Skip to contentSkip to conversation

rcu: Align stall warning 'idle=' output with documentation

PATCH In review Donglin Peng 2 participants 5 messages Jan 5 – Sep 16 RCU

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 rcu: Align stall warning 'idle=' output with documentation

From: Donglin Peng <[email protected]>

The RCU stall warning message includes an "idle=" field to indicate the dyntick-idle state of a CPU. According to Documentation/RCU/stallwarn.rst, the hexadecimal number before the first '/' represents the low-order 16 bits of the dynticks counter. An even value denotes that the CPU is in dyntick-idle mode, while an odd value indicates otherwise.

This was valid until commit 171476775d32 ("context_tracking: Convert state to atomic_t"), which merged the context-tracking state and dynticks counter into a
single atomic variable. In the new layout, the dynticks counter occupies the higher bits starting from CT_RCU_WATCHING_START.

However, the current stall warning code prints the value from ct_rcu_watching_cpu(), which returns atomic_read(&ct->state) & CT_RCU_WATCHING_MASK. This masks out (clears) the lower state bits, resulting in a value that is always even. This obscures the CPU's true idle state and makes the output inconsistent with the documentation.

To restore consistency between the code's output and the documentation, shift the atomic value right by CT_RCU_WATCHING_START bits before printing. This extracts and displays only the relevant dynticks counter portion, allowing the parity (even/odd) to correctly reflect the CPU's dyntick-idle state.

Fixes 171476775d32 ("context_tracking: Convert state to atomic_t")

---
v2:
- Use CT_RCU_WATCHING_START to replace ilog2(CT_RCU_WATCHING) to clean up the code, thanks to Zqiang.
- Clarify the commit message.

---
 kernel/rcu/tree_stall.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
kernel/rcu/tree_stall.h +1 -1
index b67532cb8770..fd9d7fd6e17c 100644
@@ -555,7 +555,7 @@ static void print_cpu_stall_info(int cpu)
 			rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
 				"!."[!delta],
 	       ticks_value, ticks_title,
-	       ct_rcu_watching_cpu(cpu) & 0xffff,
+	       (ct_rcu_watching_cpu(cpu) >> CT_RCU_WATCHING_START) & 0xffff,
 	       ct_nesting_cpu(cpu), ct_nmi_nesting_cpu(cpu),
 	       rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
 	       data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
Signature · 1 line
2.34.1
#1 lore raw
Quoted 42 lines · Donglin Peng

From: Donglin Peng <[email protected]>

The RCU stall warning message includes an "idle=" field to indicate the dyntick-idle state of a CPU. According to Documentation/RCU/stallwarn.rst, the hexadecimal number before the first '/' represents the low-order 16 bits of the dynticks counter. An even value denotes that the CPU is in dyntick-idle mode, while an odd value indicates otherwise.

This was valid until commit 171476775d32 ("context_tracking: Convert state to atomic_t"), which merged the context-tracking state and dynticks counter into a
single atomic variable. In the new layout, the dynticks counter occupies the higher bits starting from CT_RCU_WATCHING_START.

However, the current stall warning code prints the value from ct_rcu_watching_cpu(), which returns atomic_read(&ct->state) & CT_RCU_WATCHING_MASK. This masks out (clears) the lower state bits, resulting in a value that is always even. This obscures the CPU's true idle state and makes the output inconsistent with the documentation.

To restore consistency between the code's output and the documentation, shift the atomic value right by CT_RCU_WATCHING_START bits before printing. This extracts and displays only the relevant dynticks counter portion, allowing the parity (even/odd) to correctly reflect the CPU's dyntick-idle state.

Fixes 171476775d32 ("context_tracking: Convert state to atomic_t")

---
v2:
- Use CT_RCU_WATCHING_START to replace ilog2(CT_RCU_WATCHING) to clean up the code, thanks to Zqiang.
- Clarify the commit message.

---
  kernel/rcu/tree_stall.h | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
kernel/rcu/tree_stall.h +0 -0
index b67532cb8770..fd9d7fd6e17c 100644

kernel/rcu/tree_stall.h static void print_cpu_stall_info(int cpu) · line 561

Quoted 8 lines
kernel/rcu/tree_stall.h +1 -1
@@ -555,7 +555,7 @@ static void print_cpu_stall_info(int cpu)
  			rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
  				"!."[!delta],
  	       ticks_value, ticks_title,
-	       ct_rcu_watching_cpu(cpu) & 0xffff,
+	       (ct_rcu_watching_cpu(cpu) >> CT_RCU_WATCHING_START) & 0xffff,
  	       ct_nesting_cpu(cpu), ct_nmi_nesting_cpu(cpu),
  	       rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
  	       data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,

Gentle ping on this patch. Thanks!

maintainer
#2 lore raw
Quoted 38 lines · Peng Donglin
Quoted 37 lines · Donglin Peng

From: Donglin Peng <[email protected]>

The RCU stall warning message includes an "idle=" field to indicate the dyntick-idle state of a CPU. According to Documentation/RCU/stallwarn.rst, the hexadecimal number before the first '/' represents the low-order 16 bits of the dynticks counter. An even value denotes that the CPU is in dyntick-idle mode, while an odd value indicates otherwise.

This was valid until commit 171476775d32 ("context_tracking: Convert state to atomic_t"), which merged the context-tracking state and dynticks counter into a
single atomic variable. In the new layout, the dynticks counter occupies the higher bits starting from CT_RCU_WATCHING_START.

However, the current stall warning code prints the value from ct_rcu_watching_cpu(), which returns atomic_read(&ct->state) & CT_RCU_WATCHING_MASK. This masks out (clears) the lower state bits, resulting in a value that is always even. This obscures the CPU's true idle state and makes the output inconsistent with the documentation.

To restore consistency between the code's output and the documentation, shift the atomic value right by CT_RCU_WATCHING_START bits before printing. This extracts and displays only the relevant dynticks counter portion, allowing the parity (even/odd) to correctly reflect the CPU's dyntick-idle state.

Fixes 171476775d32 ("context_tracking: Convert state to atomic_t")

---
v2:
- Use CT_RCU_WATCHING_START to replace ilog2(CT_RCU_WATCHING) to clean up the code, thanks to Zqiang.
- Clarify the commit message.

---
  kernel/rcu/tree_stall.h | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 15 lines

kernel/rcu/tree_stall.h static void print_cpu_stall_info(int cpu) · line 561

Quoted 13 lines
kernel/rcu/tree_stall.h +1 -1
index b67532cb8770..fd9d7fd6e17c 100644
@@ -555,7 +555,7 @@ static void print_cpu_stall_info(int cpu)
  			rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
  				"!."[!delta],
  	       ticks_value, ticks_title,
-	       ct_rcu_watching_cpu(cpu) & 0xffff,
+	       (ct_rcu_watching_cpu(cpu) >> CT_RCU_WATCHING_START) & 0xffff,
  	       ct_nesting_cpu(cpu), ct_nmi_nesting_cpu(cpu),
  	       rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
  	       data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,

Gentle ping on this patch. Thanks!

The advantage of the current code is that the CT_STATE_KERNEL, CT_STATE_IDLE, CT_STATE_USER, and CT_STATE_GUEST information is visible in the lower two bits of that hex number.

Now, if you wanted to print that information symbolically and then shift down the counter, that might well be a valuable improvement. Given corresponding updates to stallwarn.rst, of course.

Or am I missing something subtle here?

Thanx, Paul

#3 lore raw
Quoted 56 lines · Paul E. McKenney
Quoted 38 lines · Peng Donglin
Quoted 37 lines · Donglin Peng

From: Donglin Peng <[email protected]>

The RCU stall warning message includes an "idle=" field to indicate the dyntick-idle state of a CPU. According to Documentation/RCU/stallwarn.rst, the hexadecimal number before the first '/' represents the low-order 16 bits of the dynticks counter. An even value denotes that the CPU is in dyntick-idle mode, while an odd value indicates otherwise.

This was valid until commit 171476775d32 ("context_tracking: Convert state to atomic_t"), which merged the context-tracking state and dynticks counter into a
single atomic variable. In the new layout, the dynticks counter occupies the higher bits starting from CT_RCU_WATCHING_START.

However, the current stall warning code prints the value from ct_rcu_watching_cpu(), which returns atomic_read(&ct->state) & CT_RCU_WATCHING_MASK. This masks out (clears) the lower state bits, resulting in a value that is always even. This obscures the CPU's true idle state and makes the output inconsistent with the documentation.

To restore consistency between the code's output and the documentation, shift the atomic value right by CT_RCU_WATCHING_START bits before printing. This extracts and displays only the relevant dynticks counter portion, allowing the parity (even/odd) to correctly reflect the CPU's dyntick-idle state.

Fixes 171476775d32 ("context_tracking: Convert state to atomic_t")

---
v2:
- Use CT_RCU_WATCHING_START to replace ilog2(CT_RCU_WATCHING) to clean up the code, thanks to Zqiang.
- Clarify the commit message.

---
  kernel/rcu/tree_stall.h | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 15 lines

kernel/rcu/tree_stall.h static void print_cpu_stall_info(int cpu) · line 561

Quoted 13 lines
kernel/rcu/tree_stall.h +1 -1
index b67532cb8770..fd9d7fd6e17c 100644
@@ -555,7 +555,7 @@ static void print_cpu_stall_info(int cpu)
                    rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
                            "!."[!delta],
           ticks_value, ticks_title,
-          ct_rcu_watching_cpu(cpu) & 0xffff,
+          (ct_rcu_watching_cpu(cpu) >> CT_RCU_WATCHING_START) & 0xffff,
           ct_nesting_cpu(cpu), ct_nmi_nesting_cpu(cpu),
           rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
           data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,

Gentle ping on this patch. Thanks!

Quoted 7 lines

The advantage of the current code is that the CT_STATE_KERNEL, CT_STATE_IDLE, CT_STATE_USER, and CT_STATE_GUEST information is visible in the lower two bits of that hex number.

Now, if you wanted to print that information symbolically and then shift down the counter, that might well be a valuable improvement. Given corresponding updates to stallwarn.rst, of course.

Thanks, that makes sense. I'll print the state symbolically, shift down the counter, update stallwarn.rst accordingly, and send a v3.

Quoted 4 lines

Or am I missing something subtle here?

Thanx, Paul

maintainer
#4 lore raw
Quoted 61 lines · Donglin Peng
Quoted 60 lines · Paul E. McKenney
Quoted 38 lines · Peng Donglin
Quoted 37 lines · Donglin Peng

From: Donglin Peng <[email protected]>

The RCU stall warning message includes an "idle=" field to indicate the dyntick-idle state of a CPU. According to Documentation/RCU/stallwarn.rst, the hexadecimal number before the first '/' represents the low-order 16 bits of the dynticks counter. An even value denotes that the CPU is in dyntick-idle mode, while an odd value indicates otherwise.

This was valid until commit 171476775d32 ("context_tracking: Convert state to atomic_t"), which merged the context-tracking state and dynticks counter into a
single atomic variable. In the new layout, the dynticks counter occupies the higher bits starting from CT_RCU_WATCHING_START.

However, the current stall warning code prints the value from ct_rcu_watching_cpu(), which returns atomic_read(&ct->state) & CT_RCU_WATCHING_MASK. This masks out (clears) the lower state bits, resulting in a value that is always even. This obscures the CPU's true idle state and makes the output inconsistent with the documentation.

To restore consistency between the code's output and the documentation, shift the atomic value right by CT_RCU_WATCHING_START bits before printing. This extracts and displays only the relevant dynticks counter portion, allowing the parity (even/odd) to correctly reflect the CPU's dyntick-idle state.

Fixes 171476775d32 ("context_tracking: Convert state to atomic_t")

---
v2:
- Use CT_RCU_WATCHING_START to replace ilog2(CT_RCU_WATCHING) to clean up the code, thanks to Zqiang.
- Clarify the commit message.

---
  kernel/rcu/tree_stall.h | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 15 lines

kernel/rcu/tree_stall.h static void print_cpu_stall_info(int cpu) · line 561

Quoted 13 lines
kernel/rcu/tree_stall.h +1 -1
index b67532cb8770..fd9d7fd6e17c 100644
@@ -555,7 +555,7 @@ static void print_cpu_stall_info(int cpu)
                    rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
                            "!."[!delta],
           ticks_value, ticks_title,
-          ct_rcu_watching_cpu(cpu) & 0xffff,
+          (ct_rcu_watching_cpu(cpu) >> CT_RCU_WATCHING_START) & 0xffff,
           ct_nesting_cpu(cpu), ct_nmi_nesting_cpu(cpu),
           rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
           data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,

Gentle ping on this patch. Thanks!

The advantage of the current code is that the CT_STATE_KERNEL, CT_STATE_IDLE, CT_STATE_USER, and CT_STATE_GUEST information is visible in the lower two bits of that hex number.

Quoted 6 lines
Quoted 3 lines

Now, if you wanted to print that information symbolically and then shift down the counter, that might well be a valuable improvement. Given corresponding updates to stallwarn.rst, of course.

Thanks, that makes sense. I'll print the state symbolically, shift down the counter, update stallwarn.rst accordingly, and send a v3.

Sounds good, and looking forward to it!

Thanx, Paul

Quoted 3 lines
Quoted 3 lines

Or am I missing something subtle here?

Thanx, Paul