[RFC PATCH 14/14] kprobes: Remove NMI context check

[Masami Hiramatsu]

Since the commit 9b38cc704e84 ("kretprobe: Prevent triggering
kretprobe from within kprobe_flush_task") sets a dummy current
kprobe in the trampoline handler by kprobe_busy_begin/end(),
it is not possible to run a kretprobe pre handler in kretprobe
trampoline handler context even with the NMI. If the NMI interrupts
a kretprobe_trampoline_handler() and it hits a kretprobe, the
2nd kretprobe will detect recursion correctly and it will be
skipped.
This means we have almost no double-lock issue on kretprobes by NMI.

The last one point is in cleanup_rp_inst() which also takes
kretprobe_table_lock without setting up current kprobes.
So adding kprobe_busy_begin/end() there allows us to remove
in_nmi() check.

The above commit applies kprobe_busy_begin/end() on x86, but
now all arch implementation are unified to generic one, we can
safely remove the in_nmi() check from arch independent code.

Signed-off-by: Masami Hiramatsu <mhiramat@xxxxxxxxxx>
---
 kernel/kprobes.c |   16 ++++------------
 1 file changed, 4 insertions(+), 12 deletions(-)

diff --git a/kernel/kprobes.c b/kernel/kprobes.c
index cbd2ad1af7b7..311033e4b8e4 100644
--- a/kernel/kprobes.c
+++ b/kernel/kprobes.c
@@ -1359,7 +1359,8 @@ static void cleanup_rp_inst(struct kretprobe *rp)
 	struct hlist_node *next;
 	struct hlist_head *head;
 
-	/* No race here */
+	/* To avoid recursive kretprobe by NMI, set kprobe busy here */
+	kprobe_busy_begin();
 	for (hash = 0; hash < KPROBE_TABLE_SIZE; hash++) {
 		kretprobe_table_lock(hash, &flags);
 		head = &kretprobe_inst_table[hash];
@@ -1369,6 +1370,8 @@ static void cleanup_rp_inst(struct kretprobe *rp)
 		}
 		kretprobe_table_unlock(hash, &flags);
 	}
+	kprobe_busy_end();
+
 	free_rp_inst(rp);
 }
 NOKPROBE_SYMBOL(cleanup_rp_inst);
@@ -2038,17 +2041,6 @@ static int pre_handler_kretprobe(struct kprobe *p, struct pt_regs *regs)
 	unsigned long hash, flags = 0;
 	struct kretprobe_instance *ri;
 
-	/*
-	 * To avoid deadlocks, prohibit return probing in NMI contexts,
-	 * just skip the probe and increase the (inexact) 'nmissed'
-	 * statistical counter, so that the user is informed that
-	 * something happened:
-	 */
-	if (unlikely(in_nmi())) {
-		rp->nmissed++;
-		return 0;
-	}
-
 	/* TODO: consider to only swap the RA after the last pre_handler fired */
 	hash = hash_ptr(current, KPROBE_HASH_BITS);
 	raw_spin_lock_irqsave(&rp->lock, flags);