]> git.apps.os.sepia.ceph.com Git - ceph.git/commitdiff
quincy: qa/workunits/windows: backport rbd-wnbd tests 49883/head
authorLucian Petrut <lpetrut@cloudbasesolutions.com>
Thu, 26 Jan 2023 07:39:11 +0000 (09:39 +0200)
committerLucian Petrut <lpetrut@cloudbasesolutions.com>
Thu, 23 Feb 2023 13:11:28 +0000 (15:11 +0200)
The quincy branch doesn't currently include the rbd-wnbd python
tests. For this reason, the ceph-build scripts are fetching a
certain version from the main branch and then run each test
individually [1].

One issue is that the fio test is failing on quincy and we aren't
getting any information. It seemed to be a time out, but it still
fails even after the timeout was increased.

By backporting the tests to quincy, the tests will be invoked
through the "run-tests.ps1" script, hopefully receiving more
information.

Note that we're skipping the live resize test since this feature
isn't available in Quincy.

This will also allow us to simplify the ceph-build script.

[1] https://github.com/ceph/ceph-build/blob/b5adddec2cb03059d0bc0b5d4ffa7aac7a248fc0/scripts/ceph-windows/run_tests#L70-L85

Signed-off-by: Lucian Petrut <lpetrut@cloudbasesolutions.com>
qa/workunits/windows/run-tests.ps1 [new file with mode: 0644]
qa/workunits/windows/test_rbd_wnbd.py [new file with mode: 0644]

diff --git a/qa/workunits/windows/run-tests.ps1 b/qa/workunits/windows/run-tests.ps1
new file mode 100644 (file)
index 0000000..9aef612
--- /dev/null
@@ -0,0 +1,31 @@
+$ProgressPreference = "SilentlyContinue"
+$ErrorActionPreference = "Stop"
+
+$scriptLocation = [System.IO.Path]::GetDirectoryName(
+    $myInvocation.MyCommand.Definition)
+
+$testRbdWnbd = "$scriptLocation/test_rbd_wnbd.py"
+
+function safe_exec() {
+    # Powershell doesn't check the command exit code, we'll need to
+    # do it ourselves. Also, in case of native commands, it treats stderr
+    # output as an exception, which is why we'll have to capture it.
+    cmd /c "$args 2>&1"
+    if ($LASTEXITCODE) {
+        throw "Command failed: $args"
+    }
+}
+
+safe_exec python.exe $testRbdWnbd --test-name RbdTest --iterations 100
+safe_exec python.exe $testRbdWnbd --test-name RbdFioTest --iterations 100
+safe_exec python.exe $testRbdWnbd --test-name RbdStampTest --iterations 100
+
+# It can take a while to setup the partition (~10s), we'll use fewer iterations.
+safe_exec python.exe $testRbdWnbd --test-name RbdFsTest --iterations 4
+safe_exec python.exe $testRbdWnbd --test-name RbdFsFioTest --iterations 4
+safe_exec python.exe $testRbdWnbd --test-name RbdFsStampTest --iterations 4
+
+# The live resize feature is not available in Quincy. Uncomment the following if
+# it gets backported.
+# safe_exec python.exe $testRbdWnbd `
+#     --test-name RbdResizeFioTest --image-size-mb 64
diff --git a/qa/workunits/windows/test_rbd_wnbd.py b/qa/workunits/windows/test_rbd_wnbd.py
new file mode 100644 (file)
index 0000000..f22810e
--- /dev/null
@@ -0,0 +1,919 @@
+import argparse
+import collections
+import functools
+import json
+import logging
+import math
+import os
+import prettytable
+import random
+import subprocess
+import time
+import threading
+import typing
+import uuid
+from concurrent import futures
+
+LOG = logging.getLogger()
+
+parser = argparse.ArgumentParser(description='rbd-wnbd tests')
+parser.add_argument('--test-name',
+                    help='The test to be run.',
+                    default="RbdFioTest")
+parser.add_argument('--iterations',
+                    help='Total number of test iterations',
+                    default=1, type=int)
+parser.add_argument('--concurrency',
+                    help='The number of tests to run in parallel',
+                    default=4, type=int)
+parser.add_argument('--fio-iterations',
+                    help='Total number of benchmark iterations per disk.',
+                    default=1, type=int)
+parser.add_argument('--fio-workers',
+                    help='Total number of fio workers per disk.',
+                    default=1, type=int)
+parser.add_argument('--fio-depth',
+                    help='The number of concurrent asynchronous operations '
+                         'executed per disk',
+                    default=64, type=int)
+parser.add_argument('--fio-verify',
+                    help='The mechanism used to validate the written '
+                         'data. Examples: crc32c, md5, sha1, null, etc. '
+                         'If set to null, the written data will not be '
+                         'verified.',
+                    default='crc32c')
+parser.add_argument('--bs',
+                    help='Benchmark block size.',
+                    default="2M")
+parser.add_argument('--op',
+                    help='Benchmark operation. '
+                         'Examples: read, randwrite, rw, etc.',
+                    default="rw")
+parser.add_argument('--image-prefix',
+                    help='The image name prefix.',
+                    default="cephTest-")
+parser.add_argument('--image-size-mb',
+                    help='The image size in megabytes.',
+                    default=1024, type=int)
+parser.add_argument('--map-timeout',
+                    help='Image map timeout.',
+                    default=60, type=int)
+parser.add_argument('--skip-enabling-disk', action='store_true',
+                    help='If set, the disk will not be turned online and the '
+                         'read-only flag will not be removed. Useful when '
+                         'the SAN policy is set to "onlineAll".')
+parser.add_argument('--verbose', action='store_true',
+                    help='Print info messages.')
+parser.add_argument('--debug', action='store_true',
+                    help='Print debug messages.')
+parser.add_argument('--stop-on-error', action='store_true',
+                    help='Stop testing when hitting errors.')
+parser.add_argument('--skip-cleanup-on-error', action='store_true',
+                    help='Skip cleanup when hitting errors.')
+
+
+class CephTestException(Exception):
+    msg_fmt = "An exception has been encountered."
+
+    def __init__(self, message: str = None, **kwargs):
+        self.kwargs = kwargs
+        if not message:
+            message = self.msg_fmt % kwargs
+        self.message = message
+        super(CephTestException, self).__init__(message)
+
+
+class CommandFailed(CephTestException):
+    msg_fmt = (
+        "Command failed: %(command)s. "
+        "Return code: %(returncode)s. "
+        "Stdout: %(stdout)s. Stderr: %(stderr)s.")
+
+
+class CephTestTimeout(CephTestException):
+    msg_fmt = "Operation timeout."
+
+
+def setup_logging(log_level: int = logging.INFO):
+    handler = logging.StreamHandler()
+    handler.setLevel(log_level)
+
+    log_fmt = '[%(asctime)s] %(levelname)s - %(message)s'
+    formatter = logging.Formatter(log_fmt)
+    handler.setFormatter(formatter)
+
+    LOG.addHandler(handler)
+    LOG.setLevel(logging.DEBUG)
+
+
+def retry_decorator(timeout: int = 60,
+                    retry_interval: int = 2,
+                    silent_interval: int = 10,
+                    additional_details: str = "",
+                    retried_exceptions:
+                        typing.Union[
+                            typing.Type[Exception],
+                            collections.abc.Iterable[
+                                typing.Type[Exception]]] = Exception):
+    def wrapper(f: typing.Callable[..., typing.Any]):
+        @functools.wraps(f)
+        def inner(*args, **kwargs):
+            tstart: float = time.time()
+            elapsed: float = 0
+            exc = None
+            details = additional_details or "%s failed" % f.__qualname__
+
+            while elapsed < timeout or not timeout:
+                try:
+                    return f(*args, **kwargs)
+                except retried_exceptions as ex:
+                    exc = ex
+                    elapsed = time.time() - tstart
+                    if elapsed > silent_interval:
+                        level = logging.WARNING
+                    else:
+                        level = logging.DEBUG
+                    LOG.log(level,
+                            "Exception: %s. Additional details: %s. "
+                            "Time elapsed: %d. Timeout: %d",
+                            ex, details, elapsed, timeout)
+
+                    time.sleep(retry_interval)
+                    elapsed = time.time() - tstart
+
+            msg = (
+                "Operation timed out. Exception: %s. Additional details: %s. "
+                "Time elapsed: %d. Timeout: %d.")
+            raise CephTestTimeout(
+                msg % (exc, details, elapsed, timeout))
+        return inner
+    return wrapper
+
+
+def execute(*args, **kwargs):
+    LOG.debug("Executing: %s", args)
+    result = subprocess.run(
+        args,
+        stdout=subprocess.PIPE,
+        stderr=subprocess.PIPE,
+        **kwargs)
+    LOG.debug("Command %s returned %d.", args, result.returncode)
+    if result.returncode:
+        exc = CommandFailed(
+            command=args, returncode=result.returncode,
+            stdout=result.stdout, stderr=result.stderr)
+        LOG.error(exc)
+        raise exc
+    return result
+
+
+def ps_execute(*args, **kwargs):
+    # Disable PS progress bar, causes issues when invoked remotely.
+    prefix = "$global:ProgressPreference = 'SilentlyContinue' ; "
+    return execute(
+        "powershell.exe", "-NonInteractive",
+        "-Command", prefix, *args, **kwargs)
+
+
+def array_stats(array: list):
+    mean = sum(array) / len(array) if len(array) else 0
+    variance = (sum((i - mean) ** 2 for i in array) / len(array)
+                if len(array) else 0)
+    std_dev = math.sqrt(variance)
+    sorted_array = sorted(array)
+
+    return {
+        'min': min(array) if len(array) else 0,
+        'max': max(array) if len(array) else 0,
+        'sum': sum(array) if len(array) else 0,
+        'mean': mean,
+        'median': sorted_array[len(array) // 2] if len(array) else 0,
+        'max_90': sorted_array[int(len(array) * 0.9)] if len(array) else 0,
+        'min_90': sorted_array[int(len(array) * 0.1)] if len(array) else 0,
+        'variance': variance,
+        'std_dev': std_dev,
+        'count': len(array)
+    }
+
+
+class Tracer:
+    data: collections.OrderedDict = collections.OrderedDict()
+    lock = threading.Lock()
+
+    @classmethod
+    def trace(cls, func):
+        def wrapper(*args, **kwargs):
+            tstart = time.time()
+            exc_str = None
+
+            # Preserve call order
+            with cls.lock:
+                if func.__qualname__ not in cls.data:
+                    cls.data[func.__qualname__] = list()
+
+            try:
+                return func(*args, **kwargs)
+            except Exception as exc:
+                exc_str = str(exc)
+                raise
+            finally:
+                tend = time.time()
+
+                with cls.lock:
+                    cls.data[func.__qualname__] += [{
+                        "duration": tend - tstart,
+                        "error": exc_str,
+                    }]
+
+        return wrapper
+
+    @classmethod
+    def get_results(cls):
+        stats = collections.OrderedDict()
+        for f in cls.data.keys():
+            stats[f] = array_stats([i['duration'] for i in cls.data[f]])
+            errors = []
+            for i in cls.data[f]:
+                if i['error']:
+                    errors.append(i['error'])
+
+            stats[f]['errors'] = errors
+        return stats
+
+    @classmethod
+    def print_results(cls):
+        r = cls.get_results()
+
+        table = prettytable.PrettyTable(title="Duration (s)")
+        table.field_names = [
+            "function", "min", "max", "total",
+            "mean", "median", "std_dev",
+            "max 90%", "min 90%", "count", "errors"]
+        table.float_format = ".4"
+        for f, s in r.items():
+            table.add_row([f, s['min'], s['max'], s['sum'],
+                           s['mean'], s['median'], s['std_dev'],
+                           s['max_90'], s['min_90'],
+                           s['count'], len(s['errors'])])
+        print(table)
+
+
+class RbdImage(object):
+    def __init__(self,
+                 name: str,
+                 size_mb: int,
+                 is_shared: bool = True,
+                 disk_number: int = -1,
+                 mapped: bool = False):
+        self.name = name
+        self.size_mb = size_mb
+        self.is_shared = is_shared
+        self.disk_number = disk_number
+        self.mapped = mapped
+        self.removed = False
+        self.drive_letter = ""
+
+    @classmethod
+    @Tracer.trace
+    def create(cls,
+               name: str,
+               size_mb: int = 1024,
+               is_shared: bool = True):
+        LOG.info("Creating image: %s. Size: %s.", name, "%sM" % size_mb)
+        cmd = ["rbd", "create", name, "--size", "%sM" % size_mb]
+        if is_shared:
+            cmd += ["--image-shared"]
+        execute(*cmd)
+
+        return RbdImage(name, size_mb, is_shared)
+
+    @Tracer.trace
+    def get_disk_number(self,
+                        timeout: int = 60,
+                        retry_interval: int = 2):
+        @retry_decorator(
+            retried_exceptions=CephTestException,
+            timeout=timeout,
+            retry_interval=retry_interval)
+        def _get_disk_number():
+            LOG.info("Retrieving disk number: %s", self.name)
+
+            result = execute("rbd-wnbd", "show", self.name, "--format=json")
+            disk_info = json.loads(result.stdout)
+            disk_number = disk_info["disk_number"]
+            if disk_number > 0:
+                LOG.debug("Image %s disk number: %d", self.name, disk_number)
+                return disk_number
+
+            raise CephTestException(
+                f"Could not get disk number: {self.name}.")
+
+        return _get_disk_number()
+
+    @Tracer.trace
+    def _wait_for_disk(self,
+                       timeout: int = 60,
+                       retry_interval: int = 2):
+        @retry_decorator(
+            retried_exceptions=(FileNotFoundError, OSError),
+            additional_details="the mapped disk isn't available yet",
+            timeout=timeout,
+            retry_interval=retry_interval)
+        def wait_for_disk():
+            LOG.debug("Waiting for disk to be accessible: %s %s",
+                      self.name, self.path)
+
+            with open(self.path, 'rb'):
+                pass
+
+        return wait_for_disk()
+
+    @property
+    def path(self):
+        return f"\\\\.\\PhysicalDrive{self.disk_number}"
+
+    @Tracer.trace
+    @retry_decorator(additional_details="couldn't clear disk read-only flag")
+    def set_writable(self):
+        ps_execute(
+            "Set-Disk", "-Number", str(self.disk_number),
+            "-IsReadOnly", "$false")
+
+    @Tracer.trace
+    @retry_decorator(additional_details="couldn't bring the disk online")
+    def set_online(self):
+        ps_execute(
+            "Set-Disk", "-Number", str(self.disk_number),
+            "-IsOffline", "$false")
+
+    @Tracer.trace
+    def map(self, timeout: int = 60):
+        LOG.info("Mapping image: %s", self.name)
+        tstart = time.time()
+
+        execute("rbd-wnbd", "map", self.name)
+        self.mapped = True
+
+        self.disk_number = self.get_disk_number(timeout=timeout)
+
+        elapsed = time.time() - tstart
+        self._wait_for_disk(timeout=timeout - elapsed)
+
+    @Tracer.trace
+    def unmap(self):
+        if self.mapped:
+            LOG.info("Unmapping image: %s", self.name)
+            execute("rbd-wnbd", "unmap", self.name)
+            self.mapped = False
+
+    @Tracer.trace
+    def remove(self):
+        if not self.removed:
+            LOG.info("Removing image: %s", self.name)
+            execute("rbd", "rm", self.name)
+            self.removed = True
+
+    def cleanup(self):
+        try:
+            self.unmap()
+        finally:
+            self.remove()
+
+    @Tracer.trace
+    @retry_decorator()
+    def _init_disk(self):
+        cmd = f"Get-Disk -Number {self.disk_number} | Initialize-Disk"
+        ps_execute(cmd)
+
+    @Tracer.trace
+    @retry_decorator()
+    def _create_partition(self):
+        cmd = (f"Get-Disk -Number {self.disk_number} | "
+               "New-Partition -AssignDriveLetter -UseMaximumSize")
+        ps_execute(cmd)
+
+    @Tracer.trace
+    @retry_decorator()
+    def _format_volume(self):
+        cmd = (
+            f"(Get-Partition -DiskNumber {self.disk_number}"
+            " | ? { $_.DriveLetter }) | Format-Volume -Force -Confirm:$false")
+        ps_execute(cmd)
+
+    @Tracer.trace
+    @retry_decorator()
+    def _get_drive_letter(self):
+        cmd = (f"(Get-Partition -DiskNumber {self.disk_number}"
+               " | ? { $_.DriveLetter }).DriveLetter")
+        result = ps_execute(cmd)
+
+        # The PowerShell command will place a null character if no drive letter
+        # is available. For example, we can receive "\x00\r\n".
+        self.drive_letter = result.stdout.decode().strip()
+        if not self.drive_letter.isalpha() or len(self.drive_letter) != 1:
+            raise CephTestException(
+                "Invalid drive letter received: %s" % self.drive_letter)
+
+    @Tracer.trace
+    def init_fs(self):
+        if not self.mapped:
+            raise CephTestException("Unable to create fs, image not mapped.")
+
+        LOG.info("Initializing fs, image: %s.", self.name)
+
+        self._init_disk()
+        self._create_partition()
+        self._format_volume()
+        self._get_drive_letter()
+
+    @Tracer.trace
+    def get_fs_capacity(self):
+        if not self.drive_letter:
+            raise CephTestException("No drive letter available")
+
+        cmd = f"(Get-Volume -DriveLetter {self.drive_letter}).Size"
+        result = ps_execute(cmd)
+
+        return int(result.stdout.decode().strip())
+
+    @Tracer.trace
+    def resize(self, new_size_mb, allow_shrink=False):
+        LOG.info(
+            "Resizing image: %s. New size: %s MB, old size: %s MB",
+            self.name, new_size_mb, self.size_mb)
+
+        cmd = ["rbd", "resize", self.name,
+               "--size", f"{new_size_mb}M", "--no-progress"]
+        if allow_shrink:
+            cmd.append("--allow-shrink")
+
+        execute(*cmd)
+
+        self.size_mb = new_size_mb
+
+    @Tracer.trace
+    def get_disk_size(self):
+        """Retrieve the virtual disk size (bytes) reported by Windows."""
+        cmd = f"(Get-Disk -Number {self.disk_number}).Size"
+        result = ps_execute(cmd)
+
+        disk_size = result.stdout.decode().strip()
+        if not disk_size.isdigit():
+            raise CephTestException(
+                "Invalid disk size received: %s" % disk_size)
+
+        return int(disk_size)
+
+    @Tracer.trace
+    @retry_decorator(timeout=30)
+    def wait_for_disk_resize(self):
+        # After resizing the rbd image, the daemon is expected to receive
+        # the notification, inform the WNBD driver and then trigger a disk
+        # rescan (IOCTL_DISK_UPDATE_PROPERTIES). This might take a few seconds,
+        # so we'll need to do some polling.
+        disk_size = self.get_disk_size()
+        disk_size_mb = disk_size // (1 << 20)
+
+        if disk_size_mb != self.size_mb:
+            raise CephTestException(
+                "The disk size hasn't been updated yet. Retrieved size: "
+                f"{disk_size_mb}MB. Expected size: {self.size_mb}MB.")
+
+
+class RbdTest(object):
+    image: RbdImage
+
+    requires_disk_online = False
+    requires_disk_write = False
+
+    def __init__(self,
+                 image_prefix: str = "cephTest-",
+                 image_size_mb: int = 1024,
+                 map_timeout: int = 60,
+                 **kwargs):
+        self.image_size_mb = image_size_mb
+        self.image_name = image_prefix + str(uuid.uuid4())
+        self.map_timeout = map_timeout
+        self.skip_enabling_disk = kwargs.get("skip_enabling_disk")
+
+    @Tracer.trace
+    def initialize(self):
+        self.image = RbdImage.create(
+            self.image_name,
+            self.image_size_mb)
+        self.image.map(timeout=self.map_timeout)
+
+        if not self.skip_enabling_disk:
+            if self.requires_disk_write:
+                self.image.set_writable()
+
+            if self.requires_disk_online:
+                self.image.set_online()
+
+    def run(self):
+        pass
+
+    def cleanup(self):
+        if self.image:
+            self.image.cleanup()
+
+    @classmethod
+    def print_results(cls,
+                      title: str = "Test results",
+                      description: str = None):
+        pass
+
+
+class RbdFsTestMixin(object):
+    # Windows disks must be turned online before accessing partitions.
+    requires_disk_online = True
+    requires_disk_write = True
+
+    @Tracer.trace
+    def initialize(self):
+        super(RbdFsTestMixin, self).initialize()
+
+        self.image.init_fs()
+
+    def get_subpath(self, *args):
+        drive_path = f"{self.image.drive_letter}:\\"
+        return os.path.join(drive_path, *args)
+
+
+class RbdFsTest(RbdFsTestMixin, RbdTest):
+    pass
+
+
+class RbdFioTest(RbdTest):
+    data: typing.DefaultDict[str, typing.List[typing.Dict[str, str]]] = (
+        collections.defaultdict(list))
+    lock = threading.Lock()
+
+    def __init__(self,
+                 *args,
+                 fio_size_mb: int = None,
+                 iterations: int = 1,
+                 workers: int = 1,
+                 bs: str = "2M",
+                 iodepth: int = 64,
+                 op: str = "rw",
+                 verify: str = "crc32c",
+                 **kwargs):
+
+        super(RbdFioTest, self).__init__(*args, **kwargs)
+
+        self.fio_size_mb = fio_size_mb or self.image_size_mb
+        self.iterations = iterations
+        self.workers = workers
+        self.bs = bs
+        self.iodepth = iodepth
+        self.op = op
+        if op not in ("read", "randread"):
+            self.requires_disk_write = True
+        self.verify = verify
+
+    def process_result(self, raw_fio_output: str):
+        result = json.loads(raw_fio_output)
+        with self.lock:
+            for job in result["jobs"]:
+                # Fio doesn't support trim on Windows
+                for op in ['read', 'write']:
+                    if op in job:
+                        self.data[op].append({
+                            'error': job['error'],
+                            'io_bytes': job[op]['io_bytes'],
+                            'bw_bytes': job[op]['bw_bytes'],
+                            'runtime': job[op]['runtime'] / 1000,  # seconds
+                            'total_ios': job[op]['short_ios'],
+                            'short_ios': job[op]['short_ios'],
+                            'dropped_ios': job[op]['short_ios'],
+                            'clat_ns_min': job[op]['clat_ns']['min'],
+                            'clat_ns_max': job[op]['clat_ns']['max'],
+                            'clat_ns_mean': job[op]['clat_ns']['mean'],
+                            'clat_ns_stddev': job[op]['clat_ns']['stddev'],
+                            'clat_ns_10': job[op].get('clat_ns', {})
+                                                 .get('percentile', {})
+                                                 .get('10.000000', 0),
+                            'clat_ns_90': job[op].get('clat_ns', {})
+                                                 .get('percentile', {})
+                                                 .get('90.000000', 0)
+                        })
+
+    def _get_fio_path(self):
+        return self.image.path
+
+    @Tracer.trace
+    def _run_fio(self, fio_size_mb=None):
+        LOG.info("Starting FIO test.")
+        cmd = [
+            "fio", "--thread", "--output-format=json",
+            "--randrepeat=%d" % self.iterations,
+            "--direct=1", "--name=test",
+            "--bs=%s" % self.bs, "--iodepth=%s" % self.iodepth,
+            "--size=%sM" % (fio_size_mb or self.fio_size_mb),
+            "--readwrite=%s" % self.op,
+            "--numjobs=%s" % self.workers,
+            "--filename=%s" % self._get_fio_path(),
+        ]
+        if self.verify:
+            cmd += ["--verify=%s" % self.verify]
+        result = execute(*cmd)
+        LOG.info("Completed FIO test.")
+        self.process_result(result.stdout)
+
+    @Tracer.trace
+    def run(self):
+        self._run_fio()
+
+    @classmethod
+    def print_results(cls,
+                      title: str = "Benchmark results",
+                      description: str = None):
+        if description:
+            title = "%s (%s)" % (title, description)
+
+        for op in cls.data.keys():
+            op_title = "%s op=%s" % (title, op)
+
+            table = prettytable.PrettyTable(title=op_title)
+            table.field_names = ["stat", "min", "max", "mean",
+                                 "median", "std_dev",
+                                 "max 90%", "min 90%", "total"]
+            table.float_format = ".4"
+
+            op_data = cls.data[op]
+
+            s = array_stats([float(i["bw_bytes"]) / 1000_000 for i in op_data])
+            table.add_row(["bandwidth (MB/s)",
+                           s['min'], s['max'], s['mean'],
+                           s['median'], s['std_dev'],
+                           s['max_90'], s['min_90'], 'N/A'])
+
+            s = array_stats([float(i["runtime"]) for i in op_data])
+            table.add_row(["duration (s)",
+                          s['min'], s['max'], s['mean'],
+                          s['median'], s['std_dev'],
+                          s['max_90'], s['min_90'], s['sum']])
+
+            s = array_stats([i["error"] for i in op_data])
+            table.add_row(["errors",
+                           s['min'], s['max'], s['mean'],
+                           s['median'], s['std_dev'],
+                           s['max_90'], s['min_90'], s['sum']])
+
+            s = array_stats([i["short_ios"] for i in op_data])
+            table.add_row(["incomplete IOs",
+                           s['min'], s['max'], s['mean'],
+                           s['median'], s['std_dev'],
+                           s['max_90'], s['min_90'], s['sum']])
+
+            s = array_stats([i["dropped_ios"] for i in op_data])
+            table.add_row(["dropped IOs",
+                           s['min'], s['max'], s['mean'],
+                           s['median'], s['std_dev'],
+                           s['max_90'], s['min_90'], s['sum']])
+
+            clat_min = array_stats([i["clat_ns_min"] for i in op_data])
+            clat_max = array_stats([i["clat_ns_max"] for i in op_data])
+            clat_mean = array_stats([i["clat_ns_mean"] for i in op_data])
+            clat_stddev = math.sqrt(
+                sum([float(i["clat_ns_stddev"]) ** 2 for i in op_data]) / len(op_data)
+                if len(op_data) else 0)
+            clat_10 = array_stats([i["clat_ns_10"] for i in op_data])
+            clat_90 = array_stats([i["clat_ns_90"] for i in op_data])
+            # For convenience, we'll convert it from ns to seconds.
+            table.add_row(["completion latency (s)",
+                           clat_min['min'] / 1e+9,
+                           clat_max['max'] / 1e+9,
+                           clat_mean['mean'] / 1e+9,
+                           clat_mean['median'] / 1e+9,
+                           clat_stddev / 1e+9,
+                           clat_10['mean'] / 1e+9,
+                           clat_90['mean'] / 1e+9,
+                           clat_mean['sum'] / 1e+9])
+            print(table)
+
+
+class RbdResizeFioTest(RbdFioTest):
+    """Image resize test.
+
+    This test extends and then shrinks the image, performing FIO tests to
+    validate the resized image.
+    """
+
+    @Tracer.trace
+    def run(self):
+        self.image.resize(self.image_size_mb * 2)
+        self.image.wait_for_disk_resize()
+
+        self._run_fio(fio_size_mb=self.image_size_mb * 2)
+
+        self.image.resize(self.image_size_mb // 2, allow_shrink=True)
+        self.image.wait_for_disk_resize()
+
+        self._run_fio(fio_size_mb=self.image_size_mb // 2)
+
+        # Just like rbd-nbd, rbd-wnbd is masking out-of-bounds errors.
+        # For this reason, we don't have a negative test that writes
+        # passed the disk boundary.
+
+
+class RbdFsFioTest(RbdFsTestMixin, RbdFioTest):
+    def initialize(self):
+        super(RbdFsFioTest, self).initialize()
+
+        if not self.fio_size_mb or self.fio_size_mb == self.image_size_mb:
+            # Out of caution, we'll use up to 80% of the FS by default
+            self.fio_size_mb = int(
+                self.image.get_fs_capacity() * 0.8 / (1024 * 1024))
+
+    @staticmethod
+    def _fio_escape_path(path):
+        # FIO allows specifying multiple files separated by colon.
+        # This means that ":" has to be escaped, so
+        # F:\filename becomes F\:\filename.
+        return path.replace(":", "\\:")
+
+    def _get_fio_path(self):
+        return self._fio_escape_path(self.get_subpath("test-fio"))
+
+
+class RbdStampTest(RbdTest):
+    requires_disk_write = True
+
+    _write_open_mode = "rb+"
+    _read_open_mode = "rb"
+    _expect_path_exists = True
+
+    @staticmethod
+    def _rand_float(min_val: float, max_val: float):
+        return min_val + (random.random() * max_val - min_val)
+
+    def _get_stamp(self):
+        buff = self.image_name.encode()
+        padding = 512 - len(buff)
+        buff += b'\0' * padding
+        return buff
+
+    def _get_stamp_path(self):
+        return self.image.path
+
+    @Tracer.trace
+    def _write_stamp(self):
+        with open(self._get_stamp_path(), self._write_open_mode) as disk:
+            stamp = self._get_stamp()
+            disk.write(stamp)
+
+    @Tracer.trace
+    def _read_stamp(self):
+        with open(self._get_stamp_path(), self._read_open_mode) as disk:
+            return disk.read(len(self._get_stamp()))
+
+    @Tracer.trace
+    def run(self):
+        if self._expect_path_exists:
+            # Wait up to 5 seconds and then check the disk, ensuring that
+            # nobody else wrote to it. This is particularly useful when
+            # running a high number of tests in parallel, ensuring that
+            # we aren't writing to the wrong disk.
+            time.sleep(self._rand_float(0, 5))
+
+            stamp = self._read_stamp()
+            assert stamp == b'\0' * len(self._get_stamp())
+
+        self._write_stamp()
+
+        stamp = self._read_stamp()
+        assert stamp == self._get_stamp()
+
+
+class RbdFsStampTest(RbdFsTestMixin, RbdStampTest):
+    _write_open_mode = "wb"
+    _expect_path_exists = False
+
+    def _get_stamp_path(self):
+        return self.get_subpath("test-stamp")
+
+
+class TestRunner(object):
+    def __init__(self,
+                 test_cls: typing.Type[RbdTest],
+                 test_params: dict = {},
+                 iterations: int = 1,
+                 workers: int = 1,
+                 stop_on_error: bool = False,
+                 cleanup_on_error: bool = True):
+        self.test_cls = test_cls
+        self.test_params = test_params
+        self.iterations = iterations
+        self.workers = workers
+        self.executor = futures.ThreadPoolExecutor(max_workers=workers)
+        self.lock = threading.Lock()
+        self.completed = 0
+        self.errors = 0
+        self.stopped = False
+        self.stop_on_error = stop_on_error
+        self.cleanup_on_error = cleanup_on_error
+
+    @Tracer.trace
+    def run(self):
+        tasks = []
+        for i in range(self.iterations):
+            task = self.executor.submit(self.run_single_test)
+            tasks.append(task)
+
+        LOG.info("Waiting for %d tests to complete.", self.iterations)
+        for task in tasks:
+            task.result()
+
+    def run_single_test(self):
+        failed = False
+        if self.stopped:
+            return
+
+        try:
+            test = self.test_cls(**self.test_params)
+            test.initialize()
+            test.run()
+        except KeyboardInterrupt:
+            LOG.warning("Received Ctrl-C.")
+            self.stopped = True
+        except Exception as ex:
+            failed = True
+            if self.stop_on_error:
+                self.stopped = True
+            with self.lock:
+                self.errors += 1
+                LOG.exception(
+                    "Test exception: %s. Total exceptions: %d",
+                    ex, self.errors)
+        finally:
+            if not failed or self.cleanup_on_error:
+                try:
+                    test.cleanup()
+                except KeyboardInterrupt:
+                    LOG.warning("Received Ctrl-C.")
+                    self.stopped = True
+                    # Retry the cleanup
+                    test.cleanup()
+                except Exception:
+                    LOG.exception("Test cleanup failed.")
+
+            with self.lock:
+                self.completed += 1
+                LOG.info("Completed tests: %d. Pending: %d",
+                         self.completed, self.iterations - self.completed)
+
+
+TESTS: typing.Dict[str, typing.Type[RbdTest]] = {
+    'RbdTest': RbdTest,
+    'RbdFioTest': RbdFioTest,
+    'RbdResizeFioTest': RbdResizeFioTest,
+    'RbdStampTest': RbdStampTest,
+    # FS tests
+    'RbdFsTest': RbdFsTest,
+    'RbdFsFioTest': RbdFsFioTest,
+    'RbdFsStampTest': RbdFsStampTest,
+}
+
+if __name__ == '__main__':
+    args = parser.parse_args()
+
+    log_level = logging.WARNING
+    if args.verbose:
+        log_level = logging.INFO
+    if args.debug:
+        log_level = logging.DEBUG
+    setup_logging(log_level)
+
+    test_params = dict(
+        image_size_mb=args.image_size_mb,
+        image_prefix=args.image_prefix,
+        bs=args.bs,
+        op=args.op,
+        verify=args.fio_verify,
+        iodepth=args.fio_depth,
+        map_timeout=args.map_timeout,
+        skip_enabling_disk=args.skip_enabling_disk,
+    )
+
+    try:
+        test_cls = TESTS[args.test_name]
+    except KeyError:
+        raise CephTestException("Unkown test: {}".format(args.test_name))
+
+    runner = TestRunner(
+        test_cls,
+        test_params=test_params,
+        iterations=args.iterations,
+        workers=args.concurrency,
+        stop_on_error=args.stop_on_error,
+        cleanup_on_error=not args.skip_cleanup_on_error)
+    runner.run()
+
+    Tracer.print_results()
+    test_cls.print_results(
+        description="count: %d, concurrency: %d" %
+        (args.iterations, args.concurrency))
+
+    assert runner.errors == 0, f"encountered {runner.errors} error(s)."