Skip to contentSkip to conversation

2.6.17-mm5

PATCH In review Martin J. Bligh 6 participants 15 messages Jul 2, 2006 – Jul 6, 2006 Scheduler

lore
Share & tools
lore
#0 lore raw

Panic on NUMA-Q (mm4 was fine). Presumably some new scheduler patch

divide error: 0000 [#1]
8K_STACKS SMP last sysfs file: Modules linked in:

CPU:    1
EIP:    0060:[<c0112b6e>]    Not tainted VLI
EFLAGS: 00010046   (2.6.17-mm5-autokern1 #1) EIP is at find_busiest_group+0x1a3/0x47c
eax: 00000000   ebx: 00000007   ecx: 00000000   edx: 00000000
esi: 00000000   edi: e75ff264   ebp: e7405ec8   esp: e7405e58
ds: 007b   es: 007b   ss: 0068
Process swapper (pid: 0, ti=e7404000 task=c13f8560 task.ti=e7404000)
Stack: e75ff264 00000010 c0119020 00000000 00000000 00000000 00000000 00000000 
      ffffffff 00000000 00000000 00000001 00000001 00000001 00000080 00000000 
      00000000 00000200 00000020 00000080 00000000 00000000 e75ff260 c1364960 
Call Trace:
[<c0119020>] vprintk+0x5f/0x213
[<c0112efb>] load_balance+0x54/0x1d6
[<c011332d>] rebalance_tick+0xc5/0xe3
[<c01137a3>] scheduler_tick+0x2cb/0x2d3
[<c01215b4>] update_process_times+0x51/0x5d
[<c010c224>] smp_apic_timer_interrupt+0x5a/0x61
[<c0102d5b>] apic_timer_interrupt+0x1f/0x24
[<c01006c0>] default_idle+0x0/0x59
[<c01006f1>] default_idle+0x31/0x59
[<c0100791>] cpu_idle+0x64/0x79
Code: 00 5b 83 f8 1f 89 c6 5f 0f 8e 63 ff ff ff 8b 45 e0 8b 55 e8 01 45 dc 8b 4a 08 89 c2 01 4d d4 c1 e2 07 89 d0 31 d2 89 ce c1 ee 07 <f7> f1 83 7d 9c 00 89 45 e0 74 17 89 45 d8 8b 55 e8 8b 4d a4 8b EIP: [<c0112b6e>] find_busiest_group+0x1a3/0x47c SS:ESP 0068:e7405e58
#1 lore raw
Quoted 14 lines · On Sun, 02 Jul 2006 16:27:55 -0700 "Martin J. Bligh" <[email protected]> wrote:

Panic on NUMA-Q (mm4 was fine). Presumably some new scheduler patch

divide error: 0000 [#1]
8K_STACKS SMP
last sysfs file:

Modules linked in:
CPU:    1
EIP:    0060:[<c0112b6e>]    Not tainted VLI
EFLAGS: 00010046   (2.6.17-mm5-autokern1 #1) 

EIP is at find_busiest_group+0x1a3/0x47c
eax: 00000000 ebx: 00000007 ecx: 00000000 edx: 00000000
esi: 00000000 edi: e75ff264 ebp: e7405ec8 esp: e7405e58
ds: 007b es: 007b ss: 0068
Process swapper (pid: 0, ti=e7404000 task=c13f8560 task.ti=e7404000)

Quoted 16 lines
Stack: e75ff264 00000010 c0119020 00000000 00000000 00000000 00000000 00000000 
       ffffffff 00000000 00000000 00000001 00000001 00000001 00000080 00000000 
       00000000 00000200 00000020 00000080 00000000 00000000 e75ff260 c1364960 
Call Trace:
 [<c0119020>] vprintk+0x5f/0x213
 [<c0112efb>] load_balance+0x54/0x1d6
 [<c011332d>] rebalance_tick+0xc5/0xe3
 [<c01137a3>] scheduler_tick+0x2cb/0x2d3
 [<c01215b4>] update_process_times+0x51/0x5d
 [<c010c224>] smp_apic_timer_interrupt+0x5a/0x61
 [<c0102d5b>] apic_timer_interrupt+0x1f/0x24
 [<c01006c0>] default_idle+0x0/0x59
 [<c01006f1>] default_idle+0x31/0x59
 [<c0100791>] cpu_idle+0x64/0x79
Code: 00 5b 83 f8 1f 89 c6 5f 0f 8e 63 ff ff ff 8b 45 e0 8b 55 e8 01 45 dc 8b 4a 08 89 c2 01 4d d4 c1 e2 07 89 d0 31 d2 89 ce c1 ee 07 <f7> f1 83 7d 9c 00 89 45 e0 74 17 89 45 d8 8b 55 e8 8b 4d a4 8b 
EIP: [<c0112b6e>] find_busiest_group+0x1a3/0x47c SS:ESP 0068:e7405e58

Yes, Andy's reporting that too. I asked him to identify the file-n-line and he ran away on me.

#2 lore raw

PATCH 1/1 sched: fix macro -> inline function conversion bug

Quoted 34 lines · Andrew Morton
Quoted 14 lines · On Sun, 02 Jul 2006 16:27:55 -0700 "Martin J. Bligh" <[email protected]> wrote:

Panic on NUMA-Q (mm4 was fine). Presumably some new scheduler patch

divide error: 0000 [#1]
8K_STACKS SMP
last sysfs file:

Modules linked in:
CPU:    1
EIP:    0060:[<c0112b6e>]    Not tainted VLI
EFLAGS: 00010046   (2.6.17-mm5-autokern1 #1) 

EIP is at find_busiest_group+0x1a3/0x47c
eax: 00000000 ebx: 00000007 ecx: 00000000 edx: 00000000
esi: 00000000 edi: e75ff264 ebp: e7405ec8 esp: e7405e58
ds: 007b es: 007b ss: 0068
Process swapper (pid: 0, ti=e7404000 task=c13f8560 task.ti=e7404000)

Quoted 16 lines
Stack: e75ff264 00000010 c0119020 00000000 00000000 00000000 00000000 00000000 
       ffffffff 00000000 00000000 00000001 00000001 00000001 00000080 00000000 
       00000000 00000200 00000020 00000080 00000000 00000000 e75ff260 c1364960 
Call Trace:
 [<c0119020>] vprintk+0x5f/0x213
 [<c0112efb>] load_balance+0x54/0x1d6
 [<c011332d>] rebalance_tick+0xc5/0xe3
 [<c01137a3>] scheduler_tick+0x2cb/0x2d3
 [<c01215b4>] update_process_times+0x51/0x5d
 [<c010c224>] smp_apic_timer_interrupt+0x5a/0x61
 [<c0102d5b>] apic_timer_interrupt+0x1f/0x24
 [<c01006c0>] default_idle+0x0/0x59
 [<c01006f1>] default_idle+0x31/0x59
 [<c0100791>] cpu_idle+0x64/0x79
Code: 00 5b 83 f8 1f 89 c6 5f 0f 8e 63 ff ff ff 8b 45 e0 8b 55 e8 01 45 dc 8b 4a 08 89 c2 01 4d d4 c1 e2 07 89 d0 31 d2 89 ce c1 ee 07 <f7> f1 83 7d 9c 00 89 45 e0 74 17 89 45 d8 8b 55 e8 8b 4d a4 8b 
EIP: [<c0112b6e>] find_busiest_group+0x1a3/0x47c SS:ESP 0068:e7405e58
Quoted 2 lines

Yes, Andy's reporting that too. I asked him to identify the file-n-line and he ran away on me.

i checked the scheduler queue and nothing jumped out at me, except the cleanup bug fixed by the patch below. (which should be harmless in this particular case - nr_running should never be smaller than 0 or larger than ~4 billion. A fix is warranted nevertheless.)

Ingo

-------------->
Subject: sched: fix macro -> inline function conversion bug
From: Ingo Molnar <[email protected]>

nr_running is long, not int.

---
 kernel/sched.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
kernel/sched.c +1 -1
Index: linux/kernel/sched.c
===================================================================
@@ -2480,7 +2480,7 @@ find_busiest_queue(struct sched_group *g
  */
 #define MAX_PINNED_INTERVAL	512
 
-static inline int minus_1_or_zero(int n)
+static inline unsigned long minus_1_or_zero(unsigned long n)
 {
 	return n > 0 ? n - 1 : 0;
 }
#3 lore raw

Re: [patch] sched: fix macro -> inline function conversion bug

Quoted 37 lines · On Mon, 3 Jul 2006 07:25:39 +0200 Ingo Molnar <[email protected]> wrote:
Quoted 34 lines · Andrew Morton
Quoted 30 lines · On Sun, 02 Jul 2006 16:27:55 -0700 "Martin J. Bligh" <[email protected]> wrote:

Panic on NUMA-Q (mm4 was fine). Presumably some new scheduler patch

divide error: 0000 [#1]
8K_STACKS SMP
last sysfs file:

Modules linked in:
CPU:    1
EIP:    0060:[<c0112b6e>]    Not tainted VLI
EFLAGS: 00010046   (2.6.17-mm5-autokern1 #1) 

EIP is at find_busiest_group+0x1a3/0x47c
eax: 00000000 ebx: 00000007 ecx: 00000000 edx: 00000000
esi: 00000000 edi: e75ff264 ebp: e7405ec8 esp: e7405e58
ds: 007b es: 007b ss: 0068
Process swapper (pid: 0, ti=e7404000 task=c13f8560 task.ti=e7404000)

Stack: e75ff264 00000010 c0119020 00000000 00000000 00000000 00000000 00000000 
       ffffffff 00000000 00000000 00000001 00000001 00000001 00000080 00000000 
       00000000 00000200 00000020 00000080 00000000 00000000 e75ff260 c1364960 
Call Trace:
 [<c0119020>] vprintk+0x5f/0x213
 [<c0112efb>] load_balance+0x54/0x1d6
 [<c011332d>] rebalance_tick+0xc5/0xe3
 [<c01137a3>] scheduler_tick+0x2cb/0x2d3
 [<c01215b4>] update_process_times+0x51/0x5d
 [<c010c224>] smp_apic_timer_interrupt+0x5a/0x61
 [<c0102d5b>] apic_timer_interrupt+0x1f/0x24
 [<c01006c0>] default_idle+0x0/0x59
 [<c01006f1>] default_idle+0x31/0x59
 [<c0100791>] cpu_idle+0x64/0x79
Code: 00 5b 83 f8 1f 89 c6 5f 0f 8e 63 ff ff ff 8b 45 e0 8b 55 e8 01 45 dc 8b 4a 08 89 c2 01 4d d4 c1 e2 07 89 d0 31 d2 89 ce c1 ee 07 <f7> f1 83 7d 9c 00 89 45 e0 74 17 89 45 d8 8b 55 e8 8b 4d a4 8b 
EIP: [<c0112b6e>] find_busiest_group+0x1a3/0x47c SS:ESP 0068:e7405e58
Quoted 7 lines
Quoted 2 lines

Yes, Andy's reporting that too. I asked him to identify the file-n-line and he ran away on me.

i checked the scheduler queue and nothing jumped out at me, except the cleanup bug fixed by the patch below. (which should be harmless in this particular case - nr_running should never be smaller than 0 or larger than ~4 billion. A fix is warranted nevertheless.)

Did you work out which divide is getting the div-by-zero? I started at it a bit and wasn't sure - am getting wildly different code generation over here.

#4 lore raw

Re: [patch] sched: fix macro -> inline function conversion bug

Quoted 34 lines · Andrew Morton
Quoted 31 lines · On Mon, 3 Jul 2006 07:25:39 +0200 Ingo Molnar <[email protected]> wrote:
Quoted 29 lines · Andrew Morton
Quoted 26 lines · On Sun, 02 Jul 2006 16:27:55 -0700 "Martin J. Bligh" <[email protected]> wrote:

Panic on NUMA-Q (mm4 was fine). Presumably some new scheduler patch

divide error: 0000 [#1]
8K_STACKS SMP last sysfs file: Modules linked in:

CPU:    1
EIP:    0060:[<c0112b6e>]    Not tainted VLI
EFLAGS: 00010046   (2.6.17-mm5-autokern1 #1) EIP is at find_busiest_group+0x1a3/0x47c
eax: 00000000   ebx: 00000007   ecx: 00000000   edx: 00000000
esi: 00000000   edi: e75ff264   ebp: e7405ec8   esp: e7405e58
ds: 007b   es: 007b   ss: 0068
Process swapper (pid: 0, ti=e7404000 task=c13f8560 task.ti=e7404000)
Stack: e75ff264 00000010 c0119020 00000000 00000000 00000000 00000000 00000000 
       ffffffff 00000000 00000000 00000001 00000001 00000001 00000080 00000000 
       00000000 00000200 00000020 00000080 00000000 00000000 e75ff260 c1364960 
Call Trace:
 [<c0119020>] vprintk+0x5f/0x213
 [<c0112efb>] load_balance+0x54/0x1d6
 [<c011332d>] rebalance_tick+0xc5/0xe3
 [<c01137a3>] scheduler_tick+0x2cb/0x2d3
 [<c01215b4>] update_process_times+0x51/0x5d
 [<c010c224>] smp_apic_timer_interrupt+0x5a/0x61
 [<c0102d5b>] apic_timer_interrupt+0x1f/0x24
 [<c01006c0>] default_idle+0x0/0x59
 [<c01006f1>] default_idle+0x31/0x59
 [<c0100791>] cpu_idle+0x64/0x79
Code: 00 5b 83 f8 1f 89 c6 5f 0f 8e 63 ff ff ff 8b 45 e0 8b 55 e8 01 45 dc 8b 4a 08 89 c2 01 4d d4 c1 e2 07 89 d0 31 d2 89 ce c1 ee 07 <f7> f1 83 7d 9c 00 89 45 e0 74 17 89 45 d8 8b 55 e8 8b 4d a4 8b EIP: [<c0112b6e>] find_busiest_group+0x1a3/0x47c SS:ESP 0068:e7405e58
Quoted 6 lines
Quoted 2 lines
Quoted 1 line

Yes, Andy's reporting that too. I asked him to identify the file-n-line and he ran away on me.

i checked the scheduler queue and nothing jumped out at me, except the cleanup bug fixed by the patch below. (which should be harmless in this particular case - nr_running should never be smaller than 0 or larger than ~4 billion. A fix is warranted nevertheless.)

Did you work out which divide is getting the div-by-zero? I started at it a bit and wasn't sure - am getting wildly different code generation over here.

As far as I can see all divides, except those that rely on group->cpu_power being non zero, in find_busiest_queue() are protected against divide by zero. So this would suggest that initialization of the scheduler group data would be the place to look.

Peter

Signature · 4 lines
Peter Williams                                   [email protected]

"Learning, n. The kind of ignorance distinguishing the studious."
 -- Ambrose Bierce
#5 lore raw

Re: [patch] sched: fix macro -> inline function conversion bug

In reply to Andrew Morton
Quoted 8 lines · Andrew Morton
Quoted 4 lines

i checked the scheduler queue and nothing jumped out at me, except the cleanup bug fixed by the patch below. (which should be harmless in this particular case - nr_running should never be smaller than 0 or larger than ~4 billion. A fix is warranted nevertheless.)

Did you work out which divide is getting the div-by-zero? I started at it a bit and wasn't sure - am getting wildly different code generation over here.

my bet is on sched-group-cpu-power-setup-cleanup.patch.

Ingo

#6 lore raw

Re: [patch] sched: fix macro -> inline function conversion bug

Quoted 5 lines · Ingo Molnar
Quoted 3 lines

Did you work out which divide is getting the div-by-zero? I started at it a bit and wasn't sure - am getting wildly different code generation over here.

my bet is on sched-group-cpu-power-setup-cleanup.patch.

in particular, we dont seem to initialize ->cpu_power properly. Martin, does the patch below solve your crash?

Ingo

-------------->
Subject: sched: group cpu power setup cleanup, fix
From: Ingo Molnar <[email protected]>

- fix missing initialization of ->cpu_power
- clean up the cleanup

---
 include/linux/sched.h |    2 +-
 kernel/sched.c        |    9 +++++++--
 2 files changed, 8 insertions(+), 3 deletions(-)
include/linux/sched.h +1 -1
Index: linux/include/linux/sched.h
===================================================================
@@ -636,7 +636,7 @@ enum idle_type
 	((sched_mc_power_savings || sched_smt_power_savings) ?	\
 					SD_POWERSAVINGS_BALANCE : 0)
 
-#define test_sd_flag(sd, flag)	((sd && sd->flags & flag) ? 1 : 0)
+#define test_sd_flag(sd, flag)	((sd && (sd->flags & flag)) ? 1 : 0)
 
 
 struct sched_group {
kernel/sched.c +7 -2
Index: linux/kernel/sched.c
===================================================================
@@ -1292,7 +1292,7 @@ static int sched_balance_self(int cpu, i
 		cpu = new_cpu;
 nextlevel:
 		sd = sd->child;
-		if (sd && sd->flags & flag)
+		if (test_sd_flag(sd, flag))
 			goto nextlevel;
 		/* while loop will break here if sd == NULL */
 	}
@@ -6224,6 +6224,7 @@ static int cpu_to_allnodes_group(int cpu
 {
 	return cpu_to_node(cpu);
 }
+
 static void init_numa_sched_groups_power(struct sched_group *group_head)
 {
 	struct sched_group *sg = group_head;
@@ -6314,8 +6315,12 @@ static void init_sched_groups_power(int 
 	struct sched_domain *child;
 	struct sched_group *group;
 
-	if (!sd || !sd->groups || (cpu != first_cpu(sd->groups->cpumask)))
+	WARN_ON(!sd || !sd->groups);
+
+	if (cpu != first_cpu(sd->groups->cpumask)) {
+		sd->groups->cpu_power = SCHED_LOAD_SCALE;
 		return;
+	}
 
 	child = sd->child;
 
#7 lore raw

Re: [patch] sched: fix macro -> inline function conversion bug

Martin, Andy: Can you please try the appended patch on top of 2.6.17-mm5?

thanks,
suresh

Quoted 12 lines · Ingo Molnar
Quoted 5 lines · Ingo Molnar
Quoted 3 lines

Did you work out which divide is getting the div-by-zero? I started at it a bit and wasn't sure - am getting wildly different code generation over here.

my bet is on sched-group-cpu-power-setup-cleanup.patch.

in particular, we dont seem to initialize ->cpu_power properly. Martin, does the patch below solve your crash?

Review

Quoted 3 lines
(quoted diff) +1 -1
 		sd = sd->child;
-		if (sd && sd->flags & flag)
+		if (test_sd_flag(sd, flag))

There is a bug in my patch. Appended patch fixes this.

Review

Quoted 6 lines
(quoted diff) +4 -1
-	if (!sd || !sd->groups || (cpu != first_cpu(sd->groups->cpumask)))
+	WARN_ON(!sd || !sd->groups);
+
+	if (cpu != first_cpu(sd->groups->cpumask)) {
+		sd->groups->cpu_power = SCHED_LOAD_SCALE;
 		return;

This is also not correct and will corrupt some of the groups cpu_power. NUMA sched group setup is some what different from the other domains like HT and SMP. Appended patch has the correct fix.

--

- go back to original numa sched group power initialization
- fix the sched_balance_self code
- some cleanup as suggested by Ingo.

kernel/sched.c renamed from kernel/sched.c~ +8 -7
@@ -1292,7 +1292,7 @@ static int sched_balance_self(int cpu, i
 		cpu = new_cpu;
 nextlevel:
 		sd = sd->child;
-		if (sd && sd->flags & flag)
+		if (sd && !(sd->flags & flag))
 			goto nextlevel;
 		/* while loop will break here if sd == NULL */
 	}
@@ -5534,7 +5534,7 @@ static void cpu_attach_domain(struct sch
 
 	if (sd && sd_degenerate(sd)) {
 		sd = sd->parent;
-		if(sd)
+		if (sd)
 			sd->child = NULL;
 	}
 
@@ -6224,6 +6224,7 @@ static int cpu_to_allnodes_group(int cpu
 {
 	return cpu_to_node(cpu);
 }
+
 static void init_numa_sched_groups_power(struct sched_group *group_head)
 {
 	struct sched_group *sg = group_head;
@@ -6314,7 +6315,9 @@ static void init_sched_groups_power(int 
 	struct sched_domain *child;
 	struct sched_group *group;
 
-	if (!sd || !sd->groups || (cpu != first_cpu(sd->groups->cpumask)))
+	WARN_ON(!sd || !sd->groups);
+
+	if (cpu != first_cpu(sd->groups->cpumask))
 		return;
 
 	child = sd->child;
@@ -6596,10 +6599,8 @@ static int build_sched_domains(const cpu
 	}
 
 #ifdef CONFIG_NUMA
-	for_each_cpu_mask(i, *cpu_map) {
-		sd = &per_cpu(node_domains, i);
-		init_sched_groups_power(i, sd);
-	}
+	for (i = 0; i < MAX_NUMNODES; i++)
+		init_numa_sched_groups_power(sched_group_nodes[i]);
 
 	init_numa_sched_groups_power(sched_group_allnodes);
 #endif
include/linux/sched.h renamed from include/linux/sched.h~ +1 -1
@@ -636,7 +636,7 @@ enum idle_type
 	((sched_mc_power_savings || sched_smt_power_savings) ?	\
 					SD_POWERSAVINGS_BALANCE : 0)
 
-#define test_sd_flag(sd, flag)	((sd && sd->flags & flag) ? 1 : 0)
+#define test_sd_flag(sd, flag)	((sd && (sd->flags & flag)) ? 1 : 0)
 
 
 struct sched_group {
#8 lore raw

Re: [patch] sched: fix macro -> inline function conversion bug

kernel/sched.c static int sched_balance_self(int cpu, i · line 1295

Quoted 2 lines · Siddha, Suresh B
kernel/sched.c renamed from kernel/sched.c~ +1 -1
-		if (sd && sd->flags & flag)
+		if (sd && !(sd->flags & flag))

use test_sd_flag() here, as i did in my fix patch.

include/linux/sched.h enum idle_type · line 639

Quoted 2 lines
include/linux/sched.h renamed from include/linux/sched.h~ +1 -1
-#define test_sd_flag(sd, flag)	((sd && sd->flags & flag) ? 1 : 0)
+#define test_sd_flag(sd, flag)	((sd && (sd->flags & flag)) ? 1 : 0)

remove the 'sd' check in test_sd_flag. In the other cases we know that there's an sd. (it's usually a sign of spaghetti code if tests like this include a check for the existence of the object checked)

Ingo

#9 lore raw

Re: [patch] sched: fix macro -> inline function conversion bug

Quoted 8 lines · Ingo Molnar

kernel/sched.c static int sched_balance_self(int cpu, i · line 1295

Quoted 2 lines · Siddha, Suresh B
kernel/sched.c renamed from kernel/sched.c~ +1 -1
-		if (sd && sd->flags & flag)
+		if (sd && !(sd->flags & flag))

use test_sd_flag() here, as i did in my fix patch.

Quoted 6 lines

include/linux/sched.h enum idle_type · line 639

Quoted 2 lines
include/linux/sched.h renamed from include/linux/sched.h~ +1 -1
-#define test_sd_flag(sd, flag)	((sd && sd->flags & flag) ? 1 : 0)
+#define test_sd_flag(sd, flag)	((sd && (sd->flags & flag)) ? 1 : 0)

remove the 'sd' check in test_sd_flag. In the other cases we know that there's an sd. (it's usually a sign of spaghetti code if tests like this include a check for the existence of the object checked)

In other cases, we are passing sd->parent as the first argument to test_sd_flag(). We know that there is a 'sd' but not sure about sd->parent or sd->child.

thanks,
suresh

#10 lore raw

Re: [patch] sched: fix macro -> inline function conversion bug

Quoted 16 lines · Siddha, Suresh B
Quoted 8 lines · Ingo Molnar

kernel/sched.c static int sched_balance_self(int cpu, i · line 1295

Quoted 2 lines · Siddha, Suresh B
kernel/sched.c renamed from kernel/sched.c~ +1 -1
-		if (sd && sd->flags & flag)
+		if (sd && !(sd->flags & flag))

use test_sd_flag() here, as i did in my fix patch.

Quoted 6 lines

include/linux/sched.h enum idle_type · line 639

Quoted 2 lines
include/linux/sched.h renamed from include/linux/sched.h~ +1 -1
-#define test_sd_flag(sd, flag)	((sd && sd->flags & flag) ? 1 : 0)
+#define test_sd_flag(sd, flag)	((sd && (sd->flags & flag)) ? 1 : 0)

remove the 'sd' check in test_sd_flag. In the other cases we know that there's an sd. (it's usually a sign of spaghetti code if tests like this include a check for the existence of the object checked)

Quoted 3 lines

In other cases, we are passing sd->parent as the first argument to test_sd_flag(). We know that there is a 'sd' but not sure about sd->parent or sd->child.

ok. But the first issue above should be fixed.

Ingo

#11 lore raw

Re: [patch] sched: fix macro -> inline function conversion bug

Quoted 19 lines · Ingo Molnar
Quoted 16 lines · Siddha, Suresh B
Quoted 14 lines · Ingo Molnar

kernel/sched.c static int sched_balance_self(int cpu, i · line 1295

Quoted 2 lines · Siddha, Suresh B
kernel/sched.c renamed from kernel/sched.c~ +1 -1
-		if (sd && sd->flags & flag)
+		if (sd && !(sd->flags & flag))

use test_sd_flag() here, as i did in my fix patch.

include/linux/sched.h enum idle_type · line 639

Quoted 2 lines
include/linux/sched.h renamed from include/linux/sched.h~ +1 -1
-#define test_sd_flag(sd, flag)	((sd && sd->flags & flag) ? 1 : 0)
+#define test_sd_flag(sd, flag)	((sd && (sd->flags & flag)) ? 1 : 0)

remove the 'sd' check in test_sd_flag. In the other cases we know that there's an sd. (it's usually a sign of spaghetti code if tests like this include a check for the existence of the object checked)

Quoted 5 lines
Quoted 3 lines

In other cases, we are passing sd->parent as the first argument to test_sd_flag(). We know that there is a 'sd' but not sure about sd->parent or sd->child.

ok. But the first issue above should be fixed.

I can't simply change it to test_sd_flag(). In sched_balance_self(), paths for sd == 0 and a 'flag' not set in sd->flags are different.

I can change that piece of code to (sd && !test_sd_flag(sd, flag)) though.. but that is not clean, right?

thanks,
suresh

#12 lore raw

Re: [patch] sched: fix macro -> inline function conversion bug

In reply to Siddha, Suresh B
Quoted 1 line · Suresh B

Martin, Andy: Can you please try the appended patch on top of 2.6.17-mm5?

Submitted, will let you know.

-apw

#13 lore raw
In reply to Andrew Morton
Quoted 36 lines · Andrew Morton
Quoted 14 lines · On Sun, 02 Jul 2006 16:27:55 -0700 "Martin J. Bligh" <[email protected]> wrote:

Panic on NUMA-Q (mm4 was fine). Presumably some new scheduler patch

divide error: 0000 [#1]
8K_STACKS SMP
last sysfs file:

Modules linked in:
CPU:    1
EIP:    0060:[<c0112b6e>]    Not tainted VLI
EFLAGS: 00010046   (2.6.17-mm5-autokern1 #1) 

EIP is at find_busiest_group+0x1a3/0x47c
eax: 00000000 ebx: 00000007 ecx: 00000000 edx: 00000000
esi: 00000000 edi: e75ff264 ebp: e7405ec8 esp: e7405e58
ds: 007b es: 007b ss: 0068
Process swapper (pid: 0, ti=e7404000 task=c13f8560 task.ti=e7404000)

Quoted 16 lines
Stack: e75ff264 00000010 c0119020 00000000 00000000 00000000 00000000 00000000 
      ffffffff 00000000 00000000 00000001 00000001 00000001 00000080 00000000 
      00000000 00000200 00000020 00000080 00000000 00000000 e75ff260 c1364960 
Call Trace:
[<c0119020>] vprintk+0x5f/0x213
[<c0112efb>] load_balance+0x54/0x1d6
[<c011332d>] rebalance_tick+0xc5/0xe3
[<c01137a3>] scheduler_tick+0x2cb/0x2d3
[<c01215b4>] update_process_times+0x51/0x5d
[<c010c224>] smp_apic_timer_interrupt+0x5a/0x61
[<c0102d5b>] apic_timer_interrupt+0x1f/0x24
[<c01006c0>] default_idle+0x0/0x59
[<c01006f1>] default_idle+0x31/0x59
[<c0100791>] cpu_idle+0x64/0x79
Code: 00 5b 83 f8 1f 89 c6 5f 0f 8e 63 ff ff ff 8b 45 e0 8b 55 e8 01 45 dc 8b 4a 08 89 c2 01 4d d4 c1 e2 07 89 d0 31 d2 89 ce c1 ee 07 <f7> f1 83 7d 9c 00 89 45 e0 74 17 89 45 d8 8b 55 e8 8b 4d a4 8b 
EIP: [<c0112b6e>] find_busiest_group+0x1a3/0x47c SS:ESP 0068:e7405e58
Quoted 2 lines

Yes, Andy's reporting that too. I asked him to identify the file-n-line and he ran away on me.

I went away to debug it, but then had to skip out to a BBQ. Its definatly the cpu_power on the group being zero.

group->cpu_power ZERO => c3150920

/me gives Ingo's patch a spin.

-apw

#14 lore raw
Quoted 43 lines · Andy Whitcroft
Quoted 37 lines · Andrew Morton
Quoted 14 lines · On Sun, 02 Jul 2006 16:27:55 -0700 "Martin J. Bligh" <[email protected]> wrote:

Panic on NUMA-Q (mm4 was fine). Presumably some new scheduler patch

divide error: 0000 [#1]
8K_STACKS SMP
last sysfs file:

Modules linked in:
CPU:    1
EIP:    0060:[<c0112b6e>]    Not tainted VLI
EFLAGS: 00010046   (2.6.17-mm5-autokern1 #1) 

EIP is at find_busiest_group+0x1a3/0x47c
eax: 00000000 ebx: 00000007 ecx: 00000000 edx: 00000000
esi: 00000000 edi: e75ff264 ebp: e7405ec8 esp: e7405e58
ds: 007b es: 007b ss: 0068
Process swapper (pid: 0, ti=e7404000 task=c13f8560 task.ti=e7404000)

Quoted 16 lines
Stack: e75ff264 00000010 c0119020 00000000 00000000 00000000 00000000 00000000 
     ffffffff 00000000 00000000 00000001 00000001 00000001 00000080 00000000 
     00000000 00000200 00000020 00000080 00000000 00000000 e75ff260 c1364960 
Call Trace:
[<c0119020>] vprintk+0x5f/0x213
[<c0112efb>] load_balance+0x54/0x1d6
[<c011332d>] rebalance_tick+0xc5/0xe3
[<c01137a3>] scheduler_tick+0x2cb/0x2d3
[<c01215b4>] update_process_times+0x51/0x5d
[<c010c224>] smp_apic_timer_interrupt+0x5a/0x61
[<c0102d5b>] apic_timer_interrupt+0x1f/0x24
[<c01006c0>] default_idle+0x0/0x59
[<c01006f1>] default_idle+0x31/0x59
[<c0100791>] cpu_idle+0x64/0x79
Code: 00 5b 83 f8 1f 89 c6 5f 0f 8e 63 ff ff ff 8b 45 e0 8b 55 e8 01 45 dc 8b 4a 08 89 c2 01 4d d4 c1 e2 07 89 d0 31 d2 89 ce c1 ee 07 <f7> f1 83 7d 9c 00 89 45 e0 74 17 89 45 d8 8b 55 e8 8b 4d a4 8b 
EIP: [<c0112b6e>] find_busiest_group+0x1a3/0x47c SS:ESP 0068:e7405e58
Quoted 2 lines

Yes, Andy's reporting that too. I asked him to identify the file-n-line and he ran away on me.

Quoted 6 lines

I went away to debug it, but then had to skip out to a BBQ. Its definatly the cpu_power on the group being zero.

group->cpu_power ZERO => c3150920

/me gives Ingo's patch a spin.

Ok. Thats not fixed it either.

Confirmed that it is caused by the patch below, backing it out sorts things:

sched-group-cpu-power-setup-cleanup.patch

Did a fair bit of analysis of this problem, and it seems that the issue is where we initialise the NUMA domains. For each cpu we initialise a domain spanning the whole machine, but ordered starting at the node in which we start. In the original code we used the following to initialise each of these groups:

        for (i = 0; i < MAX_NUMNODES; i++)
                init_numa_sched_groups_power(sched_group_nodes[i]);

init_numa_sched_groups_power iterated over all of the groups within the domain and added up power based on the physical packages. Now we use:

        for_each_cpu_mask(i, *cpu_map) {
                sd = &per_cpu(node_domains, i);
                init_sched_groups_power(i, sd);
        }

init_sched_groups_power only thinks in terms of the first group of the domain, which leaves the subsequent groups in the domain at power 0.

I've tried reverting just that part of the change (as attached) and that also seems to fix things. However, are we correct in all the other cases ignoring the subsequent groups? I am also not sure if this will change the purpose of the patch? It seems unlikely but ...

Suresh I guess I'll punt to you :).

-apw

sched revert numa domain init

Seems that the schedular domains for the NUMA nodea arn't being initialised correctly. Their cpu_power is left at 0 leading to panic in load balancing. Seem that just reverting this part of the change below is enough to fix booting of NUMA systems.

sched-group-cpu-power-setup-cleanup.patch

Not sure if this changes the purpose of that patch.

---
 sched.c |    6 ++----
 1 files changed, 2 insertions(+), 4 deletions(-)
kernel/sched.c +2 -4
diff -upN reference/kernel/sched.c current/kernel/sched.c
@@ -6656,10 +6656,8 @@ printk(KERN_WARNING "init CPU domains\n"
 
 printk(KERN_WARNING "init NUMA domains\n");
 #ifdef CONFIG_NUMA
-	for_each_cpu_mask(i, *cpu_map) {
-		sd = &per_cpu(node_domains, i);
-		init_sched_groups_power(i, sd);
-	}
+        for (i = 0; i < MAX_NUMNODES; i++)
+		init_numa_sched_groups_power(sched_group_nodes[i]);
 
 	init_numa_sched_groups_power(sched_group_allnodes);
 #endif