[PATCH 03/42] Unionfs: documentation for general concepts

[Erez Zadok <ezk@cs.sunysb.edu>]

Signed-off-by: Erez Zadok <ezk@cs.sunysb.edu>
---
 Documentation/filesystems/unionfs/concepts.txt |  199 ++++++++++++++++++++++++
 1 files changed, 199 insertions(+), 0 deletions(-)
 create mode 100644 Documentation/filesystems/unionfs/concepts.txt

diff --git a/Documentation/filesystems/unionfs/concepts.txt b/Documentation/filesystems/unionfs/concepts.txt
new file mode 100644
index 0000000..7654ccc
--- /dev/null
+++ b/Documentation/filesystems/unionfs/concepts.txt
@@ -0,0 +1,199 @@
+Unionfs 2.1 CONCEPTS:
+=====================
+
+This file describes the concepts needed by a namespace unification file
+system.
+
+
+Branch Priority:
+================
+
+Each branch is assigned a unique priority - starting from 0 (highest
+priority).  No two branches can have the same priority.
+
+
+Branch Mode:
+============
+
+Each branch is assigned a mode - read-write or read-only. This allows
+directories on media mounted read-write to be used in a read-only manner.
+
+
+Whiteouts:
+==========
+
+A whiteout removes a file name from the namespace. Whiteouts are needed when
+one attempts to remove a file on a read-only branch.
+
+Suppose we have a two-branch union, where branch 0 is read-write and branch
+1 is read-only. And a file 'foo' on branch 1:
+
+./b0/
+./b1/
+./b1/foo
+
+The unified view would simply be:
+
+./union/
+./union/foo
+
+Since 'foo' is stored on a read-only branch, it cannot be removed. A
+whiteout is used to remove the name 'foo' from the unified namespace. Again,
+since branch 1 is read-only, the whiteout cannot be created there. So, we
+try on a higher priority (lower numerically) branch and create the whiteout
+there.
+
+./b0/
+./b0/.wh.foo
+./b1/
+./b1/foo
+
+Later, when Unionfs traverses branches (due to lookup or readdir), it
+eliminate 'foo' from the namespace (as well as the whiteout itself.)
+
+
+Duplicate Elimination:
+======================
+
+It is possible for files on different branches to have the same name.
+Unionfs then has to select which instance of the file to show to the user.
+Given the fact that each branch has a priority associated with it, the
+simplest solution is to take the instance from the highest priority
+(numerically lowest value) and "hide" the others.
+
+
+Copyup:
+=======
+
+When a change is made to the contents of a file's data or meta-data, they
+have to be stored somewhere. The best way is to create a copy of the
+original file on a branch that is writable, and then redirect the write
+though to this copy. The copy must be made on a higher priority branch so
+that lookup and readdir return this newer "version" of the file rather than
+the original (see duplicate elimination).
+
+
+Cache Coherency:
+================
+
+Unionfs users often want to be able to modify files and directories directly
+on the lower branches, and have those changes be visible at the Unionfs
+level.  This means that data (e.g., pages) and meta-data (dentries, inodes,
+open files, etc.) have to be synchronized between the upper and lower
+layers.  In other words, the newest changes from a layer below have to be
+propagated to the Unionfs layer above.  If the two layers are not in sync, a
+cache incoherency ensues, which could lead to application failures and even
+oopses.  The Linux kernel, however, has a rather limited set of mechanisms
+to ensure this inter-layer cache coherency---so Unionfs has to do most of
+the hard work on its own.
+
+Maintaining Invariants:
+
+The way Unionfs ensures cache coherency is as follows.  At each entry point
+to a Unionfs file system method, we call a utility function to validate the
+primary objects of this method.  Generally, we call unionfs_file_revalidate
+on open files, and __unionfs_d_revalidate_chain on dentries (which also
+validates inodes).  These utility functions check to see whether the upper
+Unionfs object is in sync with any of the lower objects that it represents.
+The checks we perform include whether the Unionfs superblock has a newer
+generation number, or if any of the lower objects mtime's or ctime's are
+newer.  (Note: generation numbers change when branch-management commands are
+issued, so in a way, maintaining cache coherency is also very important for
+branch-management.)  If indeed we determine that any Unionfs object is no
+longer in sync with its lower counterparts, then we rebuild that object
+similarly to how we do so for branch-management.
+
+While rebuilding Unionfs's objects, we also purge any page mappings and
+truncate inode pages (see fs/unionfs/dentry.c:purge_inode_data).  This is to
+ensure that Unionfs will re-get the newer data from the lower branches.  We
+perform this purging only if the Unionfs operation in question is a reading
+operation; if Unionfs is performing a data writing operation (e.g., ->write,
+->commit_write, etc.) then we do NOT flush the lower mappings/pages: this is
+because (1) a self-deadlock could occur and (2) the upper Unionfs pages are
+considered more authoritative anyway, as they are newer and will overwrite
+any lower pages.
+
+Unionfs maintains the following important invariant regarding mtime's,
+ctime's, and atime's: the upper inode object's times are the max() of all of
+the lower ones.  For non-directory objects, there's only one object below,
+so the mapping is simple; for directory objects, there could me multiple
+lower objects and we have to sync up with the newest one of all the lower
+ones.  This invariant is important to maintain, especially for directories
+(besides, we need this to be POSIX compliant).  A union could comprise
+multiple writable branches, each of which could change.  If we don't reflect
+the newest possible mtime/ctime, some applications could fail.  For example,
+NFSv2/v3 exports check for newer directory mtimes on the server to determine
+if the client-side attribute cache should be purged.
+
+To maintain these important invariants, of course, Unionfs carefully
+synchronizes upper and lower times in various places.  For example, if we
+copy-up a file to a top-level branch, the parent directory where the file
+was copied up to will now have a new mtime: so after a successful copy-up,
+we sync up with the new top-level branch's parent directory mtime.
+
+Implementation:
+
+This cache-coherency implementation is efficient because it defers any
+synchronizing between the upper and lower layers until absolutely needed.
+Consider the example a common situation where users perform a lot of lower
+changes, such as untarring a whole package.  While these take place,
+typically the user doesn't access the files via Unionfs; only after the
+lower changes are done, does the user try to access the lower files.  With
+our cache-coherency implementation, the entirety of the changes to the lower
+branches will not result in a single CPU cycle spent at the Unionfs level
+until the user invokes a system call that goes through Unionfs.
+
+We have considered two alternate cache-coherency designs.  (1) Using the
+dentry/inode notify functionality to register interest in finding out about
+any lower changes.  This is a somewhat limited and also a heavy-handed
+approach which could result in many notifications to the Unionfs layer upon
+each small change at the lower layer (imagine a file being modified multiple
+times in rapid succession).  (2) Rewriting the VFS to support explicit
+callbacks from lower objects to upper objects.  We began exploring such an
+implementation, but found it to be very complicated--it would have resulted
+in massive VFS/MM changes which are unlikely to be accepted by the LKML
+community.  We therefore believe that our current cache-coherency design and
+implementation represent the best approach at this time.
+
+Limitations:
+
+Our implementation works in that as long as a user process will have caused
+Unionfs to be called, directly or indirectly, even to just do
+->d_revalidate; then we will have purged the current Unionfs data and the
+process will see the new data.  For example, a process that continually
+re-reads the same file's data will see the NEW data as soon as the lower
+file had changed, upon the next read(2) syscall (even if the file is still
+open!)  However, this doesn't work when the process re-reads the open file's
+data via mmap(2) (unless the user unmaps/closes the file and remaps/reopens
+it).  Once we respond to ->readpage(s), then the kernel maps the page into
+the process's address space and there doesn't appear to be a way to force
+the kernel to invalidate those pages/mappings, and force the process to
+re-issue ->readpage.  If there's a way to invalidate active mappings and
+force a ->readpage, let us know please (invalidate_inode_pages2 doesn't do
+the trick).
+
+Our current Unionfs code has to perform many file-revalidation calls.  It
+would be really nice if the VFS would export an optional file system hook
+->file_revalidate (similarly to dentry->d_revalidate) that will be called
+before each VFS op that has a "struct file" in it.
+
+Certain file systems have micro-second granularity (or better) for inode
+times, and asynchronous actions could cause those times to change with some
+small delay.  In such cases, Unionfs may see a changed inode time that only
+differs by a tiny fraction of a second: such a change may be a false
+positive indication that the lower object has changed, whereas if unionfs
+waits a little longer, that false indication will not be seen.  (These false
+positives are harmless, because they would at most cause unionfs to
+re-validate an object that may need no revalidation, and print a debugging
+message that clutters the console/logs.)  Therefore, to minimize the chances
+of these situations, we delay the detection of changed times by a small
+factor of a few seconds, called UNIONFS_MIN_CC_TIME (which defaults to 3
+seconds, as does NFS).  This means that we will detect the change, only a
+couple of seconds later, if indeed the time change persists in the lower
+file object.  This delayed detection has an added performance benefit: we
+reduce the number of times that unionfs has to revalidate objects, in case
+there's a lot of concurrent activity on both the upper and lower objects,
+for the same file(s).  Lastly, this delayed time attribute detection is
+similar to how NFS clients operate (e.g., acregmin).
+
+For more information, see <http://unionfs.filesystems.org/>.
-- 
1.5.2.2

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/