From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from eggs.gnu.org ([2001:4830:134:3::10]:44492) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1bIGrG-0004mo-80 for qemu-devel@nongnu.org; Wed, 29 Jun 2016 10:52:55 -0400 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1bIGrB-0005Xc-8s for qemu-devel@nongnu.org; Wed, 29 Jun 2016 10:52:53 -0400 Received: from mail-lf0-x22d.google.com ([2a00:1450:4010:c07::22d]:36783) by eggs.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1bIGrA-0005XX-Sy for qemu-devel@nongnu.org; Wed, 29 Jun 2016 10:52:49 -0400 Received: by mail-lf0-x22d.google.com with SMTP id q132so35307940lfe.3 for ; Wed, 29 Jun 2016 07:52:48 -0700 (PDT) References: <87r3bg1275.fsf@linaro.org> From: Sergey Fedorov Message-ID: <5773E0BE.9090607@gmail.com> Date: Wed, 29 Jun 2016 17:52:46 +0300 MIME-Version: 1.0 In-Reply-To: <87r3bg1275.fsf@linaro.org> Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: 8bit Subject: Re: [Qemu-devel] [RFC v3 19/19] cpu-exec: remove tb_lock from the hot-path List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: =?UTF-8?Q?Alex_Benn=c3=a9e?= Cc: mttcg@listserver.greensocs.com, qemu-devel@nongnu.org, fred.konrad@greensocs.com, a.rigo@virtualopensystems.com, cota@braap.org, bobby.prani@gmail.com, mark.burton@greensocs.com, pbonzini@redhat.com, jan.kiszka@siemens.com, rth@twiddle.net, peter.maydell@linaro.org, claudio.fontana@huawei.com, Peter Crosthwaite On 29/06/16 17:47, Alex Bennée wrote: > Sergey Fedorov writes: > >> On 03/06/16 23:40, Alex Bennée wrote: >>> Lock contention in the hot path of moving between existing patched >>> TranslationBlocks is the main drag on MTTCG performance. This patch >>> pushes the tb_lock() usage down to the two places that really need it: >>> >>> - code generation (tb_gen_code) >>> - jump patching (tb_add_jump) >>> >>> The rest of the code doesn't really need to hold a lock as it is either >>> using per-CPU structures or designed to be used in concurrent read >>> situations (qht_lookup). >>> >>> Signed-off-by: Alex Bennée >>> >>> --- >>> v3 >>> - fix merge conflicts with Sergey's patch >>> --- >>> cpu-exec.c | 59 ++++++++++++++++++++++++++++++----------------------------- >>> 1 file changed, 30 insertions(+), 29 deletions(-) >>> >>> diff --git a/cpu-exec.c b/cpu-exec.c >>> index b017643..4af0b52 100644 >>> --- a/cpu-exec.c >>> +++ b/cpu-exec.c >>> @@ -298,41 +298,38 @@ static TranslationBlock *tb_find_slow(CPUState *cpu, >>> * Pairs with smp_wmb() in tb_phys_invalidate(). */ >>> smp_rmb(); >>> tb = tb_find_physical(cpu, pc, cs_base, flags); >>> - if (tb) { >>> - goto found; >>> - } >>> + if (!tb) { >>> >>> - /* mmap_lock is needed by tb_gen_code, and mmap_lock must be >>> - * taken outside tb_lock. Since we're momentarily dropping >>> - * tb_lock, there's a chance that our desired tb has been >>> - * translated. >>> - */ >>> - tb_unlock(); >>> - mmap_lock(); >>> - tb_lock(); >>> - tb = tb_find_physical(cpu, pc, cs_base, flags); >>> - if (tb) { >>> - mmap_unlock(); >>> - goto found; >>> - } >>> + /* mmap_lock is needed by tb_gen_code, and mmap_lock must be >>> + * taken outside tb_lock. >>> + */ >>> + mmap_lock(); >>> + tb_lock(); >>> >>> - /* if no translated code available, then translate it now */ >>> - tb = tb_gen_code(cpu, pc, cs_base, flags, 0); >>> + /* There's a chance that our desired tb has been translated while >>> + * taking the locks so we check again inside the lock. >>> + */ >>> + tb = tb_find_physical(cpu, pc, cs_base, flags); >>> + if (!tb) { >>> + /* if no translated code available, then translate it now */ >>> + tb = tb_gen_code(cpu, pc, cs_base, flags, 0); >>> + } >>> >>> - mmap_unlock(); >>> + tb_unlock(); >>> + mmap_unlock(); >>> + } >>> >>> -found: >>> - /* we add the TB in the virtual pc hash table */ >>> + /* We add the TB in the virtual pc hash table for the fast lookup */ >>> cpu->tb_jmp_cache[tb_jmp_cache_hash_func(pc)] = tb; >> Hmm, seems like I forgot to convert this into atomic_set() in the >> previous patch... > OK, can you fix that in your quick fixes series? Sure. I think that patch and this are both ready-to-go into mainline. > >>> return tb; >>> } >>> >>> static inline TranslationBlock *tb_find_fast(CPUState *cpu, >>> - TranslationBlock **last_tb, >>> + TranslationBlock **ltbp, >> I'm not sure if it is more readable... > I'll revert. I was trying to keep line lengths short :-/ > >>> int tb_exit) >>> { >>> CPUArchState *env = (CPUArchState *)cpu->env_ptr; >>> - TranslationBlock *tb; >>> + TranslationBlock *tb, *last_tb; >>> target_ulong cs_base, pc; >>> uint32_t flags; >>> >>> @@ -340,7 +337,6 @@ static inline TranslationBlock *tb_find_fast(CPUState *cpu, >>> always be the same before a given translated block >>> is executed. */ >>> cpu_get_tb_cpu_state(env, &pc, &cs_base, &flags); >>> - tb_lock(); >>> tb = atomic_read(&cpu->tb_jmp_cache[tb_jmp_cache_hash_func(pc)]); >>> if (unlikely(!tb || tb->pc != pc || tb->cs_base != cs_base || >>> tb->flags != flags)) { >>> @@ -350,7 +346,7 @@ static inline TranslationBlock *tb_find_fast(CPUState *cpu, >>> /* Ensure that no TB jump will be modified as the >>> * translation buffer has been flushed. >>> */ >>> - *last_tb = NULL; >>> + *ltbp = NULL; >>> cpu->tb_flushed = false; >>> } >>> #ifndef CONFIG_USER_ONLY >>> @@ -359,14 +355,19 @@ static inline TranslationBlock *tb_find_fast(CPUState *cpu, >>> * spanning two pages because the mapping for the second page can change. >>> */ >>> if (tb->page_addr[1] != -1) { >>> - *last_tb = NULL; >>> + *ltbp = NULL; >>> } >>> #endif >>> + >>> /* See if we can patch the calling TB. */ >>> - if (*last_tb && !qemu_loglevel_mask(CPU_LOG_TB_NOCHAIN)) { >>> - tb_add_jump(*last_tb, tb_exit, tb); >>> + last_tb = *ltbp; >>> + if (!qemu_loglevel_mask(CPU_LOG_TB_NOCHAIN) && >>> + last_tb && >>> + !last_tb->jmp_list_next[tb_exit]) { >> If we're going to check this outside of tb_lock we have to do this with >> atomic_{read,set}(). However, I think it is rare case to race on >> tb_add_jump() so probably it is okay to do the check under tb_lock. > It's checking for NULL, it gets re-checked in tb_add_jump while under > lock so I the setting race should be OK I think. I think we could just skip this check and be fine. What do you think regarding this? Thanks, Sergey > >>> + tb_lock(); >>> + tb_add_jump(last_tb, tb_exit, tb); >>> + tb_unlock(); >>> } >>> - tb_unlock(); >>> return tb; >>> } >>> >> Kind regards, >> Sergey > > -- > Alex Bennée