[Cluster-devel] [GFS2 PATCH] GFS2: Rework rgrp glock congestion functions for intra-node

[Bob Peterson <rpeterso@redhat.com>]

Hi,

This patch reworks the "congestion" algorithms for resource group
glocks to take into account intra-node demands as well as inter-
node demands.

Signed-off-by: Bob Peterson <rpeterso@redhat.com>
---
diff --git a/fs/gfs2/rgrp.c b/fs/gfs2/rgrp.c
index ad62bfb..903f58b 100644
--- a/fs/gfs2/rgrp.c
+++ b/fs/gfs2/rgrp.c
@@ -1807,6 +1807,21 @@ static void try_rgrp_unlink(struct gfs2_rgrpd *rgd, u64 *last_unlinked, u64 skip
 }
 
 /**
+ * gfs2_rgrp_used_recently
+ * @rs: The block reservation with the rgrp to test
+ * @msecs: The time limit in milliseconds
+ *
+ * Returns: True if the rgrp glock has been used within the time limit
+ */
+static inline bool gfs2_rgrp_used_recently(const struct gfs2_rgrpd *rgd,
+					   u64 msecs)
+{
+	return (ktime_before(ktime_get_real(),
+			     ktime_add(rgd->rd_gl->gl_dstamp,
+				       ms_to_ktime(msecs))));
+}
+
+/**
  * gfs2_rgrp_congested - Use stats to figure out whether an rgrp is congested
  * @rgd: The rgrp in question
  * @loops: An indication of how picky we can be (0=very, 1=less so)
@@ -1833,7 +1848,7 @@ static void try_rgrp_unlink(struct gfs2_rgrpd *rgd, u64 *last_unlinked, u64 skip
  * Returns: A boolean verdict on the congestion status
  */
 
-static bool gfs2_rgrp_congested(const struct gfs2_rgrpd *rgd, int loops)
+static bool gfs2_rgrp_congested_dlm(const struct gfs2_rgrpd *rgd, int loops)
 {
 	const struct gfs2_glock *gl = rgd->rd_gl;
 	const struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
@@ -1845,6 +1860,11 @@ static bool gfs2_rgrp_congested(const struct gfs2_rgrpd *rgd, int loops)
 	u64 var;
 	int cpu, nonzero = 0;
 
+	/* If it hasn't been used recently we can't judge the statistics, so
+	   assume it's not congested. */
+	if (!gfs2_rgrp_used_recently(rgd, HZ))
+		return false;
+
 	preempt_disable();
 	for_each_present_cpu(cpu) {
 		st = &per_cpu_ptr(sdp->sd_lkstats, cpu)->lkstats[LM_TYPE_RGRP];
@@ -1880,21 +1900,66 @@ static bool gfs2_rgrp_congested(const struct gfs2_rgrpd *rgd, int loops)
 }
 
 /**
- * gfs2_rgrp_used_recently
- * @rs: The block reservation with the rgrp to test
- * @msecs: The time limit in milliseconds
+ * fast_to_acquire - determine if a resource group will be fast to acquire
  *
- * Returns: True if the rgrp glock has been used within the time limit
+ * If this is one of our preferred rgrps, it should be quicker to acquire,
+ * because we tried to set ourselves up as dlm lock master.
+ */
+static inline bool fast_to_acquire(const struct gfs2_rgrpd *rgd)
+{
+	struct gfs2_glock *gl = rgd->rd_gl;
+
+	if (gl->gl_state != LM_ST_UNLOCKED && list_empty(&gl->gl_holders) &&
+	    !test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags) &&
+	    !test_bit(GLF_DEMOTE, &gl->gl_flags))
+		return true;
+	if (rgd->rd_flags & GFS2_RDF_PREFERRED)
+		return true;
+	return false;
+}
+
+/**
+ * gfs2_rgrp_congested - decide whether a rgrp glock is congested
+ * @rgd: The rgrp in question
+ * @loops: An indication of how picky we can be (0=very, 1=less so)
+ *
+ * There are two kinds of congestion: inter-node and intra-node.
+ *
+ * Inter-node congestion is where multiple nodes all want to allocate blocks
+ * inside the same rgrp, which means they need to trade the rgrp glock back
+ * and forth, which is a slow process. To mitigate this, we use glock
+ * statistics to predict whether the glock is historically fast to acquire.
+ *
+ * Intra-node congestion is where you have multiple processes on the same
+ * node, all trying to allocate blocks in the same rgrp. There's nothing wrong
+ * with doing so, but each process needs to wait for the other to release the
+ * rgrp glock before it may proceed. We can predict whether a rgrp glock is
+ * congested by how many block reservations are currently attached.
+ *
+ * Both kinds of congestion can hurt performance, but it's faster to check
+ * intra-node, so we do that first. After all, why bother to check if we can
+ * get the glock quickly from DLM if other processes have also used that
+ * same reasoning.
+ *
+ * We know the number of loops we've been around, so we know how desperate we
+ * are to find something. On first loop, call it congested if anyone else has
+ * a block reservation. On second loop, call it congested if it's not fast to
+ * acquire.
  */
-static bool gfs2_rgrp_used_recently(const struct gfs2_blkreserv *rs,
-				    u64 msecs)
+static bool gfs2_rgrp_congested(const struct gfs2_rgrpd *rgd, int loops)
 {
-	u64 tdiff;
+	/* Check for intra-node congestion */
+	if (loops == 0 && !RB_EMPTY_ROOT(&rgd->rd_rstree))
+		return true;
+
+	if (loops == 1 && !fast_to_acquire(rgd))
+		return true;
 
-	tdiff = ktime_to_ns(ktime_sub(ktime_get_real(),
-                            rs->rs_rbm.rgd->rd_gl->gl_dstamp));
+	/* Check for inter-node congestion */
+	if (rgd->rd_sbd->sd_lockstruct.ls_ops->lm_lock) /* lock_dlm */
+		return gfs2_rgrp_congested_dlm(rgd, loops);
 
-	return tdiff > (msecs * 1000 * 1000);
+	return false;
 }
 
 static u32 gfs2_orlov_skip(const struct gfs2_inode *ip)
@@ -1921,25 +1986,6 @@ static bool gfs2_select_rgrp(struct gfs2_rgrpd **pos, const struct gfs2_rgrpd *b
 }
 
 /**
- * fast_to_acquire - determine if a resource group will be fast to acquire
- *
- * If this is one of our preferred rgrps, it should be quicker to acquire,
- * because we tried to set ourselves up as dlm lock master.
- */
-static inline int fast_to_acquire(struct gfs2_rgrpd *rgd)
-{
-	struct gfs2_glock *gl = rgd->rd_gl;
-
-	if (gl->gl_state != LM_ST_UNLOCKED && list_empty(&gl->gl_holders) &&
-	    !test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags) &&
-	    !test_bit(GLF_DEMOTE, &gl->gl_flags))
-		return 1;
-	if (rgd->rd_flags & GFS2_RDF_PREFERRED)
-		return 1;
-	return 0;
-}
-
-/**
  * gfs2_inplace_reserve - Reserve space in the filesystem
  * @ip: the inode to reserve space for
  * @ap: the allocation parameters
@@ -1995,7 +2041,6 @@ int gfs2_inplace_reserve(struct gfs2_inode *ip, struct gfs2_alloc_parms *ap)
 				    !fast_to_acquire(rs->rs_rbm.rgd))
 					goto next_rgrp;
 				if ((loops < 2) &&
-				    gfs2_rgrp_used_recently(rs, 1000) &&
 				    gfs2_rgrp_congested(rs->rs_rbm.rgd, loops))
 					goto next_rgrp;
 			}