Re: [LTP] LTP mmap01 failure due to incorrect address alignment

[Martin H. <err...@mp...>]

Looking into failures for shmat01 and shmt09 I came upon sys/shm.h,
which defines SHMLBA.
Unfortunately it is set to the pagesize, which is wong as well.
I've opened glibc bug #6471 for this.

If I patch sys/shm.h to the cache size the shmat01 test succeeds.
shmt09 needs an additional patch to correct the sbrk increment to
the SHMLBA size, which you find attached.
I'll submit a patch for mmap01 later this week.

Best regards,
Martin

	Signed-off-by: Martin Habets <err...@mp...>

On Wed, Apr 30, 2008 at 10:59:26PM +0100, Martin Habets wrote:
> LTP test mmap01 fails on sparc with:
> 
> mmap01      1  FAIL  :  mmap() Failed on mmapfile, errno=22
> 
> because the address passed in is only page aligned:
> 
> 	/*
> 	 * Initialize addr to align with the first page boundary above the
> 	 * break address of the process.
> 	 */
> 	addr = (void *) (((intptr_t)sbrk(0) + (page_sz - 1)) & ~(page_sz - 1));
> ...
> 	addr = mmap(addr, page_sz, PROT_READ | PROT_WRITE,
> 		    MAP_FILE|MAP_SHARED|MAP_FIXED, fildes, 0);
> 
> The address should be aligned to the value of SHMLBA as defined in
> asm/shmparam.h.
> See Documentation/cachetlb.txt for the explanation behind this.
> For most architectures this value is the page size, but not all.
> 
> Unfortunately asm/shmparam.h is not suitable for inclusion in user space
> programs. I looked at sysconf(LEVEL1_DCACHE_SIZE), but that returns 0 on
> sparc.
> Is there another solution other than hardcoding the exceptions in
> architecture specific code?
> 
> Thanks,
> Martin