Skip to contentSkip to conversation

Alignment on XScale ARM

DISCUSSION Marcus Tangermann 4 participants 4 messages Nov 7, 2008 – Nov 8, 2008

lore
Share & tools
lore
#0 lore raw

Hi,

we use an XScale 422 (ARM) with a big endian Linux system based on the buildroot environment. It seems there is an alignment problem we have. I've tracked down the problem to the following.
1. The data received via a netlink communication from the kernel is stored into a buffer buf. The data is correct.
2. The address of the buffer is converted to a pointer to nlmsghdr:

         struct nlmsghdr *h;
         h = (struct nlmsghdr*) buf; 
         printf("IPSECTOOLS: Length  %04d\n", h->nlmsg_len);
         printf("IPSECTOOLS: Type %02d\n", h->nlmsg_type);
   The results of printf are nonsens. The first fields seem to have an offset of 2 bytes, at the end there also seem to be swapped bytes.
3. To test what might happen, I've created an own struct
       struct aligntest{
                __u32 first;
                __u32 int second;
                __u32 third;
                __u32 fourth;
       };
     When you now set a point to a buffer, again the values are wrong:
         s1 = (struct aligntest*) buf;
         printf("1: %08X\n", s1->first);
         printf("2: %08X\n", s1->second);
         printf("3: %08X\n", s1->third);
         printf("4: %08X\n", s1->fourth);

It seems, that there is a problem with the unsigned int values. When you add the attribute "packed" to the definition of struct aligntest ( __attribute__(__packed__) ), all values are displayed correctly. So, any hints what the problem can be?

Regards
Marcus
____________________________________________________________________
Psssst! Schon vom neuen WEB.DE MultiMessenger gehört?
Der kann`s mit allen: http://www.produkte.web.de/messenger/?did=3123

#1 lore raw
Quoted 24 lines · Marcus Tangermann

Hi,

we use an XScale 422 (ARM) with a big endian Linux system based on the buildroot environment. It seems there is an alignment problem we have. I've tracked down the problem to the following.
1. The data received via a netlink communication from the kernel is stored into a buffer buf. The data is correct.
2. The address of the buffer is converted to a pointer to nlmsghdr:

        struct nlmsghdr *h;
        h = (struct nlmsghdr*) buf;
        printf("IPSECTOOLS: Length  %04d\n", h->nlmsg_len);
        printf("IPSECTOOLS: Type %02d\n", h->nlmsg_type);
  The results of printf are nonsens. The first fields seem to have an offset of 2 bytes, at the end there also seem to be swapped bytes.
3. To test what might happen, I've created an own struct
      struct aligntest{
               __u32 first;
               __u32 int second;
               __u32 third;
               __u32 fourth;
      };
    When you now set a point to a buffer, again the values are wrong:
        s1 = (struct aligntest*) buf;
        printf("1: %08X\n", s1->first);
        printf("2: %08X\n", s1->second);
        printf("3: %08X\n", s1->third);
        printf("4: %08X\n", s1->fourth);
Quoted 2 lines

It seems, that there is a problem with the unsigned int values. When you add the attribute "packed" to the definition of struct aligntest ( __attribute__(__packed__) ), all values are displayed correctly. So, any hints what the problem can be?

I'm not sure what the problem is? You can't generally misalign a struct pointer safely.

#2 lore raw
In reply to Marcus Tangermann
Quoted 28 lines · Marcus Tangermann

Hi,

we use an XScale 422 (ARM) with a big endian Linux system based on the buildroot environment. It seems there is an alignment problem we have. I've tracked down the problem to the following. 1. The data received via a netlink communication from the kernel is stored into a buffer buf. The data is correct. 2. The address of the buffer is converted to a pointer to nlmsghdr:

         struct nlmsghdr *h;
         h = (struct nlmsghdr*) buf; 
         printf("IPSECTOOLS: Length  %04d\n", h->nlmsg_len);
         printf("IPSECTOOLS: Type %02d\n", h->nlmsg_type);
The results of printf are nonsens. The first fields seem to have an 
offset of 2 bytes, at the end there also seem to be swapped bytes. 3. 
To test what might happen, I've created an own struct
       struct aligntest{
                __u32 first;
                __u32 int second;
                __u32 third;
                __u32 fourth;
       };
When you now set a point to a buffer, again the values are wrong:
         s1 = (struct aligntest*) buf;
         printf("1: %08X\n", s1->first);
         printf("2: %08X\n", s1->second);
         printf("3: %08X\n", s1->third);
         printf("4: %08X\n", s1->fourth);
Quoted 4 lines

It seems, that there is a problem with the unsigned int values. When you add the attribute "packed" to the definition of struct aligntest ( __attribute__(__packed__) ), all values are displayed correctly. So, any hints what the problem can be?

The ARM processor traditionally cannot access misaligned short and int values in memory. By using the packed attribute you tell the compiler that the structure should not have inserted padding to align its members, and that accesses are to be performed wth byte accesses and the value reconstructed that way.

Nicolas

#3 lore raw
Quoted 5 lines · Nicolas Pitre

The ARM processor traditionally cannot access misaligned short and int values in memory.  By using the packed attribute you tell the compiler that the structure should not have inserted padding to align its members, and that accesses are to be performed wth byte accesses and the value reconstructed that way.

... but then when you pass "&misaligned_ptr->member" to some other function, that other function generally won't know it needs to use byte access. So, be careful. The "packed" struct attribute isn't a panacea.

General Linux kernel policy is to use <asm/unaligned.h> accessors, but obviously "packed" has advantages when you've got to pull structures out of protocol buffers. The userspace headers don't seem to support such accessors though.

Also, the network stack is known for making assumptions about its data being properly aligned. See how NET_IP_ALIGN is used inside the kernel; you might be happier aligning your RX buffers before reading data into them from the kernel.

- Dave