From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id AA6E8C433F5 for ; Thu, 11 Nov 2021 19:05:43 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [23.128.96.18]) by mail.kernel.org (Postfix) with ESMTP id 8C58361252 for ; Thu, 11 Nov 2021 19:05:43 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S233928AbhKKTIc (ORCPT ); Thu, 11 Nov 2021 14:08:32 -0500 Received: from mx0a-001b2d01.pphosted.com ([148.163.156.1]:2562 "EHLO mx0a-001b2d01.pphosted.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S232411AbhKKTIb (ORCPT ); Thu, 11 Nov 2021 14:08:31 -0500 Received: from pps.filterd (m0098396.ppops.net [127.0.0.1]) by mx0a-001b2d01.pphosted.com (8.16.1.2/8.16.1.2) with SMTP id 1ABH0NJa030970; Thu, 11 Nov 2021 19:05:41 GMT DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=ibm.com; h=message-id : subject : from : to : cc : date : in-reply-to : references : content-type : mime-version : content-transfer-encoding; s=pp1; bh=K+hVt+CKf4OssH9jEI3HlNuwQBAj975n0xaJV3m9/fU=; b=MT81+vQTYAKRqYw+usCzt199FEwN9SgZfdVeOVV09HuXd1gIm4ztb41NKv6fXmqTh9vG ynF5B0fEaGcl9k3wT7PioW/kgJxxqDG95uQuPV6NtW+naPRvACcfMYUXX0x+JUzeK5DX OKnUsE3njlO/5Xn+6R9GfRcBn85GRXII2fyOX74zs2V7tm4fBBzu7nVpDUNinK6ovV8m RqcSJyRvvNb96BAINJX9PtpHxa0FYjaAybU5sSDTN2AGZYw0qvoUKFZ4iYZ/vghwt7GI 36l+nq7A34NGyKpK7Rdl3YcVG/Q55sxDLqXsoDzREbQeRXkR9IIlR2dUBAxizcavHcNk Vg== Received: from pps.reinject (localhost [127.0.0.1]) by mx0a-001b2d01.pphosted.com with ESMTP id 3c955gxnag-1 (version=TLSv1.2 cipher=ECDHE-RSA-AES256-GCM-SHA384 bits=256 verify=NOT); Thu, 11 Nov 2021 19:05:40 +0000 Received: from m0098396.ppops.net (m0098396.ppops.net [127.0.0.1]) by pps.reinject (8.16.0.43/8.16.0.43) with SMTP id 1ABIC6ia024486; Thu, 11 Nov 2021 19:05:40 GMT Received: from ppma03wdc.us.ibm.com (ba.79.3fa9.ip4.static.sl-reverse.com [169.63.121.186]) by mx0a-001b2d01.pphosted.com with ESMTP id 3c955gxn9w-1 (version=TLSv1.2 cipher=ECDHE-RSA-AES256-GCM-SHA384 bits=256 verify=NOT); Thu, 11 Nov 2021 19:05:40 +0000 Received: from pps.filterd (ppma03wdc.us.ibm.com [127.0.0.1]) by ppma03wdc.us.ibm.com (8.16.1.2/8.16.1.2) with SMTP id 1ABJ4giX029867; Thu, 11 Nov 2021 19:05:39 GMT Received: from b03cxnp07028.gho.boulder.ibm.com (b03cxnp07028.gho.boulder.ibm.com [9.17.130.15]) by ppma03wdc.us.ibm.com with ESMTP id 3c5hbd3tga-1 (version=TLSv1.2 cipher=ECDHE-RSA-AES256-GCM-SHA384 bits=256 verify=NOT); Thu, 11 Nov 2021 19:05:39 +0000 Received: from b03ledav003.gho.boulder.ibm.com (b03ledav003.gho.boulder.ibm.com [9.17.130.234]) by b03cxnp07028.gho.boulder.ibm.com (8.14.9/8.14.9/NCO v10.0) with ESMTP id 1ABJ5b7S30081726 (version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-GCM-SHA384 bits=256 verify=OK); Thu, 11 Nov 2021 19:05:37 GMT Received: from b03ledav003.gho.boulder.ibm.com (unknown [127.0.0.1]) by IMSVA (Postfix) with ESMTP id B271A6A05F; Thu, 11 Nov 2021 19:05:37 +0000 (GMT) Received: from b03ledav003.gho.boulder.ibm.com (unknown [127.0.0.1]) by IMSVA (Postfix) with ESMTP id BB1A26A047; Thu, 11 Nov 2021 19:05:36 +0000 (GMT) Received: from farman-thinkpad-t470p (unknown [9.211.106.148]) by b03ledav003.gho.boulder.ibm.com (Postfix) with ESMTP; Thu, 11 Nov 2021 19:05:36 +0000 (GMT) Message-ID: <19a2543b24015873db736bddb14d0e4d97712086.camel@linux.ibm.com> Subject: Re: [RFC PATCH v3 2/2] KVM: s390: Extend the USER_SIGP capability From: Eric Farman To: David Hildenbrand , Janosch Frank , Christian Borntraeger , Claudio Imbrenda , Thomas Huth Cc: Heiko Carstens , Vasily Gorbik , kvm@vger.kernel.org, linux-s390@vger.kernel.org Date: Thu, 11 Nov 2021 14:05:35 -0500 In-Reply-To: <85ba9fa3-ca25-b598-aecd-5e0c6a0308f2@redhat.com> References: <20211110203322.1374925-1-farman@linux.ibm.com> <20211110203322.1374925-3-farman@linux.ibm.com> <32836eb5-532f-962d-161a-faa2213a0691@linux.ibm.com> <85ba9fa3-ca25-b598-aecd-5e0c6a0308f2@redhat.com> Content-Type: text/plain; charset="UTF-8" X-Mailer: Evolution 3.28.5 (3.28.5-16.el8) Mime-Version: 1.0 Content-Transfer-Encoding: 7bit X-TM-AS-GCONF: 00 X-Proofpoint-GUID: tmV6IxcfVatlRe0I3janGZF93l3tTYgY X-Proofpoint-ORIG-GUID: tXZ7B2PWodQA_HJtM9TEw02KsxYT3MT- X-Proofpoint-Virus-Version: vendor=baseguard engine=ICAP:2.0.205,Aquarius:18.0.790,Hydra:6.0.425,FMLib:17.0.607.475 definitions=2021-11-11_06,2021-11-11_01,2020-04-07_01 X-Proofpoint-Spam-Details: rule=outbound_notspam policy=outbound score=0 priorityscore=1501 suspectscore=0 spamscore=0 impostorscore=0 adultscore=0 mlxscore=0 phishscore=0 mlxlogscore=999 clxscore=1015 bulkscore=0 malwarescore=0 lowpriorityscore=0 classifier=spam adjust=0 reason=mlx scancount=1 engine=8.12.0-2110150000 definitions=main-2111110100 Precedence: bulk List-ID: X-Mailing-List: kvm@vger.kernel.org On Thu, 2021-11-11 at 19:29 +0100, David Hildenbrand wrote: > On 11.11.21 18:48, Eric Farman wrote: > > On Thu, 2021-11-11 at 17:13 +0100, Janosch Frank wrote: > > > On 11/11/21 16:03, Eric Farman wrote: > > > > On Thu, 2021-11-11 at 10:15 +0100, David Hildenbrand wrote: > > > > > On 10.11.21 21:33, Eric Farman wrote: > > > > > > With commit 2444b352c3ac ("KVM: s390: forward most SIGP > > > > > > orders > > > > > > to > > > > > > user > > > > > > space") we have a capability that allows the "fast" SIGP > > > > > > orders > > > > > > (as > > > > > > defined by the Programming Notes for the SIGNAL PROCESSOR > > > > > > instruction in > > > > > > the Principles of Operation) to be handled in-kernel, while > > > > > > all > > > > > > others are > > > > > > sent to userspace for processing. > > > > > > > > > > > > This works fine but it creates a situation when, for > > > > > > example, a > > > > > > SIGP SENSE > > > > > > might return CC1 (STATUS STORED, and status bits indicating > > > > > > the > > > > > > vcpu is > > > > > > stopped), when in actuality userspace is still processing a > > > > > > SIGP > > > > > > STOP AND > > > > > > STORE STATUS order, and the vcpu is not yet actually > > > > > > stopped. > > > > > > Thus, > > > > > > the > > > > > > SIGP SENSE should actually be returning CC2 (busy) instead > > > > > > of > > > > > > CC1. > > > > > > > > > > > > To fix this, add another CPU capability, dependent on the > > > > > > USER_SIGP > > > > > > one, > > > > > > and two associated IOCTLs. One IOCTL will be used by > > > > > > userspace > > > > > > to > > > > > > mark a > > > > > > vcpu "busy" processing a SIGP order, and cause concurrent > > > > > > orders > > > > > > handled > > > > > > in-kernel to be returned with CC2 (busy). Another IOCTL > > > > > > will be > > > > > > used by > > > > > > userspace to mark the SIGP "finished", and the vcpu free to > > > > > > process > > > > > > additional orders. > > > > > > > > > > > > > > > > This looks much cleaner to me, thanks! > > > > > > > > > > [...] > > > > > > > > > > > diff --git a/arch/s390/kvm/kvm-s390.h b/arch/s390/kvm/kvm- > > > > > > s390.h > > > > > > index c07a050d757d..54371cede485 100644 > > > > > > --- a/arch/s390/kvm/kvm-s390.h > > > > > > +++ b/arch/s390/kvm/kvm-s390.h > > > > > > @@ -82,6 +82,22 @@ static inline int is_vcpu_idle(struct > > > > > > kvm_vcpu > > > > > > *vcpu) > > > > > > return test_bit(vcpu->vcpu_idx, vcpu->kvm- > > > > > > > arch.idle_mask); > > > > > > } > > > > > > > > > > > > +static inline bool kvm_s390_vcpu_is_sigp_busy(struct > > > > > > kvm_vcpu > > > > > > *vcpu) > > > > > > +{ > > > > > > + return (atomic_read(&vcpu->arch.sigp_busy) == 1); > > > > > > > > > > You can drop () > > > > > > > > > > > +} > > > > > > + > > > > > > +static inline bool kvm_s390_vcpu_set_sigp_busy(struct > > > > > > kvm_vcpu > > > > > > *vcpu) > > > > > > +{ > > > > > > + /* Return zero for success, or -EBUSY if another vcpu > > > > > > won */ > > > > > > + return (atomic_cmpxchg(&vcpu->arch.sigp_busy, 0, 1) == > > > > > > 0) ? 0 : > > > > > > -EBUSY; > > > > > > > > > > You can drop () as well. > > > > > > > > > > We might not need the -EBUSY semantics after all. User space > > > > > can > > > > > just > > > > > track if it was set, because it's in charge of setting it. > > > > > > > > Hrm, I added this to distinguish a newer kernel with an older > > > > QEMU, > > > > but > > > > of course an older QEMU won't know the difference either. I'll > > > > doublecheck that this is works fine in the different > > > > permutations. > > > > > > > > > > +} > > > > > > + > > > > > > +static inline void kvm_s390_vcpu_clear_sigp_busy(struct > > > > > > kvm_vcpu > > > > > > *vcpu) > > > > > > +{ > > > > > > + atomic_set(&vcpu->arch.sigp_busy, 0); > > > > > > +} > > > > > > + > > > > > > static inline int kvm_is_ucontrol(struct kvm *kvm) > > > > > > { > > > > > > #ifdef CONFIG_KVM_S390_UCONTROL > > > > > > diff --git a/arch/s390/kvm/sigp.c b/arch/s390/kvm/sigp.c > > > > > > index 5ad3fb4619f1..a37496ea6dfa 100644 > > > > > > --- a/arch/s390/kvm/sigp.c > > > > > > +++ b/arch/s390/kvm/sigp.c > > > > > > @@ -276,6 +276,10 @@ static int handle_sigp_dst(struct > > > > > > kvm_vcpu > > > > > > *vcpu, u8 order_code, > > > > > > if (!dst_vcpu) > > > > > > return SIGP_CC_NOT_OPERATIONAL; > > > > > > > > > > > > + if (kvm_s390_vcpu_is_sigp_busy(dst_vcpu)) { > > > > > > + return SIGP_CC_BUSY; > > > > > > + } > > > > > > > > > > You can drop {} > > > > > > > > Arg, I had some debug in there which needed the braces, and of > > > > course > > > > it's unnecessary now. Thanks. > > > > > > > > > > + > > > > > > switch (order_code) { > > > > > > case SIGP_SENSE: > > > > > > vcpu->stat.instruction_sigp_sense++; > > > > > > @@ -411,6 +415,12 @@ int kvm_s390_handle_sigp(struct > > > > > > kvm_vcpu > > > > > > *vcpu) > > > > > > if (handle_sigp_order_in_user_space(vcpu, order_code, > > > > > > cpu_addr)) > > > > > > return -EOPNOTSUPP; > > > > > > > > > > > > + /* Check the current vcpu, if it was a target from > > > > > > another vcpu > > > > > > */ > > > > > > + if (kvm_s390_vcpu_is_sigp_busy(vcpu)) { > > > > > > + kvm_s390_set_psw_cc(vcpu, SIGP_CC_BUSY); > > > > > > + return 0; > > > > > > + } > > > > > > > > > > I don't think we need this. I think the above (checking the > > > > > target of > > > > > a > > > > > SIGP order) is sufficient. Or which situation do you have in > > > > > mind? > > > > > > > > > > > > > Hrm... I think you're right. I was thinking of this: > > > > > > > > VCPU 1 - SIGP STOP CPU 2 > > > > VCPU 2 - SIGP SENSE CPU 1 > > > > > > > > But of course either CPU2 is going to be marked "busy" first, > > > > and > > > > the > > > > sense doesn't get processed until it's reset, or the sense > > > > arrives > > > > first, and the busy/notbusy doesn't matter. Let me doublecheck > > > > my > > > > tests > > > > for the non-RFC version. > > > > > > > > > I do wonder if we want to make this a kvm_arch_vcpu_ioctl() > > > > > instead, > > > > > > > > In one of my original attempts between v1 and v2, I had put > > > > this > > > > there. > > > > This reliably deadlocks my guest, because the caller > > > > (kvm_vcpu_ioctl()) > > > > tries to acquire vcpu->mutex, and racing SIGPs (via KVM_RUN) > > > > might > > > > already be holding it. Thus, it's an async ioctl. I could fold > > > > it > > > > into > > > > the existing interrupt ioctl, but as those are architected > > > > structs > > > > it > > > > seems more natural do it this way. Or I have mis-understood > > > > something > > > > along the way? > > > > > > > > > essentially just providing a KVM_S390_SET_SIGP_BUSY *and* > > > > > providing > > > > > the > > > > > order. "order == 0" sets it to !busy. > > > > > > > > I'd tried this too, since it provided some nice debug-ability. > > > > Unfortunately, I have a testcase (which I'll eventually get > > > > folded > > > > into > > > > kvm-unit-tests :)) that picks a random order between 0-255, > > > > knowing > > > > that there's only a couple handfuls of valid orders, to check > > > > the > > > > response. Zero is valid architecturally (POPS figure 4-29), > > > > even if > > > > it's unassigned. The likelihood of it becoming assigned is > > > > probably > > > > quite low, but I'm not sure that I like special-casing an order > > > > of > > > > zero > > > > in this way. > > > > > > > > > > Looking at the API I'd like to avoid having two IOCTLs > > > > Since the order is a single byte, we could have the payload of an > > ioctl > > say "0-255 is an order that we're busy processing, anything higher > > than > > that resets the busy" or something. That would remove the need for > > a > > second IOCTL. > > Maybe just pass an int and treat a negative (or just -1) value as > clearing the order. > Right, that's exactly what I had at one point. I thought it was too cumbersome, but maybe not. Will dust it off, pending my question to Janosch about 0-vs-1 IOCTLs.