xfs: increase the default parallelism levels of pwork clientspwork-parallelism_2020-03-10

Increase the default parallelism level for pwork clients so that we can
take advantage of computers with a lot of CPUs and a lot of hardware.

8x raid0 spinning rust running quotacheck:
1		39s
2		29s
4		26s
8		24s
24 (nr_cpus)	24s

4x raid0 sata ssds running quotacheck:
1		12s
2		12s
4		12s
8		13s
24 (nr_cpus)	14s

4x raid0 nvme ssds running quotacheck:
1		18s
2		18s
4		19s
8		20s
20 (nr_cpus)	20s

So, mixed results...

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

7 files changed, 82 insertions, 14 deletions

diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c
index 217e4f82a44a..11a97bc35f70 100644
--- a/fs/xfs/xfs_buf.c
+++ b/fs/xfs/xfs_buf.c
@@ -2158,3 +2158,37 @@ xfs_verify_magic16(
 		return false;
 	return dmagic == bp->b_ops->magic16[idx];
 }
+
+/* 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_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;
+}
diff --git a/fs/xfs/xfs_buf.h b/fs/xfs/xfs_buf.h
index d79a1fe5d738..1c2640771b9e 100644
--- a/fs/xfs/xfs_buf.h
+++ b/fs/xfs/xfs_buf.h
@@ -348,6 +348,7 @@ extern xfs_buftarg_t *xfs_alloc_buftarg(struct xfs_mount *,
 extern void xfs_free_buftarg(struct xfs_buftarg *);
 extern void xfs_wait_buftarg(xfs_buftarg_t *);
 extern int xfs_setsize_buftarg(xfs_buftarg_t *, unsigned int);
+unsigned int xfs_buftarg_guess_threads(struct xfs_buftarg *btp);
 
 #define xfs_getsize_buftarg(buftarg)	block_size((buftarg)->bt_bdev)
 #define xfs_readonly_buftarg(buftarg)	bdev_read_only((buftarg)->bt_bdev)
diff --git a/fs/xfs/xfs_iwalk.c b/fs/xfs/xfs_iwalk.c
index 233dcc8784db..f2f89cae1f3c 100644
--- a/fs/xfs/xfs_iwalk.c
+++ b/fs/xfs/xfs_iwalk.c
@@ -605,7 +605,7 @@ xfs_iwalk_threaded(
 	ASSERT(agno < mp->m_sb.sb_agcount);
 	ASSERT(!(flags & ~XFS_IWALK_FLAGS_ALL));
 
-	nr_threads = xfs_pwork_guess_datadev_parallelism(mp);
+	nr_threads = xfs_pwork_guess_threads(mp);
 	error = xfs_pwork_init(mp, &pctl, xfs_iwalk_ag_work, "xfs_iwalk",
 			nr_threads);
 	if (error)
diff --git a/fs/xfs/xfs_mount.c b/fs/xfs/xfs_mount.c
index 56efe140c923..ad57b4403587 100644
--- a/fs/xfs/xfs_mount.c
+++ b/fs/xfs/xfs_mount.c
@@ -1408,3 +1408,42 @@ xfs_mod_delalloc(
 	percpu_counter_add_batch(&mp->m_delalloc_blks, delta,
 			XFS_DELALLOC_BATCH);
 }
+
+/*
+ * Estimate the amount of parallelism that is available for metadata operations
+ * on this filesystem.
+ */
+unsigned int
+xfs_guess_metadata_threads(
+	struct xfs_mount	*mp)
+{
+	unsigned int		threads;
+
+	/*
+	 * Estimate the amount of parallelism for metadata operations from the
+	 * least capable of the two devices that handle metadata.  Cap that
+	 * estimate to the number of AGs to avoid unnecessary lock contention.
+	 */
+	threads = xfs_buftarg_guess_threads(mp->m_ddev_targp);
+	if (mp->m_logdev_targp != mp->m_ddev_targp)
+		threads = min(xfs_buftarg_guess_threads(mp->m_logdev_targp),
+			      threads);
+	threads = min(mp->m_sb.sb_agcount, threads);
+
+	/* If the storage told us it has fancy capabilities, we're done. */
+	if (threads > 1)
+		goto clamp;
+
+	/*
+	 * Metadata storage did not even hint that it has any parallel
+	 * capability.  If the filesystem was formatted with a stripe unit and
+	 * width, we'll treat that as evidence of a RAID setup and estimate
+	 * the number of disks.
+	 */
+	if (mp->m_sb.sb_unit > 0 && mp->m_sb.sb_width > mp->m_sb.sb_unit)
+		threads = mp->m_sb.sb_width / mp->m_sb.sb_unit;
+
+clamp:
+	/* Don't return an estimate larger than the CPU count. */
+	return min(num_online_cpus(), threads);
+}
diff --git a/fs/xfs/xfs_mount.h b/fs/xfs/xfs_mount.h
index 88ab09ed29e7..7664af01af69 100644
--- a/fs/xfs/xfs_mount.h
+++ b/fs/xfs/xfs_mount.h
@@ -412,5 +412,6 @@ struct xfs_error_cfg * xfs_error_get_cfg(struct xfs_mount *mp,
 		int error_class, int error);
 void xfs_force_summary_recalc(struct xfs_mount *mp);
 void xfs_mod_delalloc(struct xfs_mount *mp, int64_t delta);
+unsigned int xfs_guess_metadata_threads(struct xfs_mount *mp);
 
 #endif	/* __XFS_MOUNT_H__ */
diff --git a/fs/xfs/xfs_pwork.c b/fs/xfs/xfs_pwork.c
index 4bcc3e61056c..5f1a5e575a48 100644
--- a/fs/xfs/xfs_pwork.c
+++ b/fs/xfs/xfs_pwork.c
@@ -118,19 +118,12 @@ xfs_pwork_poll(
 		touch_softlockup_watchdog();
 }
 
-/*
- * Return the amount of parallelism that the data device can handle, or 0 for
- * no limit.
- */
+/* Estimate how many threads we need for a parallel work queue. */
 unsigned int
-xfs_pwork_guess_datadev_parallelism(
+xfs_pwork_guess_threads(
 	struct xfs_mount	*mp)
 {
-	struct xfs_buftarg	*btp = mp->m_ddev_targp;
-
-	/*
-	 * For now we'll go with the most conservative setting possible,
-	 * which is two threads for an SSD and 1 thread everywhere else.
-	 */
-	return blk_queue_nonrot(btp->bt_bdev->bd_queue) ? 2 : 1;
+	/* pwork queues are not unbounded, so we have to abide WQ_MAX_ACTIVE. */
+	return min_t(unsigned int, xfs_guess_metadata_threads(mp),
+			WQ_MAX_ACTIVE);
 }
diff --git a/fs/xfs/xfs_pwork.h b/fs/xfs/xfs_pwork.h
index 8133124cf3bb..f402920f7061 100644
--- a/fs/xfs/xfs_pwork.h
+++ b/fs/xfs/xfs_pwork.h
@@ -56,6 +56,6 @@ int xfs_pwork_init(struct xfs_mount *mp, struct xfs_pwork_ctl *pctl,
 void xfs_pwork_queue(struct xfs_pwork_ctl *pctl, struct xfs_pwork *pwork);
 int xfs_pwork_destroy(struct xfs_pwork_ctl *pctl);
 void xfs_pwork_poll(struct xfs_pwork_ctl *pctl);
-unsigned int xfs_pwork_guess_datadev_parallelism(struct xfs_mount *mp);
+unsigned int xfs_pwork_guess_threads(struct xfs_mount *mp);
 
 #endif /* __XFS_PWORK_H__ */