From mboxrd@z Thu Jan 1 00:00:00 1970 From: Will Deacon Subject: Re: [PATCH v5 00/20] APEI in_nmi() rework and arm64 SDEI wire-up Date: Wed, 4 Jul 2018 15:37:28 +0100 Message-ID: <20180704143727.GI4828@arm.com> References: <20180626170116.25825-1-james.morse@arm.com> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Return-path: Content-Disposition: inline In-Reply-To: <20180626170116.25825-1-james.morse@arm.com> List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Sender: "linux-arm-kernel" Errors-To: linux-arm-kernel-bounces+linux-arm-kernel=m.gmane.org@lists.infradead.org To: James Morse Cc: jonathan.zhang@cavium.com, Tony Luck , Xie XiuQi , linux-mm@kvack.org, Marc Zyngier , Catalin Marinas , Tyler Baicar , Rafael Wysocki , Christoffer Dall , Dongjiu Geng , linux-acpi@vger.kernel.org, Punit Agrawal , Borislav Petkov , Naoya Horiguchi , kvmarm@lists.cs.columbia.edu, linux-arm-kernel@lists.infradead.org, Len Brown List-Id: linux-acpi@vger.kernel.org Hi James, On Tue, Jun 26, 2018 at 06:00:56PM +0100, James Morse wrote: > The aim of this series is to wire arm64's SDEI into APEI. > > On arm64 we have three APEI notifications that are NMI-like, and > in the unlikely event that all three are supported by a platform, > they can interrupt each other. > The GHES driver shouldn't have to deal with this, so this series aims > to make it re-entrant. > > To do that, we refactor the estatus queue to allow multiple notifications > to use it, then convert NOTIFY_SEA to always be described as NMI-like, > and to use the estatus queue. > > From here we push the locking and fixmap choices out to the notification > functions, and remove the use of per-ghes estatus and flags. This removes > the in_nmi() 'timebomb' in ghes_copy_tofrom_phys(). > > Things get sticky when an NMI notification needs to know how big the > CPER records might be, before reading it. This series splits > ghes_estatus_read() to let us peek at the buffer. A side effect of this > is the 20byte header will get read twice. (how does it work today? it > reads the records into a per-ghes worst-case sized buffer, allocates > the correct size and copies the records. in_nmi() use of this per-ghes > buffer needs eliminating). > > One alternative was to trust firmware's 'max raw data length' and use > that to allocate 'enough' memory. We don't use this value today, so its > probably wrong on some sytem somewhere. > > Since v4 patches 5,8-15 are new, otherwise changes are noted in the patch. The little bits touching arch/arm64/ all look fine to me here, but it looks like other patches need review separately and ultimately I suspect you're going to route it via some other tree. Let me know if you need me to help with anything. Will