Re: [PATCH v2 14/20] nvme: allow multiple aios per command

[Klaus Birkelund <its@irrelevant.dk>]

On Tue, Nov 12, 2019 at 03:25:06PM +0000, Beata Michalska wrote:
> Hi Klaus,
> 
> On Tue, 15 Oct 2019 at 11:55, Klaus Jensen <its@irrelevant.dk> wrote:
> > @@ -341,19 +344,18 @@ static uint16_t nvme_dma_write_prp(NvmeCtrl *n, uint8_t *ptr, uint32_t len,
> Any reason why the nvme_dma_write_prp is missing the changes applied
> to nvme_dma_read_prp ?
> 

This was adressed by proxy through changes to the previous patch
(by combining the read/write functions).

> > +    case NVME_AIO_OPC_WRITE_ZEROES:
> > +        block_acct_start(stats, acct, aio->iov.size, BLOCK_ACCT_WRITE);
> > +        aio->aiocb = blk_aio_pwrite_zeroes(aio->blk, aio->offset,
> > +            aio->iov.size, BDRV_REQ_MAY_UNMAP, nvme_aio_cb, aio);
> Minor: aio->blk  => blk
> 

Thanks. Fixed this in a couple of other places as well.

> > @@ -621,8 +880,11 @@ static uint16_t nvme_del_sq(NvmeCtrl *n, NvmeCmd *cmd)
> >      sq = n->sq[qid];
> >      while (!QTAILQ_EMPTY(&sq->out_req_list)) {
> >          req = QTAILQ_FIRST(&sq->out_req_list);
> > -        assert(req->aiocb);
> > -        blk_aio_cancel(req->aiocb);
> > +        while (!QTAILQ_EMPTY(&req->aio_tailq)) {
> > +            aio = QTAILQ_FIRST(&req->aio_tailq);
> > +            assert(aio->aiocb);
> > +            blk_aio_cancel(aio->aiocb);
> What about releasing memory associated with given aio ?

I believe the callback is still called when cancelled? That should take
care of it. Or have I misunderstood that? At least for the DMAAIOCBs it
is.

> > +struct NvmeAIO {
> > +    NvmeRequest *req;
> > +
> > +    NvmeAIOOp       opc;
> > +    int64_t         offset;
> > +    BlockBackend    *blk;
> > +    BlockAIOCB      *aiocb;
> > +    BlockAcctCookie acct;
> > +
> > +    NvmeAIOCompletionFunc *cb;
> > +    void                  *cb_arg;
> > +
> > +    QEMUSGList   *qsg;
> > +    QEMUIOVector iov;
> 
> There is a bit of inconsistency on the ownership of IOVs and SGLs.
> SGLs now seem to be owned by request whereas IOVs by the aio.
> WOuld be good to have that unified or documented at least.
> 

Fixed this. The NvmeAIO only holds pointers now.

> > +#define NVME_REQ_TRANSFER_DMA  0x1
> This one does not seem to be used ....
> 

I have dropped the flags and reverted to a simple req->is_cmb as that is
all that is really needed.