Date:   Wed, 25 Aug 2021 10:25:40 -0700
From:   Dan Williams <dan.j.williams@...el.com>
To:     Pankaj Gupta <pankaj.gupta.linux@...il.com>
Cc:     Linux NVDIMM <nvdimm@...ts.linux.dev>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        jmoyer <jmoyer@...hat.com>, David Hildenbrand <david@...hat.com>,
        "Michael S. Tsirkin" <mst@...hat.com>,
        Cornelia Huck <cohuck@...hat.com>,
        Vishal L Verma <vishal.l.verma@...el.com>,
        Dave Jiang <dave.jiang@...el.com>,
        "Weiny, Ira" <ira.weiny@...el.com>,
        Pankaj Gupta <pankaj.gupta@...os.com>
Subject: Re: [RFC v2 1/2] virtio-pmem: Async virtio-pmem flush

On Sun, Jul 25, 2021 at 11:09 PM Pankaj Gupta
<pankaj.gupta.linux@...il.com> wrote:
>
> From: Pankaj Gupta <pankaj.gupta@...os.com>
>
> Implement asynchronous flush for virtio pmem using work queue
> to solve the preflush ordering issue. Also, coalesce the flush
> requests when a flush is already in process.
>
> Signed-off-by: Pankaj Gupta <pankaj.gupta@...os.com>
> ---
>  drivers/nvdimm/nd_virtio.c   | 72 ++++++++++++++++++++++++++++--------
>  drivers/nvdimm/virtio_pmem.c | 10 ++++-
>  drivers/nvdimm/virtio_pmem.h | 14 +++++++
>  3 files changed, 79 insertions(+), 17 deletions(-)
>
> diff --git a/drivers/nvdimm/nd_virtio.c b/drivers/nvdimm/nd_virtio.c
> index 10351d5b49fa..61b655b583be 100644
> --- a/drivers/nvdimm/nd_virtio.c
> +++ b/drivers/nvdimm/nd_virtio.c
> @@ -97,29 +97,69 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
>         return err;
>  };
>
> +static void submit_async_flush(struct work_struct *ws);
> +
>  /* 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();
> +
> +       spin_lock_irq(&vpmem->lock);
> +       /* flush requests wait until ongoing flush completes,
> +        * hence coalescing all the pending requests.
>          */
> -       if (bio && bio->bi_iter.bi_sector != -1) {
> -               struct bio *child = bio_alloc(GFP_ATOMIC, 0);
> -
> -               if (!child)
> -                       return -ENOMEM;
> -               bio_copy_dev(child, bio);
> -               child->bi_opf = REQ_PREFLUSH;
> -               child->bi_iter.bi_sector = -1;
> -               bio_chain(child, bio);
> -               submit_bio(child);
> -               return 0;
> +       wait_event_lock_irq(vpmem->sb_wait,
> +                           !vpmem->flush_bio ||
> +                           ktime_before(req_start, vpmem->prev_flush_start),
> +                           vpmem->lock);
> +       /* new request after previous flush is completed */
> +       if (ktime_after(req_start, vpmem->prev_flush_start)) {
> +               WARN_ON(vpmem->flush_bio);
> +               vpmem->flush_bio = bio;
> +               bio = NULL;
> +       }

Why the dance with ->prev_flush_start vs just calling queue_work()
again. queue_work() is naturally coalescing in that if the last work
request has not started execution another queue attempt will be
dropped.

> +       spin_unlock_irq(&vpmem->lock);
> +
> +       if (!bio) {
> +               INIT_WORK(&vpmem->flush_work, submit_async_flush);

I expect this only needs to be initialized once at driver init time.

> +               queue_work(vpmem->pmem_wq, &vpmem->flush_work);
> +               return 1;
> +       }
> +
> +       /* flush completed in other context while we waited */
> +       if (bio && (bio->bi_opf & REQ_PREFLUSH)) {
> +               bio->bi_opf &= ~REQ_PREFLUSH;
> +               submit_bio(bio);
> +       } else if (bio && (bio->bi_opf & REQ_FUA)) {
> +               bio->bi_opf &= ~REQ_FUA;
> +               bio_endio(bio);

It's not clear to me how this happens, shouldn't all flush completions
be driven from the work completion?

>         }
> -       if (virtio_pmem_flush(nd_region))
> -               return -EIO;
>
>         return 0;
>  };
>  EXPORT_SYMBOL_GPL(async_pmem_flush);
> +
> +static void submit_async_flush(struct work_struct *ws)
> +{
> +       struct virtio_pmem *vpmem = container_of(ws, struct virtio_pmem, flush_work);
> +       struct bio *bio = vpmem->flush_bio;
> +
> +       vpmem->start_flush = ktime_get_boottime();
> +       bio->bi_status = errno_to_blk_status(virtio_pmem_flush(vpmem->nd_region));
> +       vpmem->prev_flush_start = vpmem->start_flush;
> +       vpmem->flush_bio = NULL;
> +       wake_up(&vpmem->sb_wait);
> +
> +       /* Submit parent bio only for PREFLUSH */
> +       if (bio && (bio->bi_opf & REQ_PREFLUSH)) {
> +               bio->bi_opf &= ~REQ_PREFLUSH;
> +               submit_bio(bio);
> +       } else if (bio && (bio->bi_opf & REQ_FUA)) {
> +               bio->bi_opf &= ~REQ_FUA;
> +               bio_endio(bio);
> +       }

Shouldn't the wait_event_lock_irq() be here rather than in
async_pmem_flush()? That will cause the workqueue to back up and flush
requests to coalesce.

> +}
>  MODULE_LICENSE("GPL");
> diff --git a/drivers/nvdimm/virtio_pmem.c b/drivers/nvdimm/virtio_pmem.c
> index 726c7354d465..56780a6140c7 100644
> --- a/drivers/nvdimm/virtio_pmem.c
> +++ b/drivers/nvdimm/virtio_pmem.c
> @@ -24,6 +24,7 @@ static int init_vq(struct virtio_pmem *vpmem)
>                 return PTR_ERR(vpmem->req_vq);
>
>         spin_lock_init(&vpmem->pmem_lock);
> +       spin_lock_init(&vpmem->lock);

Why 2 locks?

Powered by blists - more mailing lists