Attachment 207916 Details for Bug 240590 – EPOLLHUP.patch

[patch] EPOLLHUP.patch

EPOLLHUP.patch (text/plain), 15.26 KB, created by Vladimir Kondratyev on 2019-09-28 12:39:18 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: Vladimir Kondratyev

Created: 2019-09-28 12:39:18 UTC

Size: 15.26 KB

patch

obsolete

>diff --git a/sys/compat/linux/linux_emul.c b/sys/compat/linux/linux_emul.c
>index e88eb554aed1..39c2a62ed60e 100644
>--- a/sys/compat/linux/linux_emul.c
>+++ b/sys/compat/linux/linux_emul.c
>@@ -92,7 +92,6 @@ linux_proc_init(struct thread *td, struct thread *newtd, int flags)
> {
> 	struct linux_emuldata *em;
> 	struct linux_pemuldata *pem;
>-	struct epoll_emuldata *emd;
> 	struct proc *p;
> 
> 	if (newtd != NULL) {
>@@ -131,15 +130,9 @@ linux_proc_init(struct thread *td, struct thread *newtd, int flags)
> 		em->child_clear_tid = NULL;
> 		em->child_set_tid = NULL;
> 
>-		 /* epoll should be destroyed in a case of exec. */
> 		pem = pem_find(p);
> 		KASSERT(pem != NULL, ("proc_exit: proc emuldata not found.\n"));
> 		pem->persona = 0;
>-		if (pem->epoll != NULL) {
>-			emd = pem->epoll;
>-			pem->epoll = NULL;
>-			free(emd, M_EPOLL);
>-		}
> 	}
> 
> }
>@@ -148,7 +141,6 @@ void
> linux_proc_exit(void *arg __unused, struct proc *p)
> {
> 	struct linux_pemuldata *pem;
>-	struct epoll_emuldata *emd;
> 	struct thread *td = curthread;
> 
> 	if (__predict_false(SV_CURPROC_ABI() != SV_ABI_LINUX))
>@@ -164,12 +156,6 @@ linux_proc_exit(void *arg __unused, struct proc *p)
> 
> 	p->p_emuldata = NULL;
> 
>-	if (pem->epoll != NULL) {
>-		emd = pem->epoll;
>-		pem->epoll = NULL;
>-		free(emd, M_EPOLL);
>-	}
>-
> 	sx_destroy(&pem->pem_sx);
> 	free(pem, M_LINUX);
> }
>@@ -214,7 +200,6 @@ int
> linux_common_execve(struct thread *td, struct image_args *eargs)
> {
> 	struct linux_pemuldata *pem;
>-	struct epoll_emuldata *emd;
> 	struct vmspace *oldvmspace;
> 	struct linux_emuldata *em;
> 	struct proc *p;
>@@ -246,12 +231,6 @@ linux_common_execve(struct thread *td, struct image_args *eargs)
> 		p->p_emuldata = NULL;
> 		PROC_UNLOCK(p);
> 
>-		if (pem->epoll != NULL) {
>-			emd = pem->epoll;
>-			pem->epoll = NULL;
>-			free(emd, M_EPOLL);
>-		}
>-
> 		free(em, M_TEMP);
> 		free(pem, M_LINUX);
> 	}
>diff --git a/sys/compat/linux/linux_emul.h b/sys/compat/linux/linux_emul.h
>index daa92e15b502..247a6c2d10b5 100644
>--- a/sys/compat/linux/linux_emul.h
>+++ b/sys/compat/linux/linux_emul.h
>@@ -68,7 +68,6 @@ int	linux_common_execve(struct thread *, struct image_args *);
> struct linux_pemuldata {
> 	uint32_t	flags;		/* process emuldata flags */
> 	struct sx	pem_sx;		/* lock for this struct */
>-	void		*epoll;		/* epoll data */
> 	uint32_t	persona;	/* process execution domain */
> 	uint32_t	ptrace_flags;	/* used by ptrace(2) */
> };
>diff --git a/sys/compat/linux/linux_event.c b/sys/compat/linux/linux_event.c
>index 7d26ee3e5051..4214727d4f98 100644
>--- a/sys/compat/linux/linux_event.c
>+++ b/sys/compat/linux/linux_event.c
>@@ -67,25 +67,8 @@ __FBSDID("$FreeBSD$");
> #include <compat/linux/linux_timer.h>
> #include <compat/linux/linux_util.h>
> 
>-/*
>- * epoll defines 'struct epoll_event' with the field 'data' as 64 bits
>- * on all architectures. But on 32 bit architectures BSD 'struct kevent' only
>- * has 32 bit opaque pointer as 'udata' field. So we can't pass epoll supplied
>- * data verbatuim. Therefore we allocate 64-bit memory block to pass
>- * user supplied data for every file descriptor.
>- */
>-
> typedef uint64_t	epoll_udata_t;
> 
>-struct epoll_emuldata {
>-	uint32_t	fdc;		/* epoll udata max index */
>-	epoll_udata_t	udata[1];	/* epoll user data vector */
>-};
>-
>-#define	EPOLL_DEF_SZ		16
>-#define	EPOLL_SIZE(fdn)			\
>-	(sizeof(struct epoll_emuldata)+(fdn) * sizeof(epoll_udata_t))
>-
> struct epoll_event {
> 	uint32_t	events;
> 	epoll_udata_t	data;
>@@ -97,15 +80,16 @@ __attribute__((packed))
> 
> #define	LINUX_MAX_EVENTS	(INT_MAX / sizeof(struct epoll_event))
> 
>-static void	epoll_fd_install(struct thread *td, int fd, epoll_udata_t udata);
> static int	epoll_to_kevent(struct thread *td, struct file *epfp,
> 		    int fd, struct epoll_event *l_event, int *kev_flags,
> 		    struct kevent *kevent, int *nkevents);
> static void	kevent_to_epoll(struct kevent *kevent, struct epoll_event *l_event);
> static int	epoll_kev_copyout(void *arg, struct kevent *kevp, int count);
> static int	epoll_kev_copyin(void *arg, struct kevent *kevp, int count);
>-static int	epoll_delete_event(struct thread *td, struct file *epfp,
>-		    int fd, int filter);
>+static int	epoll_register_kevent(struct thread *td, struct file *epfp,
>+		    int fd, int filter, unsigned int flags);
>+static int	epoll_fd_registered(struct thread *td, struct file *epfp,
>+		    struct file *fp, int fd);
> static int	epoll_delete_all_events(struct thread *td, struct file *epfp,
> 		    int fd);
> 
>@@ -220,35 +204,6 @@ static void	linux_timerfd_expire(void *);
> static void	linux_timerfd_curval(struct timerfd *, struct itimerspec *);
> 
> 
>-static void
>-epoll_fd_install(struct thread *td, int fd, epoll_udata_t udata)
>-{
>-	struct linux_pemuldata *pem;
>-	struct epoll_emuldata *emd;
>-	struct proc *p;
>-
>-	p = td->td_proc;
>-
>-	pem = pem_find(p);
>-	KASSERT(pem != NULL, ("epoll proc emuldata not found.\n"));
>-
>-	LINUX_PEM_XLOCK(pem);
>-	if (pem->epoll == NULL) {
>-		emd = malloc(EPOLL_SIZE(fd), M_EPOLL, M_WAITOK);
>-		emd->fdc = fd;
>-		pem->epoll = emd;
>-	} else {
>-		emd = pem->epoll;
>-		if (fd > emd->fdc) {
>-			emd = realloc(emd, EPOLL_SIZE(fd), M_EPOLL, M_WAITOK);
>-			emd->fdc = fd;
>-			pem->epoll = emd;
>-		}
>-	}
>-	emd->udata[fd] = udata;
>-	LINUX_PEM_XUNLOCK(pem);
>-}
>-
> static int
> epoll_create_common(struct thread *td, int flags)
> {
>@@ -258,8 +213,6 @@ epoll_create_common(struct thread *td, int flags)
> 	if (error != 0)
> 		return (error);
> 
>-	epoll_fd_install(td, EPOLL_DEF_SZ, 0);
>-
> 	return (0);
> }
> 
>@@ -296,13 +249,14 @@ linux_epoll_create1(struct thread *td, struct linux_epoll_create1_args *args)
> 
> /* Structure converting function from epoll to kevent. */
> static int
>-epoll_to_kevent(struct thread *td, struct file *epfp,
>+epoll_to_kevent(struct thread *td, struct file *fp,
>     int fd, struct epoll_event *l_event, int *kev_flags,
>     struct kevent *kevent, int *nkevents)
> {
> 	uint32_t levents = l_event->events;
> 	struct linux_pemuldata *pem;
> 	struct proc *p;
>+	bool need_prot = false;
> 
> 	/* flags related to how event is registered */
> 	if ((levents & LINUX_EPOLLONESHOT) != 0)
>@@ -316,11 +270,46 @@ epoll_to_kevent(struct thread *td, struct file *epfp,
> 
> 	/* flags related to what event is registered */
> 	if ((levents & LINUX_EPOLL_EVRD) != 0) {
>-		EV_SET(kevent++, fd, EVFILT_READ, *kev_flags, 0, 0, 0);
>+		EV_SET(kevent, fd, EVFILT_READ, *kev_flags, 0, 0, 0);
>+		kevent->ext[0] = l_event->data;
>+		kevent->ext[1] = fp->f_type;
>+		++kevent;
>+		++(*nkevents);
>+	} else if ((levents & LINUX_EPOLLRDHUP) != 0 &&
>+		    fp->f_type == DTYPE_SOCKET) {
>+		/* Set lowat arbitrary high to block data-flow events */
>+		EV_SET(kevent, fd, EVFILT_READ, *kev_flags, NOTE_LOWAT,
>+		    INT_MAX, 0);
>+		kevent->ext[0] = l_event->data;
>+		kevent->ext[1] = fp->f_type;
>+		++kevent;
> 		++(*nkevents);
>+		need_prot = true;
> 	}
> 	if ((levents & LINUX_EPOLL_EVWR) != 0) {
>-		EV_SET(kevent++, fd, EVFILT_WRITE, *kev_flags, 0, 0, 0);
>+		EV_SET(kevent, fd, EVFILT_WRITE, *kev_flags, 0, 0, 0);
>+		kevent->ext[0] = l_event->data;
>+		kevent->ext[1] = fp->f_type;
>+		++kevent;
>+		++(*nkevents);
>+	} else if (fp->f_type == DTYPE_SOCKET) {
>+		/* Always set EVFILT_WRITE to catch socket's EPOLLHUP events */
>+		EV_SET(kevent, fd, EVFILT_WRITE, *kev_flags, NOTE_LOWAT,
>+		    INT_MAX, 0);
>+		kevent->ext[0] = l_event->data;
>+		kevent->ext[1] = fp->f_type;
>+		++kevent;
>+		++(*nkevents);
>+		if ((levents & LINUX_EPOLL_EVRD) == 0)
>+			need_prot = true;
>+	}
>+	if (need_prot) {
>+		/*
>+		 * Add protective event to prevent clobbering of fflags and
>+		 * data fields of kevents in epoll_fd_registered() routine.
>+		 */
>+		EV_SET(kevent, fd, EVFILT_EMPTY, EV_ADD | EV_DISABLE, 0, 0, 0);
>+		++kevent;
> 		++(*nkevents);
> 	}
> 
>@@ -329,7 +318,6 @@ epoll_to_kevent(struct thread *td, struct file *epfp,
> 
> 		pem = pem_find(p);
> 		KASSERT(pem != NULL, ("epoll proc emuldata not found.\n"));
>-		KASSERT(pem->epoll != NULL, ("epoll proc epolldata not found.\n"));
> 
> 		LINUX_PEM_XLOCK(pem);
> 		if ((pem->flags & LINUX_XUNSUP_EPOLL) == 0) {
>@@ -354,20 +342,33 @@ static void
> kevent_to_epoll(struct kevent *kevent, struct epoll_event *l_event)
> {
> 
>+	l_event->data = kevent->ext[0];
>+
> 	if ((kevent->flags & EV_ERROR) != 0) {
> 		l_event->events = LINUX_EPOLLERR;
> 		return;
> 	}
> 
>-	/* XXX EPOLLPRI, EPOLLHUP */
>+	/* XXX EPOLLPRI */
> 	switch (kevent->filter) {
> 	case EVFILT_READ:
>-		l_event->events = LINUX_EPOLLIN;
> 		if ((kevent->flags & EV_EOF) != 0)
>-			l_event->events |= LINUX_EPOLLRDHUP;
>+			l_event->events = (kevent->ext[1] == DTYPE_SOCKET) ?
>+			    LINUX_EPOLLRDHUP | LINUX_EPOLLIN : LINUX_EPOLLHUP;
>+		else
>+			l_event->events = LINUX_EPOLLIN;
> 	break;
> 	case EVFILT_WRITE:
>-		l_event->events = LINUX_EPOLLOUT;
>+		/*
>+		 * XXX: socket's EPOLLHUP requires EV_EOF from both send and
>+		 * receive directions at the same time. As now only send is
>+		 * taken into account, it is posible to erroneously trigger
>+		 * EPOLLHUP with shutdown(fd, SHUT_WR) call.
>+		 */
>+		if ((kevent->flags & EV_EOF) != 0)
>+			l_event->events = LINUX_EPOLLHUP;
>+		else
>+			l_event->events = LINUX_EPOLLOUT;
> 	break;
> 	}
> }
>@@ -382,30 +383,15 @@ static int
> epoll_kev_copyout(void *arg, struct kevent *kevp, int count)
> {
> 	struct epoll_copyout_args *args;
>-	struct linux_pemuldata *pem;
>-	struct epoll_emuldata *emd;
> 	struct epoll_event *eep;
>-	int error, fd, i;
>+	int error, i;
> 
> 	args = (struct epoll_copyout_args*) arg;
> 	eep = malloc(sizeof(*eep) * count, M_EPOLL, M_WAITOK | M_ZERO);
> 
>-	pem = pem_find(args->p);
>-	KASSERT(pem != NULL, ("epoll proc emuldata not found.\n"));
>-	LINUX_PEM_SLOCK(pem);
>-	emd = pem->epoll;
>-	KASSERT(emd != NULL, ("epoll proc epolldata not found.\n"));
>-
>-	for (i = 0; i < count; i++) {
>+	for (i = 0; i < count; i++)
> 		kevent_to_epoll(&kevp[i], &eep[i]);
> 
>-		fd = kevp[i].ident;
>-		KASSERT(fd <= emd->fdc, ("epoll user data vector"
>-						    " is too small.\n"));
>-		eep[i].data = emd->udata[fd];
>-	}
>-	LINUX_PEM_SUNLOCK(pem);
>-
> 	error = copyout(eep, args->leventlist, count * sizeof(*eep));
> 	if (error == 0) {
> 		args->leventlist += count;
>@@ -445,7 +431,7 @@ linux_epoll_ctl(struct thread *td, struct linux_epoll_ctl_args *args)
> {
> 	struct file *epfp, *fp;
> 	struct epoll_copyin_args ciargs;
>-	struct kevent kev[2];
>+	struct kevent kev[3];
> 	struct kevent_copyops k_ops = { &ciargs,
> 					NULL,
> 					epoll_kev_copyin};
>@@ -485,7 +471,7 @@ linux_epoll_ctl(struct thread *td, struct linux_epoll_ctl_args *args)
> 
> 	if (args->op != LINUX_EPOLL_CTL_DEL) {
> 		kev_flags = EV_ADD | EV_ENABLE;
>-		error = epoll_to_kevent(td, epfp, args->fd, &le,
>+		error = epoll_to_kevent(td, fp, args->fd, &le,
> 		    &kev_flags, kev, &nchanges);
> 		if (error != 0)
> 			goto leave0;
>@@ -499,19 +485,10 @@ linux_epoll_ctl(struct thread *td, struct linux_epoll_ctl_args *args)
> 		break;
> 
> 	case LINUX_EPOLL_CTL_ADD:
>-		/*
>-		 * kqueue_register() return ENOENT if event does not exists
>-		 * and the EV_ADD flag is not set. Reset EV_ENABLE flag to
>-		 * avoid accidental activation of fired oneshot events.
>-		 */
>-		kev[0].flags &= ~(EV_ADD | EV_ENABLE);
>-		error = kqfd_register(args->epfd, &kev[0], td, M_WAITOK);
>-		if (error != ENOENT) {
>+		if (epoll_fd_registered(td, epfp, fp, args->fd)) {
> 			error = EEXIST;
> 			goto leave0;
> 		}
>-		error = 0;
>-		kev[0].flags |= (EV_ADD | EV_ENABLE);
> 		break;
> 
> 	case LINUX_EPOLL_CTL_DEL:
>@@ -524,8 +501,6 @@ linux_epoll_ctl(struct thread *td, struct linux_epoll_ctl_args *args)
> 		goto leave0;
> 	}
> 
>-	epoll_fd_install(td, args->fd, le.data);
>-
> 	error = kern_kevent_fp(td, epfp, nchanges, 0, &k_ops, NULL);
> 
> leave0:
>@@ -562,13 +537,13 @@ linux_epoll_wait_common(struct thread *td, int epfd, struct epoll_event *events,
> 		return (error);
> 	if (epfp->f_type != DTYPE_KQUEUE) {
> 		error = EINVAL;
>-		goto leave1;
>+		goto leave;
> 	}
> 	if (uset != NULL) {
> 		error = kern_sigprocmask(td, SIG_SETMASK, uset,
> 		    &omask, 0);
> 		if (error != 0)
>-			goto leave1;
>+			goto leave;
> 		td->td_pflags |= TDP_OLDMASK;
> 		/*
> 		 * Make sure that ast() is called on return to
>@@ -586,11 +561,7 @@ linux_epoll_wait_common(struct thread *td, int epfd, struct epoll_event *events,
> 	coargs.count = 0;
> 	coargs.error = 0;
> 
>-	if (timeout != -1) {
>-		if (timeout < 0) {
>-			error = EINVAL;
>-			goto leave0;
>-		}
>+	if (timeout >= 0) {
> 		/* Convert from milliseconds to timespec. */
> 		ts.tv_sec = timeout / 1000;
> 		ts.tv_nsec = (timeout % 1000) * 1000000;
>@@ -610,11 +581,10 @@ linux_epoll_wait_common(struct thread *td, int epfd, struct epoll_event *events,
> 	if (error == 0)
> 		td->td_retval[0] = coargs.count;
> 
>-leave0:
> 	if (uset != NULL)
> 		error = kern_sigprocmask(td, SIG_SETMASK, &omask,
> 		    NULL, 0);
>-leave1:
>+leave:
> 	fdrop(epfp, td);
> 	return (error);
> }
>@@ -651,7 +621,8 @@ linux_epoll_pwait(struct thread *td, struct linux_epoll_pwait_args *args)
> }
> 
> static int
>-epoll_delete_event(struct thread *td, struct file *epfp, int fd, int filter)
>+epoll_register_kevent(struct thread *td, struct file *epfp, int fd, int filter,
>+    unsigned int flags)
> {
> 	struct epoll_copyin_args ciargs;
> 	struct kevent kev;
>@@ -660,21 +631,45 @@ epoll_delete_event(struct thread *td, struct file *epfp, int fd, int filter)
> 					epoll_kev_copyin};
> 
> 	ciargs.changelist = &kev;
>-	EV_SET(&kev, fd, filter, EV_DELETE | EV_DISABLE, 0, 0, 0);
>+	EV_SET(&kev, fd, filter, flags, 0, 0, 0);
> 
> 	return (kern_kevent_fp(td, epfp, 1, 0, &k_ops, NULL));
> }
> 
>+static int
>+epoll_fd_registered(struct thread *td, struct file *epfp, struct file *fp,
>+    int fd)
>+{
>+	/*
>+	 * Set empty filter flags to avoid accidental modification of already
>+	 * registered events. In the case of re-registration:
>+	 * 1. If event does not exists kevent() does nothing and returns ENOENT
>+	 * 2. If event does exists, it's enabled/disabled state is preserved
>+	 *    but fflags, data and udata fields are overwritten.
>+	 * p.2 means that we can not store user's context pointer in udata.
>+	 * The order of checking is important for sockets and should be in line
>+	 * with epoll_to_kevent() routine.
>+	 */
>+	if ((fp->f_type == DTYPE_SOCKET &&
>+	    epoll_register_kevent(td, epfp, fd, EVFILT_EMPTY, 0) != ENOENT) ||
>+	    epoll_register_kevent(td, epfp, fd, EVFILT_READ, 0) != ENOENT ||
>+	    epoll_register_kevent(td, epfp, fd, EVFILT_WRITE, 0) != ENOENT)
>+		return (1);
>+
>+	return (0);
>+}
>+
> static int
> epoll_delete_all_events(struct thread *td, struct file *epfp, int fd)
> {
>-	int error1, error2;
>+	int error1, error2, error3;
> 
>-	error1 = epoll_delete_event(td, epfp, fd, EVFILT_READ);
>-	error2 = epoll_delete_event(td, epfp, fd, EVFILT_WRITE);
>+	error1 = epoll_register_kevent(td, epfp, fd, EVFILT_READ, EV_DELETE);
>+	error2 = epoll_register_kevent(td, epfp, fd, EVFILT_WRITE, EV_DELETE);
>+	error3 = epoll_register_kevent(td, epfp, fd, EVFILT_EMPTY, EV_DELETE);
> 
> 	/* return 0 if at least one result positive */
>-	return (error1 == 0 ? 0 : error2);
>+	return (error1 == 0 ? 0 : (error2 == 0 ? 0 : error3));
> }
> 
> static int
>diff --git a/sys/compat/linux/linux_event.h b/sys/compat/linux/linux_event.h
>index c483df58736b..e78b3320657d 100644
>--- a/sys/compat/linux/linux_event.h
>+++ b/sys/compat/linux/linux_event.h
>@@ -45,11 +45,11 @@
> #define	LINUX_EPOLLONESHOT	1u<<30
> #define	LINUX_EPOLLET		1u<<31
> 
>-#define	LINUX_EPOLL_EVRD	(LINUX_EPOLLIN|LINUX_EPOLLRDNORM	\
>-		|LINUX_EPOLLHUP|LINUX_EPOLLERR|LINUX_EPOLLPRI)
>+#define	LINUX_EPOLL_EVRD	(LINUX_EPOLLIN|LINUX_EPOLLRDNORM)
> #define	LINUX_EPOLL_EVWR	(LINUX_EPOLLOUT|LINUX_EPOLLWRNORM)
> #define	LINUX_EPOLL_EVSUP	(LINUX_EPOLLET|LINUX_EPOLLONESHOT	\
>-		|LINUX_EPOLL_EVRD|LINUX_EPOLL_EVWR|LINUX_EPOLLRDHUP)
>+		|LINUX_EPOLL_EVRD|LINUX_EPOLL_EVWR|LINUX_EPOLLRDHUP	\
>+		|LINUX_EPOLLHUP|LINUX_EPOLLERR|LINUX_EPOLLPRI)
> 
> #define	LINUX_EPOLL_CTL_ADD	1
> #define	LINUX_EPOLL_CTL_DEL	2

Actions: View | Diff

Attachments on bug 240590: 207708 | 207916 | 207926