From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail.kernel.org ([198.145.29.99]:36402 "EHLO mail.kernel.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1727124AbgJ2IMo (ORCPT ); Thu, 29 Oct 2020 04:12:44 -0400 Date: Thu, 29 Oct 2020 10:12:25 +0200 From: Mike Rapoport Subject: Re: [PATCH 0/4] arch, mm: improve robustness of direct map manipulation Message-ID: <20201029081225.GK1428094@kernel.org> References: <20201025101555.3057-1-rppt@kernel.org> <20201026090526.GA1154158@kernel.org> <20201027083816.GG1154158@kernel.org> <20201028112011.GB27927@willie-the-truck> <20201028113059.GG1428094@kernel.org> <9e77d0a939eda3029d6ae89bd14d7f1465b0559d.camel@intel.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <9e77d0a939eda3029d6ae89bd14d7f1465b0559d.camel@intel.com> List-ID: To: "Edgecombe, Rick P" Cc: "will@kernel.org" , "david@redhat.com" , "cl@linux.com" , "gor@linux.ibm.com" , "hpa@zytor.com" , "peterz@infradead.org" , "catalin.marinas@arm.com" , "dave.hansen@linux.intel.com" , "borntraeger@de.ibm.com" , "penberg@kernel.org" , "linux-mm@kvack.org" , "iamjoonsoo.kim@lge.com" , "aou@eecs.berkeley.edu" , "kirill@shutemov.name" , "rientjes@google.com" , "rppt@linux.ibm.com" , "paulus@samba.org" , "hca@linux.ibm.com" , "bp@alien8.de" , "pavel@ucw.cz" , "sparclinux@vger.kernel.org" , "akpm@linux-foundation.org" , "luto@kernel.org" , "davem@davemloft.net" , "mpe@ellerman.id.au" , "tglx@linutronix.de" , "linuxppc-dev@lists.ozlabs.org" , "linux-riscv@lists.infradead.org" , "x86@kernel.org" , "rjw@rjwysocki.net" , "linux-pm@vger.kernel.org" , "benh@kernel.crashing.org" , "linux-arm-kernel@lists.infradead.org" , "palmer@dabbelt.com" , "Brown, Len" , "mingo@redhat.com" , "linux-s390@vger.kernel.org" , "linux-kernel@vger.kernel.org" , "paul.walmsley@sifive.com" On Wed, Oct 28, 2020 at 09:03:31PM +0000, Edgecombe, Rick P wrote: > > On Wed, Oct 28, 2020 at 11:20:12AM +0000, Will Deacon wrote: > > > On Tue, Oct 27, 2020 at 10:38:16AM +0200, Mike Rapoport wrote: > > > > > > > > This is a theoretical bug, but it is still not nice :) > > > > > > > > > > Just to clarify: this patch series fixes this problem, right? > > > > Yes. > > > > Well, now I'm confused again. > > As David pointed, __vunmap() should not be executing simultaneously > with the hibernate operation because hibernate can't snapshot while > data it needs to save is still updating. If a thread was paused when a > page was in an "invalid" state, it should be remapped by hibernate > before the copy. > > To level set, before reading this mail, my takeaways from the > discussions on potential hibernate/debug page alloc problems were: > > Potential RISC-V issue: > Doesn't have hibernate support > > Potential ARM issue: > The logic around when it's cpa determines pages might be unmapped looks > correct for current callers. > > Potential x86 page break issue: > Seems to be ok for now, but a new set_memory_np() caller could violate > assumptions in hibernate. > > Non-obvious thorny logic: > General agreement it would be good to separate dependencies. > > Behavior of V1 of this patchset: > No functional change other than addition of a warn in hibernate. There is a change that adds explicit use of set_direct_map() to hibernate. Currently, in case of arm64 with DEBUG_PAGEALLOC=n if a thread was paused when a page was in an "invalid" state hibernate will access an unmapped data because __kernel_map_pages() will bail out. After the change set_direct_map_default_noflush() would be used and the page will get mapped before copy. > So "does this fix the problem", "yes" leaves me a bit confused... Not > saying there couldn't be any problems, especially due to the thorniness > and cross arch stride, but what is it exactly and how does this series > fix it? This series goal was primarily to separate dependincies and make it clearer what DEBUG_PAGEALLOC and what SET_DIRECT_MAP are. As it turned out, there is also some lack of consistency between architectures that implement either of this so I tried to improve this as well. Honestly, I don't know if a thread can be paused at the time __vunmap() left invalid pages, but it could, there is an issue on arm64 with DEBUG_PAGEALLOC=n and this set fixes it. __vunmap() vm_remove_mappings() set_direct_map_invalid() /* thread is frozen */ safe_copy_page() __kernel_map_pages() if (!debug_pagealloc()) return do_copy_page() -> fault -- Sincerely yours, Mike.