From: Steven Rostedt (VMware) rostedt@goodmis.org
commit b02414c8f045ab3b9afc816c3735bc98c5c3d262 upstream.
The recursion protection of the ring buffer depends on preempt_count() to be correct. But it is possible that the ring buffer gets called after an interrupt comes in but before it updates the preempt_count(). This will trigger a false positive in the recursion code.
Use the same trick from the ftrace function callback recursion code which uses a "transition" bit that gets set, to allow for a single recursion for to handle transitions between contexts.
Cc: stable@vger.kernel.org Fixes: 567cd4da54ff4 ("ring-buffer: User context bit recursion checking") Signed-off-by: Steven Rostedt (VMware) rostedt@goodmis.org Signed-off-by: Greg Kroah-Hartman gregkh@linuxfoundation.org
--- kernel/trace/ring_buffer.c | 58 +++++++++++++++++++++++++++++++++++---------- 1 file changed, 46 insertions(+), 12 deletions(-)
--- a/kernel/trace/ring_buffer.c +++ b/kernel/trace/ring_buffer.c @@ -422,14 +422,16 @@ struct rb_event_info {
/* * Used for which event context the event is in. - * NMI = 0 - * IRQ = 1 - * SOFTIRQ = 2 - * NORMAL = 3 + * TRANSITION = 0 + * NMI = 1 + * IRQ = 2 + * SOFTIRQ = 3 + * NORMAL = 4 * * See trace_recursive_lock() comment below for more details. */ enum { + RB_CTX_TRANSITION, RB_CTX_NMI, RB_CTX_IRQ, RB_CTX_SOFTIRQ, @@ -2660,10 +2662,10 @@ rb_wakeups(struct ring_buffer *buffer, s * a bit of overhead in something as critical as function tracing, * we use a bitmask trick. * - * bit 0 = NMI context - * bit 1 = IRQ context - * bit 2 = SoftIRQ context - * bit 3 = normal context. + * bit 1 = NMI context + * bit 2 = IRQ context + * bit 3 = SoftIRQ context + * bit 4 = normal context. * * This works because this is the order of contexts that can * preempt other contexts. A SoftIRQ never preempts an IRQ @@ -2686,6 +2688,30 @@ rb_wakeups(struct ring_buffer *buffer, s * The least significant bit can be cleared this way, and it * just so happens that it is the same bit corresponding to * the current context. + * + * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit + * is set when a recursion is detected at the current context, and if + * the TRANSITION bit is already set, it will fail the recursion. + * This is needed because there's a lag between the changing of + * interrupt context and updating the preempt count. In this case, + * a false positive will be found. To handle this, one extra recursion + * is allowed, and this is done by the TRANSITION bit. If the TRANSITION + * bit is already set, then it is considered a recursion and the function + * ends. Otherwise, the TRANSITION bit is set, and that bit is returned. + * + * On the trace_recursive_unlock(), the TRANSITION bit will be the first + * to be cleared. Even if it wasn't the context that set it. That is, + * if an interrupt comes in while NORMAL bit is set and the ring buffer + * is called before preempt_count() is updated, since the check will + * be on the NORMAL bit, the TRANSITION bit will then be set. If an + * NMI then comes in, it will set the NMI bit, but when the NMI code + * does the trace_recursive_unlock() it will clear the TRANSTION bit + * and leave the NMI bit set. But this is fine, because the interrupt + * code that set the TRANSITION bit will then clear the NMI bit when it + * calls trace_recursive_unlock(). If another NMI comes in, it will + * set the TRANSITION bit and continue. + * + * Note: The TRANSITION bit only handles a single transition between context. */
static __always_inline int @@ -2701,8 +2727,16 @@ trace_recursive_lock(struct ring_buffer_ bit = pc & NMI_MASK ? RB_CTX_NMI : pc & HARDIRQ_MASK ? RB_CTX_IRQ : RB_CTX_SOFTIRQ;
- if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) - return 1; + if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) { + /* + * It is possible that this was called by transitioning + * between interrupt context, and preempt_count() has not + * been updated yet. In this case, use the TRANSITION bit. + */ + bit = RB_CTX_TRANSITION; + if (val & (1 << (bit + cpu_buffer->nest))) + return 1; + }
val |= (1 << (bit + cpu_buffer->nest)); cpu_buffer->current_context = val; @@ -2717,8 +2751,8 @@ trace_recursive_unlock(struct ring_buffe cpu_buffer->current_context - (1 << cpu_buffer->nest); }
-/* The recursive locking above uses 4 bits */ -#define NESTED_BITS 4 +/* The recursive locking above uses 5 bits */ +#define NESTED_BITS 5
/** * ring_buffer_nest_start - Allow to trace while nested