ceph/tasks/scrub_test.py

"""Scrub testing"""
from cStringIO import StringIO

import contextlib
import json
import logging
import os
import time
import ctypes
import tempfile

import ceph_manager
from teuthology import misc as teuthology

log = logging.getLogger(__name__)


def wait_for_victim_pg(manager):
    """Return a PG with some data and its acting set"""
    # wait for some PG to have data that we can mess with
    victim = None
    while victim is None:
        stats = manager.get_pg_stats()
        for pg in stats:
            size = pg['stat_sum']['num_bytes']
            if size > 0:
                victim = pg['pgid']
                acting = pg['acting']
                return victim, acting
        time.sleep(3)


def find_victim_object(ctx, pg, osd):
    """Return a file to be fuzzed"""
    (osd_remote,) = ctx.cluster.only('osd.%d' % osd).remotes.iterkeys()
    data_path = os.path.join(
        '/var/lib/ceph/osd',
        'ceph-{id}'.format(id=osd),
        'current',
        '{pg}_head'.format(pg=pg)
        )

    # fuzz time
    with contextlib.closing(StringIO()) as ls_fp:
        osd_remote.run(
            args=['sudo', 'ls', data_path],
            stdout=ls_fp,
        )
        ls_out = ls_fp.getvalue()

    # find an object file we can mess with
    osdfilename = next(line for line in ls_out.split('\n')
                       if not line.startswith('__'))
    assert osdfilename is not None

    # Get actual object name from osd stored filename
    objname, _ = osdfilename.split('__', 1)
    objname = objname.replace(r'\u', '_')
    return osd_remote, os.path.join(data_path, osdfilename), objname


def corrupt_file(osd_remote, path):
    # put a single \0 at the beginning of the file
    osd_remote.run(
        args=['sudo', 'dd',
              'if=/dev/zero',
              'of=%s' % path,
              'bs=1', 'count=1', 'conv=notrunc']
    )


def deep_scrub(manager, victim):
    # scrub, verify inconsistent
    manager.raw_cluster_cmd('pg', 'deep-scrub', victim)
    # Give deep-scrub a chance to start
    time.sleep(60)

    while True:
        stats = manager.get_single_pg_stats(victim)
        state = stats['state']

        # wait for the scrub to finish
        if 'scrubbing' in state:
            time.sleep(3)
            continue

        inconsistent = stats['state'].find('+inconsistent') != -1
        assert inconsistent
        break


def repair(manager, victim):
    # repair, verify no longer inconsistent
    manager.raw_cluster_cmd('pg', 'repair', victim)
    # Give repair a chance to start
    time.sleep(60)

    while True:
        stats = manager.get_single_pg_stats(victim)
        state = stats['state']

        # wait for the scrub to finish
        if 'scrubbing' in state:
            time.sleep(3)
            continue

        inconsistent = stats['state'].find('+inconsistent') != -1
        assert not inconsistent
        break


def test_repair_corrupted_obj(ctx, manager, pg, osd_remote, obj_path):
    corrupt_file(osd_remote, obj_path)
    deep_scrub(manager, pg)
    repair(manager, pg)


def test_repair_bad_omap(ctx, manager, pg, osd, objname):
    # Test deep-scrub with various omap modifications
    # Modify omap on specific osd
    log.info('fuzzing omap of %s' % objname)
    manager.osd_admin_socket(osd, ['rmomapkey', 'rbd', objname, 'key'])
    manager.osd_admin_socket(osd, ['setomapval', 'rbd', objname,
                                   'badkey', 'badval'])
    manager.osd_admin_socket(osd, ['setomapheader', 'rbd', objname, 'badhdr'])

    deep_scrub(manager, pg)
    # please note, the repair here is errnomous, it rewrites the correct omap
    # digest and data digest on the replicas with the corresponding digests
    # from the primary osd which is hosting the victim object, see
    # find_victim_object().
    # so we need to either put this test and the end of this task or
    # undo the mess-up manually before the "repair()" that just ensures
    # the cleanup is sane, otherwise the succeeding tests will fail. if they
    # try set "badkey" in hope to get an "inconsistent" pg with a deep-scrub.
    manager.osd_admin_socket(osd, ['setomapheader', 'rbd', objname, 'hdr'])
    manager.osd_admin_socket(osd, ['rmomapkey', 'rbd', objname, 'badkey'])
    manager.osd_admin_socket(osd, ['setomapval', 'rbd', objname,
                                   'key', 'val'])
    repair(manager, pg)


class MessUp:
    def __init__(self, manager, osd_remote, pool, osd_id,
                 obj_name, obj_path, omap_key, omap_val):
        self.manager = manager
        self.osd = osd_remote
        self.pool = pool
        self.osd_id = osd_id
        self.obj = obj_name
        self.path = obj_path
        self.omap_key = omap_key
        self.omap_val = omap_val

    @contextlib.contextmanager
    def _test_with_file(self, messup_cmd, *checks):
        temp = tempfile.mktemp()
        backup_cmd = ['sudo', 'cp', self.path, temp]
        self.osd.run(args=backup_cmd)
        self.osd.run(args=messup_cmd.split())
        yield checks
        restore_cmd = ['sudo', 'mv', temp, self.path]
        self.osd.run(args=restore_cmd)

    def remove(self):
        cmd = 'sudo rm {path}'.format(path=self.path)
        return self._test_with_file(cmd, 'missing')

    def append(self):
        cmd = 'sudo dd if=/dev/zero of={path} bs=1 count=1 ' \
              'conv=notrunc oflag=append'.format(path=self.path)
        return self._test_with_file(cmd,
                                    'data_digest_mismatch',
                                    'size_mismatch')

    def truncate(self):
        cmd = 'sudo dd if=/dev/null of={path}'.format(path=self.path)
        return self._test_with_file(cmd,
                                    'data_digest_mismatch',
                                    'size_mismatch')

    def change_obj(self):
        cmd = 'sudo dd if=/dev/zero of={path} bs=1 count=1 ' \
              'conv=notrunc'.format(path=self.path)
        return self._test_with_file(cmd,
                                    'data_digest_mismatch')

    @contextlib.contextmanager
    def rm_omap(self):
        cmd = ['rmomapkey', self.pool, self.obj, self.omap_key]
        self.manager.osd_admin_socket(self.osd_id, cmd)
        yield ('omap_digest_mismatch',)
        cmd = ['setomapval', self.pool, self.obj,
               self.omap_key, self.omap_val]
        self.manager.osd_admin_socket(self.osd_id, cmd)

    @contextlib.contextmanager
    def add_omap(self):
        cmd = ['setomapval', self.pool, self.obj, 'badkey', 'badval']
        self.manager.osd_admin_socket(self.osd_id, cmd)
        yield ('omap_digest_mismatch',)
        cmd = ['rmomapkey', self.pool, self.obj, 'badkey']
        self.manager.osd_admin_socket(self.osd_id, cmd)

    @contextlib.contextmanager
    def change_omap(self):
        cmd = ['setomapval', self.pool, self.obj, self.omap_key, 'badval']
        self.manager.osd_admin_socket(self.osd_id, cmd)
        yield ('omap_digest_mismatch',)
        cmd = ['setomapval', self.pool, self.obj, self.omap_key, self.omap_val]
        self.manager.osd_admin_socket(self.osd_id, cmd)


class InconsistentObjChecker:
    """Check the returned inconsistents/inconsistent info"""

    CEPH_NOSNAP = ctypes.c_uint64(-2).value

    def __init__(self, osd, acting, obj_name):
        self.osd = osd
        self.acting = acting
        self.obj = obj_name
        assert self.osd in self.acting

    def basic_checks(self, inc):
        assert inc['object']['name'] == self.obj
        assert inc['object']['snap'] == self.CEPH_NOSNAP
        assert len(inc['shards']) == len(self.acting), \
            "the number of returned shard does not match with the acting set"

    def run(self, check, inc):
        func = getattr(self, check)
        func(inc)

    def _get_attrs(self, inc, attr_name):
        bad_attr = None
        good_attr = None
        for shard in inc['shards']:
            log.info('shard = %r' % shard)
            log.info('attr = %s' % attr_name)
            assert 'osd' in shard
            assert attr_name in shard
            osd = shard['osd']
            attr = shard[attr_name]
            if osd == self.osd:
                assert bad_attr is None, \
                    "multiple entries found for the given OSD"
                bad_attr = attr
            else:
                assert osd in self.acting, "shard not in acting set"
                assert good_attr is None or good_attr == attr, \
                    "multiple good attrs found"
                good_attr = attr
        assert bad_attr is not None, \
            "good {attr} not found".format(attr=attr_name)
        assert good_attr is not None, \
            "bad {attr} not found".format(attr=attr_name)
        assert good_attr != bad_attr, \
            "bad attr is identical to the good ones: " \
            "{0} == {1}".format(good_attr, bad_attr)
        return bad_attr, good_attr

    def data_digest_mismatch(self, inc):
        assert inc['data_digest_mismatch'] is True
        self._get_attrs(inc, 'data_digest')

    def missing(self, inc):
        assert inc['missing'] is True
        has_missing, _ = self._get_attrs(inc, 'missing')
        assert has_missing is True, "the removed shard is not missing"

    def size_mismatch(self, inc):
        assert inc['size_mismatch'] is True
        self._get_attrs(inc, 'size')

    def omap_digest_mismatch(self, inc):
        assert inc['omap_digest_mismatch'] is True
        self._get_attrs(inc, 'omap_digest')


def test_list_inconsistent_obj(ctx, manager, osd_remote, pg, acting, osd_id,
                               obj_name, obj_path):
    mon = manager.controller
    pool = 'rbd'
    omap_key = 'key'
    omap_val = 'val'
    manager.do_rados(mon, ['-p', pool, 'setomapval', obj_name,
                           omap_key, omap_val])
    messup = MessUp(manager, osd_remote, pool, osd_id, obj_name, obj_path,
                    omap_key, omap_val)
    for test in [messup.rm_omap, messup.add_omap, messup.change_omap,
                 messup.append, messup.truncate, messup.change_obj,
                 messup.remove]:
        with test() as checks:
            deep_scrub(manager, pg)
            cmd = 'rados list-inconsistent-pg {pool} ' \
                  '--format=json'.format(pool=pool)
            with contextlib.closing(StringIO()) as out:
                mon.run(args=cmd.split(), stdout=out)
                pgs = json.loads(out.getvalue())
            assert pgs == [pg]

            cmd = 'rados list-inconsistent-obj {pg} ' \
                  '--format=json'.format(pg=pg)
            with contextlib.closing(StringIO()) as out:
                mon.run(args=cmd.split(), stdout=out)
                objs = json.loads(out.getvalue())
            assert len(objs) == 1

            checker = InconsistentObjChecker(osd_id, acting, obj_name)
            inc_obj = objs[0]
            checker.basic_checks(inc_obj)
            for check in checks:
                log.info('inc = %r', inc_obj)
                checker.run(check, inc_obj)


def task(ctx, config):
    """
    Test [deep] scrub

    tasks:
    - chef:
    - install:
    - ceph:
        log-whitelist:
        - '!= known digest'
        - '!= known omap_digest'
        - deep-scrub 0 missing, 1 inconsistent objects
        - deep-scrub 1 errors
        - repair 0 missing, 1 inconsistent objects
        - repair 1 errors, 1 fixed
    - scrub_test:
    """
    if config is None:
        config = {}
    assert isinstance(config, dict), \
        'scrub_test task only accepts a dict for configuration'
    first_mon = teuthology.get_first_mon(ctx, config)
    (mon,) = ctx.cluster.only(first_mon).remotes.iterkeys()

    num_osds = teuthology.num_instances_of_type(ctx.cluster, 'osd')
    log.info('num_osds is %s' % num_osds)

    manager = ceph_manager.CephManager(
        mon,
        ctx=ctx,
        logger=log.getChild('ceph_manager'),
        )

    while len(manager.get_osd_status()['up']) < num_osds:
        time.sleep(10)

    for i in range(num_osds):
        manager.raw_cluster_cmd('tell', 'osd.%d' % i, 'flush_pg_stats')
    manager.wait_for_clean()

    # write some data
    p = manager.do_rados(mon, ['-p', 'rbd', 'bench', '--no-cleanup', '1',
                               'write', '-b', '4096'])
    log.info('err is %d' % p.exitstatus)

    # wait for some PG to have data that we can mess with
    pg, acting = wait_for_victim_pg(manager)
    osd = acting[0]

    osd_remote, obj_path, obj_name = find_victim_object(ctx, pg, osd)
    manager.do_rados(mon, ['-p', 'rbd', 'setomapval', obj_name, 'key', 'val'])
    log.info('err is %d' % p.exitstatus)
    manager.do_rados(mon, ['-p', 'rbd', 'setomapheader', obj_name, 'hdr'])
    log.info('err is %d' % p.exitstatus)

    log.info('messing with PG %s on osd %d' % (pg, osd))
    test_repair_corrupted_obj(ctx, manager, pg, osd_remote, obj_path)
    test_repair_bad_omap(ctx, manager, pg, osd, obj_name)
    test_list_inconsistent_obj(ctx, manager, osd_remote, pg, acting, osd,
                               obj_name, obj_path)
    log.info('test successful!')