From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from dfw.source.kernel.org (dfw.source.kernel.org [139.178.84.217]) by mx.groups.io with SMTP id smtpd.web09.33350.1643640878342324502 for ; Mon, 31 Jan 2022 06:54:38 -0800 Authentication-Results: mx.groups.io; dkim=pass header.i=@kernel.org header.s=k20201202 header.b=KyMMr/l+; spf=pass (domain: kernel.org, ip: 139.178.84.217, mailfrom: ardb@kernel.org) Received: from smtp.kernel.org (relay.kernel.org [52.25.139.140]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by dfw.source.kernel.org (Postfix) with ESMTPS id 6596C61353 for ; Mon, 31 Jan 2022 14:54:37 +0000 (UTC) Received: by smtp.kernel.org (Postfix) with ESMTPSA id C9FEEC340EE for ; Mon, 31 Jan 2022 14:54:36 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=kernel.org; s=k20201202; t=1643640876; bh=wESdwUayPNsQ2Qo9MjnSgzhJCNIXg2tGtcClCfly018=; h=References:In-Reply-To:From:Date:Subject:To:Cc:From; b=KyMMr/l+ZIuFyeGfxyjVNpNfEq5A/iK3T62GVABIsOMtdfm6p8Nv3TM5RVl0GXrOq 2UTeNCG8kBOc38Zceu64J4sGIGd0a87ImSzkmCD72DdoEY6Jw8GnJOc1IxQ3lpnu8x dV7VZOlL5Ih/hDljv4FNVeghZs1GTpbK/uegfzGzTZcjY2Mt8P7niYmiHUEk8hHK9K Jp3E7KJxDpu32UYp+8uoCBzRMFVDKycJA4dgkFIXBWJbUJ3TRp6T0SEkhNld44ztms hpezQxEPUvkyCvRrpORIqgvYxo+XLWiB0x2+St+J3bPjW7kaIwUU80BJnP456NbA/F AXf8dsPEDQsMQ== Received: by mail-wr1-f50.google.com with SMTP id w11so25910978wra.4 for ; Mon, 31 Jan 2022 06:54:36 -0800 (PST) X-Gm-Message-State: AOAM532vND0ROGwRB302O9zndYqcGCb0qz/B61vjgVZ16df0CfXHpiSD pMt7KSoWvX2TxhDAcQoOT5pWToJnhOvIteu2LnY= X-Google-Smtp-Source: ABdhPJzSFmtYGrnmTT1sUqgO54cy107w3wtNnUoIutzp4kmSKlo+N//dXpORkhuPqD6fcoCZxzLfg2v0bCuvw8+sCu4= X-Received: by 2002:adf:f386:: with SMTP id m6mr18174438wro.447.1643640875121; Mon, 31 Jan 2022 06:54:35 -0800 (PST) MIME-Version: 1.0 References: <20220128154103.20752-1-Pierre.Gondois@arm.com> <20220128154103.20752-4-Pierre.Gondois@arm.com> <87sft6xv2h.wl-maz@kernel.org> <7f29990d-7ce4-8d4e-5ba3-8eea8a960afe@arm.com> <87k0eg57wk.wl-maz@kernel.org> In-Reply-To: <87k0eg57wk.wl-maz@kernel.org> From: "Ard Biesheuvel" Date: Mon, 31 Jan 2022 15:54:23 +0100 X-Gmail-Original-Message-ID: Message-ID: Subject: Re: [PATCH v3 3/8] DynamicTablesPkg: AcpiSsdtPcieLibArm: Fix _PRT description To: Marc Zyngier Cc: Pierre Gondois , edk2-devel-groups-io , Ard Biesheuvel , Sami Mujawar Content-Type: text/plain; charset="UTF-8" On Mon, 31 Jan 2022 at 14:53, Marc Zyngier wrote: > > On Mon, 31 Jan 2022 12:59:11 +0000, > Pierre Gondois wrote: > > > > Hi, > > > > On 1/29/22 7:20 PM, Marc Zyngier wrote: > > > On Sat, 29 Jan 2022 15:52:02 +0000, > > > Ard Biesheuvel wrote: > > >> > > >> (+ Marc) > > >> > > >> On Fri, 28 Jan 2022 at 16:41, wrote: > > >>> > > >>> From: Pierre Gondois > > >>> > > >>> In ACPI 6.4, s6.2.13, _PRT objects describing PCI legacy interrupts > > >>> can be defined following 2 models. > > >>> In the first model, a _SRS object must be described to modify the PCI > > >>> interrupt. The _CRS and _PRS object allows to describe the PCI > > >>> interrupt (level/edge triggered, active high/low). > > >>> In the second model, the PCI interrupt cannot be described with a > > >>> similar granularity. PCI interrupts are by default level triggered, > > >>> active low. > > >>> > > >>> GicV2 SPI interrupts are level or edge triggered, active high. To > > >>> correctly describe PCI interrupts, the first model is used, even though > > >>> Arm Base Boot Requirements v1.0 requires to use the second mode. > > >>> > > >> > > >> There are two different issues here: > > >> > > >> - using separate 'interrupt link' device objects with an Interrupt() > > >> resource rather than a simple GSIV number > > >> - whether _PRS and _SRS need to be implemented on those link objects. > > >> > > >> The latter is simply not true - _PRS and _SRS are optional, and > > >> pointless if there is only a single possible value, so there is really > > >> no need to add them here. > > >> > > >> As for the choice between the link object or the GSIV number: I don't > > >> think INTx interrupts on ARM systems are actually level low, and the > > >> GSIV option is widely used, also in platforms that exist in > > >> edk2-platforms, without any reported issues. > > >> > > >> I've cc'ed Marc, perhaps he can shed some light on this, but I'd > > >> prefer to stick to the GSIV approach if we can, as it is much simpler. > > > > > > I don't immediately see the point either. Yes, the GIC only supports > > > level-high interrupts. However, all the PCIe implementations connected > > > to it are inverting the level. If they don't, that's even simpler (the > > > HW is broken). > > > > > > Is this to address an apparent disconnect with the spec? > > > > > > [...] > > > > > > > 1. _PRS/_SRS methods > > I agree they are optional and meaningless here as interrupts are not dynamically configurable. > > However linux is checking that: > > - the interrupt used in _CRS is one of the possible interrupts advertised in _PRS. If not, then a warning is issued and the interrupt is not used. > > - the _SRS method is present. If not, setting the interrupt fails. > > If _PRS and _SRS method are really optional, it seems these checks should not happen. > > For now, using a link object without _PRS/_SRS doesn't work. > > > > Note: > > The first check was initially done because an invalid interrupt was advertised in _CRS when valid interrupts were available in _PRS. > > https://bugzilla.kernel.org/show_bug.cgi?id=2665 > > > > 2. GSIV vs link object > > The fist motivation was to accurately describe the interrupts. > > Even though GIC interrupts must be active high, PCI interrupts are > > active low by default (according to spec), and the GSIV model > > doesn't allow to describe the polarity/activation state. > > And any operating system that groks ACPI on arm64 already knows about > this quirks. > > > Another point that came out is that in linux, GSIV interrupts for > > PCI are configured as level triggered by default. > > That's part of the PCI spec. INTx is level triggered, no ifs, no > buts. Otherwise, you can't implement interrupt sharing, which legacy > PCI requires with INTx. So Linux has nothing to do with this. > > > From "Base System Architecture 1.0", sE.4 and sE.6, PCI interrupts > > can be level or edge triggered. > > You are confusing MSIs, which *MUST* be edge, and INTx which *MUST* be > level. These are two very different thing, and you really should not > conflate the two. > > > More specifically, KvmTool configures PCI interrupts as edge > > triggered. > > Well, that's a gross bug in kvmtool. I guess that it doesn't really > matter for virtio devices, but this should be fixed. > > > So the only way to describe an edge interrupt is to use a link object. > > MSIs should never be described in ACPI, as they are entirely SW > programmable (there is no static allocation). Only INTx must be > described, and that's strictly level. > Thanks for clearing that up Marc. So we can drop the link objects after all, which means we are not forced to add 'optional' methods just to appease Linux. This is far better IMHO. In the mean time, I expect that Marc or myself will get kvmtool fixed, but this doesn't actually matter with GSIVs anyway. Thanks, Ard.