From: Guenter Roeck Subject: Re: Random crashes with i386 and efi boots Date: Tue, 11 Sep 2018 11:22:16 -0700 Message-ID: <20180911182216.GA21160@roeck-us.net> References: <20180910215659.GA17966@roeck-us.net> <877118e5-beee-4551-28d3-79e7aa52f74e@roeck-us.net> <90A7FF2E-F186-49CF-A028-CDE317BE13E1@amacapital.net> <20180911174158.7qu6f4vfnfzqqrcf@suse.de> <51FDF5AF-2FE8-4817-9A5C-8CE4A8BC23AB@amacapital.net> Mime-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: 8bit Return-path: Content-Disposition: inline In-Reply-To: <51FDF5AF-2FE8-4817-9A5C-8CE4A8BC23AB@amacapital.net> Sender: linux-kernel-owner@vger.kernel.org To: Andy Lutomirski Cc: Joerg Roedel , linux-kernel@vger.kernel.org, Ard Biesheuvel , Thomas Gleixner , Michal Hocko , Andi Kleen , Linus Torvalds , Dave Hansen , Pavel Machek , linux-efi@vger.kernel.org, x86@kernel.org List-Id: linux-efi@vger.kernel.org On Tue, Sep 11, 2018 at 11:05:25AM -0700, Andy Lutomirski wrote: > > > > On Sep 11, 2018, at 10:41 AM, Joerg Roedel wrote: > > > > On Tue, Sep 11, 2018 at 09:36:51AM -0700, Andy Lutomirski wrote: > >>> save_pgd = efi_call_phys_prolog(); > >>> local_irq_save(flags); > >>> status = efi_call_phys(...); > >>> local_irq_restore(flags); > >>> > >>> efi_call_phys_epilog(save_pgd); > >>> > >>> So, yes, interrupts are very much enabled. > >> > >> Does fixing that solve the problem? It seems more robust. > > > > The problem is still that in efi_call_phys_prolog() we load the gdt with > > its physical address, and when we reload the %cr3 in _epilog from > > initial_page_table to swapper_pg_dir again the gdt is no longer mapped. > > Blocking interrupts is more robust, but we can't block NMIs that way > > that would also trigger the issue, no? > > > > So I am in favor of changing the order in efi_call_phys_epilog() too. > > > > I’m rather confused here. We’re loading CR3 with page tables that don’t have the fixmap mapped? With interrupts on? And we expect it to work? This is *nuts*. > > There are IMO only three sane fixes here: > > 1. Load the fixmap, cpu_entry_area, etc into the EFI page table. Drop the GDT reload entirely. > > 2. Do this whole virtual map dance earlier so we don’t have IRQs and NMIs and such. Maybe while we’re still using the initial page table? > > 3. Just identity map all the EFI regions. Make EFI page tables that literally map them at their physical addresses *and* map the entire kernel, just like we do for normal user mms. > > Sure, as a stopgap, turning off IRQs and applying Guenter’s patch seems okay, but this code is not okay. I submitted a patch with the diff I suggested above; it seems to be the least invasive solution and addresses the immediate problem. I am way out of league regarding the other suggested changes. I'll be happy to test the code if someone is willing to rearrange the code accordingly, but I don't think it would make sense to even try doing it myself. Thanks, Guenter