On 24/09/2016 22:44, Richard Henderson wrote: > On 09/24/2016 04:51 AM, Paolo Bonzini wrote: >> >> >> ----- Original Message ----- >>> From: "Richard Henderson" <[email protected]> >>> To: "Paolo Bonzini" <[email protected]>, [email protected] >>> Cc: "serge fdrv" <[email protected]>, [email protected], "alex >>> bennee" <[email protected]>, "sergey fedorov" >>> <[email protected]> >>> Sent: Friday, September 23, 2016 8:06:09 PM >>> Subject: Re: [Qemu-devel] [PATCH 15/16] tcg: Make tb_flush() thread safe >>> >>> On 09/23/2016 12:31 AM, Paolo Bonzini wrote: >>>> + unsigned tb_flush_req = (unsigned) (uintptr_t) data; >>> >>> Extra cast? >>> >>>> - tcg_ctx.tb_ctx.tb_flush_count++; >>>> + atomic_inc(&tcg_ctx.tb_ctx.tb_flush_count); >>> >>> Since this is the only place this value is incremented, and we're >>> under a >>> lock, >>> it should be cheaper to use >>> >>> atomic_mb_set(&tcg_ctx.tb_ctx.tb_flush_count, tb_flush_req + 1); >> >> atomic_set will do even. Though it's not really a fast path, which is >> why I went for atomic_inc. > > Don't we need the flush to be complete before the new count is seen? > That's why I was suggesting the mb_set.
You're right in that the mb_set is more correct, or even better would be a store-release. Actually even atomic_set works, though in a fairly surprising manner. This is because the final check is done by do_tb_flush under the mutex, so the final check does wait for the flush to be complete. If tb_flush_count is exposed too early to tb_flush, all that can happen is that do_tb_flush sees a tb_flush_req that's more recent than it should. do_tb_flush then incorrectly redoes the flush, but that's not a disaster. But I'm changing it to mb_set as you suggested. Paolo
