From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S261636AbVFFT3A (ORCPT ); Mon, 6 Jun 2005 15:29:00 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S261644AbVFFTYt (ORCPT ); Mon, 6 Jun 2005 15:24:49 -0400 Received: from fmr20.intel.com ([134.134.136.19]:13990 "EHLO orsfmr005.jf.intel.com") by vger.kernel.org with ESMTP id S261640AbVFFTXD (ORCPT ); Mon, 6 Jun 2005 15:23:03 -0400 Message-Id: <20050606192113.307745000@araj-em64t> References: <20050606191433.104273000@araj-em64t> Date: Mon, 06 Jun 2005 12:14:37 -0700 From: Ashok Raj To: Andrew Morton , linux-kernel@vger.kernel.org Cc: Zwane Mwaikambo , Srivattsa Vaddagiri , discuss@x86-64.org, Rusty Russell , Ashok Raj , Andi Kleen Subject: [patch 4/5] try2: x86_64: Dont use broadcast shortcut to make it cpu hotplug safe. Content-Disposition: inline; filename=no_broadcast_ipi.patch Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org Broadcast IPI's provide un-expected behaviour for cpu hotplug. CPU's in offline state also end up receiving the IPI. Once the cpus become online they receive these stale IPI's which are bad and introduce unexpected behaviour. This is easily avoided by not sending a broadcast and addressing just the CPU's in online map. Doing prelim cycle counts it appears there is no big overhead and numbers seem around 0x3000-0x3900 on an average on x86 and x86_64 systems with CPUS running 3G, both for broadcast and mask version of the API's. The shortcuts are useful only for flat mode (where the perf shows no degradation), and in cluster mode, its unicast anyway. Its simpler to just not use broadcast anymore. Signed-off-by: Ashok Raj Acked-by: Andi Kleen Acked-by: Zwane Mwaikambo ----------------------------------------- arch/x86_64/kernel/genapic_flat.c | 41 +++++++++++++++++++++++--------------- 1 files changed, 25 insertions(+), 16 deletions(-) Index: linux-2.6.12-rc5-mm2/arch/x86_64/kernel/genapic_flat.c =================================================================== --- linux-2.6.12-rc5-mm2.orig/arch/x86_64/kernel/genapic_flat.c +++ linux-2.6.12-rc5-mm2/arch/x86_64/kernel/genapic_flat.c @@ -7,6 +7,8 @@ * Hacked for x86-64 by James Cleverdon from i386 architecture code by * Martin Bligh, Andi Kleen, James Bottomley, John Stultz, and * James Cleverdon. + * Ashok Raj + * Removed IPI broadcast shortcut to support CPU hotplug */ #include #include @@ -45,22 +47,6 @@ static void flat_init_apic_ldr(void) apic_write_around(APIC_LDR, val); } -static void flat_send_IPI_allbutself(int vector) -{ - /* - * if there are no other CPUs in the system then - * we get an APIC send error if we try to broadcast. - * thus we have to avoid sending IPIs in this case. - */ - if (num_online_cpus() > 1) - __send_IPI_shortcut(APIC_DEST_ALLBUT, vector, APIC_DEST_LOGICAL); -} - -static void flat_send_IPI_all(int vector) -{ - __send_IPI_shortcut(APIC_DEST_ALLINC, vector, APIC_DEST_LOGICAL); -} - static void flat_send_IPI_mask(cpumask_t cpumask, int vector) { unsigned long mask = cpus_addr(cpumask)[0]; @@ -93,6 +79,29 @@ static void flat_send_IPI_mask(cpumask_t local_irq_restore(flags); } +static void flat_send_IPI_allbutself(int vector) +{ + cpumask_t mask; + /* + * if there are no other CPUs in the system then + * we get an APIC send error if we try to broadcast. + * thus we have to avoid sending IPIs in this case. + */ + get_cpu(); + mask = cpu_online_map; + cpu_clear(smp_processor_id(), mask); + + if (cpus_weight(mask) >= 1) + flat_send_IPI_mask(mask, vector); + + put_cpu(); +} + +static void flat_send_IPI_all(int vector) +{ + flat_send_IPI_mask(cpu_online_map, vector); +} + static int flat_apic_id_registered(void) { return physid_isset(GET_APIC_ID(apic_read(APIC_ID)), phys_cpu_present_map); --