Skip site navigation (1)Skip section navigation (2)
Date:      Fri, 24 Jan 2020 10:24:53 +0100
From:      Hans Petter Selasky <hps@selasky.org>
To:        Gleb Smirnoff <glebius@FreeBSD.org>, src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-head@freebsd.org
Subject:   Re: svn commit: r357004 - in head/sys: kern sys
Message-ID:  <23f272a4-c997-a454-19d6-10392713e71f@selasky.org>
In-Reply-To: <fb6e84b6-85a9-3252-c27a-bb9ee7ca203f@selasky.org>
References:  <202001230124.00N1OlXi029506@repo.freebsd.org> <fb6e84b6-85a9-3252-c27a-bb9ee7ca203f@selasky.org>

next in thread | previous in thread | raw e-mail | index | archive | help
On 2020-01-23 09:36, Hans Petter Selasky wrote:
> On 2020-01-23 02:24, Gleb Smirnoff wrote:
>> Author: glebius
>> Date: Thu Jan 23 01:24:47 2020
>> New Revision: 357004
>> URL: https://svnweb.freebsd.org/changeset/base/357004
>>
>> Log:
>>    Enter the network epoch for interrupt handlers of INTR_TYPE_NET.
>>    Provide tunable to limit how many times handlers may be executed
>>    without reentering epoch.
>>    Differential Revision:    https://reviews.freebsd.org/D23242
>>
>> Modified:
>>    head/sys/kern/kern_intr.c
>>    head/sys/sys/interrupt.h
>>
>> Modified: head/sys/kern/kern_intr.c
>> ============================================================================== 
>>
>> --- head/sys/kern/kern_intr.c    Thu Jan 23 01:20:59 2020    (r357003)
>> +++ head/sys/kern/kern_intr.c    Thu Jan 23 01:24:47 2020    (r357004)
>> @@ -48,6 +48,7 @@ __FBSDID("$FreeBSD$");
>>   #include <sys/mutex.h>
>>   #include <sys/priv.h>
>>   #include <sys/proc.h>
>> +#include <sys/epoch.h>
>>   #include <sys/random.h>
>>   #include <sys/resourcevar.h>
>>   #include <sys/sched.h>
>> @@ -94,6 +95,9 @@ static int intr_storm_threshold = 0;
>>   SYSCTL_INT(_hw, OID_AUTO, intr_storm_threshold, CTLFLAG_RWTUN,
>>       &intr_storm_threshold, 0,
>>       "Number of consecutive interrupts before storm protection is 
>> enabled");
>> +static int intr_epoch_batch = 1000;
>> +SYSCTL_INT(_hw, OID_AUTO, intr_epoch_batch, CTLFLAG_RWTUN, 
>> &intr_epoch_batch,
>> +    0, "Maximum interrupt handler executions without re-entering 
>> epoch(9)");
>>   static TAILQ_HEAD(, intr_event) event_list =
>>       TAILQ_HEAD_INITIALIZER(event_list);
>>   static struct mtx event_lock;
>> @@ -587,6 +591,8 @@ intr_event_add_handler(struct intr_event *ie, 
>> const ch
>>           ih->ih_flags |= IH_MPSAFE;
>>       if (flags & INTR_ENTROPY)
>>           ih->ih_flags |= IH_ENTROPY;
>> +    if (flags & INTR_TYPE_NET)
>> +        ih->ih_flags |= IH_NET;
>>       /* We can only have one exclusive handler in a event. */
>>       mtx_lock(&ie->ie_lock);
>> @@ -1196,11 +1202,12 @@ ithread_execute_handlers(struct proc *p, 
>> struct intr_e
>>   static void
>>   ithread_loop(void *arg)
>>   {
>> +    struct epoch_tracker et;
>>       struct intr_thread *ithd;
>>       struct intr_event *ie;
>>       struct thread *td;
>>       struct proc *p;
>> -    int wake;
>> +    int wake, epoch_count;
>>       td = curthread;
>>       p = td->td_proc;
>> @@ -1235,8 +1242,21 @@ ithread_loop(void *arg)
>>            * that the load of ih_need in ithread_execute_handlers()
>>            * is ordered after the load of it_need here.
>>            */
>> -        while (atomic_cmpset_acq_int(&ithd->it_need, 1, 0) != 0)
>> +        if (ie->ie_hflags & IH_NET) {
>> +            epoch_count = 0;
>> +            NET_EPOCH_ENTER(et);
>> +        }
>> +        while (atomic_cmpset_acq_int(&ithd->it_need, 1, 0) != 0) {
>>               ithread_execute_handlers(p, ie);
>> +            if ((ie->ie_hflags & IH_NET) &&
>> +                ++epoch_count >= intr_epoch_batch) {
>> +                NET_EPOCH_EXIT(et);
>> +                epoch_count = 0;
>> +                NET_EPOCH_ENTER(et);
>> +            }
>> +        }
>> +        if (ie->ie_hflags & IH_NET)
>> +            NET_EPOCH_EXIT(et);
>>           WITNESS_WARN(WARN_PANIC, NULL, "suspending ithread");
>>           mtx_assert(&Giant, MA_NOTOWNED);

Hi Gleb,

What you want to do here is right, but how it is implemented is wrong, 
in my opinion.

1) Remove intr_epoch_batch. Most network drivers use interrupt 
moderation, and a timeout of 1000 iterations can easily become 1 second 
under heavly load !

2) You need to make a new request function for interrupts which take a 
pointer to an EPOCH and replace that IH_NET in hflags!

--HPS



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?23f272a4-c997-a454-19d6-10392713e71f>