Bad traps for unaligned access in STM instruction

[Lluís Batlle i Rossell]

On Tue, Sep 18, 2012 at 04:25:02PM +0100, Dave Martin wrote:
> On Sat, Sep 15, 2012 at 12:51:51PM +0200, Lluís Batlle i Rossell wrote:
> > On Sat, Sep 15, 2012 at 12:46:02PM +0200, Lluís Batlle i Rossell wrote:
> > > I'll start looking at the arm traps for unaligned accesses, but maybe someone
> > > here can give a quick answer.
> > 
> > Ah stupid me, here goes the quick answer. I had mode 'warn', and not
> > 'fixup+warn'. With 'fixup' enabled, all works fine.
> > 
> > I was used to mips, where I think there isn't a 'warn' that doesn't 'fixup'.
> 
> What source are you actually trying to build/run here?  The "fixup" alignment
> mode is primarily a workaround for incorrect code, because the legacy rotated
> unaligned access behaviour would cause really weird things to happen silently
> in the offending code.  (Although misaligned LDM/STM was never permitted by
> the architecture anyway, and normally indicates badly-written code.)
> 
> Pure C code should never trigger alignment faults fixups unless it
> violates the C language specification.
> 
> Assembler should not trigger faults at all, because it's arch-specific
> and so you can and should fix it not to cause faults.  Optimising code
> in assembler becomes pointless if you write or use it in a way which
> triggers unnecessary faults into the kernel.
> 
> If you are getting faults in compiled code and the source code follows
> the C standard with regard to alignment requirements, this suggests a
> bug in the compiler.

This happened in the btrfs userland code, where there are packed structs, and
structs stored at some offsets of datablocks. Offsets like '+17 bytes' since
block start loaded from disk, and so.

Then, some functions take pointers to structs. The "datablock+17bytes" is passed
to processing functions as a pointer to struct. The code in those functions,
expects that the structs are properly aligned.

In an ideal userland software world, the counters of the traps would be zero.
But they are not. And without fixup, the mkfs.btrfs fails to run properly, and
creates a broken filesystem that crashes the kernel.

Is the best way to deal with it, to allocate the structs in stack, copy from
unaligned places to stack, and then work with those stack struct pointers?

Thank you a lot for your answer,
Lluís.