Re: [RFC] thread-pool: Add option to fix the pool size

[Stefan Hajnoczi]

On Mon, Feb 07, 2022 at 01:00:46PM +0100, Nicolas Saenz Julienne wrote:
> Hi Stefan, thanks for the review. I took note of your comments.
> 
> On Thu, 2022-02-03 at 10:53 +0000, Stefan Hajnoczi wrote:
> > > Some background on my workload: I'm using IDE emulation, the guest is an
> > > old RTOS that doesn't support virtio, using 'aio=native' isn't possible
> > > either (unaligned IO accesses).
> > 
> > I thought QEMU's block layer creates bounce buffers for unaligned
> > accesses, handling both memory buffer alignment and LBA alignment
> > necessary for aio=native,cache=none?
> 
> See block/file-posix.c:raw_co_prw() {
> 
>     /*
>      * When using O_DIRECT, the request must be aligned to be able to use
>      * either libaio or io_uring interface. If not fail back to regular thread
>      * pool read/write code which emulates this for us if we set
>      * QEMU_AIO_MISALIGNED.
>      */
>     if (s->needs_alignment && !bdrv_qiov_is_aligned(bs, qiov))
>       type |= QEMU_AIO_MISALIGNED;
>     else if (s->use_linux_io_uring)
>         return luring_co_submit(...);
>     else if (s->use_linux_aio)
>         return laio_co_submit(...);
> 
>     return raw_thread_pool_submit(..., handle_aiocb_rw, ...);
> }
> 
> bdrv_qiov_is_aligned() returns 'false' on my use-case.
> 
> I believe what you're referring to happens in handle_aiocb_rw(), but it's too
> late then.

I was thinking of block/io.c:bdrv_co_preadv_part() but it only aligns
the LBA, not memory buffers.

Stefan

signature.asc
Description: PGP signature