def task(ctx, config):
"""
Test handling of lost objects.
A pretty rigid cluster is brought up and tested by this task
"""
POOL = 'unfound_pool'
if config is None:
config = {}
assert isinstance(config, dict), \
'lost_unfound task only accepts a dict for configuration'
first_mon = teuthology.get_first_mon(ctx, config)
(mon, ) = ctx.cluster.only(first_mon).remotes.keys()
manager = ceph_manager.CephManager(
mon,
ctx=ctx,
logger=log.getChild('ceph_manager'),
)
while len(manager.get_osd_status()['up']) < 3:
time.sleep(10)
manager.wait_for_clean()
manager.create_pool(POOL)
# something that is always there
dummyfile = '/etc/fstab'
# take an osd out until the very end
manager.kill_osd(2)
manager.mark_down_osd(2)
manager.mark_out_osd(2)
# kludge to make sure they get a map
rados(ctx, mon, ['-p', POOL, 'put', 'dummy', dummyfile])
manager.flush_pg_stats([0, 1])
manager.wait_for_recovery()
# create old objects
for f in range(1, 10):
rados(ctx, mon, ['-p', POOL, 'put', 'existing_%d' % f, dummyfile])
rados(ctx, mon, ['-p', POOL, 'put', 'existed_%d' % f, dummyfile])
rados(ctx, mon, ['-p', POOL, 'rm', 'existed_%d' % f])
# delay recovery, and make the pg log very long (to prevent backfill)
manager.raw_cluster_cmd(
'tell', 'osd.1', 'injectargs',
'--osd-recovery-delay-start 1000 --osd-min-pg-log-entries 100000000')
manager.kill_osd(0)
manager.mark_down_osd(0)
for f in range(1, 10):
rados(ctx, mon, ['-p', POOL, 'put', 'new_%d' % f, dummyfile])
rados(ctx, mon, ['-p', POOL, 'put', 'existed_%d' % f, dummyfile])
rados(ctx, mon, ['-p', POOL, 'put', 'existing_%d' % f, dummyfile])
# bring osd.0 back up, let it peer, but don't replicate the new
# objects...
log.info('osd.0 command_args is %s' % 'foo')
log.info(ctx.daemons.get_daemon('osd', 0).command_args)
ctx.daemons.get_daemon('osd', 0).command_kwargs['args'].extend(
['--osd-recovery-delay-start', '1000'])
manager.revive_osd(0)
manager.mark_in_osd(0)
manager.wait_till_osd_is_up(0)
manager.flush_pg_stats([1, 0])
manager.wait_till_active()
# take out osd.1 and the only copy of those objects.
manager.kill_osd(1)
manager.mark_down_osd(1)
manager.mark_out_osd(1)
manager.raw_cluster_cmd('osd', 'lost', '1', '--yes-i-really-mean-it')
# bring up osd.2 so that things would otherwise, in theory, recovery fully
manager.revive_osd(2)
manager.mark_in_osd(2)
manager.wait_till_osd_is_up(2)
manager.flush_pg_stats([0, 2])
manager.wait_till_active()
manager.flush_pg_stats([0, 2])
# verify that there are unfound objects
unfound = manager.get_num_unfound_objects()
log.info("there are %d unfound objects" % unfound)
assert unfound
testdir = teuthology.get_testdir(ctx)
procs = []
if config.get('parallel_bench', True):
procs.append(
mon.run(args=[
"/bin/sh",
"-c",
" ".join([
'adjust-ulimits',
'ceph-coverage',
'{tdir}/archive/coverage',
'rados',
'--no-log-to-stderr',
'--name',
'client.admin',
'-b',
str(4 << 10),
'-p',
POOL,
'-t',
'20',
'bench',
'240',
'write',
]).format(tdir=testdir),
],
logger=log.getChild(
'radosbench.{id}'.format(id='client.admin')),
stdin=run.PIPE,
wait=False))
time.sleep(10)
# mark stuff lost
pgs = manager.get_pg_stats()
for pg in pgs:
if pg['stat_sum']['num_objects_unfound'] > 0:
primary = 'osd.%d' % pg['acting'][0]
# verify that i can list them direct from the osd
log.info('listing missing/lost in %s state %s', pg['pgid'],
pg['state'])
m = manager.list_pg_unfound(pg['pgid'])
#log.info('%s' % m)
assert m['num_unfound'] == pg['stat_sum']['num_objects_unfound']
assert m['available_might_have_unfound'] == True
assert m['might_have_unfound'][0]['osd'] == "1"
assert m['might_have_unfound'][0]['status'] == "osd is down"
num_unfound = 0
for o in m['objects']:
if len(o['locations']) == 0:
num_unfound += 1
assert m['num_unfound'] == num_unfound
log.info("reverting unfound in %s on %s", pg['pgid'], primary)
manager.raw_cluster_cmd('pg', pg['pgid'], 'mark_unfound_lost',
'revert')
else:
log.info("no unfound in %s", pg['pgid'])
manager.raw_cluster_cmd('tell', 'osd.0', 'debug', 'kick_recovery_wq', '5')
manager.raw_cluster_cmd('tell', 'osd.2', 'debug', 'kick_recovery_wq', '5')
manager.flush_pg_stats([0, 2])
manager.wait_for_recovery()
# verify result
for f in range(1, 10):
err = rados(ctx, mon, ['-p', POOL, 'get', 'new_%d' % f, '-'])
assert err
err = rados(ctx, mon, ['-p', POOL, 'get', 'existed_%d' % f, '-'])
assert err
err = rados(ctx, mon, ['-p', POOL, 'get', 'existing_%d' % f, '-'])
assert not err
# see if osd.1 can cope
manager.mark_in_osd(1)
manager.revive_osd(1)
manager.wait_till_osd_is_up(1)
manager.wait_for_clean()
run.wait(procs)
manager.wait_for_clean()
def task(ctx, config):
"""
Test handling of object location going down
"""
if config is None:
config = {}
assert isinstance(config, dict), \
'lost_unfound task only accepts a dict for configuration'
first_mon = teuthology.get_first_mon(ctx, config)
(mon,) = ctx.cluster.only(first_mon).remotes.keys()
manager = ceph_manager.CephManager(
mon,
ctx=ctx,
logger=log.getChild('ceph_manager'),
)
while len(manager.get_osd_status()['up']) < 3:
time.sleep(10)
manager.wait_for_clean()
# something that is always there
dummyfile = '/etc/fstab'
# take 0, 1 out
manager.mark_out_osd(0)
manager.mark_out_osd(1)
manager.wait_for_clean()
# delay recovery, and make the pg log very long (to prevent backfill)
manager.raw_cluster_cmd(
'tell', 'osd.0',
'injectargs',
'--osd-recovery-delay-start 10000 --osd-min-pg-log-entries 100000000'
)
# delay recovery, and make the pg log very long (to prevent backfill)
manager.raw_cluster_cmd(
'tell', 'osd.1',
'injectargs',
'--osd-recovery-delay-start 10000 --osd-min-pg-log-entries 100000000'
)
# delay recovery, and make the pg log very long (to prevent backfill)
manager.raw_cluster_cmd(
'tell', 'osd.2',
'injectargs',
'--osd-recovery-delay-start 10000 --osd-min-pg-log-entries 100000000'
)
# delay recovery, and make the pg log very long (to prevent backfill)
manager.raw_cluster_cmd(
'tell', 'osd.3',
'injectargs',
'--osd-recovery-delay-start 10000 --osd-min-pg-log-entries 100000000'
)
# kludge to make sure they get a map
rados(ctx, mon, ['-p', 'data', 'put', 'dummy', dummyfile])
# create old objects
for f in range(1, 10):
rados(ctx, mon, ['-p', 'data', 'put', 'existing_%d' % f, dummyfile])
manager.mark_out_osd(3)
manager.wait_till_active()
manager.mark_in_osd(0)
manager.wait_till_active()
manager.flush_pg_stats([2, 0])
manager.mark_out_osd(2)
manager.wait_till_active()
# bring up 1
manager.mark_in_osd(1)
manager.wait_till_active()
manager.flush_pg_stats([0, 1])
log.info("Getting unfound objects")
unfound = manager.get_num_unfound_objects()
assert not unfound
manager.kill_osd(2)
manager.mark_down_osd(2)
manager.kill_osd(3)
manager.mark_down_osd(3)
manager.flush_pg_stats([0, 1])
log.info("Getting unfound objects")
unfound = manager.get_num_unfound_objects()
assert unfound
def task(ctx, config):
"""
Test handling of lost objects on an ec pool.
A pretty rigid cluster is brought up andtested by this task
"""
if config is None:
config = {}
assert isinstance(config, dict), \
'lost_unfound task only accepts a dict for configuration'
first_mon = teuthology.get_first_mon(ctx, config)
(mon,) = ctx.cluster.only(first_mon).remotes.keys()
manager = ceph_manager.CephManager(
mon,
ctx=ctx,
logger=log.getChild('ceph_manager'),
)
manager.wait_for_clean()
profile = config.get('erasure_code_profile', {
'k': '2',
'm': '2',
'crush-failure-domain': 'osd'
})
profile_name = profile.get('name', 'lost_unfound')
manager.create_erasure_code_profile(profile_name, profile)
pool = manager.create_pool_with_unique_name(
erasure_code_profile_name=profile_name,
min_size=2)
# something that is always there, readable and never empty
dummyfile = '/etc/group'
# kludge to make sure they get a map
rados(ctx, mon, ['-p', pool, 'put', 'dummy', dummyfile])
manager.flush_pg_stats([0, 1])
manager.wait_for_recovery()
# create old objects
for f in range(1, 10):
rados(ctx, mon, ['-p', pool, 'put', 'existing_%d' % f, dummyfile])
rados(ctx, mon, ['-p', pool, 'put', 'existed_%d' % f, dummyfile])
rados(ctx, mon, ['-p', pool, 'rm', 'existed_%d' % f])
# delay recovery, and make the pg log very long (to prevent backfill)
manager.raw_cluster_cmd(
'tell', 'osd.1',
'injectargs',
'--osd-recovery-delay-start 1000 --osd-min-pg-log-entries 100000000'
)
manager.kill_osd(0)
manager.mark_down_osd(0)
manager.kill_osd(3)
manager.mark_down_osd(3)
for f in range(1, 10):
rados(ctx, mon, ['-p', pool, 'put', 'new_%d' % f, dummyfile])
rados(ctx, mon, ['-p', pool, 'put', 'existed_%d' % f, dummyfile])
rados(ctx, mon, ['-p', pool, 'put', 'existing_%d' % f, dummyfile])
# take out osd.1 and a necessary shard of those objects.
manager.kill_osd(1)
manager.mark_down_osd(1)
manager.raw_cluster_cmd('osd', 'lost', '1', '--yes-i-really-mean-it')
manager.revive_osd(0)
manager.wait_till_osd_is_up(0)
manager.revive_osd(3)
manager.wait_till_osd_is_up(3)
manager.flush_pg_stats([0, 2, 3])
manager.wait_till_active()
manager.flush_pg_stats([0, 2, 3])
# verify that there are unfound objects
unfound = manager.get_num_unfound_objects()
log.info("there are %d unfound objects" % unfound)
assert unfound
testdir = teuthology.get_testdir(ctx)
procs = []
if config.get('parallel_bench', True):
procs.append(mon.run(
args=[
"/bin/sh", "-c",
" ".join(['adjust-ulimits',
'ceph-coverage',
'{tdir}/archive/coverage',
'rados',
'--no-log-to-stderr',
'--name', 'client.admin',
'-b', str(4<<10),
'-p' , pool,
'-t', '20',
'bench', '240', 'write',
]).format(tdir=testdir),
],
logger=log.getChild('radosbench.{id}'.format(id='client.admin')),
stdin=run.PIPE,
wait=False
))
time.sleep(10)
# mark stuff lost
pgs = manager.get_pg_stats()
for pg in pgs:
if pg['stat_sum']['num_objects_unfound'] > 0:
# verify that i can list them direct from the osd
log.info('listing missing/lost in %s state %s', pg['pgid'],
pg['state']);
m = manager.list_pg_unfound(pg['pgid'])
log.info('%s' % m)
assert m['num_unfound'] == pg['stat_sum']['num_objects_unfound']
log.info("reverting unfound in %s", pg['pgid'])
manager.raw_cluster_cmd('pg', pg['pgid'],
'mark_unfound_lost', 'delete')
else:
log.info("no unfound in %s", pg['pgid'])
manager.raw_cluster_cmd('tell', 'osd.0', 'debug', 'kick_recovery_wq', '5')
manager.raw_cluster_cmd('tell', 'osd.2', 'debug', 'kick_recovery_wq', '5')
manager.raw_cluster_cmd('tell', 'osd.3', 'debug', 'kick_recovery_wq', '5')
manager.flush_pg_stats([0, 2, 3])
manager.wait_for_recovery()
if not config.get('parallel_bench', True):
time.sleep(20)
# verify result
for f in range(1, 10):
err = rados(ctx, mon, ['-p', pool, 'get', 'new_%d' % f, '-'])
assert err
err = rados(ctx, mon, ['-p', pool, 'get', 'existed_%d' % f, '-'])
assert err
err = rados(ctx, mon, ['-p', pool, 'get', 'existing_%d' % f, '-'])
assert err
# see if osd.1 can cope
manager.revive_osd(1)
manager.wait_till_osd_is_up(1)
manager.wait_for_clean()
run.wait(procs)
def cod_setup(log, ctx, remote, NUM_OBJECTS, DATADIR, BASE_NAME, DATALINECOUNT,
POOL, db, ec):
ERRORS = 0
log.info("Creating {objs} objects in pool".format(objs=NUM_OBJECTS))
objects = range(1, NUM_OBJECTS + 1)
for i in objects:
NAME = BASE_NAME + "{num}".format(num=i)
DDNAME = os.path.join(DATADIR, NAME)
proc = rados(ctx,
remote, ['-p', POOL, 'put', NAME, DDNAME],
wait=False)
# proc = remote.run(args=['rados', '-p', POOL, 'put', NAME, DDNAME])
ret = proc.wait()
if ret != 0:
log.critical("Rados put failed with status {ret}".format(
ret=proc.exitstatus))
sys.exit(1)
db[NAME] = {}
keys = range(i)
db[NAME]["xattr"] = {}
for k in keys:
if k == 0:
continue
mykey = "key{i}-{k}".format(i=i, k=k)
myval = "val{i}-{k}".format(i=i, k=k)
proc = remote.run(
args=['rados', '-p', POOL, 'setxattr', NAME, mykey, myval])
ret = proc.wait()
if ret != 0:
log.error("setxattr failed with {ret}".format(ret=ret))
ERRORS += 1
db[NAME]["xattr"][mykey] = myval
# Erasure coded pools don't support omap
if ec:
continue
# Create omap header in all objects but REPobject1
if i != 1:
myhdr = "hdr{i}".format(i=i)
proc = remote.run(
args=['rados', '-p', POOL, 'setomapheader', NAME, myhdr])
ret = proc.wait()
if ret != 0:
log.critical("setomapheader failed with {ret}".format(ret=ret))
ERRORS += 1
db[NAME]["omapheader"] = myhdr
db[NAME]["omap"] = {}
for k in keys:
if k == 0:
continue
mykey = "okey{i}-{k}".format(i=i, k=k)
myval = "oval{i}-{k}".format(i=i, k=k)
proc = remote.run(
args=['rados', '-p', POOL, 'setomapval', NAME, mykey, myval])
ret = proc.wait()
if ret != 0:
log.critical("setomapval failed with {ret}".format(ret=ret))
db[NAME]["omap"][mykey] = myval
return ERRORS
def task(ctx, config):
"""
Test backfill reservation calculates "toofull" condition correctly.
A pretty rigid cluster is brought up and tested by this task
"""
if config is None:
config = {}
assert isinstance(config, dict), \
'backfill_toofull task only accepts a dict for configuration'
first_mon = teuthology.get_first_mon(ctx, config)
(mon, ) = ctx.cluster.only(first_mon).remotes.keys()
manager = ceph_manager.CephManager(
mon,
ctx=ctx,
logger=log.getChild('ceph_manager'),
)
profile = config.get('erasure_code_profile', {
'k': '2',
'm': '1',
'crush-failure-domain': 'osd'
})
profile_name = profile.get('name', 'backfill_toofull')
manager.create_erasure_code_profile(profile_name, profile)
pool = manager.create_pool_with_unique_name(
pg_num=1, erasure_code_profile_name=profile_name, min_size=2)
manager.raw_cluster_cmd('osd', 'pool', 'set', pool, 'pg_autoscale_mode',
'off')
manager.flush_pg_stats([0, 1, 2, 3])
manager.wait_for_clean()
pool_id = manager.get_pool_num(pool)
pgid = '%d.0' % pool_id
pgs = manager.get_pg_stats()
acting = next((pg['acting'] for pg in pgs if pg['pgid'] == pgid), None)
log.debug("acting=%s" % acting)
assert acting
primary = acting[0]
target = acting[1]
log.debug("write some data")
rados(ctx, mon, ['-p', pool, 'bench', '120', 'write', '--no-cleanup'])
df = manager.get_osd_df(target)
log.debug("target osd df: %s" % df)
total_kb = df['kb']
used_kb = df['kb_used']
log.debug("pause recovery")
manager.raw_cluster_cmd('osd', 'set', 'noout')
manager.raw_cluster_cmd('osd', 'set', 'nobackfill')
manager.raw_cluster_cmd('osd', 'set', 'norecover')
log.debug("stop tartget osd %s" % target)
manager.kill_osd(target)
manager.wait_till_active()
pgs = manager.get_pg_stats()
pg = next((pg for pg in pgs if pg['pgid'] == pgid), None)
log.debug('pg=%s' % pg)
assert pg
log.debug("re-write data")
rados(ctx, mon, ['-p', pool, 'cleanup'])
time.sleep(10)
rados(ctx, mon, ['-p', pool, 'bench', '60', 'write', '--no-cleanup'])
df = manager.get_osd_df(primary)
log.debug("primary osd df: %s" % df)
primary_used_kb = df['kb_used']
log.info("test backfill reservation rejected with toofull")
# We set backfillfull ratio less than new data size and expect the pg
# entering backfill_toofull state.
#
# We also need to update nearfull ratio to prevent "full ratio(s) out of order".
backfillfull = 0.9 * primary_used_kb / total_kb
nearfull = backfillfull * 0.9
log.debug("update nearfull ratio to %s and backfillfull ratio to %s" %
(nearfull, backfillfull))
manager.raw_cluster_cmd('osd', 'set-nearfull-ratio',
'{:.3f}'.format(nearfull + 0.001))
manager.raw_cluster_cmd('osd', 'set-backfillfull-ratio',
'{:.3f}'.format(backfillfull + 0.001))
log.debug("start tartget osd %s" % target)
manager.revive_osd(target)
manager.wait_for_active()
manager.wait_till_osd_is_up(target)
wait_for_pg_state(manager, pgid, 'backfill_toofull', target)
log.info("test pg not enter backfill_toofull after restarting backfill")
# We want to set backfillfull ratio to be big enough for the target to
# successfully backfill new data but smaller than the sum of old and new
# data, so if the osd backfill reservation incorrectly calculates "toofull"
# the test will detect this (fail).
#
# Note, we need to operate with "uncompressed" bytes because currently
# osd backfill reservation does not take compression into account.
#
# We also need to update nearfull ratio to prevent "full ratio(s) out of order".
pdf = manager.get_pool_df(pool)
log.debug("pool %s df: %s" % (pool, pdf))
assert pdf
compress_ratio = 1.0 * pdf['compress_under_bytes'] / pdf['compress_bytes_used'] \
if pdf['compress_bytes_used'] > 0 else 1.0
log.debug("compress_ratio: %s" % compress_ratio)
backfillfull = (used_kb + primary_used_kb) * compress_ratio / total_kb
assert backfillfull < 0.9
nearfull_min = max(used_kb, primary_used_kb) * compress_ratio / total_kb
assert nearfull_min < backfillfull
delta = backfillfull - nearfull_min
nearfull = nearfull_min + delta * 0.1
backfillfull = nearfull_min + delta * 0.2
log.debug("update nearfull ratio to %s and backfillfull ratio to %s" %
(nearfull, backfillfull))
manager.raw_cluster_cmd('osd', 'set-nearfull-ratio',
'{:.3f}'.format(nearfull + 0.001))
manager.raw_cluster_cmd('osd', 'set-backfillfull-ratio',
'{:.3f}'.format(backfillfull + 0.001))
wait_for_pg_state(manager, pgid, 'backfilling', target)
pgs = manager.get_pg_stats()
pg = next((pg for pg in pgs if pg['pgid'] == pgid), None)
log.debug('pg=%s' % pg)
assert pg
log.debug("interrupt %s backfill" % target)
manager.mark_down_osd(target)
# after marking the target osd down it will automatically be
# up soon again
log.debug("resume recovery")
manager.raw_cluster_cmd('osd', 'unset', 'noout')
manager.raw_cluster_cmd('osd', 'unset', 'nobackfill')
manager.raw_cluster_cmd('osd', 'unset', 'norecover')
# wait for everything to peer, backfill and recover
manager.wait_for_clean()
pgs = manager.get_pg_stats()
pg = next((pg for pg in pgs if pg['pgid'] == pgid), None)
log.info('pg=%s' % pg)
assert pg
assert 'clean' in pg['state'].split('+')
def task(ctx, config):
"""
Test handling of divergent entries with prior_version
prior to log_tail and a ceph-objectstore-tool export/import
overrides:
ceph:
conf:
osd:
debug osd: 5
Requires 3 osds on a single test node.
"""
if config is None:
config = {}
assert isinstance(config, dict), \
'divergent_priors task only accepts a dict for configuration'
manager = ctx.managers['ceph']
while len(manager.get_osd_status()['up']) < 3:
time.sleep(10)
manager.flush_pg_stats([0, 1, 2])
manager.raw_cluster_cmd('osd', 'set', 'noout')
manager.raw_cluster_cmd('osd', 'set', 'noin')
manager.raw_cluster_cmd('osd', 'set', 'nodown')
manager.wait_for_clean()
# something that is always there
dummyfile = '/etc/fstab'
dummyfile2 = '/etc/resolv.conf'
testdir = teuthology.get_testdir(ctx)
# create 1 pg pool
log.info('creating foo')
manager.raw_cluster_cmd('osd', 'pool', 'create', 'foo', '1')
osds = [0, 1, 2]
for i in osds:
manager.set_config(i, osd_min_pg_log_entries=10)
manager.set_config(i, osd_max_pg_log_entries=10)
manager.set_config(i, osd_pg_log_trim_min=5)
# determine primary
divergent = manager.get_pg_primary('foo', 0)
log.info("primary and soon to be divergent is %d", divergent)
non_divergent = list(osds)
non_divergent.remove(divergent)
log.info('writing initial objects')
first_mon = teuthology.get_first_mon(ctx, config)
(mon, ) = ctx.cluster.only(first_mon).remotes.keys()
# write 100 objects
for i in range(100):
rados(ctx, mon, ['-p', 'foo', 'put', 'existing_%d' % i, dummyfile])
manager.wait_for_clean()
# blackhole non_divergent
log.info("blackholing osds %s", str(non_divergent))
for i in non_divergent:
manager.set_config(i, objectstore_blackhole=1)
DIVERGENT_WRITE = 5
DIVERGENT_REMOVE = 5
# Write some soon to be divergent
log.info('writing divergent objects')
for i in range(DIVERGENT_WRITE):
rados(ctx,
mon, ['-p', 'foo', 'put',
'existing_%d' % i, dummyfile2],
wait=False)
# Remove some soon to be divergent
log.info('remove divergent objects')
for i in range(DIVERGENT_REMOVE):
rados(ctx,
mon, ['-p', 'foo', 'rm',
'existing_%d' % (i + DIVERGENT_WRITE)],
wait=False)
time.sleep(10)
mon.run(args=['killall', '-9', 'rados'], wait=True, check_status=False)
# kill all the osds but leave divergent in
log.info('killing all the osds')
for i in osds:
manager.kill_osd(i)
for i in osds:
manager.mark_down_osd(i)
for i in non_divergent:
manager.mark_out_osd(i)
# bring up non-divergent
log.info("bringing up non_divergent %s", str(non_divergent))
for i in non_divergent:
manager.revive_osd(i)
for i in non_divergent:
manager.mark_in_osd(i)
# write 1 non-divergent object (ensure that old divergent one is divergent)
objname = "existing_%d" % (DIVERGENT_WRITE + DIVERGENT_REMOVE)
log.info('writing non-divergent object ' + objname)
rados(ctx, mon, ['-p', 'foo', 'put', objname, dummyfile2])
manager.wait_for_recovery()
# ensure no recovery of up osds first
log.info('delay recovery')
for i in non_divergent:
manager.wait_run_admin_socket('osd', i,
['set_recovery_delay', '100000'])
# bring in our divergent friend
log.info("revive divergent %d", divergent)
manager.raw_cluster_cmd('osd', 'set', 'noup')
manager.revive_osd(divergent)
log.info('delay recovery divergent')
manager.wait_run_admin_socket('osd', divergent,
['set_recovery_delay', '100000'])
manager.raw_cluster_cmd('osd', 'unset', 'noup')
while len(manager.get_osd_status()['up']) < 3:
time.sleep(10)
log.info('wait for peering')
rados(ctx, mon, ['-p', 'foo', 'put', 'foo', dummyfile])
# At this point the divergent_priors should have been detected
log.info("killing divergent %d", divergent)
manager.kill_osd(divergent)
# Export a pg
(exp_remote,) = ctx.\
cluster.only('osd.{o}'.format(o=divergent)).remotes.keys()
FSPATH = manager.get_filepath()
JPATH = os.path.join(FSPATH, "journal")
prefix = ("sudo adjust-ulimits ceph-objectstore-tool "
"--data-path {fpath} --journal-path {jpath} "
"--log-file="
"/var/log/ceph/objectstore_tool.$$.log ".format(fpath=FSPATH,
jpath=JPATH))
pid = os.getpid()
expfile = os.path.join(testdir, "exp.{pid}.out".format(pid=pid))
cmd = ((prefix + "--op export-remove --pgid 2.0 --file {file}").format(
id=divergent, file=expfile))
try:
exp_remote.sh(cmd, wait=True)
except CommandFailedError as e:
assert e.exitstatus == 0
cmd = ((prefix + "--op import --file {file}").format(id=divergent,
file=expfile))
try:
exp_remote.sh(cmd, wait=True)
except CommandFailedError as e:
assert e.exitstatus == 0
log.info("reviving divergent %d", divergent)
manager.revive_osd(divergent)
manager.wait_run_admin_socket('osd', divergent, ['dump_ops_in_flight'])
time.sleep(20)
log.info('allowing recovery')
# Set osd_recovery_delay_start back to 0 and kick the queue
for i in osds:
manager.raw_cluster_cmd('tell', 'osd.%d' % i, 'debug',
'kick_recovery_wq', ' 0')
log.info('reading divergent objects')
for i in range(DIVERGENT_WRITE + DIVERGENT_REMOVE):
exit_status = rados(
ctx, mon, ['-p', 'foo', 'get',
'existing_%d' % i, '/tmp/existing'])
assert exit_status == 0
cmd = 'rm {file}'.format(file=expfile)
exp_remote.run(args=cmd, wait=True)
log.info("success")
def test_objectstore(ctx, config, cli_remote, REP_POOL, REP_NAME, ec=False):
manager = ctx.managers['ceph']
osds = ctx.cluster.only(teuthology.is_type('osd'))
TEUTHDIR = teuthology.get_testdir(ctx)
DATADIR = os.path.join(TEUTHDIR, "ceph.data")
DATALINECOUNT = 10000
ERRORS = 0
NUM_OBJECTS = config.get('objects', 10)
log.info("objects: {num}".format(num=NUM_OBJECTS))
pool_dump = manager.get_pool_dump(REP_POOL)
REPID = pool_dump['pool']
log.debug("repid={num}".format(num=REPID))
db = {}
LOCALDIR = tempfile.mkdtemp("cod")
cod_setup_local_data(log, ctx, NUM_OBJECTS, LOCALDIR, REP_NAME,
DATALINECOUNT)
allremote = []
allremote.append(cli_remote)
allremote += list(osds.remotes.keys())
allremote = list(set(allremote))
for remote in allremote:
cod_setup_remote_data(log, ctx, remote, NUM_OBJECTS, DATADIR, REP_NAME,
DATALINECOUNT)
ERRORS += cod_setup(log, ctx, cli_remote, NUM_OBJECTS, DATADIR, REP_NAME,
DATALINECOUNT, REP_POOL, db, ec)
pgs = {}
for stats in manager.get_pg_stats():
if stats["pgid"].find(str(REPID) + ".") != 0:
continue
if pool_dump["type"] == ceph_manager.CephManager.REPLICATED_POOL:
for osd in stats["acting"]:
pgs.setdefault(osd, []).append(stats["pgid"])
elif pool_dump["type"] == ceph_manager.CephManager.ERASURE_CODED_POOL:
shard = 0
for osd in stats["acting"]:
pgs.setdefault(osd, []).append("{pgid}s{shard}".format(
pgid=stats["pgid"], shard=shard))
shard += 1
else:
raise Exception("{pool} has an unexpected type {type}".format(
pool=REP_POOL, type=pool_dump["type"]))
log.info(pgs)
log.info(db)
for osd in manager.get_osd_status()['up']:
manager.kill_osd(osd)
time.sleep(5)
pgswithobjects = set()
objsinpg = {}
# Test --op list and generate json for all objects
log.info("Test --op list by generating json for all objects")
prefix = ("sudo ceph-objectstore-tool "
"--data-path {fpath} "
"--journal-path {jpath} ").format(fpath=FSPATH, jpath=JPATH)
for remote in osds.remotes.keys():
log.debug(remote)
log.debug(osds.remotes[remote])
for role in osds.remotes[remote]:
if not role.startswith("osd."):
continue
osdid = int(role.split('.')[1])
log.info("process osd.{id} on {remote}".format(id=osdid,
remote=remote))
cmd = (prefix + "--op list").format(id=osdid)
try:
lines = remote.sh(cmd, check_status=False).splitlines()
for pgline in lines:
if not pgline:
continue
(pg, obj) = json.loads(pgline)
name = obj['oid']
if name in db:
pgswithobjects.add(pg)
objsinpg.setdefault(pg, []).append(name)
db[name].setdefault("pg2json",
{})[pg] = json.dumps(obj)
except CommandFailedError as e:
log.error(
"Bad exit status {ret} from --op list request".format(
ret=e.exitstatus))
ERRORS += 1
log.info(db)
log.info(pgswithobjects)
log.info(objsinpg)
if pool_dump["type"] == ceph_manager.CephManager.REPLICATED_POOL:
# Test get-bytes
log.info("Test get-bytes and set-bytes")
for basename in db.keys():
file = os.path.join(DATADIR, basename)
GETNAME = os.path.join(DATADIR, "get")
SETNAME = os.path.join(DATADIR, "set")
for remote in osds.remotes.keys():
for role in osds.remotes[remote]:
if not role.startswith("osd."):
continue
osdid = int(role.split('.')[1])
if osdid not in pgs:
continue
for pg, JSON in db[basename]["pg2json"].items():
if pg in pgs[osdid]:
cmd = ((prefix + "--pgid {pg}").format(
id=osdid, pg=pg).split())
cmd.append(run.Raw("'{json}'".format(json=JSON)))
cmd += ("get-bytes {fname}".format(
fname=GETNAME).split())
proc = remote.run(args=cmd, check_status=False)
if proc.exitstatus != 0:
remote.run(args="rm -f {getfile}".format(
getfile=GETNAME).split())
log.error("Bad exit status {ret}".format(
ret=proc.exitstatus))
ERRORS += 1
continue
cmd = ("diff -q {file} {getfile}".format(
file=file, getfile=GETNAME))
proc = remote.run(args=cmd.split())
if proc.exitstatus != 0:
log.error("Data from get-bytes differ")
# log.debug("Got:")
# cat_file(logging.DEBUG, GETNAME)
# log.debug("Expected:")
# cat_file(logging.DEBUG, file)
ERRORS += 1
remote.run(args="rm -f {getfile}".format(
getfile=GETNAME).split())
data = ("put-bytes going into {file}\n".format(
file=file))
teuthology.write_file(remote, SETNAME, data)
cmd = ((prefix + "--pgid {pg}").format(
id=osdid, pg=pg).split())
cmd.append(run.Raw("'{json}'".format(json=JSON)))
cmd += ("set-bytes {fname}".format(
fname=SETNAME).split())
proc = remote.run(args=cmd, check_status=False)
proc.wait()
if proc.exitstatus != 0:
log.info(
"set-bytes failed for object {obj} "
"in pg {pg} osd.{id} ret={ret}".format(
obj=basename,
pg=pg,
id=osdid,
ret=proc.exitstatus))
ERRORS += 1
cmd = ((prefix + "--pgid {pg}").format(
id=osdid, pg=pg).split())
cmd.append(run.Raw("'{json}'".format(json=JSON)))
cmd += "get-bytes -".split()
try:
output = remote.sh(cmd, wait=True)
if data != output:
log.error("Data inconsistent after "
"set-bytes, got:")
log.error(output)
ERRORS += 1
except CommandFailedError as e:
log.error("get-bytes after "
"set-bytes ret={ret}".format(
ret=e.exitstatus))
ERRORS += 1
cmd = ((prefix + "--pgid {pg}").format(
id=osdid, pg=pg).split())
cmd.append(run.Raw("'{json}'".format(json=JSON)))
cmd += ("set-bytes {fname}".format(
fname=file).split())
proc = remote.run(args=cmd, check_status=False)
proc.wait()
if proc.exitstatus != 0:
log.info(
"set-bytes failed for object {obj} "
"in pg {pg} osd.{id} ret={ret}".format(
obj=basename,
pg=pg,
id=osdid,
ret=proc.exitstatus))
ERRORS += 1
log.info("Test list-attrs get-attr")
for basename in db.keys():
file = os.path.join(DATADIR, basename)
GETNAME = os.path.join(DATADIR, "get")
SETNAME = os.path.join(DATADIR, "set")
for remote in osds.remotes.keys():
for role in osds.remotes[remote]:
if not role.startswith("osd."):
continue
osdid = int(role.split('.')[1])
if osdid not in pgs:
continue
for pg, JSON in db[basename]["pg2json"].items():
if pg in pgs[osdid]:
cmd = ((prefix + "--pgid {pg}").format(id=osdid,
pg=pg).split())
cmd.append(run.Raw("'{json}'".format(json=JSON)))
cmd += ["list-attrs"]
try:
keys = remote.sh(cmd, wait=True,
stderr=BytesIO()).split()
except CommandFailedError as e:
log.error("Bad exit status {ret}".format(
ret=e.exitstatus))
ERRORS += 1
continue
values = dict(db[basename]["xattr"])
for key in keys:
if (key == "_" or key == "snapset"
or key == "hinfo_key"):
continue
key = key.strip("_")
if key not in values:
log.error(
"The key {key} should be present".format(
key=key))
ERRORS += 1
continue
exp = values.pop(key)
cmd = ((prefix + "--pgid {pg}").format(
id=osdid, pg=pg).split())
cmd.append(run.Raw("'{json}'".format(json=JSON)))
cmd += ("get-attr {key}".format(key="_" +
key).split())
try:
val = remote.sh(cmd, wait=True)
except CommandFailedError as e:
log.error("get-attr failed with {ret}".format(
ret=e.exitstatus))
ERRORS += 1
continue
if exp != val:
log.error("For key {key} got value {got} "
"instead of {expected}".format(
key=key, got=val, expected=exp))
ERRORS += 1
if "hinfo_key" in keys:
cmd_prefix = prefix.format(id=osdid)
cmd = """
expected=$({prefix} --pgid {pg} '{json}' get-attr {key} | base64)
echo placeholder | {prefix} --pgid {pg} '{json}' set-attr {key} -
test $({prefix} --pgid {pg} '{json}' get-attr {key}) = placeholder
echo $expected | base64 --decode | \
{prefix} --pgid {pg} '{json}' set-attr {key} -
test $({prefix} --pgid {pg} '{json}' get-attr {key} | base64) = $expected
""".format(prefix=cmd_prefix,
pg=pg,
json=JSON,
key="hinfo_key")
log.debug(cmd)
proc = remote.run(
args=['bash', '-e', '-x', '-c', cmd],
check_status=False,
stdout=BytesIO(),
stderr=BytesIO())
proc.wait()
if proc.exitstatus != 0:
log.error("failed with " +
str(proc.exitstatus))
log.error(" ".join([
six.ensure_str(proc.stdout.getvalue()),
six.ensure_str(proc.stderr.getvalue()),
]))
ERRORS += 1
if len(values) != 0:
log.error("Not all keys found, remaining keys:")
log.error(values)
log.info("Test pg info")
for remote in osds.remotes.keys():
for role in osds.remotes[remote]:
if not role.startswith("osd."):
continue
osdid = int(role.split('.')[1])
if osdid not in pgs:
continue
for pg in pgs[osdid]:
cmd = ((prefix + "--op info --pgid {pg}").format(
id=osdid, pg=pg).split())
try:
info = remote.sh(cmd, wait=True)
except CommandFailedError as e:
log.error("Failure of --op info command with {ret}".format(
e.exitstatus))
ERRORS += 1
continue
if not str(pg) in info:
log.error("Bad data from info: {info}".format(info=info))
ERRORS += 1
log.info("Test pg logging")
for remote in osds.remotes.keys():
for role in osds.remotes[remote]:
if not role.startswith("osd."):
continue
osdid = int(role.split('.')[1])
if osdid not in pgs:
continue
for pg in pgs[osdid]:
cmd = ((prefix + "--op log --pgid {pg}").format(id=osdid,
pg=pg).split())
try:
output = remote.sh(cmd, wait=True)
except CommandFailedError as e:
log.error("Getting log failed for pg {pg} "
"from osd.{id} with {ret}".format(
pg=pg, id=osdid, ret=e.exitstatus))
ERRORS += 1
continue
HASOBJ = pg in pgswithobjects
MODOBJ = "modify" in output
if HASOBJ != MODOBJ:
log.error("Bad log for pg {pg} from osd.{id}".format(
pg=pg, id=osdid))
MSG = (HASOBJ and [""] or ["NOT "])[0]
log.error(
"Log should {msg}have a modify entry".format(msg=MSG))
ERRORS += 1
log.info("Test pg export")
EXP_ERRORS = 0
for remote in osds.remotes.keys():
for role in osds.remotes[remote]:
if not role.startswith("osd."):
continue
osdid = int(role.split('.')[1])
if osdid not in pgs:
continue
for pg in pgs[osdid]:
fpath = os.path.join(DATADIR, "osd{id}.{pg}".format(id=osdid,
pg=pg))
cmd = ((prefix +
"--op export --pgid {pg} --file {file}").format(
id=osdid, pg=pg, file=fpath))
try:
remote.sh(cmd, wait=True)
except CommandFailedError as e:
log.error("Exporting failed for pg {pg} "
"on osd.{id} with {ret}".format(
pg=pg, id=osdid, ret=e.exitstatus))
EXP_ERRORS += 1
ERRORS += EXP_ERRORS
log.info("Test pg removal")
RM_ERRORS = 0
for remote in osds.remotes.keys():
for role in osds.remotes[remote]:
if not role.startswith("osd."):
continue
osdid = int(role.split('.')[1])
if osdid not in pgs:
continue
for pg in pgs[osdid]:
cmd = ((prefix + "--force --op remove --pgid {pg}").format(
pg=pg, id=osdid))
try:
remote.sh(cmd, wait=True)
except CommandFailedError as e:
log.error("Removing failed for pg {pg} "
"on osd.{id} with {ret}".format(
pg=pg, id=osdid, ret=e.exitstatus))
RM_ERRORS += 1
ERRORS += RM_ERRORS
IMP_ERRORS = 0
if EXP_ERRORS == 0 and RM_ERRORS == 0:
log.info("Test pg import")
for remote in osds.remotes.keys():
for role in osds.remotes[remote]:
if not role.startswith("osd."):
continue
osdid = int(role.split('.')[1])
if osdid not in pgs:
continue
for pg in pgs[osdid]:
fpath = os.path.join(
DATADIR, "osd{id}.{pg}".format(id=osdid, pg=pg))
cmd = ((prefix + "--op import --file {file}").format(
id=osdid, file=fpath))
try:
remote.sh(cmd, wait=True)
except CommandFailedError as e:
log.error(
"Import failed from {file} with {ret}".format(
file=fpath, ret=e.exitstatus))
IMP_ERRORS += 1
else:
log.warning("SKIPPING IMPORT TESTS DUE TO PREVIOUS FAILURES")
ERRORS += IMP_ERRORS
if EXP_ERRORS == 0 and RM_ERRORS == 0 and IMP_ERRORS == 0:
log.info("Restarting OSDs....")
# They are still look to be up because of setting nodown
for osd in manager.get_osd_status()['up']:
manager.revive_osd(osd)
# Wait for health?
time.sleep(5)
# Let scrub after test runs verify consistency of all copies
log.info("Verify replicated import data")
objects = range(1, NUM_OBJECTS + 1)
for i in objects:
NAME = REP_NAME + "{num}".format(num=i)
TESTNAME = os.path.join(DATADIR, "gettest")
REFNAME = os.path.join(DATADIR, NAME)
proc = rados(ctx,
cli_remote, ['-p', REP_POOL, 'get', NAME, TESTNAME],
wait=False)
ret = proc.wait()
if ret != 0:
log.error("After import, rados get failed with {ret}".format(
ret=proc.exitstatus))
ERRORS += 1
continue
cmd = "diff -q {gettest} {ref}".format(gettest=TESTNAME,
ref=REFNAME)
proc = cli_remote.run(args=cmd, check_status=False)
proc.wait()
if proc.exitstatus != 0:
log.error("Data comparison failed for {obj}".format(obj=NAME))
ERRORS += 1
return ERRORS
def task(ctx, config):
"""
Test handling of divergent entries with prior_version
prior to log_tail
overrides:
ceph:
conf:
osd:
debug osd: 5
Requires 3 osds on a single test node.
"""
if config is None:
config = {}
assert isinstance(config, dict), \
'divergent_priors task only accepts a dict for configuration'
manager = ctx.managers['ceph']
while len(manager.get_osd_status()['up']) < 3:
time.sleep(10)
manager.flush_pg_stats([0, 1, 2])
manager.raw_cluster_cmd('osd', 'set', 'noout')
manager.raw_cluster_cmd('osd', 'set', 'noin')
manager.raw_cluster_cmd('osd', 'set', 'nodown')
manager.wait_for_clean()
# something that is always there
dummyfile = '/etc/fstab'
dummyfile2 = '/etc/resolv.conf'
# create 1 pg pool
log.info('creating foo')
manager.raw_cluster_cmd('osd', 'pool', 'create', 'foo', '1')
osds = [0, 1, 2]
for i in osds:
manager.set_config(i, osd_min_pg_log_entries=10)
manager.set_config(i, osd_max_pg_log_entries=10)
manager.set_config(i, osd_pg_log_trim_min=5)
# determine primary
divergent = manager.get_pg_primary('foo', 0)
log.info("primary and soon to be divergent is %d", divergent)
non_divergent = list(osds)
non_divergent.remove(divergent)
log.info('writing initial objects')
first_mon = teuthology.get_first_mon(ctx, config)
(mon,) = ctx.cluster.only(first_mon).remotes.keys()
# write 100 objects
for i in range(100):
rados(ctx, mon, ['-p', 'foo', 'put', 'existing_%d' % i, dummyfile])
manager.wait_for_clean()
# blackhole non_divergent
log.info("blackholing osds %s", str(non_divergent))
for i in non_divergent:
manager.set_config(i, objectstore_blackhole=1)
DIVERGENT_WRITE = 5
DIVERGENT_REMOVE = 5
# Write some soon to be divergent
log.info('writing divergent objects')
for i in range(DIVERGENT_WRITE):
rados(ctx, mon, ['-p', 'foo', 'put', 'existing_%d' % i,
dummyfile2], wait=False)
# Remove some soon to be divergent
log.info('remove divergent objects')
for i in range(DIVERGENT_REMOVE):
rados(ctx, mon, ['-p', 'foo', 'rm',
'existing_%d' % (i + DIVERGENT_WRITE)], wait=False)
time.sleep(10)
mon.run(
args=['killall', '-9', 'rados'],
wait=True,
check_status=False)
# kill all the osds but leave divergent in
log.info('killing all the osds')
for i in osds:
manager.kill_osd(i)
for i in osds:
manager.mark_down_osd(i)
for i in non_divergent:
manager.mark_out_osd(i)
# bring up non-divergent
log.info("bringing up non_divergent %s", str(non_divergent))
for i in non_divergent:
manager.revive_osd(i)
for i in non_divergent:
manager.mark_in_osd(i)
# write 1 non-divergent object (ensure that old divergent one is divergent)
objname = "existing_%d" % (DIVERGENT_WRITE + DIVERGENT_REMOVE)
log.info('writing non-divergent object ' + objname)
rados(ctx, mon, ['-p', 'foo', 'put', objname, dummyfile2])
manager.wait_for_recovery()
# ensure no recovery of up osds first
log.info('delay recovery')
for i in non_divergent:
manager.wait_run_admin_socket(
'osd', i, ['set_recovery_delay', '100000'])
# bring in our divergent friend
log.info("revive divergent %d", divergent)
manager.raw_cluster_cmd('osd', 'set', 'noup')
manager.revive_osd(divergent)
log.info('delay recovery divergent')
manager.wait_run_admin_socket(
'osd', divergent, ['set_recovery_delay', '100000'])
manager.raw_cluster_cmd('osd', 'unset', 'noup')
while len(manager.get_osd_status()['up']) < 3:
time.sleep(10)
log.info('wait for peering')
rados(ctx, mon, ['-p', 'foo', 'put', 'foo', dummyfile])
# At this point the divergent_priors should have been detected
log.info("killing divergent %d", divergent)
manager.kill_osd(divergent)
log.info("reviving divergent %d", divergent)
manager.revive_osd(divergent)
time.sleep(20)
log.info('allowing recovery')
# Set osd_recovery_delay_start back to 0 and kick the queue
for i in osds:
manager.raw_cluster_cmd('tell', 'osd.%d' % i, 'debug',
'kick_recovery_wq', ' 0')
log.info('reading divergent objects')
for i in range(DIVERGENT_WRITE + DIVERGENT_REMOVE):
exit_status = rados(ctx, mon, ['-p', 'foo', 'get', 'existing_%d' % i,
'/tmp/existing'])
assert exit_status == 0
log.info("success")
def task(ctx, config):
"""
Test handling of osd_failsafe_nearfull_ratio and osd_failsafe_full_ratio
configuration settings
In order for test to pass must use log-ignorelist as follows
tasks:
- chef:
- install:
- ceph:
log-ignorelist: ['OSD near full', 'OSD full dropping all updates']
- osd_failsafe_enospc:
"""
if config is None:
config = {}
assert isinstance(config, dict), \
'osd_failsafe_enospc task only accepts a dict for configuration'
# Give 2 seconds for injectargs + osd_op_complaint_time (30) + 2 * osd_heartbeat_interval (6) + 6 padding
sleep_time = 50
# something that is always there
dummyfile = '/etc/fstab'
dummyfile2 = '/etc/resolv.conf'
manager = ctx.managers['ceph']
# create 1 pg pool with 1 rep which can only be on osd.0
osds = manager.get_osd_dump()
for osd in osds:
if osd['osd'] != 0:
manager.mark_out_osd(osd['osd'])
log.info('creating pool foo')
manager.create_pool("foo")
manager.raw_cluster_cmd('osd', 'pool', 'set', 'foo', 'size', '1')
# State NONE -> NEAR
log.info('1. Verify warning messages when exceeding nearfull_ratio')
first_mon = teuthology.get_first_mon(ctx, config)
(mon, ) = ctx.cluster.only(first_mon).remotes.keys()
proc = mon.run(
args=['sudo', 'daemon-helper', 'kill', 'ceph', '-w'],
stdin=run.PIPE,
stdout=StringIO(),
wait=False,
)
manager.raw_cluster_cmd('tell', 'osd.0', 'injectargs',
'--osd_failsafe_nearfull_ratio .00001')
time.sleep(sleep_time)
proc.stdin.close() # causes daemon-helper send SIGKILL to ceph -w
proc.wait()
lines = proc.stdout.getvalue().split('\n')
count = len(filter(lambda line: '[WRN] OSD near full' in line, lines))
assert count == 2, 'Incorrect number of warning messages expected 2 got %d' % count
count = len(
filter(lambda line: '[ERR] OSD full dropping all updates' in line,
lines))
assert count == 0, 'Incorrect number of error messages expected 0 got %d' % count
# State NEAR -> FULL
log.info('2. Verify error messages when exceeding full_ratio')
proc = mon.run(
args=['sudo', 'daemon-helper', 'kill', 'ceph', '-w'],
stdin=run.PIPE,
stdout=StringIO(),
wait=False,
)
manager.raw_cluster_cmd('tell', 'osd.0', 'injectargs',
'--osd_failsafe_full_ratio .00001')
time.sleep(sleep_time)
proc.stdin.close() # causes daemon-helper send SIGKILL to ceph -w
proc.wait()
lines = proc.stdout.getvalue().split('\n')
count = len(
filter(lambda line: '[ERR] OSD full dropping all updates' in line,
lines))
assert count == 2, 'Incorrect number of error messages expected 2 got %d' % count
log.info('3. Verify write failure when exceeding full_ratio')
# Write data should fail
ret = rados(ctx, mon, ['-p', 'foo', 'put', 'newfile1', dummyfile])
assert ret != 0, 'Expected write failure but it succeeded with exit status 0'
# Put back default
manager.raw_cluster_cmd('tell', 'osd.0', 'injectargs',
'--osd_failsafe_full_ratio .97')
time.sleep(10)
# State FULL -> NEAR
log.info('4. Verify write success when NOT exceeding full_ratio')
# Write should succeed
ret = rados(ctx, mon, ['-p', 'foo', 'put', 'newfile2', dummyfile2])
assert ret == 0, 'Expected write to succeed, but got exit status %d' % ret
log.info('5. Verify warning messages again when exceeding nearfull_ratio')
proc = mon.run(
args=['sudo', 'daemon-helper', 'kill', 'ceph', '-w'],
stdin=run.PIPE,
stdout=StringIO(),
wait=False,
)
time.sleep(sleep_time)
proc.stdin.close() # causes daemon-helper send SIGKILL to ceph -w
proc.wait()
lines = proc.stdout.getvalue().split('\n')
count = len(filter(lambda line: '[WRN] OSD near full' in line, lines))
assert count == 1 or count == 2, 'Incorrect number of warning messages expected 1 or 2 got %d' % count
count = len(
filter(lambda line: '[ERR] OSD full dropping all updates' in line,
lines))
assert count == 0, 'Incorrect number of error messages expected 0 got %d' % count
manager.raw_cluster_cmd('tell', 'osd.0', 'injectargs',
'--osd_failsafe_nearfull_ratio .90')
time.sleep(10)
# State NONE -> FULL
log.info('6. Verify error messages again when exceeding full_ratio')
proc = mon.run(
args=['sudo', 'daemon-helper', 'kill', 'ceph', '-w'],
stdin=run.PIPE,
stdout=StringIO(),
wait=False,
)
manager.raw_cluster_cmd('tell', 'osd.0', 'injectargs',
'--osd_failsafe_full_ratio .00001')
time.sleep(sleep_time)
proc.stdin.close() # causes daemon-helper send SIGKILL to ceph -w
proc.wait()
lines = proc.stdout.getvalue().split('\n')
count = len(filter(lambda line: '[WRN] OSD near full' in line, lines))
assert count == 0, 'Incorrect number of warning messages expected 0 got %d' % count
count = len(
filter(lambda line: '[ERR] OSD full dropping all updates' in line,
lines))
assert count == 2, 'Incorrect number of error messages expected 2 got %d' % count
# State FULL -> NONE
log.info('7. Verify no messages settings back to default')
manager.raw_cluster_cmd('tell', 'osd.0', 'injectargs',
'--osd_failsafe_full_ratio .97')
time.sleep(10)
proc = mon.run(
args=['sudo', 'daemon-helper', 'kill', 'ceph', '-w'],
stdin=run.PIPE,
stdout=StringIO(),
wait=False,
)
time.sleep(sleep_time)
proc.stdin.close() # causes daemon-helper send SIGKILL to ceph -w
proc.wait()
lines = proc.stdout.getvalue().split('\n')
count = len(filter(lambda line: '[WRN] OSD near full' in line, lines))
assert count == 0, 'Incorrect number of warning messages expected 0 got %d' % count
count = len(
filter(lambda line: '[ERR] OSD full dropping all updates' in line,
lines))
assert count == 0, 'Incorrect number of error messages expected 0 got %d' % count
log.info('Test Passed')
# Bring all OSDs back in
manager.remove_pool("foo")
for osd in osds:
if osd['osd'] != 0:
manager.mark_in_osd(osd['osd'])
def task(ctx, config):
"""
Test peering.
"""
if config is None:
config = {}
assert isinstance(config, dict), \
'peer task only accepts a dict for configuration'
first_mon = teuthology.get_first_mon(ctx, config)
(mon, ) = ctx.cluster.only(first_mon).remotes.keys()
manager = ceph_manager.CephManager(
mon,
ctx=ctx,
logger=log.getChild('ceph_manager'),
)
while len(manager.get_osd_status()['up']) < 3:
time.sleep(10)
manager.flush_pg_stats([0, 1, 2])
manager.wait_for_clean()
for i in range(3):
manager.set_config(i, osd_recovery_delay_start=120)
# take on osd down
manager.kill_osd(2)
manager.mark_down_osd(2)
# kludge to make sure they get a map
rados(ctx, mon, ['-p', 'data', 'get', 'dummy', '-'])
manager.flush_pg_stats([0, 1])
manager.wait_for_recovery()
# kill another and revive 2, so that some pgs can't peer.
manager.kill_osd(1)
manager.mark_down_osd(1)
manager.revive_osd(2)
manager.wait_till_osd_is_up(2)
manager.flush_pg_stats([0, 2])
manager.wait_for_active_or_down()
manager.flush_pg_stats([0, 2])
# look for down pgs
num_down_pgs = 0
pgs = manager.get_pg_stats()
for pg in pgs:
out = manager.raw_cluster_cmd('pg', pg['pgid'], 'query')
log.debug("out string %s", out)
j = json.loads(out)
log.info("pg is %s, query json is %s", pg, j)
if pg['state'].count('down'):
num_down_pgs += 1
# verify that it is blocked on osd.1
rs = j['recovery_state']
assert len(rs) >= 2
assert rs[0]['name'] == 'Started/Primary/Peering/Down'
assert rs[1]['name'] == 'Started/Primary/Peering'
assert rs[1]['blocked']
assert rs[1]['down_osds_we_would_probe'] == [1]
assert len(rs[1]['peering_blocked_by']) == 1
assert rs[1]['peering_blocked_by'][0]['osd'] == 1
assert num_down_pgs > 0
# bring it all back
manager.revive_osd(1)
manager.wait_till_osd_is_up(1)
manager.flush_pg_stats([0, 1, 2])
manager.wait_for_clean()
def task(ctx, config):
"""
Test handling resolve stuck peering
requires 3 osds on a single test node
"""
if config is None:
config = {}
assert isinstance(config, dict), \
'Resolve stuck peering only accepts a dict for config'
manager = ctx.managers['ceph']
while len(manager.get_osd_status()['up']) < 3:
time.sleep(10)
manager.wait_for_clean()
dummyfile = '/etc/fstab'
dummyfile1 = '/etc/resolv.conf'
#create 1 PG pool
pool='foo'
log.info('creating pool foo')
manager.raw_cluster_cmd('osd', 'pool', 'create', '%s' % pool, '1')
#set min_size of the pool to 1
#so that we can continue with I/O
#when 2 osds are down
manager.set_pool_property(pool, "min_size", 1)
osds = [0, 1, 2]
primary = manager.get_pg_primary('foo', 0)
log.info("primary osd is %d", primary)
others = list(osds)
others.remove(primary)
log.info('writing initial objects')
first_mon = teuthology.get_first_mon(ctx, config)
(mon,) = ctx.cluster.only(first_mon).remotes.keys()
#create few objects
for i in range(100):
rados(ctx, mon, ['-p', 'foo', 'put', 'existing_%d' % i, dummyfile])
manager.wait_for_clean()
#kill other osds except primary
log.info('killing other osds except primary')
for i in others:
manager.kill_osd(i)
for i in others:
manager.mark_down_osd(i)
for i in range(100):
rados(ctx, mon, ['-p', 'foo', 'put', 'new_%d' % i, dummyfile1])
#kill primary osd
manager.kill_osd(primary)
manager.mark_down_osd(primary)
#revive other 2 osds
for i in others:
manager.revive_osd(i)
#make sure that pg is down
#Assuming pg number for single pg pool will start from 0
pgnum=0
pgstr = manager.get_pgid(pool, pgnum)
stats = manager.get_single_pg_stats(pgstr)
print(stats['state'])
timeout=60
start=time.time()
while 'down' not in stats['state']:
assert time.time() - start < timeout, \
'failed to reach down state before timeout expired'
stats = manager.get_single_pg_stats(pgstr)
#mark primary as lost
manager.raw_cluster_cmd('osd', 'lost', '%d' % primary,\
'--yes-i-really-mean-it')
#expect the pg status to be active+undersized+degraded
#pg should recover and become active+clean within timeout
stats = manager.get_single_pg_stats(pgstr)
print(stats['state'])
timeout=10
start=time.time()
while manager.get_num_down():
assert time.time() - start < timeout, \
'failed to recover before timeout expired'
manager.revive_osd(primary)
def task(ctx, config):
"""
Test handling of objects with inconsistent hash info during backfill and deep-scrub.
A pretty rigid cluster is brought up and tested by this task
"""
if config is None:
config = {}
assert isinstance(config, dict), \
'ec_inconsistent_hinfo task only accepts a dict for configuration'
first_mon = teuthology.get_first_mon(ctx, config)
(mon,) = ctx.cluster.only(first_mon).remotes.keys()
manager = ceph_manager.CephManager(
mon,
ctx=ctx,
logger=log.getChild('ceph_manager'),
)
profile = config.get('erasure_code_profile', {
'k': '2',
'm': '1',
'crush-failure-domain': 'osd'
})
profile_name = profile.get('name', 'backfill_unfound')
manager.create_erasure_code_profile(profile_name, profile)
pool = manager.create_pool_with_unique_name(
pg_num=1,
erasure_code_profile_name=profile_name,
min_size=2)
manager.raw_cluster_cmd('osd', 'pool', 'set', pool,
'pg_autoscale_mode', 'off')
manager.flush_pg_stats([0, 1, 2, 3])
manager.wait_for_clean()
pool_id = manager.get_pool_num(pool)
pgid = '%d.0' % pool_id
pgs = manager.get_pg_stats()
acting = next((pg['acting'] for pg in pgs if pg['pgid'] == pgid), None)
log.info("acting=%s" % acting)
assert acting
primary = acting[0]
# something that is always there, readable and never empty
dummyfile = '/etc/group'
# kludge to make sure they get a map
rados(ctx, mon, ['-p', pool, 'put', 'dummy', dummyfile])
manager.flush_pg_stats([0, 1])
manager.wait_for_recovery()
log.debug("create test object")
obj = 'test'
rados(ctx, mon, ['-p', pool, 'put', obj, dummyfile])
victim = acting[1]
log.info("remove test object hash info from osd.%s shard and test deep-scrub and repair"
% victim)
manager.objectstore_tool(pool, options='', args='rm-attr hinfo_key',
object_name=obj, osd=victim)
check_time_now = time.strftime('%s')
manager.raw_cluster_cmd('pg', 'deep-scrub', pgid)
wait_for_deep_scrub_complete(manager, pgid, check_time_now, True)
check_time_now = time.strftime('%s')
manager.raw_cluster_cmd('pg', 'repair', pgid)
wait_for_deep_scrub_complete(manager, pgid, check_time_now, False)
log.info("remove test object hash info from primary osd.%s shard and test backfill"
% primary)
log.debug("write some data")
rados(ctx, mon, ['-p', pool, 'bench', '30', 'write', '-b', '4096',
'--no-cleanup'])
manager.objectstore_tool(pool, options='', args='rm-attr hinfo_key',
object_name=obj, osd=primary)
# mark the osd out to trigger a rebalance/backfill
source = acting[1]
target = [x for x in [0, 1, 2, 3] if x not in acting][0]
manager.mark_out_osd(source)
# wait for everything to peer, backfill and recover
wait_for_backfilling_complete(manager, pgid, source, target)
manager.wait_for_clean()
manager.flush_pg_stats([0, 1, 2, 3])
pgs = manager.get_pg_stats()
pg = next((pg for pg in pgs if pg['pgid'] == pgid), None)
log.debug('pg=%s' % pg)
assert pg
assert 'clean' in pg['state'].split('+')
assert 'inconsistent' not in pg['state'].split('+')
unfound = manager.get_num_unfound_objects()
log.debug("there are %d unfound objects" % unfound)
assert unfound == 0
source, target = target, source
log.info("remove test object hash info from non-primary osd.%s shard and test backfill"
% source)
manager.objectstore_tool(pool, options='', args='rm-attr hinfo_key',
object_name=obj, osd=source)
# mark the osd in to trigger a rebalance/backfill
manager.mark_in_osd(target)
# wait for everything to peer, backfill and recover
wait_for_backfilling_complete(manager, pgid, source, target)
manager.wait_for_clean()
manager.flush_pg_stats([0, 1, 2, 3])
pgs = manager.get_pg_stats()
pg = next((pg for pg in pgs if pg['pgid'] == pgid), None)
log.debug('pg=%s' % pg)
assert pg
assert 'clean' in pg['state'].split('+')
assert 'inconsistent' not in pg['state'].split('+')
unfound = manager.get_num_unfound_objects()
log.debug("there are %d unfound objects" % unfound)
assert unfound == 0
log.info("remove hash info from two shards and test backfill")
source = acting[2]
target = [x for x in [0, 1, 2, 3] if x not in acting][0]
manager.objectstore_tool(pool, options='', args='rm-attr hinfo_key',
object_name=obj, osd=primary)
manager.objectstore_tool(pool, options='', args='rm-attr hinfo_key',
object_name=obj, osd=source)
# mark the osd out to trigger a rebalance/backfill
manager.mark_out_osd(source)
# wait for everything to peer, backfill and detect unfound object
wait_for_backfilling_complete(manager, pgid, source, target)
# verify that there is unfound object
manager.flush_pg_stats([0, 1, 2, 3])
pgs = manager.get_pg_stats()
pg = next((pg for pg in pgs if pg['pgid'] == pgid), None)
log.debug('pg=%s' % pg)
assert pg
assert 'backfill_unfound' in pg['state'].split('+')
unfound = manager.get_num_unfound_objects()
log.debug("there are %d unfound objects" % unfound)
assert unfound == 1
m = manager.list_pg_unfound(pgid)
log.debug('list_pg_unfound=%s' % m)
assert m['num_unfound'] == pg['stat_sum']['num_objects_unfound']
# mark stuff lost
pgs = manager.get_pg_stats()
manager.raw_cluster_cmd('pg', pgid, 'mark_unfound_lost', 'delete')
# wait for everything to peer and be happy...
manager.flush_pg_stats([0, 1, 2, 3])
manager.wait_for_recovery()
