OS-4855 lxbrand convert epoll_ctl to IKE

Reviewed by: Jerry Jelinek <jerry.jelinek@joyent.com>

usr/src/lib/brand/lx/lx_brand/common/lx_brand.c

@@ -1173,7 +1173,7 @@ static lx_syscall_handler_t lx_handlers[] = {
 	lx_clock_nanosleep,		/* 230: clock_nanosleep */
 	lx_group_exit,			/* 231: exit_group */
 	NULL,				/* 232: epoll_wait */
-	lx_epoll_ctl,			/* 233: epoll_ctl */
+	NULL,				/* 233: epoll_ctl */
 	NULL,				/* 234: tgkill */
 	lx_utimes,			/* 235: utimes */
 	NULL,				/* 236: vserver */
@@ -1526,7 +1526,7 @@ static lx_syscall_handler_t lx_handlers[] = {
 	lx_group_exit,			/* 252: group_exit */
 	NULL,				/* 253: lookup_dcookie */
 	NULL,				/* 254: epoll_create */
-	lx_epoll_ctl,			/* 255: epoll_ctl */
+	NULL,				/* 255: epoll_ctl */
 	NULL,				/* 256: epoll_wait */
 	NULL,				/* 257: remap_file_pages */
 	NULL,				/* 258: set_tid_address */

usr/src/lib/brand/lx/lx_brand/common/poll_select.c

@@ -40,7 +40,6 @@
 #include <sys/brand.h>
 #include <sys/poll.h>
 #include <sys/syscall.h>
-#include <sys/epoll.h>
 #include <sys/lx_debug.h>
 #include <sys/lx_poll.h>
 #include <sys/lx_syscall.h>
@@ -288,35 +287,3 @@ lx_poll(uintptr_t p1, uintptr_t p2, uintptr_t p3)
 		free(sfds);
 	return (res);
 }
-
-long
-lx_epoll_ctl(uintptr_t p1, uintptr_t p2, uintptr_t p3, uintptr_t p4)
-{
-	int epfd = (int)p1;
-	int op = (int)p2;
-	int fd = (int)p3;
-	int rval;
-	struct epoll_event ev;
-
-	/*
-	 * Linux limits the max number of open files to 1m so we can also test
-	 * for this.
-	 */
-	if (epfd < 0 || fd < 0 || epfd > (1024 * 1024) || fd > (1024 * 1024))
-		return (-EBADF);
-
-	if (epfd == fd)
-		return (-EINVAL);
-
-	/*
-	 * Unlike the base epoll_ctl, we need to return a fault if the
-	 * event pointer is invalid.
-	 */
-	if (op != EPOLL_CTL_DEL) {
-		if (uucopy((void *)p4, &ev, sizeof (ev)) != 0)
-			return (-errno);
-	}
-
-	rval = epoll_ctl(epfd, op, fd, (struct epoll_event *)p4);
-	return ((rval < 0) ? -errno : rval);
-}

usr/src/lib/brand/lx/lx_brand/sys/lx_syscall.h

@@ -103,7 +103,6 @@ extern long lx_pselect6(uintptr_t, uintptr_t, uintptr_t, uintptr_t,
     uintptr_t, uintptr_t);
 extern long lx_poll(uintptr_t, uintptr_t, uintptr_t);
 extern long lx_ppoll(uintptr_t, uintptr_t, uintptr_t, uintptr_t, uintptr_t);
-extern long lx_epoll_ctl(uintptr_t, uintptr_t, uintptr_t, uintptr_t);
 extern long lx_settimeofday(uintptr_t, uintptr_t);
 extern long lx_getrusage(uintptr_t, uintptr_t);
 extern long lx_mknod(uintptr_t, uintptr_t, uintptr_t);

usr/src/uts/common/brand/lx/os/lx_syscall.c

@@ -778,7 +778,7 @@ lx_sysent_t lx_sysent32[] = {
 	{"group_exit",	NULL,			0,		1}, /* 252 */
 	{"lookup_dcookie", NULL,		NOSYS_NO_EQUIV,	0}, /* 253 */
 	{"epoll_create", lx_epoll_create,	0,		1}, /* 254 */
-	{"epoll_ctl",	NULL,			0,		4}, /* 255 */
+	{"epoll_ctl",	lx_epoll_ctl,		0,		4}, /* 255 */
 	{"epoll_wait",	lx_epoll_wait,		0,		4}, /* 256 */
 	{"remap_file_pages", NULL,		NOSYS_NO_EQUIV,	0}, /* 257 */
 	{"set_tid_address", lx_set_tid_address,	0,		1}, /* 258 */
@@ -1127,7 +1127,7 @@ lx_sysent_t lx_sysent64[] = {
 	{"clock_nanosleep", NULL,		0,		4}, /* 230 */
 	{"exit_group",	NULL,			0,		1}, /* 231 */
 	{"epoll_wait",	lx_epoll_wait,		0,		4}, /* 232 */
-	{"epoll_ctl",	NULL,			0,		4}, /* 233 */
+	{"epoll_ctl",	lx_epoll_ctl,		0,		4}, /* 233 */
 	{"tgkill",	lx_tgkill,		0,		3}, /* 234 */
 	{"utimes",	NULL,			0,		2}, /* 235 */
 	{"vserver",	NULL,			NOSYS_NULL,	0}, /* 236 */

usr/src/uts/common/brand/lx/sys/lx_syscalls.h

@@ -46,6 +46,7 @@ extern long lx_clock_settime();
 extern long lx_connect();
 extern long lx_epoll_create();
 extern long lx_epoll_create1();
+extern long lx_epoll_ctl();
 extern long lx_epoll_pwait();
 extern long lx_epoll_wait();
 extern long lx_fallocate();

usr/src/uts/common/brand/lx/syscall/lx_epoll.c

@@ -23,12 +23,25 @@
 #include <sys/devpoll.h>
 #include <sys/fcntl.h>
 #include <sys/file.h>
+#include <sys/ddi.h>
 #include <sys/sunddi.h>
 #include <sys/sunldi.h>
 #include <sys/vnode.h>
 #include <sys/lx_brand.h>
 #include <sys/lx_types.h>
 
+static major_t devpoll_major = 0;
+
+static boolean_t
+lx_epoll_isvalid(file_t *fp)
+{
+	vnode_t *vp = fp->f_vnode;
+
+	if (vp->v_type == VCHR && getmajor(vp->v_rdev) == devpoll_major)
+		return (B_TRUE);
+	return (B_FALSE);
+}
+
 long
 lx_epoll_create1(int flags)
 {
@@ -64,6 +77,8 @@ lx_epoll_create1(int flags)
 		goto error;
 	}
 
+	devpoll_major = getmajor(vp->v_rdev);
+
 	fp->f_vnode = vp;
 	mutex_exit(&fp->f_tlock);
 	setf(fd, fp);
@@ -93,6 +108,93 @@ lx_epoll_create(int size)
 	return (lx_epoll_create1(0));
 }
 
+
+/* Match values from libc implementation */
+#define	EPOLLIGNORED 	(EPOLLMSG | EPOLLWAKEUP)
+#define	EPOLLSWIZZLED	\
+	(EPOLLRDHUP | EPOLLONESHOT | EPOLLET | EPOLLWRBAND | EPOLLWRNORM)
+
+long
+lx_epoll_ctl(int fd, int op, int pfd, void *event)
+{
+	epoll_event_t epevent;
+	dvpoll_epollfd_t dpevent[2];
+	file_t *fp;
+	iovec_t aiov;
+	uio_t auio;
+	uint32_t events, ev = 0;
+	int error = 0, i = 0;
+
+	dpevent[i].dpep_pollfd.fd = pfd;
+	switch (op) {
+	case EPOLL_CTL_DEL:
+		dpevent[i].dpep_pollfd.events = POLLREMOVE;
+		break;
+
+	case EPOLL_CTL_MOD:
+		/*
+		 * In the modify case, we pass down two events:  one to
+		 * remove the event and another to add it back.
+		 */
+		dpevent[i++].dpep_pollfd.events = POLLREMOVE;
+		dpevent[i].dpep_pollfd.fd = pfd;
+		/* FALLTHROUGH */
+
+	case EPOLL_CTL_ADD:
+		if (copyin(event, &epevent, sizeof (epevent)) != 0)
+			return (set_errno(EFAULT));
+
+		/*
+		 * Mask off the events that we ignore, and then swizzle the
+		 * events for which our values differ from their epoll(7)
+		 * equivalents.
+		 */
+		events = epevent.events;
+		ev = events & ~(EPOLLIGNORED | EPOLLSWIZZLED);
+
+		if (events & EPOLLRDHUP)
+			ev |= POLLRDHUP;
+		if (events & EPOLLET)
+			ev |= POLLET;
+		if (events & EPOLLONESHOT)
+			ev |= POLLONESHOT;
+		if (events & EPOLLWRNORM)
+			ev |= POLLWRNORM;
+		if (events & EPOLLWRBAND)
+			ev |= POLLWRBAND;
+
+		dpevent[i].dpep_data = epevent.data.u64;
+		dpevent[i].dpep_pollfd.events = ev;
+		break;
+
+	default:
+		return (set_errno(EINVAL));
+	}
+
+	if ((fp = getf(fd)) == NULL) {
+		return (set_errno(EBADF));
+	} else if (!lx_epoll_isvalid(fp)) {
+		releasef(fd);
+		return (set_errno(EINVAL));
+	}
+
+	aiov.iov_base = (void *)dpevent;
+	aiov.iov_len = sizeof (dvpoll_epollfd_t) * (i + 1);
+	auio.uio_iov = &aiov;
+	auio.uio_iovcnt = 1;
+	auio.uio_resid = aiov.iov_len;
+	auio.uio_segflg = UIO_SYSSPACE;
+	auio.uio_loffset = 0;
+	auio.uio_fmode = fp->f_flag;
+
+	error = VOP_WRITE(fp->f_vnode, &auio, 1, fp->f_cred, NULL);
+
+	releasef(fd);
+	if (error)
+		return (set_errno(error));
+	return (0);
+}
+
 long
 lx_epoll_wait(int fd, void *events, int maxevents, int timeout)
 {
@@ -105,6 +207,9 @@ lx_epoll_wait(int fd, void *events, int maxevents, int timeout)
 	}
 	if ((fp = getf(fd)) == NULL) {
 		return (set_errno(EBADF));
+	} else if (!lx_epoll_isvalid(fp)) {
+		releasef(fd);
+		return (set_errno(EINVAL));
 	}
 
 	arg.dp_nfds = maxevents;
@@ -133,6 +238,9 @@ lx_epoll_pwait(int fd, void *events, int maxevents, int timeout, void *sigmask)
 	}
 	if ((fp = getf(fd)) == NULL) {
 		return (set_errno(EBADF));
+	} else if (!lx_epoll_isvalid(fp)) {
+		releasef(fd);
+		return (set_errno(EINVAL));
 	}
 
 	arg.dp_nfds = maxevents;