From: Ben Hutchings Date: Fri, 9 Mar 2018 00:11:14 +0000 (+0000) Subject: x86/syscall: Sanitize syscall table de-references under speculation X-Git-Url: http://git.openpandora.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=429edb965dd15507f143cdd08637b0dbfa282f0e;p=pandora-kernel.git x86/syscall: Sanitize syscall table de-references under speculation commit 2fbd7af5af8665d18bcefae3e9700be07e22b681 upstream. The upstream version of this, touching C code, was written by Dan Williams, with the following description: > The syscall table base is a user controlled function pointer in kernel > space. Use array_index_nospec() to prevent any out of bounds speculation. > > While retpoline prevents speculating into a userspace directed target it > does not stop the pointer de-reference, the concern is leaking memory > relative to the syscall table base, by observing instruction cache > behavior. The x86_64 assembly version for 4.4 was written by Jiri Slaby, with the following description: > In 4.4.118, we have commit c8961332d6da (x86/syscall: Sanitize syscall > table de-references under speculation), which is a backport of upstream > commit 2fbd7af5af86. But it fixed only the C part of the upstream patch > -- the IA32 sysentry. So it ommitted completely the assembly part -- the > 64bit sysentry. > > Fix that in this patch by explicit array_index_mask_nospec written in > assembly. The same was used in lib/getuser.S. > > However, to have "sbb" working properly, we have to switch from "cmp" > against (NR_syscalls-1) to (NR_syscalls), otherwise the last syscall > number would be "and"ed by 0. It is because the original "ja" relies on > "CF" or "ZF", but we rely only on "CF" in "sbb". That means: switch to > "jae" conditional jump too. > > Final note: use rcx for mask as this is exactly what is overwritten by > the 4th syscall argument (r10) right after. In 3.2 the x86_32 syscall table lookup is also written in assembly. So I've taken Jiri's version and added similar masking in entry_32.S, using edx as the temporary. edx is clobbered by SAVE_REGS and seems to be free at this point. The ia32 compat syscall table lookup on x86_64 is also written in assembly, so I've added the same masking in ia32entry.S, using r8 as the temporary since it is always clobbered by the following instructions. The x86_64 entry code also lacks syscall masking for x32. Cc: Dan Williams Cc: Jiri Slaby Cc: Jan Beulich Cc: Linus Torvalds Cc: Thomas Gleixner Cc: linux-arch@vger.kernel.org Cc: kernel-hardening@lists.openwall.com Cc: gregkh@linuxfoundation.org Cc: Andy Lutomirski Cc: alan@linux.intel.com Cc: Jinpu Wang Signed-off-by: Ben Hutchings --- Reading git-diff-tree failed