Re: [RFC v3 1/2] virtio-pmem: Async virtio-pmem flush

[Dan Williams <dan.j.williams@intel.com>]

On Wed, Feb 16, 2022 at 12:47 AM Pankaj Gupta
<pankaj.gupta.linux@gmail.com> wrote:
>
> > >
> > > Enable asynchronous flush for virtio pmem using work queue. Also,
> > > coalesce the flush requests when a flush is already in process.
> > > This functionality is copied from md/RAID code.
> > >
> > > When a flush is already in process, new flush requests wait till
> > > previous flush completes in another context (work queue). For all
> > > the requests come between ongoing flush and new flush start time, only
> > > single flush executes, thus adhers to flush coalscing logic. This is
> >
> > s/adhers/adheres/
> >
> > s/coalscing/coalescing/
> >
> > > important for maintaining the flush request order with request coalscing.
> >
> > s/coalscing/coalescing/
>
> o.k. Sorry for the spelling mistakes.
>
> >
> > >
> > > Signed-off-by: Pankaj Gupta <pankaj.gupta.linux@gmail.com>
> > > ---
> > >  drivers/nvdimm/nd_virtio.c   | 74 +++++++++++++++++++++++++++---------
> > >  drivers/nvdimm/virtio_pmem.c | 10 +++++
> > >  drivers/nvdimm/virtio_pmem.h | 16 ++++++++
> > >  3 files changed, 83 insertions(+), 17 deletions(-)
> > >
> > > diff --git a/drivers/nvdimm/nd_virtio.c b/drivers/nvdimm/nd_virtio.c
> > > index 10351d5b49fa..179ea7a73338 100644
> > > --- a/drivers/nvdimm/nd_virtio.c
> > > +++ b/drivers/nvdimm/nd_virtio.c
> > > @@ -100,26 +100,66 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
> > >  /* The asynchronous flush callback function */
> > >  int async_pmem_flush(struct nd_region *nd_region, struct bio *bio)
> > >  {
> > > -       /*
> > > -        * Create child bio for asynchronous flush and chain with
> > > -        * parent bio. Otherwise directly call nd_region flush.
> > > +       /* queue asynchronous flush and coalesce the flush requests */
> > > +       struct virtio_device *vdev = nd_region->provider_data;
> > > +       struct virtio_pmem *vpmem  = vdev->priv;
> > > +       ktime_t req_start = ktime_get_boottime();
> > > +       int ret = -EINPROGRESS;
> > > +
> > > +       spin_lock_irq(&vpmem->lock);
> >
> > Why a new lock and not continue to use ->pmem_lock?
>
> This spinlock is to protect entry in 'wait_event_lock_irq'
> and the Other spinlock is to protect virtio queue data.

Understood, but md shares the mddev->lock for both purposes, so I
would ask that you either document what motivates the locking split,
or just reuse the lock until a strong reason to split them arises.