facebook/sapling
1
1
Fork 0
mirror of https://github.com/facebook/sapling.git synced 2024-10-05 14:28:17 +03:00
Code Issues Projects Releases Wiki Activity
56e2440c1d
sapling/eden/integration/takeover_test.py
Katie Mancini c97abca976 fix graceful restart crash bundle 
Summary:
The current serialization type we use for the graceful restart protocol only
allows us to send mount specific information in the non error case.

If we want to send general takeover information, like the order of the file
descriptors we send along with the takeover message, we are out of luck.

I am changing the final type we send during takeover to be a struct so that
we can add non-mount specific information. This allows us to create an explicit
order of our file descriptors and will let us send a dynamic number of general
file descriptors.

Currently, we have to send all the expected general file descriptors
(thift socket, lock file, mountd socket) in a fixed order.
And those file descriptors have to be valid because we cannot send an invalid
file descriptor with sendmsg.

When NFS is not enabled we do not initialize the mountd socket so we cannot
send a valid filedescriptor for the mountd socket (unless we sent some dummy
file descriptor which feels really hacky and not a good permanent solution).

To be able to fix the bug we have in graceful restart due to not having
a mountd socket we need to be able to send only a portion of the general
file descriptors.

When NFS is not enabled, we do not initialize the NFS Server. This causes
a seg fault during takeover because we try to shutdown the nfs server.

We should only shutdown the NFS Server when we initialized it. This
means we only pass the mountd server socket during takeover when NFS was enabled
when we started the server.

To allow use to enable the NFS Server across graceful restart we only
perform a takeover start when the mountd server socket was received during
takeover.

Reviewed By: xavierd

Differential Revision: D33755447

fbshipit-source-id: cc471c3051de8e41c21a4d1d8e42d93d5295e90e
2022-01-25 17:25:37 -08:00

502 lines
19 KiB
Python
Raw Blame History

#!/usr/bin/env python3
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This software may be used and distributed according to the terms of the
# GNU General Public License version 2.
import os
import signal
import sys
import threading
from multiprocessing import Process
from pathlib import Path
from typing import List, Dict, Optional
import pexpect
from eden.fs.cli.util import get_pid_using_lockfile, poll_until
from eden.thrift.legacy import EdenClient
from facebook.eden.ttypes import FaultDefinition, UnblockFaultArg
from fb303_core.ttypes import fb303_status
from .lib import testcase
from .lib.find_executables import FindExe
# pyre-ignore[13]: T62487924
class TakeoverTestBase(testcase.EdenRepoTest):
pagesize: int
page1: str
page2: str
commit1: str
commit2: str
enable_fault_injection: bool = True
def populate_repo(self) -> None:
import resource
self.pagesize = resource.getpagesize()
self.page1 = "1" * self.pagesize
self.page2 = "2" * self.pagesize
self.repo.write_file("tree/hello", self.page1 + self.page2)
self.repo.write_file("tree/deleted", self.page1 + self.page2)
self.repo.write_file("src/main.c", "hello world")
self.commit1 = self.repo.commit("Initial commit.")
self.repo.write_file("src/main.c", "hello world v2")
self.repo.write_file("src/test/test1.py", "test1")
self.repo.write_file("src/test/test2.py", "test2")
self.commit2 = self.repo.commit("Initial commit.")
def select_storage_engine(self) -> str:
"""we need to persist data across restarts"""
return "sqlite"
def edenfs_logging_settings(self) -> Dict[str, str]:
if self._testMethodName == "test_takeover_with_io":
# test_takeover_with_io causes lots of I/O, so do not enable
# verbose logging of I/O operations in this test.
return {}
return {
"eden.strace": "DBG7",
"eden.fs.fuse": "DBG7",
"eden.fs.inodes.InodeMap": "DBG6",
"eden.fs.takeover": "DBG7",
"eden.fs.service": "DBG4",
}
def do_takeover_test(self) -> None:
hello = os.path.join(self.mount, "tree/hello")
deleted = os.path.join(self.mount, "tree/deleted")
deleted_local = os.path.join(self.mount, "deleted-local")
# To test our handling of unlinked inodes, in addition
# to unlinking something that is in the manifest we
# need to check that we handle the case of a local
# file being deleted to make sure that we cover both
# code paths for FileInode.
with open(deleted_local, "w") as dl:
dl.write(self.page1)
dl.write(self.page2)
# We'd like to make sure that we do something reasonable
# for directories that have been unlinked and that are
# still referenced via a file descriptor. Ideally we'd call
# opendir() here and then readdir() it after we've performed
# the graceful restart, but we can't directly call those
# functions from python. The approach used here is to
# open a file descriptor to the directory and then try
# to stat() it after the restart. Since the directory
# has to be empty in order to be unlinked, a readdir
# from it wouldn't return any interesting results anyway.
deleted_dir = os.path.join(self.mount, "deleted-dir")
os.mkdir(deleted_dir)
deleted_dir_fd = os.open(deleted_dir, 0)
os.rmdir(deleted_dir)
with open(hello, "r") as f, open(deleted, "r") as d, open(
deleted_local, "r"
) as dl:
# Read the first page only (rather than the whole file)
# before we restart the process.
# This is so that we can check that the kernel really
# does call in to us for the second page and that we're
# really servicing the read for the second page and that
# it isn't just getting served from the kernel buffer cache
self.assertEqual(self.page1, f.read(self.pagesize))
# Let's make sure that unlinked inodes continue to
# work appropriately too. We've opened the file
# handles and are holding them alive in `d` and `dl`,
# so now let's unlink it from the filesystem
os.unlink(deleted)
os.unlink(deleted_local)
print("=== beginning restart ===", file=sys.stderr)
self.eden.graceful_restart()
print("=== restart complete ===", file=sys.stderr)
# Ensure that our file handle is still live across
# the restart boundary
f.seek(0)
self.assertEqual(self.page1, f.read(self.pagesize))
self.assertEqual(self.page2, f.read(self.pagesize))
# We should be able to read from the `d` file handle
# even though we deleted the file from the tree
self.assertEqual(self.page1, d.read(self.pagesize))
self.assertEqual(self.page2, d.read(self.pagesize))
# Likewise for the `dl` file handle
self.assertEqual(self.page1, dl.read(self.pagesize))
self.assertEqual(self.page2, dl.read(self.pagesize))
# Now check that the unlinked directory handle still seems
# connected. This is difficult to do directly in python;
# the directory had to be empty in order to be removed
# so even if we could read its directory entries there
# wouldn't be anything to read.
# Note that os.stat() will throw if the fd is deemed
# bad either by the kernel or the eden instance,
# so we're just calling it and discarding the return
# value.
os.stat(deleted_dir_fd)
os.close(deleted_dir_fd)
# Let's also test opening the same file up again,
# just to make sure that that is still working after
# the graceful restart.
with open(hello, "r") as f:
self.assertEqual(self.page1, f.read(self.pagesize))
self.assertEqual(self.page2, f.read(self.pagesize))
@testcase.eden_repo_test
class TakeoverTest(TakeoverTestBase):
def test_takeover(self) -> None:
return self.do_takeover_test()
def test_takeover_after_diff_revisions(self) -> None:
# Make a getScmStatusBetweenRevisions() call to Eden.
# Previously this thrift call caused Eden to create temporary inode
# objects outside of the normal root inode tree, and this would cause
# Eden to crash when shutting down afterwards.
with self.get_thrift_client() as client:
client.getScmStatusBetweenRevisions(
os.fsencode(self.mount),
self.commit1.encode("utf-8"),
self.commit2.encode("utf-8"),
)
return self.do_takeover_test()
def test_takeover_with_io(self) -> None:
num_threads = 4
write_chunk_size = 1024 * 1024
max_file_length = write_chunk_size * 100
# TODO: Setting this higher than 1 currently makes it likely that
# edenfs will crash during restart.
# There are still some other bugs we need to track down in the restart
# ordering.
num_restarts = 1
stop = threading.Event()
bufs = [b"x" * write_chunk_size, b"y" * write_chunk_size]
def do_io(thread_id: int, running_event: threading.Event) -> None:
path = os.path.join(self.mount, "src", "test", "data%d.log" % thread_id)
with open(path, "wb") as f:
# Use raw file descriptors to avoid going through python's I/O
# buffering code.
fd = f.fileno()
buf_idx = 0
buf = bufs[buf_idx]
offset = 0
# Repeatedly write and rewrite the same file,
# alternating between two different data buffers.
running_event.set()
while True:
os.pwrite(fd, buf, offset)
if stop.is_set():
return
offset += len(buf)
if offset >= max_file_length:
buf_idx += 1
buf = bufs[buf_idx % len(bufs)]
offset = 0
# Log the mount points device ID at the start of the test
# (Just in case anything hangs and we need to abort the mount
# using /sys/fs/fuse/connections/<dev>/)
st = os.lstat(self.mount)
print("=== eden mount device=%d ===" % st.st_dev, file=sys.stderr)
# Start several threads doing I/O while we we perform a takeover
threads = []
try:
running_events = []
for n in range(num_threads):
running = threading.Event()
thread = threading.Thread(target=do_io, args=(n, running))
thread.start()
threads.append(thread)
running_events.append(running)
# Wait until all threads have started and are doing I/O
for event in running_events:
event.wait()
# Restart edenfs
for n in range(num_restarts):
print("=== beginning restart %d ===" % n, file=sys.stderr)
self.eden.graceful_restart()
print("=== restart %d complete ===" % n, file=sys.stderr)
finally:
stop.set()
for thread in threads:
thread.join()
def test_takeover_updates_process_id_in_lock_file(self) -> None:
self.assertEqual(
self.eden.get_pid_via_thrift(),
get_pid_using_lockfile(Path(self.eden.eden_dir)),
)
self.eden.graceful_restart()
self.assertEqual(
self.eden.get_pid_via_thrift(),
get_pid_using_lockfile(Path(self.eden.eden_dir)),
)
def test_takeover_preserves_inode_numbers_for_open_nonmaterialized_files(
self,
) -> None:
hello = os.path.join(self.mount, "tree/hello")
fd = os.open(hello, os.O_RDONLY)
try:
inode_number = os.fstat(fd).st_ino
self.eden.graceful_restart()
self.assertEqual(inode_number, os.fstat(fd).st_ino)
finally:
os.close(fd)
fd = os.open(hello, os.O_RDONLY)
try:
self.assertEqual(inode_number, os.fstat(fd).st_ino)
finally:
os.close(fd)
def test_contents_are_the_same_if_handle_is_held_open(self) -> None:
with open(os.path.join(self.mount, "tree", "hello")) as c2_hello_file, open(
os.path.join(self.mount, "src", "main.c")
) as c2_mainc_file:
self.eden.graceful_restart()
self.eden.run_cmd(
"debug", "flush_cache", os.path.join("tree", "hello"), cwd=self.mount
)
self.eden.run_cmd(
"debug", "flush_cache", os.path.join("src", "main.c"), cwd=self.mount
)
self.assertEqual(self.page1 + self.page2, c2_hello_file.read())
self.assertEqual("hello world v2", c2_mainc_file.read())
def test_readdir_after_graceful_restart(self) -> None:
# Ensure capability flags (e.g. FUSE_NO_OPENDIR_SUPPORT) survive
# graceful restart
self.eden.graceful_restart()
self.assertEqual(
["test1.py", "test2.py"],
sorted(os.listdir(os.path.join(self.mount, "src", "test"))),
)
def test_readdir_before_and_after_graceful_restart(self) -> None:
self.assertEqual(
["test1.py", "test2.py"],
sorted(os.listdir(os.path.join(self.mount, "src", "test"))),
)
self.eden.graceful_restart()
self.assertEqual(
["test1.py", "test2.py"],
sorted(os.listdir(os.path.join(self.mount, "src", "test"))),
)
def test_takeover_doesnt_send_ping(self) -> None:
"""
tests that if we try a takeover with a version that doesn't know
how to accept a ping, we don't send one. This test should not fail
in either case since it is running against a client that knows how
to listen for a ping. It just is used to look at logs to make sure
the correct code path is entered.
"""
self.eden.fake_takeover_with_version(3)
def test_takeover_failure(self) -> None:
print("=== beginning restart ===", file=sys.stderr)
self.eden.takeover_without_ping_response()
print("=== restart complete ===", file=sys.stderr)
self.assertTrue(self.eden.wait_for_is_healthy())
def run_restart(self) -> "pexpect.spawn[bytes]":
edenfsctl, env = FindExe.get_edenfsctl_env()
restart_cmd = [
edenfsctl,
"--config-dir",
str(self.eden_dir),
"--etc-eden-dir",
str(self.etc_eden_dir),
"--home-dir",
str(self.home_dir),
"restart",
"--daemon-binary",
FindExe.FAKE_EDENFS,
]
print("Restarting eden: %r" % (restart_cmd,))
return pexpect.spawn(
restart_cmd[0],
restart_cmd[1:],
logfile=sys.stdout.buffer,
timeout=5,
env=env,
)
def assert_restart_fails_with_in_progress_graceful_restart(
self, client: EdenClient
) -> None:
pid = self.eden.get_pid_via_thrift()
p = self.run_restart()
p.expect_exact(
f"The current edenfs daemon (pid {pid}) is in the middle of stopping."
f"\r\nUse `eden restart --force` if you want to forcibly restart the current daemon\r\n"
)
p.wait()
self.assertEqual(p.exitstatus, 4)
self.assertEqual(client.getDaemonInfo().status, fb303_status.STOPPING)
def assert_shutdown_fails_with_in_progress_graceful_restart(
self, client: EdenClient
) -> None:
# call initiateShutdown. This should not throw.
try:
client.initiateShutdown("shutdown requested during graceful restart")
except Exception:
self.fail(
"initiateShutdown should not throw when graceful restart is in progress"
)
self.assertEqual(client.getDaemonInfo().status, fb303_status.STOPPING)
def assert_sigkill_fails_with_in_progress_graceful_restart(
self, client: EdenClient
) -> None:
# send SIGTERM to process. This should not throw.
pid = self.eden.get_pid_via_thrift()
try:
os.kill(pid, signal.SIGTERM)
except Exception:
self.fail(
"sending SIGTERM should not throw when graceful restart is in progress"
)
self.assertEqual(client.getDaemonInfo().status, fb303_status.STOPPING)
def test_stop_during_takeover(self) -> None:
# block graceful restart
with self.eden.get_thrift_client() as client:
client.injectFault(
FaultDefinition(
keyClass="takeover", keyValueRegex="server_shutdown", block=True
)
)
self.eden.wait_for_is_healthy()
# Run a graceful restart
# This won't succeed until we unblock the shutdown.
p = Process(target=self.eden.graceful_restart)
p.start()
# Wait for the state to be shutting down
def state_shutting_down() -> Optional[bool]:
if not p.is_alive():
raise Exception(
"eden restart --graceful command finished while "
"graceful restart was still blocked"
)
if client.getDaemonInfo().status is fb303_status.STOPPING:
return True
return None
poll_until(state_shutting_down, timeout=60)
# Normal restart should be rejected while a graceful restart
# is in progress
self.assert_restart_fails_with_in_progress_graceful_restart(client)
# Normal shutdown should be rejected while a graceful restart
# is in progress
self.assert_shutdown_fails_with_in_progress_graceful_restart(client)
# Getting SIGTERM should not kill process while a graceful restart is in
# progress
self.assert_sigkill_fails_with_in_progress_graceful_restart(client)
# Unblock the server shutdown and wait for the graceful restart to complete.
client.unblockFault(
UnblockFaultArg(keyClass="takeover", keyValueRegex="server_shutdown")
)
p.join()
@testcase.eden_repo_test(run_on_nfs=False)
class TakeoverTestNoNFSServer(TakeoverTestBase):
def edenfs_extra_config(self) -> Optional[Dict[str, List[str]]]:
return {}
def test_takeover(self) -> None:
return self.do_takeover_test()
@testcase.eden_repo_test
class TakeoverRocksDBStressTest(testcase.EdenRepoTest):
enable_fault_injection: bool = True
def populate_repo(self) -> None:
self.repo.write_file("test-directory/file", "")
# pyre-fixme[16]: `TakeoverRocksDBStressTest` has no attribute `commit1`.
self.commit1 = self.repo.commit("Initial commit.")
def select_storage_engine(self) -> str:
return "rocksdb"
def test_takeover_with_tree_inode_loading_from_local_store(self) -> None:
"""
Restart edenfs while a tree inode is being loaded asynchronously. Ensure
restarting does not deadlock.
"""
def load_test_directory_inode_from_local_store_asynchronously() -> None:
"""
Make edenfs start loading "/test-directory" from the local store.
To ensure that the local store is in use during takeover, load the tree
inode using a prefetch.
At the time of writing, os.listdir("foo") causes edenfs to prefetch
the tree inodes of foo/*. Exploit this to load the tree inode for
"/directory".
Other options considered:
* At the time of writing, if we load the tree inode using a FUSE
request (e.g. os.stat), edenfs would wait for the FUSE request to
finish before starting the inode shutdown procedure.
* At the time of writing, 'edenfsctl prefetch' does not prefetch
tree inodes asynchronously.
"""
os.listdir(self.mount)
graceful_restart_startup_time = 5.0
with self.eden.get_thrift_client() as client:
for key_class in ["local store get single", "local store get batch"]:
client.injectFault(
FaultDefinition(
keyClass=key_class,
keyValueRegex=".*",
delayMilliseconds=int(graceful_restart_startup_time * 1000),
count=100,
)
)
load_test_directory_inode_from_local_store_asynchronously()
self.eden.graceful_restart()
Reference in New Issue View Git Blame Copy Permalink