Mitch reported a problem where you could get an ENOSPC error when untarring
a kernel git tree onto a 16gb file system with compress-force=zlib.  This is
because compression is a huge pain, it will return from ->writepages()
without having actually created any ordered extents.  To get around this we
check to see if the async submit counter is up, and if it is wait until it
drops to 0 before doing our normal ordered wait dance.  With this patch I
can now untar a kernel git tree onto a 16gb file system without getting
ENOSPC errors.  Thanks,
Signed-off-by: NJosef Bacik <jbacik@fusionio.com>

We should insert/update 6 items(root ref, root backref, dir item, dir index,
root item and parent inode) when creating a snapshot, not 5 items, fix it.
Signed-off-by: NMiao Xie <miaox@cn.fujitsu.com>

Sometimes we need choose the method of the reservation according to the type
of the block reservation, such as the reservation for the delayed inode update.
Now we identify the type just by comparing the address of the reservation
variants, it is very ugly if it is a temporary one because we need compare it
with all the common reservation variants. So we add a new "type" field to keep
the type the reservation variants.
Signed-off-by: NMiao Xie <miaox@cn.fujitsu.com>

This patch adds hole punching via fallocate.  Thanks,
Signed-off-by: NJosef Bacik <jbacik@fusionio.com>

We will stop and restart a transaction every time we move to a different leaf
when truncating a file.  This is for enospc reasons, but really we could
probably get away with doing this a little better by actually working until we
hit an ENOSPC.  So add a ->failfast flag to the block_rsv and set it when we do
truncates which will fail as soon as the block rsv runs out of space, and then
at that point we can stop and restart the transaction and refill the block rsv
and carry on.  This will make rm'ing of a file with lots of extents a bit
faster.  Thanks,
Signed-off-by: NJosef Bacik <jbacik@fusionio.com>

Swinging this pendulum back the other way.  We've been allocating chunks up
to 2% of the disk no matter how much we actually have allocated.  So instead
fix this calculation to only allocate chunks if we have more than 80% of the
space available allocated.  Please test this as it will likely cause all
sorts of ENOSPC problems to pop up suddenly.  Thanks,
Signed-off-by: NJosef Bacik <jbacik@fusionio.com>

Daniel Blueman reported a bug with fio+balance on a ramdisk setup.
Basically what happens is the balance relocates a tree block which will drop
the implicit refs for all of its children and adds a full backref.  Once the
block is relocated we have to add the implicit refs back, so when we cow the
block again we add the implicit refs for its children back.  The problem
comes when the original drop ref doesn't get run before we add the implicit
refs back.  The delayed ref stuff will specifically prefer ADD operations
over DROP to keep us from freeing up an extent that will have references to
it, so we try to add the implicit ref before it is actually removed and we
panic.  This worked fine before because the add would have just canceled the
drop out and we would have been fine.  But the backref walking work needs to
be able to freeze the delayed ref stuff in time so we have this ever
increasing sequence number that gets attached to all new delayed ref updates
which makes us not merge refs and we run into this issue.

So to fix this we need to merge delayed refs.  So everytime we run a
clustered ref we need to try and merge all of its delayed refs.  The backref
walking stuff locks the delayed ref head before processing, so if we have it
locked we are safe to merge any refs inside of the sequence number.  If
there is no sequence number we can merge all refs.  Doing this not only
fixes our bug but keeps the delayed ref code from adding and removing
useless refs and batching together multiple refs into one search instead of
one search per delayed ref, which will really help our commit times.  I ran
this with Daniels test and 276 and I haven't seen any problems.  Thanks,
Reported-by: NDaniel J Blueman <daniel@quora.org>
Signed-off-by: NJosef Bacik <jbacik@fusionio.com>

With commit

commit d1270cd9
Author: Arne Jansen <sensille@gmx.net>
Date:   Tue Sep 13 15:16:43 2011 +0200

     Btrfs: put back delayed refs that are too new

I added a window where the delayed_ref's head->ref_mod code can diverge
from the sum of the remaining refs, because we release the head->mutex
in the middle. This leads to btrfs_lookup_extent_info returning wrong
numbers. This patch fixes this by adjusting the head's ref_mod with each
delayed ref we run.
Signed-off-by: NArne Jansen <sensille@gmx.net>
Signed-off-by: NChris Mason <chris.mason@fusionio.com>

Arne was complaining about the space cache having mismatching generation
numbers when debugging a deadlock.  This is because we can run out of space
in our preallocated range for our space cache if you have a pretty
fragmented amount of space in your pinned space.  So just increase the
amount of space we preallocate for space cache so we can be sure to have
enough space.  This will only really affect data ranges since their the only
chunks that end up larger than 256MB.  Thanks,
Signed-off-by: NJosef Bacik <jbacik@fusionio.com>
Signed-off-by: NChris Mason <chris.mason@fusionio.com>

Commit a168650c introduced a waiting mechanism to prevent busy waiting in
btrfs_run_delayed_refs. This can deadlock with btrfs_run_ordered_operations,
where a tree_mod_seq is held while waiting for the io to complete, while
the end_io calls btrfs_run_delayed_refs.
This whole mechanism is unnecessary. If not enough runnable refs are
available to satisfy count, just return as count is more like a guideline
than a strict requirement.
In case we have to run all refs, commit transaction makes sure that no
other threads are working in the transaction anymore, so we just assert
here that no refs are blocked.
Signed-off-by: NArne Jansen <sensille@gmx.net>
Signed-off-by: NChris Mason <chris.mason@fusionio.com>

We should release this mutex before returning the error code.
Signed-off-by: NDan Carpenter <dan.carpenter@oracle.com>

We were missing wakeups on the delayed ref waitqueue due
to races on waitqueue_active.
Signed-off-by: NChris Mason <chris.mason@fusionio.com>

For backref walking, we've introduce delayed ref's sequence.  However,
it changes our preallocation behavior.

The story is that when we preallocate an extent and then mark it written
piece by piece, the ideal case should be that we don't need to COW the
extent, which is why we use 'preallocate'.

But we may not make use of preallocation, since when we check for cross refs on
the extent, we may have two ref entries which have the same content except
the sequence value, and we recognize them as cross refs and do COW to allocate
another extent.

So we end up with several pieces of space instead of an whole extent.
Signed-off-by: NLiu Bo <liubo2009@cn.fujitsu.com>
Signed-off-by: NJosef Bacik <jbacik@fusionio.com>

Inodes always allocate free space with BTRFS_BLOCK_GROUP_DATA type,
which means every inode has the same BTRFS_I(inode)->free_space pointer.

This shrinks struct btrfs_inode by 4 bytes (or 8 bytes on 64 bits).
Signed-off-by: NLi Zefan <lizefan@huawei.com>

Block group has ro attributes, make dump_space_info show it.
Signed-off-by: NLiu Bo <liubo2009@cn.fujitsu.com>
Signed-off-by: NJosef Bacik <jbacik@fusionio.com>

Here is the whole story:
1)
A free space cache consists of two parts:
o  free space cache inode, which is special becase it's stored in root tree.
o  free space info, which is stored as the above inode's file data.

But we only build up another new inode and does not flush its free space info
onto disk when we _clear and setup_ free space cache, and this ends up with
that the block group cache's cache_state remains DC_SETUP instead of DC_WRITTEN.

And holding DC_SETUP means that we will not truncate this free space cache inode,
which means the disk offset of its file extent will remain _unchanged_ at least
until next transaction finishes committing itself.

2)
We can set a block group readonly when we relocate the block group.

However,
if the readonly block group covers the disk offset where our free space cache
inode is going to write, it will force the free space cache inode into
cow_file_range() and it'll end up hitting a BUG_ON.

3)
Due to the above analysis, we fix this bug by adding the missing dirty flag.

4)
However, it's not over, there is still another case, nospace_cache.

With nospace_cache, we do not want to set dirty flag, instead we just truncate
free space cache inode and bail out with setting cache state DC_WRITTEN.

We can benifit from it since it saves us another 'pre-allocation' part which
usually costs a lot.
Signed-off-by: NLiu Bo <liubo2009@cn.fujitsu.com>
Signed-off-by: NMiao Xie <miaox@cn.fujitsu.com>
Signed-off-by: NJosef Bacik <jbacik@fusionio.com>

During disk balance, we prealloc new file extent for file data relocation,
but we may fail in 'no available space' case, and it leads to flipping btrfs
into readonly.

It is not necessary to bail out and abort transaction since we do have several
ways to rescue ourselves from ENOSPC case.
Signed-off-by: NLiu Bo <liubo2009@cn.fujitsu.com>
Signed-off-by: NJosef Bacik <jbacik@fusionio.com>

Since root can be fetched via BTRFS_I macro directly, we can save an args
for btrfs_is_free_space_inode().
Signed-off-by: NLiu Bo <liubo2009@cn.fujitsu.com>
Signed-off-by: NJosef Bacik <jbacik@fusionio.com>

So shrink_delalloc has grown all sorts of cruft over the years thanks to
many reworkings of how we track enospc.  What happens now as we fill up the
disk is we will loop for freaking ever hoping to reclaim a arbitrary amount
of space of metadata, this was from when everybody flushed at the same time.
Now we only have people flushing one at a time.  So instead of trying to
reclaim a huge amount of space, just try to flush a decent chunk of space,
and stop looping as soon as we have enough free space to satisfy our
reservation.  This makes xfstests 224 go much faster.  Thanks,
Signed-off-by: NJosef Bacik <jbacik@fusionio.com>

There is weird logic I had to put in place to make sure that when we were
adding csums that we'd used the delalloc block rsv instead of the global
block rsv.  Part of this meant that we had to free up our transaction
reservation before we ran the delayed refs since csum deletion happens
during the delayed ref work.  The problem with this is that when we release
a reservation we will add it to the global reserve if it is not full in
order to keep us going along longer before we have to force a transaction
commit.  By releasing our reservation before we run delayed refs we don't
get the opportunity to drain down the global reserve for the work we did, so
we won't refill it as often.  This isn't a problem per-se, it just results
in us possibly committing transactions more and more often, and in rare
cases could cause those WARN_ON()'s to pop in use_block_rsv because we ran
out of space in our block rsv.

This also helps us by holding onto space while the delayed refs run so we
don't end up with as many people trying to do things at the same time, which
again will help us not force commits or hit the use_block_rsv warnings.
Thanks,
Signed-off-by: NJosef Bacik <jbacik@fusionio.com>

Those crazy gentoo guys have been complaining about ENOSPC errors on their
portage volumes.  This is because doing things like untar tends to create
lots of new files which will soak up all the reservation space in the
delayed inodes.  Usually this gets papered over by the fact that we will try
and commit the transaction, however if this happens in the wrong spot or we
choose not to commit the transaction you will be screwed.  So add the
ability to expclitly flush delayed inodes to free up space.  Please test
this out guys to make sure it works since as usual I cannot reproduce.
Thanks,
Signed-off-by: NJosef Bacik <jbacik@fusionio.com>

Like block reserves, reserve a small piece of space on each
transaction start and for delalloc. These are the hooks that
can actually return EDQUOT to the user.
The amount of space reserved is tracked in the transaction
handle.
Signed-off-by: NArne Jansen <sensille@gmx.net>

Signed-off-by: NJan Schmidt <list.btrfs@jan-o-sch.net>

Signed-off-by: NArne Jansen <sensille@gmx.net>
Signed-off-by: NJan Schmidt <list.btrfs@jan-o-sch.net>

Normally delayed refs get processed in ascending bytenr order. This
correlates in most cases to the order added. To expose dependencies
on this order, we start to process the tree in the middle instead of
the beginning.
This code is only effective when SCRAMBLE_DELAYED_REFS is defined.
Signed-off-by: NArne Jansen <sensille@gmx.net>

We've got two mechanisms both required for reliable backref resolving (tree
mod log and holding back delayed refs). You cannot make use of one without
the other. So instead of requiring the user of this mechanism to setup both
correctly, we join them into a single interface.

Additionally, we stop inserting non-blockers into fs_info->tree_mod_seq_list
as we did before, which was of no value.
Signed-off-by: NJan Schmidt <list.btrfs@jan-o-sch.net>

We track two conditions to decide if we should sleep while waiting for more
delayed refs, the number of delayed refs (num_refs) and the first entry in
the list of blockers (first_seq).

When we suspect staleness, we save num_refs and do one more cycle. If
nothing changes, we then save first_seq for later comparison and do
wait_event. We ought to save first_seq the very same moment we're saving
num_refs. Otherwise we cannot be sure that nothing has changed and we might
start waiting when we shouldn't, which could lead to starvation.
Signed-off-by: NJan Schmidt <list.btrfs@jan-o-sch.net>

Miao pointed this out while I was working on an orphan problem that messing
with a bitfield where different ranges are protected by different locks
doesn't work out right.  Turns out we've been doing this forever where we
have different parts of the bit field protected by either no lock at all or
different locks which could cause all sorts of weird problems including the
issue I was hitting.  So instead make a runtime_flags thing that we use the
normal bit operations on that are all atomic so we can keep having our
no/different locking for the different flags and then make force_compress
it's own thing so it can be treated normally.  Thanks,
Signed-off-by: NJosef Bacik <josef@redhat.com>

Three callers of btrfs_free_tree_block or btrfs_alloc_tree_block passed
parameter for_cow = 1. In fact, these two functions should never mark
their tree modification operations as for_cow, because they can change
the number of blocks referenced by a tree.

Hence, we remove the extra for_cow parameter from these functions and
make them pass a zero down.
Signed-off-by: NJan Schmidt <list.btrfs@jan-o-sch.net>

It confuses Smatch that we use two names for the same lock.  Plus the
shorter name is nicer.  This doesn't change how the code works, it's
just a cleanup.
Signed-off-by: NDan Carpenter <dan.carpenter@oracle.com>

verify_parent_transid needs to lock the extent range to make
sure no IO is underway, and so it can safely clear the
uptodate bits if our checks fail.

But, a few callers are using it with spinlocks held.  Most
of the time, the generation numbers are going to match, and
we don't want to switch to a blocking lock just for the error
case.  This adds an atomic flag to verify_parent_transid,
and changes it to return EAGAIN if it needs to block to
properly verifiy things.
Signed-off-by: NChris Mason <chris.mason@oracle.com>

may_commit_transaction() calls
        spin_lock(&space_info->lock);
        spin_lock(&delayed_rsv->lock);
and update_global_block_rsv() calls
        spin_lock(&block_rsv->lock);
        spin_lock(&sinfo->lock);

Lockdep complains about this at run time.
Everywhere except in update_global_block_rsv(), the space_info lock is
the outer lock, therefore the locking order in update_global_block_rsv()
is changed.
Signed-off-by: NStefan Behrens <sbehrens@giantdisaster.de>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

It is basically a good thing if we are interruptible when waiting for
free space, but the generality in which it is implemented currently
leads to system calls being interruptible that are not documented this
way. For example git can't handle interrupted unlink(), leading to
corrupt repos under space pressure.
Instead we raise the bar to only be interruptible by SIGKILL.
Thanks to David Sterba for suggesting this.
Signed-off-by: NArne Jansen <sensille@gmx.net>

The caller expects this function to return with the lock held and
releases it immediately on error.
Signed-off-by: NDan Carpenter <dan.carpenter@oracle.com>

A user reported that booting his box up with btrfs root on 3.4 was way
slower than on 3.3 because I removed the ideal caching code.  It turns out
that we don't load the free space cache if we're in a commit for deadlock
reasons, but since we're reading the cache and it hasn't changed yet we are
safe reading the inode and free space item from the commit root, so do that
and remove all of the deadlock checks so we don't unnecessarily skip loading
the free space cache.  The user reported this fixed the slowness.  Thanks,
Tested-by: NCalvin Walton <calvin.walton@kepstin.ca>
Signed-off-by: NJosef Bacik <josef@redhat.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

This fixes a regression introduced by fc67c450.  spin_is_locked() always
returns 0 on UP kernels, which caused assert in get_restripe_target() to
be fired on every call from btrfs_reduce_alloc_profile() on UP systems.
Remove it completely for now, it's not clear if it's going to be needed
in future.
Reported-by: NBobby Powers <bobbypowers@gmail.com>
Reported-by: NMitch Harder <mitch.harder@sabayonlinux.org>
Tested-by: NMitch Harder <mitch.harder@sabayonlinux.org>
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

This reverts commit 5500cdbe.

We've had a number of complaints of early enospc that bisect down
to this patch.  We'll hae to fix the reservations differently.

CC: stable@kernel.org
Signed-off-by: NChris Mason <chris.mason@oracle.com>

This deadlock comes from xfstests 251.

We'll hold the chunk_mutex throughout the whole of a chunk allocation.
But if we find that we've used up system chunk space, we need to allocate a
new system chunk, but this will lead to a recursion of chunk allocation and end
up with a deadlock on chunk_mutex.
So instead we need to allocate the system chunk first if we find we're in ENOSPC.
Signed-off-by: NLiu Bo <liubo2009@cn.fujitsu.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

o For space info, the type of space info is useful for debug.
o For transaction handle, its transid is useful.
Signed-off-by: NLiu Bo <liubo2009@cn.fujitsu.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

Currently if we don't have enough space allocated we go ahead and loop
though devices in the hopes of finding enough space for a chunk of the
*same* type as the one we are trying to relocate.  The problem with that
is that if we are trying to restripe the chunk its target type can be
more relaxed than the current one (eg require less devices or less
space).  So, when restriping, run checks against the target profile
instead of the current one.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>