Skip to contentSkip to conversation

Follow-up to remaining issue with alignment of __log_buf in printk.c

DISCUSSION Robert Jarzmik 3 participants 4 messages Jun 14, 2012 – Jun 15, 2012

lore
Share & tools
lore
#0 lore raw

Hi Stephen and others,

I have a XScale PXA based board with has the alignement issue which makes the kernel trap during its early stage.

I wonder what is the status now, is there a fix available ?

I have tracked what happens on PXA. The pxa is an ARM v5TE chip. The new printk version you submitted is translated to the following assembly on the line :

        msg->ts_nsec = local_clock();

Into:
=> 0xc001bbe0 <log_store+496>: strd r0, [r4, r5]

In ARMv5, the "strd" assembly opcode expects the address to be 64bits aligned, hence the bug.

Now the solutions I have seen so far in the mailing lists :
- #define LOG_ALIGN (__alignof__(u64))
Does always work.
- #define LOG_ALIGN (__alignof__(struct log))
Doesn't work with my toolchain, as __alignof__(struct log) is 4, not 8

What are you intending to do to solve the ARMv5 issue ? Are you waiting for someone to submit a patch ?

Cheers.

Signature · 1 line
Robert
#1 lore raw
Quoted 14 lines · Robert Jarzmik

Hi Stephen and others,

I have a XScale PXA based board with has the alignement issue which makes the kernel trap during its early stage.

I wonder what is the status now, is there a fix available ?

I have tracked what happens on PXA. The pxa is an ARM v5TE chip. The new printk version you submitted is translated to the following assembly on the line :

        msg->ts_nsec = local_clock();

Into:
=> 0xc001bbe0 <log_store+496>: strd r0, [r4, r5]

Quoted 8 lines

In ARMv5, the "strd" assembly opcode expects the address to be 64bits aligned, hence the bug.

Now the solutions I have seen so far in the mailing lists :
- #define LOG_ALIGN (__alignof__(u64))
Does always work.
- #define LOG_ALIGN (__alignof__(struct log))
Doesn't work with my toolchain, as __alignof__(struct log) is 4, not 8

Isn't that a bug in the toolchain; isn't the alignment of a struct required to be the greatest alignment of any of its members? Otherwise, this problem could arise with any usage of that struct.

I suppose this could be worked around with something like:

#define LOG_ALIGN max(__alignof__(struct log), __alignof__(u64))

#2 lore raw
Quoted 3 lines · Stephen Warren

Isn't that a bug in the toolchain; isn't the alignment of a struct required to be the greatest alignment of any of its members? Otherwise, this problem could arise with any usage of that struct.

I'm not a compiler expert, but I think EABI requires 64-bit data types to be 8-bytes aligned. If that assertion is true, then clearly my eabi compiler is a little liar with its __alignof__(struct log) result.
And I think also that ISO C standard requires a structure to be aligned by default to the <quote>least perfect multiple of the lowest common multiple of the alignments of all the members of the struct</quote>. Which here means 8 bytes ...

Quoted 2 lines

I suppose this could be worked around with something like:
#define LOG_ALIGN max(__alignof__(struct log), __alignof__(u64))

Well, I wonder what is the correct approach here. In the end, the compiler should be fixed ... But meanwhile, your proposition would solve the PXA problem. The real deal is to know whether we want a compiler work-around in printk.c ... and that's rather a policy issue than a technical one.

I wonder what others think about it ... let's wait for another pair of point of views.

Cheers.

Signature · 6 lines
Robert

PS: For Andrew, I saw your patch. It won't solve my problem as my compiler
believes __alignof__(struct log) == 4 ... bad luck for me.

PPS: My toolchain is arm-2011.09
#3 lore raw
Quoted 4 lines

PS: For Andrew, I saw your patch. It won't solve my problem as my compiler believes __alignof__(struct log) == 4 ... bad luck for me.

PPS: My toolchain is arm-2011.09

Hi Robert

I have gcc version 4.4.5 (Debian 4.4.5-8), both native and cross.

Andrew