From e8424e2947dd8a0e1d3f08295432941b5c5aafb9 Mon Sep 17 00:00:00 2001
From: John Baldwin <jhb@FreeBSD.org>
Date: Fri, 11 Jun 2021 17:58:54 -0700
Subject: bhyve: Register new kevents synchronously.

Change mevent_add*() to synchronously add the new kevent.  This
permits reporting event registration failures to the caller and avoids
failing the registration of other, unrelated events queued up in the
same batch.

Reviewed by:	grehan, markj
MFC after:	1 month
Differential Revision:	https://reviews.freebsd.org/D30502
---
 usr.sbin/bhyve/mevent.c | 96 ++++++++++++++++++++++++++++++++-----------------
 1 file changed, 63 insertions(+), 33 deletions(-)

(limited to 'usr.sbin/bhyve')

diff --git a/usr.sbin/bhyve/mevent.c b/usr.sbin/bhyve/mevent.c
index 2e9059837011..a394b7c5d69f 100644
--- a/usr.sbin/bhyve/mevent.c
+++ b/usr.sbin/bhyve/mevent.c
@@ -64,8 +64,10 @@ __FBSDID("$FreeBSD$");
 #define	MEVENT_MAX	64
 
 static pthread_t mevent_tid;
+static pthread_once_t mevent_once = PTHREAD_ONCE_INIT;
 static int mevent_timid = 43;
 static int mevent_pipefd[2];
+static int mfd;
 static pthread_mutex_t mevent_lmutex = PTHREAD_MUTEX_INITIALIZER;
 
 struct mevent {
@@ -124,6 +126,26 @@ mevent_notify(void)
 	}
 }
 
+static void
+mevent_init(void)
+{
+#ifndef WITHOUT_CAPSICUM
+	cap_rights_t rights;
+#endif
+
+	mfd = kqueue();
+	assert(mfd > 0);
+
+#ifndef WITHOUT_CAPSICUM
+	cap_rights_init(&rights, CAP_KQUEUE);
+	if (caph_rights_limit(mfd, &rights) == -1)
+		errx(EX_OSERR, "Unable to apply rights for sandbox");
+#endif
+
+	LIST_INIT(&change_head);
+	LIST_INIT(&global_head);
+}
+
 static int
 mevent_kq_filter(struct mevent *mevp)
 {
@@ -159,8 +181,24 @@ mevent_kq_fflags(struct mevent *mevp)
 	return (0);
 }
 
+static void
+mevent_populate(struct mevent *mevp, struct kevent *kev)
+{
+	if (mevp->me_type == EVF_TIMER) {
+		kev->ident = mevp->me_timid;
+		kev->data = mevp->me_msecs;
+	} else {
+		kev->ident = mevp->me_fd;
+		kev->data = 0;
+	}
+	kev->filter = mevent_kq_filter(mevp);
+	kev->flags = mevent_kq_flags(mevp);
+	kev->fflags = mevent_kq_fflags(mevp);
+	kev->udata = mevp;
+}
+
 static int
-mevent_build(int mfd, struct kevent *kev)
+mevent_build(struct kevent *kev)
 {
 	struct mevent *mevp, *tmpp;
 	int i;
@@ -177,17 +215,8 @@ mevent_build(int mfd, struct kevent *kev)
 			 */
 			close(mevp->me_fd);
 		} else {
-			if (mevp->me_type == EVF_TIMER) {
-				kev[i].ident = mevp->me_timid;
-				kev[i].data = mevp->me_msecs;
-			} else {
-				kev[i].ident = mevp->me_fd;
-				kev[i].data = 0;
-			}
-			kev[i].filter = mevent_kq_filter(mevp);
-			kev[i].flags = mevent_kq_flags(mevp);
-			kev[i].fflags = mevent_kq_fflags(mevp);
-			kev[i].udata = mevp;
+			assert((mevp->me_state & EV_ADD) == 0);
+			mevent_populate(mevp, &kev[i]);
 			i++;
 		}
 
@@ -197,12 +226,6 @@ mevent_build(int mfd, struct kevent *kev)
 		if (mevp->me_state & EV_DELETE) {
 			free(mevp);
 		} else {
-			/*
-			 * We need to add the event only once, so we can
-			 * reset the EV_ADD bit after it has been propagated
-			 * to the kevent() arguments the first time.
-			 */
-			mevp->me_state &= ~EV_ADD;
 			LIST_INSERT_HEAD(&global_head, mevp, me_list);
 		}
 
@@ -234,7 +257,9 @@ mevent_add_state(int tfd, enum ev_type type,
 	   void (*func)(int, enum ev_type, void *), void *param,
 	   int state)
 {
+	struct kevent kev;
 	struct mevent *lp, *mevp;
+	int ret;
 
 	if (tfd < 0 || func == NULL) {
 		return (NULL);
@@ -242,6 +267,8 @@ mevent_add_state(int tfd, enum ev_type type,
 
 	mevp = NULL;
 
+	pthread_once(&mevent_once, mevent_init);
+
 	mevent_qlock();
 
 	/*
@@ -262,7 +289,7 @@ mevent_add_state(int tfd, enum ev_type type,
 	}
 
 	/*
-	 * Allocate an entry, populate it, and add it to the change list.
+	 * Allocate an entry and populate it.
 	 */
 	mevp = calloc(1, sizeof(struct mevent));
 	if (mevp == NULL) {
@@ -277,11 +304,22 @@ mevent_add_state(int tfd, enum ev_type type,
 	mevp->me_type = type;
 	mevp->me_func = func;
 	mevp->me_param = param;
-
-	LIST_INSERT_HEAD(&change_head, mevp, me_list);
-	mevp->me_cq = 1;
 	mevp->me_state = state;
-	mevent_notify();
+
+	/*
+	 * Try to add the event.  If this fails, report the failure to
+	 * the caller.
+	 */
+	mevent_populate(mevp, &kev);
+	ret = kevent(mfd, &kev, 1, NULL, 0, NULL);
+	if (ret == -1) {
+		free(mevp);
+		mevp = NULL;
+		goto exit;
+	}
+
+	mevp->me_state &= ~EV_ADD;
+	LIST_INSERT_HEAD(&global_head, mevp, me_list);
 
 exit:
 	mevent_qunlock();
@@ -415,7 +453,6 @@ mevent_dispatch(void)
 	struct kevent changelist[MEVENT_MAX];
 	struct kevent eventlist[MEVENT_MAX];
 	struct mevent *pipev;
-	int mfd;
 	int numev;
 	int ret;
 #ifndef WITHOUT_CAPSICUM
@@ -425,14 +462,7 @@ mevent_dispatch(void)
 	mevent_tid = pthread_self();
 	mevent_set_name();
 
-	mfd = kqueue();
-	assert(mfd > 0);
-
-#ifndef WITHOUT_CAPSICUM
-	cap_rights_init(&rights, CAP_KQUEUE);
-	if (caph_rights_limit(mfd, &rights) == -1)
-		errx(EX_OSERR, "Unable to apply rights for sandbox");
-#endif
+	pthread_once(&mevent_once, mevent_init);
 
 	/*
 	 * Open the pipe that will be used for other threads to force
@@ -466,7 +496,7 @@ mevent_dispatch(void)
 		 * to eliminate the extra syscall. Currently better for
 		 * debug.
 		 */
-		numev = mevent_build(mfd, changelist);
+		numev = mevent_build(changelist);
 		if (numev) {
 			ret = kevent(mfd, changelist, numev, NULL, 0, NULL);
 			if (ret == -1) {
-- 
cgit v1.2.3