Skip to contentSkip to conversation

Kernel panic at i686

linux41 In review Philip Müller 9 participants 23 messages Jul 23, 2015 – Sep 17, 2015 Driver core

lore
Share & tools
lore
#0 lore raw

Hi all,

I started to test linux 4.1 series with rc6. However, I was never able to boot that kernel in i686 architecture. Trying it again with VirtualBox gave me more conclusions. Using one core it simply boots up. Using more than one CPU core it crashes with:

Failed to access perfctr msr (MSR c0010007 is 0)

task: f58e0000 ti: f58e8000 task.ti: f58e800
EIP: 0060:[<c135a903>] EFLAGS: 00010206 CPU: 0 EIP is at free_cache_attributes+0x83/0xd0
EAX: 00000001 EBX: f589d46c ECX: 00000090 EDX: 360c2000
ESI: 00000000 EDI: c1724a80 EBP: f58e9ec0 ESP: f58e9ea0
DS: 007b ES: 007b FS: 00d8 GS: 00e0 SS: 0068
CR0: 8005003b CR2: 000000ac CR3: 01731000 CR4: 000006d0

In more rich detail you can find that problem on my bug-tracker for Manjaro Linux:

https://github.com/manjaro/packages-core/issues/14

I just want to know if you are aware of it. With current 4.1.3 release I still face that issue ...

kind regards
Philip Müller
--------------------------
Manjaro Project-Lead

#1 lore raw

[linux41] regression with 'x86/cacheinfo: Move cacheinfo sysfs code to generic infrastructure' on AMD i686

Hi Guenter, Sudeep,

It now came down to 'x86/cacheinfo: Move cacheinfo sysfs code to generic infrastructure'[1] as you can see here[2].

The facts are:

- You can't boot on i686 only with more than one CPU core on AMD hardware (x86_64 however works)
- Using Ubuntu config[3] it boots but creates an kernel >= 1 GB in size on Manjaro
- By reverting 0d55ba4[4] the kernel boots.

So we have to find out what is causing this issue[5].

kind regards
Philip Müller
--------------------------
Manjaro Project-Lead

[1]
https://github.com/torvalds/linux/commit/0d55ba46bfbee64fd2b492b87bfe2ec172e7b056
[2]
https://raw.githubusercontent.com/philmmanjaro/linux41/master/git-bisect.txt
[3]
https://raw.githubusercontent.com/philmmanjaro/linux41/master/config.4.1.3-040103-generic
[4]
https://github.com/manjaro/packages-core/commit/f7b77f3e84295a6313a9181d520fb48e60453b64
[5] https://github.com/manjaro/packages-core/issues/14

Quoted 30 lines · Philip Müller

Hi all,

I started to test linux 4.1 series with rc6. However, I was never able to boot that kernel in i686 architecture. Trying it again with VirtualBox gave me more conclusions. Using one core it simply boots up. Using more than one CPU core it crashes with:

Failed to access perfctr msr (MSR c0010007 is 0)

task: f58e0000 ti: f58e8000 task.ti: f58e800
EIP: 0060:[<c135a903>] EFLAGS: 00010206 CPU: 0 EIP is at free_cache_attributes+0x83/0xd0
EAX: 00000001 EBX: f589d46c ECX: 00000090 EDX: 360c2000
ESI: 00000000 EDI: c1724a80 EBP: f58e9ec0 ESP: f58e9ea0
DS: 007b ES: 007b FS: 00d8 GS: 00e0 SS: 0068
CR0: 8005003b CR2: 000000ac CR3: 01731000 CR4: 000006d0

In more rich detail you can find that problem on my bug-tracker for Manjaro Linux:

https://github.com/manjaro/packages-core/issues/14

I just want to know if you are aware of it. With current 4.1.3 release I still face that issue ...

kind regards
Philip Müller
--------------------------
Manjaro Project-Lead

#2 lore raw

Re: [linux41] regression with 'x86/cacheinfo: Move cacheinfo sysfs code to generic infrastructure' on AMD i686

Quoted 7 lines · Philip Müller
Quoted 7 lines

task: f58e0000 ti: f58e8000 task.ti: f58e800
EIP: 0060:[<c135a903>] EFLAGS: 00010206 CPU: 0 EIP is at free_cache_attributes+0x83/0xd0
EAX: 00000001 EBX: f589d46c ECX: 00000090 EDX: 360c2000
ESI: 00000000 EDI: c1724a80 EBP: f58e9ec0 ESP: f58e9ea0
DS: 007b ES: 007b FS: 00d8 GS: 00e0 SS: 0068
CR0: 8005003b CR2: 000000ac CR3: 01731000 CR4: 000006d0

That's a trivial NULL pointer dereference in the error/cleanup path. Patch below should fix it.

Thanks,

tglx
---

drivers/base/cacheinfo.c +4 -2
index 764280a91776..f09b106d8b81 100644
@@ -159,6 +159,9 @@ static void cache_shared_cpu_map_remove(unsigned int cpu)
 
 static void free_cache_attributes(unsigned int cpu)
 {
+	if (!per_cpu_cacheinfo(cpu))
+		return;
+
 	cache_shared_cpu_map_remove(cpu);
 
 	kfree(per_cpu_cacheinfo(cpu));
@@ -514,8 +517,7 @@ static int cacheinfo_cpu_callback(struct notifier_block *nfb,
 		break;
 	case CPU_DEAD:
 		cache_remove_dev(cpu);
-		if (per_cpu_cacheinfo(cpu))
-			free_cache_attributes(cpu);
+		free_cache_attributes(cpu);
 		break;
 	}
 	return notifier_from_errno(rc);
#3 lore raw

Re: [linux41] regression with 'x86/cacheinfo: Move cacheinfo sysfs code to generic infrastructure' on AMD i686

Quoted 11 lines · Thomas Gleixner
Quoted 7 lines · Philip Müller
Quoted 7 lines

task: f58e0000 ti: f58e8000 task.ti: f58e800
EIP: 0060:[<c135a903>] EFLAGS: 00010206 CPU: 0 EIP is at free_cache_attributes+0x83/0xd0
EAX: 00000001 EBX: f589d46c ECX: 00000090 EDX: 360c2000
ESI: 00000000 EDI: c1724a80 EBP: f58e9ec0 ESP: f58e9ea0
DS: 007b ES: 007b FS: 00d8 GS: 00e0 SS: 0068
CR0: 8005003b CR2: 000000ac CR3: 01731000 CR4: 000006d0

That's a trivial NULL pointer dereference in the error/cleanup path. Patch below should fix it.

Well, I got a bit different, and of course totally untested possible solution:

cache_shared_cpu_map_setup() does check sib_cpu_ci->info_list before setting cpumask bits while cache_shared_cpu_map_remove() doesn't. Ballancing this out would mean:

---

drivers/base/cacheinfo.c +4 -0
index 764280a91776..8a4546dc25e3 100644
@@ -148,7 +148,11 @@ static void cache_shared_cpu_map_remove(unsigned int cpu)
 
 			if (sibling == cpu) /* skip itself */
 				continue;
+
 			sib_cpu_ci = get_cpu_cacheinfo(sibling);
+			if (!sib_cpu_ci->info_list)
+				continue;
+
 			sib_leaf = sib_cpu_ci->info_list + index;
 			cpumask_clear_cpu(cpu, &sib_leaf->shared_cpu_map);
 			cpumask_clear_cpu(sibling, &this_leaf->shared_cpu_map);

---

Now Philip can have some more fun testing :-)

Signature · 5 lines
Regards/Gruss,
    Boris.

ECO tip #101: Trim your mails when you reply.
--
#4 lore raw

Re: [linux41] regression with 'x86/cacheinfo: Move cacheinfo sysfs code to generic infrastructure' on AMD i686

Hi Borislav,

I can confirm your patch working. However, it might be good to use yours and Thomas' in combination to solve this properly.

kind regards
Philip

Quoted 41 lines · Borislav Petkov
Quoted 11 lines · Thomas Gleixner
Quoted 7 lines · Philip Müller
Quoted 7 lines

task: f58e0000 ti: f58e8000 task.ti: f58e800
EIP: 0060:[<c135a903>] EFLAGS: 00010206 CPU: 0 EIP is at free_cache_attributes+0x83/0xd0
EAX: 00000001 EBX: f589d46c ECX: 00000090 EDX: 360c2000
ESI: 00000000 EDI: c1724a80 EBP: f58e9ec0 ESP: f58e9ea0
DS: 007b ES: 007b FS: 00d8 GS: 00e0 SS: 0068
CR0: 8005003b CR2: 000000ac CR3: 01731000 CR4: 000006d0

That's a trivial NULL pointer dereference in the error/cleanup path. Patch below should fix it.

Well, I got a bit different, and of course totally untested possible solution:

cache_shared_cpu_map_setup() does check sib_cpu_ci->info_list before setting cpumask bits while cache_shared_cpu_map_remove() doesn't. Ballancing this out would mean:

---

drivers/base/cacheinfo.c +4 -1
index 764280a91776..8a4546dc25e3 100644
@@ -148,7 +148,11 @@ static void cache_shared_cpu_map_remove(unsigned int cpu)
 
 			if (sibling == cpu) /* skip itself */
 				continue;
+
 			sib_cpu_ci = get_cpu_cacheinfo(sibling);
+			if (!sib_cpu_ci->info_list)
+				continue;
+
 			sib_leaf = sib_cpu_ci->info_list + index;
 			cpumask_clear_cpu(cpu, &sib_leaf->shared_cpu_map);
 			cpumask_clear_cpu(sibling, &this_leaf->shared_cpu_map);
---

Now Philip can have some more fun testing :-)

#5 lore raw

Re: [linux41] regression with 'x86/cacheinfo: Move cacheinfo sysfs code to generic infrastructure' on AMD i686

Quoted 2 lines · Philip Müller

I can confirm your patch working. However, it might be good to use yours and Thomas' in combination to solve this properly.

Please do not top-post.

We could use Thomas' too although from looking at it, detect_cache_attributes() allocates a per-CPU per_cpu_cacheinfo thing for each CPU. By the time we hit cache_shared_cpu_map_remove() in free_cache_attributes(), those per_cpu_cacheinfo(cpu) things are still allocated. We kfree them in the next step only.

But I like the moving of the check from the CPU hotplug callback to free_cache_attributes().

So I'll merge the two patches and write up a proper commit message, unless someone objects.

I'll add your Tested-by too.

Thanks.

Signature · 5 lines
Regards/Gruss,
    Boris.

ECO tip #101: Trim your mails when you reply.
--
#6 lore raw

Re: [linux41] regression with 'x86/cacheinfo: Move cacheinfo sysfs code to generic infrastructure' on AMD i686

Hi Borislav,

I'm fine with that decision. I tested your patch alone and the combination with Thomas' changes. Both work to solve this problem.

Do whatever suits best for this matter. Thx to you too for providing solutions so fast.

kind regards
Philip

p.s. what do you mean by top-post?

Am 26.07.2015 um 16:42 schrieb Borislav Petkov:

Quoted 22 lines · Philip Müller
Quoted 2 lines

I can confirm your patch working. However, it might be good to use yours and Thomas' in combination to solve this properly.

Please do not top-post.

We could use Thomas' too although from looking at it, detect_cache_attributes() allocates a per-CPU per_cpu_cacheinfo thing for each CPU. By the time we hit cache_shared_cpu_map_remove() in free_cache_attributes(), those per_cpu_cacheinfo(cpu) things are still allocated. We kfree them in the next step only.

But I like the moving of the check from the CPU hotplug callback to free_cache_attributes().

So I'll merge the two patches and write up a proper commit message, unless someone objects.

I'll add your Tested-by too.

Thanks.

#7 lore raw

Re: [linux41] regression with 'x86/cacheinfo: Move cacheinfo sysfs code to generic infrastructure' on AMD i686

Quoted 5 lines · Philip Müller

Hi Borislav,

I'm fine with that decision. I tested your patch alone and the combination with Thomas' changes. Both work to solve this problem.

Quoted 8 lines

Do whatever suits best for this matter. Thx to you too for providing solutions so fast.

kind regards
Philip

p.s. what do you mean by top-post?

What you just did ;-).

http://ck.wikia.com/wiki/TopPosting

Guenter

Quoted 25 lines

Am 26.07.2015 um 16:42 schrieb Borislav Petkov:

Quoted 22 lines · Philip Müller
Quoted 2 lines

I can confirm your patch working. However, it might be good to use yours and Thomas' in combination to solve this properly.

Please do not top-post.

We could use Thomas' too although from looking at it, detect_cache_attributes() allocates a per-CPU per_cpu_cacheinfo thing for each CPU. By the time we hit cache_shared_cpu_map_remove() in free_cache_attributes(), those per_cpu_cacheinfo(cpu) things are still allocated. We kfree them in the next step only.

But I like the moving of the check from the CPU hotplug callback to free_cache_attributes().

So I'll merge the two patches and write up a proper commit message, unless someone objects.

I'll add your Tested-by too.

Thanks.

#12 lore raw

PATCH 1/1 cpu/cacheinfo: Fix teardown path

In reply to Philip Müller

From: Borislav Petkov <[email protected]>
Date: Mon, 27 Jul 2015 08:36:27 +0200
Subject: [PATCH] cpu/cacheinfo: Fix teardown path
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Philip Müller reported a hang when booting 32-bit 4.1 kernel on an AMD box. A fragment of the splat was enough to pinpoint the issue:

task: f58e0000 ti: f58e8000 task.ti: f58e800
EIP: 0060:[<c135a903>] EFLAGS: 00010206 CPU: 0 EIP is at free_cache_attributes+0x83/0xd0
EAX: 00000001 EBX: f589d46c ECX: 00000090 EDX: 360c2000
ESI: 00000000 EDI: c1724a80 EBP: f58e9ec0 ESP: f58e9ea0
DS: 007b ES: 007b FS: 00d8 GS: 00e0 SS: 0068
CR0: 8005003b CR2: 000000ac CR3: 01731000 CR4: 000006d0

cache_shared_cpu_map_setup() did check sibling CPUs cacheinfo descriptor while the respective teardown path cache_shared_cpu_map_remove() didn't. Fix that.

Quoted 1 line

From tglx's version: to be on the safe side, move the cacheinfo

descriptor check to free_cache_attributes(), thus cleaning up the hotplug path a little and making this even more robust.

---

Moin Thomas,

I've merged both patches and tagged it for stable. Which means, tip-urgent.

Thanks.

 drivers/base/cacheinfo.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)
drivers/base/cacheinfo.c +8 -2
index 764280a91776..e9fd32e91668 100644
@@ -148,7 +148,11 @@ static void cache_shared_cpu_map_remove(unsigned int cpu)
 
 			if (sibling == cpu) /* skip itself */
 				continue;
+
 			sib_cpu_ci = get_cpu_cacheinfo(sibling);
+			if (!sib_cpu_ci->info_list)
+				continue;
+
 			sib_leaf = sib_cpu_ci->info_list + index;
 			cpumask_clear_cpu(cpu, &sib_leaf->shared_cpu_map);
 			cpumask_clear_cpu(sibling, &this_leaf->shared_cpu_map);
@@ -159,6 +163,9 @@ static void cache_shared_cpu_map_remove(unsigned int cpu)
 
 static void free_cache_attributes(unsigned int cpu)
 {
+	if (!per_cpu_cacheinfo(cpu))
+		return;
+
 	cache_shared_cpu_map_remove(cpu);
 
 	kfree(per_cpu_cacheinfo(cpu));
@@ -514,8 +521,7 @@ static int cacheinfo_cpu_callback(struct notifier_block *nfb,
 		break;
 	case CPU_DEAD:
 		cache_remove_dev(cpu);
-		if (per_cpu_cacheinfo(cpu))
-			free_cache_attributes(cpu);
+		free_cache_attributes(cpu);
 		break;
 	}
 	return notifier_from_errno(rc);
Signature · 8 lines
2.5.0.rc2.28.g6003e7f

-- 
Regards/Gruss,
    Boris.

ECO tip #101: Trim your mails when you reply.
--
Acked-by on v1 1/1
#13 lore raw

Re: [PATCH] cpu/cacheinfo: Fix teardown path

Quoted 26 lines · Borislav Petkov

From: Borislav Petkov <[email protected]>
Date: Mon, 27 Jul 2015 08:36:27 +0200
Subject: [PATCH] cpu/cacheinfo: Fix teardown path
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Philip Müller reported a hang when booting 32-bit 4.1 kernel on an AMD box. A fragment of the splat was enough to pinpoint the issue:

   task: f58e0000 ti: f58e8000 task.ti: f58e800
   EIP: 0060:[<c135a903>] EFLAGS: 00010206 CPU: 0
   EIP is at free_cache_attributes+0x83/0xd0
   EAX: 00000001 EBX: f589d46c ECX: 00000090 EDX: 360c2000
   ESI: 00000000 EDI: c1724a80 EBP: f58e9ec0 ESP: f58e9ea0
    DS: 007b ES: 007b FS: 00d8 GS: 00e0 SS: 0068
   CR0: 8005003b CR2: 000000ac CR3: 01731000 CR4: 000006d0

cache_shared_cpu_map_setup() did check sibling CPUs cacheinfo descriptor while the respective teardown path cache_shared_cpu_map_remove() didn't. Fix that.

From tglx's version: to be on the safe side, move the cacheinfo
descriptor check to free_cache_attributes(), thus cleaning up the hotplug path a little and making this even more robust.

Quoted 10 lines
Cc Andre Przywara <[email protected]>
Cc Guenter Roeck <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Philip Müller <[email protected]>
Cc Sudeep Holla <[email protected]>

Looks good to me. If not too late

Regards,
Sudeep

Reviewed-by on v1 1/1
#14 lore raw

Re: [PATCH] cpu/cacheinfo: Fix teardown path

In reply to Borislav Petkov
Quoted 39 lines · Borislav Petkov

From: Borislav Petkov <[email protected]>
Date: Mon, 27 Jul 2015 08:36:27 +0200
Subject: [PATCH] cpu/cacheinfo: Fix teardown path
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Philip Müller reported a hang when booting 32-bit 4.1 kernel on an AMD box. A fragment of the splat was enough to pinpoint the issue:

task: f58e0000 ti: f58e8000 task.ti: f58e800
EIP: 0060:[<c135a903>] EFLAGS: 00010206 CPU: 0 EIP is at free_cache_attributes+0x83/0xd0
EAX: 00000001 EBX: f589d46c ECX: 00000090 EDX: 360c2000
ESI: 00000000 EDI: c1724a80 EBP: f58e9ec0 ESP: f58e9ea0
DS: 007b ES: 007b FS: 00d8 GS: 00e0 SS: 0068
CR0: 8005003b CR2: 000000ac CR3: 01731000 CR4: 000006d0

cache_shared_cpu_map_setup() did check sibling CPUs cacheinfo descriptor while the respective teardown path cache_shared_cpu_map_remove() didn't. Fix that.

Quoted 1 line

From tglx's version: to be on the safe side, move the cacheinfo

descriptor check to free_cache_attributes(), thus cleaning up the hotplug path a little and making this even more robust.

Quoted 6 lines

---

Moin Thomas,

I've merged both patches and tagged it for stable. Which means, tip-urgent.

#16 lore raw

[tip:x86/urgent] x86/cpu/cacheinfo: Fix teardown path

In reply to Borislav Petkov
x86/cpu/cacheinfo: Fix teardown path
Commit-ID:  680ac028240f8747f31c03986fbcf18b2b521e93
Gitweb:     http://git.kernel.org/tip/680ac028240f8747f31c03986fbcf18b2b521e93
Author:     Borislav Petkov <[email protected]>
AuthorDate: Mon, 27 Jul 2015 09:58:05 +0200
Committer:  Ingo Molnar <[email protected]>
CommitDate: Wed, 5 Aug 2015 10:08:17 +0200

x86/cpu/cacheinfo: Fix teardown path

Philip Müller reported a hang when booting 32-bit 4.1 kernel on an AMD box. A fragment of the splat was enough to pinpoint the issue:

task: f58e0000 ti: f58e8000 task.ti: f58e800
EIP: 0060:[<c135a903>] EFLAGS: 00010206 CPU: 0 EIP is at free_cache_attributes+0x83/0xd0
EAX: 00000001 EBX: f589d46c ECX: 00000090 EDX: 360c2000
ESI: 00000000 EDI: c1724a80 EBP: f58e9ec0 ESP: f58e9ea0
DS: 007b ES: 007b FS: 00d8 GS: 00e0 SS: 0068
CR0: 8005003b CR2: 000000ac CR3: 01731000 CR4: 000006d0

cache_shared_cpu_map_setup() did check sibling CPUs cacheinfo descriptor while the respective teardown path cache_shared_cpu_map_remove() didn't. Fix that.

Quoted 1 line

From tglx's version: to be on the safe side, move the cacheinfo

descriptor check to free_cache_attributes(), thus cleaning up the hotplug path a little and making this even more robust.

---
 drivers/base/cacheinfo.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)
drivers/base/cacheinfo.c +8 -2
index 764280a..e9fd32e 100644
@@ -148,7 +148,11 @@ static void cache_shared_cpu_map_remove(unsigned int cpu)
 
 			if (sibling == cpu) /* skip itself */
 				continue;
+
 			sib_cpu_ci = get_cpu_cacheinfo(sibling);
+			if (!sib_cpu_ci->info_list)
+				continue;
+
 			sib_leaf = sib_cpu_ci->info_list + index;
 			cpumask_clear_cpu(cpu, &sib_leaf->shared_cpu_map);
 			cpumask_clear_cpu(sibling, &this_leaf->shared_cpu_map);
@@ -159,6 +163,9 @@ static void cache_shared_cpu_map_remove(unsigned int cpu)
 
 static void free_cache_attributes(unsigned int cpu)
 {
+	if (!per_cpu_cacheinfo(cpu))
+		return;
+
 	cache_shared_cpu_map_remove(cpu);
 
 	kfree(per_cpu_cacheinfo(cpu));
@@ -514,8 +521,7 @@ static int cacheinfo_cpu_callback(struct notifier_block *nfb,
 		break;
 	case CPU_DEAD:
 		cache_remove_dev(cpu);
-		if (per_cpu_cacheinfo(cpu))
-			free_cache_attributes(cpu);
+		free_cache_attributes(cpu);
 		break;
 	}
 	return notifier_from_errno(rc);
Reviewed-by Thomas Gleixner Acked-by Sudeep Holla on v1 1/1
#17 lore raw

Re: [PATCH] cpu/cacheinfo: Fix teardown path

In reply to Borislav Petkov
Quoted 2 lines · Borislav Petkov

Philip Müller reported a hang when booting 32-bit 4.1 kernel on an AMD box. A fragment of the splat was enough to pinpoint the issue:

Bah, this goes to Greg and not to tip. Anyway, here's a version with updated tags.

Greg, please queue for 4.2 as it fixes a hang.

Thanks.

---
From: Borislav Petkov <[email protected]>
Date: Mon, 27 Jul 2015 08:36:27 +0200
Subject: [PATCH] cpu/cacheinfo: Fix teardown path
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Philip Müller reported a hang when booting 32-bit 4.1 kernel on an AMD box. A fragment of the splat was enough to pinpoint the issue:

task: f58e0000 ti: f58e8000 task.ti: f58e800
EIP: 0060:[<c135a903>] EFLAGS: 00010206 CPU: 0 EIP is at free_cache_attributes+0x83/0xd0
EAX: 00000001 EBX: f589d46c ECX: 00000090 EDX: 360c2000
ESI: 00000000 EDI: c1724a80 EBP: f58e9ec0 ESP: f58e9ea0
DS: 007b ES: 007b FS: 00d8 GS: 00e0 SS: 0068
CR0: 8005003b CR2: 000000ac CR3: 01731000 CR4: 000006d0

cache_shared_cpu_map_setup() did check sibling CPUs cacheinfo descriptor while the respective teardown path cache_shared_cpu_map_remove() didn't. Fix that.

Quoted 1 line

From tglx's version: to be on the safe side, move the cacheinfo

descriptor check to free_cache_attributes(), thus cleaning up the hotplug path a little and making this even more robust.

---
 drivers/base/cacheinfo.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)
drivers/base/cacheinfo.c +8 -2
index 764280a91776..e9fd32e91668 100644
@@ -148,7 +148,11 @@ static void cache_shared_cpu_map_remove(unsigned int cpu)
 
 			if (sibling == cpu) /* skip itself */
 				continue;
+
 			sib_cpu_ci = get_cpu_cacheinfo(sibling);
+			if (!sib_cpu_ci->info_list)
+				continue;
+
 			sib_leaf = sib_cpu_ci->info_list + index;
 			cpumask_clear_cpu(cpu, &sib_leaf->shared_cpu_map);
 			cpumask_clear_cpu(sibling, &this_leaf->shared_cpu_map);
@@ -159,6 +163,9 @@ static void cache_shared_cpu_map_remove(unsigned int cpu)
 
 static void free_cache_attributes(unsigned int cpu)
 {
+	if (!per_cpu_cacheinfo(cpu))
+		return;
+
 	cache_shared_cpu_map_remove(cpu);
 
 	kfree(per_cpu_cacheinfo(cpu));
@@ -514,8 +521,7 @@ static int cacheinfo_cpu_callback(struct notifier_block *nfb,
 		break;
 	case CPU_DEAD:
 		cache_remove_dev(cpu);
-		if (per_cpu_cacheinfo(cpu))
-			free_cache_attributes(cpu);
+		free_cache_attributes(cpu);
 		break;
 	}
 	return notifier_from_errno(rc);
Signature · 8 lines
2.5.0.rc2.28.g6003e7f

-- 
Regards/Gruss,
    Boris.

ECO tip #101: Trim your mails when you reply.
--
maintainer
#18 lore raw

Re: [PATCH] cpu/cacheinfo: Fix teardown path

Quoted 8 lines · Borislav Petkov
Quoted 2 lines · Borislav Petkov

Philip Müller reported a hang when booting 32-bit 4.1 kernel on an AMD box. A fragment of the splat was enough to pinpoint the issue:

Bah, this goes to Greg and not to tip. Anyway, here's a version with updated tags.

Greg, please queue for 4.2 as it fixes a hang.

What commit caused this issue?

And it's a bit late for 4.2, as you say 4.1 is also affected, I'll wait for 4.3-rc1 to give this a chance to get some testing.

thanks,

greg k-h

#19 lore raw

Re: [PATCH] cpu/cacheinfo: Fix teardown path

Hi Greg,

I bi-sected it to following commit:

0d55ba46bfbee64fd2b492b87bfe2ec172e7b056 is the first bad commit commit 0d55ba46bfbee64fd2b492b87bfe2ec172e7b056

Author: Sudeep Holla <[email protected]>
Date:   Wed Mar 4 12:00:16 2015 +0000

x86/cacheinfo: Move cacheinfo sysfs code to generic infrastructure

You can follow it on my github repo in rich detail:

https://github.com/philmmanjaro/linux41/blob/master/git-bisect.txt

kind regards
Philip Müller

Am 08.08.2015 um 17:41 schrieb Greg KH:

Quoted 19 lines · Borislav Petkov
Quoted 8 lines · Borislav Petkov
Quoted 2 lines

Philip Müller reported a hang when booting 32-bit 4.1 kernel on an AMD box. A fragment of the splat was enough to pinpoint the issue:

Bah, this goes to Greg and not to tip. Anyway, here's a version with updated tags.

Greg, please queue for 4.2 as it fixes a hang.

What commit caused this issue?

And it's a bit late for 4.2, as you say 4.1 is also affected, I'll wait for 4.3-rc1 to give this a chance to get some testing.

thanks,

greg k-h

#20 lore raw

Re: [PATCH] cpu/cacheinfo: Fix teardown path

Quoted 6 lines · Philip Müller

Hi Greg,

I bi-sected it to following commit:

0d55ba46bfbee64fd2b492b87bfe2ec172e7b056 is the first bad commit commit 0d55ba46bfbee64fd2b492b87bfe2ec172e7b056

Quoted 8 lines
Author: Sudeep Holla <[email protected]>
Date:   Wed Mar 4 12:00:16 2015 +0000

x86/cacheinfo: Move cacheinfo sysfs code to generic infrastructure

You can follow it on my github repo in rich detail:

https://github.com/philmmanjaro/linux41/blob/master/git-bisect.txt

Philip,

what is with you and top-posting? How hard is it not to do it?!

Please stop with the top-posting already.

Signature · 5 lines
Regards/Gruss,
    Boris.

ECO tip #101: Trim your mails when you reply.
--
#21 lore raw

Re: [PATCH] cpu/cacheinfo: Fix teardown path

In reply to Greg KH
Quoted 1 line · Greg KH

What commit caused this issue?

Apparently

0d55ba46bfbe ("x86/cacheinfo: Move cacheinfo sysfs code to generic infrastructure")

Looks like moving x86 to the generic cacheinfo stuff uncovered this shortcoming there...

Quoted 2 lines

And it's a bit late for 4.2, as you say 4.1 is also affected, I'll wait for 4.3-rc1 to give this a chance to get some testing.

Right, I guess that's fine too as it'll trickle to stable eventually...

Thanks.

Signature · 5 lines
Regards/Gruss,
    Boris.

ECO tip #101: Trim your mails when you reply.
--
#22 lore raw

Re: [PATCH] cpu/cacheinfo: Fix teardown path

Quoted 10 lines · Borislav Petkov
Quoted 1 line · Greg KH

What commit caused this issue?

Apparently

0d55ba46bfbe ("x86/cacheinfo: Move cacheinfo sysfs code to generic infrastructure")

Looks like moving x86 to the generic cacheinfo stuff uncovered this shortcoming there...

Quoted 7 lines
Quoted 2 lines

And it's a bit late for 4.2, as you say 4.1 is also affected, I'll wait for 4.3-rc1 to give this a chance to get some testing.

Right, I guess that's fine too as it'll trickle to stable eventually...

Thanks.

Just a note from my end. Seems this patch didn't made it into 4.3-rc1. Any reason why?

drivers/base/cacheinfo.c +8 -2
index 764280a91776..e9fd32e91668 100644
@@ -148,7 +148,11 @@ static void cache_shared_cpu_map_remove(unsigned int cpu)
 
 			if (sibling == cpu) /* skip itself */
 				continue;
+
 			sib_cpu_ci = get_cpu_cacheinfo(sibling);
+			if (!sib_cpu_ci->info_list)
+				continue;
+
 			sib_leaf = sib_cpu_ci->info_list + index;
 			cpumask_clear_cpu(cpu, &sib_leaf->shared_cpu_map);
 			cpumask_clear_cpu(sibling, &this_leaf->shared_cpu_map);
@@ -159,6 +163,9 @@ static void cache_shared_cpu_map_remove(unsigned int cpu)
 
 static void free_cache_attributes(unsigned int cpu)
 {
+	if (!per_cpu_cacheinfo(cpu))
+		return;
+
 	cache_shared_cpu_map_remove(cpu);
 
 	kfree(per_cpu_cacheinfo(cpu));
@@ -514,8 +521,7 @@ static int cacheinfo_cpu_callback(struct notifier_block *nfb,
 		break;
 	case CPU_DEAD:
 		cache_remove_dev(cpu);
-		if (per_cpu_cacheinfo(cpu))
-			free_cache_attributes(cpu);
+		free_cache_attributes(cpu);
 		break;
 	}
 	return notifier_from_errno(rc);
#8 lore raw

Re: [linux41] regression with 'x86/cacheinfo: Move cacheinfo sysfs code to generic infrastructure' on AMD i686

In reply to Borislav Petkov
Quoted 12 lines · Borislav Petkov
Quoted 2 lines · Philip Müller

I can confirm your patch working. However, it might be good to use yours and Thomas' in combination to solve this properly.

Please do not top-post.

We could use Thomas' too although from looking at it, detect_cache_attributes() allocates a per-CPU per_cpu_cacheinfo thing for each CPU. By the time we hit cache_shared_cpu_map_remove() in free_cache_attributes(), those per_cpu_cacheinfo(cpu) things are still allocated. We kfree them in the next step only.

Quoted 7 lines

But I like the moving of the check from the CPU hotplug callback to free_cache_attributes().

So I'll merge the two patches and write up a proper commit message, unless someone objects.

I'll add your Tested-by too.

Did this actually happen? I don't see either fix in Linus' tree yet, the merge window is closed, and the bug happens on 4.1 and 4.2 stable kernels..

josh

#9 lore raw

Re: [linux41] regression with 'x86/cacheinfo: Move cacheinfo sysfs code to generic infrastructure' on AMD i686

Am 17.09.2015 um 01:52 schrieb Josh Boyer:

Quoted 7 lines

Did this actually happen? I don't see either fix in Linus' tree yet, the merge window is closed, and the bug happens on 4.1 and 4.2 stable kernels..

josh

Seems not yet. I don't see it neither in 4.3-rc1. Seems 4.3 will have the same issues then ...

#10 lore raw

Re: [linux41] regression with 'x86/cacheinfo: Move cacheinfo sysfs code to generic infrastructure' on AMD i686

In reply to Josh Boyer
Quoted 16 lines · Josh Boyer
Quoted 15 lines · Borislav Petkov
Quoted 2 lines · Philip Müller

I can confirm your patch working. However, it might be good to use yours and Thomas' in combination to solve this properly.

Please do not top-post.

We could use Thomas' too although from looking at it, detect_cache_attributes() allocates a per-CPU per_cpu_cacheinfo thing for each CPU. By the time we hit cache_shared_cpu_map_remove() in free_cache_attributes(), those per_cpu_cacheinfo(cpu) things are still allocated. We kfree them in the next step only.

But I like the moving of the check from the CPU hotplug callback to free_cache_attributes().

Quoted 8 lines
Quoted 4 lines

So I'll merge the two patches and write up a proper commit message, unless someone objects.

I'll add your Tested-by too.

Did this actually happen? I don't see either fix in Linus' tree yet, the merge window is closed, and the bug happens on 4.1 and 4.2 stable kernels..

Greg wanted to pick it up...

Greg, what's up?

Signature · 5 lines
Regards/Gruss,
    Boris.

ECO tip #101: Trim your mails when you reply.
--
maintainer
#11 lore raw

Re: [linux41] regression with 'x86/cacheinfo: Move cacheinfo sysfs code to generic infrastructure' on AMD i686

Quoted 22 lines · Borislav Petkov
Quoted 21 lines · Josh Boyer
Quoted 19 lines · Borislav Petkov
Quoted 2 lines · Philip Müller

I can confirm your patch working. However, it might be good to use yours and Thomas' in combination to solve this properly.

Please do not top-post.

We could use Thomas' too although from looking at it, detect_cache_attributes() allocates a per-CPU per_cpu_cacheinfo thing for each CPU. By the time we hit cache_shared_cpu_map_remove() in free_cache_attributes(), those per_cpu_cacheinfo(cpu) things are still allocated. We kfree them in the next step only.

But I like the moving of the check from the CPU hotplug callback to free_cache_attributes().

So I'll merge the two patches and write up a proper commit message, unless someone objects.

I'll add your Tested-by too.

Quoted 7 lines
Quoted 3 lines

Did this actually happen? I don't see either fix in Linus' tree yet, the merge window is closed, and the bug happens on 4.1 and 4.2 stable kernels..

Greg wanted to pick it up...

Greg, what's up?

It's in my "to-apply" queue, let me go dig it up now...

thanks for the reminder.

greg k-h