=5 pos=23302796
i.e. I asked for 8 bytes, there were 4 in the buffer, it called fill buffer,
which got one
more byte, and thus it returned me 5.
I think what vmstate_subsection_load wants (and what I want) is something
like qemu_read_buffer but which doesn't advance it's pointer, i.e. to read
a header, decide it's not for me and let the next function along use it.
vmstate_subsection_load doesn't look like it flags an error if it
doesn't read enough; I guess the effect will be just to fail to
load a migration in some interesting way.
Dave
--
Dr. David Alan Gilbert / dgilb...@redhat.com / Manchester, UK
