Fix: lttng perf counter deadlock
authorMathieu Desnoyers <mathieu.desnoyers@efficios.com>
Mon, 7 Oct 2019 19:45:46 +0000 (15:45 -0400)
committerMathieu Desnoyers <mathieu.desnoyers@efficios.com>
Fri, 11 Oct 2019 14:53:56 +0000 (10:53 -0400)
Using the ust_lock() to lazily setup the perf counters introduces
a scenario where this lock is nested within the urcu-bp read-side
lock.

However, the LTTNG_UST_WAIT_QUIESCENT ust command requires that
urcu-bp synchronize_rcu() is performed with the ust_lock() held.

This inter-dependency introduces a deadlock:

Thread A                          Thread B

rcu_read_lock()
                                  ust_lock()
                                  synchronize_rcu() (blocked by rcu
                                                     read-side lock)
ust_lock()   <-- deadlock

Introduce a new lttng_perf_lock to protect the lttng perf context
data structures from concurrent modifications and from fork. This
lock can be nested within the ust_lock, but never the opposite.

This removes the circular locking dependency involving urcu bp.

Fixes: #1202
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
liblttng-ust/lttng-context-perf-counters.c
liblttng-ust/lttng-tracer-core.h
liblttng-ust/lttng-ust-comm.c

index 4816f89b5a4a536e4cc2741e6e5d658379692baf..a6ff55b678d472d29fb157f7b0527918432121e2 100644 (file)
@@ -39,6 +39,7 @@
 #include <urcu/ref.h>
 #include <usterr-signal-safe.h>
 #include <signal.h>
+#include <urcu/tls-compat.h>
 #include "perf_event.h"
 #include "lttng-tracer-core.h"
 
@@ -71,6 +72,98 @@ struct lttng_perf_counter_field {
 
 static pthread_key_t perf_counter_key;
 
+/*
+ * lttng_perf_lock - Protect lttng-ust perf counter data structures
+ *
+ * Nests within the ust_lock, and therefore within the libc dl lock.
+ * Therefore, we need to fixup the TLS before nesting into this lock.
+ * Nests inside RCU bp read-side lock. Protects against concurrent
+ * fork.
+ */
+static pthread_mutex_t ust_perf_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+/*
+ * Cancel state when grabbing the ust_perf_mutex. Saved when locking,
+ * restored on unlock. Protected by ust_perf_mutex.
+ */
+static int ust_perf_saved_cancelstate;
+
+/*
+ * Track whether we are tracing from a signal handler nested on an
+ * application thread.
+ */
+static DEFINE_URCU_TLS(int, ust_perf_mutex_nest);
+
+/*
+ * Force a read (imply TLS fixup for dlopen) of TLS variables.
+ */
+void lttng_ust_fixup_perf_counter_tls(void)
+{
+       asm volatile ("" : : "m" (URCU_TLS(ust_perf_mutex_nest)));
+}
+
+void lttng_perf_lock(void)
+{
+       sigset_t sig_all_blocked, orig_mask;
+       int ret, oldstate;
+
+       ret = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate);
+       if (ret) {
+               ERR("pthread_setcancelstate: %s", strerror(ret));
+       }
+       sigfillset(&sig_all_blocked);
+       ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
+       if (ret) {
+               ERR("pthread_sigmask: %s", strerror(ret));
+       }
+       if (!URCU_TLS(ust_perf_mutex_nest)++) {
+               /*
+                * Ensure the compiler don't move the store after the close()
+                * call in case close() would be marked as leaf.
+                */
+               cmm_barrier();
+               pthread_mutex_lock(&ust_perf_mutex);
+               ust_perf_saved_cancelstate = oldstate;
+       }
+       ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
+       if (ret) {
+               ERR("pthread_sigmask: %s", strerror(ret));
+       }
+}
+
+void lttng_perf_unlock(void)
+{
+       sigset_t sig_all_blocked, orig_mask;
+       int ret, newstate, oldstate;
+       bool restore_cancel = false;
+
+       sigfillset(&sig_all_blocked);
+       ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
+       if (ret) {
+               ERR("pthread_sigmask: %s", strerror(ret));
+       }
+       /*
+        * Ensure the compiler don't move the store before the close()
+        * call, in case close() would be marked as leaf.
+        */
+       cmm_barrier();
+       if (!--URCU_TLS(ust_perf_mutex_nest)) {
+               newstate = ust_perf_saved_cancelstate;
+               restore_cancel = true;
+               pthread_mutex_unlock(&ust_perf_mutex);
+       }
+       ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
+       if (ret) {
+               ERR("pthread_sigmask: %s", strerror(ret));
+       }
+       if (restore_cancel) {
+               ret = pthread_setcancelstate(newstate, &oldstate);
+               if (ret) {
+                       ERR("pthread_setcancelstate: %s", strerror(ret));
+               }
+       }
+}
+
 static
 size_t perf_counter_get_size(struct lttng_ctx_field *field, size_t offset)
 {
@@ -308,12 +401,12 @@ struct lttng_perf_counter_thread_field *
         * Note: thread_field->pc can be NULL if setup_perf() fails.
         * Also, thread_field->fd can be -1 if open_perf_fd() fails.
         */
-       ust_lock_nocheck();
+       lttng_perf_lock();
        cds_list_add_rcu(&thread_field->rcu_field_node,
                        &perf_thread->rcu_field_list);
        cds_list_add(&thread_field->thread_field_node,
                        &perf_field->thread_field_list);
-       ust_unlock();
+       lttng_perf_unlock();
 skip:
        ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
        if (ret)
@@ -370,7 +463,7 @@ void perf_counter_get_value(struct lttng_ctx_field *field,
        value->u.s64 = wrapper_perf_counter_read(field);
 }
 
-/* Called with UST lock held */
+/* Called with perf lock held */
 static
 void lttng_destroy_perf_thread_field(
                struct lttng_perf_counter_thread_field *thread_field)
@@ -388,11 +481,11 @@ void lttng_destroy_perf_thread_key(void *_key)
        struct lttng_perf_counter_thread *perf_thread = _key;
        struct lttng_perf_counter_thread_field *pos, *p;
 
-       ust_lock_nocheck();
+       lttng_perf_lock();
        cds_list_for_each_entry_safe(pos, p, &perf_thread->rcu_field_list,
                        rcu_field_node)
                lttng_destroy_perf_thread_field(pos);
-       ust_unlock();
+       lttng_perf_unlock();
        free(perf_thread);
 }
 
@@ -408,11 +501,15 @@ void lttng_destroy_perf_counter_field(struct lttng_ctx_field *field)
        /*
         * This put is performed when no threads can concurrently
         * perform a "get" concurrently, thanks to urcu-bp grace
-        * period.
+        * period. Holding the lttng perf lock protects against
+        * concurrent modification of the per-thread thread field
+        * list.
         */
+       lttng_perf_lock();
        cds_list_for_each_entry_safe(pos, p, &perf_field->thread_field_list,
                        thread_field_node)
                lttng_destroy_perf_thread_field(pos);
+       lttng_perf_unlock();
        free(perf_field);
 }
 
index ba232f32acf5dc20067a41d15b496653ca1ee1a8..f750f4b01e89a32e7c2237da8847e1c9e42ce98f 100644 (file)
@@ -64,4 +64,23 @@ void lttng_ust_dummy_get_value(struct lttng_ctx_field *field,
 int lttng_context_is_app(const char *name);
 void lttng_ust_fixup_tls(void);
 
+#ifdef LTTNG_UST_HAVE_PERF_EVENT
+void lttng_ust_fixup_perf_counter_tls(void);
+void lttng_perf_lock(void);
+void lttng_perf_unlock(void);
+#else /* #ifdef LTTNG_UST_HAVE_PERF_EVENT */
+static inline
+void lttng_ust_fixup_perf_counter_tls(void)
+{
+}
+static inline
+void lttng_perf_lock(void)
+{
+}
+static inline
+void lttng_perf_unlock(void)
+{
+}
+#endif /* #else #ifdef LTTNG_UST_HAVE_PERF_EVENT */
+
 #endif /* _LTTNG_TRACER_CORE_H */
index 47ba36e5c9f1d9f1871c267469049708a7dd1e1e..b067b3da1bbc9b71084b121c46a7a530976db041 100644 (file)
@@ -423,6 +423,7 @@ void lttng_ust_fixup_tls(void)
        lttng_fixup_nest_count_tls();
        lttng_fixup_procname_tls();
        lttng_fixup_ust_mutex_nest_tls();
+       lttng_ust_fixup_perf_counter_tls();
        lttng_ust_fixup_fd_tracker_tls();
 }
 
@@ -2068,6 +2069,7 @@ void ust_before_fork(sigset_t *save_sigset)
        ust_lock_nocheck();
        urcu_bp_before_fork();
        lttng_ust_lock_fd_tracker();
+       lttng_perf_lock();
 }
 
 static void ust_after_fork_common(sigset_t *restore_sigset)
@@ -2075,6 +2077,7 @@ static void ust_after_fork_common(sigset_t *restore_sigset)
        int ret;
 
        DBG("process %d", getpid());
+       lttng_perf_unlock();
        lttng_ust_unlock_fd_tracker();
        ust_unlock();
 
This page took 0.028609 seconds and 4 git commands to generate.