On Tue, Aug 18, 2020 at 04:19:58PM +0100, Stefan Hajnoczi wrote:
On Mon, Aug 17, 2020 at 2:13 PM Coiby Xu <coiby...@gmail.com> wrote:
I'm working on vhost-user block device backend [1]. When I tested my
patches on the latest version of QEMU (commit
d0ed6a69d399ae193959225cdeaa9382746c91cc) with --enable-sanitizer turned
on, the follow error will be thrown out,
==46631==ERROR: LeakSanitizer: detected memory leaks
Direct leak of 312 byte(s) in 1 object(s) allocated from:
#0 0x55d0a7813921 in calloc (qemu/qemu-storage-daemon+0x646921)
#1 0x7f8bd6f7b591 in g_malloc0 (/usr/lib/libglib-2.0.so.0+0x4f591)
#2 0x55d0a815f3fb in qemu_coroutine_create
qemu/util/qemu-coroutine.c:75:14
#3 0x55d0a817321e in vu_client_start
qemu/util/vhost-user-server.c:225:23
#4 0x55d0a816fcf3 in vu_accept qemu/util/vhost-user-server.c:341:5
#5 0x55d0a7f1293f in qio_net_listener_channel_func
qemu/io/net-listener.c:54:9
#6 0x55d0a7efb0f3 in qio_channel_fd_source_dispatch
qemu/io/channel-watch.c:84:12
#7 0x7f8bd6f7e33f in g_main_context_dispatch
(/usr/lib/libglib-2.0.so.0+0x5233f)
I have to call qemu_coroutine_delete to eliminate this kind of errors.
qemu_coroutine_delete() is an internal API and shouldn't be called directly.
If you are sure that the coroutine terminated (returned from its entry
function) then the leak is probably caused by the following:
Thank you for this clue! Yesterday I realized the execution of coroutine
has been yielded since qio_channel_yield is called. Now I give the socket
server enough time to do the clean-up work after the client socket is
closed and the issue is gone.
util/qemu-coroutine.c has a pool of Coroutine objects that are reused.
It's likely that the "leaked" object is in that pool.
I'm surprised that this issue hasn't been reported before, but
release_pool should be cleaned up when the program terminates.
I still can't figure out why the issue hasn't been reported before. I guess
I have to put it aside.
Stefan
--
Best regards,
Coiby
