OS-5431 under many conditions, ptrace (strace) on Ubuntu 16 shows syscalls return ENOSYS

Reviewed by: Patrick Mooney <patrick.mooney@joyent.com>

5 files changed, 109 insertions, 11 deletions

diff --git a/usr/src/lib/brand/lx/lx_brand/common/signal.c b/usr/src/lib/brand/lx/lx_brand/common/signal.c
index 80f1bf09a1..7d07dfe66d 100644
--- a/usr/src/lib/brand/lx/lx_brand/common/signal.c
+++ b/usr/src/lib/brand/lx/lx_brand/common/signal.c
@@ -25,7 +25,7 @@
  */
 
 /*
- * Copyright 2015 Joyent, Inc. All rights reserved.
+ * Copyright 2016 Joyent, Inc. All rights reserved.
  */
 
 #include <sys/types.h>
@@ -1077,6 +1077,12 @@ lx_rt_sigreturn(void)
 	ucontext_t *retucp;
 	uintptr_t sp;
 
+	/*
+	 * Since we don't take the normal return path from this syscall, we
+	 * inform the kernel that we're returning, for the sake of ptrace.
+	 */
+	(void) syscall(SYS_brand, B_PTRACE_SIG_RETURN);
+
 	/* Get the registers at the emulated Linux rt_sigreturn syscall */
 	ucp = lx_syscall_regs();
 
diff --git a/usr/src/uts/common/brand/lx/os/lx_brand.c b/usr/src/uts/common/brand/lx/os/lx_brand.c
index c74d5e8ebd..71290b0bdc 100644
--- a/usr/src/uts/common/brand/lx/os/lx_brand.c
+++ b/usr/src/uts/common/brand/lx/os/lx_brand.c
@@ -1485,6 +1485,22 @@ lx_brandsys(int cmd, int64_t *rval, uintptr_t arg1, uintptr_t arg2,
 		return (lx_ptrace_set_clone_inherit((int)arg1, arg2 == 0 ?
 		    B_FALSE : B_TRUE));
 
+	case B_PTRACE_SIG_RETURN: {
+		/*
+		 * Our ptrace emulation must emit PR_SYSEXIT for rt_sigreturn.
+		 * Since that syscall does not pass through the normal
+		 * emulation, which would call lx_syscall_return, the event is
+		 * emitted manually. A successful result of the syscall is
+		 * assumed since there is little to be done in the face of
+		 * failure.
+		 */
+		struct regs *rp = lwptoregs(lwp);
+
+		rp->r_r0 = 0;
+		lx_ptrace_stop(LX_PR_SYSEXIT);
+		return (0);
+	}
+
 	case B_UNSUPPORTED: {
 		char dmsg[256];
 
diff --git a/usr/src/uts/common/brand/lx/os/lx_misc.c b/usr/src/uts/common/brand/lx/os/lx_misc.c
index 267c720e67..5932495655 100644
--- a/usr/src/uts/common/brand/lx/os/lx_misc.c
+++ b/usr/src/uts/common/brand/lx/os/lx_misc.c
@@ -550,6 +550,11 @@ lx_forklwp(klwp_t *srclwp, klwp_t *dstlwp)
 	 */
 	dst->br_lwp_flags = src->br_lwp_flags & BR_CPU_BOUND;
 	dst->br_scall_args = NULL;
+
+	/*
+	 * Flag so child doesn't ptrace-stop on syscall exit.
+	 */
+	dst->br_ptrace_flags |= LX_PTF_NOSTOP;
 }
 
 /*
@@ -1078,10 +1083,27 @@ lx_regs_location(lx_lwp_data_t *lwpd, void **ucp, boolean_t for_write)
 			/* setting registers not allowed in this state */
 			break;
 		}
-		if (lwpd->br_ptrace_whystop == PR_BRAND &&
-		    lwpd->br_ptrace_whatstop == LX_PR_EVENT) {
+		if (lwpd->br_ptrace_whystop == PR_BRAND) {
 			/* Called while ptrace-event-stopped by lx_exec. */
-			return (LX_REG_LOC_LWP);
+			if (lwpd->br_ptrace_whatstop == LX_PR_EVENT) {
+				return (LX_REG_LOC_LWP);
+			}
+
+			/* Called while ptrace-event-stopped after clone. */
+			if (lwpd->br_ptrace_whatstop == LX_PR_SIGNALLED &&
+			    lwpd->br_ptrace_stopsig == LX_SIGSTOP &&
+			    (lwpd->br_ptrace_flags & LX_PTF_STOPPED)) {
+				return (LX_REG_LOC_LWP);
+			}
+
+			/*
+			 * Called to obtain syscall exit for other cases
+			 * (e.g. pseudo return from rt_sigreturn).
+			 */
+			if (lwpd->br_ptrace_whatstop == LX_PR_SYSEXIT &&
+			    (lwpd->br_ptrace_flags & LX_PTF_STOPPED)) {
+				return (LX_REG_LOC_LWP);
+			}
 		}
 		break;
 	default:
diff --git a/usr/src/uts/common/brand/lx/os/lx_ptrace.c b/usr/src/uts/common/brand/lx/os/lx_ptrace.c
index 0f521df61b..e619149bb8 100644
--- a/usr/src/uts/common/brand/lx/os/lx_ptrace.c
+++ b/usr/src/uts/common/brand/lx/os/lx_ptrace.c
@@ -127,10 +127,10 @@
  *
  * Various actions, either directly ptrace(2) related or commonly associated
  * with tracing, cause process- or thread-directed SIGSTOP signals to be sent
- * to tracees.  These signals, and indeed any signal other than SIGKILL, can
- * be suppressed by the tracer when using a restarting request (including
- * PTRACE_DETACH) on a child.  The signal may also be substituted for a
- * different signal.
+ * to tracees (a "signal-delivery-stop"). These signals, and indeed any signal
+ * other than SIGKILL, can be suppressed by the tracer when using a restarting
+ * request (including PTRACE_DETACH) on a child.  The signal may also be
+ * substituted for a different signal.
  *
  * If a SIGSTOP (or other stopping signal) is not suppressed by the tracer,
  * it will induce the regular illumos native job control stop of the entire
@@ -148,6 +148,20 @@
  * cleared after each stop; for ongoing system call tracing the tracee must
  * be continuously restarted with PTRACE_SYSCALL.
  *
+ * SPECIAL CASES FOR STOP EVENTS
+ *
+ * The strace command is one of the primary consumers of ptrace. In order for
+ * strace to properly understand what is actually happening when it receives a
+ * signal associated with a stop event, these signals must match Linux behavior
+ * exactly or the strace consumer will get out of sync and report incorrect
+ * state. There are a couple of special cases we have to handle to provide
+ * proper interaction of the syscall-entry-stop, syscall-exit-stop, and
+ * signal-delivery-stop events:
+ * 1) The child process of a clone/fork does not emit a syscall-exit-stop event.
+ * 2) A signal that arrives between syscall-enter-stop & syscall-exit-stop must
+ *    not immediately emit signal-delivery-stop. This event must be emitted
+ *    after the syscall is interrupted and syscall-exit-stop has been emitted.
+ *
  * EVENT STOPS
  *
  * Various events (particularly FORK, VFORK, CLONE, EXEC and EXIT) are
@@ -1042,6 +1056,9 @@ lx_ptrace_attach(pid_t lx_pid)
 		rlwpd->br_ptrace_attach = LX_PTA_ATTACH;
 		rlwpd->br_ptrace_tracer = accord;
 
+		/* Don't emit ptrace syscall-stop-exit event on kernel exit. */
+		rlwpd->br_ptrace_flags |= LX_PTF_NOSTOP;
+
 		/*
 		 * We had no tracer, and are thus not in the tracees list.
 		 * It is safe to take the tracee list lock while we insert
@@ -1568,6 +1585,16 @@ lx_ptrace_stop(ushort_t what)
 	 * Lock this process and re-check the condition.
 	 */
 	mutex_enter(&p->p_lock);
+
+	/*
+	 * The child after a fork/clone doesn't emit syscall-exit-stop event.
+	 */
+	if (what == LX_PR_SYSEXIT && (lwpd->br_ptrace_flags & LX_PTF_NOSTOP)) {
+		lwpd->br_ptrace_flags &= ~LX_PTF_NOSTOP;
+		mutex_exit(&p->p_lock);
+		return (B_FALSE);
+	}
+
 	if (lwpd->br_ptrace_tracer == NULL) {
 		VERIFY0(lwpd->br_ptrace_flags & LX_PTF_SYSCALL);
 		mutex_exit(&p->p_lock);
@@ -1575,6 +1602,12 @@ lx_ptrace_stop(ushort_t what)
 	}
 
 	if (what == LX_PR_SYSENTRY || what == LX_PR_SYSEXIT) {
+		if (what == LX_PR_SYSENTRY) {
+			lwpd->br_ptrace_flags |= LX_PTF_INSYSCALL;
+		} else {
+			lwpd->br_ptrace_flags &= ~LX_PTF_INSYSCALL;
+		}
+
 		/*
 		 * This is a syscall-entry-stop or syscall-exit-stop point.
 		 */
@@ -1642,6 +1675,17 @@ lx_ptrace_issig_stop(proc_t *p, klwp_t *lwp)
 	}
 
 	/*
+	 * We can't deliver the signal-delivery-stop condition while we're
+	 * between the syscall-enter-stop and syscall-exit-stop conditions.
+	 * We must first let the signal interrupt the in-progress syscall, let
+	 * it emit syscall-exit-stop with the interrupted result, then we'll
+	 * come back here to emit signal-delivery-stop.
+	 */
+	if (lwpd->br_ptrace_flags & LX_PTF_INSYSCALL) {
+		return (0);
+	}
+
+	/*
 	 * We stash the signal on the LWP where our waitid_helper will find it
 	 * and enter the ptrace "signal-delivery-stop" condition.
 	 */
@@ -2401,6 +2445,14 @@ lx_ptrace_kernel(int ptrace_op, pid_t lxpid, uintptr_t addr, uintptr_t data)
 		return (ESRCH);
 	}
 
+	if (ptrace_op == LX_PTRACE_DETACH) {
+		/*
+		 * We're detaching, make sure in-syscall flag is off so that
+		 * signal will stop the process directly.
+		 */
+		remote->br_ptrace_flags &= ~LX_PTF_INSYSCALL;
+	}
+
 	/*
 	 * Attempt to lock the target LWP.
 	 */
diff --git a/usr/src/uts/common/brand/lx/sys/lx_brand.h b/usr/src/uts/common/brand/lx/sys/lx_brand.h
index 88546dbe80..01b74f727e 100644
--- a/usr/src/uts/common/brand/lx/sys/lx_brand.h
+++ b/usr/src/uts/common/brand/lx/sys/lx_brand.h
@@ -108,7 +108,7 @@ extern "C" {
 #define	B_SET_NATIVE_STACK	147
 #define	B_SIGEV_THREAD_ID	148
 #define	B_OVERRIDE_KERN_VER	149
-/* formerly B_NOTIFY_VDSO_LOC	150 */
+#define	B_PTRACE_SIG_RETURN	150
 #define	B_GET_PERSONALITY	151
 
 #ifndef _ASM
@@ -379,7 +379,7 @@ typedef enum lx_accord_flags {
  * Flags values for "br_ptrace_flags" in the LWP-specific data.
  */
 typedef enum lx_ptrace_flags {
-	LX_PTF_SYSCALL = 0x01,
+	LX_PTF_SYSCALL = 0x01,		/* handling syscall or a trap */
 	LX_PTF_EXITING = 0x02,
 	LX_PTF_STOPPING = 0x04,
 	LX_PTF_INHERIT = 0x08,
@@ -387,7 +387,9 @@ typedef enum lx_ptrace_flags {
 	LX_PTF_PARENT_WAIT = 0x20,
 	LX_PTF_CLDPEND = 0x40,
 	LX_PTF_CLONING = 0x80,
-	LX_PTF_WAITPEND = 0x100
+	LX_PTF_WAITPEND = 0x100,
+	LX_PTF_NOSTOP = 0x200,		/* disable syscall stop event */
+	LX_PTF_INSYSCALL = 0x400	/* between syscall enter & exit */
 } lx_ptrace_flags_t;
 
 /*