Re: [PATCHv2 1/7] zram: introduce compressed data writeback

[zhangdongdong]

On 1/7/26 12:28, Sergey Senozhatsky wrote:
> On (26/01/07 11:50), zhangdongdong wrote:
> > Hi Sergey,
> >
> > Thanks for the work on decompression-on-demand.
> >
> > One concern I’d like to raise is the use of a workqueue for readback
> > decompression. In our measurements, deferring decompression to a worker
> > introduces non-trivial scheduling overhead, and under memory pressure
> > the added latency can be noticeable (tens of milliseconds in some cases).
>
> The problem is those bio completions happen in atomic context, and zram
> requires both compression and decompression to be non-atomic.  And we
> can't do sync read on the zram side, because those bio-s are chained.
> So the current plan is to look how system hi-prio per-cpu workqueue
> will handle this.
>
> Did you try high priority workqueue?
Hi,Sergey

Yes, we have tried high priority workqueues. In fact, our current
implementation already uses a dedicated workqueue created with
WQ_HIGHPRI and marked as UNBOUND, which handles the read/decompression
path for swap-in.

Below is a simplified snippet of the queue we are currently using:

zgroup_read_wq = alloc_workqueue("zgroup_read",
				 WQ_HIGHPRI | WQ_UNBOUND, 0);

static int zgroup_submit_zio_async(struct zgroup_io *zio,
				   struct zram_group *zgroup)
{
	struct zgroup_req req = {
		.zio = zio,
	};

	if (!zgroup_io_step_chg(zio, ZIO_STARTED, ZIO_INFLIGHT)) {
		wait_for_completion(&zio->wait);
		if (zio->status)
			zgroup_put_io(zio);
		return zio->status;
	}

	INIT_WORK_ONSTACK(&req.work, zgroup_submit_zio_work);
	queue_work(zgroup_read_wq, &req.work);
	flush_work(&req.work);
	destroy_work_on_stack(&req.work);

	return req.status ?: zgroup_decrypt_pages(zio);
}

Thanks,
dongdong