Skip to contentSkip to conversation

alignment exceptionhandler sleeps in invalid context

DISCUSSION Olaf Hering 4 participants 8 messages Apr 24, 2006 – May 2, 2006

lore
Share & tools
lore
#0 lore raw

I'm not sure where the bug is. Does it mean the network stack does something nasty, or is the exception handler itself broken? (probably the latter) This is 2.6.16.9 on a p270.

<3>Debug: sleeping function called from invalid context at arch/powerpc/kernel/align.c:440 <4>in_atomic():1, irqs_disabled():0
<4>Call Trace:
<4>[C00000000FFE6D10] [C00000000000EE20] .show_stack+0x68/0x1b0 (unreliable) <4>[C00000000FFE6DB0] [C000000000054248] .__might_sleep+0xd8/0xf4 <4>[C00000000FFE6E30] [C00000000000C490] .fix_alignment+0x50c/0x9c8 <4>[C00000000FFE6F10] [C000000000024E20] .alignment_exception+0x18/0xc8 <4>[C00000000FFE6F90] [C00000000000430C] alignment_common+0x10c/0x180 <4>--- Exception: 600 at .csum_partial+0x34/0x9c
<4> LR = .skb_checksum+0x194/0x35c
<4>[C00000000FFE7280] [D00000000033F3D0] 0xd00000000033f3d0 (unreliable) <4>[C00000000FFE7340] [C0000000002EB558] .skb_checksum_help+0x98/0x124 <4>[C00000000FFE73D0] [D000000000330748] .ip_nat_fn+0x84/0x278 [iptable_nat] <4>[C00000000FFE7490] [D000000000330C34] .ip_nat_local_fn+0x40/0xec [iptable_nat] <4>[C00000000FFE7520] [C00000000030DAE0] .nf_iterate+0x80/0x11c <4>[C00000000FFE75E0] [C00000000030DEC8] .nf_hook_slow+0x8c/0x17c <4>[C00000000FFE76B0] [C00000000031D2B0] .ip_queue_xmit+0x4bc/0x5d0 <4>[C00000000FFE7800] [C00000000032FBA0] .tcp_transmit_skb+0x7d8/0x844 <4>[C00000000FFE78B0] [C0000000003318F0] .__tcp_push_pending_frames+0x2d4/0x3fc <4>[C00000000FFE7970] [C00000000032EAD4] .tcp_rcv_established+0x894/0x964 <4>[C00000000FFE7A20] [C0000000003355C8] .tcp_v4_do_rcv+0x5c/0x48c <4>[C00000000FFE7AF0] [C000000000338A48] .tcp_v4_rcv+0xc78/0xcfc <4>[C00000000FFE7BC0] [C000000000316310] .ip_local_deliver+0x208/0x36c <4>[C00000000FFE7C50] [C000000000316024] .ip_rcv+0x600/0x6e4 <4>[C00000000FFE7CF0] [C0000000002EBB34] .netif_receive_skb+0x500/0x598 <4>[C00000000FFE7D90] [C0000000002EE37C] .process_backlog+0xcc/0x1c8 <4>[C00000000FFE7E40] [C0000000002EE55C] .net_rx_action+0xe4/0x230 <4>[C00000000FFE7EF0] [C000000000066524] .__do_softirq+0x98/0x164 <4>[C00000000FFE7F90] [C000000000025854] .call_do_softirq+0x14/0x24 <4>[C0000003BA3C3370] [C00000000000BCC8] .do_softirq+0x8c/0xd8 <4>[C0000003BA3C3400] [C0000000000663E4] .local_bh_enable+0x58/0x8c <4>[C0000003BA3C3480] [C0000000002EEA34] .dev_queue_xmit+0x2f0/0x318 <4>[C0000003BA3C3510] [C00000000031CA60] .ip_output+0x348/0x3f0 <4>[C0000003BA3C35B0] [C00000000031D2E0] .ip_queue_xmit+0x4ec/0x5d0 <4>[C0000003BA3C3700] [C00000000032FBA0] .tcp_transmit_skb+0x7d8/0x844 <4>[C0000003BA3C37B0] [C0000000003231A8] .tcp_cleanup_rbuf+0x148/0x168 <4>[C0000003BA3C3840] [C000000000325B24] .tcp_recvmsg+0x8a8/0xa00 <4>[C0000003BA3C3930] [C0000000002E0C64] .sock_common_recvmsg+0x5c/0x84 <4>[C0000003BA3C39C0] [C0000000002DD408] .do_sock_read+0x120/0x14c <4>[C0000003BA3C3A60] [C0000000002DF034] .sock_aio_read+0x58/0x74 <4>[C0000003BA3C3B70] [C0000000000C6050] .do_sync_read+0xd4/0x130 <4>[C0000003BA3C3CF0] [C0000000000C6ECC] .vfs_read+0x134/0x1fc <4>[C0000003BA3C3D90] [C0000000000C7390] .sys_read+0x4c/0x8c <4>[C0000003BA3C3E30] [C00000000000871C] syscall_exit+0x0/0x40

#1 lore raw
Quoted 3 lines · Olaf Hering

I'm not sure where the bug is. Does it mean the network stack does something nasty, or is the exception handler itself broken? (probably the latter) This is 2.6.16.9 on a p270.

The alignment handler does a get_user to read the instruction. I suppose it will have to read the instruction directly if the exception occurred in kernel mode.

Paul.

#2 lore raw
In reply to Olaf Hering
Quoted 3 lines · Olaf Hering

I'm not sure where the bug is. Does it mean the network stack does something nasty, or is the exception handler itself broken? (probably the latter) This is 2.6.16.9 on a p270.

This patch should fix it, I hope. If you can verify that it fixes it I'll send it to Linus.

Paul.

include/asm-powerpc/uaccess.h +6 -3
index 3872e92..b02d858 100644
@@ -179,7 +179,8 @@ do {								\
 #define __put_user_nocheck(x, ptr, size)			\
 ({								\
 	long __pu_err;						\
-	might_sleep();						\
+	if ((unsigned long)ptr < PAGE_OFFSET)			\
+		might_sleep();					\
 	__chk_user_ptr(ptr);					\
 	__put_user_size((x), (ptr), (size), __pu_err);		\
 	__pu_err;						\
@@ -259,7 +260,8 @@ ({								\
 	long __gu_err;						\
 	unsigned long __gu_val;					\
 	__chk_user_ptr(ptr);					\
-	might_sleep();						\
+	if ((unsigned long)ptr < PAGE_OFFSET)			\
+		might_sleep();					\
 	__get_user_size(__gu_val, (ptr), (size), __gu_err);	\
 	(x) = (__typeof__(*(ptr)))__gu_val;			\
 	__gu_err;						\
@@ -271,7 +273,8 @@ ({								\
 	long __gu_err;						\
 	long long __gu_val;					\
 	__chk_user_ptr(ptr);					\
-	might_sleep();						\
+	if ((unsigned long)ptr < PAGE_OFFSET)			\
+		might_sleep();					\
 	__get_user_size(__gu_val, (ptr), (size), __gu_err);	\
 	(x) = (__typeof__(*(ptr)))__gu_val;			\
 	__gu_err;						\
#3 lore raw
Quoted 8 lines · Paul Mackeras
Quoted 3 lines · Olaf Hering

I'm not sure where the bug is. Does it mean the network stack does something nasty, or is the exception handler itself broken? (probably the latter) This is 2.6.16.9 on a p270.

This patch should fix it, I hope. If you can verify that it fixes it I'll send it to Linus.

I dont have a testcase, it just happend once on the p270.

#4 lore raw
Quoted 4 lines · Olaf Hering
Quoted 2 lines · Paul Mackeras

Olaf Hering writes:

Quoted 8 lines
Quoted 6 lines
Quoted 3 lines

I'm not sure where the bug is. Does it mean the network stack does something nasty, or is the exception handler itself broken? (probably the latter) This is 2.6.16.9 on a p270.

This patch should fix it, I hope. If you can verify that it fixes it I'll send it to Linus.

I dont have a testcase, it just happend once on the p270.

I was going to ask how this was actually hit. It seems odd that we would have an alignment error in the network stack. From the oops can we determine what the instruction was that faulted and what address it was trying to access.

- k

#5 lore raw
Quoted 4 lines · Kumar Gala

I was going to ask how this was actually hit. It seems odd that we would have an alignment error in the network stack. From the oops can we determine what the instruction was that faulted and what address it was trying to access.

No, it's not particularly odd, given that our get_unaligned() and put_unaligned() in include/asm-powerpc/unaligned.h just do a single access. That makes the common case (address is actually aligned, or cpu handles it in hardware) fast at the expense of occasionally taking an alignment exception.

Paul.

#6 lore raw
In reply to Paul Mackerras
Quoted 16 lines · Paul Mackerras
Quoted 3 lines · Olaf Hering

I'm not sure where the bug is. Does it mean the network stack does something nasty, or is the exception handler itself broken? (probably the latter) This is 2.6.16.9 on a p270.

This patch should fix it, I hope. If you can verify that it fixes it I'll send it to Linus.

Paul.

include/asm-powerpc/uaccess.h +0 -0
index 3872e92..b02d858 100644

include/asm-powerpc/uaccess.h do { \ · line 183

Quoted 6 lines
include/asm-powerpc/uaccess.h +2 -1
@@ -179,7 +179,8 @@ do {								\
 #define __put_user_nocheck(x, ptr, size)			\
 ({								\
 	long __pu_err;						\
-	might_sleep();						\
+	if ((unsigned long)ptr < PAGE_OFFSET)			\
+		might_sleep();					\

+ if (!is_kernel_addr((unsigned long)ptr)) \
+ might_sleep(); \

In asm/page.h :)

cheers

Signature · 8 lines
Michael Ellerman
IBM OzLabs

wwweb: http://michael.ellerman.id.au
phone: +61 2 6212 1183 (tie line 70 21183)

We do not inherit the earth from our ancestors,
we borrow it from our children. - S.M.A.R.T Person
#7 lore raw
In reply to Paul Mackerras
Quoted 2 lines · I wrote:

This patch should fix it, I hope. If you can verify that it fixes it I'll send it to Linus.

but init barfs at startup with that patch, due to the double evaluation of `ptr'. This one runs a bit better...

Paul.

include/asm-powerpc/uaccess.h +15 -8
index 3872e92..d83fc29 100644
@@ -7,6 +7,7 @@ #ifndef __ASSEMBLY__
 #include <linux/sched.h>
 #include <linux/errno.h>
 #include <asm/processor.h>
+#include <asm/page.h>
 
 #define VERIFY_READ	0
 #define VERIFY_WRITE	1
@@ -179,9 +180,11 @@ do {								\
 #define __put_user_nocheck(x, ptr, size)			\
 ({								\
 	long __pu_err;						\
-	might_sleep();						\
-	__chk_user_ptr(ptr);					\
-	__put_user_size((x), (ptr), (size), __pu_err);		\
+	__typeof__(*(ptr)) __user *__pu_addr = (ptr);		\
+	if (!is_kernel_addr((unsigned long)__pu_addr))		\
+		might_sleep();					\
+	__chk_user_ptr(ptr);					\
+	__put_user_size((x), __pu_addr, (size), __pu_err);	\
 	__pu_err;						\
 })
 
@@ -258,9 +261,11 @@ #define __get_user_nocheck(x, ptr, size)
 ({								\
 	long __gu_err;						\
 	unsigned long __gu_val;					\
+	const __typeof__(*(ptr)) __user *__gu_addr = (ptr);	\
 	__chk_user_ptr(ptr);					\
-	might_sleep();						\
-	__get_user_size(__gu_val, (ptr), (size), __gu_err);	\
+	if (!is_kernel_addr((unsigned long)__gu_addr))		\
+		might_sleep();					\
+	__get_user_size(__gu_val, __gu_addr, (size), __gu_err);	\
 	(x) = (__typeof__(*(ptr)))__gu_val;			\
 	__gu_err;						\
 })
@@ -270,9 +275,11 @@ #define __get_user64_nocheck(x, ptr, siz
 ({								\
 	long __gu_err;						\
 	long long __gu_val;					\
-	__chk_user_ptr(ptr);					\
-	might_sleep();						\
-	__get_user_size(__gu_val, (ptr), (size), __gu_err);	\
+	const __typeof__(*(ptr)) __user *__gu_addr = (ptr);	\
+	__chk_user_ptr(ptr);					\
+	if (!is_kernel_addr((unsigned long)__gu_addr))		\
+		might_sleep();					\
+	__get_user_size(__gu_val, __gu_addr, (size), __gu_err);	\
 	(x) = (__typeof__(*(ptr)))__gu_val;			\
 	__gu_err;						\
 })