man.bsd.lv manual page server

Manual Page Search Parameters
NAMEI(9) Kernel Developer's Manual NAMEI(9)

namei, NDINIT, NDAT, namei_simple_kernel, namei_simple_user, relookup, lookup_for_nfsd, lookup_for_nfsd_index — pathname lookup

#include <sys/namei.h>
#include <sys/uio.h>
#include <sys/vnode.h>

NDINIT(struct nameidata *ndp, u_long op, u_long flags, struct pathbuf *pathbuf);

NDAT(struct nameidata *ndp, struct vnode *dvp);

int
namei(struct nameidata *ndp);

int
namei_simple_kernel(const char *path, namei_simple_flags_t sflags, struct vnode **ret);

int
namei_simple_user(const char *path, namei_simple_flags_t sflags, struct vnode **ret);

int
relookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, int dummy);

int
lookup_for_nfsd(struct nameidata *ndp, struct vnode *startdir, int neverfollow);

int
lookup_for_nfsd_index(struct nameidata *ndp, struct vnode *startdir);

The namei interface is used to convert pathnames to file system vnodes. The name of the interface is actually a contraction of the words and for name-to-inode conversion, in the days before the vfs(9) interface was implemented.

All access to the namei interface must be in process context. Pathname lookups cannot be done in interrupt context.

In the general form of namei, a caller must:

  1. Allocate storage for a struct nameidata object nd.
  2. Initialize nd with () and optionally () to specify the arguments to a lookup.
  3. Call namei() and handle failure if it returns a nonzero error code.
  4. Read the resulting vnode out of nd.ni_vp. If requested with LOCKPARENT, read the directory vnode out of nd.ni_dvp.
  5. For directory operations, use the struct componentname object stored at nd.ni_cnd.

The other fields of struct nameidata should not be examined or altered directly.

Note that the nfs(4) code misuses struct nameidata and currently has an incestuous relationship with the namei code. This is gradually being cleaned up.

The struct componentname type has the following layout:

struct componentname {
	/*
	 * Arguments to VOP_LOOKUP and directory VOP routines.
	 */
	uint32_t	cn_nameiop;	/* namei operation */
	uint32_t	cn_flags;	/* flags to namei */
	kauth_cred_t 	cn_cred;	/* credentials */
	const char 	*cn_nameptr;	/* pointer to looked up name */
	size_t		cn_namelen;	/* length of looked up comp */
	/*
	 * Side result from VOP_LOOKUP.
	 */
	size_t		cn_consume;	/* chars to consume in lookup */
};

This structure contains the information about a single directory component name, along with certain other information required by vnode operations. See vnodeops(9) for more information about these vnode operations.

The members:

cn_nameiop
The type of operation in progress; indicates the basic operating mode of namei. May be one of LOOKUP, CREATE, DELETE, or RENAME. These modes are described below.
cn_flags
Additional flags affecting the operation of namei. These are described below as well.
cn_cred
The credentials to use for the lookup or other operation the componentname is passed to. This may match the credentials of the current process or it may not, depending on where the original operation request came from and how it has been routed.
cn_nameptr
The name of this directory component, followed by the rest of the path being looked up.
cn_namelen
The length of the name of this directory component. The name is not in general null terminated, although the complete string (the full remaining path) always is.
cn_consume
This field starts at zero; it may be set to a larger value by implementations of VOP_LOOKUP(9) to indicate how many more characters beyond are being consumed. New uses of this feature are discouraged and should be discussed.

Each lookup happens in one of the following modes, specified by callers of namei with () and specified internally by namei to VOP_LOOKUP(9):

  • Callers of namei specify the mode for the last component of a lookup.
  • Internally, namei recursively calls VOP_LOOKUP(9) in LOOKUP mode for each directory component, and then finally calls VOP_LOOKUP(9) in the caller-specified mode for the last component.
Each mode can fail in different ways — for example, LOOKUP mode fails with ENOENT if no entry exists, but CREATE mode succeeds with a NULL vnode.
Yield the vnode for an existing entry. Callers specify LOOKUP for operations on existing vnodes: stat(2), open(2) without O_CREATE, etc.

File systems:

  • MUST refuse if user lacks lookup permission for directory.
  • SHOULD use namecache(9) to cache lookup results.

[ENOENT]
No entry exists.
Yield the vnode for an existing entry; or, if there is none, yield NULL and hint that it will soon be created. Callers specify CREATE for operations that may create directory entries: mkdir(2), open(2) with O_CREATE, etc.

File systems:

  • MUST refuse if user lacks lookup permission for directory.
  • MUST refuse if no entry exists and user lacks write permission for directory.
  • MUST refuse if no entry exists and file system is read-only.
  • SHOULD NOT use namecache(9) to cache negative lookup results.
  • SHOULD save lookup hints internally in the directory for a subsequent operation to create a directory entry.

[EPERM]
The user lacks lookup permission for the directory.
[EPERM]
No entry exists and the user lacks write permission for the directory.
[EROFS]
No entry exists and the file system is read-only.
Yield the vnode of an existing entry, and hint that it will soon be deleted. Callers specify DELETE for operations that delete directory entries: unlink(2), rmdir(2), etc.

File systems:

  • MUST refuse if user lacks lookup permission for directory.
  • MUST refuse if entry exists and user lacks write permission for directory.
  • MUST refuse if entry exists and file system is read-only.
  • SHOULD NOT use namecache(9) to cache lookup results.
  • SHOULD save lookup hints internally in the directory for a subsequent operation to delete a directory entry.

[ENOENT]
No entry exists.
[EPERM]
The user lacks lookup permission for the directory.
[EPERM]
An entry exists and the user lacks write permission for the directory.
[EROFS]
An entry exists and the file system is read-only.
Yield the vnode of an existing entry, and hint that it will soon be overwritten; or, if there is none, yield NULL, and hint that it will soon be created.

Callers specify RENAME for an entry that is about to be created or overwritten, namely for the target of rename(2).

File systems:

  • MUST refuse if user lacks lookup permission for directory.
  • MUST refuse if user lacks write permission for directory.
  • MUST refuse if file system is read-only.
  • SHOULD NOT use namecache(9) to cache lookup results.
  • SHOULD save lookup hints internally in the directory for a subsequent operation to create or overwrite a directory entry.

[EPERM]
The user lacks lookup permission for the directory.
[EPERM]
The user lacks write permission for the directory.
[EROFS]
The file system is read-only.

If a caller decides not to perform an operation it hinted at by a destructive operating mode (CREATE, DELETE, or RENAME), it SHOULD call VOP_ABORTOP(9) to release the hints. If a file system fails to perform such an operation, it SHOULD call VOP_ABORTOP(9) to release the hints. However, the current code is inconsistent about this, and every implementation of VOP_ABORTOP(9) does nothing.

The following flags may be specified by of namei, and MUST NOT be used by file systems:

Follow symbolic links in the last path component. Used by operations that do not address symbolic links directly, such as stat(2). (Does not affect symbolic links found in the middle of a path.)
Do not follow symbolic links in the last path component. Used by operations that address symbolic links directly, such as lstat(2).

Note: The value of NOFOLLOW is 0. We define the constant to let callers say either FOLLOW or NOFOLLOW explicitly.

On successful lookup, lock the vnode, if any, in ndp->ni_vp. Without this flag, it would be unlocked.
On successful lookup, lock and return the directory vnode in ndp->ni_dvp. Without this flag, it is not returned at all.
If set, the path is looked up in the emulation root of the current process first. If that fails, the system root is used.
Indicates that the caller has set ndp->ni_erootdir prior to calling namei. This is only useful or permitted when the emulation in the current process is partway through being set up.
Bypass normal chroot(8) handling for absolute paths.
Do not cross mount points.
Enforce read-only behavior.
Accept slashes after a component name that does not exist. This only makes sense in CREATE mode and when creating a directory.
Do not cache the lookup result for the last component name. This is used only with the RENAME mode for the target; the cache entry would be invalidated immediately.

The following flag may be set by a caller of namei and tested by a file system in VOP_LOOKUP(9) or other subsequent directory operations:

Allow whiteouts to be seen as objects instead of functioning as “nothing there”.

The following flags are set by namei for calling VOP_LOOKUP(9):

The current pathname component is “..”. May be tested by subsequent directory operations too.
The current pathname component is the last component found in the pathname. Guaranteed to remain set in subsequent directory operations.
The current object to be looked up must be a directory. May not be used by subsequent directory operations.
The lookup result for the current pathname component should be added to the namecache(9). May be used to make additional caching decisions, e.g. to store an mtime for determining whether our cache for a remote vnode is stale. May not be used by subsequent directory operatoins.

A file system may set the following flag on return from VOP_LOOKUP(9) for use by namei, namecache(9), and subsequent directory operations:

The object at the current pathname component is a whiteout.

The following additional historic flags have been removed from NetBSD and should be handled as follows if porting code from elsewhere:

Part of a misbegotten and incorrect locking scheme. Any file-system-level code using this is presumptively incorrect. File systems should use the genfs_rename(9) interface to handle locking in ().
Used at one point for signaling to puffs(3) to work around a protocol deficiency that was later rectified.
Useless internal state.
Unclean setting affect vnode reference counting. Now effectively never in effect. Any code referring to this is suspect.
Unclean setting relating to responsibility for freeing pathname buffers in the days before the pathbuf structure. Now effectively always in effect; the caller of namei owns the pathbuf structure and is always responsible for destroying it.
Related to SAVENAME. Any uses can be replaced with “true”.

NDINIT(ndp, op, flags, pathbuf)
Initialise a nameidata structure pointed to by ndp for use by the namei interface. The operating mode and flags (as documented above) are specified by op and flags respectively. The pathname is passed as a pathbuf structure, which should be initialized using one of the pathbuf(9) operations. Destroying the pathbuf is the responsibility of the caller; this must not be done until the caller is finished with all of the namei results and all of the nameidata contents except for the result vnode.

This routine stores the credentials of the calling thread (curlwp) in ndp. () sets the credentials using kauth_cred_get(9). In the rare case that another set of credentials is required for the namei operation, must be set manually after NDINIT().

NDAT(ndp, dvp)
This macro is used after NDINIT() to set the starting directory. This supersedes the current process's current working directory as the initial point of departure for looking up relative paths. This mechanism is used by openat(2) and related calls.
namei(ndp)
Convert a pathname into a pointer to a vnode. The nameidata structure pointed to by ndp should be initialized with the NDINIT() macro, and perhaps also the NDAT() macro. Direct initialization of members of struct nameidata is supported and may (will) break silently in the future.

The vnode for the pathname is returned in . The parent directory is returned locked in iff LOCKPARENT is specified.

Any or all of the flags documented above as set by the caller can be enabled by passing them (OR'd together) as the flags argument of (). As discussed above every such call should explicitly contain either FOLLOW or NOFOLLOW to control the behavior regarding final symbolic links.

(path, sflags, ret)
Look up the path path and translate it to a vnode, returned in ret. The path argument must be a kernel (UIO_SYSSPACE) pointer. The sflags argument chooses the precise behavior. It may be set to one of the following symbols:
 
 
 
 
These select (or not) the FOLLOW/NOFOLLOW and TRYEMULROOT flags. Other flags are not available through this interface, which is nonetheless sufficient for more than half the namei() usage in the kernel. Note that the encoding of sflags has deliberately been arranged to be type-incompatible with anything else. This prevents various possible accidents while the namei() interface is being rototilled.
(path, sflags, ret)
This function is the same as namei_simple_kernel() except that the path argument shall be a user pointer (UIO_USERSPACE) rather than a kernel pointer.
(dvp, vpp, cnp, dummy)
Reacquire a path name component is a directory. This is a quicker way to lookup a pathname component when the parent directory is known. The locked parent directory vnode is specified by dvp and the pathname component by cnp. The vnode of the pathname is returned in the address specified by vpp. The dummy argument is unused. Note that one may only use relookup() to repeat a lookup of a final path component previously done by namei, and one must use the same componentname structure that call produced. Otherwise the behavior is undefined and likely adverse.
(ndp, startdir, neverfollow)
This is a private entry point into namei used by the NFS server code. It looks up a path starting from startdir. If neverfollow is set, symbolic link (not just at the end of the path) will cause an error. Otherwise, it follows symlinks normally. It should not be used by new code.
(ndp, startdir)
This is a (second) private entry point into namei used by the NFS server code. It looks up a single path component starting from startdir. It should not be used by new code.

The nameidata structure has the following layout:

struct nameidata {
	/*
	 * Arguments to namei.
	 */
	struct vnode *ni_atdir;		/* startup dir, cwd if null */
	struct pathbuf *ni_pathbuf;	/* pathname container */
	char *ni_pnbuf;			/* extra pathname buffer ref (XXX) */
	/*
	 * Internal starting state. (But see notes.)
	 */
	struct	vnode *ni_rootdir;	/* logical root directory */
	struct	vnode *ni_erootdir;	/* emulation root directory */
	/*
	 * Results from namei.
	 */
	struct	vnode *ni_vp;		/* vnode of result */
	struct	vnode *ni_dvp;		/* vnode of intermediate directory */
	/*
	 * Internal current state.
	 */
	size_t		ni_pathlen;	/* remaining chars in path */
	const char	*ni_next;	/* next location in pathname */
	unsigned int	ni_loopcnt;	/* count of symlinks encountered */
	/*
	 * Lookup parameters: this structure describes the subset of
	 * information from the nameidata structure that is passed
	 * through the VOP interface.
	 */
	struct componentname ni_cnd;
};

These fields are:

ni_atdir
The directory to use for the starting point of relative paths. If null, the current process's current directory is used. This is initialized to NULL by () and set by ().
ni_pathbuf
The abstract path buffer in use, passed as an argument to NDINIT(). The name pointers that appear elsewhere, such as in the componentname structure, point into this buffer. It is owned by the caller and must not be destroyed until all namei operations are complete. See pathbuf(9).
ni_pnbuf
This is the name pointer used during namei. It points into ni_pathbuf. It is not initialized until entry into namei.
ni_rootdir
The root directory to use as the starting point for absolute paths. This is retrieved from the current process's current root directory when namei starts up. It is not initialized by NDINIT().
ni_erootdir
The root directory to use as the emulation root, for processes running in emulation. This is retrieved from the current process's emulation root directory when namei starts up and not initialized by NDINIT(). As described elsewhere, it may be set by the caller if the EMULROOTSET flag is used, but this should only be done when the current process's emulation root directory is not yet initialized. (And ideally in the future things would be tidied so that this is not necessary.)
ni_vp
 
ni_dvp
Returned vnodes, as described above. These only contain valid values if namei returns successfully.
ni_pathlen
The length of the full current remaining path string in ni_pnbuf. This is not initialized by NDINIT() and is used only internally.
ni_next
The remaining part of the path, after the current component found in the componentname structure. This is not initialized by NDINIT() and is used only internally.
ni_loopcnt
The number of symbolic links encountered (and traversed) so far. If this exceeds a limit, namei fails with ELOOP. This is not initialized by NDINIT() and is used only internally.
ni_cnd
The componentname structure holding the current directory component, and also the mode, flags, and credentials. The mode, flags, and credentials are initialized by NDINIT(); the rest is not initialized until namei runs.

There is also a namei_state structure that is hidden within vfs_lookup.c. This contains the following additional state:

docache
A flag indicating whether to cache the last pathname component.
rdonly
The read-only state, initialized from the RDONLY flag.
slashes
The number of trailing slashes found after the current pathname component.
attempt_retry
Set on some error cases (and not others) to indicate that a failure in the emulation root should be followed by a retry in the real system root.

The state in namei_state is genuinely private to namei. Note that much of the state in nameidata should also be private, but is currently not because it is misused in some fashion by outside code, usually nfs(4).

The control flow within the namei portions of vfs_lookup.c is as follows.

()
does a complete path lookup by calling (), (), and ().
namei_init()
sets up the basic internal state and makes some (precondition-type) assertions.
namei_cleanup()
makes some postcondition-type assertions; it currently does nothing besides this.
namei_tryemulroot()
handles TRYEMULROOT by calling () once or twice as needed, and attends to making sure the original pathname is preserved for the second try.
namei_oneroot()
does a complete path search from a single root directory. It begins with (), then calls () (and if necessary, ()) repeatedly until done. It also handles returning the result vnode(s) in the requested state.
namei_start()
sets up the initial state and locking; it calls ().
namei_getstartdir()
initializes the root directory state (both ni_rootdir and ni_erootdir) and picks the starting directory, consuming the leading slashes of an absolute path and handling the magic “/../” string for bypassing the emulation root. A different version () is used for lookups coming from nfsd(8) as those are required to have different semantics.
lookup_once()
calls VOP_LOOKUP() for one path component, also handling any needed crossing of mount points (either up or down) and coping with locking requirements.
()
is called prior to each lookup_once() call to examine the pathname and find where the next component starts.
namei_follow()
reads the contents of a symbolic link and updates both the path buffer and the search directory accordingly.

As a final note be advised that the magic return value associated with CREATE mode is different for namei than it is for (). The latter “fails” with EJUSTRETURN. namei translates this into succeeding and returning a null vnode.

The name lookup subsystem is implemented within the file sys/kern/vfs_lookup.c.

intro(9), namecache(9), vfs(9), vnode(9), vnodeops(9)

There should be no such thing as operating modes. Only LOOKUP is actually needed. The behavior where removing an object looks it up within namei and then calls into the file system (which must look it up again internally or cache state from VOP_LOOKUP()) is particularly contorted.

Most of the flags are equally bogus.

Most of the contents of the nameidata structure should be private and hidden within namei; currently it cannot be because of abuse elsewhere.

The EMULROOTSET flag is messy.

There is no good way to support file systems that want to use a more elaborate pathname schema than the customary slash-delimited components.

May 5, 2019 NetBSD-9.2