Date:   Sun, 2 Aug 2020 16:27:53 +0200
From:   Sebastian Gottschall <s.gottschall@...wrt.com>
To:     Eric Dumazet <eric.dumazet@...il.com>,
        Felix Fietkau <nbd@....name>, netdev@...r.kernel.org
Cc:     Hillf Danton <hdanton@...a.com>
Subject: Re: [PATCH] net: add support for threaded NAPI polling


Am 31.07.2020 um 18:36 schrieb Eric Dumazet:
>
> On 7/30/20 10:21 AM, Sebastian Gottschall wrote:
>> Am 30.07.2020 um 18:08 schrieb Eric Dumazet:
>>> On 7/30/20 7:30 AM, Sebastian Gottschall wrote:
>>>> Am 29.07.2020 um 19:44 schrieb Eric Dumazet:
>>>>> On 7/29/20 9:50 AM, Felix Fietkau wrote:
>>>>>> For some drivers (especially 802.11 drivers), doing a lot of work in the NAPI
>>>>>> poll function does not perform well. Since NAPI poll is bound to the CPU it
>>>>>> was scheduled from, we can easily end up with a few very busy CPUs spending
>>>>>> most of their time in softirq/ksoftirqd and some idle ones.
>>>>>>
>>>>>> Introduce threaded NAPI for such drivers based on a workqueue. The API is the
>>>>>> same except for using netif_threaded_napi_add instead of netif_napi_add.
>>>>>>
>>>>>> In my tests with mt76 on MT7621 using threaded NAPI + a thread for tx scheduling
>>>>>> improves LAN->WLAN bridging throughput by 10-50%. Throughput without threaded
>>>>>> NAPI is wildly inconsistent, depending on the CPU that runs the tx scheduling
>>>>>> thread.
>>>>>>
>>>>>> With threaded NAPI, throughput seems stable and consistent (and higher than
>>>>>> the best results I got without it).
>>>>>>
>>>>>> Based on a patch by Hillf Danton
>>>>>>
>>>>>> Cc: Hillf Danton <hdanton@...a.com>
>>>>>> Signed-off-by: Felix Fietkau <nbd@....name>
>>>>>> ---
>>>>>> Changes since RFC v2:
>>>>>> - fix unused but set variable reported by kbuild test robot
>>>>>>
>>>>>> Changes since RFC:
>>>>>> - disable softirq around threaded poll functions
>>>>>> - reuse most parts of napi_poll()
>>>>>> - fix re-schedule condition
>>>>>>
>>>>>>     include/linux/netdevice.h |  23 ++++++
>>>>>>     net/core/dev.c            | 162 ++++++++++++++++++++++++++------------
>>>>>>     2 files changed, 133 insertions(+), 52 deletions(-)
>>>>>>
>>>>>> diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
>>>>>> index ac2cd3f49aba..3a39211c7598 100644
>>>>>> --- a/include/linux/netdevice.h
>>>>>> +++ b/include/linux/netdevice.h
>>>>>> @@ -347,6 +347,7 @@ struct napi_struct {
>>>>>>         struct list_head    dev_list;
>>>>>>         struct hlist_node    napi_hash_node;
>>>>>>         unsigned int        napi_id;
>>>>>> +    struct work_struct    work;
>>>>>>     };
>>>>>>       enum {
>>>>>> @@ -357,6 +358,7 @@ enum {
>>>>>>         NAPI_STATE_HASHED,    /* In NAPI hash (busy polling possible) */
>>>>>>         NAPI_STATE_NO_BUSY_POLL,/* Do not add in napi_hash, no busy polling */
>>>>>>         NAPI_STATE_IN_BUSY_POLL,/* sk_busy_loop() owns this NAPI */
>>>>>> +    NAPI_STATE_THREADED,    /* Use threaded NAPI */
>>>>>>     };
>>>>>>       enum {
>>>>>> @@ -367,6 +369,7 @@ enum {
>>>>>>         NAPIF_STATE_HASHED     = BIT(NAPI_STATE_HASHED),
>>>>>>         NAPIF_STATE_NO_BUSY_POLL = BIT(NAPI_STATE_NO_BUSY_POLL),
>>>>>>         NAPIF_STATE_IN_BUSY_POLL = BIT(NAPI_STATE_IN_BUSY_POLL),
>>>>>> +    NAPIF_STATE_THREADED     = BIT(NAPI_STATE_THREADED),
>>>>>>     };
>>>>>>       enum gro_result {
>>>>>> @@ -2315,6 +2318,26 @@ static inline void *netdev_priv(const struct net_device *dev)
>>>>>>     void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
>>>>>>                 int (*poll)(struct napi_struct *, int), int weight);
>>>>>>     +/**
>>>>>> + *    netif_threaded_napi_add - initialize a NAPI context
>>>>>> + *    @dev:  network device
>>>>>> + *    @napi: NAPI context
>>>>>> + *    @poll: polling function
>>>>>> + *    @weight: default weight
>>>>>> + *
>>>>>> + * This variant of netif_napi_add() should be used from drivers using NAPI
>>>>>> + * with CPU intensive poll functions.
>>>>>> + * This will schedule polling from a high priority workqueue that
>>>>>> + */
>>>>>> +static inline void netif_threaded_napi_add(struct net_device *dev,
>>>>>> +                       struct napi_struct *napi,
>>>>>> +                       int (*poll)(struct napi_struct *, int),
>>>>>> +                       int weight)
>>>>>> +{
>>>>>> +    set_bit(NAPI_STATE_THREADED, &napi->state);
>>>>>> +    netif_napi_add(dev, napi, poll, weight);
>>>>>> +}
>>>>>> +
>>>>>>     /**
>>>>>>      *    netif_tx_napi_add - initialize a NAPI context
>>>>>>      *    @dev:  network device
>>>>>> diff --git a/net/core/dev.c b/net/core/dev.c
>>>>>> index 19f1abc26fcd..11b027f3a2b9 100644
>>>>>> --- a/net/core/dev.c
>>>>>> +++ b/net/core/dev.c
>>>>>> @@ -158,6 +158,7 @@ static DEFINE_SPINLOCK(offload_lock);
>>>>>>     struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly;
>>>>>>     struct list_head ptype_all __read_mostly;    /* Taps */
>>>>>>     static struct list_head offload_base __read_mostly;
>>>>>> +static struct workqueue_struct *napi_workq __read_mostly;
>>>>>>       static int netif_rx_internal(struct sk_buff *skb);
>>>>>>     static int call_netdevice_notifiers_info(unsigned long val,
>>>>>> @@ -6286,6 +6287,11 @@ void __napi_schedule(struct napi_struct *n)
>>>>>>     {
>>>>>>         unsigned long flags;
>>>>>>     +    if (test_bit(NAPI_STATE_THREADED, &n->state)) {
>>>>>> +        queue_work(napi_workq, &n->work);
>>>>>> +        return;
>>>>>> +    }
>>>>>> +
>>>>> Where is the corresponding cancel_work_sync() or flush_work() at device dismantle ?
>>>>>
>>>>> Just hoping the thread will eventually run seems optimistic to me.
>>>>>
>>>>>
>>>>> Quite frankly, I do believe this STATE_THREADED status should be a generic NAPI attribute
>>>>> that can be changed dynamically, at admin request, instead of having to change/recompile
>>>>> a driver.
>>>> thats not that easy. wifi devices do use dummy netdev devices. they are not visible to sysfs and other administrative options.
>>>> so changing it would just be possible if a special mac80211 based control would be implemented for these drivers.
>>>> for standard netdev devices it isnt a big thing to implement a administrative control by sysfs (if you are talking about such a feature)
>>> We do not want to add code in fast path only for one device. We need something truly generic.
>>>
>>> I am not saying only the admin can chose, it is fine if a driver does not give the choice
>>> and will simply call netif_threaded_napi_add()
>> what could make sense if the feature can be disabled / enabled, but it will only affect drivers using the netif_threaded_napi_add call, but it should not affect drivers
>> using the old api in any way since not all drivers will work with this feature.
>
> If we provide something in core NAPI stack, we want to make sure we can test/use it with other drivers.
>
> ethtool, or a /sys/class/net/ethXXX entry could be used.
but this doesnt work for wifi drivers. since wifi drivers are using 
dummy netdev devices. we are running in circles here
i mean a sane way could be also that dummy netdev devices are present in 
sysfs too which is not the case right now.
so changing the api, so the driver is forced to set sane virtual dummy 
netdev name (like the driver name for instance). so it can be accessed 
by sysfs.
>
> The argument about not affecting other drivers is misleading, since the patch adds another conditional test in
> standard NAPI layer.
>
> Lets keep NAPI generic please.
>
> Lets make sure syzbot will find bugs without having to attach a specific mac80211 hardware.
the patch is not mac80211 specific. i  tested it already with network 
drivers. it is generic.
>
> Another concern I have with this patch is that we no longer can contain NIC processing is done
> on a selected set of cpus (as commanded in /proc/irq/XXX/smp_affinity).
> Or can we ?
i had this discussion already with felix in a phonecall last week. 
kthread vs. workq. his oppinion is that workq works more effective than 
kthread's
since kthreads required application support for good balancing like 
irqbalance. personally i have no real oppinion here. the good point on 
kthreads is
that i'm able to track the system load per thread with simple process 
watching. and its possible to force the thread on a specific cpuset.

the good thing on workq is, its more simple to implement and usually 
more risc free, even if i havent seen any problems with kthreads.
maybe felix should say something here about this. the background of this 
patch is simply performance. especially on embedded devices.
it started with a ath10k patch which was introducing napi threading for 
a specific chipset which leaded to some research on my side until i found
a historic napi kthread patch from 2016 which you have denied at that 
time. i tested that patch and saw heavy performance boost for ath10k.
which leaded to this workq patch at the end

see also this discussion here. (you may remember this discussion since 
you where involved in it)
https://www.mail-archive.com/linux-kernel@vger.kernel.org/msg1142611.html

Sebastian

>
>

Powered by blists - more mailing lists