Re: [Qemu-devel] [PATCH] use bdrv_aio functions in fdc.c

[Stefan Hajnoczi]

On Mon, Mar 19, 2012 at 05:17:15PM +0800, Li Zhi Hui wrote:
> @@ -1196,107 +1322,108 @@ static int fdctrl_transfer_handler (void *opaque, 
> int nchan,
>          return 0;
>      }
>      cur_drv = get_cur_drv(fdctrl);
> -    if (fdctrl->data_dir == FD_DIR_SCANE || fdctrl->data_dir == FD_DIR_SCANL 
> ||
> -        fdctrl->data_dir == FD_DIR_SCANH)
> +    if ((fdctrl->data_dir == FD_DIR_SCANE) ||
> +        (fdctrl->data_dir == FD_DIR_SCANL) ||
> +        (fdctrl->data_dir == FD_DIR_SCANH)) {
>          status2 = FD_SR2_SNS;
> -    if (dma_len > fdctrl->data_len)
> +    }
> +    if (dma_len > fdctrl->data_len) {
>          dma_len = fdctrl->data_len;
> +    }
>      if (cur_drv->bs == NULL) {
> -        if (fdctrl->data_dir == FD_DIR_WRITE)
> -            fdctrl_stop_transfer(fdctrl, FD_SR0_ABNTERM | FD_SR0_SEEK, 0x00, 
> 0x00);
> -        else
> +        if (fdctrl->data_dir == FD_DIR_WRITE) {
> +            fdctrl_stop_transfer(fdctrl,
> +                FD_SR0_ABNTERM | FD_SR0_SEEK, 0x00, 0x00);
> +        } else {
>              fdctrl_stop_transfer(fdctrl, FD_SR0_ABNTERM, 0x00, 0x00);
> +        }
>          len = 0;
>          goto transfer_error;
>      }
> +
> +    if ((fdctrl->data_dir != FD_DIR_WRITE) && (fdctrl->data_pos < dma_len)) {
> +        fdc_sector_num = (dma_len + FD_SECTOR_LEN - 1) / FD_SECTOR_LEN;
> +        opaque_cb = g_malloc0(sizeof(FDC_opaque));

I think we can only have 1 I/O pending at a time.  Therefore it's
probably not necessary to create a separate struct, we can just pass the
FDrive/FDCtrl.

> +        qiov = g_malloc0(sizeof(QEMUIOVector));

This is leaked.  I think it can be a field in opaque_cb, there's no need
to allocate this separately on the heap.

> +        pfdc_string = g_malloc0(fdc_sector_num * FD_SECTOR_LEN);

Looks like this buffer is leaked.  A block I/O buffer should be
allocated with qemu_blockalign() instead of g_malloc0() so that memory
alignment requirements for O_DIRECT image files are met.

Would it be possible to use the fifo[] buffer instead of allocating a
new buffer?

> +
> +        qemu_iovec_init(qiov, 1);
> +        qiov->iov->iov_base = pfdc_string;
> +        qiov->iov->iov_len = fdc_sector_num * FD_SECTOR_LEN;
> +        qiov->niov = 1;
> +        qiov->size = fdc_sector_num * FD_SECTOR_LEN;

The easiest way to do this is:

iov.iov_base = fifo;
iov.iov_len  = fdc_sector_num * FD_SECTOR_LEN;
qemu_iovec_init_external(qiov, iov, 1);

We shouldn't duplicate the qiov->size calculation - that's already
provided by qemu_iovec_init_external() or qemu_iovec_add().

> +        opaque_cb->fdctrl = fdctrl;
> +        opaque_cb->qiov = qiov;
> +        opaque_cb->nchan = nchan;
> +        opaque_cb->dma_len = dma_len;
> +        bdrv_aio_readv(cur_drv->bs, fd_sector(cur_drv),
> +            qiov, fdc_sector_num, fdctrl_read_DMA_cb, opaque_cb);
> +        return dma_len;

We are returning dma_len but the I/O has not yet happened.  The guest
could see that the DMA controller register has updated before we've
actually transferred data.  This seems risky.

Have you checked what hw/dma.c does after we return?  The DMA operation
has not completed yet so I wonder if it will call
fdctrl_transfer_handler() again from DMA_run()?

Stefan