Re: [Qemu-devel] [PATCH] NVMe: Initial commit to add an NVM Express device

[Kevin Wolf <kwolf@redhat.com>]

Am 13.12.2012 01:13, schrieb Busch, Keith:
> On Mon, Dec 10, 2012 at 7:11 AM, Stefan Hajnoczi <stefanha@gmail.com> wrote:
>> Quick pointers to get started on Kevin's suggestion:
>>
>> bdrv_aio_readv(), bdrv_aio_writev(), bdrv_aio_flush(), and
>> bdrv_aio_discard() provide the block device operations that emulated
>> storage controllers use.
> 
> There seems to be an issue with the bdrv_aio_[readv/writev] routines. I added some additional tracing in the block and coroutine code, and well, long story short, the coroutine "bdrv_co_io_em" needs to call "qemu_coroutine_yield" before his aio callback "bdrv_co_io_em_complete" invokes "qemu_coroutine_enter". It does not always win this race in my experiments, and qemu aborts with a recursive re-entry error. I don't know this coroutine code well enough to propose a fix -- I'd say maybe use locks but I think that defeats the purpose of using coroutines if I understand them correctly?

The block layer, like most other qemu functions, is supposed to run
under the qemu_global_mutex (also called BQL). Do you call into it from
a different thread that doesn't take this lock?

Kevin