On Sat, Jul 30, 2016 at 01:44:48PM -0700, Linus Torvalds wrote:
> No. Let's not. "smp_rmb()" is completely free on x86 (ok, so it's a
> instruction scheduling barrer - close enough), so trying to optimize
> away rmb's and replacing them with double compares sounds entirely
> misdesigned.
>
> Yes, yes, there are other architectures where rmb is much more
> expensive. But quite frankly, in most cases those architectures have
> broken synchronization to begin with ("synchronization is unusual, so
> let's not optimize it"). They'll fix it eventually.
>
> Instead, what we should look at, is to make raw_seqcount_begin() use a
> smp_load_acquire() on architectures where that is cheaper than the
> rmb.
>
> But again, I don't see the point of double-testing "parent" when a
> load-acquire or load+rmb _should_ be cheap (and absolutely is on x86).
Umm... Even on x86, a lot of hash chain elements will have ->d_parent
mismatch. Suppose rmb was a no-op; current code does
fetch ->d_seq
fetch ->d_parent
compare with register
branch taken to the end of body
while this would avoid the first fetch. On the entries with the same
->d_parent we'd do
fetch ->d_parent
compare with register
branch not taken
fetch ->d_seq
fetch ->d_parent
compare with register
branch (expectedly) not taken
which is the same as the mainline in terms of actual memory accesses and extra
3 insns. I suspect that the win on the entries with ->d_parent mismatch can
outweight that, but that needs profiling to verify.
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
