Re: [PATCH proxmox v5 08/13] s3-client: request counters: periodically persist counters to file

[Hannes Laimer <h.laimer@proxmox.com>]

On 2026-03-11 14:08, Christian Ebner wrote:
> According to the mmap man page [0] shared memory mapped contents are
> not guaranteed to be written back to the underlying file until
> unmapped, unless the MAP_SYNC flag is used. This would however lead
> to excessive I/O, writing counter states on each and every update.
> 
> Therefore, periodically persist the counter state via asynchronous
> msync calls. To reduce the number of such requests, do not perform
> this unconditionally and for each mmapped counter individually, but
> rather provide a global flusher which only does the flush if the
> s3 client registered a callback function. Further, do not perform
> the call for each counter update, but rather let the s3 client signal
> when counter updates happened via an async channel.
> 
> The task for flush request processing is instantiated and executed on
> demand once callbacks are registered.
> 
> [0] https://man7.org/linux/man-pages/man2/mmap.2.html
> 
> Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
> ---

[..]

> @@ -455,6 +465,10 @@ impl S3Client {
>                          .context("failed to account for upload traffic")?;
>                      let _prev_uploaded =
>                          counters.add_upload_traffic(transferred, Ordering::AcqRel);
> +
> +                    tokio::task::spawn_blocking(|| {
> +                        let _ = SHARED_COUNTER_FLUSHER.read().unwrap().request_flush();
> +                    });

`request_flush()` does `try_send()`, so we don't need `spawn_blocking`
here

>                  }
>                  return Ok(response?);
>              }

[..]