From mboxrd@z Thu Jan 1 00:00:00 1970 From: James Morse Subject: Re: [PATCH v6 00/18] APEI in_nmi() rework Date: Wed, 3 Oct 2018 18:50:38 +0100 Message-ID: References: <20180921221705.6478-1-james.morse@arm.com> <20180925124526.GD23986@zn.tnic> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: <20180925124526.GD23986@zn.tnic> Content-Language: en-US List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: kvmarm-bounces@lists.cs.columbia.edu Sender: kvmarm-bounces@lists.cs.columbia.edu To: Borislav Petkov Cc: jonathan.zhang@cavium.com, Rafael Wysocki , Tony Luck , linux-mm@kvack.org, Marc Zyngier , Catalin Marinas , Tyler Baicar , Will Deacon , Dongjiu Geng , linux-acpi@vger.kernel.org, Punit Agrawal , Naoya Horiguchi , kvmarm@lists.cs.columbia.edu, linux-arm-kernel@lists.infradead.org, Len Brown List-Id: linux-acpi@vger.kernel.org Hi Boris, On 25/09/18 13:45, Borislav Petkov wrote: > On Fri, Sep 21, 2018 at 11:16:47PM +0100, James Morse wrote: >> Hello, >> >> The GHES driver has collected quite a few bugs: >> >> ghes_proc() at ghes_probe() time can be interrupted by an NMI that >> will clobber the ghes->estatus fields, flags, and the buffer_paddr. >> >> ghes_copy_tofrom_phys() uses in_nmi() to decide which path to take. arm64's >> SEA taking both paths, depending on what it interrupted. >> >> There is no guarantee that queued memory_failure() errors will be processed >> before this CPU returns to user-space. >> >> x86 can't TLBI from interrupt-masked code which this driver does all the >> time. >> >> >> This series aims to fix the first three, with an eye to fixing the >> last one with a follow-up series. >> >> Previous postings included the SDEI notification calls, which I haven't >> finished re-testing. This series is big enough as it is. > Yeah, and everywhere I look, this thing looks overengineered. Like, > for example, what's the purpose of this ghes_esource_prealloc_size() > computing a size each time the pool changes size? The size to grow the pool by, because each error-source described by a GHES entry has its own worst-case size. Today ghes_nmi_add() does this each time its called. You could have multiple GHES entries in the HEST that describe NMI as the notification. The worst-case size for the records is described in the GHES entry, and could be different for each one. (error_block_length and records_to_preallocate, or table 18-379 of acpi v6.2) These different error-sources could be delivered on different CPUs at the same time, so need their own pre-allocated reserved memory. ghes_notify_nmi()'s atomic_add_unless() suggests this can happen on x86, but I don't know the arch-specifics. It definitely can happen on arm64. > AFAICT, this size can be computed exactly *once* at driver init and be > done with it. Right? We could do two passes of the HEST to pre-compute the total size of this estatus-queue memory, allocate it, then do the notification registration stuff. But this doesn't really work with the way this driver acts as platform-driver for a ghes device... The non-ghes HEST entries have a "number of records to pre-allocate" too, we could make this memory pool something hest.c looks after, but I can't see if the other error sources use those values. Hmmm, The size is capped to 64K, we could ignore the firmware description of the memory requirements, and allocate SZ_64K each time. Doing it per-GHES is still the only way to avoid allocating nmi-safe memory for irqs. Thanks, James