Fiona asked why it's correct that bdrv_graph_co_rd_lock/unlock() do nothing if qemu_in_main_thread() returns true. As far as I can tell, it's not correct. The coroutine can yield while it believes to have the lock, and then the main loop can call some code that takes a write lock without waiting for the coroutine to finish.
So this series - or more specifically the last patch - fixes this problem by enabling the locking even in the main thread. The patches before this are fixes for bugs that we hadn't discovered while they were only triggered with iothreads, and which are necessary so that all test cases still pass after the final patch. Kevin Wolf (8): block: Call .bdrv_co_create(_opts) unlocked block/export: Fix null pointer dereference in error path qcow2: Unlock the graph in qcow2_do_open() where necessary qemu-img: Take graph lock more selectively test-bdrv-drain: Take graph lock more selectively test-bdrv-drain: Call bdrv_co_unref() in coroutine context blockjob: Adhere to rate limit even when reentered early graph-lock: Honour read locks even in the main thread include/block/block-global-state.h | 8 +++--- include/block/block_int-common.h | 4 +-- include/block/blockjob_int.h | 14 +++++++--- block.c | 1 - block/commit.c | 7 ++--- block/create.c | 1 - block/crypto.c | 25 +++++++++-------- block/export/export.c | 6 +++-- block/graph-lock.c | 10 ------- block/mirror.c | 23 +++++++--------- block/parallels.c | 6 ++--- block/qcow.c | 6 ++--- block/qcow2.c | 43 +++++++++++++++++++++--------- block/qed.c | 6 ++--- block/raw-format.c | 2 +- block/stream.c | 7 ++--- block/vdi.c | 11 ++++---- block/vhdx.c | 8 +++--- block/vmdk.c | 27 +++++++++---------- block/vpc.c | 6 ++--- blockjob.c | 22 +++++++++++++-- qemu-img.c | 5 ++-- tests/unit/test-bdrv-drain.c | 6 ++--- 23 files changed, 138 insertions(+), 116 deletions(-) -- 2.40.1
