[2.27 COMMITTED] Don't write beyond destination in __mempcpy_avx512_no_vzeroupper (bug 23196)

[Florian Weimer]

From: Andreas Schwab <schwab@suse.de>

When compiled as mempcpy, the return value is the end of the destination
buffer, thus it cannot be used to refer to the start of it.

(cherry picked from commit 9aaaab7c6e4176e61c59b0a63c6ba906d875dc0e)

2018-05-23  Andreas Schwab  <schwab@suse.de>

	[BZ #23196]
	CVE-2018-11237
	* sysdeps/x86_64/multiarch/memmove-avx512-no-vzeroupper.S
	(L(preloop_large)): Save initial destination pointer in %r11 and
	use it instead of %rax after the loop.
	* string/test-mempcpy.c (MIN_PAGE_SIZE): Define.

diff --git a/NEWS b/NEWS
index 57f6714d89..ee08fc3d23 100644
--- a/NEWS
+++ b/NEWS
@@ -48,6 +48,13 @@ The following bugs are resolved with this release:
     build with -Os)
   [23152] gd_GB: Fix typo in "May" (abbreviated)
   [23166] sunrpc: Remove stray exports without --enable-obsolete-rpc
+  [23196] __mempcpy_avx512_no_vzeroupper mishandles large copies
+
+Security related changes:
+
+  CVE-2018-11237: The mempcpy implementation for the Intel Xeon Phi
+  architecture could write beyond the target buffer, resulting in a buffer
+  overflow.  Reported by Andreas Schwab.
 
 
 Version 2.27
diff --git a/string/test-mempcpy.c b/string/test-mempcpy.c
index c08fba895e..d98ecdd2d9 100644
--- a/string/test-mempcpy.c
+++ b/string/test-mempcpy.c
@@ -18,6 +18,7 @@
    <http://www.gnu.org/licenses/>.  */
 
 #define MEMCPY_RESULT(dst, len) (dst) + (len)
+#define MIN_PAGE_SIZE 131072
 #define TEST_MAIN
 #define TEST_NAME "mempcpy"
 #include "test-string.h"
diff --git a/sysdeps/x86_64/multiarch/memmove-avx512-no-vzeroupper.S b/sysdeps/x86_64/multiarch/memmove-avx512-no-vzeroupper.S
index 23c0f7a9ed..effc3ac2de 100644
--- a/sysdeps/x86_64/multiarch/memmove-avx512-no-vzeroupper.S
+++ b/sysdeps/x86_64/multiarch/memmove-avx512-no-vzeroupper.S
@@ -336,6 +336,7 @@ L(preloop_large):
 	vmovups	(%rsi), %zmm4
 	vmovups	0x40(%rsi), %zmm5
 
+	mov	%rdi, %r11
 /* Align destination for access with non-temporal stores in the loop.  */
 	mov	%rdi, %r8
 	and	$-0x80, %rdi
@@ -366,8 +367,8 @@ L(gobble_256bytes_nt_loop):
 	cmp	$256, %rdx
 	ja	L(gobble_256bytes_nt_loop)
 	sfence
-	vmovups	%zmm4, (%rax)
-	vmovups	%zmm5, 0x40(%rax)
+	vmovups	%zmm4, (%r11)
+	vmovups	%zmm5, 0x40(%r11)
 	jmp	L(check)
 
 L(preloop_large_bkw):