context_tracking: Ensure that the critical path cannot be instrumented
authorThomas Gleixner <tglx@linutronix.de>
Wed, 4 Mar 2020 10:05:22 +0000 (11:05 +0100)
committerThomas Gleixner <tglx@linutronix.de>
Thu, 11 Jun 2020 13:14:36 +0000 (15:14 +0200)
context tracking lacks a few protection mechanisms against instrumentation:

 - While the core functions are marked NOKPROBE they lack protection
   against function tracing which is required as the function entry/exit
   points can be utilized by BPF.

 - static functions invoked from the protected functions need to be marked
   as well as they can be instrumented otherwise.

 - using plain inline allows the compiler to emit traceable and probable
   functions.

Fix this by marking the functions noinstr and converting the plain inlines
to __always_inline.

The NOKPROBE_SYMBOL() annotations are removed as the .noinstr.text section
is already excluded from being probed.

Cures the following objtool warnings:

 vmlinux.o: warning: objtool: enter_from_user_mode()+0x34: call to __context_tracking_exit() leaves .noinstr.text section
 vmlinux.o: warning: objtool: prepare_exit_to_usermode()+0x29: call to __context_tracking_enter() leaves .noinstr.text section
 vmlinux.o: warning: objtool: syscall_return_slowpath()+0x29: call to __context_tracking_enter() leaves .noinstr.text section
 vmlinux.o: warning: objtool: do_syscall_64()+0x7f: call to __context_tracking_enter() leaves .noinstr.text section
 vmlinux.o: warning: objtool: do_int80_syscall_32()+0x3d: call to __context_tracking_enter() leaves .noinstr.text section
 vmlinux.o: warning: objtool: do_fast_syscall_32()+0x9c: call to __context_tracking_enter() leaves .noinstr.text section

and generates new ones...

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Masami Hiramatsu <mhiramat@kernel.org>
Reviewed-by: Alexandre Chartre <alexandre.chartre@oracle.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200505134340.811520478@linutronix.de
include/linux/context_tracking.h
include/linux/context_tracking_state.h
kernel/context_tracking.c

index 8cac62ee6add1d96a8e7bc5e9a0d6919293e2413..981b880d5b605e8e5a46f1b4b016751d2b8be323 100644 (file)
@@ -33,13 +33,13 @@ static inline void user_exit(void)
 }
 
 /* Called with interrupts disabled.  */
-static inline void user_enter_irqoff(void)
+static __always_inline void user_enter_irqoff(void)
 {
        if (context_tracking_enabled())
                __context_tracking_enter(CONTEXT_USER);
 
 }
-static inline void user_exit_irqoff(void)
+static __always_inline void user_exit_irqoff(void)
 {
        if (context_tracking_enabled())
                __context_tracking_exit(CONTEXT_USER);
@@ -75,7 +75,7 @@ static inline void exception_exit(enum ctx_state prev_ctx)
  * is enabled.  If context tracking is disabled, returns
  * CONTEXT_DISABLED.  This should be used primarily for debugging.
  */
-static inline enum ctx_state ct_state(void)
+static __always_inline enum ctx_state ct_state(void)
 {
        return context_tracking_enabled() ?
                this_cpu_read(context_tracking.state) : CONTEXT_DISABLED;
index e7fe6678b7ad135f0d74bb9f5335417a8b1479a8..65a60d3313b006e40d27da7418b8174564b82d39 100644 (file)
@@ -26,12 +26,12 @@ struct context_tracking {
 extern struct static_key_false context_tracking_key;
 DECLARE_PER_CPU(struct context_tracking, context_tracking);
 
-static inline bool context_tracking_enabled(void)
+static __always_inline bool context_tracking_enabled(void)
 {
        return static_branch_unlikely(&context_tracking_key);
 }
 
-static inline bool context_tracking_enabled_cpu(int cpu)
+static __always_inline bool context_tracking_enabled_cpu(int cpu)
 {
        return context_tracking_enabled() && per_cpu(context_tracking.active, cpu);
 }
@@ -41,7 +41,7 @@ static inline bool context_tracking_enabled_this_cpu(void)
        return context_tracking_enabled() && __this_cpu_read(context_tracking.active);
 }
 
-static inline bool context_tracking_in_user(void)
+static __always_inline bool context_tracking_in_user(void)
 {
        return __this_cpu_read(context_tracking.state) == CONTEXT_USER;
 }
index ce430885c26c6d79d63ee86cf5e6c66ed32a396c..36a98c48aedc7c6c14e732a7ce3f360a1b8f3087 100644 (file)
@@ -31,7 +31,7 @@ EXPORT_SYMBOL_GPL(context_tracking_key);
 DEFINE_PER_CPU(struct context_tracking, context_tracking);
 EXPORT_SYMBOL_GPL(context_tracking);
 
-static bool context_tracking_recursion_enter(void)
+static noinstr bool context_tracking_recursion_enter(void)
 {
        int recursion;
 
@@ -45,7 +45,7 @@ static bool context_tracking_recursion_enter(void)
        return false;
 }
 
-static void context_tracking_recursion_exit(void)
+static __always_inline void context_tracking_recursion_exit(void)
 {
        __this_cpu_dec(context_tracking.recursion);
 }
@@ -59,7 +59,7 @@ static void context_tracking_recursion_exit(void)
  * instructions to execute won't use any RCU read side critical section
  * because this function sets RCU in extended quiescent state.
  */
-void __context_tracking_enter(enum ctx_state state)
+void noinstr __context_tracking_enter(enum ctx_state state)
 {
        /* Kernel threads aren't supposed to go to userspace */
        WARN_ON_ONCE(!current->mm);
@@ -77,8 +77,10 @@ void __context_tracking_enter(enum ctx_state state)
                         * on the tick.
                         */
                        if (state == CONTEXT_USER) {
+                               instrumentation_begin();
                                trace_user_enter(0);
                                vtime_user_enter(current);
+                               instrumentation_end();
                        }
                        rcu_user_enter();
                }
@@ -99,7 +101,6 @@ void __context_tracking_enter(enum ctx_state state)
        }
        context_tracking_recursion_exit();
 }
-NOKPROBE_SYMBOL(__context_tracking_enter);
 EXPORT_SYMBOL_GPL(__context_tracking_enter);
 
 void context_tracking_enter(enum ctx_state state)
@@ -142,7 +143,7 @@ NOKPROBE_SYMBOL(context_tracking_user_enter);
  * This call supports re-entrancy. This way it can be called from any exception
  * handler without needing to know if we came from userspace or not.
  */
-void __context_tracking_exit(enum ctx_state state)
+void noinstr __context_tracking_exit(enum ctx_state state)
 {
        if (!context_tracking_recursion_enter())
                return;
@@ -155,15 +156,16 @@ void __context_tracking_exit(enum ctx_state state)
                         */
                        rcu_user_exit();
                        if (state == CONTEXT_USER) {
+                               instrumentation_begin();
                                vtime_user_exit(current);
                                trace_user_exit(0);
+                               instrumentation_end();
                        }
                }
                __this_cpu_write(context_tracking.state, CONTEXT_KERNEL);
        }
        context_tracking_recursion_exit();
 }
-NOKPROBE_SYMBOL(__context_tracking_exit);
 EXPORT_SYMBOL_GPL(__context_tracking_exit);
 
 void context_tracking_exit(enum ctx_state state)