From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1757154AbaEPO7o (ORCPT ); Fri, 16 May 2014 10:59:44 -0400 Received: from smtp02.citrix.com ([66.165.176.63]:54020 "EHLO SMTP02.CITRIX.COM" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1756780AbaEPO7n (ORCPT ); Fri, 16 May 2014 10:59:43 -0400 X-IronPort-AV: E=Sophos;i="4.97,1067,1389744000"; d="scan'208";a="131202651" From: David Vrabel To: CC: Thomas Gleixner , Ingo Molnar , "H. Peter Anvin" , , David Vrabel , Dave Hansen Subject: [PATCHv2] x86: skip check for spurious faults for non-present faults Date: Fri, 16 May 2014 15:44:18 +0100 Message-ID: <1400251458-14891-1-git-send-email-david.vrabel@citrix.com> X-Mailer: git-send-email 1.7.10.4 MIME-Version: 1.0 Content-Type: text/plain X-DLP: MIA1 Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org If a fault on a kernel address is due to a non-present page, then it cannot be the result of stale TLB entry from a protection change (RO to RW or NX to X). Thus the pagetable walk in spurious_fault() can be skipped. See the initial if in spurious_fault() and the tests in spurious_fault_check()) for the set of possible error codes checked for spurious faults. These are: IRUWP Before x00xx && ( 1xxxx || xxx1x ) After ( 10001 || 00011 ) && ( 1xxxx || xxx1x ) Thus the new condition is a subset of the previous one, excluding only non-present faults (I == 1 and W == 1 are mutually exclusive). This avoids spurious_fault() oopsing in some cases if the pagetables it attempts to walk are not accessible. This obscures the location of the original fault. This also fixes a crash with Xen PV guests when they access entries in the M2P corresponding to device MMIO regions. The M2P is mapped (read-only) by Xen into the kernel address space of the guest and this mapping may contains holes for non-RAM regions. Read faults will result in calls to spurious_fault(), but because the page tables for the M2P mappings are not accessible by the guest the pagetable walk would fault. This was not normally a problem as MMIO mappings would not normally result in a M2P lookup because of the use of the _PAGE_IOMAP bit the PTE. However, removing the _PAGE_IOMAP bit requires M2P lookups for MMIO mappings as well. Signed-off-by: David Vrabel Reported-by: Konrad Rzeszutek Wilk Tested-by: Konrad Rzeszutek Wilk Cc: Dave Hansen --- x86 maintainers, this is a prerequisite for removing Xen's usage of _PAGE_IOMAP so I think this is best merged via the Xen tree. v2: - improve comments --- arch/x86/mm/fault.c | 22 ++++++++++++++++++++-- 1 file changed, 20 insertions(+), 2 deletions(-) diff --git a/arch/x86/mm/fault.c b/arch/x86/mm/fault.c index 8e57229..7f790e4 100644 --- a/arch/x86/mm/fault.c +++ b/arch/x86/mm/fault.c @@ -924,8 +924,17 @@ static int spurious_fault_check(unsigned long error_code, pte_t *pte) * cross-processor TLB flush, even if no stale TLB entries exist * on other processors. * + * Spurious faults may only occur if the TLB contains an entry with + * fewer permission than the page table entry. Non-present (P = 0) + * and reserved bit (R = 1) faults are never spurious. + * * There are no security implications to leaving a stale TLB when * increasing the permissions on a page. + * + * Returns non-zero if a spurious fault was handled, zero otherwise. + * + * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3 + * (Optional Invalidation). */ static noinline __kprobes int spurious_fault(unsigned long error_code, unsigned long address) @@ -936,8 +945,17 @@ spurious_fault(unsigned long error_code, unsigned long address) pte_t *pte; int ret; - /* Reserved-bit violation or user access to kernel space? */ - if (error_code & (PF_USER | PF_RSVD)) + /* + * Only writes to RO or instruction fetches from NX may cause + * spurious faults. + * + * These could be from user or supervisor accesses but the TLB + * is only lazily flushed after a kernel mapping protection + * change, so user accesses are not expected to cause spurious + * faults. + */ + if (error_code != (PF_WRITE | PF_PROT) + && error_code != (PF_INSTR | PF_PROT)) return 0; pgd = init_mm.pgd + pgd_index(address); -- 1.7.10.4