On Feb  9 03:59, Keith Busch wrote:
> On Mon, Jan 25, 2021 at 09:42:31PM +0100, Klaus Jensen wrote:
> > @@ -1644,10 +1679,56 @@ static uint16_t nvme_compare(NvmeCtrl *n, NvmeRequest *req)
> >  
> >  static uint16_t nvme_flush(NvmeCtrl *n, NvmeRequest *req)
> >  {
> > -    block_acct_start(blk_get_stats(req->ns->blkconf.blk), &req->acct, 0,
> > -                     BLOCK_ACCT_FLUSH);
> > -    req->aiocb = blk_aio_flush(req->ns->blkconf.blk, nvme_rw_cb, req);
> > -    return NVME_NO_COMPLETE;
> > +    uint32_t nsid = le32_to_cpu(req->cmd.nsid);
> > +    uintptr_t *num_flushes = (uintptr_t *)&req->opaque;
> > +    uint16_t status;
> > +    struct nvme_aio_flush_ctx *ctx;
> > +    NvmeNamespace *ns;
> > +
> > +    trace_pci_nvme_flush(nvme_cid(req), nsid);
> > +
> > +    if (nsid != NVME_NSID_BROADCAST) {
> > +        req->ns = nvme_ns(n, nsid);
> > +        if (unlikely(!req->ns)) {
> > +            return NVME_INVALID_FIELD | NVME_DNR;
> > +        }
> > +
> > +        block_acct_start(blk_get_stats(req->ns->blkconf.blk), &req->acct, 0,
> > +                         BLOCK_ACCT_FLUSH);
> > +        req->aiocb = blk_aio_flush(req->ns->blkconf.blk, nvme_rw_cb, req);
> > +        return NVME_NO_COMPLETE;
> > +    }
> > +
> > +    /* 1-initialize; see comment in nvme_dsm */
> > +    *num_flushes = 1;
> > +
> > +    for (int i = 1; i <= n->num_namespaces; i++) {
> > +        ns = nvme_ns(n, i);
> > +        if (!ns) {
> > +            continue;
> > +        }
> > +
> > +        ctx = g_new(struct nvme_aio_flush_ctx, 1);
> > +        ctx->req = req;
> > +        ctx->ns = ns;
> > +
> > +        (*num_flushes)++;
> > +
> > +        block_acct_start(blk_get_stats(ns->blkconf.blk), &ctx->acct, 0,
> > +                         BLOCK_ACCT_FLUSH);
> > +        req->aiocb = blk_aio_flush(ns->blkconf.blk, nvme_aio_flush_cb, ctx);
> > +    }
> 
> Overwriting req->aiocb with the most recent flush request doesn't seem
> right.
> 

It doesn't really matter if this[1] patch is merged, but it is wrong and
to align with the other multi-aio commands, this should just ignore the
returned aiocb.

> This whole implementation would be much simpler with the synchronous
> blk_flush() routine instead of the AIO equivalent. This is not really a
> performant feature, so I don't think it's critical to get these
> operations happening in parallel. What do you think?

It would definitely be simpler, but I believe that if there is a lot to
flush, then we won't just block the nvme device. We are holding the Big
QEMU Lock and will block most other devices as well.


  [1]: hw/block/nvme: drain namespaces on sq deletion