Skip to contentSkip to conversation

srcu: Optimize when srcu_gp_start_if_needed() holds read lock

PATCH In review Pingfan Liu 3 participants 20 messages Nov 23, 2022 – Nov 29, 2022

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

v2 cover srcu: Optimize when srcu_gp_start_if_needed() holds read lock

This series optimizes and remove some code under the condition that srcu_gp_start_if_needed() holds the srcu read lock.

v1 -> v2:
add debug check in [1/2]
drop original "[PATCH 2/2] srcu: Remove needless updating of
srcu_have_cbs in srcu_gp_end()"

Pingfan Liu (2):
srcu: Remove needless rcu_seq_done() check while holding read lock
srcu: Eliminate the case that snp_seq bigger than snap in
srcu_funnel_gp_start()

Cc: Lai Jiangshan <[email protected]>
Cc: "Paul E. McKenney" <[email protected]>
Cc: Frederic Weisbecker <[email protected]>
Cc: Josh Triplett <[email protected]>
Cc: Steven Rostedt <[email protected]>
Cc: Mathieu Desnoyers <[email protected]>
To: [email protected]

 include/linux/rcupdate.h |  1 +
 kernel/rcu/srcutree.c    | 27 +++++++++++++++------------
 2 files changed, 16 insertions(+), 12 deletions(-)
Signature · 1 line
2.31.1
#1 lore raw

PATCH v2 1/2 srcu: Remove needless rcu_seq_done() check while holding read lock

As the code changes, now, srcu_gp_start_if_needed() holds the srcu read lock, gets the gp_seq snap and call srcu_funnel_gp_start() passing that snap value.

As the rcu_seq_snap() promises "a full grace period has elapsed since the current time". In srcu_funnel_gp_start(), the statement
rcu_seq_done(&ssp->srcu_gp_seq, s)
always return false.

The same condition stands for srcu_funnel_exp_start(). Hence removing all the checks of rcu_seq_done().

Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]

---
 kernel/rcu/srcutree.c | 16 +++++++++-------
 1 file changed, 9 insertions(+), 7 deletions(-)
kernel/rcu/srcutree.c +9 -7
index 1c304fec89c0..2fc0e775ade4 100644
@@ -851,8 +851,7 @@ static void srcu_funnel_exp_start(struct srcu_struct *ssp, struct srcu_node *snp
 	if (snp)
 		for (; snp != NULL; snp = snp->srcu_parent) {
 			sgsne = READ_ONCE(snp->srcu_gp_seq_needed_exp);
-			if (rcu_seq_done(&ssp->srcu_gp_seq, s) ||
-			    (!srcu_invl_snp_seq(sgsne) && ULONG_CMP_GE(sgsne, s)))
+			if (!srcu_invl_snp_seq(sgsne) && ULONG_CMP_GE(sgsne, s))
 				return;
 			spin_lock_irqsave_rcu_node(snp, flags);
 			sgsne = snp->srcu_gp_seq_needed_exp;
@@ -878,6 +877,9 @@ static void srcu_funnel_exp_start(struct srcu_struct *ssp, struct srcu_node *snp
  *
  * Note that this function also does the work of srcu_funnel_exp_start(),
  * in some cases by directly invoking it.
+ *
+ * The srcu read lock should be hold around this function. And s is a seq snap
+ * after holding that lock.
  */
 static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
 				 unsigned long s, bool do_norm)
@@ -895,11 +897,12 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
 	else
 		snp_leaf = sdp->mynode;
 
+	BUG_ON(!srcu_read_lock_held(ssp));
+	WARN_ON_ONCE(rcu_seq_done(&ssp->srcu_gp_seq, s));
+
 	if (snp_leaf)
 		/* Each pass through the loop does one level of the srcu_node tree. */
 		for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
-			if (rcu_seq_done(&ssp->srcu_gp_seq, s) && snp != snp_leaf)
-				return; /* GP already done and CBs recorded. */
 			spin_lock_irqsave_rcu_node(snp, flags);
 			snp_seq = snp->srcu_have_cbs[idx];
 			if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
@@ -935,9 +938,8 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
 	if (!do_norm && ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, s))
 		WRITE_ONCE(ssp->srcu_gp_seq_needed_exp, s);
 
-	/* If grace period not already done and none in progress, start it. */
-	if (!rcu_seq_done(&ssp->srcu_gp_seq, s) &&
-	    rcu_seq_state(ssp->srcu_gp_seq) == SRCU_STATE_IDLE) {
+	/* If grace period not already in progress, start it. */
+	if (rcu_seq_state(ssp->srcu_gp_seq) == SRCU_STATE_IDLE) {
 		WARN_ON_ONCE(ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed));
 		srcu_gp_start(ssp);
 
Signature · 1 line
2.31.1
maintainer
#2 lore raw

Re: [PATCHv2 1/2] srcu: Remove needless rcu_seq_done() check while holding read lock

Quoted 12 lines · Pingfan Liu

As the code changes, now, srcu_gp_start_if_needed() holds the srcu read lock, gets the gp_seq snap and call srcu_funnel_gp_start() passing that snap value.

As the rcu_seq_snap() promises "a full grace period has elapsed since the current time". In srcu_funnel_gp_start(), the statement
rcu_seq_done(&ssp->srcu_gp_seq, s)
always return false.

The same condition stands for srcu_funnel_exp_start(). Hence removing all the checks of rcu_seq_done().

Quoted 8 lines
Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]

Thank you! I have queued this with some wordsmithing and modifications, as shown below. Please check these.

Also, please test future patches with the debug options shown in Documentation/process/submit-checklist.rst. In this case, testing with CONFIG_PROVE_RCU=y would have located a bug with this repeat-by:

tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration 2m --configs "SRCU-P SRCU-N" --trust-make

Thanx, Paul

------------------------------------------------------------------------

commit 17a874a71faed9e38b69c6ec7bae6aa309a2530e

Author: Pingfan Liu <[email protected]>
Date:   Wed Nov 23 21:56:37 2022 +0800

srcu: Remove needless rcu_seq_done() check while holding read lock

The srcu_gp_start_if_needed() function now read-holds the srcu_struct whose grace period is being started, which means that the corresponding SRCU grace period cannot end. This in turn means that the SRCU grace-period sequence number returned by rcu_seq_snap() cannot expire during this time. And that means that the calls to rcu_seq_done() in srcu_funnel_exp_start() and srcu_funnel_gp_start() can never return true.

This commit therefore removes these rcu_seq_done() checks, but adds checks in kernels built with CONFIG_PROVE_RCU=y that splats if rcu_seq_done() does somehow return true.

[ paulmck: Rearrange checks to handle kernels built with lockdep. ]

Cc Lai Jiangshan <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]

kernel/rcu/srcutree.c +11 -7
index 6af0312005801..b25f77df9e5e7 100644
@@ -915,8 +915,9 @@ static void srcu_funnel_exp_start(struct srcu_struct *ssp, struct srcu_node *snp
 	if (snp)
 		for (; snp != NULL; snp = snp->srcu_parent) {
 			sgsne = READ_ONCE(snp->srcu_gp_seq_needed_exp);
-			if (rcu_seq_done(&ssp->srcu_gp_seq, s) ||
-			    (!srcu_invl_snp_seq(sgsne) && ULONG_CMP_GE(sgsne, s)))
+			WARN_ON_ONCE(IS_ENABLED(CONFIG_PROVE_RCU) &&
+				     rcu_seq_done(&ssp->srcu_gp_seq, s));
+			if (!srcu_invl_snp_seq(sgsne) && ULONG_CMP_GE(sgsne, s))
 				return;
 			spin_lock_irqsave_rcu_node(snp, flags);
 			sgsne = snp->srcu_gp_seq_needed_exp;
@@ -942,6 +943,9 @@ static void srcu_funnel_exp_start(struct srcu_struct *ssp, struct srcu_node *snp
  *
  * Note that this function also does the work of srcu_funnel_exp_start(),
  * in some cases by directly invoking it.
+ *
+ * The srcu read lock should be hold around this function. And s is a seq snap
+ * after holding that lock.
  */
 static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
 				 unsigned long s, bool do_norm)
@@ -962,8 +966,8 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
 	if (snp_leaf)
 		/* Each pass through the loop does one level of the srcu_node tree. */
 		for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
-			if (rcu_seq_done(&ssp->srcu_gp_seq, s) && snp != snp_leaf)
-				return; /* GP already done and CBs recorded. */
+			WARN_ON_ONCE(IS_ENABLED(CONFIG_PROVE_RCU) &&
+				     rcu_seq_done(&ssp->srcu_gp_seq, s));
 			spin_lock_irqsave_rcu_node(snp, flags);
 			snp_seq = snp->srcu_have_cbs[idx];
 			if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
@@ -999,9 +1003,9 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
 	if (!do_norm && ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, s))
 		WRITE_ONCE(ssp->srcu_gp_seq_needed_exp, s);
 
-	/* If grace period not already done and none in progress, start it. */
-	if (!rcu_seq_done(&ssp->srcu_gp_seq, s) &&
-	    rcu_seq_state(ssp->srcu_gp_seq) == SRCU_STATE_IDLE) {
+	/* If grace period not already in progress, start it. */
+	WARN_ON_ONCE(IS_ENABLED(CONFIG_PROVE_RCU) && rcu_seq_done(&ssp->srcu_gp_seq, s));
+	if (rcu_seq_state(ssp->srcu_gp_seq) == SRCU_STATE_IDLE) {
 		WARN_ON_ONCE(ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed));
 		srcu_gp_start(ssp);
 
#3 lore raw

PATCH v2 2/2 srcu: Eliminate the case that snp_seq bigger than snap in srcu_funnel_gp_start()

In reply to Pingfan Liu

Since the srcu read lock is still held during srcu_funnel_gp_start(), the seq snap should be the largest number for the slot srcu_have_cbs[idx].

Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]

---
 include/linux/rcupdate.h |  1 +
 kernel/rcu/srcutree.c    | 11 ++++++-----
 2 files changed, 7 insertions(+), 5 deletions(-)
include/linux/rcupdate.h +1 -0
index 08605ce7379d..a09007236660 100644
@@ -32,6 +32,7 @@
 #include <linux/context_tracking_irq.h>
 
 #define ULONG_CMP_GE(a, b)	(ULONG_MAX / 2 >= (a) - (b))
+#define ULONG_CMP_GT(a, b)	(ULONG_MAX / 2 > (a) - (b))
 #define ULONG_CMP_LT(a, b)	(ULONG_MAX / 2 < (a) - (b))
 #define ulong2long(a)		(*(long *)(&(a)))
 #define USHORT_CMP_GE(a, b)	(USHRT_MAX / 2 >= (unsigned short)((a) - (b)))
kernel/rcu/srcutree.c +6 -5
index 2fc0e775ade4..41902b823687 100644
@@ -905,14 +905,15 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
 		for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
 			spin_lock_irqsave_rcu_node(snp, flags);
 			snp_seq = snp->srcu_have_cbs[idx];
-			if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
+			/*
+			 * s should be the biggest in the current slot. Hence only LE is
+			 * valid
+			 */
+			BUG_ON(ULONG_CMP_GT(snp_seq, s));
+			if (!srcu_invl_snp_seq(snp_seq) && (snp_seq == s)) {
 				if (snp == snp_leaf && snp_seq == s)
 					snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
 				spin_unlock_irqrestore_rcu_node(snp, flags);
-				if (snp == snp_leaf && snp_seq != s) {
-					srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
-					return;
-				}
 				if (!do_norm)
 					srcu_funnel_exp_start(ssp, snp, s);
 				return;
Signature · 1 line
2.31.1
maintainer
#4 lore raw

Re: [PATCHv2 2/2] srcu: Eliminate the case that snp_seq bigger than snap in srcu_funnel_gp_start()

Quoted 22 lines · Pingfan Liu

Since the srcu read lock is still held during srcu_funnel_gp_start(), the seq snap should be the largest number for the slot srcu_have_cbs[idx].

Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]

---
 include/linux/rcupdate.h |  1 +
 kernel/rcu/srcutree.c    | 11 ++++++-----
 2 files changed, 7 insertions(+), 5 deletions(-)
include/linux/rcupdate.h +0 -0
index 08605ce7379d..a09007236660 100644

include/linux/rcupdate.h line 35

Quoted 4 lines
include/linux/rcupdate.h +1 -0
@@ -32,6 +32,7 @@
 #include <linux/context_tracking_irq.h>
 
 #define ULONG_CMP_GE(a, b)	(ULONG_MAX / 2 >= (a) - (b))
+#define ULONG_CMP_GT(a, b)	(ULONG_MAX / 2 > (a) - (b))

Please see below...

Quoted 8 lines
include/linux/rcupdate.h +0 -0
 #define ULONG_CMP_LT(a, b)	(ULONG_MAX / 2 < (a) - (b))
 #define ulong2long(a)		(*(long *)(&(a)))
 #define USHORT_CMP_GE(a, b)	(USHRT_MAX / 2 >= (unsigned short)((a) - (b)))
kernel/rcu/srcutree.c +0 -0
index 2fc0e775ade4..41902b823687 100644

kernel/rcu/srcutree.c static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp, · line 912

Quoted 9 lines
kernel/rcu/srcutree.c +5 -1
@@ -905,14 +905,15 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
 		for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
 			spin_lock_irqsave_rcu_node(snp, flags);
 			snp_seq = snp->srcu_have_cbs[idx];
-			if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
+			/*
+			 * s should be the biggest in the current slot. Hence only LE is
+			 * valid
+			 */
+			BUG_ON(ULONG_CMP_GT(snp_seq, s));

Why not this? (Plus adjusting the comment above, of course.)

			WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

That way we don't need ULONG_CMP_GT(). Plus if we are confused about s being the biggest in the current slot, we get a splat and can debug further. We both might be quite sure that we are not confused, but that is exactly when we are most prone to making mistakes. ;-)

I also ask that you run with this check for some time. After all, if the assumption is incorrect, the resulting SRCU hangs off in various systems around the world will not be so much fun to debug.

Given that this is slowpath code, it is much better to take an extra compare and branch than to introduce even an extremely small risk of hanging SRCU.

Thanx, Paul

Quoted 14 lines
kernel/rcu/srcutree.c +1 -4
+			if (!srcu_invl_snp_seq(snp_seq) && (snp_seq == s)) {
 				if (snp == snp_leaf && snp_seq == s)
 					snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
 				spin_unlock_irqrestore_rcu_node(snp, flags);
-				if (snp == snp_leaf && snp_seq != s) {
-					srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
-					return;
-				}
 				if (!do_norm)
 					srcu_funnel_exp_start(ssp, snp, s);
 				return;
Signature · 1 line
2.31.1
#5 lore raw

Re: [PATCHv2 2/2] srcu: Eliminate the case that snp_seq bigger than snap in srcu_funnel_gp_start()

Quoted 50 lines · Paul E. McKenney
Quoted 22 lines · Pingfan Liu

Since the srcu read lock is still held during srcu_funnel_gp_start(), the seq snap should be the largest number for the slot srcu_have_cbs[idx].

Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]

---
 include/linux/rcupdate.h |  1 +
 kernel/rcu/srcutree.c    | 11 ++++++-----
 2 files changed, 7 insertions(+), 5 deletions(-)
include/linux/rcupdate.h +0 -0
index 08605ce7379d..a09007236660 100644

include/linux/rcupdate.h line 35

Quoted 4 lines
include/linux/rcupdate.h +1 -0
@@ -32,6 +32,7 @@
 #include <linux/context_tracking_irq.h>
 
 #define ULONG_CMP_GE(a, b)	(ULONG_MAX / 2 >= (a) - (b))
+#define ULONG_CMP_GT(a, b)	(ULONG_MAX / 2 > (a) - (b))

Please see below...

Quoted 8 lines
include/linux/rcupdate.h +0 -0
 #define ULONG_CMP_LT(a, b)	(ULONG_MAX / 2 < (a) - (b))
 #define ulong2long(a)		(*(long *)(&(a)))
 #define USHORT_CMP_GE(a, b)	(USHRT_MAX / 2 >= (unsigned short)((a) - (b)))
kernel/rcu/srcutree.c +0 -0
index 2fc0e775ade4..41902b823687 100644

kernel/rcu/srcutree.c static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp, · line 912

Quoted 9 lines
kernel/rcu/srcutree.c +5 -1
@@ -905,14 +905,15 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
 		for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
 			spin_lock_irqsave_rcu_node(snp, flags);
 			snp_seq = snp->srcu_have_cbs[idx];
-			if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
+			/*
+			 * s should be the biggest in the current slot. Hence only LE is
+			 * valid
+			 */
+			BUG_ON(ULONG_CMP_GT(snp_seq, s));

Why not this? (Plus adjusting the comment above, of course.)

Quoted 2 lines
			WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

A neat solution!

And what about the comment

			/* s should be the biggest in the current slot. */
Quoted 5 lines

That way we don't need ULONG_CMP_GT(). Plus if we are confused about s being the biggest in the current slot, we get a splat and can debug further. We both might be quite sure that we are not confused, but that is exactly when we are most prone to making mistakes. ;-)

Same feeling :)

Quoted 1 line

I also ask that you run with this check for some time. After all, if

Sure. I will try it immediately.

Quoted 7 lines

the assumption is incorrect, the resulting SRCU hangs off in various systems around the world will not be so much fun to debug.

Given that this is slowpath code, it is much better to take an extra compare and branch than to introduce even an extremely small risk of hanging SRCU.

Agree.

Thanks,

Pingfan

Quoted 16 lines

Thanx, Paul

Quoted 14 lines
kernel/rcu/srcutree.c +1 -4
+			if (!srcu_invl_snp_seq(snp_seq) && (snp_seq == s)) {
 				if (snp == snp_leaf && snp_seq == s)
 					snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
 				spin_unlock_irqrestore_rcu_node(snp, flags);
-				if (snp == snp_leaf && snp_seq != s) {
-					srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
-					return;
-				}
 				if (!do_norm)
 					srcu_funnel_exp_start(ssp, snp, s);
 				return;
Signature · 1 line
2.31.1
#6 lore raw

Re: [PATCHv2 2/2] srcu: Eliminate the case that snp_seq bigger than snap in srcu_funnel_gp_start()

Quoted 60 lines · Pingfan Liu
Quoted 50 lines · Paul E. McKenney
Quoted 22 lines · Pingfan Liu

Since the srcu read lock is still held during srcu_funnel_gp_start(), the seq snap should be the largest number for the slot srcu_have_cbs[idx].

Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]

---
 include/linux/rcupdate.h |  1 +
 kernel/rcu/srcutree.c    | 11 ++++++-----
 2 files changed, 7 insertions(+), 5 deletions(-)
include/linux/rcupdate.h +0 -0
index 08605ce7379d..a09007236660 100644

include/linux/rcupdate.h line 35

Quoted 4 lines
include/linux/rcupdate.h +1 -0
@@ -32,6 +32,7 @@
 #include <linux/context_tracking_irq.h>
 
 #define ULONG_CMP_GE(a, b) (ULONG_MAX / 2 >= (a) - (b))
+#define ULONG_CMP_GT(a, b) (ULONG_MAX / 2 > (a) - (b))

Please see below...

Quoted 8 lines
include/linux/rcupdate.h +0 -0
 #define ULONG_CMP_LT(a, b) (ULONG_MAX / 2 < (a) - (b))
 #define ulong2long(a)              (*(long *)(&(a)))
 #define USHORT_CMP_GE(a, b)        (USHRT_MAX / 2 >= (unsigned short)((a) - (b)))
kernel/rcu/srcutree.c +0 -0
index 2fc0e775ade4..41902b823687 100644

kernel/rcu/srcutree.c static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp, · line 912

Quoted 9 lines
kernel/rcu/srcutree.c +5 -1
@@ -905,14 +905,15 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
            for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
                    spin_lock_irqsave_rcu_node(snp, flags);
                    snp_seq = snp->srcu_have_cbs[idx];
-                   if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
+                   /*
+                    * s should be the biggest in the current slot. Hence only LE is
+                    * valid
+                    */
+                   BUG_ON(ULONG_CMP_GT(snp_seq, s));

Why not this? (Plus adjusting the comment above, of course.)

Quoted 2 lines
                      WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

A neat solution!

And what about the comment

                        /* s should be the biggest in the current slot. */
Quoted 10 lines
Quoted 5 lines

That way we don't need ULONG_CMP_GT(). Plus if we are confused about s being the biggest in the current slot, we get a splat and can debug further. We both might be quite sure that we are not confused, but that is exactly when we are most prone to making mistakes. ;-)

Same feeling :)

Quoted 1 line

I also ask that you run with this check for some time. After all, if

Forget to ask if the test
"tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration
10h --configs 18*SRCU-P"
satisfies your requirement?

Thanks

Pingfan

Quoted 33 lines

Sure. I will try it immediately.

Quoted 7 lines

the assumption is incorrect, the resulting SRCU hangs off in various systems around the world will not be so much fun to debug.

Given that this is slowpath code, it is much better to take an extra compare and branch than to introduce even an extremely small risk of hanging SRCU.

Agree.

Thanks,

Pingfan

Quoted 16 lines

Thanx, Paul

Quoted 14 lines
kernel/rcu/srcutree.c +1 -5
+                   if (!srcu_invl_snp_seq(snp_seq) && (snp_seq == s)) {
                            if (snp == snp_leaf && snp_seq == s)
                                    snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
                            spin_unlock_irqrestore_rcu_node(snp, flags);
-                           if (snp == snp_leaf && snp_seq != s) {
-                                   srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
-                                   return;
-                           }
                            if (!do_norm)
                                    srcu_funnel_exp_start(ssp, snp, s);
                            return;
--

2.31.1

maintainer
#7 lore raw

Re: [PATCHv2 2/2] srcu: Eliminate the case that snp_seq bigger than snap in srcu_funnel_gp_start()

Quoted 67 lines · Pingfan Liu
Quoted 66 lines · Pingfan Liu
Quoted 50 lines · Paul E. McKenney
Quoted 22 lines · Pingfan Liu

Since the srcu read lock is still held during srcu_funnel_gp_start(), the seq snap should be the largest number for the slot srcu_have_cbs[idx].

Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]

---
 include/linux/rcupdate.h |  1 +
 kernel/rcu/srcutree.c    | 11 ++++++-----
 2 files changed, 7 insertions(+), 5 deletions(-)
include/linux/rcupdate.h +0 -0
index 08605ce7379d..a09007236660 100644

include/linux/rcupdate.h line 35

Quoted 4 lines
include/linux/rcupdate.h +1 -0
@@ -32,6 +32,7 @@
 #include <linux/context_tracking_irq.h>
 
 #define ULONG_CMP_GE(a, b) (ULONG_MAX / 2 >= (a) - (b))
+#define ULONG_CMP_GT(a, b) (ULONG_MAX / 2 > (a) - (b))

Please see below...

Quoted 8 lines
include/linux/rcupdate.h +0 -0
 #define ULONG_CMP_LT(a, b) (ULONG_MAX / 2 < (a) - (b))
 #define ulong2long(a)              (*(long *)(&(a)))
 #define USHORT_CMP_GE(a, b)        (USHRT_MAX / 2 >= (unsigned short)((a) - (b)))
kernel/rcu/srcutree.c +0 -0
index 2fc0e775ade4..41902b823687 100644

kernel/rcu/srcutree.c static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp, · line 912

Quoted 9 lines
kernel/rcu/srcutree.c +5 -1
@@ -905,14 +905,15 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
            for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
                    spin_lock_irqsave_rcu_node(snp, flags);
                    snp_seq = snp->srcu_have_cbs[idx];
-                   if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
+                   /*
+                    * s should be the biggest in the current slot. Hence only LE is
+                    * valid
+                    */
+                   BUG_ON(ULONG_CMP_GT(snp_seq, s));

Why not this? (Plus adjusting the comment above, of course.)

Quoted 2 lines
                      WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

A neat solution!

And what about the comment

                        /* s should be the biggest in the current slot. */
Quoted 5 lines

That way we don't need ULONG_CMP_GT(). Plus if we are confused about s being the biggest in the current slot, we get a splat and can debug further. We both might be quite sure that we are not confused, but that is exactly when we are most prone to making mistakes. ;-)

Quoted 8 lines
Quoted 3 lines

Same feeling :)

Quoted 1 line

I also ask that you run with this check for some time. After all, if

Forget to ask if the test
"tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration
10h --configs 18*SRCU-P"
satisfies your requirement?

I am thinking more in terms of adding the WARN_ON_ONCE(), letting it be in -next and possibly mainline for a couple of years, and then if there are no splats, start feeling more confident in the asserted relationship.

If there was a significant performance, scalability, energy-efficiency, or simplification benefit, I would feel justified in being more aggressive.

But I am not seeing a significant benefit.

Thanx, Paul

Quoted 36 lines

Thanks

Pingfan

Quoted 33 lines

Sure. I will try it immediately.

Quoted 7 lines

the assumption is incorrect, the resulting SRCU hangs off in various systems around the world will not be so much fun to debug.

Given that this is slowpath code, it is much better to take an extra compare and branch than to introduce even an extremely small risk of hanging SRCU.

Agree.

Thanks,

Pingfan

Quoted 16 lines

Thanx, Paul

Quoted 14 lines
kernel/rcu/srcutree.c +1 -5
+                   if (!srcu_invl_snp_seq(snp_seq) && (snp_seq == s)) {
                            if (snp == snp_leaf && snp_seq == s)
                                    snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
                            spin_unlock_irqrestore_rcu_node(snp, flags);
-                           if (snp == snp_leaf && snp_seq != s) {
-                                   srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
-                                   return;
-                           }
                            if (!do_norm)
                                    srcu_funnel_exp_start(ssp, snp, s);
                            return;
--

2.31.1

maintainer
#8 lore raw

Re: [PATCHv2 2/2] srcu: Eliminate the case that snp_seq bigger than snap in srcu_funnel_gp_start()

Quoted 77 lines · Paul E. McKenney
Quoted 67 lines · Pingfan Liu
Quoted 66 lines · Pingfan Liu
Quoted 50 lines · Paul E. McKenney
Quoted 22 lines · Pingfan Liu

Since the srcu read lock is still held during srcu_funnel_gp_start(), the seq snap should be the largest number for the slot srcu_have_cbs[idx].

Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]

---
 include/linux/rcupdate.h |  1 +
 kernel/rcu/srcutree.c    | 11 ++++++-----
 2 files changed, 7 insertions(+), 5 deletions(-)
include/linux/rcupdate.h +0 -0
index 08605ce7379d..a09007236660 100644

include/linux/rcupdate.h line 35

Quoted 4 lines
include/linux/rcupdate.h +1 -0
@@ -32,6 +32,7 @@
 #include <linux/context_tracking_irq.h>
 
 #define ULONG_CMP_GE(a, b) (ULONG_MAX / 2 >= (a) - (b))
+#define ULONG_CMP_GT(a, b) (ULONG_MAX / 2 > (a) - (b))

Please see below...

Quoted 8 lines
include/linux/rcupdate.h +0 -0
 #define ULONG_CMP_LT(a, b) (ULONG_MAX / 2 < (a) - (b))
 #define ulong2long(a)              (*(long *)(&(a)))
 #define USHORT_CMP_GE(a, b)        (USHRT_MAX / 2 >= (unsigned short)((a) - (b)))
kernel/rcu/srcutree.c +0 -0
index 2fc0e775ade4..41902b823687 100644

kernel/rcu/srcutree.c static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp, · line 912

Quoted 9 lines
kernel/rcu/srcutree.c +5 -1
@@ -905,14 +905,15 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
            for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
                    spin_lock_irqsave_rcu_node(snp, flags);
                    snp_seq = snp->srcu_have_cbs[idx];
-                   if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
+                   /*
+                    * s should be the biggest in the current slot. Hence only LE is
+                    * valid
+                    */
+                   BUG_ON(ULONG_CMP_GT(snp_seq, s));

Why not this? (Plus adjusting the comment above, of course.)

Quoted 2 lines
                      WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

A neat solution!

And what about the comment

                        /* s should be the biggest in the current slot. */
Quoted 5 lines

That way we don't need ULONG_CMP_GT(). Plus if we are confused about s being the biggest in the current slot, we get a splat and can debug further. We both might be quite sure that we are not confused, but that is exactly when we are most prone to making mistakes. ;-)

Quoted 8 lines
Quoted 3 lines

Same feeling :)

Quoted 1 line

I also ask that you run with this check for some time. After all, if

Forget to ask if the test
"tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration
10h --configs 18*SRCU-P"
satisfies your requirement?

Quoted 8 lines

I am thinking more in terms of adding the WARN_ON_ONCE(), letting it be in -next and possibly mainline for a couple of years, and then if there are no splats, start feeling more confident in the asserted relationship.

If there was a significant performance, scalability, energy-efficiency, or simplification benefit, I would feel justified in being more aggressive.

But I am not seeing a significant benefit.

Ah, and before I forget -again-, have you thought through the counter-wrap scenarios? Please keep in mind that on a 32-bit system, those counters can wrap quite quickly compared to typical uptimes.

Thanx, Paul

Quoted 36 lines
Quoted 36 lines

Thanks

Pingfan

Quoted 33 lines

Sure. I will try it immediately.

Quoted 7 lines

the assumption is incorrect, the resulting SRCU hangs off in various systems around the world will not be so much fun to debug.

Given that this is slowpath code, it is much better to take an extra compare and branch than to introduce even an extremely small risk of hanging SRCU.

Agree.

Thanks,

Pingfan

Quoted 16 lines

Thanx, Paul

Quoted 14 lines
kernel/rcu/srcutree.c +1 -5
+                   if (!srcu_invl_snp_seq(snp_seq) && (snp_seq == s)) {
                            if (snp == snp_leaf && snp_seq == s)
                                    snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
                            spin_unlock_irqrestore_rcu_node(snp, flags);
-                           if (snp == snp_leaf && snp_seq != s) {
-                                   srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
-                                   return;
-                           }
                            if (!do_norm)
                                    srcu_funnel_exp_start(ssp, snp, s);
                            return;
--

2.31.1

#9 lore raw

Re: [PATCHv2 2/2] srcu: Eliminate the case that snp_seq bigger than snap in srcu_funnel_gp_start()

Quoted 82 lines · Paul E. McKenney
Quoted 81 lines · Paul E. McKenney
Quoted 67 lines · Pingfan Liu
Quoted 66 lines · Pingfan Liu
Quoted 50 lines · Paul E. McKenney
Quoted 22 lines · Pingfan Liu

Since the srcu read lock is still held during srcu_funnel_gp_start(), the seq snap should be the largest number for the slot srcu_have_cbs[idx].

Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]

---
 include/linux/rcupdate.h |  1 +
 kernel/rcu/srcutree.c    | 11 ++++++-----
 2 files changed, 7 insertions(+), 5 deletions(-)
include/linux/rcupdate.h +0 -0
index 08605ce7379d..a09007236660 100644

include/linux/rcupdate.h line 35

Quoted 4 lines
include/linux/rcupdate.h +1 -0
@@ -32,6 +32,7 @@
 #include <linux/context_tracking_irq.h>
 
 #define ULONG_CMP_GE(a, b) (ULONG_MAX / 2 >= (a) - (b))
+#define ULONG_CMP_GT(a, b) (ULONG_MAX / 2 > (a) - (b))

Please see below...

Quoted 8 lines
include/linux/rcupdate.h +0 -0
 #define ULONG_CMP_LT(a, b) (ULONG_MAX / 2 < (a) - (b))
 #define ulong2long(a)              (*(long *)(&(a)))
 #define USHORT_CMP_GE(a, b)        (USHRT_MAX / 2 >= (unsigned short)((a) - (b)))
kernel/rcu/srcutree.c +0 -0
index 2fc0e775ade4..41902b823687 100644

kernel/rcu/srcutree.c static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp, · line 912

Quoted 9 lines
kernel/rcu/srcutree.c +5 -1
@@ -905,14 +905,15 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
            for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
                    spin_lock_irqsave_rcu_node(snp, flags);
                    snp_seq = snp->srcu_have_cbs[idx];
-                   if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
+                   /*
+                    * s should be the biggest in the current slot. Hence only LE is
+                    * valid
+                    */
+                   BUG_ON(ULONG_CMP_GT(snp_seq, s));

Why not this? (Plus adjusting the comment above, of course.)

Quoted 2 lines
                      WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

A neat solution!

And what about the comment

                        /* s should be the biggest in the current slot. */
Quoted 5 lines

That way we don't need ULONG_CMP_GT(). Plus if we are confused about s being the biggest in the current slot, we get a splat and can debug further. We both might be quite sure that we are not confused, but that is exactly when we are most prone to making mistakes. ;-)

Quoted 8 lines
Quoted 3 lines

Same feeling :)

Quoted 1 line

I also ask that you run with this check for some time. After all, if

Forget to ask if the test
"tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration
10h --configs 18*SRCU-P"
satisfies your requirement?

I am thinking more in terms of adding the WARN_ON_ONCE(), letting it be in -next and possibly mainline for a couple of years, and then if there are no splats, start feeling more confident in the asserted relationship.

Quoted 9 lines
Quoted 4 lines

If there was a significant performance, scalability, energy-efficiency, or simplification benefit, I would feel justified in being more aggressive.

But I am not seeing a significant benefit.

Ah, and before I forget -again-, have you thought through the counter-wrap scenarios? Please keep in mind that on a 32-bit system, those counters can wrap quite quickly compared to typical uptimes.

I think 32-bit has not significant different since the statement
WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

where 's < snp_seq' is not associated with bits. Am I missing anything?

Thanks,

Pingfan

Quoted 38 lines

Thanx, Paul

Quoted 36 lines
Quoted 36 lines

Thanks

Pingfan

Quoted 33 lines

Sure. I will try it immediately.

Quoted 7 lines

the assumption is incorrect, the resulting SRCU hangs off in various systems around the world will not be so much fun to debug.

Given that this is slowpath code, it is much better to take an extra compare and branch than to introduce even an extremely small risk of hanging SRCU.

Agree.

Thanks,

Pingfan

Quoted 16 lines

Thanx, Paul

Quoted 14 lines
kernel/rcu/srcutree.c +1 -5
+                   if (!srcu_invl_snp_seq(snp_seq) && (snp_seq == s)) {
                            if (snp == snp_leaf && snp_seq == s)
                                    snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
                            spin_unlock_irqrestore_rcu_node(snp, flags);
-                           if (snp == snp_leaf && snp_seq != s) {
-                                   srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
-                                   return;
-                           }
                            if (!do_norm)
                                    srcu_funnel_exp_start(ssp, snp, s);
                            return;
--

2.31.1

maintainer
#10 lore raw

Re: [PATCHv2 2/2] srcu: Eliminate the case that snp_seq bigger than snap in srcu_funnel_gp_start()

Quoted 87 lines · Pingfan Liu
Quoted 86 lines · Paul E. McKenney
Quoted 84 lines · Paul E. McKenney
Quoted 66 lines · Pingfan Liu
Quoted 65 lines · Pingfan Liu
Quoted 50 lines · Paul E. McKenney
Quoted 22 lines · Pingfan Liu

Since the srcu read lock is still held during srcu_funnel_gp_start(), the seq snap should be the largest number for the slot srcu_have_cbs[idx].

Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]

---
 include/linux/rcupdate.h |  1 +
 kernel/rcu/srcutree.c    | 11 ++++++-----
 2 files changed, 7 insertions(+), 5 deletions(-)
include/linux/rcupdate.h +0 -0
index 08605ce7379d..a09007236660 100644

include/linux/rcupdate.h line 35

Quoted 4 lines
include/linux/rcupdate.h +1 -0
@@ -32,6 +32,7 @@
 #include <linux/context_tracking_irq.h>
 
 #define ULONG_CMP_GE(a, b) (ULONG_MAX / 2 >= (a) - (b))
+#define ULONG_CMP_GT(a, b) (ULONG_MAX / 2 > (a) - (b))

Please see below...

Quoted 8 lines
include/linux/rcupdate.h +0 -0
 #define ULONG_CMP_LT(a, b) (ULONG_MAX / 2 < (a) - (b))
 #define ulong2long(a)              (*(long *)(&(a)))
 #define USHORT_CMP_GE(a, b)        (USHRT_MAX / 2 >= (unsigned short)((a) - (b)))
kernel/rcu/srcutree.c +0 -0
index 2fc0e775ade4..41902b823687 100644

kernel/rcu/srcutree.c static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp, · line 912

Quoted 9 lines
kernel/rcu/srcutree.c +5 -1
@@ -905,14 +905,15 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
            for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
                    spin_lock_irqsave_rcu_node(snp, flags);
                    snp_seq = snp->srcu_have_cbs[idx];
-                   if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
+                   /*
+                    * s should be the biggest in the current slot. Hence only LE is
+                    * valid
+                    */
+                   BUG_ON(ULONG_CMP_GT(snp_seq, s));

Why not this? (Plus adjusting the comment above, of course.)

Quoted 2 lines
                      WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

A neat solution!

And what about the comment

                        /* s should be the biggest in the current slot. */
Quoted 5 lines

That way we don't need ULONG_CMP_GT(). Plus if we are confused about s being the biggest in the current slot, we get a splat and can debug further. We both might be quite sure that we are not confused, but that is exactly when we are most prone to making mistakes. ;-)

Quoted 8 lines
Quoted 3 lines

Same feeling :)

Quoted 1 line

I also ask that you run with this check for some time. After all, if

Forget to ask if the test
"tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration
10h --configs 18*SRCU-P"
satisfies your requirement?

I am thinking more in terms of adding the WARN_ON_ONCE(), letting it be in -next and possibly mainline for a couple of years, and then if there are no splats, start feeling more confident in the asserted relationship.

If there was a significant performance, scalability, energy-efficiency, or simplification benefit, I would feel justified in being more aggressive.

But I am not seeing a significant benefit.

Quoted 8 lines
Quoted 3 lines

Ah, and before I forget -again-, have you thought through the counter-wrap scenarios? Please keep in mind that on a 32-bit system, those counters can wrap quite quickly compared to typical uptimes.

I think 32-bit has not significant different since the statement
WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

where 's < snp_seq' is not associated with bits. Am I missing anything?

Your point is that long before there was any danger of wrap, this WARN_ON_ONCE() would trigger?

If so, the point is to avoid that WARN_ON_ONCE() from triggering.

Or am I missing your point?

Thanx, Paul

Quoted 42 lines

Thanks,

Pingfan

Quoted 38 lines

Thanx, Paul

Quoted 36 lines
Quoted 36 lines

Thanks

Pingfan

Quoted 33 lines

Sure. I will try it immediately.

Quoted 7 lines

the assumption is incorrect, the resulting SRCU hangs off in various systems around the world will not be so much fun to debug.

Given that this is slowpath code, it is much better to take an extra compare and branch than to introduce even an extremely small risk of hanging SRCU.

Agree.

Thanks,

Pingfan

Quoted 16 lines

Thanx, Paul

Quoted 14 lines
kernel/rcu/srcutree.c +1 -5
+                   if (!srcu_invl_snp_seq(snp_seq) && (snp_seq == s)) {
                            if (snp == snp_leaf && snp_seq == s)
                                    snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
                            spin_unlock_irqrestore_rcu_node(snp, flags);
-                           if (snp == snp_leaf && snp_seq != s) {
-                                   srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
-                                   return;
-                           }
                            if (!do_norm)
                                    srcu_funnel_exp_start(ssp, snp, s);
                            return;
--

2.31.1

maintainer
#11 lore raw

Re: [PATCHv2 2/2] srcu: Eliminate the case that snp_seq bigger than snap in srcu_funnel_gp_start()

Quoted 95 lines · Paul E. McKenney
Quoted 94 lines · Pingfan Liu
Quoted 81 lines · Paul E. McKenney
Quoted 80 lines · Paul E. McKenney
Quoted 66 lines · Pingfan Liu
Quoted 65 lines · Pingfan Liu
Quoted 50 lines · Paul E. McKenney
Quoted 22 lines · Pingfan Liu

Since the srcu read lock is still held during srcu_funnel_gp_start(), the seq snap should be the largest number for the slot srcu_have_cbs[idx].

Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]

---
 include/linux/rcupdate.h |  1 +
 kernel/rcu/srcutree.c    | 11 ++++++-----
 2 files changed, 7 insertions(+), 5 deletions(-)
include/linux/rcupdate.h +0 -0
index 08605ce7379d..a09007236660 100644

include/linux/rcupdate.h line 35

Quoted 4 lines
include/linux/rcupdate.h +1 -0
@@ -32,6 +32,7 @@
 #include <linux/context_tracking_irq.h>
 
 #define ULONG_CMP_GE(a, b) (ULONG_MAX / 2 >= (a) - (b))
+#define ULONG_CMP_GT(a, b) (ULONG_MAX / 2 > (a) - (b))

Please see below...

Quoted 8 lines
include/linux/rcupdate.h +0 -0
 #define ULONG_CMP_LT(a, b) (ULONG_MAX / 2 < (a) - (b))
 #define ulong2long(a)              (*(long *)(&(a)))
 #define USHORT_CMP_GE(a, b)        (USHRT_MAX / 2 >= (unsigned short)((a) - (b)))
kernel/rcu/srcutree.c +0 -0
index 2fc0e775ade4..41902b823687 100644

kernel/rcu/srcutree.c static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp, · line 912

Quoted 9 lines
kernel/rcu/srcutree.c +5 -1
@@ -905,14 +905,15 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
            for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
                    spin_lock_irqsave_rcu_node(snp, flags);
                    snp_seq = snp->srcu_have_cbs[idx];
-                   if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
+                   /*
+                    * s should be the biggest in the current slot. Hence only LE is
+                    * valid
+                    */
+                   BUG_ON(ULONG_CMP_GT(snp_seq, s));

Why not this? (Plus adjusting the comment above, of course.)

Quoted 2 lines
                      WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

A neat solution!

And what about the comment

                        /* s should be the biggest in the current slot. */
Quoted 5 lines

That way we don't need ULONG_CMP_GT(). Plus if we are confused about s being the biggest in the current slot, we get a splat and can debug further. We both might be quite sure that we are not confused, but that is exactly when we are most prone to making mistakes. ;-)

Quoted 8 lines
Quoted 3 lines

Same feeling :)

Quoted 1 line

I also ask that you run with this check for some time. After all, if

Forget to ask if the test
"tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration
10h --configs 18*SRCU-P"
satisfies your requirement?

I am thinking more in terms of adding the WARN_ON_ONCE(), letting it be in -next and possibly mainline for a couple of years, and then if there are no splats, start feeling more confident in the asserted relationship.

Quoted 8 lines
Quoted 4 lines

If there was a significant performance, scalability, energy-efficiency, or simplification benefit, I would feel justified in being more aggressive.

But I am not seeing a significant benefit.

Ah, and before I forget -again-, have you thought through the counter-wrap scenarios? Please keep in mind that on a 32-bit system, those counters can wrap quite quickly compared to typical uptimes.

I think 32-bit has not significant different since the statement
WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

Quoted 8 lines
Quoted 1 line

where 's < snp_seq' is not associated with bits. Am I missing anything?

Your point is that long before there was any danger of wrap, this WARN_ON_ONCE() would trigger?

If so, the point is to avoid that WARN_ON_ONCE() from triggering.

Or am I missing your point?

For example, are you instead saying that the WARN_ON_ONCE() proposed in your patch will detect the problem either way? If so, what I am asking is whether your earlier analysis considered wrap without the addition of that WARN_ON_ONCE().

Thanx, Paul

Quoted 42 lines
Quoted 42 lines

Thanks,

Pingfan

Quoted 38 lines

Thanx, Paul

Quoted 36 lines
Quoted 36 lines

Thanks

Pingfan

Quoted 33 lines

Sure. I will try it immediately.

Quoted 7 lines

the assumption is incorrect, the resulting SRCU hangs off in various systems around the world will not be so much fun to debug.

Given that this is slowpath code, it is much better to take an extra compare and branch than to introduce even an extremely small risk of hanging SRCU.

Agree.

Thanks,

Pingfan

Quoted 16 lines

Thanx, Paul

Quoted 14 lines
kernel/rcu/srcutree.c +1 -5
+                   if (!srcu_invl_snp_seq(snp_seq) && (snp_seq == s)) {
                            if (snp == snp_leaf && snp_seq == s)
                                    snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
                            spin_unlock_irqrestore_rcu_node(snp, flags);
-                           if (snp == snp_leaf && snp_seq != s) {
-                                   srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
-                                   return;
-                           }
                            if (!do_norm)
                                    srcu_funnel_exp_start(ssp, snp, s);
                            return;
--

2.31.1

#12 lore raw

Re: [PATCHv2 2/2] srcu: Eliminate the case that snp_seq bigger than snap in srcu_funnel_gp_start()

Quoted 105 lines · Paul E. McKenney
Quoted 100 lines · Paul E. McKenney
Quoted 88 lines · Pingfan Liu
Quoted 86 lines · Paul E. McKenney
Quoted 84 lines · Paul E. McKenney
Quoted 66 lines · Pingfan Liu
Quoted 65 lines · Pingfan Liu
Quoted 50 lines · Paul E. McKenney
Quoted 22 lines · Pingfan Liu

Since the srcu read lock is still held during srcu_funnel_gp_start(), the seq snap should be the largest number for the slot srcu_have_cbs[idx].

Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]
---
include/linux/rcupdate.h | 1 +
kernel/rcu/srcutree.c | 11 ++++++-----

2 files changed, 7 insertions(+), 5 deletions(-)
include/linux/rcupdate.h +0 -0
index 08605ce7379d..a09007236660 100644

include/linux/rcupdate.h line 35

Quoted 4 lines
include/linux/rcupdate.h +1 -0
@@ -32,6 +32,7 @@
 #include <linux/context_tracking_irq.h>
 
 #define ULONG_CMP_GE(a, b) (ULONG_MAX / 2 >= (a) - (b))
+#define ULONG_CMP_GT(a, b) (ULONG_MAX / 2 > (a) - (b))

Please see below...

Quoted 8 lines
include/linux/rcupdate.h +0 -0
 #define ULONG_CMP_LT(a, b) (ULONG_MAX / 2 < (a) - (b))
 #define ulong2long(a)              (*(long *)(&(a)))
 #define USHORT_CMP_GE(a, b)        (USHRT_MAX / 2 >= (unsigned short)((a) - (b)))
kernel/rcu/srcutree.c +0 -0
index 2fc0e775ade4..41902b823687 100644

kernel/rcu/srcutree.c static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp, · line 912

Quoted 9 lines
kernel/rcu/srcutree.c +5 -1
@@ -905,14 +905,15 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
           for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
                   spin_lock_irqsave_rcu_node(snp, flags);
                   snp_seq = snp->srcu_have_cbs[idx];
-                   if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
+                   /*
+                    * s should be the biggest in the current slot. Hence only LE is
+                    * valid
+                    */
+                   BUG_ON(ULONG_CMP_GT(snp_seq, s));

Why not this? (Plus adjusting the comment above, of course.)

Quoted 2 lines
                     WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

A neat solution!

And what about the comment

                       /* s should be the biggest in the current slot. */
Quoted 5 lines

That way we don't need ULONG_CMP_GT(). Plus if we are confused about s being the biggest in the current slot, we get a splat and can debug further. We both might be quite sure that we are not confused, but that is exactly when we are most prone to making mistakes. ;-)

Quoted 8 lines
Quoted 3 lines

Same feeling :)

Quoted 1 line

I also ask that you run with this check for some time. After all, if

Forget to ask if the test
"tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration
10h --configs 18*SRCU-P"
satisfies your requirement?

I am thinking more in terms of adding the WARN_ON_ONCE(), letting it be in -next and possibly mainline for a couple of years, and then if there are no splats, start feeling more confident in the asserted relationship.

If there was a significant performance, scalability, energy-efficiency, or simplification benefit, I would feel justified in being more aggressive.

But I am not seeing a significant benefit.

Quoted 8 lines
Quoted 3 lines

Ah, and before I forget -again-, have you thought through the counter-wrap scenarios? Please keep in mind that on a 32-bit system, those counters can wrap quite quickly compared to typical uptimes.

I think 32-bit has not significant different since the statement
WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

where 's < snp_seq' is not associated with bits. Am I missing anything?

Your point is that long before there was any danger of wrap, this WARN_ON_ONCE() would trigger?

Quoted 8 lines
Quoted 3 lines

If so, the point is to avoid that WARN_ON_ONCE() from triggering.

Or am I missing your point?

For example, are you instead saying that the WARN_ON_ONCE() proposed in your patch will detect the problem either way? If so, what I am asking is whether your earlier analysis considered wrap without the addition of that WARN_ON_ONCE().

Do we want to make some rules on when to delete WARNs and fold in the change, such as a future date after scanning the list for reports? Such date can be embedded in a comment on top of the warning.

Thanks,

- Joel

Quoted 44 lines

Thanx, Paul

Quoted 42 lines
Quoted 42 lines

Thanks,

Pingfan

Quoted 38 lines

Thanx, Paul

Quoted 36 lines
Quoted 36 lines

Thanks

Pingfan

Quoted 33 lines

Sure. I will try it immediately.

Quoted 7 lines

the assumption is incorrect, the resulting SRCU hangs off in various systems around the world will not be so much fun to debug.

Given that this is slowpath code, it is much better to take an extra compare and branch than to introduce even an extremely small risk of hanging SRCU.

Agree.

Thanks,

Pingfan

Quoted 16 lines

Thanx, Paul

Quoted 14 lines
kernel/rcu/srcutree.c +1 -5
+                   if (!srcu_invl_snp_seq(snp_seq) && (snp_seq == s)) {
                           if (snp == snp_leaf && snp_seq == s)
                                   snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
                           spin_unlock_irqrestore_rcu_node(snp, flags);
-                           if (snp == snp_leaf && snp_seq != s) {
-                                   srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
-                                   return;
-                           }
                           if (!do_norm)
                                   srcu_funnel_exp_start(ssp, snp, s);
                           return;
--

2.31.1

maintainer
#13 lore raw

Re: [PATCHv2 2/2] srcu: Eliminate the case that snp_seq bigger than snap in srcu_funnel_gp_start()

Quoted 106 lines · Joel Fernandes
Quoted 106 lines · Paul E. McKenney
Quoted 103 lines · Paul E. McKenney
Quoted 88 lines · Pingfan Liu
Quoted 86 lines · Paul E. McKenney
Quoted 84 lines · Paul E. McKenney
Quoted 66 lines · Pingfan Liu
Quoted 65 lines · Pingfan Liu
Quoted 50 lines · Paul E. McKenney
Quoted 22 lines · Pingfan Liu

Since the srcu read lock is still held during srcu_funnel_gp_start(), the seq snap should be the largest number for the slot srcu_have_cbs[idx].

Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]
---
include/linux/rcupdate.h | 1 +
kernel/rcu/srcutree.c | 11 ++++++-----

2 files changed, 7 insertions(+), 5 deletions(-)
include/linux/rcupdate.h +0 -0
index 08605ce7379d..a09007236660 100644

include/linux/rcupdate.h line 35

Quoted 4 lines
include/linux/rcupdate.h +1 -0
@@ -32,6 +32,7 @@
 #include <linux/context_tracking_irq.h>
 
 #define ULONG_CMP_GE(a, b) (ULONG_MAX / 2 >= (a) - (b))
+#define ULONG_CMP_GT(a, b) (ULONG_MAX / 2 > (a) - (b))

Please see below...

Quoted 8 lines
include/linux/rcupdate.h +0 -0
 #define ULONG_CMP_LT(a, b) (ULONG_MAX / 2 < (a) - (b))
 #define ulong2long(a)              (*(long *)(&(a)))
 #define USHORT_CMP_GE(a, b)        (USHRT_MAX / 2 >= (unsigned short)((a) - (b)))
kernel/rcu/srcutree.c +0 -0
index 2fc0e775ade4..41902b823687 100644

kernel/rcu/srcutree.c static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp, · line 912

Quoted 9 lines
kernel/rcu/srcutree.c +5 -1
@@ -905,14 +905,15 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
           for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
                   spin_lock_irqsave_rcu_node(snp, flags);
                   snp_seq = snp->srcu_have_cbs[idx];
-                   if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
+                   /*
+                    * s should be the biggest in the current slot. Hence only LE is
+                    * valid
+                    */
+                   BUG_ON(ULONG_CMP_GT(snp_seq, s));

Why not this? (Plus adjusting the comment above, of course.)

Quoted 2 lines
                     WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

A neat solution!

And what about the comment

                       /* s should be the biggest in the current slot. */
Quoted 5 lines

That way we don't need ULONG_CMP_GT(). Plus if we are confused about s being the biggest in the current slot, we get a splat and can debug further. We both might be quite sure that we are not confused, but that is exactly when we are most prone to making mistakes. ;-)

Quoted 8 lines
Quoted 3 lines

Same feeling :)

Quoted 1 line

I also ask that you run with this check for some time. After all, if

Forget to ask if the test
"tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration
10h --configs 18*SRCU-P"
satisfies your requirement?

I am thinking more in terms of adding the WARN_ON_ONCE(), letting it be in -next and possibly mainline for a couple of years, and then if there are no splats, start feeling more confident in the asserted relationship.

If there was a significant performance, scalability, energy-efficiency, or simplification benefit, I would feel justified in being more aggressive.

But I am not seeing a significant benefit.

Quoted 8 lines
Quoted 3 lines

Ah, and before I forget -again-, have you thought through the counter-wrap scenarios? Please keep in mind that on a 32-bit system, those counters can wrap quite quickly compared to typical uptimes.

I think 32-bit has not significant different since the statement
WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

where 's < snp_seq' is not associated with bits. Am I missing anything?

Your point is that long before there was any danger of wrap, this WARN_ON_ONCE() would trigger?

If so, the point is to avoid that WARN_ON_ONCE() from triggering.

Or am I missing your point?

Quoted 8 lines
Quoted 4 lines

For example, are you instead saying that the WARN_ON_ONCE() proposed in your patch will detect the problem either way? If so, what I am asking is whether your earlier analysis considered wrap without the addition of that WARN_ON_ONCE().

Do we want to make some rules on when to delete WARNs and fold in the change, such as a future date after scanning the list for reports? Such date can be embedded in a comment on top of the warning.

Possibly, but this question is about adding a warning. If wrap was not considered, let's consider wrap and see if removing the store is appropriate. Only if removing the store is appropriate does it make sense to add the warning.

Thanx, Paul

Quoted 48 lines

Thanks,

- Joel

Quoted 44 lines

Thanx, Paul

Quoted 42 lines
Quoted 42 lines

Thanks,

Pingfan

Quoted 38 lines

Thanx, Paul

Quoted 36 lines
Quoted 36 lines

Thanks

Pingfan

Quoted 33 lines

Sure. I will try it immediately.

Quoted 7 lines

the assumption is incorrect, the resulting SRCU hangs off in various systems around the world will not be so much fun to debug.

Given that this is slowpath code, it is much better to take an extra compare and branch than to introduce even an extremely small risk of hanging SRCU.

Agree.

Thanks,

Pingfan

Quoted 16 lines

Thanx, Paul

Quoted 14 lines
kernel/rcu/srcutree.c +1 -5
+                   if (!srcu_invl_snp_seq(snp_seq) && (snp_seq == s)) {
                           if (snp == snp_leaf && snp_seq == s)
                                   snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
                           spin_unlock_irqrestore_rcu_node(snp, flags);
-                           if (snp == snp_leaf && snp_seq != s) {
-                                   srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
-                                   return;
-                           }
                           if (!do_norm)
                                   srcu_funnel_exp_start(ssp, snp, s);
                           return;
--

2.31.1

#14 lore raw

Re: [PATCHv2 2/2] srcu: Eliminate the case that snp_seq bigger than snap in srcu_funnel_gp_start()

In reply to Paul E. McKenney
Quoted 96 lines · Paul E. McKenney
Quoted 95 lines · Paul E. McKenney
Quoted 94 lines · Pingfan Liu
Quoted 81 lines · Paul E. McKenney
Quoted 80 lines · Paul E. McKenney
Quoted 66 lines · Pingfan Liu
Quoted 65 lines · Pingfan Liu
Quoted 50 lines · Paul E. McKenney
Quoted 22 lines · Pingfan Liu

Since the srcu read lock is still held during srcu_funnel_gp_start(), the seq snap should be the largest number for the slot srcu_have_cbs[idx].

Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]

---
 include/linux/rcupdate.h |  1 +
 kernel/rcu/srcutree.c    | 11 ++++++-----
 2 files changed, 7 insertions(+), 5 deletions(-)
include/linux/rcupdate.h +0 -0
index 08605ce7379d..a09007236660 100644

include/linux/rcupdate.h line 35

Quoted 4 lines
include/linux/rcupdate.h +1 -0
@@ -32,6 +32,7 @@
 #include <linux/context_tracking_irq.h>
 
 #define ULONG_CMP_GE(a, b) (ULONG_MAX / 2 >= (a) - (b))
+#define ULONG_CMP_GT(a, b) (ULONG_MAX / 2 > (a) - (b))

Please see below...

Quoted 8 lines
include/linux/rcupdate.h +0 -0
 #define ULONG_CMP_LT(a, b) (ULONG_MAX / 2 < (a) - (b))
 #define ulong2long(a)              (*(long *)(&(a)))
 #define USHORT_CMP_GE(a, b)        (USHRT_MAX / 2 >= (unsigned short)((a) - (b)))
kernel/rcu/srcutree.c +0 -0
index 2fc0e775ade4..41902b823687 100644

kernel/rcu/srcutree.c static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp, · line 912

Quoted 9 lines
kernel/rcu/srcutree.c +5 -1
@@ -905,14 +905,15 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
            for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
                    spin_lock_irqsave_rcu_node(snp, flags);
                    snp_seq = snp->srcu_have_cbs[idx];
-                   if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
+                   /*
+                    * s should be the biggest in the current slot. Hence only LE is
+                    * valid
+                    */
+                   BUG_ON(ULONG_CMP_GT(snp_seq, s));

Why not this? (Plus adjusting the comment above, of course.)

Quoted 2 lines
                      WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

A neat solution!

And what about the comment

                        /* s should be the biggest in the current slot. */
Quoted 5 lines

That way we don't need ULONG_CMP_GT(). Plus if we are confused about s being the biggest in the current slot, we get a splat and can debug further. We both might be quite sure that we are not confused, but that is exactly when we are most prone to making mistakes. ;-)

Quoted 8 lines
Quoted 3 lines

Same feeling :)

Quoted 1 line

I also ask that you run with this check for some time. After all, if

Forget to ask if the test
"tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration
10h --configs 18*SRCU-P"
satisfies your requirement?

I am thinking more in terms of adding the WARN_ON_ONCE(), letting it be in -next and possibly mainline for a couple of years, and then if there are no splats, start feeling more confident in the asserted relationship.

Quoted 8 lines
Quoted 4 lines

If there was a significant performance, scalability, energy-efficiency, or simplification benefit, I would feel justified in being more aggressive.

But I am not seeing a significant benefit.

Ah, and before I forget -again-, have you thought through the counter-wrap scenarios? Please keep in mind that on a 32-bit system, those counters can wrap quite quickly compared to typical uptimes.

I think 32-bit has not significant different since the statement
WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

Quoted 7 lines
Quoted 7 lines
Quoted 1 line

where 's < snp_seq' is not associated with bits. Am I missing anything?

Your point is that long before there was any danger of wrap, this WARN_ON_ONCE() would trigger?

If so, the point is to avoid that WARN_ON_ONCE() from triggering.

Oops! I failed to figure out the scene you pointed out, where snp_x did not have call_srcu event for a very long time, then comes the wrap, and when s is close enough to snp_x's snp_seq, ULONG_CMP_LT(s, snp_seq) will raise false alarm.

Yes, this is a bug.

Quoted 1 line
Quoted 1 line

Or am I missing your point?

Sorry that I missed your point and confused you.

Can I remedy it using the root snp? Because the root snp's snp_seq advances each GP. And it is the biggest one among snp-s.

Quoted 6 lines

For example, are you instead saying that the WARN_ON_ONCE() proposed in your patch will detect the problem either way? If so, what I am asking is whether your earlier analysis considered wrap without the addition of that WARN_ON_ONCE().

Have I answer all of your concerns?

Thanks,

Pingfan

Quoted 44 lines

Thanx, Paul

Quoted 42 lines
Quoted 42 lines

Thanks,

Pingfan

Quoted 38 lines

Thanx, Paul

Quoted 36 lines
Quoted 36 lines

Thanks

Pingfan

Quoted 33 lines

Sure. I will try it immediately.

Quoted 7 lines

the assumption is incorrect, the resulting SRCU hangs off in various systems around the world will not be so much fun to debug.

Given that this is slowpath code, it is much better to take an extra compare and branch than to introduce even an extremely small risk of hanging SRCU.

Agree.

Thanks,

Pingfan

Quoted 16 lines

Thanx, Paul

Quoted 14 lines
kernel/rcu/srcutree.c +1 -5
+                   if (!srcu_invl_snp_seq(snp_seq) && (snp_seq == s)) {
                            if (snp == snp_leaf && snp_seq == s)
                                    snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
                            spin_unlock_irqrestore_rcu_node(snp, flags);
-                           if (snp == snp_leaf && snp_seq != s) {
-                                   srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
-                                   return;
-                           }
                            if (!do_norm)
                                    srcu_funnel_exp_start(ssp, snp, s);
                            return;
--

2.31.1

maintainer
#15 lore raw

Re: [PATCHv2 2/2] srcu: Eliminate the case that snp_seq bigger than snap in srcu_funnel_gp_start()

Quoted 102 lines · Pingfan Liu
Quoted 101 lines · Paul E. McKenney
Quoted 100 lines · Paul E. McKenney
Quoted 87 lines · Pingfan Liu
Quoted 86 lines · Paul E. McKenney
Quoted 84 lines · Paul E. McKenney
Quoted 66 lines · Pingfan Liu
Quoted 65 lines · Pingfan Liu
Quoted 50 lines · Paul E. McKenney
Quoted 22 lines · Pingfan Liu

Since the srcu read lock is still held during srcu_funnel_gp_start(), the seq snap should be the largest number for the slot srcu_have_cbs[idx].

Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]

---
 include/linux/rcupdate.h |  1 +
 kernel/rcu/srcutree.c    | 11 ++++++-----
 2 files changed, 7 insertions(+), 5 deletions(-)
include/linux/rcupdate.h +0 -0
index 08605ce7379d..a09007236660 100644

include/linux/rcupdate.h line 35

Quoted 4 lines
include/linux/rcupdate.h +1 -0
@@ -32,6 +32,7 @@
 #include <linux/context_tracking_irq.h>
 
 #define ULONG_CMP_GE(a, b) (ULONG_MAX / 2 >= (a) - (b))
+#define ULONG_CMP_GT(a, b) (ULONG_MAX / 2 > (a) - (b))

Please see below...

Quoted 8 lines
include/linux/rcupdate.h +0 -0
 #define ULONG_CMP_LT(a, b) (ULONG_MAX / 2 < (a) - (b))
 #define ulong2long(a)              (*(long *)(&(a)))
 #define USHORT_CMP_GE(a, b)        (USHRT_MAX / 2 >= (unsigned short)((a) - (b)))
kernel/rcu/srcutree.c +0 -0
index 2fc0e775ade4..41902b823687 100644

kernel/rcu/srcutree.c static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp, · line 912

Quoted 9 lines
kernel/rcu/srcutree.c +5 -1
@@ -905,14 +905,15 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
            for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
                    spin_lock_irqsave_rcu_node(snp, flags);
                    snp_seq = snp->srcu_have_cbs[idx];
-                   if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
+                   /*
+                    * s should be the biggest in the current slot. Hence only LE is
+                    * valid
+                    */
+                   BUG_ON(ULONG_CMP_GT(snp_seq, s));

Why not this? (Plus adjusting the comment above, of course.)

Quoted 2 lines
                      WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

A neat solution!

And what about the comment

                        /* s should be the biggest in the current slot. */
Quoted 5 lines

That way we don't need ULONG_CMP_GT(). Plus if we are confused about s being the biggest in the current slot, we get a splat and can debug further. We both might be quite sure that we are not confused, but that is exactly when we are most prone to making mistakes. ;-)

Quoted 8 lines
Quoted 3 lines

Same feeling :)

Quoted 1 line

I also ask that you run with this check for some time. After all, if

Forget to ask if the test
"tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration
10h --configs 18*SRCU-P"
satisfies your requirement?

I am thinking more in terms of adding the WARN_ON_ONCE(), letting it be in -next and possibly mainline for a couple of years, and then if there are no splats, start feeling more confident in the asserted relationship.

If there was a significant performance, scalability, energy-efficiency, or simplification benefit, I would feel justified in being more aggressive.

But I am not seeing a significant benefit.

Quoted 8 lines
Quoted 3 lines

Ah, and before I forget -again-, have you thought through the counter-wrap scenarios? Please keep in mind that on a 32-bit system, those counters can wrap quite quickly compared to typical uptimes.

I think 32-bit has not significant different since the statement
WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

where 's < snp_seq' is not associated with bits. Am I missing anything?

Your point is that long before there was any danger of wrap, this WARN_ON_ONCE() would trigger?

Quoted 8 lines
Quoted 1 line
Quoted 1 line

If so, the point is to avoid that WARN_ON_ONCE() from triggering.

Oops! I failed to figure out the scene you pointed out, where snp_x did not have call_srcu event for a very long time, then comes the wrap, and when s is close enough to snp_x's snp_seq, ULONG_CMP_LT(s, snp_seq) will raise false alarm.

Yes, this is a bug.

Very good, you did get there. ;-)

Quoted 6 lines
Quoted 1 line
Quoted 1 line

Or am I missing your point?

Sorry that I missed your point and confused you.

Can I remedy it using the root snp? Because the root snp's snp_seq advances each GP. And it is the biggest one among snp-s.

Yes, the root srcu_node structure is always there, but it will tend to be less warm in the cache, and thus will tend to be more expensive to access.

So, does keeping the current unconditional store solve the wrap problem?

Quoted 6 lines
Quoted 4 lines

For example, are you instead saying that the WARN_ON_ONCE() proposed in your patch will detect the problem either way? If so, what I am asking is whether your earlier analysis considered wrap without the addition of that WARN_ON_ONCE().

Have I answer all of your concerns?

Quite possibly.

However, to quote George Bernard Shaw, "The single biggest problem in communication is the illusion that it has taken place."

So I guess we will see! ;-)

Thanx, Paul

Quoted 48 lines

Thanks,

Pingfan

Quoted 44 lines

Thanx, Paul

Quoted 42 lines
Quoted 42 lines

Thanks,

Pingfan

Quoted 38 lines

Thanx, Paul

Quoted 36 lines
Quoted 36 lines

Thanks

Pingfan

Quoted 33 lines

Sure. I will try it immediately.

Quoted 7 lines

the assumption is incorrect, the resulting SRCU hangs off in various systems around the world will not be so much fun to debug.

Given that this is slowpath code, it is much better to take an extra compare and branch than to introduce even an extremely small risk of hanging SRCU.

Agree.

Thanks,

Pingfan

Quoted 16 lines

Thanx, Paul

Quoted 14 lines
kernel/rcu/srcutree.c +1 -5
+                   if (!srcu_invl_snp_seq(snp_seq) && (snp_seq == s)) {
                            if (snp == snp_leaf && snp_seq == s)
                                    snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
                            spin_unlock_irqrestore_rcu_node(snp, flags);
-                           if (snp == snp_leaf && snp_seq != s) {
-                                   srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
-                                   return;
-                           }
                            if (!do_norm)
                                    srcu_funnel_exp_start(ssp, snp, s);
                            return;
--

2.31.1

#16 lore raw

Re: [PATCHv2 2/2] srcu: Eliminate the case that snp_seq bigger than snap in srcu_funnel_gp_start()

Quoted 118 lines · Paul E. McKenney
Quoted 102 lines · Pingfan Liu
Quoted 101 lines · Paul E. McKenney
Quoted 100 lines · Paul E. McKenney
Quoted 87 lines · Pingfan Liu
Quoted 86 lines · Paul E. McKenney
Quoted 84 lines · Paul E. McKenney
Quoted 66 lines · Pingfan Liu
Quoted 65 lines · Pingfan Liu
Quoted 50 lines · Paul E. McKenney
Quoted 22 lines · Pingfan Liu

Since the srcu read lock is still held during srcu_funnel_gp_start(), the seq snap should be the largest number for the slot srcu_have_cbs[idx].

Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]

---
 include/linux/rcupdate.h |  1 +
 kernel/rcu/srcutree.c    | 11 ++++++-----
 2 files changed, 7 insertions(+), 5 deletions(-)
include/linux/rcupdate.h +0 -0
index 08605ce7379d..a09007236660 100644

include/linux/rcupdate.h line 35

Quoted 4 lines
include/linux/rcupdate.h +1 -0
@@ -32,6 +32,7 @@
 #include <linux/context_tracking_irq.h>
 
 #define ULONG_CMP_GE(a, b) (ULONG_MAX / 2 >= (a) - (b))
+#define ULONG_CMP_GT(a, b) (ULONG_MAX / 2 > (a) - (b))

Please see below...

Quoted 8 lines
include/linux/rcupdate.h +0 -0
 #define ULONG_CMP_LT(a, b) (ULONG_MAX / 2 < (a) - (b))
 #define ulong2long(a)              (*(long *)(&(a)))
 #define USHORT_CMP_GE(a, b)        (USHRT_MAX / 2 >= (unsigned short)((a) - (b)))
kernel/rcu/srcutree.c +0 -0
index 2fc0e775ade4..41902b823687 100644

kernel/rcu/srcutree.c static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp, · line 912

Quoted 9 lines
kernel/rcu/srcutree.c +5 -1
@@ -905,14 +905,15 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
            for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
                    spin_lock_irqsave_rcu_node(snp, flags);
                    snp_seq = snp->srcu_have_cbs[idx];
-                   if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
+                   /*
+                    * s should be the biggest in the current slot. Hence only LE is
+                    * valid
+                    */
+                   BUG_ON(ULONG_CMP_GT(snp_seq, s));

Why not this? (Plus adjusting the comment above, of course.)

Quoted 2 lines
                      WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

A neat solution!

And what about the comment

                        /* s should be the biggest in the current slot. */
Quoted 5 lines

That way we don't need ULONG_CMP_GT(). Plus if we are confused about s being the biggest in the current slot, we get a splat and can debug further. We both might be quite sure that we are not confused, but that is exactly when we are most prone to making mistakes. ;-)

Quoted 8 lines
Quoted 3 lines

Same feeling :)

Quoted 1 line

I also ask that you run with this check for some time. After all, if

Forget to ask if the test
"tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration
10h --configs 18*SRCU-P"
satisfies your requirement?

I am thinking more in terms of adding the WARN_ON_ONCE(), letting it be in -next and possibly mainline for a couple of years, and then if there are no splats, start feeling more confident in the asserted relationship.

If there was a significant performance, scalability, energy-efficiency, or simplification benefit, I would feel justified in being more aggressive.

But I am not seeing a significant benefit.

Quoted 8 lines
Quoted 3 lines

Ah, and before I forget -again-, have you thought through the counter-wrap scenarios? Please keep in mind that on a 32-bit system, those counters can wrap quite quickly compared to typical uptimes.

I think 32-bit has not significant different since the statement
WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

where 's < snp_seq' is not associated with bits. Am I missing anything?

Your point is that long before there was any danger of wrap, this WARN_ON_ONCE() would trigger?

Quoted 8 lines
Quoted 1 line
Quoted 1 line

If so, the point is to avoid that WARN_ON_ONCE() from triggering.

Oops! I failed to figure out the scene you pointed out, where snp_x did not have call_srcu event for a very long time, then comes the wrap, and when s is close enough to snp_x's snp_seq, ULONG_CMP_LT(s, snp_seq) will raise false alarm.

Yes, this is a bug.

Very good, you did get there. ;-)

Quoted 4 lines
Quoted 1 line
Quoted 1 line

Or am I missing your point?

Sorry that I missed your point and confused you.

Quoted 8 lines
Quoted 2 lines

Can I remedy it using the root snp? Because the root snp's snp_seq advances each GP. And it is the biggest one among snp-s.

Yes, the root srcu_node structure is always there, but it will tend to be less warm in the cache, and thus will tend to be more expensive to access.

So, does keeping the current unconditional store solve the wrap problem?

I think you are talking about the "snp->srcu_have_cbs[idx] = gpseq" in srcu_gp_end(). Yes, it can serve the purpose. And I can also apply wrap check on that logic later.

So if I keep that store, is PATCHv3 good enough?

Thanks,

Pingfan

Quoted 64 lines
Quoted 6 lines
Quoted 4 lines

For example, are you instead saying that the WARN_ON_ONCE() proposed in your patch will detect the problem either way? If so, what I am asking is whether your earlier analysis considered wrap without the addition of that WARN_ON_ONCE().

Have I answer all of your concerns?

Quite possibly.

However, to quote George Bernard Shaw, "The single biggest problem in communication is the illusion that it has taken place."

So I guess we will see! ;-)

Thanx, Paul

Quoted 48 lines

Thanks,

Pingfan

Quoted 44 lines

Thanx, Paul

Quoted 42 lines
Quoted 42 lines

Thanks,

Pingfan

Quoted 38 lines

Thanx, Paul

Quoted 36 lines
Quoted 36 lines

Thanks

Pingfan

Quoted 33 lines

Sure. I will try it immediately.

Quoted 7 lines

the assumption is incorrect, the resulting SRCU hangs off in various systems around the world will not be so much fun to debug.

Given that this is slowpath code, it is much better to take an extra compare and branch than to introduce even an extremely small risk of hanging SRCU.

Agree.

Thanks,

Pingfan

Quoted 16 lines

Thanx, Paul

Quoted 14 lines
kernel/rcu/srcutree.c +1 -5
+                   if (!srcu_invl_snp_seq(snp_seq) && (snp_seq == s)) {
                            if (snp == snp_leaf && snp_seq == s)
                                    snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
                            spin_unlock_irqrestore_rcu_node(snp, flags);
-                           if (snp == snp_leaf && snp_seq != s) {
-                                   srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
-                                   return;
-                           }
                            if (!do_norm)
                                    srcu_funnel_exp_start(ssp, snp, s);
                            return;
--

2.31.1

maintainer
#17 lore raw

Re: [PATCHv2 2/2] srcu: Eliminate the case that snp_seq bigger than snap in srcu_funnel_gp_start()

Quoted 125 lines · Pingfan Liu
Quoted 124 lines · Paul E. McKenney
Quoted 102 lines · Pingfan Liu
Quoted 101 lines · Paul E. McKenney
Quoted 100 lines · Paul E. McKenney
Quoted 87 lines · Pingfan Liu
Quoted 86 lines · Paul E. McKenney
Quoted 84 lines · Paul E. McKenney
Quoted 66 lines · Pingfan Liu
Quoted 65 lines · Pingfan Liu
Quoted 50 lines · Paul E. McKenney
Quoted 22 lines · Pingfan Liu

Since the srcu read lock is still held during srcu_funnel_gp_start(), the seq snap should be the largest number for the slot srcu_have_cbs[idx].

Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]

---
 include/linux/rcupdate.h |  1 +
 kernel/rcu/srcutree.c    | 11 ++++++-----
 2 files changed, 7 insertions(+), 5 deletions(-)
include/linux/rcupdate.h +0 -0
index 08605ce7379d..a09007236660 100644

include/linux/rcupdate.h line 35

Quoted 4 lines
include/linux/rcupdate.h +1 -0
@@ -32,6 +32,7 @@
 #include <linux/context_tracking_irq.h>
 
 #define ULONG_CMP_GE(a, b) (ULONG_MAX / 2 >= (a) - (b))
+#define ULONG_CMP_GT(a, b) (ULONG_MAX / 2 > (a) - (b))

Please see below...

Quoted 8 lines
include/linux/rcupdate.h +0 -0
 #define ULONG_CMP_LT(a, b) (ULONG_MAX / 2 < (a) - (b))
 #define ulong2long(a)              (*(long *)(&(a)))
 #define USHORT_CMP_GE(a, b)        (USHRT_MAX / 2 >= (unsigned short)((a) - (b)))
kernel/rcu/srcutree.c +0 -0
index 2fc0e775ade4..41902b823687 100644

kernel/rcu/srcutree.c static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp, · line 912

Quoted 9 lines
kernel/rcu/srcutree.c +5 -1
@@ -905,14 +905,15 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
            for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
                    spin_lock_irqsave_rcu_node(snp, flags);
                    snp_seq = snp->srcu_have_cbs[idx];
-                   if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
+                   /*
+                    * s should be the biggest in the current slot. Hence only LE is
+                    * valid
+                    */
+                   BUG_ON(ULONG_CMP_GT(snp_seq, s));

Why not this? (Plus adjusting the comment above, of course.)

Quoted 2 lines
                      WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

A neat solution!

And what about the comment

                        /* s should be the biggest in the current slot. */
Quoted 5 lines

That way we don't need ULONG_CMP_GT(). Plus if we are confused about s being the biggest in the current slot, we get a splat and can debug further. We both might be quite sure that we are not confused, but that is exactly when we are most prone to making mistakes. ;-)

Quoted 8 lines
Quoted 3 lines

Same feeling :)

Quoted 1 line

I also ask that you run with this check for some time. After all, if

Forget to ask if the test
"tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration
10h --configs 18*SRCU-P"
satisfies your requirement?

I am thinking more in terms of adding the WARN_ON_ONCE(), letting it be in -next and possibly mainline for a couple of years, and then if there are no splats, start feeling more confident in the asserted relationship.

If there was a significant performance, scalability, energy-efficiency, or simplification benefit, I would feel justified in being more aggressive.

But I am not seeing a significant benefit.

Quoted 8 lines
Quoted 3 lines

Ah, and before I forget -again-, have you thought through the counter-wrap scenarios? Please keep in mind that on a 32-bit system, those counters can wrap quite quickly compared to typical uptimes.

I think 32-bit has not significant different since the statement
WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

where 's < snp_seq' is not associated with bits. Am I missing anything?

Your point is that long before there was any danger of wrap, this WARN_ON_ONCE() would trigger?

Quoted 8 lines
Quoted 1 line
Quoted 1 line

If so, the point is to avoid that WARN_ON_ONCE() from triggering.

Oops! I failed to figure out the scene you pointed out, where snp_x did not have call_srcu event for a very long time, then comes the wrap, and when s is close enough to snp_x's snp_seq, ULONG_CMP_LT(s, snp_seq) will raise false alarm.

Yes, this is a bug.

Very good, you did get there. ;-)

Quoted 6 lines
Quoted 1 line
Quoted 1 line

Or am I missing your point?

Sorry that I missed your point and confused you.

Can I remedy it using the root snp? Because the root snp's snp_seq advances each GP. And it is the biggest one among snp-s.

Yes, the root srcu_node structure is always there, but it will tend to be less warm in the cache, and thus will tend to be more expensive to access.

Quoted 8 lines
Quoted 2 lines

So, does keeping the current unconditional store solve the wrap problem?

I think you are talking about the "snp->srcu_have_cbs[idx] = gpseq" in srcu_gp_end(). Yes, it can serve the purpose. And I can also apply wrap check on that logic later.

So if I keep that store, is PATCHv3 good enough?

It certainly needs an improved commit log. For example, removing that call to srcu_schedule_cbs_sdp() can increase apparent SRCU grace-period latency. I could easily believe that this is not a problem, but the commit log at the very least needs to mention this possibility.

I also need to see at least one ack, given that this patch is removing software-engineering redundancy.

Thanx, Paul

Quoted 68 lines

Thanks,

Pingfan

Quoted 64 lines
Quoted 6 lines
Quoted 4 lines

For example, are you instead saying that the WARN_ON_ONCE() proposed in your patch will detect the problem either way? If so, what I am asking is whether your earlier analysis considered wrap without the addition of that WARN_ON_ONCE().

Have I answer all of your concerns?

Quite possibly.

However, to quote George Bernard Shaw, "The single biggest problem in communication is the illusion that it has taken place."

So I guess we will see! ;-)

Thanx, Paul

Quoted 48 lines

Thanks,

Pingfan

Quoted 44 lines

Thanx, Paul

Quoted 42 lines
Quoted 42 lines

Thanks,

Pingfan

Quoted 38 lines

Thanx, Paul

Quoted 36 lines
Quoted 36 lines

Thanks

Pingfan

Quoted 33 lines

Sure. I will try it immediately.

Quoted 7 lines

the assumption is incorrect, the resulting SRCU hangs off in various systems around the world will not be so much fun to debug.

Given that this is slowpath code, it is much better to take an extra compare and branch than to introduce even an extremely small risk of hanging SRCU.

Agree.

Thanks,

Pingfan

Quoted 16 lines

Thanx, Paul

Quoted 14 lines
kernel/rcu/srcutree.c +1 -5
+                   if (!srcu_invl_snp_seq(snp_seq) && (snp_seq == s)) {
                            if (snp == snp_leaf && snp_seq == s)
                                    snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
                            spin_unlock_irqrestore_rcu_node(snp, flags);
-                           if (snp == snp_leaf && snp_seq != s) {
-                                   srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
-                                   return;
-                           }
                            if (!do_norm)
                                    srcu_funnel_exp_start(ssp, snp, s);
                            return;
--

2.31.1

#18 lore raw

Re: [PATCHv2 2/2] srcu: Eliminate the case that snp_seq bigger than snap in srcu_funnel_gp_start()

In reply to Paul E. McKenney
Quoted 72 lines · Paul E. McKenney
Quoted 71 lines · Pingfan Liu
Quoted 69 lines · Pingfan Liu
Quoted 50 lines · Paul E. McKenney
Quoted 22 lines · Pingfan Liu

Since the srcu read lock is still held during srcu_funnel_gp_start(), the seq snap should be the largest number for the slot srcu_have_cbs[idx].

Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]

---
 include/linux/rcupdate.h |  1 +
 kernel/rcu/srcutree.c    | 11 ++++++-----
 2 files changed, 7 insertions(+), 5 deletions(-)
include/linux/rcupdate.h +0 -0
index 08605ce7379d..a09007236660 100644

include/linux/rcupdate.h line 35

Quoted 4 lines
include/linux/rcupdate.h +1 -0
@@ -32,6 +32,7 @@
 #include <linux/context_tracking_irq.h>
 
 #define ULONG_CMP_GE(a, b) (ULONG_MAX / 2 >= (a) - (b))
+#define ULONG_CMP_GT(a, b) (ULONG_MAX / 2 > (a) - (b))

Please see below...

Quoted 8 lines
include/linux/rcupdate.h +0 -0
 #define ULONG_CMP_LT(a, b) (ULONG_MAX / 2 < (a) - (b))
 #define ulong2long(a)              (*(long *)(&(a)))
 #define USHORT_CMP_GE(a, b)        (USHRT_MAX / 2 >= (unsigned short)((a) - (b)))
kernel/rcu/srcutree.c +0 -0
index 2fc0e775ade4..41902b823687 100644

kernel/rcu/srcutree.c static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp, · line 912

Quoted 9 lines
kernel/rcu/srcutree.c +5 -1
@@ -905,14 +905,15 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
            for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
                    spin_lock_irqsave_rcu_node(snp, flags);
                    snp_seq = snp->srcu_have_cbs[idx];
-                   if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
+                   /*
+                    * s should be the biggest in the current slot. Hence only LE is
+                    * valid
+                    */
+                   BUG_ON(ULONG_CMP_GT(snp_seq, s));

Why not this? (Plus adjusting the comment above, of course.)

Quoted 2 lines
                      WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));

A neat solution!

And what about the comment

                        /* s should be the biggest in the current slot. */
Quoted 5 lines

That way we don't need ULONG_CMP_GT(). Plus if we are confused about s being the biggest in the current slot, we get a splat and can debug further. We both might be quite sure that we are not confused, but that is exactly when we are most prone to making mistakes. ;-)

Same feeling :)

Quoted 1 line

I also ask that you run with this check for some time. After all, if

Quoted 9 lines
Quoted 4 lines

Forget to ask if the test
"tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration
10h --configs 18*SRCU-P"
satisfies your requirement?

I am thinking more in terms of adding the WARN_ON_ONCE(), letting it be in -next and possibly mainline for a couple of years, and then if there are no splats, start feeling more confident in the asserted relationship.

Sounds reasonable.

Quoted 5 lines

If there was a significant performance, scalability, energy-efficiency, or simplification benefit, I would feel justified in being more aggressive.

But I am not seeing a significant benefit.

No, not a significant benefit.

I will send out V3 for this patch

Thanks,

Pingfan

Quoted 38 lines

Thanx, Paul

Quoted 36 lines

Thanks

Pingfan

Quoted 33 lines

Sure. I will try it immediately.

Quoted 7 lines

the assumption is incorrect, the resulting SRCU hangs off in various systems around the world will not be so much fun to debug.

Given that this is slowpath code, it is much better to take an extra compare and branch than to introduce even an extremely small risk of hanging SRCU.

Agree.

Thanks,

Pingfan

Quoted 16 lines

Thanx, Paul

Quoted 14 lines
kernel/rcu/srcutree.c +1 -5
+                   if (!srcu_invl_snp_seq(snp_seq) && (snp_seq == s)) {
                            if (snp == snp_leaf && snp_seq == s)
                                    snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
                            spin_unlock_irqrestore_rcu_node(snp, flags);
-                           if (snp == snp_leaf && snp_seq != s) {
-                                   srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
-                                   return;
-                           }
                            if (!do_norm)
                                    srcu_funnel_exp_start(ssp, snp, s);
                            return;
--

2.31.1

#19 lore raw

PATCH v3 1/1 srcu: Eliminate the case that snp_seq bigger than snap in srcu_funnel_gp_start()

In reply to Pingfan Liu

Since the srcu read lock is still held during srcu_funnel_gp_start(), the seq snap should be the largest number for the slot srcu_have_cbs[idx].

Test info
Running the following test without any failure
tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration 10h --configs 18*SRCU-P
And
cd res && find . -name console.log -exec grep WARNING {} \;
prints nothing

Cc Lai Jiangshan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Steven Rostedt <[email protected]>
Cc Mathieu Desnoyers <[email protected]>

To: [email protected]

---
 kernel/rcu/srcutree.c | 8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)
kernel/rcu/srcutree.c +3 -5
index 7e450bb4ddcf..687f243dd2e2 100644
@@ -905,14 +905,12 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
 		for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
 			spin_lock_irqsave_rcu_node(snp, flags);
 			snp_seq = snp->srcu_have_cbs[idx];
-			if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
+			/* s should be the biggest in the current slot. */
+			WARN_ON_ONCE(ULONG_CMP_LT(s, snp_seq));
+			if (!srcu_invl_snp_seq(snp_seq) && (snp_seq == s)) {
 				if (snp == snp_leaf && snp_seq == s)
 					snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
 				spin_unlock_irqrestore_rcu_node(snp, flags);
-				if (snp == snp_leaf && snp_seq != s) {
-					srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
-					return;
-				}
 				if (!do_norm)
 					srcu_funnel_exp_start(ssp, snp, s);
 				return;
Signature · 1 line
2.31.1