ext3's fully data-journalled mode has been broken for a year.  This patch
fixes it up.

The prepare_write/commit_write/writepage implementations have been split up.
Instead of having each function handle all three journalling mode we now have
three separate sets of address_space_operations.

The problematic part of data=journal is MAP_SHARED writepage traffic: pages
which don't have buffers.  In 2.4 these were cheatingly treated as
data-ordered buffers and that caused several nasty problems.

Here we do it properly: writepage traffic is fully journalled.  This means
that the various workarounds for the 2.4 scheme can be removed, when I
remember where they all are.

The PG_checked flag has been borrowed: it it set in the atomic set_page_dirty
a_op to tell the subsequent writepage() that this page needs to have buffers
attached, dirtied and journalled.

This rather defines PG_checked as "fs-private info in page->flags" and it
should be renamed sometime.

Avoid holding the journal's j_list_lock while copying the buffer_head's data.
We hold jbd_lock_bh_state() during the copy, which is all that is needed.

In start_this_handle() the caller does not have a handle ref pinning the
transaction open, and so the call to log_start_commit() is racy because some
other CPU could take the transaction into commit state independently.

Fix that by holding j_state_lock (which pins j_running_transaction) across
the log_start_commit() call.

Plug a conceivable race with the freeing up of trasnactions, and add some
more debug checks.

Drop in a few assertions to ensure that the locking rules are being adhered
to.

Add a comment describing why a race isn't there.

After ext3_writepage() has called block_write_full_page() it will walk the
page's buffer ring dropping the buffer_head refcounts.

It does this wrong - on the final loop it will dereference the buffer_head
which it just dropped the refcount on.  Poisoned oopses have been seen
against bh->b_this_page.

Change it to take a local copy of b_this_page prior to dropping the bh's
refcount.

We need to check that buffer is still journalled _after_ taking the right
locks.

There's a bug: a caller tries to journal a buffer and then decides he didn't
want to after all.  He calls journal_release_buffer().

But journal_release_buffer() is only allowed to give the caller a buffer
credit back if it was the caller who added the buffer in the first place.

journal_release_buffer() currently looks at the buffer state to work that
out, but gets it wrong: if the buffer has been moved onto a different list by
some other part of ext3 the credit is bogusly not returned to the caller and
the fs can later go BUG due to handle credit exhaustion.


The fix:

Change journal_get_undo_access() to return the number of buffers which the
caller actually added to the journal.  (one or zero).

When the caller later calls journal_release_buffer(), he passes in that
count, to tell journal_release_buffer() how many credits the caller should
get back.

For API consistency this change should also be made to
journal_get_create_access() and journal_get_write_access().  But there is no
requirement for that in ext3 at this time.


The remaining bug:

This logic effectively gives another transaction handle a free buffer credit.
These could conceivably accumulate and cause a journal overflow.  This is a
separate problem and needs changes to the t_outstanding_credits accounting
and the logic in start_this_handle.

This filesystem-wide sleeping lock is no longer needed.  Remove it.

lock_kernel() is no longer needed in JBD.  Remove all the lock_kernel() calls
from fs/jbd/.

Here is where I get to say "ex-parrot".

Remove the remaining sleep_on() calls from JBD.

From: Alex Tomas <bzzz@tmi.comex.ru>

We're about to remove lock_journal(), and it is lock_journal which separates
the running and committing transaction's revokes on the single revoke table.

So implement two revoke tables and rotate them at commit time.

Impement the designed locking around journal->j_commit_request.

Implement the designed locking around journal->j_commit_sequence.

Implement the designed locking around journal->j_free.

Things get a lot better here, too.

Implement the designed locking around journal->j_tail.

Implement the designed locking around journal->j_head.

Implement the designed locking around j_checkpoint_transactions.  It was all
pretty much there actually.

Go through all sites which use j_committing_transaction and ensure that the
deisgned locking is correctly implemented there.

Implement the designed locking around journal->j_running_transaction.

A lot more of the new locking scheme falls into place.

We now start to move onto the fields of the topmost JBD data structure: the
journal.

The patch implements the designed locking around the j_barrier_count member.
And as a part of that, a lot of the new locking scheme is implemented.
Several lock_kernel()s and sleep_on()s go away.

Provide the designed locking around the transaction's t_jcb callback list.

It turns out that this is wholly redundant at present.

Implement the designed locking for t_outstanding_credits

Provide the designating locking for transaction_t.t_updates.

Now we move more into the locking of the transaction_t fields.

t_nr_buffers locking is just an audit-and-commentary job.

This was a system-wide spinlock.

Simple transformation: make it a filesystem-wide spinlock, in the JBD
journal.

That's a bit lame, and later it might be nice to make it per-transaction_t.
But there are interesting ranking and ordering problems with that, especially
around __journal_refile_buffer().

Implement the designated b_tnext locking.

This also covers b_tprev locking.

Go through all b_next_transaction instances, implement locking rules.
(Nothing to do here - b_transaction locking covered it)

Go through all use of b_transaction and implement the rules.

Fairly straightforward.

Implement the designed locking schema around the
journal_head.b_committed_data field.

We now start to move across the JBD data structure's fields, from "innermost"
and outwards.

Start with journal_head.b_frozen_data, because the locking for this field was
partially implemented in jbd-010-b_committed_data-race-fix.patch.

It is protected by jbd_lock_bh_state().  We keep the lock_journal() and
spin_lock(&journal_datalist_lock) calls in place.  Later,
spin_lock(&journal_datalist_lock) is replaced by
spin_lock(&journal->j_list_lock).

Of course, this completion of the locking around b_frozen_data also puts a
lot of the locking for other fields in place.

journal_unlock_journal_head() is misnamed: what it does is to drop a ref on
the journal_head and free it if that ref fell to zero.  It doesn't actually
unlock anything.

Rename it to journal_put_journal_head().

buffer_heads and journal_heads are joined at the hip.  We need a lock to
protect the joint and its refcounts.

JBD is currently using a global spinlock for that.  Change it to use one bit
in bh->b_state.

This was a strange spinlock which was designed to prevent another CPU from
ripping a buffer's journal_head away while this CPU was inspecting its state.

Really, we don't need it - we can inspect that state directly from bh->b_state.

So kill it off, along with a few things which used it which are themselves
not actually used any more.

This is the start of the JBD locking rework.

The aims of all this are to remove all lock_kernel() calls from JBD, to
remove all lock_journal() calls (the context switch rate is astonishing when
the lock_kernel()s are removed) and to remove all sleep_on() instances.




The strategy which is taken is:

a) Define the lcoking schema (this patch)

b) Work through every JBD data structure and implement its locking fully,
   according to the above schema.  We work from "innermost" data structures
   and outwards.

It isn't guaranteed that the filesystem will work very well at all stages of
this patch series.



In this patch:


Add commentary and various locks to jbd.h describing the locking scheme which
is about to be implemented.

Initialise the new locks.

Coding-style goodness in jbd.h

From: Alex Tomas <bzzz@tmi.comex.ru>

We have a race wherein the block allocator can decide that
journal_head.b_committed_data is present and then will use it.  But kjournald
can concurrently free it and set the pointer to NULL.  It goes oops.

We introduce per-buffer_head "spinlocking" based on a bit in b_state.  To do
this we abstract out pte_chain_lock() and reuse the implementation.

The bit-based spinlocking is pretty inefficient CPU-wise (hence the warning
in there) and we may move this to a hashed spinlock later.

From: Alex Tomas <bzzz@tmi.comex.ru>

This is a port from ext2 of the fuzzy counters (for Orlov allocator
heuristics) and the hashed spinlocking (for the inode and bloock allocators).

From: Alex Tomas <bzzz@tmi.comex.ru>


This patch weans ext3 off lock_super()-based protection for the inode and
block allocators.

It's basically the same as the ext2 changes.


1) each group has own spinlock, which is used for group counter
   modifications

2) sb->s_free_blocks_count isn't used any more.  ext2_statfs() and
   find_group_orlov() loop over groups to count free blocks

3) sb->s_free_blocks_count is recalculated at mount/umount/sync_super time
   in order to check consistency and to avoid fsck warnings

4) reserved blocks are distributed over last groups

5) ext3_new_block() tries to use non-reserved blocks and if it fails then
   tries to use reserved blocks

6) ext3_new_block() and ext3_free_blocks do not modify sb->s_free_blocks,
   therefore they do not call mark_buffer_dirty() for superblock's
   buffer_head. this should reduce I/O a bit


Also fix orlov allocator boundary case:

In the interests of SMP scalability the ext2 free blocks and free inodes
counters are "approximate".  But there is a piece of code in the Orlov
allocator which fails due to boundary conditions on really small
filesystems.

Fix that up via a final allocation pass which simply uses first-fit for
allocatiopn of a directory inode.

Move some lock_kernel() calls from the caller to the callee, reducing
holdtimes.