block allocator needs to try harder

[yomimono]

With the following target:

(rule
  (alias big)
  (deps colossal)
  (locks colossal)
  (action (progn
            (run dd if=/dev/zero of=for_colossal.img bs=512 count=1000)
            (setenv OCAMLRUNPARAM b (run %{exe:chamelon.exe} format for_colossal.img 512))
            (setenv OCAMLRUNPARAM b (with-stdin-from colossal (run %{exe:chamelon.exe} write --verbosity=debug for_colossal.img 512 "colossal.bak" -)))
            (setenv OCAMLRUNPARAM b (with-stdin-from colossal (run %{exe:chamelon.exe} write --verbosity=debug for_colossal.img 512 "colossal.bak.bak" -)))

            (with-stdout-to colossal-ls.expected (run echo "colossal.bak : file"))
            (setenv OCAMLRUNPARAM b (with-stdout-to colossal-ls.output (run %{exe:chamelon.exe} ls for_colossal.img 512 "/")))
            (diff colossal-ls.expected colossal-ls.output)

            (setenv OCAMLRUNPARAM b (with-stdout-to colossal.bak (run %{exe:chamelon.exe} read for_colossal.img 512 "colossal.bak")))
            (run cmp colossal colossal.bak)
            (run %{exe:chamelon.exe} rm for_colossal.img 512 "/colossal.bak")

The write of colossal.bak.bak fails. colossal is the output of dd if=/dev/urandom of=colossal bs=128K count=1, so we would expect to be able to write this at least twice (2 * 128K = 256K) in our 512K filesystem, even with skip-list overhead. I suspect the allocator is finding an entire subsection full and then giving up.