[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-devel] [RFC v3 19/19] cpu-exec: remove tb_lock from the hot-pa
From: |
Sergey Fedorov |
Subject: |
Re: [Qemu-devel] [RFC v3 19/19] cpu-exec: remove tb_lock from the hot-path |
Date: |
Wed, 29 Jun 2016 17:52:46 +0300 |
User-agent: |
Mozilla/5.0 (X11; Linux x86_64; rv:38.0) Gecko/20100101 Thunderbird/38.8.0 |
On 29/06/16 17:47, Alex Bennée wrote:
> Sergey Fedorov <address@hidden> 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 <address@hidden>
>>>
>>> ---
>>> 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
- [Qemu-devel] [RFC v3 10/19] tcg: cpus rm tcg_exec_all(), (continued)
- [Qemu-devel] [RFC v3 10/19] tcg: cpus rm tcg_exec_all(), Alex Bennée, 2016/06/03
- [Qemu-devel] [RFC v3 18/19] tcg: Ensure safe TB lookup out of 'tb_lock', Alex Bennée, 2016/06/03
- [Qemu-devel] [RFC v3 17/19] tcg: enable thread-per-vCPU, Alex Bennée, 2016/06/03
- [Qemu-devel] [RFC v3 15/19] tcg: drop global lock during TCG code execution, Alex Bennée, 2016/06/03
- [Qemu-devel] [RFC v3 19/19] cpu-exec: remove tb_lock from the hot-path, Alex Bennée, 2016/06/03
[Qemu-devel] [RFC v3 16/19] tcg: move locking for tb_invalidate_phys_page_range up, Alex Bennée, 2016/06/03
Re: [Qemu-devel] [RFC v3 00/19] Base enabling patches for MTTCG, Pranith Kumar, 2016/06/04