From: Mykola Golub <mgolub@suse.com>
Date: Thu, 9 Sep 2021 11:44:25 +0000 (+0300)
Subject: qa/suites/rados: add backfill_toofull test
X-Git-Tag: v16.2.7~82^2~1
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=ccd1c40433120a39daf2a0398a9da9517da2e5bb;p=ceph.git

qa/suites/rados: add backfill_toofull test

Signed-off-by: Mykola Golub <mgolub@suse.com>
(cherry picked from commit 76743e005866664795e9240460734b31108824e2)
---

diff --git a/qa/suites/rados/singleton/all/backfill-toofull.yaml b/qa/suites/rados/singleton/all/backfill-toofull.yaml
new file mode 100644
index 00000000000..fcc3d0e2942
--- /dev/null
+++ b/qa/suites/rados/singleton/all/backfill-toofull.yaml
@@ -0,0 +1,37 @@
+roles:
+- - mon.a
+  - mon.b
+  - mon.c
+  - mgr.x
+  - osd.0
+  - osd.1
+  - osd.2
+  - osd.3
+openstack:
+  - volumes: # attached to each instance
+      count: 4
+      size: 10 # GB
+tasks:
+- install:
+- ceph:
+    create_rbd_pool: false
+    pre-mgr-commands:
+      - sudo ceph config set mgr mgr/devicehealth/enable_monitoring false --force
+    log-ignorelist:
+      - Error
+      - overall HEALTH_
+      - \(OBJECT_
+      - \(OSDMAP_FLAGS\)
+      - \(OSD_
+      - \(PG_
+      - \(POOL_BACKFILLFULL\)
+      - \(POOL_NEARFULL\)
+      - \(SLOW_OPS\)
+      - \(TOO_FEW_PGS\)
+      - Monitor daemon marked osd\.[[:digit:]]+ down, but it is still running
+      - slow request
+    conf:
+      osd:
+        osd min pg log entries: 5
+        osd max pg log entries: 5
+- backfill_toofull:
diff --git a/qa/tasks/backfill_toofull.py b/qa/tasks/backfill_toofull.py
new file mode 100644
index 00000000000..f2156bea601
--- /dev/null
+++ b/qa/tasks/backfill_toofull.py
@@ -0,0 +1,181 @@
+"""
+Backfill_toofull
+"""
+import logging
+import time
+from tasks import ceph_manager
+from tasks.util.rados import rados
+from teuthology import misc as teuthology
+
+log = logging.getLogger(__name__)
+
+def wait_for_pg_state(manager, pgid, state, to_osd):
+    log.debug("waiting for pg %s state is %s" % (pgid, state))
+    for i in range(300):
+        time.sleep(5)
+        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.info('pg=%s' % pg);
+        assert pg
+        status = pg['state'].split('+')
+        if 'active' not in status:
+            log.debug('not active')
+            continue
+        if state not in status:
+            log.debug('not %s' % state)
+            continue
+        assert to_osd in pg['up']
+        return
+    assert False, '%s not in %s' % (pgid, state)
+
+
+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', '60', '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))
+    manager.raw_cluster_cmd('osd', 'set-backfillfull-ratio',
+                            '{:.3f}'.format(backfillfull))
+
+    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).
+    #
+    # We also need to update nearfull ratio to prevent "full ratio(s) out of order".
+
+    backfillfull = min(used_kb + primary_used_kb, total_kb * 0.9) / total_kb
+    nearfull_min = max(used_kb, primary_used_kb) / total_kb
+    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))
+    manager.raw_cluster_cmd('osd', 'set-backfillfull-ratio',
+                            '{:.3f}'.format(backfillfull))
+
+    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('+')
diff --git a/qa/tasks/ceph_manager.py b/qa/tasks/ceph_manager.py
index eb078c478c1..2ff26a9cbf0 100644
--- a/qa/tasks/ceph_manager.py
+++ b/qa/tasks/ceph_manager.py
@@ -2324,6 +2324,15 @@ class CephManager:
         except KeyError:
             return j['pg_stats']
 
+    def get_osd_df(self, osdid):
+        """
+        Get the osd df stats
+        """
+        out = self.raw_cluster_cmd('osd', 'df', 'name', 'osd.{}'.format(osdid),
+                                   '--format=json')
+        j = json.loads('\n'.join(out.split('\n')[1:]))
+        return j['nodes'][0]
+
     def get_pgids_to_force(self, backfill):
         """
         Return the randomized list of PGs that can have their recovery/backfill forced