On Wed, Feb 11, 2015 at 01:26:41PM +0100, Marc Sune wrote: > > On 11/02/15 02:54, Zhang, Helin wrote: > > > >>-----Original Message----- > >>From: Richardson, Bruce > >>Sent: Tuesday, February 10, 2015 9:24 PM > >>To: Marc Sune > >>Cc: dev at dpdk.org; Zhang, Helin > >>Subject: Re: [dpdk-dev] [PATCH] Adding RTE_KNI_PREEMPT configuration > >>option > >> > >>On Tue, Feb 10, 2015 at 12:59:29PM +0100, Marc Sune wrote: > >>>This patch of Nov 2014 hasn't been yet ACKed/NACKed. Could someone > >>>please give some quick feedback? > >>> > >>>Thanks > >>>marc > >>> > >>>On 07/11/14 12:00, Marc Sune wrote: > >>>>This patch introduces CONFIG_RTE_KNI_PREEMPT flag. When set to 'no', > >>>>KNI kernel thread(s) do not call schedule_timeout_interruptible(), > >>>>which improves overall KNI performance at the expense of CPU cycles > >>(polling). > >>>>Default values is 'yes', maintaining the same behaviour as of now. > >>>> > >>>>Signed-off-by: Marc Sune <marc.sune at bisdn.de> > >>Although a better option would be to have a runtime setting, this is still > >>an > >>improvement over what we have. > >> > >>Acked-by: Bruce Richardson <bruce.richardson at intel.com> > >> > >>>>--- > >>>> config/common_linuxapp | 1 + > >>>> lib/librte_eal/linuxapp/kni/kni_misc.c | 4 ++++ > >>>> 2 files changed, 5 insertions(+) > >>>> > >>>>diff --git a/config/common_linuxapp b/config/common_linuxapp index > >>>>57b61c9..24b529d 100644 > >>>>--- a/config/common_linuxapp > >>>>+++ b/config/common_linuxapp > >>>>@@ -380,6 +380,7 @@ CONFIG_RTE_LIBRTE_PIPELINE=y > >>>> # Compile librte_kni > >>>> # > >>>> CONFIG_RTE_LIBRTE_KNI=y > >>>>+CONFIG_RTE_KNI_PREEMPT=y > >>>> CONFIG_RTE_KNI_KO_DEBUG=n > >>>> CONFIG_RTE_KNI_VHOST=n > >>>> CONFIG_RTE_KNI_VHOST_MAX_CACHE_SIZE=1024 > >>>>diff --git a/lib/librte_eal/linuxapp/kni/kni_misc.c > >>>>b/lib/librte_eal/linuxapp/kni/kni_misc.c > >>>>index ba77776..e7e6c27 100644 > >>>>--- a/lib/librte_eal/linuxapp/kni/kni_misc.c > >>>>+++ b/lib/librte_eal/linuxapp/kni/kni_misc.c > >>>>@@ -229,9 +229,11 @@ kni_thread_single(void *unused) > >>>> } > >>>> } > >>>> up_read(&kni_list_lock); > >>>>+#ifdef RTE_KNI_PREEMPT > >>>> /* reschedule out for a while */ > >>>> schedule_timeout_interruptible(usecs_to_jiffies( \ > >>>> KNI_KTHREAD_RESCHEDULE_INTERVAL)); > >>>>+#endif > >>>> } > >>>> return 0; > >>>>@@ -252,8 +254,10 @@ kni_thread_multiple(void *param) > >>>> #endif > >>>> kni_net_poll_resp(dev); > >>>> } > >>>>+#ifdef RTE_KNI_PREEMPT > >>>> schedule_timeout_interruptible(usecs_to_jiffies( \ > >>>> KNI_KTHREAD_RESCHEDULE_INTERVAL)); > >>>>+#endif > >>>> } > >>>> return 0; > >As Bruce indicated, it would be better to do that at runtime, we can > >add a config in struct rte_kni_conf, which will be copied to > >struct rte_kni_device_info, then kernel space will know the configuration. > >This way, we can enable/disable PREEMPT during KNI instance allocation time. > > As I said before, I see the point on having it at runtime and I agree. > > However, rte_kni_device_info is a struct that is per interface, not for the > entire KNI subsystem, so it is kind of abusing to add a flag there when only > will be used once, at bootstrap. To do it "properly" we should create a new > ioctl() call IMHO. > > >Anyway, it can be done now or later. > > So, do we integrate the current patch, as acked by Bruce before? > > Marc
I'd like to see it go in, as a step forward, rather than not having it at all because it's not quite perfect. One suggestion might be to have the compile time setting be "CONFIG_RTE_KNI_PREEMPT_DEFAULT" rather than just "CONFIG_RTE_KNI_PREEMPT". This would mean that we would not need to remove the setting later if we do add a runtime option, as the setting only specified the default value rather than the final value. :-) /Bruce
