Re: [Qemu-devel] [PATCH v6 1/8] atomic: introduce atomic operations

[Richard Henderson <rth@twiddle.net>]

On 2012-11-14 23:47, liu ping fan wrote:
> Probably I made a mistake here, in vhost,  log =
> __sync_fetch_and_and(from, 0)  is used to fetch 64bits atomically in
> the case  32bits qemu running on 64bits linux.  Right?   But how can
> we read 32bits twice in atomic?  Seem that no instruction like "_lock
> xchg" for this ops.  So I guess _sync_fetch_and_and() based on
> something like spinlock.

... or for gcc 4.7 and later,

  log = __atomic_load_n(from, memory_model)

For i386, we will not perform 2 32-bit reads of course.  Paulo suggests
using cmpxchg8b, but that's a tad slow.  Instead we'll perform a 64-bit
read into either the fpu or the sse units, and from there copy the data
wherever it's needed.  Such 64-bit aligned reads are guaranteed to be
atomic for i586 (pentium) and later.

For other 32-bit architectures other possibilities exist.  Recent arm can
use its ldrexd insn.  Many of the 32-bit linux architectures have special
kernel entry points or schemes to perform atomic operations.  These are
generally based on the assumption of a single-processor system, and are
arranged to either disable interrupts or notice that no interrupt occurred,
while executing a code region.

As an ultimate fallback, yes we would use locks.  But none of the host
architectures that QEMU supports needs to do so.


r~