[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
JFFS2 locking documentation.
The 2.5 kernel pushed down the BKL into file systems. In theory, we ought
to be able to just drop it from JFFS2 because we have our own locking.
In practice, it does seem to survive like that in initial testing too.
It wants a proper audit before I actually take the BKL out of the version in
Linus' tree though. To that end, and partly because of some stupid mistakes
I made this week when I forgot the ordering constraints, I've finally
managed to document the locking.
I've fixed some bugs already, and implemented a fix for the GC of deletion
dirent case which I'm fairly unhappy with but can't see an alternative.
If anyone feels like casting an eye over the documentation and the code,
and trying to spot any remaining problems, that would be very much
appreciated. Once I'm fairly convinced it's OK I'll ask Al Viro to look at
it - but I'd rather be fairly sure it's correct before I'm brave enough to
do that :)
$Id: README.Locking,v 1.4 2002/03/08 16:20:06 dwmw2 Exp $
JFFS2 LOCKING DOCUMENTATION
At least theoretically, JFFS2 does not require the Big Kernel Lock
(BKL), which was always helpfully obtained for it by Linux 2.4 VFS
code. It has its own locking, as described below.
This document attempts to describe the existing locking rules for
JFFS2. It is not expected to remain perfectly up to date, but ought to
be fairly close.
The alloc_sem is a per-filesystem semaphore, used primarily to ensure
contiguous allocation of space on the medium. It is automatically
obtained during space allocations (jffs2_reserve_space()) and freed
upon write completion (jffs2_complete_reservation()). Note that
the garbage collector will obtain this right at the beginning of
jffs2_garbage_collect_pass() and release it at the end, thereby
preventing any other write activity on the file system during a
garbage collect pass.
When writing new nodes, the alloc_sem must be held until the new nodes
have been properly linked into the data structures for the inode to
which they belong. This is for the benefit of NAND flash - adding new
nodes to an inode may obsolete old ones, and by holding the alloc_sem
until this happens we ensure that any data in the write-buffer at the
time this happens are part of the new node, not just something that
was written afterwards. Hence, we can ensure the newly-obsoleted nodes
don't actually get erased until the write-buffer has been flushed to
With the introduction of NAND flash support and the write-buffer,
the alloc_sem is also used to protect the wbuf-related members of the
jffs2_sb_info structure. Atomically reading the wbuf_len member to see
if the wbuf is currently holding any data is permitted, though.
Ordering constraints: See f->sem.
File Semaphore f->sem
This is the JFFS2-internal equivalent of the inode semaphore i->i_sem.
It protects the contents of the jffs2_inode_info private inode data,
including the linked list of node fragments (but see the notes below on
The reason that the i_sem itself isn't used for this purpose is to
avoid deadlocks with garbage collection -- the VFS will lock the i_sem
before calling a function which may need to allocate space. The
allocation may trigger garbage-collection, which may need to move a
node belonging to the inode which was locked in the first place by the
VFS. If the garbage collection code were to attempt to lock the i_sem
of the inode from which it's garbage-collecting a physical node, this
lead to deadlock, unless we played games with unlocking the i_sem
before calling the space allocation functions.
Instead of playing such games, we just have an extra internal
semaphore, which is obtained by the garbage collection code and also
by the normal file system code _after_ allocation of space.
1. Never attempt to allocate space or lock alloc_sem with
any f->sem held.
2. Never attempt to lock two file semaphores in one thread.
No ordering rules have been made for doing so.
This is used to serialise access to the eraseblock lists, to the
per-eraseblock lists of physical jffs2_raw_node_ref structures, and
(NB) the per-inode list of physical nodes. The latter is a special
case - see below.
As the MTD API permits erase-completion callback functions to be
called from bottom-half (timer) context, and these functions access
the data structures protected by this lock, it must be locked with
Note that the per-inode list of physical nodes (f->nodes) is a special
case. Any changes to _valid_ nodes (i.e. ->flash_offset & 1 == 0) in
the list are protected by the file semaphore f->sem. But the erase
code may remove _obsolete_ nodes from the list while holding only the
erase_completion_lock. So you can walk the list only while holding the
erase_completion_lock, and can drop the lock temporarily mid-walk as
long as the pointer you're holding is to a _valid_ node, not an
The erase_completion_lock is also used to protect the c->gc_task
pointer when the garbage collection thread exits. The code to kill the
GC thread locks it, sends the signal, then unlocks it - while the GC
thread itself locks it, zeroes c->gc_task, then unlocks on the exit path.
This semaphore is only used by the erase code which frees obsolete
node references and the jffs2_garbage_collect_deletion_dirent()
function. The latter function on NAND flash must read _obsolete_ nodes
to determine whether the 'deletion dirent' under consideration can be
discarded or whether it is still required to show that an inode has
been unlinked. Because reading from the flash may sleep, the
erase_completion_lock cannot be held, so an alternative, more
heavyweight lock was required to prevent the erase code from freeing
the jffs2_raw_node_ref structures in question while the garbage
collection code is looking at them.
Suggestions for alternative solutions to this problem would be welcomed.
To unsubscribe from this list: send the line "unsubscribe jffs-dev" in
the body of a message to firstname.lastname@example.org