Re: [Qemu-devel] QEMU issues on SPARC hosts due to more-than-page-alignment requirement for MAP_SHARED|MAP_FIXED...

[Laszlo Ersek <lersek@redhat.com>]

On 06/15/17 20:17, Peter Maydell wrote:
> It turns out that SPARC hosts impose a requirement that
> if you mmap MAP_SHARED|MAP_FIXED then the address must be
> aligned to SHMLBA alignment, which for SPARC is 16K and
> larger than the 8K page size:
>  http://lxr.linux.no/#linux+v4.10.1/arch/sparc/kernel/sys_sparc_64.c#L158
> 
> QEMU doesn't really cope with this. In particular, ivmshm
> doesn't work:
> 
> $ ppc64-softmmu/qemu-system-ppc64 -display none -device
> ivshmem,shm=/qtest-20273-1334572184,size=0x1
> qemu-system-ppc64: System page size 0x2000 is not enabled in
> page_size_mask (0x11000). Performance may be slow
> qemu-system-ppc64: -device
> ivshmem,shm=/qtest-20273-1334572184,size=0x1: ivshmem is deprecated,
> please use ivshmem-plain or ivshmem-doorbell instead
> Unexpected error in file_ram_alloc() at /home/pm215/qemu/exec.c:1599:
> qemu-system-ppc64: -device
> ivshmem,shm=/qtest-20273-1334572184,size=0x1: unable to map backing
> store for guest RAM: Invalid argument
> Aborted
> 
> (noticed because this is a make check failure).
> 
> exec.c calls qemu_fd_getpagesize() to figure out the alignment
> here (on a file in /dev/shm/) and ends up in its default codepath
> that returns getpagesize(), which is too small.
> 
> What's the best way to fix this? As far as I can tell
> the restriction applies to any mapping, not just /dev/shm files.
> QEMU's assumptions that pagesize is the alignment requirement
> seem fairly widespread :-(

In POSIX, SHMLBA is specified for shmat(), and not specified for mmap().

http://pubs.opengroup.org/onlinepubs/9699919799/functions/shmat.html
http://pubs.opengroup.org/onlinepubs/9699919799/functions/mmap.html

Apparently, /dev/shm (tmpfs) stands for the same sysv memory sharing
that shmat() stands for. Perhaps we can collect the /dev/shm mappings
and apply the shmat() restriction(s) to them explicitly -- but only to them.

If that's not the case, i.e. Linux on SPARC enforces SHMLBA for all
kinds of mmap() -- which I would call a POSIX conformance bug --, then
we can perhaps generally round up to SHMLBA within qemu_fd_getpagesize()
and qemu_mempath_getpagesize(), on SPARC hosts.

Laszlo