kcsan: Address missing case with KCSAN_REPORT_VALUE_CHANGE_ONLY
authorMarco Elver <elver@google.com>
Wed, 29 Jan 2020 15:01:02 +0000 (16:01 +0100)
committerIngo Molnar <mingo@kernel.org>
Sat, 21 Mar 2020 08:41:29 +0000 (09:41 +0100)
Even with KCSAN_REPORT_VALUE_CHANGE_ONLY, KCSAN still reports data
races between reads and watchpointed writes, even if the writes wrote
values already present.  This commit causes KCSAN to unconditionally
skip reporting in this case.

Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
kernel/kcsan/report.c

index 33bdf8b..7cd3428 100644 (file)
@@ -130,12 +130,25 @@ static bool rate_limit_report(unsigned long frame1, unsigned long frame2)
  * Special rules to skip reporting.
  */
 static bool
-skip_report(int access_type, bool value_change, unsigned long top_frame)
+skip_report(bool value_change, unsigned long top_frame)
 {
-       const bool is_write = (access_type & KCSAN_ACCESS_WRITE) != 0;
-
-       if (IS_ENABLED(CONFIG_KCSAN_REPORT_VALUE_CHANGE_ONLY) && is_write &&
-           !value_change) {
+       /*
+        * The first call to skip_report always has value_change==true, since we
+        * cannot know the value written of an instrumented access. For the 2nd
+        * call there are 6 cases with CONFIG_KCSAN_REPORT_VALUE_CHANGE_ONLY:
+        *
+        * 1. read watchpoint, conflicting write (value_change==true): report;
+        * 2. read watchpoint, conflicting write (value_change==false): skip;
+        * 3. write watchpoint, conflicting write (value_change==true): report;
+        * 4. write watchpoint, conflicting write (value_change==false): skip;
+        * 5. write watchpoint, conflicting read (value_change==false): skip;
+        * 6. write watchpoint, conflicting read (value_change==true): impossible;
+        *
+        * Cases 1-4 are intuitive and expected; case 5 ensures we do not report
+        * data races where the write may have rewritten the same value; and
+        * case 6 is simply impossible.
+        */
+       if (IS_ENABLED(CONFIG_KCSAN_REPORT_VALUE_CHANGE_ONLY) && !value_change) {
                /*
                 * The access is a write, but the data value did not change.
                 *
@@ -228,7 +241,7 @@ static bool print_report(const volatile void *ptr, size_t size, int access_type,
        /*
         * Must check report filter rules before starting to print.
         */
-       if (skip_report(access_type, true, stack_entries[skipnr]))
+       if (skip_report(true, stack_entries[skipnr]))
                return false;
 
        if (type == KCSAN_REPORT_RACE_SIGNAL) {
@@ -237,7 +250,7 @@ static bool print_report(const volatile void *ptr, size_t size, int access_type,
                other_frame = other_info.stack_entries[other_skipnr];
 
                /* @value_change is only known for the other thread */
-               if (skip_report(other_info.access_type, value_change, other_frame))
+               if (skip_report(value_change, other_frame))
                        return false;
        }