Btrfs: fix quick exhaustion of the system array in the superblock

Omar reported that after commit 4fbcdf66 ("Btrfs: fix -ENOSPC when
finishing block group creation"), introduced in 4.2-rc1, the following
test was failing due to exhaustion of the system array in the superblock:

  #!/bin/bash

  truncate -s 100T big.img
  mkfs.btrfs big.img
  mount -o loop big.img /mnt/loop

  num=5
  sz=10T
  for ((i = 0; i < $num; i++)); do
      echo fallocate $i $sz
      fallocate -l $sz /mnt/loop/testfile$i
  done
  btrfs filesystem sync /mnt/loop

  for ((i = 0; i < $num; i++)); do
        echo rm $i
        rm /mnt/loop/testfile$i
        btrfs filesystem sync /mnt/loop
  done
  umount /mnt/loop

This made btrfs_add_system_chunk() fail with -EFBIG due to excessive
allocation of system block groups. This happened because the test creates
a large number of data block groups per transaction and when committing
the transaction we start the writeout of the block group caches for all
the new new (dirty) block groups, which results in pre-allocating space
for each block group's free space cache using the same transaction handle.
That in turn often leads to creation of more block groups, and all get
attached to the new_bgs list of the same transaction handle to the point
of getting a list with over 1500 elements, and creation of new block groups
leads to the need of reserving space in the chunk block reserve and often
creating a new system block group too.

So that made us quickly exhaust the chunk block reserve/system space info,
because as of the commit mentioned before, we do reserve space for each
new block group in the chunk block reserve, unlike before where we would
not and would at most allocate one new system block group and therefore
would only ensure that there was enough space in the system space info to
allocate 1 new block group even if we ended up allocating thousands of
new block groups using the same transaction handle. That worked most of
the time because the computed required space at check_system_chunk() is
very pessimistic (assumes a chunk tree height of BTRFS_MAX_LEVEL/8 and
that all nodes/leafs in a path will be COWed and split) and since the
updates to the chunk tree all happen at btrfs_create_pending_block_groups
it is unlikely that a path needs to be COWed more than once (unless
writepages() for the btree inode is called by mm in between) and that
compensated for the need of creating any new nodes/leads in the chunk
tree.

So fix this by ensuring we don't accumulate a too large list of new block
groups in a transaction's handles new_bgs list, inserting/updating the
chunk tree for all accumulated new block groups and releasing the unused
space from the chunk block reserve whenever the list becomes sufficiently
large. This is a generic solution even though the problem currently can
only happen when starting the writeout of the free space caches for all
dirty block groups (btrfs_start_dirty_block_groups()).

Reported-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Tested-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>