Implement the openat() linux syscall

Submitted by:	Roman Divacky (rdivacky@)
MFC after:	2 weeks

sys/amd64/linux32/linux.h

@@ -531,6 +531,7 @@ int	linux_ioctl_unregister_handler(struct linux_ioctl_handler *h);
 #define	LINUX_O_RDONLY		00000000
 #define	LINUX_O_WRONLY		00000001
 #define	LINUX_O_RDWR		00000002
+#define	LINUX_O_ACCMODE		00000003
 #define	LINUX_O_CREAT		00000100
 #define	LINUX_O_EXCL		00000200
 #define	LINUX_O_NOCTTY		00000400
@@ -565,6 +566,8 @@ int	linux_ioctl_unregister_handler(struct linux_ioctl_handler *h);
 #define	LINUX_F_WRLCK		1
 #define	LINUX_F_UNLCK		2
 
+#define	LINUX_AT_FDCWD		-100
+
 /*
  * mount flags
  */

sys/amd64/linux32/linux32_dummy.c

@@ -97,7 +97,6 @@ DUMMY(inotify_init);
 DUMMY(inotify_add_watch);
 DUMMY(inotify_rm_watch);
 DUMMY(migrate_pages);
-DUMMY(openat);
 DUMMY(mkdirat);
 DUMMY(mknodat);
 DUMMY(fchownat);

sys/amd64/linux32/linux32_proto.h

@@ -871,7 +871,10 @@ struct linux_migrate_pages_args {
 	register_t dummy;
 };
 struct linux_openat_args {
-	register_t dummy;
+	char dfd_l_[PADL_(l_int)]; l_int dfd; char dfd_r_[PADR_(l_int)];
+	char filename_l_[PADL_(char *)]; char * filename; char filename_r_[PADR_(char *)];
+	char flags_l_[PADL_(l_int)]; l_int flags; char flags_r_[PADR_(l_int)];
+	char mode_l_[PADL_(l_int)]; l_int mode; char mode_r_[PADR_(l_int)];
 };
 struct linux_mkdirat_args {
 	register_t dummy;
@@ -1381,7 +1384,7 @@ int	linux_unshare(struct thread *, struct linux_unshare_args *);
 #define	LINUX_SYS_AUE_linux_inotify_add_watch	AUE_NULL
 #define	LINUX_SYS_AUE_linux_inotify_rm_watch	AUE_NULL
 #define	LINUX_SYS_AUE_linux_migrate_pages	AUE_NULL
-#define	LINUX_SYS_AUE_linux_openat	AUE_NULL
+#define	LINUX_SYS_AUE_linux_openat	AUE_OPEN_RWTC
 #define	LINUX_SYS_AUE_linux_mkdirat	AUE_NULL
 #define	LINUX_SYS_AUE_linux_mknodat	AUE_NULL
 #define	LINUX_SYS_AUE_linux_fchownat	AUE_NULL

sys/amd64/linux32/linux32_sysent.c

@@ -315,7 +315,7 @@ struct sysent linux_sysent[] = {
 	{ 0, (sy_call_t *)linux_inotify_add_watch, AUE_NULL, NULL, 0, 0 },	/* 292 = linux_inotify_add_watch */
 	{ 0, (sy_call_t *)linux_inotify_rm_watch, AUE_NULL, NULL, 0, 0 },	/* 293 = linux_inotify_rm_watch */
 	{ 0, (sy_call_t *)linux_migrate_pages, AUE_NULL, NULL, 0, 0 },	/* 294 = linux_migrate_pages */
-	{ 0, (sy_call_t *)linux_openat, AUE_NULL, NULL, 0, 0 },	/* 295 = linux_openat */
+	{ AS(linux_openat_args), (sy_call_t *)linux_openat, AUE_OPEN_RWTC, NULL, 0, 0 },	/* 295 = linux_openat */
 	{ 0, (sy_call_t *)linux_mkdirat, AUE_NULL, NULL, 0, 0 },	/* 296 = linux_mkdirat */
 	{ 0, (sy_call_t *)linux_mknodat, AUE_NULL, NULL, 0, 0 },	/* 297 = linux_mknodat */
 	{ 0, (sy_call_t *)linux_fchownat, AUE_NULL, NULL, 0, 0 },	/* 298 = linux_fchownat */

sys/amd64/linux32/syscalls.master

@@ -463,7 +463,8 @@
 292	AUE_NULL	STD	{ int linux_inotify_add_watch(void); }
 293	AUE_NULL	STD	{ int linux_inotify_rm_watch(void); }
 294	AUE_NULL	STD	{ int linux_migrate_pages(void); }
-295	AUE_NULL	STD	{ int linux_openat(void); }
+295	AUE_OPEN_RWTC	STD	{ int linux_openat(l_int dfd, char *filename, \
+					l_int flags, l_int mode); }
 296	AUE_NULL	STD	{ int linux_mkdirat(void); }
 297	AUE_NULL	STD	{ int linux_mknodat(void); }
 298	AUE_NULL	STD	{ int linux_fchownat(void); }

sys/compat/linux/linux_file.c

@@ -43,6 +43,7 @@ __FBSDID("$FreeBSD$");
 #include <sys/malloc.h>
 #include <sys/mount.h>
 #include <sys/mutex.h>
+#include <sys/namei.h>
 #include <sys/proc.h>
 #include <sys/stat.h>
 #include <sys/sx.h>
@@ -85,57 +86,51 @@ linux_creat(struct thread *td, struct linux_creat_args *args)
     return (error);
 }
 
-int
+
-linux_open(struct thread *td, struct linux_open_args *args)
+static int
+linux_common_open(struct thread *td, char *path, int l_flags, int mode, int openat)
 {
     struct proc *p = td->td_proc;
     struct file *fp;
     int fd;
-    char *path;
     int bsd_flags, error;
 
-    if (args->flags & LINUX_O_CREAT)
-	LCONVPATHCREAT(td, args->path, &path);
-    else
-	LCONVPATHEXIST(td, args->path, &path);
-
-#ifdef DEBUG
-	if (ldebug(open))
-		printf(ARGS(open, "%s, 0x%x, 0x%x"),
-		    path, args->flags, args->mode);
-#endif
     bsd_flags = 0;
-    if (args->flags & LINUX_O_RDONLY)
+    switch (l_flags & LINUX_O_ACCMODE) {
-	bsd_flags |= O_RDONLY;
+    case LINUX_O_WRONLY:
-    if (args->flags & LINUX_O_WRONLY)
 	bsd_flags |= O_WRONLY;
-    if (args->flags & LINUX_O_RDWR)
+	break;
+    case LINUX_O_RDWR:
 	bsd_flags |= O_RDWR;
-    if (args->flags & LINUX_O_NDELAY)
+	break;
+    default:
+	bsd_flags |= O_RDONLY;
+    }
+    if (l_flags & LINUX_O_NDELAY)
 	bsd_flags |= O_NONBLOCK;
-    if (args->flags & LINUX_O_APPEND)
+    if (l_flags & LINUX_O_APPEND)
 	bsd_flags |= O_APPEND;
-    if (args->flags & LINUX_O_SYNC)
+    if (l_flags & LINUX_O_SYNC)
 	bsd_flags |= O_FSYNC;
-    if (args->flags & LINUX_O_NONBLOCK)
+    if (l_flags & LINUX_O_NONBLOCK)
 	bsd_flags |= O_NONBLOCK;
-    if (args->flags & LINUX_FASYNC)
+    if (l_flags & LINUX_FASYNC)
 	bsd_flags |= O_ASYNC;
-    if (args->flags & LINUX_O_CREAT)
+    if (l_flags & LINUX_O_CREAT)
 	bsd_flags |= O_CREAT;
-    if (args->flags & LINUX_O_TRUNC)
+    if (l_flags & LINUX_O_TRUNC)
 	bsd_flags |= O_TRUNC;
-    if (args->flags & LINUX_O_EXCL)
+    if (l_flags & LINUX_O_EXCL)
 	bsd_flags |= O_EXCL;
-    if (args->flags & LINUX_O_NOCTTY)
+    if (l_flags & LINUX_O_NOCTTY)
 	bsd_flags |= O_NOCTTY;
-    if (args->flags & LINUX_O_DIRECT)
+    if (l_flags & LINUX_O_DIRECT)
 	bsd_flags |= O_DIRECT;
-    if (args->flags & LINUX_O_NOFOLLOW)
+    if (l_flags & LINUX_O_NOFOLLOW)
 	bsd_flags |= O_NOFOLLOW;
     /* XXX LINUX_O_NOATIME: unable to be easily implemented. */
 
-    error = kern_open(td, path, UIO_SYSSPACE, bsd_flags, args->mode);
+    error = kern_open(td, path, UIO_SYSSPACE, bsd_flags, mode);
     if (!error) {
 	    fd = td->td_retval[0];
 	    /*
@@ -158,7 +153,7 @@ linux_open(struct thread *td, struct linux_open_args *args)
 			    PROC_UNLOCK(p);
 			    sx_sunlock(&proctree_lock);
 		    }
-		    if (args->flags & LINUX_O_DIRECTORY) {
+		    if (l_flags & LINUX_O_DIRECTORY) {
 			    if (fp->f_type != DTYPE_VNODE ||
 				fp->f_vnode->v_type != VDIR) {
 				    error = ENOTDIR;
@@ -177,10 +172,121 @@ linux_open(struct thread *td, struct linux_open_args *args)
     if (ldebug(open))
 	    printf(LMSG("open returns error %d"), error);
 #endif
-    LFREEPATH(path);
+    if (!openat)
+	LFREEPATH(path);
     return error;
 }
 
+/*
+ * common code for linux *at set of syscalls
+ *
+ * works like this:
+ * if filename is absolute 
+ *    ignore dirfd
+ * else
+ *    if dirfd == AT_FDCWD 
+ *       return CWD/filename
+ *    else
+ *       return DIRFD/filename
+ */
+static int
+linux_at(struct thread *td, int dirfd, char *filename, char **newpath, char **freebuf)
+{
+   	struct file *fp;
+	int error = 0;
+	struct vnode *dvp;
+	struct filedesc *fdp = td->td_proc->p_fd;
+	char *fullpath = "unknown";
+	char *freepath = NULL;
+
+	/* don't do anything if the pathname is absolute */
+	if (*filename == '/') {
+	   	*newpath= filename;
+	   	return (0);
+	}
+
+	/* check for AT_FDWCD */
+	if (dirfd == LINUX_AT_FDCWD) {
+	   	FILEDESC_LOCK(fdp);
+		dvp = fdp->fd_cdir;
+	   	FILEDESC_UNLOCK(fdp);
+	} else {
+	   	error = fget(td, dirfd, &fp);
+		if (error)
+		   	return (error);
+		dvp = fp->f_vnode;
+		/* only a dir can be dfd */
+		if (dvp->v_type != VDIR) {
+		   	fdrop(fp, td);
+			return (ENOTDIR);
+		}
+		fdrop(fp, td);
+	}
+
+	error = vn_fullpath(td, dvp, &fullpath, &freepath);
+	if (!error) {
+	   	*newpath = malloc(strlen(fullpath) + strlen(filename) + 2, M_TEMP, M_WAITOK | M_ZERO);
+		*freebuf = freepath;
+		sprintf(*newpath, "%s/%s", fullpath, filename);
+	}
+
+	return (error);
+}
+
+int
+linux_openat(struct thread *td, struct linux_openat_args *args)
+{
+   	char *newpath, *oldpath, *freebuf = NULL, *path;
+	int error;
+
+	oldpath = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
+	error = copyinstr(args->filename, oldpath, MAXPATHLEN, NULL);
+
+#ifdef DEBUG
+	if (ldebug(openat))
+		printf(ARGS(openat, "%i, %s, 0x%x, 0x%x"), args->dfd,
+		    oldpath, args->flags, args->mode);
+#endif
+
+	error = linux_at(td, args->dfd, oldpath, &newpath, &freebuf);
+	if (error)
+	   	return (error);
+#ifdef DEBUG
+	printf(LMSG("newpath: %s"), newpath);
+#endif
+    	if (args->flags & LINUX_O_CREAT)
+		LCONVPATH_SEG(td, newpath, &path, 1, UIO_SYSSPACE);
+    	else
+		LCONVPATH_SEG(td, newpath, &path, 0, UIO_SYSSPACE);
+	if (freebuf)
+	   	free(freebuf, M_TEMP);
+	if (*oldpath != '/')
+   	   	free(newpath, M_TEMP);
+
+	error = linux_common_open(td, path, args->flags, args->mode, 1);
+	free(oldpath, M_TEMP);
+	return (error);
+}
+
+int
+linux_open(struct thread *td, struct linux_open_args *args)
+{
+    char *path;
+
+    if (args->flags & LINUX_O_CREAT)
+	LCONVPATHCREAT(td, args->path, &path);
+    else
+	LCONVPATHEXIST(td, args->path, &path);
+
+#ifdef DEBUG
+	if (ldebug(open))
+		printf(ARGS(open, "%s, 0x%x, 0x%x"),
+		    path, args->flags, args->mode);
+#endif
+
+    return linux_common_open(td, path, args->flags, args->mode, 0);
+}
+
 int
 linux_lseek(struct thread *td, struct linux_lseek_args *args)
 {

sys/compat/linux/linux_util.h

@@ -53,16 +53,19 @@ extern const char linux_emul_path[];
 
 int linux_emul_convpath(struct thread *, char *, enum uio_seg, char **, int);
 
-#define LCONVPATH(td, upath, pathp, i) 					\
+#define LCONVPATH_SEG(td, upath, pathp, i, seg)				\
 	do {								\
 		int _error;						\
 									\
-		_error = linux_emul_convpath(td, upath, UIO_USERSPACE,  \
+		_error = linux_emul_convpath(td, upath, seg,		\
 		    pathp, i);						\
 		if (*(pathp) == NULL)					\
 			return (_error);				\
 	} while (0)
 
+#define LCONVPATH(td, upath, pathp, i) 	\
+   LCONVPATH_SEG(td, upath, pathp, i, UIO_USERSPACE)
+
 #define LCONVPATHEXIST(td, upath, pathp) LCONVPATH(td, upath, pathp, 0)
 #define LCONVPATHCREAT(td, upath, pathp) LCONVPATH(td, upath, pathp, 1)
 #define LFREEPATH(path)	free(path, M_TEMP)

sys/i386/linux/linux.h

@@ -502,6 +502,7 @@ int	linux_ioctl_unregister_handler(struct linux_ioctl_handler *h);
 #define	LINUX_O_RDONLY		00000000
 #define	LINUX_O_WRONLY		00000001
 #define	LINUX_O_RDWR		00000002
+#define	LINUX_O_ACCMODE		00000003
 #define	LINUX_O_CREAT		00000100
 #define	LINUX_O_EXCL		00000200
 #define	LINUX_O_NOCTTY		00000400
@@ -536,6 +537,8 @@ int	linux_ioctl_unregister_handler(struct linux_ioctl_handler *h);
 #define	LINUX_F_WRLCK		1
 #define	LINUX_F_UNLCK		2
 
+#define	LINUX_AT_FDCWD		-100
+
 /*
  * mount flags
  */

sys/i386/linux/linux_dummy.c

@@ -87,7 +87,6 @@ DUMMY(inotify_init);
 DUMMY(inotify_add_watch);
 DUMMY(inotify_rm_watch);
 DUMMY(migrate_pages);
-DUMMY(openat);
 DUMMY(mkdirat);
 DUMMY(mknodat);
 DUMMY(fchownat);

sys/i386/linux/linux_proto.h

@@ -893,7 +893,10 @@ struct linux_migrate_pages_args {
 	register_t dummy;
 };
 struct linux_openat_args {
-	register_t dummy;
+	char dfd_l_[PADL_(l_int)]; l_int dfd; char dfd_r_[PADR_(l_int)];
+	char filename_l_[PADL_(char *)]; char * filename; char filename_r_[PADR_(char *)];
+	char flags_l_[PADL_(l_int)]; l_int flags; char flags_r_[PADR_(l_int)];
+	char mode_l_[PADL_(l_int)]; l_int mode; char mode_r_[PADR_(l_int)];
 };
 struct linux_mkdirat_args {
 	register_t dummy;
@@ -1407,7 +1410,7 @@ int	linux_unshare(struct thread *, struct linux_unshare_args *);
 #define	LINUX_SYS_AUE_linux_inotify_add_watch	AUE_NULL
 #define	LINUX_SYS_AUE_linux_inotify_rm_watch	AUE_NULL
 #define	LINUX_SYS_AUE_linux_migrate_pages	AUE_NULL
-#define	LINUX_SYS_AUE_linux_openat	AUE_NULL
+#define	LINUX_SYS_AUE_linux_openat	AUE_OPEN_RWTC
 #define	LINUX_SYS_AUE_linux_mkdirat	AUE_NULL
 #define	LINUX_SYS_AUE_linux_mknodat	AUE_NULL
 #define	LINUX_SYS_AUE_linux_fchownat	AUE_NULL

sys/i386/linux/linux_sysent.c

@@ -314,7 +314,7 @@ struct sysent linux_sysent[] = {
 	{ 0, (sy_call_t *)linux_inotify_add_watch, AUE_NULL, NULL, 0, 0 },	/* 292 = linux_inotify_add_watch */
 	{ 0, (sy_call_t *)linux_inotify_rm_watch, AUE_NULL, NULL, 0, 0 },	/* 293 = linux_inotify_rm_watch */
 	{ 0, (sy_call_t *)linux_migrate_pages, AUE_NULL, NULL, 0, 0 },	/* 294 = linux_migrate_pages */
-	{ 0, (sy_call_t *)linux_openat, AUE_NULL, NULL, 0, 0 },	/* 295 = linux_openat */
+	{ AS(linux_openat_args), (sy_call_t *)linux_openat, AUE_OPEN_RWTC, NULL, 0, 0 },	/* 295 = linux_openat */
 	{ 0, (sy_call_t *)linux_mkdirat, AUE_NULL, NULL, 0, 0 },	/* 296 = linux_mkdirat */
 	{ 0, (sy_call_t *)linux_mknodat, AUE_NULL, NULL, 0, 0 },	/* 297 = linux_mknodat */
 	{ 0, (sy_call_t *)linux_fchownat, AUE_NULL, NULL, 0, 0 },	/* 298 = linux_fchownat */

sys/i386/linux/syscalls.master

@@ -473,7 +473,8 @@
 292	AUE_NULL	STD	{ int linux_inotify_add_watch(void); }
 293	AUE_NULL	STD	{ int linux_inotify_rm_watch(void); }
 294	AUE_NULL	STD	{ int linux_migrate_pages(void); }
-295	AUE_NULL	STD	{ int linux_openat(void); }
+295	AUE_OPEN_RWTC	STD	{ int linux_openat(l_int dfd, char *filename, \
+					l_int flags, l_int mode); }
 296	AUE_NULL	STD	{ int linux_mkdirat(void); }
 297	AUE_NULL	STD	{ int linux_mknodat(void); }
 298	AUE_NULL	STD	{ int linux_fchownat(void); }