Josef,

Have you actually seen this problem on 3.5 with no compression?

youre right, no problems with no compression.

problem in the 3.5 timeframe that should have fixed this for no-compression, and
then I've since fixed the compression part of it in btrfs-next.  Can you retest

i cherry picked on top of 3.5.4 + for-linus all enospc patches
and dependencies from your btrfs-next in that order:

ddabdd663aa819f206aaaf689b12475b1b11d71e
Btrfs: do not allocate chunks as agressively

dd314c3ef057dde30d1c8ae67c98019a71713ea4
Btrfs: turbo charge fsync

4d9f18a35d456c27e81c8e0de29c858f85d35c91
Btrfs: do not needlessly restart the transaction for enospc

82b8aa2bef5c5267d74ad8e5a82572d4fa85bf89
Btrfs: wait on async pages when shrinking delalloc

4af89eb8ce93a1ec64fc698f82f67a0e3bd8cbef
Btrfs: delay block group item insertion

i hope those are the patches you wanted me to test.

with btrfs-next and see if you still see the problem?

did and got some nice results: no problems and its faster!

these results are from an old i686 Pentium M Notebook
with 30GB HDD and 1.2GB RAM:

3.5.4 + for-linus + lzo: 6x enospc problems at 30% unsquashfs
*snip*
[================|                                         ] 269645/910412  29%
Write on output file failed because No space left on device
writer: failed to write data block 0
Failed to write /mnt/point/something, skipping
*snip*

# df -h /mnt/point/
Filesystem      Size  Used Avail Use% Mounted on
/dev/sda2        28G   12G   16G  43% /mnt/point

one run:
real    28m44.882s
user    7m4.542s
sys     4m18.353s

3.5.4 + for-linus + no compression: no problems
# df -h /mnt/point/
Filesystem      Size  Used Avail Use% Mounted on
/dev/sda2        28G   18G  9.0G  67% /mnt/point

one run:
real    20m57.157s
user    5m37.915s
sys     3m23.913s

3.5.4 + for-linus + enospc patches + lzo: no problems
# df -h /mnt/point/
Filesystem      Size  Used Avail Use% Mounted on
/dev/sda2        28G   12G   17G  42% /mnt/point

first run:
real    19m35.852s
user    4m33.359s
sys     3m2.808s

second run:
real    23m12.371s
user    5m37.398s
sys     3m47.525s

3.5.4 + for-linus + enospc patches + no compression: no problems
# df -h /mnt/point/
Filesystem      Size  Used Avail Use% Mounted on
/dev/sda2        28G   18G  9.5G  66% /mnt/point

one run:
real    22m56.940s
user    4m53.571s
sys     3m4.188s

great work.
im gonna keep those patches and deploy it on my systems here soon
after some more testing.

Ahmet