Unexpected unaligned access on arm

Unexpected unaligned access on arm

[Lars Poeschel]

[-- Attachment #1: Type: text/plain, Size: 1956 bytes --]

Hi!

I am experiencing a problem on ARM with unaligned access where I think
it should not happen. I could track this down to the attached example. I
compile the code using the following command:

arm-linux-gnueabihf-gcc -Wall -Os -g -Wa,-mthumb -mthumb-interwork -mno-unaligned-access -mhard-float -march=armv7-a -static -o gcc_example gcc_example.c

The problem is in net_set_ip_header function. The relevant produced
assembler output is this:

   1044c:	b513      	push	{r0, r1, r4, lr}
   1044e:	4604      	mov	r4, r0
   10450:	9200      	str	r2, [sp, #0]
   10452:	300c      	adds	r0, #12
   10454:	4a11      	ldr	r2, [pc, #68]	; (1049c <net_set_ip_header+0x50>)
   10456:	4b10      	ldr	r3, [pc, #64]	; (10498 <net_set_ip_header+0x4c>)
   10458:	447a      	add	r2, pc
   1045a:	9101      	str	r1, [sp, #4]
   1045c:	f840 3c0c 	str.w	r3, [r0, #-12]
   ...
   10498:	14000045 	.word	0x14000045
   1049c:	00060a48 	.word	0x00060a48

The compiler is smart and optimizes setting the constants to the struct
ip_udp_hdr fields ip_hl_v ip_tos and ip_len to one single 32bit (word)
access. It stores the three constants pc relative in a word. This gets
loaded to r3 and then stored to the structure in memory (str.w
instruction).
Since I compile with -mno-unaligned-access and the relevant structure
members are 8 / 16bit and even the function argument uint8_t *pkt pointer
tell the compiler that these are not neccessarily 32bit aligned, I think
it should not do this.
The relevant code crashes on my cpu depending on the value of the pkt
pointer when calling the function of course. The
pkt += ETH_HDR_SIZE;
line can be used to modify the pointer if needed.
If I compile with -O0 or -O1 it does not happen. It does also not
happen when commenting out the asm("nop"); as a optimization barrier.
I tested this with gcc 8.3.1, 9.3.0 and 10.1.0. They all behave more or
less the same in this regard.

What am I missing ? What am I doing wrong ?

Thank you,
Lars

[-- Attachment #2: gcc_test.c --]
[-- Type: text/x-csrc, Size: 1708 bytes --]

#include <arpa/inet.h>
#include <string.h>
#include <sys/types.h>

#define PKTBUFSRX 4
#define PKTSIZE_ALIGN 1536
#define PKTALIGN 16

#define ETH_HDR_SIZE 14
#define IP_HDR_SIZE 20

#define IP_FLAGS_DFRAG 0x4000

static uint8_t net_pkt_buf[(PKTBUFSRX+1) * PKTSIZE_ALIGN + PKTALIGN];

static unsigned net_ip_id;

/*
 *	Internet Protocol (IP) + UDP header.
 */
struct ip_udp_hdr {
	uint8_t		ip_hl_v;	/* header length and version	*/
	uint8_t		ip_tos;		/* type of service		*/
	uint16_t	ip_len;		/* total length			*/
	uint16_t	ip_id;		/* identification		*/
	uint16_t	ip_off;		/* fragment offset field	*/
	uint8_t		ip_ttl;		/* time to live			*/
	uint8_t		ip_p;		/* protocol			*/
	uint16_t	ip_sum;		/* checksum			*/
	struct in_addr	ip_src;		/* Source IP address		*/
	struct in_addr	ip_dst;		/* Destination IP address	*/
};

void net_set_ip_header(uint8_t *pkt, struct in_addr dest, struct in_addr source)
{
	struct ip_udp_hdr *ip = (struct ip_udp_hdr *)pkt;

	/*
	 *	Construct an IP header.
	 */
	ip->ip_hl_v  = 0x45;
    //asm("nop");
	ip->ip_tos   = 0;
	ip->ip_len   = htons(IP_HDR_SIZE);
	ip->ip_id    = htons(net_ip_id++);
	ip->ip_off   = htons(IP_FLAGS_DFRAG);	/* Don't fragment */
	ip->ip_ttl   = 255;
	ip->ip_sum   = 0;
	/* already in network byte order */
	memcpy((void *)&ip->ip_src, &source, sizeof(struct in_addr));
	/* already in network byte order */
	memcpy((void *)&ip->ip_dst, &dest, sizeof(struct in_addr));
}

int main(int argc, char** argv) {
	uint8_t *pkt;
	struct in_addr bcast_ip, src_ip;

	bcast_ip.s_addr = 0xFFFFFFFFL;
	src_ip.s_addr = 0x0L;

	pkt = net_pkt_buf;
	memset((void *)pkt, 0, sizeof(net_pkt_buf));
	pkt += ETH_HDR_SIZE;

	net_set_ip_header(pkt, bcast_ip, src_ip);

	return 1;
}

Re: Unexpected unaligned access on arm

[Alexander Monakov]

On Mon, 8 Jun 2020, Lars Poeschel wrote:


> What am I missing ? What am I doing wrong ?

The field

    struct in_addr  ip_src;

is 4-byte-aligned, forcing the containing structure to also be 4-byte aligned.
From that the compiler deduces the minimum required alignment of individual
members.

Alexander

Re: Unexpected unaligned access on arm

[Oleg Endo]

On Mon, 2020-06-08 at 15:34 +0200, Lars Poeschel wrote:
> 
> What am I missing ? What am I doing wrong ?
> 

You're casting an address of some byte array to a point to struct,
which has an alignment > 1 byte.  Try adding a #pragma pack (1) or
respective attribute.

Cheers,
Oleg

Re: Unexpected unaligned access on arm

[Lars Poeschel]

On Mon, Jun 08, 2020 at 10:49:22PM +0900, Oleg Endo wrote:
> On Mon, 2020-06-08 at 15:34 +0200, Lars Poeschel wrote:
> > 
> > What am I missing ? What am I doing wrong ?
> > 
> 
> You're casting an address of some byte array to a point to struct,
> which has an alignment > 1 byte.  Try adding a #pragma pack (1) or
> respective attribute.

Thank you! The pragma does indeed the right thing. gcc now produces
code, that accesses the fields in question individually.
But shouldn't the option
-mno-unaligned-access
I use for compiling also do the same ? Why is this not working ?

Thanks again,
Lars

Re: Unexpected unaligned access on arm

[Andrew Haley]

On 09/06/2020 09:05, Lars Poeschel wrote:
> On Mon, Jun 08, 2020 at 10:49:22PM +0900, Oleg Endo wrote:
>> On Mon, 2020-06-08 at 15:34 +0200, Lars Poeschel wrote:
>>>
>>> What am I missing ? What am I doing wrong ?
>>>
>>
>> You're casting an address of some byte array to a point to struct,
>> which has an alignment > 1 byte.  Try adding a #pragma pack (1) or
>> respective attribute.
>
> Thank you! The pragma does indeed the right thing. gcc now produces
> code, that accesses the fields in question individually.
> But shouldn't the option
> -mno-unaligned-access
> I use for compiling also do the same ?

No.

'-munaligned-access'
'-mno-unaligned-access'
     Enables (or disables) reading and writing of 16- and 32- bit values
     from addresses that are not 16- or 32- bit aligned.  By default
     unaligned access is disabled for all pre-ARMv6, all ARMv6-M and for
     ARMv8-M Baseline architectures, and enabled for all other
     architectures.  If unaligned access is not enabled then words in

-mno-unaligned-access tells the compiler not to generate accesses to
unaligned words. In this test case, the compiler didn't generate the
unaligned access, *you* did. You did this by casting the address of an
unaligned byte array to the address of a struct.

If you take the address of a struct, cast it to a char*, then cast it
back to a pointer to the struct type, that'll work. Likewise if you
cast the result of malloc() to any struct pointer.

  If you lie to the compiler, it will get its revenge.
    - Henry Spencer

-- 
Andrew Haley  (he/him)
Java Platform Lead Engineer
Red Hat UK Ltd. <https://www.redhat.com>
https://keybase.io/andrewhaley
EAC8 43EB D3EF DB98 CC77 2FAD A5CD 6035 332F A671