Re: [PATCH 1/7] xfs: increase the default parallelism levels of pwork clients

On Thu, Jan 14, 2021 at 01:32:59PM -0800, Darrick J. Wong wrote:

On Wed, Jan 13, 2021 at 03:49:32PM +0100, Christoph Hellwig wrote:

quoted

quoted

+/* Estimate the amount of parallelism available for a given device. */
+unsigned int
+xfs_buftarg_guess_threads(
+	struct xfs_buftarg	*btp)
+{
+	int			iomin;
+	int			ioopt;
+
+	/*
+	 * The device tells us that it is non-rotational, and we take that to
+	 * mean there are no moving parts and that the device can handle all
+	 * the CPUs throwing IO requests at it.
+	 */
+	if (blk_queue_nonrot(btp->bt_bdev->bd_disk->queue))
+		return num_online_cpus();
+
+	/*
+	 * The device has a preferred and minimum IO size that suggest a RAID
+	 * setup, so infer the number of disks and assume that the parallelism
+	 * is equal to the disk count.
+	 */
+	iomin = bdev_io_min(btp->bt_bdev);
+	ioopt = bdev_io_opt(btp->bt_bdev);
+	if (iomin > 0 && ioopt > iomin)
+		return ioopt / iomin;
+
+	/*
+	 * The device did not indicate that it has any capabilities beyond that
+	 * of a rotating disk with a single drive head, so we estimate no
+	 * parallelism at all.
+	 */
+	return 1;
+}

Why is this in xfs_buf.c despite having nothing to do with the buffer
cache?

Initially I assigned it to the buftarg code on the grounds that this is
how you'd estimate the level of parallelism available through the data
device buftarg.

I don't really care where it goes, though.  xfs_pwork_guess_threads
would be fine too.

quoted

Also I think we need some sort of manual override in case the guess is
wrong.

There already /is/ a pwork_threads sysctl knob for controlling
quotacheck parallelism; and for the block gc workqueue I add WQ_SYSFS so
that you can set /sys/bus/workqueue/devices/xfs-*/max_active.

--D

Hm, where would we put it?  One of the xfs sysctls?  And would we be
able to resize the background blockgc workqueue size too?

--D