tools/upmap: upmap-remapped.py correctness and error handling fixes

[dvanders]

A review of upmap-remapped.py turned up three ways it can fail to do its job,
plus a handful of smaller issues. One fix per commit, so they can be taken or
dropped individually.

Bugs

• The script can hang forever. The bubble sort which orders the mappings on
  erasure-coded pools never terminates when the mappings form a cycle, e.g. up
  [4,5,3] against acting [3,4,5]. Because stdout is block buffered when piped
  into sh, the commands generated before the hang never run either, so the run
  does nothing at all.
• It emits mappings the mon silently ignores. Mappings are dropped one at a
  time when an osd is out or gone, which leaves the remaining ones free to map
  onto an osd that is already in the up set. The mon ignores such a mapping, so
  the pg stays remapped no matter how many times the script is run. With osd.7
  out, up [1,2] and acting [2,7] produced pg-upmap-items 1 2, a no-op.
• Reverse pairs were only partly removed. The loop removing pairs such as
  (314, 272) & (272, 314) mutated the list it was iterating over, so pairs
  after the first were left in.
• A pool snapshot with pool in its name is parsed as a pool definition and
  overwrites that pool's type, after which the script exits with
  Unknown pool type having done nothing.
• A pool deleted between reading the pgs and reading the pools ended the run with
  a KeyError.
• Mon command return codes were never checked, so a refused osd pool ls detail
  left the pool types empty and the script died with a KeyError on the first pg.
• The 5s mon command timeout is too short for pg ls remapped on a large cluster
  with many remapped pgs, and the resulting librados error was not caught.
• Any argument other than exactly --ignore-backfilling was ignored, so a typo
  quietly did a full production run.

Also

• jq is no longer required; the jq -r . pipes were no-ops that also masked
  failing ceph commands, since check=True only saw jq's exit status.
• The osd weights are indexed instead of searched: building the mappings for
  30000 remapped 8+3 pgs on a 2000 osd cluster goes from 1.94s to 0.03s.
• The four if use_shell / else mon_command blocks became one helper, bare
  except: clauses were narrowed, and the librados connection is now closed on
  every exit path rather than only the last one.

Testing

gen_upmap() was checked against a simulation of the mon's pg-upmap-items
handling over 120000 random up/acting combinations, including out osds, unknown
osds and missing erasure-coded shards. The old code produced mappings the mon
would ignore in ~28000 of them and hung in ~3800; the new code has none of
either, and fixes slightly more shards because a pg containing a cycle now still
gets its other mappings. Both the shell and librados paths were run end to end
against a stub cluster.

Worth knowing

• Unknown command line options are now an error instead of being ignored.
• The mon command timeout goes from 5s to 300s.
• A pg whose pool has an unrecognised type still exits the run, as before; a pg
  whose pool is missing is now skipped with a message, since that is a benign
  race.

[dvanders]

@bstillwell ^