[PATCH][DEBUG] x86/refcount: split up refcount saturation handling

[Kees Cook]

In support of debugging the problems Mike Galbraith has seen with
x86-refcount vs gcc vs network refcounts...

This minimizes the differences between unchecked-refcount and x86-refcount
by changing the refcount_dec() failure case to not saturate. The reporting
of negative values is reduced to pr_warn from WARN to avoid spamming dmesg
(which may impact race conditions). Ratelimiting is disabled just to be
sure no reports are being dropped.

Signed-off-by: Kees Cook <keescook@xxxxxxxxxxxx>
---
 arch/x86/mm/extable.c | 51 ++++++++++++++++++++++++++++++++++-----------------
 kernel/panic.c        |  2 +-
 2 files changed, 35 insertions(+), 18 deletions(-)

diff --git a/arch/x86/mm/extable.c b/arch/x86/mm/extable.c
index c076f710de4c..dcb498668370 100644
--- a/arch/x86/mm/extable.c
+++ b/arch/x86/mm/extable.c
@@ -1,6 +1,7 @@
 #include <linux/extable.h>
 #include <linux/uaccess.h>
 #include <linux/sched/debug.h>
+#include <linux/ratelimit.h>
 
 #include <asm/traps.h>
 #include <asm/kdebug.h>
@@ -43,19 +44,7 @@ EXPORT_SYMBOL_GPL(ex_handler_fault);
 bool ex_handler_refcount(const struct exception_table_entry *fixup,
 			 struct pt_regs *regs, int trapnr)
 {
-	/* First unconditionally saturate the refcount. */
-	*(int *)regs->cx = INT_MIN / 2;
-
-	/*
-	 * Strictly speaking, this reports the fixup destination, not
-	 * the fault location, and not the actually overflowing
-	 * instruction, which is the instruction before the "js", but
-	 * since that instruction could be a variety of lengths, just
-	 * report the location after the overflow, which should be close
-	 * enough for finding the overflow, as it's at least back in
-	 * the function, having returned from .text.unlikely.
-	 */
-	regs->ip = ex_fixup_addr(fixup);
+	const char *msg;
 
 	/*
 	 * This function has been called because either a negative refcount
@@ -68,12 +57,40 @@ bool ex_handler_refcount(const struct exception_table_entry *fixup,
 	 * these cases we want a report, since it's a boundary condition.
 	 *
 	 */
-	if (regs->flags & (X86_EFLAGS_OF | X86_EFLAGS_ZF)) {
-		bool zero = regs->flags & X86_EFLAGS_ZF;
-
-		refcount_error_report(regs, zero ? "hit zero" : "overflow");
+	if (regs->flags & X86_EFLAGS_OF) {
+		/* Always saturate on overflow detection. */
+		*(int *)regs->cx = INT_MIN / 2;
+		msg = "saturated due to overflow";
+	} else if (regs->flags & X86_EFLAGS_SF) {
+		/* Do not generate traceback on re-saturation. */
+		*(int *)regs->cx = INT_MIN / 2;
+		regs->ip = ex_fixup_addr(fixup);
+		pr_warn("refcount_t saturated due to negative result at %pB in %s[%d]\n",
+				(void *)instruction_pointer(regs),
+				current->comm, task_pid_nr(current));
+		return true;
+	} else if (regs->flags & X86_EFLAGS_ZF) {
+		/* An unchecked dec-to-zero happened. WARN only. */
+		msg = "hit zero without test";
+	} else {
+		/* Impossible state. */
+		*(int *)regs->cx = INT_MIN / 2;
+		msg = "saturated due to unknown state";
 	}
 
+	/*
+	 * Strictly speaking, this reports the fixup destination, not
+	 * the fault location, and not the actually overflowing
+	 * instruction, which is the instruction before the "js", but
+	 * since that instruction could be a variety of lengths, just
+	 * report the location after the overflow, which should be close
+	 * enough for finding the overflow, as it's at least back in
+	 * the function, having returned from .text.unlikely.
+	 */
+	regs->ip = ex_fixup_addr(fixup);
+
+	refcount_error_report(regs, msg);
+
 	return true;
 }
 EXPORT_SYMBOL_GPL(ex_handler_refcount);
diff --git a/kernel/panic.c b/kernel/panic.c
index bdd18afa19a4..966ade491543 100644
--- a/kernel/panic.c
+++ b/kernel/panic.c
@@ -605,7 +605,7 @@ EXPORT_SYMBOL(__stack_chk_fail);
 #ifdef CONFIG_ARCH_HAS_REFCOUNT
 void refcount_error_report(struct pt_regs *regs, const char *err)
 {
-	WARN_RATELIMIT(1, "refcount_t %s at %pB in %s[%d], uid/euid: %u/%u\n",
+	WARN(1, "refcount_t %s at %pB in %s[%d], uid/euid: %u/%u\n",
 		err, (void *)instruction_pointer(regs),
 		current->comm, task_pid_nr(current),
 		from_kuid_munged(&init_user_ns, current_uid()),
-- 
2.7.4


-- 
Kees Cook
Pixel Security