This write-up is based on three articles published at lwn.net:
- <https://lwn.net/Articles/649115/> Pathname lookup in Linux
- <https://lwn.net/Articles/649729/> RCU-walk: faster pathname lookup in Linux
- <https://lwn.net/Articles/650786/> A walk among the symlinks
Written by Neil Brown with help from Al Viro and Jon Corbet. It has subsequently been updated to reflect changes in the kernel including:
- per-directory parallel name lookup.
openat2()resolution restriction flags.
Introduction to pathname lookup¶
The most obvious aspect of pathname lookup, which very little exploration is needed to discover, is that it is complex. There are many rules, special cases, and implementation alternatives that all combine to confuse the unwary reader. Computer science has long been acquainted with such complexity and has tools to help manage it. One tool that we will make extensive use of is “divide and conquer”. For the early parts of the analysis we will divide off symlinks - leaving them until the final part. Well before we get to symlinks we have another major division based on the VFS’s approach to locking which will allow us to review “REF-walk” and “RCU-walk” separately. But we are getting ahead of ourselves. There are some important low level distinctions we need to clarify first.
There are two sorts of …¶
Pathnames (sometimes “file names”), used to identify objects in the
filesystem, will be familiar to most readers. They contain two sorts
of elements: “slashes” that are sequences of one or more “
characters, and “components” that are sequences of one or more
/” characters. These form two kinds of paths. Those that
start with slashes are “absolute” and start from the filesystem root.
The others are “relative” and start from the current directory, or
from some other location specified by a file descriptor given to
*at()” system calls such as openat().
It is tempting to describe the second kind as starting with a
component, but that isn’t always accurate: a pathname can lack both
slashes and components, it can be empty, in other words. This is
generally forbidden in POSIX, but some of those “
*at()” system calls
in Linux permit it when the
AT_EMPTY_PATH flag is given. For
example, if you have an open file descriptor on an executable file you
can execute it by calling execveat() passing
the file descriptor, an empty path, and the
These paths can be divided into two sections: the final component and
everything else. The “everything else” is the easy bit. In all cases
it must identify a directory that already exists, otherwise an error
ENOTDIR will be reported.
The final component is not so simple. Not only do different system
calls interpret it quite differently (e.g. some create it, some do
not), but it might not even exist: neither the empty pathname nor the
pathname that is just slashes have a final component. If it does
exist, it could be “
.” or “
..” which are handled quite differently
from other components.
If a pathname ends with a slash, such as “
/tmp/foo/” it might be
tempting to consider that to have an empty final component. In many
ways that would lead to correct results, but not always. In
rmdir() each create or remove a directory named
by the final component, and they are required to work with pathnames
ending in “
/”. According to POSIX:
A pathname that contains at least one non-<slash> character and that ends with one or more trailing <slash> characters shall not be resolved successfully unless the last pathname component before the trailing <slash> characters names an existing directory or a directory entry that is to be created for a directory immediately after the pathname is resolved.
The Linux pathname walking code (mostly in
fs/namei.c) deals with
all of these issues: breaking the path into components, handling the
“everything else” quite separately from the final component, and
checking that the trailing slash is not used where it isn’t
permitted. It also addresses the important issue of concurrent
While one process is looking up a pathname, another might be making changes that affect that lookup. One fairly extreme case is that if “a/b” were renamed to “a/c/b” while another process were looking up “a/b/..”, that process might successfully resolve on “a/c”. Most races are much more subtle, and a big part of the task of pathname lookup is to prevent them from having damaging effects. Many of the possible races are seen most clearly in the context of the “dcache” and an understanding of that is central to understanding pathname lookup.
More than just a cache¶
The “dcache” caches information about names in each filesystem to
make them quickly available for lookup. Each entry (known as a
“dentry”) contains three significant fields: a component name, a
pointer to a parent dentry, and a pointer to the “inode” which
contains further information about the object in that parent with
the given name. The inode pointer can be
NULL indicating that the
name doesn’t exist in the parent. While there can be linkage in the
dentry of a directory to the dentries of the children, that linkage is
not used for pathname lookup, and so will not be considered here.
The dcache has a number of uses apart from accelerating lookup. One that will be particularly relevant is that it is closely integrated with the mount table that records which filesystem is mounted where. What the mount table actually stores is which dentry is mounted on top of which other dentry.
When considering the dcache, we have another of our “two types” distinctions: there are two types of filesystems.
Some filesystems ensure that the information in the dcache is always completely accurate (though not necessarily complete). This can allow the VFS to determine if a particular file does or doesn’t exist without checking with the filesystem, and means that the VFS can protect the filesystem against certain races and other problems. These are typically “local” filesystems such as ext3, XFS, and Btrfs.
Other filesystems don’t provide that guarantee because they cannot.
These are typically filesystems that are shared across a network,
whether remote filesystems like NFS and 9P, or cluster filesystems
like ocfs2 or cephfs. These filesystems allow the VFS to revalidate
cached information, and must provide their own protection against
awkward races. The VFS can detect these filesystems by the
DCACHE_OP_REVALIDATE flag being set in the dentry.
REF-walk: simple concurrency management with refcounts and spinlocks¶
With all of those divisions carefully classified, we can now start
looking at the actual process of walking along a path. In particular
we will start with the handling of the “everything else” part of a
pathname, and focus on the “REF-walk” approach to concurrency
management. This code is found in the
link_path_walk() function, if
you ignore all the places that only run when “
(indicating the use of RCU-walk) is set.
REF-walk is fairly heavy-handed with locks and reference counts. Not as heavy-handed as in the old “big kernel lock” days, but certainly not afraid of taking a lock when one is needed. It uses a variety of different concurrency controls. A background understanding of the various primitives is assumed, or can be gleaned from elsewhere such as in Meet the Lockers.
The locking mechanisms used by REF-walk include:
This uses the lockref primitive to provide both a spinlock and a reference count. The special-sauce of this primitive is that the conceptual sequence “lock; inc_ref; unlock;” can often be performed with a single atomic memory operation.
Holding a reference on a dentry ensures that the dentry won’t suddenly
be freed and used for something else, so the values in various fields
will behave as expected. It also protects the
to the inode to some extent.
The association between a dentry and its inode is fairly permanent.
For example, when a file is renamed, the dentry and inode move
together to the new location. When a file is created the dentry will
initially be negative (i.e.
NULL), and will be assigned
to the new inode as part of the act of creation.
When a file is deleted, this can be reflected in the cache either by
NULL, or by removing it from the hash table
(described shortly) used to look up the name in the parent directory.
If the dentry is still in use the second option is used as it is
perfectly legal to keep using an open file after it has been deleted
and having the dentry around helps. If the dentry is not otherwise in
use (i.e. if the refcount in
d_lockref is one), only then will
d_inode be set to
NULL. Doing it this way is more efficient for a
very common case.
So as long as a counted reference is held to a dentry, a non-
value will never be changed.
d_lock is a synonym for the spinlock that is part of
For our purposes, holding this lock protects against the dentry being
renamed or unlinked. In particular, its parent (
d_parent), and its
d_name) cannot be changed, and it cannot be removed from the
dentry hash table.
When looking for a name in a directory, REF-walk takes
each candidate dentry that it finds in the hash table and then checks
that the parent and name are correct. So it doesn’t lock the parent
while searching in the cache; it only locks children.
When looking for the parent for a given name (to handle “
REF-walk can take
d_lock to get a stable reference to
but it first tries a more lightweight approach. As seen in
dget_parent(), if a reference can be claimed on the parent, and if
d_parent can be seen to have not changed, then there is
no need to actually take the lock on the child.
Looking up a given name in a given directory involves computing a hash from the two values (the name and the dentry of the directory), accessing that slot in a hash table, and searching the linked list that is found there.
When a dentry is renamed, the name and the parent dentry can both change so the hash will almost certainly change too. This would move the dentry to a different chain in the hash table. If a filename search happened to be looking at a dentry that was moved in this way, it might end up continuing the search down the wrong chain, and so miss out on part of the correct chain.
The name-lookup process (
d_lookup()) does not try to prevent this
from happening, but only to detect when it happens.
rename_lock is a seqlock that is updated whenever any dentry is
d_lookup finds that a rename happened while it
unsuccessfully scanned a chain in the hash table, it simply tries
rename_lock is also used to detect and defend against potential attacks
LOOKUP_IN_ROOT when resolving “..” (where
the parent directory is moved outside the root, bypassing the
rename_lock is updated during the lookup and the path encounters
a “..”, a potential attack occurred and
handle_dots() will bail out with
i_rwsem is a read/write semaphore that serializes all changes to a particular
directory. This ensures that, for example, an
unlink() and a
cannot both happen at the same time. It also keeps the directory
stable while the filesystem is asked to look up a name that is not
currently in the dcache or, optionally, when the list of entries in a
directory is being retrieved with
This has a complementary role to that of
i_rwsem on a
directory protects all of the names in that directory, while
on a name protects just one name in a directory. Most changes to the
i_rwsem on the relevant directory inode and briefly take
d_lock on one or more the dentries while the change happens. One
exception is when idle dentries are removed from the dcache due to
memory pressure. This uses
i_rwsem plays no role.
The semaphore affects pathname lookup in two distinct ways. Firstly it
prevents changes during lookup of a name in a directory.
lookup_fast() first which, in turn, checks to see if the name is in the cache,
d_lock locking. If the name isn’t found, then
falls back to
lookup_slow() which takes a shared lock on
i_rwsem, checks again that
the name isn’t in the cache, and then calls in to the filesystem to get a
definitive answer. A new dentry will be added to the cache regardless of
Secondly, when pathname lookup reaches the final component, it will
sometimes need to take an exclusive lock on
i_rwsem before performing the last lookup so
that the required exclusion can be achieved. How path lookup chooses
to take, or not take,
i_rwsem is one of the
issues addressed in a subsequent section.
If two threads attempt to look up the same name at the same time - a
name that is not yet in the dcache - the shared lock on
not prevent them both adding new dentries with the same name. As this
would result in confusion an extra level of interlocking is used,
based around a secondary hash table (
in_lookup_hashtable) and a
per-dentry flag bit (
To add a new dentry to the cache while only holding a shared lock on
i_rwsem, a thread must call
d_alloc_parallel(). This allocates a
dentry, stores the required name and parent in it, checks if there
is already a matching dentry in the primary or secondary hash
tables, and if not, stores the newly allocated dentry in the secondary
hash table, with
If a matching dentry was found in the primary hash table then that is
returned and the caller can know that it lost a race with some other
thread adding the entry. If no matching dentry is found in either
cache, the newly allocated dentry is returned and the caller can
detect this from the presence of
DCACHE_PAR_LOOKUP. In this case it
knows that it has won any race and now is responsible for asking the
filesystem to perform the lookup and find the matching inode. When
the lookup is complete, it must call
d_lookup_done() which clears
the flag and does some other house keeping, including removing the
dentry from the secondary hash table - it will normally have been
added to the primary hash table already. Note that a
waitqueue_head is passed to
d_lookup_done() must be called while this
waitqueue_head is still
If a matching dentry is found in the secondary hash table,
d_alloc_parallel() has a little more work to do. It first waits for
DCACHE_PAR_LOOKUP to be cleared, using a wait_queue that was passed
to the instance of
d_alloc_parallel() that won the race and that
will be woken by the call to
d_lookup_done(). It then checks to see
if the dentry has now been added to the primary hash table. If it
has, the dentry is returned and the caller just sees that it lost any
race. If it hasn’t been added to the primary hash table, the most
likely explanation is that some other dentry was added instead using
d_splice_alias(). In any case,
d_alloc_parallel() repeats all the
look ups from the start and will normally return something from the
primary hash table.
mnt_count is a per-CPU reference counter on “
Per-CPU here means that incrementing the count is cheap as it only
uses CPU-local memory, but checking if the count is zero is expensive as
it needs to check with every CPU. Taking a
prevents the mount structure from disappearing as the result of regular
unmount operations, but does not prevent a “lazy” unmount. So holding
mnt_count doesn’t ensure that the mount remains in the namespace and,
in particular, doesn’t stabilize the link to the mounted-on dentry. It
does, however, ensure that the
mount data structure remains coherent,
and it provides a reference to the root dentry of the mounted
filesystem. So a reference through
->mnt_count provides a stable
reference to the mounted dentry, but not the mounted-on dentry.
mount_lock is a global seqlock, a bit like
rename_lock. It can be used to
check if any change has been made to any mount points.
While walking down the tree (away from the root) this lock is used when
crossing a mount point to check that the crossing was safe. That is,
the value in the seqlock is read, then the code finds the mount that
is mounted on the current directory, if there is one, and increments
mnt_count. Finally the value in
mount_lock is checked against
the old value. If there is no change, then the crossing was safe. If there
was a change, the
mnt_count is decremented and the whole process is
When walking up the tree (towards the root) by following a “..” link, a little more care is needed. In this case the seqlock (which contains both a counter and a spinlock) is fully locked to prevent any changes to any mount points while stepping up. This locking is needed to stabilize the link to the mounted-on dentry, which the refcount on the mount itself doesn’t ensure.
mount_lock is also used to detect and defend against potential attacks
LOOKUP_IN_ROOT when resolving “..” (where
the parent directory is moved outside the root, bypassing the
mount_lock is updated during the lookup and the path encounters
a “..”, a potential attack occurred and
handle_dots() will bail out with
Finally the global (but extremely lightweight) RCU read lock is held from time to time to ensure certain data structures don’t get freed unexpectedly.
In particular it is held while scanning chains in the dcache hash table, and the mount point hash table.
Bringing it together with
Throughout the process of walking a path, the current status is stored
struct nameidata, “namei” being the traditional name - dating
all the way back to First Edition Unix - of the function that
converts a “name” to an “inode”.
struct nameidata contains (among
struct path path¶
path contains a
struct vfsmount (which is
embedded in a
struct mount) and a
struct dentry. Together these
record the current status of the walk. They start out referring to the
starting point (the current working directory, the root directory, or some other
directory identified by a file descriptor), and are updated on each
step. A reference through
mnt_count is always
struct qstr last¶
This is a string together with a length (i.e. not
that is the “next” component in the pathname.
This is one of
last field is only valid if the type is
struct path root¶
This is used to hold a reference to the effective root of the
filesystem. Often that reference won’t be needed, so this field is
only assigned the first time it is used, or when a non-standard root
is requested. Keeping a reference in the
nameidata ensures that
only one root is in effect for the entire path walk, even if it races
chroot() system call.
It should be noted that in the case of
LOOKUP_BENEATH, the effective root becomes the directory file descriptor
openat2() (which exposes these
The root is needed when either of two conditions holds: (1) either the
pathname or a symbolic link starts with a “’/’”, or (2) a “
component is being handled, since “
..” from the root must always stay
at the root. The value used is usually the current root directory of
the calling process. An alternate root can be provided as when
file_open_root(), and when NFSv4 or Btrfs call
mount_subtree(). In each case a pathname is being looked up in a very
specific part of the filesystem, and the lookup must not be allowed to
escape that subtree. It works a bit like a local
Ignoring the handling of symbolic links, we can now describe the
link_path_walk()” function, which handles the lookup of everything
except the final component as:
Given a path (
name) and a nameidata structure (
nd), check that the current directory has execute permission and then advance
nameover one component while updating
last. If that was the final component, then return, otherwise call
walk_component()and repeat from the top.
walk_component() is even easier. If the component is
handle_dots() which does the necessary locking as already
described. If it finds a
LAST_NORM component it first calls
lookup_fast()” which only looks in the dcache, but will ask the
filesystem to revalidate the result if it is that sort of filesystem.
If that doesn’t get a good result, it calls “
i_rwsem, rechecks the cache, and then asks the filesystem
to find a definitive answer. Each of these will call
follow_managed() (as described below) to handle any mount points.
In the absence of symbolic links,
walk_component() creates a new
struct path containing a counted reference to the new dentry and a
reference to the new
vfsmount which is only counted if it is
different from the previous
vfsmount. It then calls
path_to_nameidata() to install the new
struct path in the
struct nameidata and drop the unneeded references.
This “hand-over-hand” sequencing of getting a reference to the new dentry before dropping the reference to the previous dentry may seem obvious, but is worth pointing out so that we will recognize its analogue in the “RCU-walk” version.
Handling the final component¶
link_path_walk() only walks as far as setting
nd->last_type to refer to the final component of the path. It does
walk_component() that last time. Handling that final
component remains for the caller to sort out. Those callers are
path_openat() each of which handles the differing requirements of
different system calls.
path_parentat() is clearly the simplest - it just wraps a little bit
of housekeeping around
link_path_walk() and returns the parent
directory and final component to the caller. The caller will be either
aiming to create a name (via
filename_create()) or remove or rename
a name (in which case
user_path_parent() is used). They will use
i_rwsem to exclude other changes while they validate and then
perform their operation.
path_lookupat() is nearly as simple - it is used when an existing
object is wanted such as by
chmod(). It essentially just
walk_component() on the final component through a call to
path_lookupat() returns just the final dentry.
path_mountpoint() handles the special case of unmounting which must
not try to revalidate the mounted filesystem. It effectively
contains, through a call to
mountpoint_last(), an alternate
lookup_slow() which skips that step. This is
important when unmounting a filesystem that is inaccessible, such as
one provided by a dead NFS server.
path_openat() is used for the
open() system call; it
contains, in support functions starting with “
do_last()”, all the
complexity needed to handle the different subtleties of O_CREAT (with
or without O_EXCL), final “
/” characters, and trailing symbolic
links. We will revisit this in the final part of this series, which
focuses on those symbolic links. “
do_last()” will sometimes, but
not always, take
i_rwsem, depending on what it finds.
Each of these, or the functions which call them, need to be alert to
the possibility that the final component is not
LAST_NORM. If the
goal of the lookup is to create something, then any value for
last_type other than
LAST_NORM will result in an error. For
LAST_DOTDOT, then the caller
won’t try to create that name. They also check for trailing slashes
last.name[last.len]. If there is any character beyond
the final component, it must be a trailing slash.
Revalidation and automounts¶
Apart from symbolic links, there are only two parts of the “REF-walk” process not yet covered. One is the handling of stale cache entries and the other is automounts.
On filesystems that require it, the lookup routines will call the
->d_revalidate() dentry method to ensure that the cached information
is current. This will often confirm validity or update a few details
from a server. In some cases it may find that there has been change
further up the path and that something that was thought to be valid
previously isn’t really. When this happens the lookup of the whole
path is aborted and retried with the “
LOOKUP_REVAL” flag set. This
forces revalidation to be more thorough. We will see more details of
this retry process in the next article.
Automount points are locations in the filesystem where an attempt to lookup a name can trigger changes to how that lookup should be handled, in particular by mounting a filesystem there. These are covered in greater detail in autofs.txt in the Linux documentation tree, but a few notes specifically related to path lookup are in order here.
The Linux VFS has a concept of “managed” dentries which is reflected
in function names such as “
follow_managed()”. There are three
potentially interesting things about these dentries corresponding
to three different flags that might be set in
If this flag has been set, then the filesystem has requested that the
d_manage() dentry operation be called before handling any possible
mount point. This can perform two particular services:
It can block to avoid races. If an automount point is being
d_manage() function will usually wait for that
process to complete before letting the new lookup proceed and possibly
trigger a new automount.
It can selectively allow only some processes to transit through a
mount point. When a server process is managing automounts, it may
need to access a directory without triggering normal automount
processing. That server process can identify itself to the
filesystem, which will then give it a special pass through
d_manage() by returning
This flag is set on every dentry that is mounted on. As Linux supports multiple filesystem namespaces, it is possible that the dentry may not be mounted on in this namespace, just in some other. So this flag is seen as a hint, not a promise.
If this flag is set, and
d_manage() didn’t return
lookup_mnt() is called to examine the mount hash table (honoring the
mount_lock described earlier) and possibly return a new
and a new
dentry (both with counted references).
d_manage() allowed us to get this far, and
find a mount point, then this flag causes the
operation to be called.
d_automount() operation can be arbitrarily complex and may
communicate with server processes etc. but it should ultimately either
report that there was an error, that there was nothing to mount, or
should provide an updated
struct path with new
In the latter case,
finish_automount() will be called to safely
install the new mount point into the mount table.
There is no new locking of import here and it is important that no locks (only counted references) are held over this processing due to the very real possibility of extended delays. This will become more important next time when we examine RCU-walk which is particularly sensitive to delays.
RCU-walk - faster pathname lookup in Linux¶
RCU-walk is another algorithm for performing pathname lookup in Linux. It is in many ways similar to REF-walk and the two share quite a bit of code. The significant difference in RCU-walk is how it allows for the possibility of concurrent access.
We noted that REF-walk is complex because there are numerous details and special cases. RCU-walk reduces this complexity by simply refusing to handle a number of cases – it instead falls back to REF-walk. The difficulty with RCU-walk comes from a different direction: unfamiliarity. The locking rules when depending on RCU are quite different from traditional locking, so we will spend a little extra time when we come to those.
Clear demarcation of roles¶
The easiest way to manage concurrency is to forcibly stop any other thread from changing the data structures that a given thread is looking at. In cases where no other thread would even think of changing the data and lots of different threads want to read at the same time, this can be very costly. Even when using locks that permit multiple concurrent readers, the simple act of updating the count of the number of current readers can impose an unwanted cost. So the goal when reading a shared data structure that no other process is changing is to avoid writing anything to memory at all. Take no locks, increment no counts, leave no footprints.
The REF-walk mechanism already described certainly doesn’t follow this principle, but then it is really designed to work when there may well be other threads modifying the data. RCU-walk, in contrast, is designed for the common situation where there are lots of frequent readers and only occasional writers. This may not be common in all parts of the filesystem tree, but in many parts it will be. For the other parts it is important that RCU-walk can quickly fall back to using REF-walk.
Pathname lookup always starts in RCU-walk mode but only remains there as long as what it is looking for is in the cache and is stable. It dances lightly down the cached filesystem image, leaving no footprints and carefully watching where it is, to be sure it doesn’t trip. If it notices that something has changed or is changing, or if something isn’t in the cache, then it tries to stop gracefully and switch to REF-walk.
This stopping requires getting a counted reference on the current
dentry, and ensuring that these are still valid -
that a path walk with REF-walk would have found the same entries.
This is an invariant that RCU-walk must guarantee. It can only make
decisions, such as selecting the next step, that are decisions which
REF-walk could also have made if it were walking down the tree at the
same time. If the graceful stop succeeds, the rest of the path is
processed with the reliable, if slightly sluggish, REF-walk. If
RCU-walk finds it cannot stop gracefully, it simply gives up and
restarts from the top with REF-walk.
This pattern of “try RCU-walk, if that fails try REF-walk” can be
clearly seen in functions like
do_file_open_root(). These five
correspond roughly to the four
path_*() functions we met earlier,
each of which calls
path_*() functions are
called using different mode flags until a mode is found which works.
They are first called with
LOOKUP_RCU set to request “RCU-walk”. If
that fails with the error
ECHILD they are called again with no
special flag to request “REF-walk”. If either of those report the
ESTALE a final attempt is made with
LOOKUP_REVAL set (and no
LOOKUP_RCU) to ensure that entries found in the cache are forcibly
revalidated - normally entries are only revalidated if the filesystem
determines that they are too old to trust.
LOOKUP_RCU attempt may drop that flag internally and switch to
REF-walk, but will never then try to switch back to RCU-walk. Places
that trip up RCU-walk are much more likely to be near the leaves and
so it is very unlikely that there will be much, if any, benefit from
RCU and seqlocks: fast and light¶
RCU is, unsurprisingly, critical to RCU-walk mode. The
rcu_read_lock() is held for the entire time that RCU-walk is walking
down a path. The particular guarantee it provides is that the key
data structures - dentries, inodes, super_blocks, and mounts - will
not be freed while the lock is held. They might be unlinked or
invalidated in one way or another, but the memory will not be
repurposed so values in various fields will still be meaningful. This
is the only guarantee that RCU provides; everything else is done using
As we saw above, REF-walk holds a counted reference to the current
dentry and the current vfsmount, and does not release those references
before taking references to the “next” dentry or vfsmount. It also
sometimes takes the
d_lock spinlock. These references and locks are
taken to prevent certain changes from happening. RCU-walk must not
take those references or locks and so cannot prevent such changes.
Instead, it checks to see if a change has been made, and aborts or
retries if it has.
To preserve the invariant mentioned above (that RCU-walk may only make
decisions that REF-walk could have made), it must make the checks at
or near the same places that REF-walk holds the references. So, when
REF-walk increments a reference count or takes a spinlock, RCU-walk
samples the status of a seqlock using
read_seqcount_begin() or a
similar function. When REF-walk decrements the count or drops the
lock, RCU-walk checks if the sampled status is still valid using
read_seqcount_retry() or similar.
However, there is a little bit more to seqlocks than that. If
RCU-walk accesses two different fields in a seqlock-protected
structure, or accesses the same field twice, there is no a priori
guarantee of any consistency between those accesses. When consistency
is needed - which it usually is - RCU-walk must take a copy and then
read_seqcount_retry() to validate that copy.
read_seqcount_retry() not only checks the sequence number, but also
imposes a memory barrier so that no memory-read instruction from
before the call can be delayed until after the call, either by the
CPU or by the compiler. A simple example of this can be seen in
slow_dentry_cmp() which, for filesystems which do not use simple
byte-wise name equality, calls into the filesystem to compare a name
against a dentry. The length and name pointer are copied into local
read_seqcount_retry() is called to confirm the two
are consistent, and only then is
->d_compare() called. When
standard filename comparison is used,
dentry_cmp() is called
instead. Notably it does not use
instead has a large comment explaining why the consistency guarantee
isn’t necessary. A subsequent
read_seqcount_retry() will be
sufficient to catch any problem that could occur at this point.
With that little refresher on seqlocks out of the way we can look at the bigger picture of how RCU-walk uses seqlocks.
We already met the
mount_lock seqlock when REF-walk used it to
ensure that crossing a mount point is performed safely. RCU-walk uses
it for that too, but for quite a bit more.
Instead of taking a counted reference to each
vfsmount as it
descends the tree, RCU-walk samples the state of
mount_lock at the
start of the walk and stores this initial sequence number in the
struct nameidata in the
m_seq field. This one lock and one
sequence number are used to validate all accesses to all
and all mount point crossings. As changes to the mount table are
relatively rare, it is reasonable to fall back on REF-walk any time
that any “mount” or “unmount” happens.
m_seq is checked (using
read_seqretry()) at the end of an RCU-walk
sequence, whether switching to REF-walk for the rest of the path or
when the end of the path is reached. It is also checked when stepping
down over a mount point (in
__follow_mount_rcu()) or up (in
follow_dotdot_rcu()). If it is ever found to have changed, the
whole RCU-walk sequence is aborted and the path is processed again by
If RCU-walk finds that
mount_lock hasn’t changed then it can be sure
that, had REF-walk taken counted references on each vfsmount, the
results would have been the same. This ensures the invariant holds,
at least for vfsmount structures.
In place of taking a count or lock on
d_reflock, RCU-walk samples
d_seq seqlock, and stores the sequence number in the
seq field of the nameidata structure, so
nd->seq should always be
the current sequence number of
nd->dentry. This number needs to be
revalidated after copying, and before using, the name, parent, or
inode of the dentry.
The handling of the name we have already looked at, and the parent is
only accessed in
follow_dotdot_rcu() which fairly trivially follows
the required pattern, though it does so for three different cases.
When not at a mount point,
d_parent is followed and its
collected. When we are at a mount point, we instead follow the
mnt->mnt_mountpoint link to get a new dentry and collect its
d_seq. Then, after finally finding a
d_parent to follow, we must
check if we have landed on a mount point and, if so, must find that
mount point and follow the
mnt->mnt_root link. This would imply a
somewhat unusual, but certainly possible, circumstance where the
starting point of the path lookup was in part of the filesystem that
was mounted on, and so not visible from the root.
The inode pointer, stored in
->d_inode, is a little more
interesting. The inode will always need to be accessed at least
twice, once to determine if it is NULL and once to verify access
permissions. Symlink handling requires a validated inode pointer too.
Rather than revalidating on each access, a copy is made on the first
access and it is stored in the
inode field of
nameidata from where
it can be safely accessed without further validation.
lookup_fast() is the only lookup routine that is used in RCU-mode,
lookup_slow() being too slow and requiring locks. It is in
lookup_fast() that we find the important “hand over hand” tracking
of the current dentry.
dentry and current
seq number are passed to
__d_lookup_rcu() which, on success, returns a new
dentry and a
lookup_fast() then copies the inode pointer and
revalidates the new
seq number. It then validates the old
with the old
seq number one last time and only then continues. This
process of getting the
seq number of the new dentry and then
seq number of the old exactly mirrors the process of
getting a counted reference to the new dentry before dropping that for
the old dentry which we saw in REF-walk.
inode->i_rwsem or even
A semaphore is a fairly heavyweight lock that can only be taken when it is
permissible to sleep. As
rcu_read_lock() forbids sleeping,
inode->i_rwsem plays no role in RCU-walk. If some other thread does
i_rwsem and modifies the directory in a way that RCU-walk needs
to notice, the result will be either that RCU-walk fails to find the
dentry that it is looking for, or it will find a dentry which
read_seqretry() won’t validate. In either case it will drop down to
REF-walk mode which can take whatever locks are needed.
rename_lock could be used by RCU-walk as it doesn’t require
any sleeping, RCU-walk doesn’t bother. REF-walk uses
protect against the possibility of hash chains in the dcache changing
while they are being searched. This can result in failing to find
something that actually is there. When RCU-walk fails to find
something in the dentry cache, whether it is really there or not, it
already drops down to REF-walk and tries again with appropriate
locking. This neatly handles all cases, so adding extra checks on
rename_lock would bring no significant value.
unlazy walk() and
That “dropping down to REF-walk” typically involves a call to
unlazy_walk(), so named because “RCU-walk” is also sometimes
referred to as “lazy walk”.
unlazy_walk() is called when
following the path down to the current vfsmount/dentry pair seems to
have proceeded successfully, but the next step is problematic. This
can happen if the next name cannot be found in the dcache, if
permission checking or name revalidation couldn’t be achieved while
rcu_read_lock() is held (which forbids sleeping), if an
automount point is found, or in a couple of cases involving symlinks.
It is also called from
complete_walk() when the lookup has reached
the final component, or the very end of the path, depending on which
particular flavor of lookup is used.
Other reasons for dropping out of RCU-walk that do not trigger a call
unlazy_walk() are when some inconsistency is found that cannot be
handled immediately, such as
mount_lock or one of the
seqlocks reporting a change. In these cases the relevant function
-ECHILD which will percolate up until it triggers a new
attempt from the top using REF-walk.
For those cases where
unlazy_walk() is an option, it essentially
takes a reference on each of the pointers that it holds (vfsmount,
dentry, and possibly some symbolic links) and then verifies that the
relevant seqlocks have not been changed. If there have been changes,
it, too, aborts with
-ECHILD, otherwise the transition to REF-walk
has been a success and the lookup process continues.
Taking a reference on those pointers is not quite as simple as just
incrementing a counter. That works to take a second reference if you
already have one (often indirectly through another object), but it
isn’t sufficient if you don’t actually have a counted reference at
dentry->d_lockref, it is safe to increment the reference
counter to get a reference unless it has been explicitly marked as
“dead” which involves setting the counter to
lockref_get_not_dead() achieves this.
mnt->mnt_count it is safe to take a reference as long as
mount_lock is then used to validate the reference. If that
validation fails, it may not be safe to just drop that reference in
the standard way of calling
mnt_put() - an unmount may have
progressed too far. So the code in
legitimize_mnt(), when it
finds that the reference it got might not be safe, checks the
MNT_SYNC_UMOUNT flag to determine if a simple
correct, or if it should just decrement the count and pretend none of
this ever happened.
Taking care in filesystems¶
RCU-walk depends almost entirely on cached information and often will not call into the filesystem at all. However there are two places, besides the already-mentioned component-name comparison, where the file system might be included in RCU-walk, and it must know to be careful.
If the filesystem has non-standard permission-checking requirements -
such as a networked filesystem which may need to check with the server
i_op->permission interface might be called during RCU-walk.
In this case an extra “
MAY_NOT_BLOCK” flag is passed so that it
knows not to sleep, but to return
-ECHILD if it cannot complete
i_op->permission is given the inode pointer, not the
dentry, so it doesn’t need to worry about further consistency checks.
However if it accesses any other filesystem data structures, it must
ensure they are safe to be accessed with only the
held. This typically means they must be freed using
If the filesystem may need to revalidate dcache entries, then
d_op->d_revalidate may be called in RCU-walk too. This interface
is passed the dentry but does not have access to the
inode or the
seq number from the
nameidata, so it needs to be extra careful
when accessing fields in the dentry. This “extra care” typically
involves using READ_ONCE() to access fields, and verifying the
result is not NULL before using it. This pattern can be seen in
A pair of patterns¶
In various places in the details of REF-walk and RCU-walk, and also in the big picture, there are a couple of related patterns that are worth being aware of.
The first is “try quickly and check, if that fails try slowly”. We
can see that in the high-level approach of first trying RCU-walk and
then trying REF-walk, and in places where
unlazy_walk() is used to
switch to REF-walk for the rest of the path. We also saw it earlier
dget_parent() when following a “
..” link. It tries a quick way
to get a reference, then falls back to taking locks if needed.
The second pattern is “try quickly and check, if that fails try
again - repeatedly”. This is seen with the use of
mount_lock in REF-walk. RCU-walk doesn’t make use of this pattern -
if anything goes wrong it is much safer to just abort and try a more
The emphasis here is “try quickly and check”. It should probably be “try quickly and carefully, then check”. The fact that checking is needed is a reminder that the system is dynamic and only a limited number of things are safe at all. The most likely cause of errors in this whole process is assuming something is safe when in reality it isn’t. Careful consideration of what exactly guarantees the safety of each access is sometimes necessary.