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

[Minwoo Im]

On 21-02-11 13:07:08, Klaus Jensen wrote:
> On Feb 11 11:49, Minwoo Im wrote:
> > On 21-01-27 14:15:05, Klaus Jensen wrote:
> > > From: Klaus Jensen <k.jensen@samsung.com>
> > > 
> > > For most commands, when issuing an AIO, the BlockAIOCB is stored in the
> > > NvmeRequest aiocb pointer when the AIO is issued. The purpose of storing
> > > this is to allow the AIO to be cancelled when deleting submission
> > > queues (it is currently not used for Abort).
> > > 
> > > Since the addition of the Dataset Management command and Zoned
> > > Namespaces, NvmeRequests may involve more than one AIO and the AIOs are
> > > issued without saving a reference to the BlockAIOCB. This is a problem
> > > since nvme_del_sq will attempt to cancel outstanding AIOs, potentially
> > > with an invalid BlockAIOCB.
> > > 
> > > Fix this by instead of explicitly cancelling the requests, just allow
> > > the AIOs to complete by draining the namespace blockdevs.
> > > 
> > > Signed-off-by: Klaus Jensen <k.jensen@samsung.com>
> > > ---
> > >  hw/block/nvme.c | 18 +++++++++++++-----
> > >  1 file changed, 13 insertions(+), 5 deletions(-)
> > > 
> > > diff --git a/hw/block/nvme.c b/hw/block/nvme.c
> > > index 316858fd8adf..91f6fb6da1e2 100644
> > > --- a/hw/block/nvme.c
> > > +++ b/hw/block/nvme.c
> > > @@ -403,6 +403,7 @@ static void nvme_req_clear(NvmeRequest *req)
> > >  {
> > >      req->ns = NULL;
> > >      req->opaque = NULL;
> > > +    req->aiocb = NULL;
> > >      memset(&req->cqe, 0x0, sizeof(req->cqe));
> > >      req->status = NVME_SUCCESS;
> > >  }
> > > @@ -2396,6 +2397,7 @@ static uint16_t nvme_del_sq(NvmeCtrl *n, 
> > > NvmeRequest *req)
> > >      NvmeSQueue *sq;
> > >      NvmeCQueue *cq;
> > >      uint16_t qid = le16_to_cpu(c->qid);
> > > +    int i;
> > >  
> > >      if (unlikely(!qid || nvme_check_sqid(n, qid))) {
> > >          trace_pci_nvme_err_invalid_del_sq(qid);
> > > @@ -2404,12 +2406,18 @@ static uint16_t nvme_del_sq(NvmeCtrl *n, 
> > > NvmeRequest *req)
> > >  
> > >      trace_pci_nvme_del_sq(qid);
> > >  
> > > -    sq = n->sq[qid];
> > > -    while (!QTAILQ_EMPTY(&sq->out_req_list)) {
> > > -        r = QTAILQ_FIRST(&sq->out_req_list);
> > > -        assert(r->aiocb);
> > > -        blk_aio_cancel(r->aiocb);
> > > +    for (i = 1; i <= n->num_namespaces; i++) {
> > > +        NvmeNamespace *ns = nvme_ns(n, i);
> > > +        if (!ns) {
> > > +            continue;
> > > +        }
> > > +
> > > +        nvme_ns_drain(ns);
> > 
> > If we just drain the entire namespaces here, commands which has nothing
> > to do with the target sq to be deleted will be drained.  And this might
> > be a burden for a single SQ deletion.
> > 
> 
> That is true. But how often would you dynamically delete and create I/O
> submission queues in the fast path?

Delete I/O queues are not that often in the working NVMe controller, but
it might be a good case for the exception test from the host side like:
I/O queue deletion during I/O workloads.  If delete I/O queues are
returning by aborting their own requests only and quickly respond to the
host, then I think it might be a good one to test with.  Handling
requests gracefully sometimes don't cause corner cases from the host
point-of-view.  But, QEMU is not only for the host testing, so I am not
sure that QEMU NVMe device should handle things gracefully or try to do
something exactly as the real hardware(but, we don't know all the
hardware behavior ;)).

(But, Right. If I'm only talking about the kernel, then kernel does not
delete queues during the fast-path hot workloads.  But it's sometimes
great to test something on their own driver or application)

> > By the way, agree with the multiple AIOs references problem for newly added
> > commands.  But, shouldn't we manage the inflight AIO request references for
> > the newlly added commands with some other way and kill them all
> > explicitly as it was?  Maybe some of list for AIOCBs?
> 
> I was hesitant to add more stuff to NvmeRequest (like a QTAILQ to track
> this). Getting a steady-state with draining was an easy fix.

Graceful handling is easy to go with.  I am not expert for the overall
purpose of the QEMU NVMe device model, but I'm curious that which one we
need to take first between `Easy to go vs. What device should do`.