On Wed, 2018-07-11 at 12:37 -0700, Jann Horn wrote:

On Tue, Jul 10, 2018 at 3:31 PM Yu-cheng Yu [off-list ref]
wrote:

quoted

Look in .note.gnu.property of an ELF file and check if shadow stack
needs
to be enabled for the task.

Signed-off-by: H.J. Lu <redacted>
Signed-off-by: Yu-cheng Yu <redacted>

[...]

quoted

diff --git a/arch/x86/kernel/elf.c b/arch/x86/kernel/elf.c
new file mode 100644
index 000000000000..233f6dad9c1f
--- /dev/null
+++ b/arch/x86/kernel/elf.c

[...]

quoted

+#define NOTE_SIZE_BAD(n, align, max) \
+       ((n->n_descsz < 8) || ((n->n_descsz % align) != 0) || \
+        (((u8 *)(n + 1) + 4 + n->n_descsz) > (max)))

Please do not compute out-of-bounds pointers and then compare them
against an expected maximum pointer. Computing an out-of-bounds
pointer is undefined behavior according to the C99 specification,
section "6.5.6 Additive operators", paragraph 8; and in this case,
n->n_descsz is 32 bits wide, which means that even if the compiler
isn't doing anything funny, if you're operating on addresses in the
last 4GiB of virtual memory and the pointer wraps around, this could
break.
In particular, if anyone ever uses this code in a 32-bit kernel, this
is going to blow up.
Please use size comparisons instead of pointer comparisons.

I will fix it.

[...]

quoted

diff --git a/fs/binfmt_elf.c b/fs/binfmt_elf.c
index 0ac456b52bdd..3395f6a631d5 100644
--- a/fs/binfmt_elf.c
+++ b/fs/binfmt_elf.c

@@ -1081,6 +1081,22 @@ static int load_elf_binary(struct

linux_binprm *bprm)
                goto out_free_dentry;
        }

+#ifdef CONFIG_ARCH_HAS_PROGRAM_PROPERTIES
+
+       if (interpreter) {
+               retval = arch_setup_features(&loc->interp_elf_ex,
+                                            interp_elf_phdata,
+                                            interpreter, true);
+       } else {
+               retval = arch_setup_features(&loc->elf_ex,
+                                            elf_phdata,
+                                            bprm->file, false);
+       }

So for non-static binaries, the ELF headers of ld.so determine
whether
CET will be on or off for the entire system, right? Is the intent
here
that ld.so should start with CET enabled, and then either use the
compatibility bitmap or turn CET off at runtime if the executable or
one of the libraries doesn't actually work with CET?

The kernel command-line options "no_cet_shstk" and "no_cet_ibt" turn
off CET features for the whole system.  The GLIBC tunable
"glibc.tune.hwcap=-SHSTK,-IBT" turns off CET features for the current
shell.  Another GLIBC tunable "glibc.tune.x86_shstk=<on, permissive>"
determines, in the current shell, how dlopen() deals with SHSTK legacy
lib's.

So, if ld.so's ELF header has SHSTK/IBT, and CET is enabled in the
current shell, it will run with CET enabled.  If the application
executable and all its dependent libraries have CET, ld.so runs the
application with CET enabled.  Otherwise ld.so turns off SHSTK (and/or
sets up legacy bitmap for IBT) before passing to the application.

Yu-cheng