alignment faults in 3.6

Quoted 26 lines · Russell King - ARM Linux

Quoted 25 lines · Mans Rullgard
Quoted 22 lines · Russell King - ARM Linux
Quoted 9 lines · Mans Rullgard
Quoted 7 lines · Rob Herring
This struct is the IP header, so a struct ptr is just set to the beginning of the received data. Since ethernet headers are 14 bytes, often the IP header is not aligned unless the NIC can place the frame at a 2 byte offset (which is something I need to investigate). So this function cannot make any assumptions about the alignment. Does the ABI define structs have some minimum alignment? Does the struct need to be declared as packed or something?
Quoted 4 lines
The ABI defines the alignment of structs as the maximum alignment of its members. Since this struct contains 32-bit members, the alignment for the whole struct becomes 32 bits as well. Declaring it as packed tells gcc it might be unaligned (in addition to removing any holes within).
This has come up before in the past.
The Linux network folk will _not_ allow - in any shape or form - for this struct to be marked packed (it's the struct which needs to be marked packed) because by doing so, it causes GCC to issue byte loads/ stores on architectures where there isn't a problem, and that decreases the performance of the Linux IP stack unnecessarily.

Quoted 5 lines

Quoted 1 line
Which architectures? I have never seen anything like that.
Does it matter? I'm just relaying the argument against adding __packed which was used before we were forced (by the networking folk) to implement the alignment fault handler.

It doesn't really matter what will be accepted or not as adding __packed to struct iphdr doesn't fix the problem anyway. gcc still emits a ldm. The only way I've found to eliminate the alignment fault is adding a barrier between the 2 loads. That seems like a compiler issue to me if there is not a better fix.

Rob

Mikael Pettersson

2012-10-05 13:51 UTC #16 lore raw

Quoted 33 lines · Rob Herring

Quoted 26 lines · Russell King - ARM Linux
Quoted 25 lines · Mans Rullgard
Quoted 22 lines · Russell King - ARM Linux
Quoted 9 lines · Mans Rullgard
Quoted 7 lines · Rob Herring
This struct is the IP header, so a struct ptr is just set to the beginning of the received data. Since ethernet headers are 14 bytes, often the IP header is not aligned unless the NIC can place the frame at a 2 byte offset (which is something I need to investigate). So this function cannot make any assumptions about the alignment. Does the ABI define structs have some minimum alignment? Does the struct need to be declared as packed or something?
Quoted 4 lines
The ABI defines the alignment of structs as the maximum alignment of its members. Since this struct contains 32-bit members, the alignment for the whole struct becomes 32 bits as well. Declaring it as packed tells gcc it might be unaligned (in addition to removing any holes within).
This has come up before in the past.
The Linux network folk will _not_ allow - in any shape or form - for this struct to be marked packed (it's the struct which needs to be marked packed) because by doing so, it causes GCC to issue byte loads/ stores on architectures where there isn't a problem, and that decreases the performance of the Linux IP stack unnecessarily.
Quoted 5 lines
Quoted 1 line
Which architectures? I have never seen anything like that.
Does it matter? I'm just relaying the argument against adding __packed which was used before we were forced (by the networking folk) to implement the alignment fault handler.

Quoted 5 lines

It doesn't really matter what will be accepted or not as adding __packed to struct iphdr doesn't fix the problem anyway. gcc still emits a ldm. The only way I've found to eliminate the alignment fault is adding a barrier between the 2 loads. That seems like a compiler issue to me if there is not a better fix.

If you suspect a GCC bug, please prepare a standalone user-space test case and submit it to GCC's bugzilla (I can do the latter if you absolutely do not want to). It wouldn't be the first alignment-related GCC bug...

Rob Herring

author

2012-10-05 16:01 UTC #17 lore raw

Quoted 40 lines · Mikael Pettersson

Quoted 33 lines · Rob Herring
Quoted 26 lines · Russell King - ARM Linux
Quoted 25 lines · Mans Rullgard
Quoted 22 lines · Russell King - ARM Linux
Quoted 9 lines · Mans Rullgard
Quoted 7 lines · Rob Herring
This struct is the IP header, so a struct ptr is just set to the beginning of the received data. Since ethernet headers are 14 bytes, often the IP header is not aligned unless the NIC can place the frame at a 2 byte offset (which is something I need to investigate). So this function cannot make any assumptions about the alignment. Does the ABI define structs have some minimum alignment? Does the struct need to be declared as packed or something?
Quoted 4 lines
The ABI defines the alignment of structs as the maximum alignment of its members. Since this struct contains 32-bit members, the alignment for the whole struct becomes 32 bits as well. Declaring it as packed tells gcc it might be unaligned (in addition to removing any holes within).
This has come up before in the past.
The Linux network folk will _not_ allow - in any shape or form - for this struct to be marked packed (it's the struct which needs to be marked packed) because by doing so, it causes GCC to issue byte loads/ stores on architectures where there isn't a problem, and that decreases the performance of the Linux IP stack unnecessarily.
Quoted 5 lines
Quoted 1 line
Which architectures? I have never seen anything like that.
Does it matter? I'm just relaying the argument against adding __packed which was used before we were forced (by the networking folk) to implement the alignment fault handler.
Quoted 5 lines
It doesn't really matter what will be accepted or not as adding __packed to struct iphdr doesn't fix the problem anyway. gcc still emits a ldm. The only way I've found to eliminate the alignment fault is adding a barrier between the 2 loads. That seems like a compiler issue to me if there is not a better fix.

Quoted 4 lines

If you suspect a GCC bug, please prepare a standalone user-space test case and submit it to GCC's bugzilla (I can do the latter if you absolutely do not want to). It wouldn't be the first alignment-related GCC bug...

Here's a testcase. Compiled on ubuntu precise with "arm-linux-gnueabihf-gcc -O2 -marm -march=armv7-a test.c".

typedef unsigned short u16;
typedef unsigned short __sum16;
typedef unsigned int __u32;
typedef unsigned char __u8;
typedef __u32 __be32;
typedef u16 __be16;

struct iphdr {
	__u8	ihl:4,
		version:4;
	__u8	tos;
	__be16	tot_len;
	__be16	id;
	__be16	frag_off;
	__u8	ttl;
	__u8	protocol;
	__sum16	check;
	__be32	saddr;
	__be32	daddr;
	/*The options start here. */
};

#define ntohl(x) __swab32((__u32)(__be32)(x))
#define IP_DF 0x4000 /* Flag: "Don't Fragment" */

static inline __attribute__((const)) __u32 __swab32(__u32 x)

{
	__asm__ ("rev %0, %1" : "=r" (x) : "r" (x));
	return x;
}

int main(void * buffer, unsigned int *p_id)

{
	unsigned int id;
	int flush = 1;
	const struct iphdr *iph = buffer;
	__u32 len = *p_id;
	
	id = ntohl(*(__be32 *)&iph->id);
	flush = (u16)((ntohl(*(__be32 *)iph) ^ len) | (id ^ IP_DF));
	id >>= 16;
	
	*p_id = id;
	return flush;
}

Mans Rullgard

2012-10-05 22:37 UTC #18 lore raw

Quoted 35 lines · Rob Herring

Here's a testcase. Compiled on ubuntu precise with "arm-linux-gnueabihf-gcc -O2 -marm -march=armv7-a test.c".
typedef unsigned short u16;
typedef unsigned short __sum16;
typedef unsigned int __u32;
typedef unsigned char __u8;
typedef __u32 __be32;
typedef u16 __be16;
struct iphdr {
        __u8    ihl:4,
                version:4;
        __u8    tos;
        __be16  tot_len;
        __be16  id;
        __be16  frag_off;
        __u8    ttl;
        __u8    protocol;
        __sum16 check;
        __be32  saddr;
        __be32  daddr;
        /*The options start here. */
};
#define ntohl(x) __swab32((__u32)(__be32)(x))
#define IP_DF 0x4000 /* Flag: "Don't Fragment" */
static inline __attribute__((const)) __u32 __swab32(__u32 x)
{
        __asm__ ("rev %0, %1" : "=r" (x) : "r" (x));
        return x;
}
int main(void * buffer, unsigned int *p_id)

Quoted 13 lines

{
        unsigned int id;
        int flush = 1;
        const struct iphdr *iph = buffer;
        __u32 len = *p_id;

        id = ntohl(*(__be32 *)&iph->id);
        flush = (u16)((ntohl(*(__be32 *)iph) ^ len) | (id ^ IP_DF));
        id >>= 16;

        *p_id = id;
        return flush;
}

The problem is the (__be32 *) casts. This is a normal pointer to a 32-bit, which is assumed to be aligned, and the cast overrides the packed attribute from the struct. Dereferencing these cast expressions must be done with the macros from asm/unaligned.h

Signature · 1 line

Mans Rullgard / mru

Russell King - ARM Linux

2012-10-05 22:42 UTC #19 lore raw

Quoted 4 lines · Mans Rullgard

The problem is the (__be32 *) casts. This is a normal pointer to a 32-bit, which is assumed to be aligned, and the cast overrides the packed attribute from the struct. Dereferencing these cast expressions must be done with the macros from asm/unaligned.h

Again, not going to happen. DaveM is on record for saying as much, but I guess you're going to reject that as well, so I'm not sure why I'm even bothering to reply.

Nicolas Pitre

2012-10-06 01:41 UTC #20 lore raw

Quoted 9 lines · Russell King - ARM Linux

Quoted 4 lines · Mans Rullgard
The problem is the (__be32 *) casts. This is a normal pointer to a 32-bit, which is assumed to be aligned, and the cast overrides the packed attribute from the struct. Dereferencing these cast expressions must be done with the macros from asm/unaligned.h
Again, not going to happen. DaveM is on record for saying as much, but I guess you're going to reject that as well, so I'm not sure why I'm even bothering to reply.

May I suggest to send a recap of this thread to netdev at vger.kernel.org and see what davem has to say this time? Otherwise we are just going in circle.

Nicolas

Mans Rullgard

2012-10-06 16:04 UTC #21 lore raw

In reply to Mans Rullgard

Quoted 7 lines · Russell King - ARM Linux

Quoted 4 lines · Mans Rullgard
The problem is the (__be32 *) casts. This is a normal pointer to a 32-bit, which is assumed to be aligned, and the cast overrides the packed attribute from the struct. Dereferencing these cast expressions must be done with the macros from asm/unaligned.h
Again, not going to happen.

There are only two options for fixing this:

1. Ensure the struct is always aligned.
2. Declare it packed (and fix casts).

Refusing to do either leaves us with a broken kernel. Is that what you want?

Signature · 1 line

Mans Rullgard / mru

Nicolas Pitre

2012-10-06 16:19 UTC #22 lore raw

Quoted 8 lines · Mans Rullgard

Quoted 6 lines · Russell King - ARM Linux
Quoted 4 lines · Mans Rullgard
The problem is the (__be32 *) casts. This is a normal pointer to a 32-bit, which is assumed to be aligned, and the cast overrides the packed attribute from the struct. Dereferencing these cast expressions must be done with the macros from asm/unaligned.h

Quoted 8 lines

Quoted 1 line
Again, not going to happen.
There are only two options for fixing this:
1. Ensure the struct is always aligned.
2. Declare it packed (and fix casts).
Refusing to do either leaves us with a broken kernel. Is that what you want?

Once again, please bring this up with davem and also CC netdev at vger.kernel.org as he's the point of contention here.

Nicolas

Russell King - ARM Linux

2012-10-06 16:31 UTC #23 lore raw

Quoted 8 lines · Mans Rullgard

Quoted 6 lines · Russell King - ARM Linux
Quoted 4 lines · Mans Rullgard
The problem is the (__be32 *) casts. This is a normal pointer to a 32-bit, which is assumed to be aligned, and the cast overrides the packed attribute from the struct. Dereferencing these cast expressions must be done with the macros from asm/unaligned.h

Quoted 8 lines

Quoted 1 line
Again, not going to happen.
There are only two options for fixing this:
1. Ensure the struct is always aligned.
2. Declare it packed (and fix casts).
Refusing to do either leaves us with a broken kernel. Is that what you want?

How about you start reading the emails that you receive rather than seemingly insisting that I somehow fix the Linux networking stack - which would involve me talking to someone who has publically tried to oust me from being ARM maintainer, and whom would probably reject any attempt to fix this _because_ I was the person involved in proposing the fix?

It's far better that someone else sorts this out, they are much more likely to get a favourable response.

Mikael Pettersson

2012-10-06 10:58 UTC #24 lore raw

Quoted 81 lines · Rob Herring

Quoted 40 lines · Mikael Pettersson
Quoted 33 lines · Rob Herring
Quoted 26 lines · Russell King - ARM Linux
Quoted 25 lines · Mans Rullgard
Quoted 22 lines · Russell King - ARM Linux
Quoted 9 lines · Mans Rullgard
Quoted 7 lines · Rob Herring
This struct is the IP header, so a struct ptr is just set to the beginning of the received data. Since ethernet headers are 14 bytes, often the IP header is not aligned unless the NIC can place the frame at a 2 byte offset (which is something I need to investigate). So this function cannot make any assumptions about the alignment. Does the ABI define structs have some minimum alignment? Does the struct need to be declared as packed or something?
Quoted 4 lines
The ABI defines the alignment of structs as the maximum alignment of its members. Since this struct contains 32-bit members, the alignment for the whole struct becomes 32 bits as well. Declaring it as packed tells gcc it might be unaligned (in addition to removing any holes within).
This has come up before in the past.
The Linux network folk will _not_ allow - in any shape or form - for this struct to be marked packed (it's the struct which needs to be marked packed) because by doing so, it causes GCC to issue byte loads/ stores on architectures where there isn't a problem, and that decreases the performance of the Linux IP stack unnecessarily.
Quoted 5 lines
Quoted 1 line
Which architectures? I have never seen anything like that.
Does it matter? I'm just relaying the argument against adding __packed which was used before we were forced (by the networking folk) to implement the alignment fault handler.
Quoted 5 lines
It doesn't really matter what will be accepted or not as adding __packed to struct iphdr doesn't fix the problem anyway. gcc still emits a ldm. The only way I've found to eliminate the alignment fault is adding a barrier between the 2 loads. That seems like a compiler issue to me if there is not a better fix.
Quoted 4 lines
If you suspect a GCC bug, please prepare a standalone user-space test case and submit it to GCC's bugzilla (I can do the latter if you absolutely do not want to). It wouldn't be the first alignment-related GCC bug...
Here's a testcase. Compiled on ubuntu precise with "arm-linux-gnueabihf-gcc -O2 -marm -march=armv7-a test.c".
typedef unsigned short u16;
typedef unsigned short __sum16;
typedef unsigned int __u32;
typedef unsigned char __u8;
typedef __u32 __be32;
typedef u16 __be16;
struct iphdr {
	__u8	ihl:4,
		version:4;
	__u8	tos;
	__be16	tot_len;
	__be16	id;
	__be16	frag_off;
	__u8	ttl;
	__u8	protocol;
	__sum16	check;
	__be32	saddr;
	__be32	daddr;
	/*The options start here. */
};
#define ntohl(x) __swab32((__u32)(__be32)(x))
#define IP_DF 0x4000 /* Flag: "Don't Fragment" */
static inline __attribute__((const)) __u32 __swab32(__u32 x)
{
	__asm__ ("rev %0, %1" : "=r" (x) : "r" (x));
	return x;
}
int main(void * buffer, unsigned int *p_id)

Quoted 14 lines

{
	unsigned int id;
	int flush = 1;
	const struct iphdr *iph = buffer;
	__u32 len = *p_id;
	
	id = ntohl(*(__be32 *)&iph->id);
	flush = (u16)((ntohl(*(__be32 *)iph) ^ len) | (id ^ IP_DF));
	id >>= 16;
	
	*p_id = id;
	return flush;
}

I was referring to your statement that adding __packed to the types involved didn't prevent GCC from emitting aligned memory accesses. The test case above only shows that if the source code lies to GCC then things break...

Scott Bambrough

2012-10-09 14:05 UTC #25 lore raw

Quoted 56 lines · Rob Herring

Quoted 40 lines · Mikael Pettersson
Quoted 33 lines · Rob Herring
Quoted 26 lines · Russell King - ARM Linux
Quoted 25 lines · Mans Rullgard
Quoted 22 lines · Russell King - ARM Linux
Quoted 9 lines · Mans Rullgard
Quoted 7 lines · Rob Herring
This struct is the IP header, so a struct ptr is just set to the beginning of the received data. Since ethernet headers are 14 bytes, often the IP header is not aligned unless the NIC can place the frame at a 2 byte offset (which is something I need to investigate). So this function cannot make any assumptions about the alignment. Does the ABI define structs have some minimum alignment? Does the struct need to be declared as packed or something?
Quoted 4 lines
The ABI defines the alignment of structs as the maximum alignment of its members. Since this struct contains 32-bit members, the alignment for the whole struct becomes 32 bits as well. Declaring it as packed tells gcc it might be unaligned (in addition to removing any holes within).
This has come up before in the past.
The Linux network folk will _not_ allow - in any shape or form - for this struct to be marked packed (it's the struct which needs to be marked packed) because by doing so, it causes GCC to issue byte loads/ stores on architectures where there isn't a problem, and that decreases the performance of the Linux IP stack unnecessarily.
Quoted 5 lines
Quoted 1 line
Which architectures? I have never seen anything like that.
Does it matter? I'm just relaying the argument against adding __packed which was used before we were forced (by the networking folk) to implement the alignment fault handler.
Quoted 5 lines
It doesn't really matter what will be accepted or not as adding __packed to struct iphdr doesn't fix the problem anyway. gcc still emits a ldm. The only way I've found to eliminate the alignment fault is adding a barrier between the 2 loads. That seems like a compiler issue to me if there is not a better fix.
Quoted 4 lines
If you suspect a GCC bug, please prepare a standalone user-space test case and submit it to GCC's bugzilla (I can do the latter if you absolutely do not want to). It wouldn't be the first alignment-related GCC bug...
Here's a testcase. Compiled on ubuntu precise with "arm-linux-gnueabihf-gcc -O2 -marm -march=armv7-a test.c".
typedef unsigned short u16;
typedef unsigned short __sum16;
typedef unsigned int __u32;
typedef unsigned char __u8;
typedef __u32 __be32;
typedef u16 __be16;

Quoted 14 lines

struct iphdr {
	__u8	ihl:4,
		version:4;
	__u8	tos;
	__be16	tot_len;
	__be16	id;
	__be16	frag_off;
	__u8	ttl;
	__u8	protocol;
	__sum16	check;
	__be32	saddr;
	__be32	daddr;
	/*The options start here. */
};

I was reading this thread with some interest. AFAIK, with the default alignment rules the above struct is packed; there will be no holes in it.

Quoted 10 lines

#define ntohl(x) __swab32((__u32)(__be32)(x))
#define IP_DF 0x4000 /* Flag: "Don't Fragment" */
static inline __attribute__((const)) __u32 __swab32(__u32 x)
{
	__asm__ ("rev %0, %1" : "=r" (x) : "r" (x));
	return x;
}

Quoted 8 lines

int main(void * buffer, unsigned int *p_id)

{
	unsigned int id;
	int flush = 1;
	const struct iphdr *iph = buffer;
	__u32 len = *p_id;
	
	id = ntohl(*(__be32 *)&iph->id);

The above statement is the problem. I think it is poorly written networking code. It takes the address of a 16 bit quantity (aligned on a halfword address), attempts to do a type conversion using pointers, then dereference it. I would have thought:

id = ntohs(iph->id);

would have been enough.

Scott

Signature · 7 lines

Scott Bambrough
Technical Director, Member Services
Linaro Ltd.
email: scott.bambrough at linaro.org
irc: scottb (freenode, irc.linaro.org)
web: http://www.linaro.org
Linaro: The future of Linux on ARM.

Mans Rullgard

2012-10-09 14:18 UTC #26 lore raw

Quoted 27 lines · Scott Bambrough

Quoted 10 lines · Rob Herring
Here's a testcase. Compiled on ubuntu precise with "arm-linux-gnueabihf-gcc -O2 -marm -march=armv7-a test.c".
typedef unsigned short u16;
typedef unsigned short __sum16;
typedef unsigned int __u32;
typedef unsigned char __u8;
typedef __u32 __be32;
typedef u16 __be16;
Quoted 14 lines
struct iphdr {
        __u8    ihl:4,
                version:4;
        __u8    tos;
        __be16  tot_len;
        __be16  id;
        __be16  frag_off;
        __u8    ttl;
        __u8    protocol;
        __sum16 check;
        __be32  saddr;
        __be32  daddr;
        /*The options start here. */
};

Quoted 2 lines

I was reading this thread with some interest. AFAIK, with the default alignment rules the above struct is packed; there will be no holes in it.

Correct. The problem here is that something is passing around a pointer to such a struct which is not 4-byte aligned as required by the ABI rules.

Quoted 16 lines

Quoted 16 lines
#define ntohl(x) __swab32((__u32)(__be32)(x))
#define IP_DF 0x4000 /* Flag: "Don't Fragment" */
static inline __attribute__((const)) __u32 __swab32(__u32 x)
{
        __asm__ ("rev %0, %1" : "=r" (x) : "r" (x));
        return x;
}
int main(void * buffer, unsigned int *p_id)
{
        unsigned int id;
        int flush = 1;
        const struct iphdr *iph = buffer;
        __u32 len = *p_id;

Quoted 6 lines

Quoted 1 line
        id = ntohl(*(__be32 *)&iph->id);
The above statement is the problem. I think it is poorly written networking code. It takes the address of a 16 bit quantity (aligned on a halfword address),

The 'id' member starts 4 bytes into the struct, so if the struct is properly aligned, there will be no fault here. The problem is that networking code does not always align these structs correctly.

Quoted 6 lines

attempts to do a type conversion using pointers, then dereference it. I would have thought:
id = ntohs(iph->id);
would have been enough.

I'm assuming the is intentionally merging two 16-bit fields into one 32-bit value. What you suggest would do something rather different.

Signature · 1 line

Mans Rullgard / mru

Russell King - ARM Linux

2012-10-05 14:05 UTC #27 lore raw

In reply to Måns Rullgård

Quoted 10 lines · Rob Herring

Quoted 3 lines · Russell King - ARM Linux
Does it matter? I'm just relaying the argument against adding __packed which was used before we were forced (by the networking folk) to implement the alignment fault handler.
It doesn't really matter what will be accepted or not as adding __packed to struct iphdr doesn't fix the problem anyway. gcc still emits a ldm. The only way I've found to eliminate the alignment fault is adding a barrier between the 2 loads. That seems like a compiler issue to me if there is not a better fix.

Even so, please test the patch I've sent you in the sub-thread - that needs testing whether or not GCC is at fault. Will's patch to add the warnings _has_ uncovered a potential issue with the use of __get_user() in some parts of the ARM specific kernel, and I really need you to test that while you're experiencing this problem.

Rob Herring

author

2012-10-05 14:33 UTC #28 lore raw

Quoted 12 lines · Russell King - ARM Linux

Quoted 10 lines · Rob Herring
Quoted 3 lines · Russell King - ARM Linux
Does it matter? I'm just relaying the argument against adding __packed which was used before we were forced (by the networking folk) to implement the alignment fault handler.
It doesn't really matter what will be accepted or not as adding __packed to struct iphdr doesn't fix the problem anyway. gcc still emits a ldm. The only way I've found to eliminate the alignment fault is adding a barrier between the 2 loads. That seems like a compiler issue to me if there is not a better fix.

Quoted 5 lines

Even so, please test the patch I've sent you in the sub-thread - that needs testing whether or not GCC is at fault. Will's patch to add the warnings _has_ uncovered a potential issue with the use of __get_user() in some parts of the ARM specific kernel, and I really need you to test that while you're experiencing this problem.

I've tested your patch and it appears to fix things. Thanks!

Now on to getting rid of faults on practically every single received IP packet:

Multi: 9871002

RX packets:9872010 errors:0 dropped:0 overruns:0 frame:0

Rob

Jon Masters

2012-10-11 00:59 UTC #29 lore raw

Hi everyone,

Quoted 13 lines · Rob Herring

Quoted 12 lines · Russell King - ARM Linux
Quoted 10 lines · Rob Herring
Quoted 3 lines · Russell King - ARM Linux
Does it matter? I'm just relaying the argument against adding __packed which was used before we were forced (by the networking folk) to implement the alignment fault handler.
It doesn't really matter what will be accepted or not as adding __packed to struct iphdr doesn't fix the problem anyway. gcc still emits a ldm. The only way I've found to eliminate the alignment fault is adding a barrier between the 2 loads. That seems like a compiler issue to me if there is not a better fix.

Quoted 7 lines

Quoted 5 lines
Even so, please test the patch I've sent you in the sub-thread - that needs testing whether or not GCC is at fault. Will's patch to add the warnings _has_ uncovered a potential issue with the use of __get_user() in some parts of the ARM specific kernel, and I really need you to test that while you're experiencing this problem.
I've tested your patch and it appears to fix things. Thanks!

Ok. I'm looking for a short term solution in the Fedora kernel because we've been bitten by this bug (I've been following this thread). I considered just reverting Will's patch, but that only sweeps the issue under the rug back to where we were in our 3.4 kernel. So, should we pull in rmk's fix? It seems there are two problems here:

1). Missaligned access fault handling atomicity in general 2). Assumptions around struct alignment that turn out not to be true at runtime due to the way that the structs are actually then aligned.

Quoted 6 lines

Now on to getting rid of faults on practically every single received IP packet:
Multi: 9871002
RX packets:9872010 errors:0 dropped:0 overruns:0 frame:0

This will still be a problem, indeed. At least we can be aware we're taking a large number of faults and hope for a netdev solution.

Jon.

Måns Rullgård

2012-10-11 02:27 UTC #30 lore raw

Quoted 15 lines · Jon Masters

Hi everyone,
Quoted 12 lines · Rob Herring
Quoted 11 lines · Russell King - ARM Linux
Quoted 10 lines · Rob Herring
Quoted 3 lines · Russell King - ARM Linux
Does it matter? I'm just relaying the argument against adding __packed which was used before we were forced (by the networking folk) to implement the alignment fault handler.
It doesn't really matter what will be accepted or not as adding __packed to struct iphdr doesn't fix the problem anyway. gcc still emits a ldm. The only way I've found to eliminate the alignment fault is adding a barrier between the 2 loads. That seems like a compiler issue to me if there is not a better fix.

This turns out to be caused by pointers being typecast to normal (aligned) types.

Quoted 7 lines

Quoted 7 lines
Quoted 5 lines
Even so, please test the patch I've sent you in the sub-thread - that needs testing whether or not GCC is at fault. Will's patch to add the warnings _has_ uncovered a potential issue with the use of __get_user() in some parts of the ARM specific kernel, and I really need you to test that while you're experiencing this problem.
I've tested your patch and it appears to fix things. Thanks!

[...]

Quoted 9 lines

Quoted 6 lines
Now on to getting rid of faults on practically every single received IP packet:
Multi: 9871002
RX packets:9872010 errors:0 dropped:0 overruns:0 frame:0
This will still be a problem, indeed. At least we can be aware we're taking a large number of faults and hope for a netdev solution.

There are exactly two possible solutions:

1. Change the networking code so those structs are always aligned. This might not be (easily) possible.
2. Mark the structs __packed and fix any typecasts like the ones seen in this thread. This will have an adverse effect in cases where the structs are in fact aligned.

Both solutions lie squarely in the networking code. It's time to involve that list, or we'll never get anywhere.

Signature · 2 lines

M?ns Rullg?rd
mans at mansr.com

Jon Masters

2012-10-11 02:34 UTC #31 lore raw

Quoted 10 lines · M?ns Rullg?rd

There are exactly two possible solutions:
1. Change the networking code so those structs are always aligned. This might not be (easily) possible.
2. Mark the structs __packed and fix any typecasts like the ones seen in this thread. This will have an adverse effect in cases where the structs are in fact aligned.
Both solutions lie squarely in the networking code. It's time to involve that list, or we'll never get anywhere.

Sure, please do let's figure out the plan. But my question is tangential
- I am after input from rmk on whether that patch he posted to fix the atomicity of missaligned faults is going to be something we should plan to be pulling into 3.6 for Fedora (after there's an official version) to correct the might_fault warnings, etc.

Meanwhile, a separate fix of some kind is still likely to be needed because we don't want to take a large number of alignment traps.

Jon.

David Laight

2012-10-11 08:21 UTC #32 lore raw

Quoted 23 lines

-----Original Message-----
From: netdev-owner at vger.kernel.org [mailto:netdev-owner at vger.kernel.org] On Behalf Of M?ns Rullg?rd
Sent: 11 October 2012 03:27
To: Jon Masters
Cc: linux-arm-kernel at lists.infradead.org; netdev at vger.kernel.org
Subject: Re: alignment faults in 3.6
Quoted 12 lines · Jon Masters
Hi everyone,
Quoted 9 lines · Rob Herring
Quoted 8 lines · Russell King - ARM Linux
Quoted 7 lines · Rob Herring
Quoted 3 lines · Russell King - ARM Linux
Does it matter? I'm just relaying the argument against adding __packed which was used before we were forced (by the networking folk) to implement the alignment fault handler.
It doesn't really matter what will be accepted or not as adding __packed to struct iphdr doesn't fix the problem anyway.

...

Quoted 10 lines

There are exactly two possible solutions:
1. Change the networking code so those structs are always aligned. This might not be (easily) possible.
2. Mark the structs __packed and fix any typecasts like the ones seen in this thread. This will have an adverse effect in cases where the structs are in fact aligned.
Both solutions lie squarely in the networking code. It's time to involve that list, or we'll never get anywhere.

It might be enough to use __attribute__((aligned(2))) on some structure members (actually does 'ldm' need 8 byte alignment?? - in which case aligned(4) is enough).

One on my bugbears is hardware that will once receive ethernet frames onto a 4n boundary - it needs to be 4n+2. two bytes of 'junk' will do.

David

Russell King - ARM Linux

2012-10-11 08:53 UTC #33 lore raw

Quoted 3 lines · David Laight

It might be enough to use __attribute__((aligned(2))) on some structure members (actually does 'ldm' need 8 byte alignment?? - in which case aligned(4) is enough).

No, ldm just needs 4 byte alignment, the same as normal word loads/stores not to fault. The only instructions which needs 8 byte alignment not to fault are the double-word load/stores.

Quoted 2 lines

One on my bugbears is hardware that will once receive ethernet frames onto a 4n boundary - it needs to be 4n+2. two bytes of 'junk' will do.

Indeed, but remember that is just a mere optimisation for IPv4. What may be true of IPv4 is not necessarily true of other protocols, though IPv4 is currently the dominant protocol today. IPv6 follows the same alignment rules as IPv4, so it's unlikely to change.

Måns Rullgård

2012-10-11 09:45 UTC #34 lore raw

In reply to David Laight

Quoted 30 lines · David Laight

Quoted 21 lines
-----Original Message-----
From: netdev-owner at vger.kernel.org [mailto:netdev-owner at vger.kernel.org] On Behalf Of M?ns Rullg?rd
Sent: 11 October 2012 03:27
To: Jon Masters
Cc: linux-arm-kernel at lists.infradead.org; netdev at vger.kernel.org
Subject: Re: alignment faults in 3.6
Quoted 12 lines · Jon Masters
Hi everyone,
Quoted 9 lines · Rob Herring
Quoted 8 lines · Russell King - ARM Linux
Quoted 7 lines · Rob Herring
Quoted 3 lines · Russell King - ARM Linux
Does it matter? I'm just relaying the argument against adding __packed which was used before we were forced (by the networking folk) to implement the alignment fault handler.
It doesn't really matter what will be accepted or not as adding __packed to struct iphdr doesn't fix the problem anyway.
...
Quoted 8 lines
There are exactly two possible solutions:
1. Change the networking code so those structs are always aligned. This might not be (easily) possible.
2. Mark the structs __packed and fix any typecasts like the ones seen in this thread. This will have an adverse effect in cases where the structs are in fact aligned.

Quoted 6 lines

Quoted 2 lines
Both solutions lie squarely in the networking code. It's time to involve that list, or we'll never get anywhere.
It might be enough to use __attribute__((aligned(2))) on some structure members (actually does 'ldm' need 8 byte alignment?? - in which case aligned(4) is enough).

The aligned attribute can only increase alignment.

Signature · 2 lines

M?ns Rullg?rd
mans at mansr.com

Eric Dumazet

2012-10-11 10:00 UTC #35 lore raw

Quoted 32 lines · M?ns Rullg?rd

Quoted 30 lines · David Laight
Quoted 21 lines
-----Original Message-----
From: netdev-owner at vger.kernel.org [mailto:netdev-owner at vger.kernel.org] On Behalf Of M?ns Rullg?rd
Sent: 11 October 2012 03:27
To: Jon Masters
Cc: linux-arm-kernel at lists.infradead.org; netdev at vger.kernel.org
Subject: Re: alignment faults in 3.6
Quoted 12 lines · Jon Masters
Hi everyone,
Quoted 9 lines · Rob Herring
Quoted 8 lines · Russell King - ARM Linux
Quoted 7 lines · Rob Herring
Quoted 3 lines · Russell King - ARM Linux
Does it matter? I'm just relaying the argument against adding __packed which was used before we were forced (by the networking folk) to implement the alignment fault handler.
It doesn't really matter what will be accepted or not as adding __packed to struct iphdr doesn't fix the problem anyway.
...
Quoted 8 lines
There are exactly two possible solutions:
1. Change the networking code so those structs are always aligned. This might not be (easily) possible.
2. Mark the structs __packed and fix any typecasts like the ones seen in this thread. This will have an adverse effect in cases where the structs are in fact aligned.

Quoted 9 lines

Quoted 6 lines
Quoted 2 lines
Both solutions lie squarely in the networking code. It's time to involve that list, or we'll never get anywhere.
It might be enough to use __attribute__((aligned(2))) on some structure members (actually does 'ldm' need 8 byte alignment?? - in which case aligned(4) is enough).
The aligned attribute can only increase alignment.

I have no idea what is the problem,

-ENOTENOUGHCONTEXT

Måns Rullgård

2012-10-11 10:20 UTC #36 lore raw

Quoted 40 lines · Eric Dumazet

Quoted 39 lines · M?ns Rullg?rd
Quoted 36 lines · David Laight
Quoted 21 lines
-----Original Message-----
From: netdev-owner at vger.kernel.org [mailto:netdev-owner at vger.kernel.org] On Behalf Of M?ns Rullg?rd
Sent: 11 October 2012 03:27
To: Jon Masters
Cc: linux-arm-kernel at lists.infradead.org; netdev at vger.kernel.org
Subject: Re: alignment faults in 3.6
Quoted 12 lines · Jon Masters
Hi everyone,
Quoted 9 lines · Rob Herring
Quoted 8 lines · Russell King - ARM Linux
Quoted 7 lines · Rob Herring
Quoted 3 lines · Russell King - ARM Linux
Does it matter? I'm just relaying the argument against adding __packed which was used before we were forced (by the networking folk) to implement the alignment fault handler.
It doesn't really matter what will be accepted or not as adding __packed to struct iphdr doesn't fix the problem anyway.
...
Quoted 10 lines
There are exactly two possible solutions:
1. Change the networking code so those structs are always aligned. This might not be (easily) possible.
2. Mark the structs __packed and fix any typecasts like the ones seen in this thread. This will have an adverse effect in cases where the structs are in fact aligned.
Both solutions lie squarely in the networking code. It's time to involve that list, or we'll never get anywhere.
It might be enough to use __attribute__((aligned(2))) on some structure members (actually does 'ldm' need 8 byte alignment?? - in which case aligned(4) is enough).

Quoted 5 lines

Quoted 1 line
The aligned attribute can only increase alignment.
I have no idea what is the problem,
-ENOTENOUGHCONTEXT

The thread starts here:
http://marc.info/?l=linux-arm-kernel&m=134939228120020

Summary: a pointer to "struct iphdr" is not 4-byte aligned as required by the ARM ABI rules, and this causes traps to the unaligned access fault handler. A recent change makes the kernel print "scheduling while atomic" warnings on some of these traps, which may or may not be benign. Either way, this is bad for performance and should be fixed one way or another.

Signature · 2 lines

M?ns Rullg?rd
mans at mansr.com

Eric Dumazet

2012-10-11 10:22 UTC #37 lore raw

Quoted 8 lines · Eric Dumazet

On Thu, 2012-10-11 at 10:45 +0100, M?ns Rullg?rd wrote:
I have no idea what is the problem,
-ENOTENOUGHCONTEXT

I took a look, and I dont see why/how gcc could use a ldm instruction

Doing so assumed the alignment of the structure was 8 bytes, but its not.

Networking stack mandates that IP headers are aligned on 4 bytes boundaries, not 8 bytes.

(Some arches like x86 dont care, so we might have some bugs in some drivers, but not in the GRO code)

Some drivers are not aware of the NET_IP_ALIGN stuff. They should be fixed, or else you have alignment faults.

struct iphdr {
__u8 ihl:4,

		version:4;
	__u8	tos;
	__be16	tot_len;
	__be16	id;
	__be16	frag_off;
	__u8	ttl;
	__u8	protocol;
	__sum16	check;
	__be32	saddr;
	__be32	daddr;
	/*The options start here. */
};

The alignment of iphdr is 4, not 8

doing id = ntohl(*(__be32 *)&iph->id); is valid

doing flush = (u16)((ntohl(*(__be32 *)iph) ^ len) | (id ^ IP_DF)); is valid as well.

If arm compiler decided to use a 8bytes load, thats a compiler bug.

(unless compiler was specifically told that alignment did not matter)

c02ac020:       e8920840        ldm     r2, {r6, fp}    // HERE
c02ac024:       e6bfbf3b        rev     fp, fp
c02ac028:       e6bf6f36        rev     r6, r6
c02ac02c:       e22bc901        eor     ip, fp, #16384  ; 0x4000
c02ac030:       e0266008        eor     r6, r6, r8
c02ac034:       e18c6006        orr     r6, ip, r6

Russell King - ARM Linux

2012-10-11 10:32 UTC #38 lore raw

Quoted 7 lines · Eric Dumazet

I took a look, and I dont see why/how gcc could use a ldm instruction
Doing so assumed the alignment of the structure was 8 bytes, but its not.
Networking stack mandates that IP headers are aligned on 4 bytes boundaries, not 8 bytes.

Err, no. ldm is "load multiple" not "load double". It loads multiple 32-bit registers, and its requirement for non-faulting behaviour is for the pointer to be 4 byte aligned. However, "load double" requires 8 byte alignment.

Eric Dumazet

2012-10-11 10:49 UTC #39 lore raw

Quoted 6 lines · Russell King - ARM Linux

Quoted 5 lines · Eric Dumazet
I took a look, and I dont see why/how gcc could use a ldm instruction
Doing so assumed the alignment of the structure was 8 bytes, but its not.

Quoted 7 lines

Quoted 2 lines
Networking stack mandates that IP headers are aligned on 4 bytes boundaries, not 8 bytes.
Err, no. ldm is "load multiple" not "load double". It loads multiple 32-bit registers, and its requirement for non-faulting behaviour is for the pointer to be 4 byte aligned. However, "load double" requires 8 byte alignment.

So if you have an alignment fault, thats because IP header is not aligned on 4 bytes ?

If so a driver is buggy and must be fixed.

Please send us full stack trace

Maxime Bizon

2012-10-11 10:56 UTC #40 lore raw

Quoted 4 lines · Eric Dumazet

So if you have an alignment fault, thats because IP header is not aligned on 4 bytes ?
If so a driver is buggy and must be fixed.

So a driver that does not align the ip header is buggy ?

I always thought it was ok not to do so (with a potential performance penalty).

I have some MIPS hardware that is not able to DMA on anything but 32bits aligned addresses (bcm63xx). I tried once to add a memcpy instead of taking unaligned faults and the result was *much slower* on a ipv4 forwarding test (which is what the hardware is used for).

Signature · 1 line

Maxime

Eric Dumazet

2012-10-11 11:28 UTC #41 lore raw

Quoted 10 lines · Maxime Bizon

Quoted 4 lines · Eric Dumazet
So if you have an alignment fault, thats because IP header is not aligned on 4 bytes ?
If so a driver is buggy and must be fixed.
So a driver that does not align the ip header is buggy ?

exactly.

Quoted 3 lines

I always thought it was ok not to do so (with a potential performance penalty).

Apparently not for some arches.

Quoted 5 lines

I have some MIPS hardware that is not able to DMA on anything but 32bits aligned addresses (bcm63xx). I tried once to add a memcpy instead of taking unaligned faults and the result was *much slower* on a ipv4 forwarding test (which is what the hardware is used for).

You probably are aware that a driver can use :

- a fragment to hold the frame given by the hardware, with whatever alignment is needed by the hardware.

Then allocate an skb with enough room (128 bytes) to pull the headers as needed later.

skb = netdev_alloc_skb_ip_align(dev, 128);

Attach the fragment to the skb, and feed stack with this skb. (pull the ethernet header before calling eth_type_trans()

Most modern drivers exactly use this strategy and get nice performance.

Of course, if hardware doesnt need a strange alignment, we can avoid the fragment and directly use

skb = netdev_alloc_skb_ip_align(dev, 1536);

So all drivers can be fixed if needed.

Maxime Bizon

2012-10-11 11:47 UTC #42 lore raw

Quoted 9 lines · Eric Dumazet

You probably are aware that a driver can use :
- a fragment to hold the frame given by the hardware, with whatever alignment is needed by the hardware.
Then allocate an skb with enough room (128 bytes) to pull the headers as needed later.
skb = netdev_alloc_skb_ip_align(dev, 128);

What happen at tx time, supposing that same hardware cannot do SG ?

Aren't we going to memcpy the data into the head ?

Signature · 1 line

Maxime

Eric Dumazet

2012-10-11 11:54 UTC #43 lore raw

Quoted 7 lines · Maxime Bizon

Quoted 5 lines · Eric Dumazet
You probably are aware that a driver can use :
- a fragment to hold the frame given by the hardware, with whatever alignment is needed by the hardware.

Quoted 9 lines

Quoted 4 lines
Then allocate an skb with enough room (128 bytes) to pull the headers as needed later.
skb = netdev_alloc_skb_ip_align(dev, 128);
What happen at tx time, supposing that same hardware cannot do SG ?
Aren't we going to memcpy the data into the head ?

Of course, if you use a forwarding setup, and the tx driver is not SG capable, performance will be bad (You have to copy the data into a single skb (linearize the skb))

But in 2012, having to use hardware without SG for a router sounds a bad joke (if cpu speed is _also_ too low)

Adding get_unaligned() everywhere in linux network stacks is not an option.

We actually want to be able to read the code and fix the bugs, not only run it on a cheap low end router.

Eric Dumazet

2012-10-11 12:00 UTC #44 lore raw

Quoted 13 lines · Eric Dumazet

Quoted 12 lines · Maxime Bizon
Quoted 9 lines · Eric Dumazet
You probably are aware that a driver can use :
- a fragment to hold the frame given by the hardware, with whatever alignment is needed by the hardware.
Then allocate an skb with enough room (128 bytes) to pull the headers as needed later.
skb = netdev_alloc_skb_ip_align(dev, 128);

Quoted 8 lines

Quoted 4 lines
What happen at tx time, supposing that same hardware cannot do SG ?
Aren't we going to memcpy the data into the head ?
Of course, if you use a forwarding setup, and the tx driver is not SG capable, performance will be bad (You have to copy the data into a single skb (linearize the skb))

By the way, if said driver also has alignments issues, it will probably copy the packet given by the stack. (Think you added an IPIP encapsulation)

It would be better for such driver to still pretend it can do SG, and it does the copy itself, avoiding the prior copies in core stack.

Maxime Bizon

2012-10-11 12:51 UTC #45 lore raw

Quoted 6 lines · Eric Dumazet

Of course, if you use a forwarding setup, and the tx driver is not SG capable, performance will be bad (You have to copy the data into a single skb (linearize the skb))
But in 2012, having to use hardware without SG for a router sounds a bad joke (if cpu speed is _also_ too low)

Hey I cannot go back in time, when that hardware was built in 2004 (mips @250Mhz), it was considered good, and we did manufacture a lot of it, so it's still maintained.

People run recent kernels on older hardware because they are *encouraged to do so*.

I fought inside my company to be good kernel citizen, not using proprietary BSP, rewrite & mainline the drivers, because that was the community promise: mainline it, we will support it for you, you will get the latest kernel features for free.

That worked, but with some drawbacks:

- kernel footprint grew that much (we started from 2.4) that it does
not fit in device flash anymore

- performance took a hit each time we upgrade, mostly because of cache
footprint growth.

- as kernel footprint grew, available RAM for conntrack & route cache
entries was smaller each time

But I had to stop upgrading after 2.6.20. Everything below is not anybody's fault. Bloat is unavoidable for software project that big.

I'm perfectly ok with that, but I don't want to be ridiculed for running mainline kernel on old hardware.

Quoted 5 lines

Adding get_unaligned() everywhere in linux network stacks is not an option.
We actually want to be able to read the code and fix the bugs, not only run it on a cheap low end router.

That was not a request, I just needed a clarification.

Documentation/unaligned-memory-access.txt does not say it's a big no-no, it says you can give unaligned pointers to the networking stack if you arch can do unaligned access (with an "efficiency" notion).

MIPS and ARM have a software handler for this, and performance wise in my case it's better to take the faults, a driver writer may think a benchmark will dictate what to do.

Signature · 1 line

Maxime

Eric Dumazet

2012-10-11 12:59 UTC #46 lore raw

Quoted 25 lines · Maxime Bizon

Hey I cannot go back in time, when that hardware was built in 2004 (mips @250Mhz), it was considered good, and we did manufacture a lot of it, so it's still maintained.
People run recent kernels on older hardware because they are *encouraged to do so*.
I fought inside my company to be good kernel citizen, not using proprietary BSP, rewrite & mainline the drivers, because that was the community promise: mainline it, we will support it for you, you will get the latest kernel features for free.
That worked, but with some drawbacks:
- kernel footprint grew that much (we started from 2.4) that it does
not fit in device flash anymore
- performance took a hit each time we upgrade, mostly because of cache
footprint growth.
- as kernel footprint grew, available RAM for conntrack & route cache
entries was smaller each time

Quoted 5 lines

But I had to stop upgrading after 2.6.20. Everything below is not anybody's fault. Bloat is unavoidable for software project that big.
I'm perfectly ok with that, but I don't want to be ridiculed for running mainline kernel on old hardware.

Hmm, I am sorry if you felt that, it was not my intent.

Quoted 10 lines

Quoted 7 lines
Adding get_unaligned() everywhere in linux network stacks is not an option.
We actually want to be able to read the code and fix the bugs, not only run it on a cheap low end router.
That was not a request, I just needed a clarification.

Quoted 8 lines

Documentation/unaligned-memory-access.txt does not say it's a big no-no, it says you can give unaligned pointers to the networking stack if you arch can do unaligned access (with an "efficiency" notion).
MIPS and ARM have a software handler for this, and performance wise in my case it's better to take the faults, a driver writer may think a benchmark will dictate what to do.

Sure, but all this discussion started because one arch apparently did not like these mis alignments, and some people complained that network guys would not add _needed_ get_unaligned_xxx() wrappers ...

So far, linux is 20 years old, I dont think we are going to add wrappers right now. Machines that could have cared are dying anyway.

Please note we still support NET_IP_ALIGN, even if its 0 on x86.

Arnd Bergmann

2012-10-11 12:28 UTC #47 lore raw

Quoted 15 lines · Eric Dumazet

Quoted 6 lines · Russell King - ARM Linux
Quoted 5 lines · Eric Dumazet
I took a look, and I dont see why/how gcc could use a ldm instruction
Doing so assumed the alignment of the structure was 8 bytes, but its not.
Quoted 7 lines
Quoted 2 lines
Networking stack mandates that IP headers are aligned on 4 bytes boundaries, not 8 bytes.
Err, no. ldm is "load multiple" not "load double". It loads multiple 32-bit registers, and its requirement for non-faulting behaviour is for the pointer to be 4 byte aligned. However, "load double" requires 8 byte alignment.

Quoted 6 lines

So if you have an alignment fault, thats because IP header is not aligned on 4 bytes ?
If so a driver is buggy and must be fixed.
Please send us full stack trace

Rob Herring as the original reporter has dropped off the Cc list, adding him back.

I assume that the calxeda xgmac driver is the culprit then. It uses netdev_alloc_skb() rather than netdev_alloc_skb_ip_align() in xgmac_rx_refill but it is not clear whether it does so intentionally or by accident.

Arnd

Eric Dumazet

2012-10-11 12:40 UTC #48 lore raw

Quoted 8 lines · Arnd Bergmann

Rob Herring as the original reporter has dropped off the Cc list, adding him back.
I assume that the calxeda xgmac driver is the culprit then. It uses netdev_alloc_skb() rather than netdev_alloc_skb_ip_align() in xgmac_rx_refill but it is not clear whether it does so intentionally or by accident.

Thanks Arnd

It seems an accident, since driver doesnt check skb->data alignment at all (this can change with SLAB debug on/off)

It also incorrectly adds 64 bytes to bfsize, there is no need for this.

(or if its needed, a comment would be nice, because on prior kernels, this makes skb->head allocations uses kmalloc-4096 instead of kmalloc-2048 slab cache... With 3.7 its less an issue now we use order-3 pages to deliver fragments for rx skbs

So the following patch should fix the alignment, and makes driver uses half memory than before for stable kernels

drivers/net/ethernet/calxeda/xgmac.c +3 -2

index 16814b3..a895e18 100644
@@ -671,7 +671,8 @@ static void xgmac_rx_refill(struct xgmac_priv *priv)
 		p = priv->dma_rx + entry;
 
 		if (priv->rx_skbuff[entry] == NULL) {
-			skb = netdev_alloc_skb(priv->dev, priv->dma_buf_sz);
+			skb = netdev_alloc_skb_ip_align(priv->dev,
+							priv->dma_buf_sz);
 			if (unlikely(skb == NULL))
 				break;

@@ -703,7 +704,7 @@ static int xgmac_dma_desc_rings_init(struct net_device *dev)
 	/* Set the Buffer size according to the MTU;
 	 * indeed, in case of jumbo we need to bump-up the buffer sizes.
 	 */
-	bfsize = ALIGN(dev->mtu + ETH_HLEN + ETH_FCS_LEN + NET_IP_ALIGN + 64,
+	bfsize = ALIGN(dev->mtu + ETH_HLEN + ETH_FCS_LEN,
 		       64);
 
 	netdev_dbg(priv->dev, "mtu [%d] bfsize [%d]\n", dev->mtu, bfsize);

Rob Herring

author

2012-10-11 13:20 UTC #49 lore raw

Quoted 10 lines · Eric Dumazet

Quoted 8 lines · Arnd Bergmann
Rob Herring as the original reporter has dropped off the Cc list, adding him back.
I assume that the calxeda xgmac driver is the culprit then. It uses netdev_alloc_skb() rather than netdev_alloc_skb_ip_align() in xgmac_rx_refill but it is not clear whether it does so intentionally or by accident.

This in fact does work and eliminates the unaligned traps. However, not all h/w can do IP aligned DMA (i.MX FEC for example), so I still think this is a questionable optimization by the compiler. We're saving 1 load instruction here for data that is likely already in the cache. It may be legal per the ABI, but the downside of this optimization is much greater than the upside.

Quoted 7 lines

Thanks Arnd
It seems an accident, since driver doesnt check skb->data alignment at all (this can change with SLAB debug on/off)
It also incorrectly adds 64 bytes to bfsize, there is no need for this.

I'm pretty sure this was needed as the h/w writes out full bursts of data, but I'll go back and check.

Rob

Quoted 33 lines

So the following patch should fix the alignment, and makes driver uses half memory than before for stable kernels

drivers/net/ethernet/calxeda/xgmac.c +3 -2

index 16814b3..a895e18 100644
@@ -671,7 +671,8 @@ static void xgmac_rx_refill(struct xgmac_priv *priv)
 		p = priv->dma_rx + entry;
 
 		if (priv->rx_skbuff[entry] == NULL) {
-			skb = netdev_alloc_skb(priv->dev, priv->dma_buf_sz);
+			skb = netdev_alloc_skb_ip_align(priv->dev,
+							priv->dma_buf_sz);
 			if (unlikely(skb == NULL))
 				break;

@@ -703,7 +704,7 @@ static int xgmac_dma_desc_rings_init(struct net_device *dev)
 	/* Set the Buffer size according to the MTU;
 	 * indeed, in case of jumbo we need to bump-up the buffer sizes.
 	 */
-	bfsize = ALIGN(dev->mtu + ETH_HLEN + ETH_FCS_LEN + NET_IP_ALIGN + 64,
+	bfsize = ALIGN(dev->mtu + ETH_HLEN + ETH_FCS_LEN,
 		       64);
 
 	netdev_dbg(priv->dev, "mtu [%d] bfsize [%d]\n", dev->mtu, bfsize);

Måns Rullgård

2012-10-11 13:32 UTC #50 lore raw

Quoted 12 lines · Rob Herring

Quoted 10 lines · Eric Dumazet
Quoted 8 lines · Arnd Bergmann
Rob Herring as the original reporter has dropped off the Cc list, adding him back.
I assume that the calxeda xgmac driver is the culprit then. It uses netdev_alloc_skb() rather than netdev_alloc_skb_ip_align() in xgmac_rx_refill but it is not clear whether it does so intentionally or by accident.

Quoted 6 lines

This in fact does work and eliminates the unaligned traps. However, not all h/w can do IP aligned DMA (i.MX FEC for example), so I still think this is a questionable optimization by the compiler. We're saving 1 load instruction here for data that is likely already in the cache. It may be legal per the ABI, but the downside of this optimization is much greater than the upside.

The compiler is working *exactly* as it should. Merging the loads saves cycles *and* code size. Many of these added up can make a real difference.

When writing code, you must follow all the rules, whether you like them or not. Without rules, the compiler would be very limited in the optimisations it could perform.

Unfortunately, new optimisations occasionally uncover broken code violating some constraint or other. When this happens, the correct course of action is to fix the code, not cripple the compiler.

Signature · 2 lines

M?ns Rullg?rd
mans at mansr.com

Arnd Bergmann

2012-10-11 13:35 UTC #51 lore raw

Quoted 13 lines · Rob Herring

Quoted 10 lines · Eric Dumazet
Quoted 8 lines · Arnd Bergmann
Rob Herring as the original reporter has dropped off the Cc list, adding him back.
I assume that the calxeda xgmac driver is the culprit then. It uses netdev_alloc_skb() rather than netdev_alloc_skb_ip_align() in xgmac_rx_refill but it is not clear whether it does so intentionally or by accident.

Quoted 7 lines

This in fact does work and eliminates the unaligned traps. However, not all h/w can do IP aligned DMA (i.MX FEC for example), so I still think this is a questionable optimization by the compiler. We're saving 1 load instruction here for data that is likely already in the cache. It may be legal per the ABI, but the downside of this optimization is much greater than the upside.

What about the other approach that Eric suggested for such hardware in http://www.spinics.net/lists/arm-kernel/msg200206.html ?

Arnd

Eric Dumazet

2012-10-11 13:47 UTC #52 lore raw

Quoted 12 lines · Rob Herring

Quoted 10 lines · Eric Dumazet
Quoted 8 lines · Arnd Bergmann
Rob Herring as the original reporter has dropped off the Cc list, adding him back.
I assume that the calxeda xgmac driver is the culprit then. It uses netdev_alloc_skb() rather than netdev_alloc_skb_ip_align() in xgmac_rx_refill but it is not clear whether it does so intentionally or by accident.

Quoted 6 lines

This in fact does work and eliminates the unaligned traps. However, not all h/w can do IP aligned DMA (i.MX FEC for example), so I still think this is a questionable optimization by the compiler. We're saving 1 load instruction here for data that is likely already in the cache. It may be legal per the ABI, but the downside of this optimization is much greater than the upside.

Compiler is asked to perform a 32bit load, it does it.

There is no questionable optimization here. Really. Please stop pretending this, this makes no sense.

As I said, if some h/w cannot do IP aligned DMA, driver can use a workaround, or a plain memmove() (some drivers seems to do this to work around this h/w limitation, just grep for memmove() in drivers/net)

Quoted 11 lines

Quoted 8 lines
Thanks Arnd
It seems an accident, since driver doesnt check skb->data alignment at all (this can change with SLAB debug on/off)
It also incorrectly adds 64 bytes to bfsize, there is no need for this.
I'm pretty sure this was needed as the h/w writes out full bursts of data, but I'll go back and check.

Maybe the ALIGN() was needed then. But the 64 + NE_IP_ALIGN sounds like the head room that we allocate/reserve in netdev_alloc_skb_ip_align()

So you allocate this extra room twice.

Thanks

Rob Herring

author

2012-10-11 15:23 UTC #53 lore raw

Quoted 13 lines · Eric Dumazet

Quoted 12 lines · Rob Herring
Quoted 10 lines · Eric Dumazet
Quoted 8 lines · Arnd Bergmann
Rob Herring as the original reporter has dropped off the Cc list, adding him back.
I assume that the calxeda xgmac driver is the culprit then. It uses netdev_alloc_skb() rather than netdev_alloc_skb_ip_align() in xgmac_rx_refill but it is not clear whether it does so intentionally or by accident.

Quoted 8 lines

Quoted 6 lines
This in fact does work and eliminates the unaligned traps. However, not all h/w can do IP aligned DMA (i.MX FEC for example), so I still think this is a questionable optimization by the compiler. We're saving 1 load instruction here for data that is likely already in the cache. It may be legal per the ABI, but the downside of this optimization is much greater than the upside.
Compiler is asked to perform a 32bit load, it does it.

Not exactly. It is asked to to perform 2 32-bit loads which are combined into a single ldm (load multiple) which cannot handle unaligned accesses. Here's a simple example that does the same thing:

void test(char * buf)

{
	printf("%d, %d\n", *((unsigned int *)&buf[0]), *((unsigned int *)&buf[4]));
}

So I guess the only ABI legal unaligned access is in a packed struct.

Quoted 2 lines

There is no questionable optimization here. Really. Please stop pretending this, this makes no sense.

I'm not the one calling the networking stack bad code.

I can fix my h/w, so I'll stop caring about this. Others can all get bitten by this new behavior in gcc 4.7.

Rob

Quoted 24 lines

As I said, if some h/w cannot do IP aligned DMA, driver can use a workaround, or a plain memmove() (some drivers seems to do this to work around this h/w limitation, just grep for memmove() in drivers/net)
Quoted 11 lines
Quoted 8 lines
Thanks Arnd
It seems an accident, since driver doesnt check skb->data alignment at all (this can change with SLAB debug on/off)
It also incorrectly adds 64 bytes to bfsize, there is no need for this.
I'm pretty sure this was needed as the h/w writes out full bursts of data, but I'll go back and check.
Maybe the ALIGN() was needed then. But the 64 + NE_IP_ALIGN sounds like the head room that we allocate/reserve in netdev_alloc_skb_ip_align()
So you allocate this extra room twice.
Thanks

David Laight

2012-10-11 15:39 UTC #54 lore raw

Quoted 9 lines

Not exactly. It is asked to to perform 2 32-bit loads which are combined into a single ldm (load multiple) which cannot handle unaligned accesses. Here's a simple example that does the same thing:
void test(char * buf)
{
	printf("%d, %d\n", *((unsigned int *)&buf[0]), *((unsigned int *)&buf[4]));
}

Have you actually looked at what an ARM processor traditionally did with misaligned memory reads?
While useful, it probably wasn't what was intended.

Actually, and IIRC, some very recent ARM cpus will do the 'expected' thing for single-word loads from misaligned addesses. However they almost certainly won't for ldm/stm.

The 'ldm' optimisation for adjacent memory loads is also dubious. On at least some ARMs it is very slow (might only be strongarms).

Quoted 1 line

So I guess the only ABI legal unaligned access is in a packed struct.

Correct. And you mustn't try casting the address, the compiler is allowed to remember where it came from.
(This causes a lot of grief...)

If you are targeting the ARM cpu that can do misaligned transfers, then gcc should generate single instructions for misaligned structure members, and never do the 'ldm' optimisations.

But, the IP header is expected to be aligned.

David

Måns Rullgård

2012-10-11 16:18 UTC #55 lore raw

Quoted 9 lines · David Laight

Quoted 8 lines
Not exactly. It is asked to to perform 2 32-bit loads which are combined into a single ldm (load multiple) which cannot handle unaligned accesses. Here's a simple example that does the same thing:
void test(char * buf)
{
	printf("%d, %d\n", *((unsigned int *)&buf[0]), *((unsigned int *)&buf[4]));
}

Quoted 6 lines

Have you actually looked at what an ARM processor traditionally did with misaligned memory reads?
While useful, it probably wasn't what was intended.
Actually, and IIRC, some very recent ARM cpus will do the 'expected' thing for single-word loads from misaligned addesses.

What various CPUs do with unaligned accesses is not the issue here. The casts in the code above act as a promise to the compiler that the address is in fact properly align for the pointer type.

Quoted 3 lines

However they almost certainly won't for ldm/stm.
The 'ldm' optimisation for adjacent memory loads is also dubious.

There is nothing whatsoever dubious about the compiler using the most efficient instruction sequence to accomplish what the code asks for.

Quoted 1 line

On at least some ARMs it is very slow (might only be strongarms).

The compiler will pick instructions suitable for the CPU you specify.

Quoted 5 lines

Quoted 1 line
So I guess the only ABI legal unaligned access is in a packed struct.
Correct. And you mustn't try casting the address, the compiler is allowed to remember where it came from.
(This causes a lot of grief...)

It is only a problem when you try to outsmart the compiler.

Quoted 3 lines

If you are targeting the ARM cpu that can do misaligned transfers, then gcc should generate single instructions for misaligned structure members, and never do the 'ldm' optimisations.

That is exactly how gcc works.

Quoted 1 line

But, the IP header is expected to be aligned.

Everything tells the compiler the struct is perfectly aligned. When the buggy driver passes a misaligned pointer, bad things happen.

Signature · 2 lines

M?ns Rullg?rd
mans at mansr.com

Arnd Bergmann

2012-10-12 08:11 UTC #56 lore raw

Quoted 4 lines · M?ns Rullg?rd

Quoted 1 line
But, the IP header is expected to be aligned.
Everything tells the compiler the struct is perfectly aligned. When the buggy driver passes a misaligned pointer, bad things happen.

Would it be appropriate to add a WARN_ON_ONCE() in the alignment fault path then? If all alignment faults in the kernel are caused by broken drivers, that would at least give us some hope of finding those drivers while at the same time not causing much overhead in the case where we need to do the fixup in the meantime.

Arnd

Russell King - ARM Linux

2012-10-12 09:03 UTC #57 lore raw

Quoted 11 lines · Arnd Bergmann

Quoted 4 lines · M?ns Rullg?rd
Quoted 1 line
But, the IP header is expected to be aligned.
Everything tells the compiler the struct is perfectly aligned. When the buggy driver passes a misaligned pointer, bad things happen.
Would it be appropriate to add a WARN_ON_ONCE() in the alignment fault path then? If all alignment faults in the kernel are caused by broken drivers, that would at least give us some hope of finding those drivers while at the same time not causing much overhead in the case where we need to do the fixup in the meantime.

No. It is my understanding that various IP option processing can also cause the alignment fault handler to be invoked, even when the packet is properly aligned, and then there's jffs2/mtd which also relies upon alignment faults being fixed up.

Eric Dumazet

2012-10-12 10:04 UTC #58 lore raw

Quoted 4 lines · Russell King - ARM Linux

No. It is my understanding that various IP option processing can also cause the alignment fault handler to be invoked, even when the packet is properly aligned, and then there's jffs2/mtd which also relies upon alignment faults being fixed up.

Oh well.

We normally make sure we dont have alignment faults on arches that dont have CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS (or a non null NET_IP_ALIGN)

So if you find an offender, please report a bug, because I can guarantee you we will _fix_ it.

One example of a fix was the following subtle one.

commit 117632e64d2a5f464e491fe221d7169a3814a77b
tcp: take care of misalignments

We discovered that TCP stack could retransmit misaligned skbs if a malicious peer acknowledged sub MSS frame. This currently can happen only if output interface is non SG enabled : If SG is enabled, tcp builds headless skbs (all payload is included in fragments), so the tcp trimming process only removes parts of skb fragments, header stay aligned.

Some arches cant handle misalignments, so force a head reallocation and shrink headroom to MAX_TCP_HEADER.

Dont care about misaligments on x86 and PPC (or other arches setting NET_IP_ALIGN to 0)

This patch introduces __pskb_copy() which can specify the headroom of new head, and pskb_copy() becomes a wrapper on top of __pskb_copy()

Russell King - ARM Linux

2012-10-12 12:24 UTC #59 lore raw

Quoted 7 lines · Eric Dumazet

Quoted 4 lines · Russell King - ARM Linux
No. It is my understanding that various IP option processing can also cause the alignment fault handler to be invoked, even when the packet is properly aligned, and then there's jffs2/mtd which also relies upon alignment faults being fixed up.

Quoted 7 lines

Oh well.
We normally make sure we dont have alignment faults on arches that dont have CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS (or a non null NET_IP_ALIGN)
So if you find an offender, please report a bug, because I can guarantee you we will _fix_ it.

I think one change I will make to the ARM alignment fixup is to get it to record the last PC where a misaligned kernel fault occurred, and report it via our statistics procfs file. That should allow us to track down where some of these occur.

They aren't anywhere near regular though - looking at the statistics, my firewall seems to do an average of around 2-3 a day, and a web server around 7-8 a day.

Måns Rullgård

2012-10-12 11:00 UTC #60 lore raw

See: http://lists.arm.linux.org.uk/lurker/message/20020225.195925.02bdbd47.en.html
and: http://lists.arm.linux.org.uk/lurker/message/20020313.150932.081a7592.en.html

Quoted 9 lines · Russell King - ARM Linux

Quoted 8 lines · Arnd Bergmann
Quoted 4 lines · M?ns Rullg?rd
Quoted 1 line
But, the IP header is expected to be aligned.
Everything tells the compiler the struct is perfectly aligned. When the buggy driver passes a misaligned pointer, bad things happen.
Would it be appropriate to add a WARN_ON_ONCE() in the alignment fault path then?

I think that's an excellent idea.

Quoted 9 lines

Quoted 4 lines
If all alignment faults in the kernel are caused by broken drivers, that would at least give us some hope of finding those drivers while at the same time not causing much overhead in the case where we need to do the fixup in the meantime.
No. It is my understanding that various IP option processing can also cause the alignment fault handler to be invoked, even when the packet is properly aligned, and then there's jffs2/mtd which also relies upon alignment faults being fixed up.

As far as I'm concerned, this is all hearsay, and I've only ever heard it from you. Why can't you let those who care fix these bugs instead?

Signature · 2 lines

M?ns Rullg?rd
mans at mansr.com

Russell King - ARM Linux

2012-10-12 11:07 UTC #61 lore raw

Quoted 6 lines · M?ns Rullg?rd

Quoted 4 lines · Russell King - ARM Linux
Quoted 3 lines · Arnd Bergmann
Quoted 2 lines · M?ns Rullg?rd
Quoted 1 line
But, the IP header is expected to be aligned.

Quoted 7 lines

Quoted 5 lines
Quoted 5 lines
Quoted 2 lines
Everything tells the compiler the struct is perfectly aligned. When the buggy driver passes a misaligned pointer, bad things happen.
Would it be appropriate to add a WARN_ON_ONCE() in the alignment fault path then?
I think that's an excellent idea.

Well, I get the last word here and it's no.

Quoted 5 lines

Quoted 5 lines
Quoted 4 lines
If all alignment faults in the kernel are caused by broken drivers, that would at least give us some hope of finding those drivers while at the same time not causing much overhead in the case where we need to do the fixup in the meantime.

Quoted 7 lines

Quoted 4 lines
No. It is my understanding that various IP option processing can also cause the alignment fault handler to be invoked, even when the packet is properly aligned, and then there's jffs2/mtd which also relies upon alignment faults being fixed up.
As far as I'm concerned, this is all hearsay, and I've only ever heard it from you. Why can't you let those who care fix these bugs instead?

You know, I'm giving you the benefit of my _knowledge_ which has been built over the course of the last 20 years. I've been in these discussions with networking people before. I ended up having to develop the alignment fault handler because of those discussions. And oh look, Eric confirmed that the networking code isn't going to get "fixed" as you were demanding, which is exactly what I said.

I've been in discussions with MTD people over these issues before, I've discussed this with David Woodhouse when it came up in JFFS2. I *KNOW* these things.

You can call it hearsay if you wish, but it seems to be more accurate than your wild outlandish and pathetic statements.

Måns Rullgård

2012-10-12 11:18 UTC #62 lore raw

Quoted 10 lines · Russell King - ARM Linux

Quoted 9 lines · M?ns Rullg?rd
Quoted 7 lines · Russell King - ARM Linux
Quoted 6 lines · Arnd Bergmann
Quoted 4 lines · M?ns Rullg?rd
Quoted 1 line
But, the IP header is expected to be aligned.
Everything tells the compiler the struct is perfectly aligned. When the buggy driver passes a misaligned pointer, bad things happen.

Quoted 6 lines

Quoted 4 lines
Quoted 2 lines
Quoted 2 lines
Would it be appropriate to add a WARN_ON_ONCE() in the alignment fault path then?
I think that's an excellent idea.
Well, I get the last word here and it's no.

Sadly, yes.

Quoted 10 lines

Quoted 10 lines
Quoted 9 lines
Quoted 4 lines
If all alignment faults in the kernel are caused by broken drivers, that would at least give us some hope of finding those drivers while at the same time not causing much overhead in the case where we need to do the fixup in the meantime.
No. It is my understanding that various IP option processing can also cause the alignment fault handler to be invoked, even when the packet is properly aligned, and then there's jffs2/mtd which also relies upon alignment faults being fixed up.

Quoted 5 lines

Quoted 2 lines
As far as I'm concerned, this is all hearsay, and I've only ever heard it from you. Why can't you let those who care fix these bugs instead?
You know, I'm giving you the benefit of my _knowledge_ which has been built over the course of the last 20 years.

How proud you sound. Now could you say something of substance instead?

Quoted 5 lines

I've been in these discussions with networking people before. I ended up having to develop the alignment fault handler because of those discussions. And oh look, Eric confirmed that the networking code isn't going to get "fixed" as you were demanding, which is exactly what I said.

Funny, I saw him say the exact opposite:

So if you find an offender, please report a bug, because I can guarantee you we will _fix_ it.

Quoted 3 lines

I've been in discussions with MTD people over these issues before, I've discussed this with David Woodhouse when it came up in JFFS2. I *KNOW* these things.

In the same way you "know" the networking people won't fix their code, despite them _clearly_ stating the opposite?

Quoted 2 lines

You can call it hearsay if you wish, but it seems to be more accurate than your wild outlandish and pathetic statements.

So you're resorting to name-calling. Not taking that bait.

Signature · 2 lines

M?ns Rullg?rd
mans at mansr.com

Russell King - ARM Linux

2012-10-12 11:44 UTC #63 lore raw

Quoted 4 lines · M?ns Rullg?rd

Quoted 1 line · Russell King - ARM Linux
Well, I get the last word here and it's no.
Sadly, yes.

It's not "sadly" - it's a matter of fact that the kernel does from time to time generate misaligned accesses and they are _not_ bugs. If they were bugs, then the code to fix up misaligned accesses would not have been developed, and we'd instead take the fault and panic or something like that.

Quoted 10 lines

Quoted 10 lines
Quoted 10 lines
Quoted 9 lines
Quoted 4 lines
If all alignment faults in the kernel are caused by broken drivers, that would at least give us some hope of finding those drivers while at the same time not causing much overhead in the case where we need to do the fixup in the meantime.
No. It is my understanding that various IP option processing can also cause the alignment fault handler to be invoked, even when the packet is properly aligned, and then there's jffs2/mtd which also relies upon alignment faults being fixed up.

Quoted 7 lines

Quoted 5 lines
Quoted 2 lines
As far as I'm concerned, this is all hearsay, and I've only ever heard it from you. Why can't you let those who care fix these bugs instead?
You know, I'm giving you the benefit of my _knowledge_ which has been built over the course of the last 20 years.
How proud you sound. Now could you say something of substance instead?

You're proving yourself to be idiot? There, that's substance.

Quoted 10 lines

Quoted 5 lines
I've been in these discussions with networking people before. I ended up having to develop the alignment fault handler because of those discussions. And oh look, Eric confirmed that the networking code isn't going to get "fixed" as you were demanding, which is exactly what I said.
Funny, I saw him say the exact opposite:
So if you find an offender, please report a bug, because I can guarantee you we will _fix_ it.

No, let's go back.

- You were demanding that the ipv4 header structure should be packed. I said that wasn't going to happen because the networking people wouldn't allow it, and it seems that's been proven correct.
- You were demanding that the ipv4 code used the unaligned accessors. I said that networking people wouldn't allow it either, and that's also been proven correct.

Both these points have been proven correct because Eric has said that the core networking code is _not_ going to be changed to suit this.

What Eric _has_ said is that networking people consider packets supplied to the networking layer where the IPv4 header is not aligned on architectures where misaligned accesses are a problem to be a bug in the network driver, not the network code, and proposed a solution.

That's entirely different from all your claims that the core networking code needs fixing to avoid these misaligned accesses.

Quoted 6 lines

Quoted 3 lines
I've been in discussions with MTD people over these issues before, I've discussed this with David Woodhouse when it came up in JFFS2. I *KNOW* these things.
In the same way you "know" the networking people won't fix their code, despite them _clearly_ stating the opposite?

I'll tell you exactly how I *KNOW* this. The issue came up because of noMMU, which does not have any way to fix up alignment faults. JFFS2 passes randomly aligned buffers to the MTD drivers, and the MTD drivers assume that they're aligned and they do word accesses on them.

See the thread http://lists.arm.linux.org.uk/lurker/thread/20021204.191632.4473796b.en.html

There's several other threads where it's also discussed.

And while you're there, note the date. There is nothing recent about this issue. It's well known, and well understood by those who have a grasp of the issues that alignment faults are a part of normal operation by the ARM kernel - and is one of the penalties of being tied into architecture independent code.

Compromises have to be sought, and that's the compromise we get to live with.

Quoted 4 lines

Quoted 2 lines
You can call it hearsay if you wish, but it seems to be more accurate than your wild outlandish and pathetic statements.
So you're resorting to name-calling. Not taking that bait.

Sorry? So what you're saying is that it's fine for you to call my comments hearsay, but I'm not allowed to express a view on your comments. How arrogant of you.

Eric Dumazet

2012-10-12 12:08 UTC #64 lore raw

Quoted 10 lines · Russell King - ARM Linux

- You were demanding that the ipv4 header structure should be packed. I said that wasn't going to happen because the networking people wouldn't allow it, and it seems that's been proven correct.
- You were demanding that the ipv4 code used the unaligned accessors. I said that networking people wouldn't allow it either, and that's also been proven correct.
Both these points have been proven correct because Eric has said that the core networking code is _not_ going to be changed to suit this.

Quoted 7 lines

What Eric _has_ said is that networking people consider packets supplied to the networking layer where the IPv4 header is not aligned on architectures where misaligned accesses are a problem to be a bug in the network driver, not the network code, and proposed a solution.
That's entirely different from all your claims that the core networking code needs fixing to avoid these misaligned accesses.

It seems we agree then, but your words were a bit misleading (at least for me)

So yes, we built network stack with the prereq that IP headers are aligned, but unfortunately many developers use x86 which doesnt care, so its possible some bugs are added.

But its not by intent, only by accident.

Thanks

Benjamin LaHaise

2012-10-12 14:22 UTC #65 lore raw

Quoted 3 lines · Eric Dumazet

So yes, we built network stack with the prereq that IP headers are aligned, but unfortunately many developers use x86 which doesnt care, so its possible some bugs are added.

x86 does have an alignment check flag that can be set in the flags register. Somehow, I doubt anyone would be willing to walk through all the noise the faults would likely trigger.

-ben

Signature · 1 line

"Thought is the essence of where you are now."

David Laight

2012-10-12 14:36 UTC #66 lore raw

Quoted 3 lines

x86 does have an alignment check flag that can be set in the flags register. Somehow, I doubt anyone would be willing to walk through all the noise the faults would likely trigger.

Someone has tried to set that (in userspace) on NetBSD. The fault reporting has been fixed, but really nothing works since optimised parts of libc deliberately do misaligned transfers.

David

Eric Dumazet

2012-10-12 14:48 UTC #67 lore raw

In reply to Benjamin LaHaise

Quoted 8 lines · Benjamin LaHaise

Quoted 3 lines · Eric Dumazet
So yes, we built network stack with the prereq that IP headers are aligned, but unfortunately many developers use x86 which doesnt care, so its possible some bugs are added.
x86 does have an alignment check flag that can be set in the flags register. Somehow, I doubt anyone would be willing to walk through all the noise the faults would likely trigger.

If this can be mapped to an event that can be used by perf tool, that might be useful ?

Benjamin LaHaise

2012-10-12 15:00 UTC #68 lore raw

Quoted 5 lines · Eric Dumazet

Quoted 2 lines
Somehow, I doubt anyone would be willing to walk through all the noise the faults would likely trigger.
If this can be mapped to an event that can be used by perf tool, that might be useful ?

There are performance counters for the various different types of alignment faults supported by perf. Modern x86 makes the vast majority of unaligned accesses very low overhead -- the only ones that really hurt are those straddling different vm pages, but even those have little cost compared to obsolete microarchitectures (*cough* P4 *cough*).

-ben

Signature · 1 line

"Thought is the essence of where you are now."

Ben Hutchings

2012-10-12 15:04 UTC #69 lore raw

Quoted 6 lines · Eric Dumazet

Quoted 5 lines · Benjamin LaHaise
Quoted 3 lines · Eric Dumazet
So yes, we built network stack with the prereq that IP headers are aligned, but unfortunately many developers use x86 which doesnt care, so its possible some bugs are added.

Quoted 6 lines

Quoted 3 lines
x86 does have an alignment check flag that can be set in the flags register. Somehow, I doubt anyone would be willing to walk through all the noise the faults would likely trigger.
If this can be mapped to an event that can be used by perf tool, that might be useful ?

AC was so useless that it has now been reallocated to use by SMAP
<https://lwn.net/Articles/517251/>.

Ben.

Signature · 3 lines

Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

David Laight

2012-10-12 15:47 UTC #70 lore raw

Quoted 2 lines

AC was so useless that it has now been reallocated to use by SMAP
<https://lwn.net/Articles/517251/>.

That is a long time coming!
Wonder when it will appear in any cpus.

How am I going to get root access when I can't get the kernel to execute code at address 0 any more :-)

Ben Hutchings

2012-10-12 16:13 UTC #71 lore raw

Quoted 8 lines · David Laight

Quoted 2 lines
AC was so useless that it has now been reallocated to use by SMAP
<https://lwn.net/Articles/517251/>.
That is a long time coming!
Wonder when it will appear in any cpus.
How am I going to get root access when I can't get the kernel to execute code at address 0 any more :-)

Moving further off the topic: that is supposed to be prevented by a separate feature, SMEP, which I think is available in current Intel CPUs (Ivy Bridge). Also, unprivileged users are generally not permitted to mmap() at address 0.

Ben.

Signature · 3 lines

Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

Måns Rullgård

2012-10-12 12:16 UTC #72 lore raw

Quoted 38 lines · Russell King - ARM Linux

Quoted 4 lines · M?ns Rullg?rd
Quoted 1 line · Russell King - ARM Linux
Well, I get the last word here and it's no.
Sadly, yes.
It's not "sadly" - it's a matter of fact that the kernel does from time to time generate misaligned accesses and they are _not_ bugs. If they were bugs, then the code to fix up misaligned accesses would not have been developed, and we'd instead take the fault and panic or something like that.
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 9 lines
Quoted 4 lines
If all alignment faults in the kernel are caused by broken drivers, that would at least give us some hope of finding those drivers while at the same time not causing much overhead in the case where we need to do the fixup in the meantime.
No. It is my understanding that various IP option processing can also cause the alignment fault handler to be invoked, even when the packet is properly aligned, and then there's jffs2/mtd which also relies upon alignment faults being fixed up.
Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
As far as I'm concerned, this is all hearsay, and I've only ever heard it from you. Why can't you let those who care fix these bugs instead?
You know, I'm giving you the benefit of my _knowledge_ which has been built over the course of the last 20 years.
How proud you sound. Now could you say something of substance instead?
You're proving yourself to be idiot? There, that's substance.
Quoted 6 lines
Quoted 5 lines
I've been in these discussions with networking people before. I ended up having to develop the alignment fault handler because of those discussions. And oh look, Eric confirmed that the networking code isn't going to get "fixed" as you were demanding, which is exactly what I said.

Quoted 8 lines

Quoted 4 lines
Funny, I saw him say the exact opposite:
So if you find an offender, please report a bug, because I can guarantee you we will _fix_ it.
No, let's go back.
- You were demanding that the ipv4 header structure should be packed.

I demanded no such thing.

Quoted 3 lines

I said that wasn't going to happen because the networking people wouldn't allow it, and it seems that's been proven correct.
- You were demanding that the ipv4 code used the unaligned accessors.

Again, I made no such demand.

Quoted 2 lines

I said that networking people wouldn't allow it either, and that's also been proven correct.

I did not, in fact, _demand_ anything at all. What I did say was that to fix the problem of unaligned access traps the OP was having, the (driver) code supplying the unaligned pointer should be fixed, or *if that is not possible* mark the structs __packed. As it turns out, fixing the driver is easy, so there is no need to change the structs or how they are accessed.

Quoted 10 lines

Both these points have been proven correct because Eric has said that the core networking code is _not_ going to be changed to suit this.
What Eric _has_ said is that networking people consider packets supplied to the networking layer where the IPv4 header is not aligned on architectures where misaligned accesses are a problem to be a bug in the network driver, not the network code, and proposed a solution.
That's entirely different from all your claims that the core networking code needs fixing to avoid these misaligned accesses.

I never said that. I said whatever code is responsible for the pointer should be fixed. That code turns out to be the driver.

Quoted 11 lines

Quoted 6 lines
Quoted 3 lines
I've been in discussions with MTD people over these issues before, I've discussed this with David Woodhouse when it came up in JFFS2. I *KNOW* these things.
In the same way you "know" the networking people won't fix their code, despite them _clearly_ stating the opposite?
I'll tell you exactly how I *KNOW* this. The issue came up because of noMMU, which does not have any way to fix up alignment faults. JFFS2 passes randomly aligned buffers to the MTD drivers, and the MTD drivers assume that they're aligned and they do word accesses on them.

So jffs2 is broken. Why can't it be fixed?

Quoted 1 line

See the thread http://lists.arm.linux.org.uk/lurker/thread/20021204.191632.4473796b.en.html

That thread is about detecting misaligned accesses and what to do with them if they do occur. I don't see anyone successfully arguing against fixing the code causing them.

Quoted 1 line

See: http://lists.arm.linux.org.uk/lurker/message/20020225.195925.02bdbd47.en.html

Yes, there seems to be a problem in jffs2. Or at least there was one 10 years ago.

Quoted 1 line

and: http://lists.arm.linux.org.uk/lurker/message/20020313.150932.081a7592.en.html

Yes, disabling the alignment trap on armv5 is a bad idea. Nobody has argued otherwise.

Quoted 10 lines

There's several other threads where it's also discussed.
And while you're there, note the date. There is nothing recent about this issue. It's well known, and well understood by those who have a grasp of the issues that alignment faults are a part of normal operation by the ARM kernel - and is one of the penalties of being tied into architecture independent code.
Compromises have to be sought, and that's the compromise we get to live with.

So because of some ancient history with jffs2, we should deny present-day developers the tools to quickly identify problems in their code? I just _love_ such compromises.

Quoted 8 lines

Quoted 4 lines
Quoted 2 lines
You can call it hearsay if you wish, but it seems to be more accurate than your wild outlandish and pathetic statements.
So you're resorting to name-calling. Not taking that bait.
Sorry? So what you're saying is that it's fine for you to call my comments hearsay, but I'm not allowed to express a view on your comments. How arrogant of you.

Go ahead, pile it on. I'm not falling into this trap.

Signature · 2 lines

M?ns Rullg?rd
mans at mansr.com

Russell King - ARM Linux

2012-10-12 11:19 UTC #73 lore raw

Quoted 27 lines · Russell King - ARM Linux

Quoted 6 lines · M?ns Rullg?rd
Quoted 4 lines · Russell King - ARM Linux
Quoted 3 lines · Arnd Bergmann
Quoted 2 lines · M?ns Rullg?rd
Quoted 1 line
But, the IP header is expected to be aligned.
Quoted 7 lines
Quoted 5 lines
Quoted 5 lines
Quoted 2 lines
Everything tells the compiler the struct is perfectly aligned. When the buggy driver passes a misaligned pointer, bad things happen.
Would it be appropriate to add a WARN_ON_ONCE() in the alignment fault path then?
I think that's an excellent idea.
Well, I get the last word here and it's no.
Quoted 10 lines
Quoted 9 lines
Quoted 4 lines
If all alignment faults in the kernel are caused by broken drivers, that would at least give us some hope of finding those drivers while at the same time not causing much overhead in the case where we need to do the fixup in the meantime.
No. It is my understanding that various IP option processing can also cause the alignment fault handler to be invoked, even when the packet is properly aligned, and then there's jffs2/mtd which also relies upon alignment faults being fixed up.

Quoted 7 lines

Quoted 2 lines
As far as I'm concerned, this is all hearsay, and I've only ever heard it from you. Why can't you let those who care fix these bugs instead?
You know, I'm giving you the benefit of my _knowledge_ which has been built over the course of the last 20 years. I've been in these discussions with networking people before. I ended up having to develop the alignment fault handler because of those discussions.

Correction: San Mehat at Corel Inc had to develop it, but I was involved in those discussions over it nevertheless, as I was the person who merged the code and then subsequently maintained it.

Eric Dumazet

2012-10-11 16:15 UTC #74 lore raw

Quoted 7 lines · Rob Herring

Quoted 1 line · Eric Dumazet
Compiler is asked to perform a 32bit load, it does it.
Not exactly. It is asked to to perform 2 32-bit loads which are combined into a single ldm (load multiple) which cannot handle unaligned accesses. Here's a simple example that does the same thing:

Thats simply not true. You are severely mistaken.

ldm does a load of seeral 32bit words.

And the compiler would not use it if the alignment was not matching the prereq (alignment >= 4)

Quoted 5 lines

void test(char * buf)

{
	printf("%d, %d\n", *((unsigned int *)&buf[0]), *((unsigned int *)&buf[4]));
}

But you completely miss the fact that network doesnt pass a "char *buf" but a "be32 *buf". Your example is not relevant at all.

So the compiler is absolutely right, and network stack is _right_ too.

The prereq is that IP header are aligned to 4 bytes boundary.

Denying this fact is not going to help you

Quoted 7 lines

So I guess the only ABI legal unaligned access is in a packed struct.
Quoted 2 lines
There is no questionable optimization here. Really. Please stop pretending this, this makes no sense.
I'm not the one calling the networking stack bad code.

Once you understand the issues, you can explain us where is the bad code. But given you say "Bug is in compiler, and/or network stack, but my driver is fine", its not very wise.

For the moment, the bug is in your driver.

Quoted 3 lines

I can fix my h/w, so I'll stop caring about this. Others can all get bitten by this new behavior in gcc 4.7.

Again compiler is fine. How should we say that so that you stop your rants ?

Stop trying to find an excuse, dont try to fool us, this is really embarrassing. Just fix the driver, there is no shame to fix an error.

Catalin Marinas

2012-10-11 16:59 UTC #75 lore raw

In reply to Måns Rullgård

Quoted 6 lines · Russell King - ARM Linux

Quoted 5 lines · Eric Dumazet
I took a look, and I dont see why/how gcc could use a ldm instruction
Doing so assumed the alignment of the structure was 8 bytes, but its not.

Quoted 7 lines

Quoted 2 lines
Networking stack mandates that IP headers are aligned on 4 bytes boundaries, not 8 bytes.
Err, no. ldm is "load multiple" not "load double". It loads multiple 32-bit registers, and its requirement for non-faulting behaviour is for the pointer to be 4 byte aligned. However, "load double" requires 8 byte alignment.

It got better with ARMv6 where LDRD/STRD only require 4 byte alignment (the only 8 byte alignment is required by LDREXD/STREXD). But on ARMv5 LDRD/STRD 8 byte alignment is indeed required (otherwise unpredictable).

Signature · 1 line

Catalin

David Laight

2012-10-11 10:16 UTC #76 lore raw

Quoted 5 lines

Quoted 3 lines
It might be enough to use __attribute__((aligned(2))) on some structure members (actually does 'ldm' need 8 byte alignment?? - in which case aligned(4) is enough).
The aligned attribute can only increase alignment.

Not true.
If you have:

struct foo {
    short   a;
    int     b __attribute__((aligned(2)));
    short   c;
};

You'll find sizeof (struct foo) is 8, and that, on arm/sparc etc, gcc will generate 2 16bit accesses (and shifts) for foo.b;

David

Måns Rullgård

2012-10-11 10:46 UTC #77 lore raw

Quoted 9 lines · David Laight

Quoted 5 lines
Quoted 3 lines
It might be enough to use __attribute__((aligned(2))) on some structure members (actually does 'ldm' need 8 byte alignment?? - in which case aligned(4) is enough).
The aligned attribute can only increase alignment.
Not true.
If you have:

Quoted 8 lines

struct foo {
    short   a;
    int     b __attribute__((aligned(2)));
    short   c;
};
You'll find sizeof (struct foo) is 8, and that, on arm/sparc etc, gcc will generate 2 16bit accesses (and shifts) for foo.b;

That is not what the manual says. Quoting:

When used on a struct, or struct member, the `aligned' attribute can only increase the alignment

My gcc agrees with the manual:

$ cat foo.c
struct foo {
    short a;
    int b __attribute__((aligned(2)));
    short c;
};

int foo(struct foo *f)

{
    return f->b;
}
$ arm-unknown-linux-gnueabi-gcc-4.7.2 -O2 -o- -S foo.c
        .text
        .align  2
        .global foo
        .type   foo, %function
foo:
        ldr     r0, [r0, #4]
        bx      lr

(compiler output edited for clarity)

This clearly says the 'b' member resides at an offset of 4 bytes into the struct.

Signature · 2 lines

M?ns Rullg?rd
mans at mansr.com

Rob Herring

author

2012-10-05 16:08 UTC #78 lore raw

In reply to Michael Hope

Quoted 31 lines · Michael Hope

Quoted 20 lines · Rob Herring
I've been scratching my head with a "scheduling while atomic" bug I started seeing on 3.6. I can easily reproduce this problem when doing a wget on my system. It ultimately seems to be a combination of factors. The "scheduling while atomic" bug is triggered in do_alignment which gets triggered by this code in net/ipv4/af_inet.c, line 1356:
id = ntohl(*(__be32 *)&iph->id);
flush = (u16)((ntohl(*(__be32 *)iph) ^ skb_gro_len(skb)) | (id ^ IP_DF)); id >>= 16;
This code compiles into this using "gcc version 4.6.3 (Ubuntu/Linaro 4.6.3-1ubuntu5)":
c02ac020:       e8920840        ldm     r2, {r6, fp}
c02ac024:       e6bfbf3b        rev     fp, fp
c02ac028:       e6bf6f36        rev     r6, r6
c02ac02c:       e22bc901        eor     ip, fp, #16384  ; 0x4000
c02ac030:       e0266008        eor     r6, r6, r8
c02ac034:       e18c6006        orr     r6, ip, r6
Quoted 2 lines
which generates alignment faults on the ldm. These are silent until this commit is applied:
Hi Rob. I assume that iph is something like:
struct foo {
    u32 x;
    char id[8];
};

Quoted 5 lines

struct foo *iph;
GCC merged the two adjacent loads of x and id into one ldm. This is an ARM specific optimisation done in load_multiple_sequence() and enabled with -fpeephole2.

I'm probably on some watch list now after searching for peephole...

It's not clear what all turning this off would affect. Is it just struct or array loading? Or does this turn off lots of different optimizations?

Rob

Russell King - ARM Linux

2012-10-05 07:29 UTC #79 lore raw

Quoted 6 lines · Rob Herring

I would think the scheduling while atomic messages are harmless in this case. However, in addition to spewing out BUG messages this commit also seems to eventually cause a kernel panic in __napi_complete. That panic seems to go away if I put barrier() between the 2 accesses above which eliminates the alignment faults. I haven't figured that part out yet.

Quoted 6 lines

There's at least a couple of problems here:
This seems like an overly aggressive compiler optimization considering unaligned accesses are not supported by ldm/stm.
The alignment fault handler should handle kernel address faults atomically.

This is bad news. do_alignment() can be called in almost any kernel context, and it must work. die() and oops dumps - specifically dump_mem() and dump_instr() will suffer from exactly the same problem.

Will, can you take a look please?

Russell King - ARM Linux

2012-10-05 10:51 UTC #80 lore raw

Quoted 9 lines · Russell King - ARM Linux

Quoted 8 lines · Rob Herring
I would think the scheduling while atomic messages are harmless in this case. However, in addition to spewing out BUG messages this commit also seems to eventually cause a kernel panic in __napi_complete. That panic seems to go away if I put barrier() between the 2 accesses above which eliminates the alignment faults. I haven't figured that part out yet.
There's at least a couple of problems here:

Quoted 8 lines

Quoted 4 lines
This seems like an overly aggressive compiler optimization considering unaligned accesses are not supported by ldm/stm.
The alignment fault handler should handle kernel address faults atomically.
This is bad news. do_alignment() can be called in almost any kernel context, and it must work. die() and oops dumps - specifically dump_mem() and dump_instr() will suffer from exactly the same problem.

Okay, this should fix the issue... I've only compile tested it so far. Rob, as you have a way to trigger this easily, can you give this patch a go and let me know if it solves your problem? Thanks.

 arch/arm/kernel/traps.c |   34 +++++++---------------------------
 arch/arm/mm/alignment.c |   11 ++++-------
 2 files changed, 11 insertions(+), 34 deletions(-)

arch/arm/kernel/traps.c +7 ~~-27~~

index b0179b8..62f429e 100644
@@ -89,17 +89,8 @@ static void dump_mem(const char *lvl, const char *str, unsigned long bottom,
 		     unsigned long top)
 {
 	unsigned long first;
-	mm_segment_t fs;
 	int i;
 
-	/*
-	 * We need to switch to kernel mode so that we can use __get_user
-	 * to safely read from kernel space.  Note that we now dump the
-	 * code first, just in case the backtrace kills us.
-	 */
-	fs = get_fs();
-	set_fs(KERNEL_DS);
-
 	printk("%s%s(0x%08lx to 0x%08lx)\n", lvl, str, bottom, top);
 
 	for (first = bottom & ~31; first < top; first += 32) {

@@ -112,7 +103,7 @@ static void dump_mem(const char *lvl, const char *str, unsigned long bottom,
 		for (p = first, i = 0; i < 8 && p < top; i++, p += 4) {
 			if (p >= bottom && p < top) {
 				unsigned long val;
-				if (__get_user(val, (unsigned long *)p) == 0)
+				if (probe_kernel_address(p, val) == 0)
 					sprintf(str + i * 9, " %08lx", val);
 				else
 					sprintf(str + i * 9, " ????????");

@@ -120,8 +111,6 @@ static void dump_mem(const char *lvl, const char *str, unsigned long bottom,
 		}
 		printk("%s%04lx:%s\n", lvl, first & 0xffff, str);
 	}
-
-	set_fs(fs);
 }
 
 static void dump_instr(const char *lvl, struct pt_regs *regs)

@@ -129,25 +118,18 @@ static void dump_instr(const char *lvl, struct pt_regs *regs)
 	unsigned long addr = instruction_pointer(regs);
 	const int thumb = thumb_mode(regs);
 	const int width = thumb ? 4 : 8;
-	mm_segment_t fs;
 	char str[sizeof("00000000 ") * 5 + 2 + 1], *p = str;
 	int i;
 
-	/*
-	 * We need to switch to kernel mode so that we can use __get_user
-	 * to safely read from kernel space.  Note that we now dump the
-	 * code first, just in case the backtrace kills us.
-	 */
-	fs = get_fs();
-	set_fs(KERNEL_DS);
-
 	for (i = -4; i < 1 + !!thumb; i++) {
 		unsigned int val, bad;
 
-		if (thumb)
-			bad = __get_user(val, &((u16 *)addr)[i]);
-		else
-			bad = __get_user(val, &((u32 *)addr)[i]);
+		if (thumb) {
+			u16 instr;
+			bad = probe_kernel_address(addr, instr);
+			val = instr;
+		} else
+			bad = probe_kernel_address(addr, val);
 
 		if (!bad)
 			p += sprintf(p, i == 0 ? "(%0*x) " : "%0*x ",

@@ -158,8 +140,6 @@ static void dump_instr(const char *lvl, struct pt_regs *regs)
 		}
 	}
 	printk("%sCode: %s\n", lvl, str);
-
-	set_fs(fs);
 }
 
 #ifdef CONFIG_ARM_UNWIND

arch/arm/mm/alignment.c +4 -7

index b9f60eb..f8f14fc 100644
@@ -749,7 +749,6 @@ do_alignment(unsigned long addr, unsigned int fsr, struct pt_regs *regs)
 	unsigned long instr = 0, instrptr;
 	int (*handler)(unsigned long addr, unsigned long instr, struct pt_regs *regs);
 	unsigned int type;
-	mm_segment_t fs;
 	unsigned int fault;
 	u16 tinstr = 0;
 	int isize = 4;

@@ -760,16 +759,15 @@ do_alignment(unsigned long addr, unsigned int fsr, struct pt_regs *regs)
 
 	instrptr = instruction_pointer(regs);
 
-	fs = get_fs();
-	set_fs(KERNEL_DS);
 	if (thumb_mode(regs)) {
-		fault = __get_user(tinstr, (u16 *)(instrptr & ~1));
+		unsigned long ptr = instrptr;
+		fault = probe_kernel_address(ptr, tinstr);
 		if (!fault) {
 			if (cpu_architecture() >= CPU_ARCH_ARMv7 &&
 			    IS_T32(tinstr)) {
 				/* Thumb-2 32-bit */
 				u16 tinst2 = 0;
-				fault = __get_user(tinst2, (u16 *)(instrptr+2));
+				fault = probe_kernel_address(ptr + 2, tinst2);
 				instr = (tinstr << 16) | tinst2;
 				thumb2_32b = 1;
 			} else {

@@ -778,8 +776,7 @@ do_alignment(unsigned long addr, unsigned int fsr, struct pt_regs *regs)
 			}
 		}
 	} else
-		fault = __get_user(instr, (u32 *)instrptr);
-	set_fs(fs);
+		fault = probe_kernel_address(instrptr, instr);
 
 	if (fault) {
 		type = TYPE_FAULT;

Jon Masters

2012-10-23 16:30 UTC #81 lore raw

On 10/05/2012 06:51 AM, Russell King - ARM Linux wrote:

Quoted 3 lines

Okay, this should fix the issue... I've only compile tested it so far. Rob, as you have a way to trigger this easily, can you give this patch a go and let me know if it solves your problem? Thanks.

Russell, can you let me know the status of this patch? I don't see it in your tree, and I think I might have missed a followup patch from you? It seems to solve a real problem with in-kernel faults during the dump_mem in general, even beyond just an alignment fault.

Jon.

Russell King - ARM Linux

2012-10-23 16:58 UTC #82 lore raw

Quoted 5 lines · Jon Masters

On 10/05/2012 06:51 AM, Russell King - ARM Linux wrote:
<snip switch out the might_fault __get_user for atomic probe_kernel_address related discussion>

Quoted 8 lines

Quoted 3 lines
Okay, this should fix the issue... I've only compile tested it so far. Rob, as you have a way to trigger this easily, can you give this patch a go and let me know if it solves your problem? Thanks.
Russell, can you let me know the status of this patch? I don't see it in your tree, and I think I might have missed a followup patch from you? It seems to solve a real problem with in-kernel faults during the dump_mem in general, even beyond just an alignment fault.

I was waiting for someone to say "I saw a problem and it works"... Would be nice to have a Tested-by tag against it...

Jon Masters

2012-10-23 17:15 UTC #83 lore raw

Quoted 10 lines · Russell King - ARM Linux

Quoted 9 lines · Jon Masters
On 10/05/2012 06:51 AM, Russell King - ARM Linux wrote:
<snip switch out the might_fault __get_user for atomic probe_kernel_address related discussion>
Quoted 3 lines
Okay, this should fix the issue... I've only compile tested it so far. Rob, as you have a way to trigger this easily, can you give this patch a go and let me know if it solves your problem? Thanks.

Quoted 7 lines

Quoted 4 lines
Russell, can you let me know the status of this patch? I don't see it in your tree, and I think I might have missed a followup patch from you? It seems to solve a real problem with in-kernel faults during the dump_mem in general, even beyond just an alignment fault.
I was waiting for someone to say "I saw a problem and it works"... Would be nice to have a Tested-by tag against it...

I've poked at the patch, made one against Fedora, and asked someone in my team to do a scratch build and test on our local highbank/other systems. We'll come back with that tag for ya asap, thanks!

Jon.

Rob Herring

author Tested-by

2012-10-23 19:14 UTC #84 lore raw