While we lack a compiler attribute to add to noinstr that would disable
KCOV, make the KCOV runtime functions return if the caller is in a
noinstr section. We then whitelist __sanitizer_cov_*() functions in
objtool. __sanitizer_cov_*() cannot safely become safe noinstr functions
as-is, as they may fault due to accesses to vmalloc's memory.
Declare write_comp_data() as __always_inline to ensure it is inlined,
and reduce stack usage and remove one extra call from the fast-path.
In future, our compilers may provide an attribute to implement
__no_sanitize_coverage, which can then be added to noinstr, and the
checks added in this patch can be guarded by an #ifdef checking if the
compiler has such an attribute or not.
Signed-off-by: Marco Elver <[email protected]>
---
Apply after:
https://lkml.kernel.org/r/[email protected]
v2:
* Rewrite based on Peter's and Andrey's feedback -- v1 worked because we
got lucky. Let's not rely on luck, as it will be difficult to ensure the
same conditions remain true in future.
v1: https://lkml.kernel.org/r/[email protected]
Note: There are a set of KCOV patches from Andrey in -next:
https://lkml.kernel.org/r/[email protected] --
Git cleanly merges this patch with those patches, and no merge conflict
is expected.
---
kernel/kcov.c | 19 +++++++++++++++++--
tools/objtool/check.c | 7 +++++++
2 files changed, 24 insertions(+), 2 deletions(-)
diff --git a/kernel/kcov.c b/kernel/kcov.c
index 8accc9722a81..3329a0fdb868 100644
--- a/kernel/kcov.c
+++ b/kernel/kcov.c
@@ -24,6 +24,7 @@
#include <linux/refcount.h>
#include <linux/log2.h>
#include <asm/setup.h>
+#include <asm/sections.h>
#define kcov_debug(fmt, ...) pr_debug("%s: " fmt, __func__, ##__VA_ARGS__)
@@ -172,6 +173,12 @@ static notrace unsigned long canonicalize_ip(unsigned long ip)
return ip;
}
+static __always_inline bool in_noinstr_section(unsigned long ip)
+{
+ return (unsigned long)__noinstr_text_start <= ip &&
+ ip < (unsigned long)__noinstr_text_end;
+}
+
/*
* Entry point from instrumented code.
* This is called once per basic-block/edge.
@@ -180,13 +187,18 @@ void notrace __sanitizer_cov_trace_pc(void)
{
struct task_struct *t;
unsigned long *area;
- unsigned long ip = canonicalize_ip(_RET_IP_);
+ unsigned long ip;
unsigned long pos;
+ if (unlikely(in_noinstr_section(_RET_IP_)))
+ return;
+
t = current;
if (!check_kcov_mode(KCOV_MODE_TRACE_PC, t))
return;
+ ip = canonicalize_ip(_RET_IP_);
+
area = t->kcov_area;
/* The first 64-bit word is the number of subsequent PCs. */
pos = READ_ONCE(area[0]) + 1;
@@ -198,12 +210,15 @@ void notrace __sanitizer_cov_trace_pc(void)
EXPORT_SYMBOL(__sanitizer_cov_trace_pc);
#ifdef CONFIG_KCOV_ENABLE_COMPARISONS
-static void notrace write_comp_data(u64 type, u64 arg1, u64 arg2, u64 ip)
+static __always_inline void write_comp_data(u64 type, u64 arg1, u64 arg2, u64 ip)
{
struct task_struct *t;
u64 *area;
u64 count, start_index, end_pos, max_pos;
+ if (unlikely(in_noinstr_section(ip)))
+ return;
+
t = current;
if (!check_kcov_mode(KCOV_MODE_TRACE_CMP, t))
return;
diff --git a/tools/objtool/check.c b/tools/objtool/check.c
index 3e214f879ada..cb208959f560 100644
--- a/tools/objtool/check.c
+++ b/tools/objtool/check.c
@@ -2213,6 +2213,13 @@ static inline bool noinstr_call_dest(struct symbol *func)
if (!strncmp(func->name, "__ubsan_handle_", 15))
return true;
+ /*
+ * The __sanitizer_cov_*() calls include a check if the caller is in the
+ * noinstr section, and simply return if that is the case.
+ */
+ if (!strncmp(func->name, "__sanitizer_cov_", 16))
+ return true;
+
return false;
}
--
2.27.0.rc2.251.g90737beb825-goog
On Thu, Jun 04, 2020 at 04:56:34PM +0200, Marco Elver wrote:
> While we lack a compiler attribute to add to noinstr that would disable
> KCOV, make the KCOV runtime functions return if the caller is in a
> noinstr section. We then whitelist __sanitizer_cov_*() functions in
> objtool.
> __sanitizer_cov_*() cannot safely become safe noinstr functions
> as-is, as they may fault due to accesses to vmalloc's memory.
I would feel very much better with those actually in noinstr, because
without it, there is nothing stopping us from adding a kprobe/hw-
breakpoint or other funny to the function.
Even if they almost instra-return, having a kprobe on the function entry
or condition check is enough to utterly wreck things.
So something like:
void noinstr __sanitizer_cov_trace_*(...)
{
if (within_noinstr_section(ip))
return;
instrumentation_begin();
write_comp_data(...);
instrumentation_end();
}
Would make me feel a whole lot better.
> +static __always_inline bool in_noinstr_section(unsigned long ip)
> +{
> + return (unsigned long)__noinstr_text_start <= ip &&
> + ip < (unsigned long)__noinstr_text_end;
> +}
.entry.text is also considered noinstr, although I suppose that all
being in .S files avoids it having annotations inserted, but perhaps a
comment?
[-Cc most]
[+Cc Mark]
On Thu, 4 Jun 2020 at 17:25, Peter Zijlstra <[email protected]> wrote:
[...]
> I would feel very much better with those actually in noinstr, because
> without it, there is nothing stopping us from adding a kprobe/hw-
> breakpoint or other funny to the function.
>
> Even if they almost instra-return, having a kprobe on the function entry
> or condition check is enough to utterly wreck things.
>
> So something like:
>
> void noinstr __sanitizer_cov_trace_*(...)
> {
> if (within_noinstr_section(ip))
> return;
>
> instrumentation_begin();
> write_comp_data(...);
> instrumentation_end();
> }
Apologies for resurrecting this. :-)
It seems I'll need to use this approach soon for upcoming KCSAN
instrumentation for memory barriers. I'm able to use the same objtool
feature that erases __sanitizer_cov* calls on x86 to erase memory
barrier instrumentation, but arm64 will still be a problem because of
lack of objtool support.
Mark, on arm64, is the approach above that Peter proposed ~1y ago
acceptable in general to make instrumentation noinstr-safe?
Thanks,
-- Marco