From: Yonghong Song <yhs@fb.com>
To: Andrii Nakryiko <andrii.nakryiko@gmail.com>
Cc: Song Liu <songliubraving@fb.com>, bpf <bpf@vger.kernel.org>,
Networking <netdev@vger.kernel.org>,
"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
"mingo@redhat.com" <mingo@redhat.com>,
"peterz@infradead.org" <peterz@infradead.org>,
"ast@kernel.org" <ast@kernel.org>,
"daniel@iogearbox.net" <daniel@iogearbox.net>,
"andrii@kernel.org" <andrii@kernel.org>,
"john.fastabend@gmail.com" <john.fastabend@gmail.com>,
"kpsingh@chromium.org" <kpsingh@chromium.org>,
Kernel Team <Kernel-team@fb.com>,
"haoluo@google.com" <haoluo@google.com>
Subject: Re: [PATCH bpf-next 4/4] bpf: runqslower: use task local storage
Date: Mon, 11 Jan 2021 23:33:42 -0800 [thread overview]
Message-ID: <8d9983c4-2842-e2f8-94ce-1676977bb720@fb.com> (raw)
In-Reply-To: <CAEf4BzZivGBmDbUxfiDwAC3aFoTWNfyWaiZRA4Vu16ZT9kzE8A@mail.gmail.com>
On 1/11/21 11:14 PM, Andrii Nakryiko wrote:
> On Mon, Jan 11, 2021 at 7:24 PM Yonghong Song <yhs@fb.com> wrote:
>>
>>
>>
>> On 1/11/21 2:54 PM, Song Liu wrote:
>>>
>>>
>>>> On Jan 11, 2021, at 9:49 AM, Yonghong Song <yhs@fb.com> wrote:
>>>>
>>>>
>>>>
>>>> On 1/8/21 3:19 PM, Song Liu wrote:
>>>>> Replace hashtab with task local storage in runqslower. This improves the
>>>>> performance of these BPF programs. The following table summarizes average
>>>>> runtime of these programs, in nanoseconds:
>>>>> task-local hash-prealloc hash-no-prealloc
>>>>> handle__sched_wakeup 125 340 3124
>>>>> handle__sched_wakeup_new 2812 1510 2998
>>>>> handle__sched_switch 151 208 991
>>>>> Note that, task local storage gives better performance than hashtab for
>>>>> handle__sched_wakeup and handle__sched_switch. On the other hand, for
>>>>> handle__sched_wakeup_new, task local storage is slower than hashtab with
>>>>> prealloc. This is because handle__sched_wakeup_new accesses the data for
>>>>> the first time, so it has to allocate the data for task local storage.
>>>>> Once the initial allocation is done, subsequent accesses, as those in
>>>>> handle__sched_wakeup, are much faster with task local storage. If we
>>>>> disable hashtab prealloc, task local storage is much faster for all 3
>>>>> functions.
>>>>> Signed-off-by: Song Liu <songliubraving@fb.com>
>>>>> ---
>>>>> tools/bpf/runqslower/runqslower.bpf.c | 26 +++++++++++++++-----------
>>>>> 1 file changed, 15 insertions(+), 11 deletions(-)
>>>>> diff --git a/tools/bpf/runqslower/runqslower.bpf.c b/tools/bpf/runqslower/runqslower.bpf.c
>>>>> index 1f18a409f0443..c4de4179a0a17 100644
>>>>> --- a/tools/bpf/runqslower/runqslower.bpf.c
>>>>> +++ b/tools/bpf/runqslower/runqslower.bpf.c
>>>>> @@ -11,9 +11,9 @@ const volatile __u64 min_us = 0;
>>>>> const volatile pid_t targ_pid = 0;
>>>>> struct {
>>>>> - __uint(type, BPF_MAP_TYPE_HASH);
>>>>> - __uint(max_entries, 10240);
>>>>> - __type(key, u32);
>>>>> + __uint(type, BPF_MAP_TYPE_TASK_STORAGE);
>>>>> + __uint(map_flags, BPF_F_NO_PREALLOC);
>>>>> + __type(key, int);
>>>>> __type(value, u64);
>>>>> } start SEC(".maps");
>>>>> @@ -25,15 +25,19 @@ struct {
>>>>> /* record enqueue timestamp */
>>>>> __always_inline
>>>>> -static int trace_enqueue(u32 tgid, u32 pid)
>>>>> +static int trace_enqueue(struct task_struct *t)
>>>>> {
>>>>> - u64 ts;
>>>>> + u32 pid = t->pid;
>>>>> + u64 ts, *ptr;
>>>>> if (!pid || (targ_pid && targ_pid != pid))
>>>>> return 0;
>>>>> ts = bpf_ktime_get_ns();
>>>>> - bpf_map_update_elem(&start, &pid, &ts, 0);
>>>>> + ptr = bpf_task_storage_get(&start, t, 0,
>>>>> + BPF_LOCAL_STORAGE_GET_F_CREATE);
>>>>> + if (ptr)
>>>>> + *ptr = ts;
>>>>> return 0;
>>>>> }
>>>>> @@ -43,7 +47,7 @@ int handle__sched_wakeup(u64 *ctx)
>>>>> /* TP_PROTO(struct task_struct *p) */
>>>>> struct task_struct *p = (void *)ctx[0];
>>>>> - return trace_enqueue(p->tgid, p->pid);
>>>>> + return trace_enqueue(p);
>>>>> }
>>>>> SEC("tp_btf/sched_wakeup_new")
>>>>> @@ -52,7 +56,7 @@ int handle__sched_wakeup_new(u64 *ctx)
>>>>> /* TP_PROTO(struct task_struct *p) */
>>>>> struct task_struct *p = (void *)ctx[0];
>>>>> - return trace_enqueue(p->tgid, p->pid);
>>>>> + return trace_enqueue(p);
>>>>> }
>>>>> SEC("tp_btf/sched_switch")
>>>>> @@ -70,12 +74,12 @@ int handle__sched_switch(u64 *ctx)
>>>>> /* ivcsw: treat like an enqueue event and store timestamp */
>>>>> if (prev->state == TASK_RUNNING)
>>>>> - trace_enqueue(prev->tgid, prev->pid);
>>>>> + trace_enqueue(prev);
>>>>> pid = next->pid;
>>>>> /* fetch timestamp and calculate delta */
>>>>> - tsp = bpf_map_lookup_elem(&start, &pid);
>>>>> + tsp = bpf_task_storage_get(&start, next, 0, 0);
>>>>> if (!tsp)
>>>>> return 0; /* missed enqueue */
>>>>
>>>> Previously, hash table may overflow so we may have missed enqueue.
>>>> Here with task local storage, is it possible to add additional pid
>>>> filtering in the beginning of handle__sched_switch such that
>>>> missed enqueue here can be treated as an error?
>>>
>>> IIUC, hashtab overflow is not the only reason of missed enqueue. If the
>>> wakeup (which calls trace_enqueue) happens before runqslower starts, we
>>> may still get missed enqueue in sched_switch, no?
>>
>> the wakeup won't happen before runqslower starts since runqslower needs
>> to start to do attachment first and then trace_enqueue() can run.
>
> I think Song is right. Given wakeup and sched_switch need to be
> matched, depending at which exact time we attach BPF programs, we can
> end up missing wakeup, but not missing sched_switch, no? So it's not
> an error.
The current approach works fine. What I suggested is to
tighten sched_switch only for target_pid. wakeup (doing queuing) will
be more relaxed than sched_switch to ensure task local storage creation
is always there for target_pid regardless of attachment timing.
I think it should work, but we have to experiment to see actual
results...
>
>>
>> For the current implementation trace_enqueue() will happen for any non-0
>> pid before setting test_progs tgid, and will happen for any non-0 and
>> test_progs tgid if it is set, so this should be okay if we do filtering
>> in handle__sched_switch. Maybe you can do an experiment to prove whether
>> my point is correct or not.
>>
>>>
>>> Thanks,
>>> Song
>>>
next prev parent reply other threads:[~2021-01-12 7:35 UTC|newest]
Thread overview: 29+ messages / expand[flat|nested] mbox.gz Atom feed top
[not found] <20210108231950.3844417-1-songliubraving@fb.com>
[not found] ` <20210108231950.3844417-2-songliubraving@fb.com>
2021-01-11 6:27 ` [PATCH bpf-next 1/4] bpf: enable task local storage for tracing programs Yonghong Song
2021-01-11 10:17 ` KP Singh
2021-01-11 15:56 ` Yonghong Song
2021-01-11 10:14 ` KP Singh
2021-01-11 23:16 ` Song Liu
2021-01-11 17:16 ` Yonghong Song
2021-01-11 18:56 ` Martin KaFai Lau
2021-01-11 21:35 ` KP Singh
2021-01-11 21:58 ` Martin KaFai Lau
2021-01-11 23:45 ` Song Liu
2021-01-12 16:32 ` Yonghong Song
2021-01-12 16:53 ` KP Singh
2021-01-15 23:34 ` Song Liu
2021-01-16 0:55 ` Yonghong Song
2021-01-16 1:12 ` Song Liu
2021-01-16 1:50 ` Yonghong Song
2021-01-11 23:41 ` Song Liu
2021-01-12 18:21 ` Martin KaFai Lau
[not found] ` <20210108231950.3844417-4-songliubraving@fb.com>
2021-01-11 17:37 ` [PATCH bpf-next 3/4] bpf: runqslower: prefer use local vmlinux Yonghong Song
[not found] ` <20210108231950.3844417-5-songliubraving@fb.com>
2021-01-11 17:49 ` [PATCH bpf-next 4/4] bpf: runqslower: use task local storage Yonghong Song
2021-01-11 22:54 ` Song Liu
2021-01-12 3:24 ` Yonghong Song
2021-01-12 7:14 ` Andrii Nakryiko
2021-01-12 7:33 ` Yonghong Song [this message]
[not found] ` <20210108231950.3844417-3-songliubraving@fb.com>
2021-01-11 17:30 ` [PATCH bpf-next 2/4] selftests/bpf: add non-BPF_LSM test for " Yonghong Song
2021-01-11 17:44 ` KP Singh
2021-01-11 22:50 ` Song Liu
2021-01-11 22:49 ` Song Liu
2021-01-12 7:06 ` Andrii Nakryiko
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=8d9983c4-2842-e2f8-94ce-1676977bb720@fb.com \
--to=yhs@fb.com \
--cc=Kernel-team@fb.com \
--cc=andrii.nakryiko@gmail.com \
--cc=andrii@kernel.org \
--cc=ast@kernel.org \
--cc=bpf@vger.kernel.org \
--cc=daniel@iogearbox.net \
--cc=haoluo@google.com \
--cc=john.fastabend@gmail.com \
--cc=kpsingh@chromium.org \
--cc=linux-kernel@vger.kernel.org \
--cc=mingo@redhat.com \
--cc=netdev@vger.kernel.org \
--cc=peterz@infradead.org \
--cc=songliubraving@fb.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).