From patchwork Thu Jul 22 05:41:44 2021 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: "Zhu, Lingshan" X-Patchwork-Id: 12393049 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-16.8 required=3.0 tests=BAYES_00, HEADER_FROM_DIFFERENT_DOMAINS,INCLUDES_CR_TRAILER,INCLUDES_PATCH, MAILING_LIST_MULTI,SPF_HELO_NONE,SPF_PASS,URIBL_BLOCKED,USER_AGENT_GIT autolearn=ham autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id 28996C63797 for ; Thu, 22 Jul 2021 05:42:40 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [23.128.96.18]) by mail.kernel.org (Postfix) with ESMTP id 12ABA60FED for ; Thu, 22 Jul 2021 05:42:40 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S230365AbhGVFCC (ORCPT ); Thu, 22 Jul 2021 01:02:02 -0400 Received: from mga06.intel.com ([134.134.136.31]:24019 "EHLO mga06.intel.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S230310AbhGVFCB (ORCPT ); Thu, 22 Jul 2021 01:02:01 -0400 X-IronPort-AV: E=McAfee;i="6200,9189,10052"; a="272686896" X-IronPort-AV: E=Sophos;i="5.84,260,1620716400"; d="scan'208";a="272686896" Received: from fmsmga002.fm.intel.com ([10.253.24.26]) by orsmga104.jf.intel.com with ESMTP/TLS/ECDHE-RSA-AES256-GCM-SHA384; 21 Jul 2021 22:42:36 -0700 X-IronPort-AV: E=Sophos;i="5.84,260,1620716400"; d="scan'208";a="512372322" Received: from vmm_a4_icx.sh.intel.com (HELO localhost.localdomain) ([10.239.53.245]) by fmsmga002-auth.fm.intel.com with ESMTP/TLS/ECDHE-RSA-AES256-GCM-SHA384; 21 Jul 2021 22:42:32 -0700 From: Zhu Lingshan To: peterz@infradead.org, pbonzini@redhat.com Cc: bp@alien8.de, seanjc@google.com, vkuznets@redhat.com, wanpengli@tencent.com, jmattson@google.com, joro@8bytes.org, kan.liang@linux.intel.com, ak@linux.intel.com, wei.w.wang@intel.com, eranian@google.com, liuxiangdong5@huawei.com, linux-kernel@vger.kernel.org, x86@kernel.org, kvm@vger.kernel.org, like.xu.linux@gmail.com, boris.ostrvsky@oracle.com, Like Xu , Zhu Lingshan Subject: [PATCH V9 03/18] perf/x86/intel: Handle guest PEBS overflow PMI for KVM guest Date: Thu, 22 Jul 2021 13:41:44 +0800 Message-Id: <20210722054159.4459-4-lingshan.zhu@intel.com> X-Mailer: git-send-email 2.27.0 In-Reply-To: <20210722054159.4459-1-lingshan.zhu@intel.com> References: <20210722054159.4459-1-lingshan.zhu@intel.com> MIME-Version: 1.0 Precedence: bulk List-ID: X-Mailing-List: kvm@vger.kernel.org From: Like Xu With PEBS virtualization, the guest PEBS records get delivered to the guest DS, and the host pmi handler uses perf_guest_cbs->is_in_guest() to distinguish whether the PMI comes from the guest code like Intel PT. No matter how many guest PEBS counters are overflowed, only triggering one fake event is enough. The fake event causes the KVM PMI callback to be called, thereby injecting the PEBS overflow PMI into the guest. KVM may inject the PMI with BUFFER_OVF set, even if the guest DS is empty. That should really be harmless. Thus guest PEBS handler would retrieve the correct information from its own PEBS records buffer. Originally-by: Andi Kleen Co-developed-by: Kan Liang Signed-off-by: Kan Liang Signed-off-by: Like Xu Signed-off-by: Zhu Lingshan --- arch/x86/events/intel/core.c | 45 ++++++++++++++++++++++++++++++++++++ 1 file changed, 45 insertions(+) diff --git a/arch/x86/events/intel/core.c b/arch/x86/events/intel/core.c index da835f5a37e2..2eceb73cd303 100644 --- a/arch/x86/events/intel/core.c +++ b/arch/x86/events/intel/core.c @@ -2783,6 +2783,50 @@ static void intel_pmu_reset(void) } DECLARE_STATIC_CALL(x86_guest_handle_intel_pt_intr, *(perf_guest_cbs->handle_intel_pt_intr)); +DECLARE_STATIC_CALL(x86_guest_state, *(perf_guest_cbs->state)); + +/* + * We may be running with guest PEBS events created by KVM, and the + * PEBS records are logged into the guest's DS and invisible to host. + * + * In the case of guest PEBS overflow, we only trigger a fake event + * to emulate the PEBS overflow PMI for guest PBES counters in KVM. + * The guest will then vm-entry and check the guest DS area to read + * the guest PEBS records. + * + * The contents and other behavior of the guest event do not matter. + */ +static void x86_pmu_handle_guest_pebs(struct pt_regs *regs, + struct perf_sample_data *data) +{ + struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); + u64 guest_pebs_idxs = cpuc->pebs_enabled & ~cpuc->intel_ctrl_host_mask; + struct perf_event *event = NULL; + unsigned int guest = 0; + int bit; + + guest = static_call(x86_guest_state)(); + if (!(guest & PERF_GUEST_ACTIVE)) + return; + + if (!x86_pmu.pebs_vmx || !x86_pmu.pebs_active || + !(cpuc->pebs_enabled & ~cpuc->intel_ctrl_host_mask)) + return; + + for_each_set_bit(bit, (unsigned long *)&guest_pebs_idxs, + INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed) { + event = cpuc->events[bit]; + if (!event->attr.precise_ip) + continue; + + perf_sample_data_init(data, 0, event->hw.last_period); + if (perf_event_overflow(event, data, regs)) + x86_pmu_stop(event, 0); + + /* Inject one fake event is enough. */ + break; + } +} static int handle_pmi_common(struct pt_regs *regs, u64 status) { @@ -2835,6 +2879,7 @@ static int handle_pmi_common(struct pt_regs *regs, u64 status) u64 pebs_enabled = cpuc->pebs_enabled; handled++; + x86_pmu_handle_guest_pebs(regs, &data); x86_pmu.drain_pebs(regs, &data); status &= intel_ctrl | GLOBAL_STATUS_TRACE_TOPAPMI;