[DTrace-devel] [PATCH v3] dtrace provider: add a predicate against the current tgid

[Nick Alcock]

If we don't put a predicate on, BEGIN/END probes fire in every running
dtrace at the same time, messing up the activity state of all but the one it
was meant to fire in and often causing the others to fail to exit on exit()
(they hang until ended by some other means, like an interrupt or -c
termination).

Thankfully the -xcpu run-BEGIN/END-in-a-thread complexities can be ignored
because we can match on DTrace's tgid instead of its PID (thread ID),
which will always catch exactly our BEGIN/END firings and no-one else's.

(In theory this might cause trouble if you run multiple consumers in
different threads in the same process, but that's not going to work as it
is, and has never been considered a sane thing to do.)

Signed-off-by: Nick Alcock <nick.alcock at oracle.com>
---
 libdtrace/dt_prov_dtrace.c | 19 +++++++++++++++++--
 1 file changed, 17 insertions(+), 2 deletions(-)

diff --git a/libdtrace/dt_prov_dtrace.c b/libdtrace/dt_prov_dtrace.c
index 307ad558874f3..3e6ae88a37848 100644
--- a/libdtrace/dt_prov_dtrace.c
+++ b/libdtrace/dt_prov_dtrace.c
@@ -93,7 +93,8 @@ static int trampoline(dt_pcb_t *pcb, uint_t exitlbl)
 	}
 
 	/*
-	 * The BEGIN probe should only run when the activity state is INACTIVE.
+	 * The BEGIN probe should only run when the activity state is INACTIVE,
+	 * for this process's PID (TGID).
 	 * At the end of the trampoline (after executing any clauses), the
 	 * state must be advanced to the next state (INACTIVE -> ACTIVE, or if
 	 * there was an exit() action in the clause, DRAINING -> STOPPED).
@@ -102,7 +103,8 @@ static int trampoline(dt_pcb_t *pcb, uint_t exitlbl)
 	 * on in state[DT_STATE_BEGANON] to ensure that we know which trace
 	 * data buffer to process first.
 	 *
-	 * The END probe should only run when the activity state is DRAINING.
+	 * The END probe should only run when the activity state is DRAINING,
+	 * for this process's PID (TGID).
 	 * At the end of the trampoline (after executing any clauses), the
 	 * state must be advanced to the next state (DRAINING -> STOPPED).
 	 *
@@ -118,6 +120,19 @@ static int trampoline(dt_pcb_t *pcb, uint_t exitlbl)
 		key = DT_STATE_ENDEDON;
 	}
 
+	/*
+	 * Retrieve the PID (TGID) of the process that caused the probe to fire,
+	 * and check it against the PID we're meant to be using.  Do it before
+	 * the trampoline to minimize the cost of pointless firings in other
+	 * tracers, even though this means preserving the context in %r1 around
+	 * the call.
+	 */
+	emit(dlp, BPF_MOV_REG(BPF_REG_6, BPF_REG_1));
+	emit(dlp, BPF_CALL_HELPER(BPF_FUNC_get_current_pid_tgid));
+	emit(dlp, BPF_ALU64_IMM(BPF_RSH, BPF_REG_0, 32));
+	emit(dlp, BPF_BRANCH_IMM(BPF_JNE, BPF_REG_0, getpid(), pcb->pcb_exitlbl));
+	emit(dlp, BPF_MOV_REG(BPF_REG_1, BPF_REG_6));
+
 	dt_cg_tramp_prologue_act(pcb, act);
 
 	/*
-- 
2.43.0.272.gce700b77fd