Merge branch 'master' into cve_fixes

Author: awherr

gprofiler/main.py

@@ -31,7 +31,7 @@
 
 import configargparse
 import humanfriendly
-from granulate_utils.linux.ns import is_running_in_init_pid
+from granulate_utils.linux.ns import is_root, is_running_in_init_pid
 from granulate_utils.linux.process import is_process_running
 from granulate_utils.metadata.cloud import get_aws_execution_env
 from psutil import NoSuchProcess, Process
@@ -70,12 +70,11 @@
     atomically_symlink,
     get_iso8601_format_time,
     grab_gprofiler_mutex,
-    is_root,
     reset_umask,
     resource_path,
     run_process,
 )
-from gprofiler.utils.fs import escape_filename, mkdir_owned_root
+from gprofiler.utils.fs import escape_filename, mkdir_owned_root_wrapper
 from gprofiler.utils.proxy import get_https_proxy
 
 if is_linux():
@@ -121,6 +120,7 @@ def __init__(
         output_dir: str,
         flamegraph: bool,
         rotating_output: bool,
+        rootless: bool,
         profiler_api_client: Optional[ProfilerAPIClient],
         collect_metrics: bool,
         collect_metadata: bool,
@@ -141,6 +141,7 @@ def __init__(
         self._output_dir = output_dir
         self._flamegraph = flamegraph
         self._rotating_output = rotating_output
+        self._rootless = rootless
         self._profiler_api_client = profiler_api_client
         self._state = state
         self._remote_logs_handler = remote_logs_handler
@@ -566,6 +567,15 @@ def parse_cmd_args() -> configargparse.Namespace:
         help="Comma separated list of processes that will be filtered to profile,"
         " given multiple times will append pids to one list",
     )
+    parser.add_argument(
+        "--rootless",
+        action="store_true",
+        default=False,
+        help="Run without root/sudo with limited functionality"
+        "Profiling is limted to only processes owned by this user that are passed with --pids. Logs and pid file "
+        "may be directed to user owned directory with --log-file and --pid-file respectively. Some additional "
+        "configuration (e.g. kernel.perf_event_paranoid) may be required to operate without root.",
+    )
 
     _add_profilers_arguments(parser)
 
@@ -903,8 +913,14 @@ def _add_profilers_arguments(parser: configargparse.ArgumentParser) -> None:
 
 
 def verify_preconditions(args: configargparse.Namespace, processes_to_profile: Optional[List[Process]]) -> None:
-    if not is_root():
-        print("Must run gprofiler as root, please re-run.", file=sys.stderr)
+    if not args.rootless and not is_root():
+        print("Not running as root, rerun with --rootless or as root.", file=sys.stderr)
+        sys.exit(1)
+    elif args.rootless and is_root():
+        print(
+            "Conflict, running with --rootless and as root, rerun with --rootless or as root (but not both).",
+            file=sys.stderr,
+        )
         sys.exit(1)
 
     if args.pid_ns_check and not is_running_in_init_pid():
@@ -1102,7 +1118,7 @@ def main() -> None:
                 )
                 sys.exit(1)
 
-        mkdir_owned_root(TEMPORARY_STORAGE_PATH)
+        mkdir_owned_root_wrapper(TEMPORARY_STORAGE_PATH)
 
         try:
             client_kwargs = {}
@@ -1154,6 +1170,7 @@ def main() -> None:
             output_dir=args.output_dir,
             flamegraph=args.flamegraph,
             rotating_output=args.rotating_output,
+            rootless=args.rootless,
             profiler_api_client=profiler_api_client,
             collect_metrics=args.collect_metrics,
             collect_metadata=args.collect_metadata,

gprofiler/metadata/metadata_collector.py

@@ -2,7 +2,7 @@
 from pathlib import Path
 from typing import Optional
 
-from granulate_utils.linux.ns import run_in_ns
+from granulate_utils.linux.ns import run_in_ns_wrapper
 from granulate_utils.metadata import Metadata
 from granulate_utils.metadata.bigdata import get_bigdata_info
 from granulate_utils.metadata.cloud import get_static_cloud_metadata
@@ -20,7 +20,7 @@ def get_static_metadata(spawn_time: float, run_args: UserArgs, external_metadata
     formatted_spawn_time = datetime.datetime.utcfromtimestamp(spawn_time).replace(microsecond=0).isoformat()
     static_system_metadata = get_static_system_info()
     cloud_metadata = get_static_cloud_metadata(logger)
-    bigdata = run_in_ns(["mnt"], get_bigdata_info)
+    bigdata = run_in_ns_wrapper(["mnt"], get_bigdata_info)
 
     metadata_dict: Metadata = {
         "cloud_provider": cloud_metadata.pop("provider") if cloud_metadata is not None else "unknown",

gprofiler/metadata/system_metadata.py

@@ -15,7 +15,7 @@
 
 import distro
 import psutil
-from granulate_utils.linux.ns import run_in_ns
+from granulate_utils.linux.ns import run_in_ns_wrapper
 
 from gprofiler.log import get_logger_adapter
 from gprofiler.platform import is_linux, is_windows
@@ -357,7 +357,7 @@ def get_infos() -> Any:
         except Exception:
             logger.exception("Failed to get the local IP")
 
-    run_in_ns(["mnt", "uts", "net"], get_infos)
+    run_in_ns_wrapper(["mnt", "uts", "net"], get_infos)
 
     return hostname, distribution, libc_version, mac_address, local_ip
 

gprofiler/metadata/versions.py

@@ -16,7 +16,7 @@
 from subprocess import CompletedProcess
 from threading import Event
 
-from granulate_utils.linux.ns import get_process_nspid, run_in_ns
+from granulate_utils.linux.ns import get_process_nspid, run_in_ns_wrapper
 from psutil import NoSuchProcess, Process
 
 from gprofiler.utils import run_process
@@ -38,7 +38,7 @@ def _run_get_version() -> "CompletedProcess[bytes]":
         return run_process([exe_path, version_arg], stop_event=stop_event, timeout=get_version_timeout)
 
     try:
-        cp = run_in_ns(["pid", "mnt"], _run_get_version, process.pid)
+        cp = run_in_ns_wrapper(["pid", "mnt"], _run_get_version, process.pid)
     except FileNotFoundError as e:
         if not process.is_running():
             raise NoSuchProcess(process.pid)

gprofiler/profilers/java.py

@@ -49,7 +49,12 @@
 if is_linux():
     from granulate_utils.linux import proc_events
     from granulate_utils.linux.kernel_messages import KernelMessage
-    from granulate_utils.linux.ns import get_proc_root_path, get_process_nspid, resolve_proc_root_links, run_in_ns
+    from granulate_utils.linux.ns import (
+        get_proc_root_path,
+        get_process_nspid,
+        resolve_proc_root_links,
+        run_in_ns_wrapper,
+    )
     from granulate_utils.linux.oom import get_oom_entry
     from granulate_utils.linux.process import (
         get_mapped_dso_elf_id,
@@ -354,7 +359,7 @@ def _run_java_version() -> "CompletedProcess[bytes]":
 
     # doesn't work without changing PID NS as well (I'm getting ENOENT for libjli.so)
     # Version is printed to stderr
-    return run_in_ns(["pid", "mnt"], _run_java_version, process.pid).stderr.decode().strip()
+    return run_in_ns_wrapper(["pid", "mnt"], _run_java_version, process.pid).stderr.decode().strip()
 
 
 def get_java_version_logged(process: Process, stop_event: Event) -> Optional[str]:

gprofiler/profilers/node.py

@@ -27,7 +27,7 @@
 
 import psutil
 import requests
-from granulate_utils.linux.ns import get_proc_root_path, get_process_nspid, resolve_proc_root_links, run_in_ns
+from granulate_utils.linux.ns import get_proc_root_path, get_process_nspid, resolve_proc_root_links, run_in_ns_wrapper
 from granulate_utils.linux.process import is_musl, is_process_running
 from retry import retry
 from websocket import create_connection
@@ -246,7 +246,7 @@ def generate_map_for_node_processes(processes: List[psutil.Process]) -> List[psu
             nspid = get_process_nspid(process.pid)
             ns_link_name = os.readlink(f"/proc/{process.pid}/ns/pid")
             _start_debugger(process.pid)
-            run_in_ns(
+            run_in_ns_wrapper(
                 ["pid", "mnt", "net"],
                 lambda: _generate_perf_map(dest, nspid, ns_link_name, process.pid),
                 process.pid,
@@ -268,7 +268,7 @@ def clean_up_node_maps(processes: List[psutil.Process]) -> None:
             ns_link_name = os.readlink(f"/proc/{process.pid}/ns/pid")
             dest = _get_dest_inside_container(is_musl(process), node_major_version)
             _start_debugger(process.pid)
-            run_in_ns(
+            run_in_ns_wrapper(
                 ["pid", "mnt", "net"],
                 lambda: _clean_up(dest, nspid, ns_link_name, process.pid),
                 process.pid,

gprofiler/profilers/perf.py

@@ -194,7 +194,9 @@ def __init__(
         try:
             # We want to be certain that `perf record` will collect samples.
             discovered_perf_event = discover_appropriate_perf_event(
-                Path(self._profiler_state.storage_dir), self._profiler_state.stop_event
+                Path(self._profiler_state.storage_dir),
+                self._profiler_state.stop_event,
+                self._profiler_state.processes_to_profile,
             )
             logger.debug("Discovered perf event", discovered_perf_event=discovered_perf_event.name)
             extra_args.extend(discovered_perf_event.perf_extra_args())

gprofiler/profilers/python.py

@@ -22,7 +22,7 @@
 from typing import Any, Dict, List, Match, Optional, cast
 
 from granulate_utils.linux.elf import get_elf_id
-from granulate_utils.linux.ns import get_process_nspid, run_in_ns
+from granulate_utils.linux.ns import get_process_nspid, run_in_ns_wrapper
 from granulate_utils.linux.process import (
     get_mapped_dso_elf_id,
     is_process_basename_matching,
@@ -140,7 +140,7 @@ def _run_python_process_in_ns() -> "CompletedProcess[bytes]":
                     timeout=self._PYTHON_TIMEOUT,
                 )
 
-            return run_in_ns(["pid", "mnt"], _run_python_process_in_ns, process.pid).stdout.decode().strip()
+            return run_in_ns_wrapper(["pid", "mnt"], _run_python_process_in_ns, process.pid).stdout.decode().strip()
         except Exception:
             return None
 

gprofiler/utils/__init__.py

@@ -39,7 +39,7 @@
 import psutil
 from granulate_utils.exceptions import CouldNotAcquireMutex
 from granulate_utils.linux.mutex import try_acquire_mutex
-from granulate_utils.linux.ns import run_in_ns
+from granulate_utils.linux.ns import is_root, run_in_ns_wrapper
 from granulate_utils.linux.process import is_kernel_thread, process_exe
 from psutil import Process
 
@@ -82,14 +82,6 @@ def resource_path(relative_path: str = "") -> str:
         raise Exception(f"Resource {relative_path!r} not found!") from e
 
 
-@lru_cache(maxsize=None)
-def is_root() -> bool:
-    if is_windows():
-        return cast(int, ctypes.windll.shell32.IsUserAnAdmin()) == 1  # type: ignore
-    else:
-        return os.geteuid() == 0
-
-
 libc: Optional[ctypes.CDLL] = None
 
 
@@ -376,7 +368,11 @@ def pgrep_maps(match: str) -> List[Process]:
     for line in result.stderr.splitlines():
         if not (
             line.startswith(b"grep: /proc/")
-            and (line.endswith(b"/maps: No such file or directory") or line.endswith(b"/maps: No such process"))
+            and (
+                line.endswith(b"/maps: No such file or directory")
+                or line.endswith(b"/maps: No such process")
+                or (not is_root() and b"/maps: Permission denied" in line)
+            )
         ):
             error_lines.append(line)
     if error_lines:
@@ -455,7 +451,7 @@ def grab_gprofiler_mutex() -> bool:
     GPROFILER_LOCK = "\x00gprofiler_lock"
 
     try:
-        run_in_ns(["net"], lambda: try_acquire_mutex(GPROFILER_LOCK))
+        run_in_ns_wrapper(["net"], lambda: try_acquire_mutex(GPROFILER_LOCK))
     except CouldNotAcquireMutex:
         print(
             "Could not acquire gProfiler's lock. Is it already running?"

gprofiler/utils/fs.py

@@ -21,8 +21,10 @@
 from secrets import token_hex
 from typing import Union
 
+from granulate_utils.linux.ns import is_root
+
 from gprofiler.platform import is_windows
-from gprofiler.utils import is_root, remove_path, run_process
+from gprofiler.utils import remove_path, run_process
 
 
 def safe_copy(src: str, dst: str) -> None:
@@ -76,6 +78,32 @@ def is_owned_by_root(path: Path) -> bool:
     return statbuf.st_uid == 0 and statbuf.st_gid == 0
 
 
+def mkdir_owned_root_wrapper(path: Union[str, Path], mode: int = 0o755) -> None:
+    """
+    Ensures a directory exists and writable.
+
+    If the directory exists and is not writable, the function raises.
+    If the directory exists and is not writable, it is left as is.
+    If the directory doesn't exist, it is created.
+    """
+    if is_root():
+        return mkdir_owned_root(path)
+
+    path = path if isinstance(path, Path) else Path(path)
+    if path.exists():
+        if not os.access(path, os.W_OK):
+            raise Exception(f"{str(path)} is not writable by current user")
+        return
+
+    try:
+        os.mkdir(path, mode=mode)
+    except FileExistsError:
+        # likely racing with another thread of gprofiler. as long as the directory is the user after all, we're good.
+        if not os.access(path, os.W_OK):
+            raise Exception(f"{str(path)} is not writable by current user")
+        pass
+
+
 def mkdir_owned_root(path: Union[str, Path], mode: int = 0o755) -> None:
     """
     Ensures a directory exists and is owned by root.

gprofiler/utils/perf.py

@@ -21,6 +21,8 @@
 from threading import Event
 from typing import List, Optional
 
+from psutil import Process
+
 from gprofiler.exceptions import CalledProcessError, PerfNoSupportedEvent
 from gprofiler.gprofiler_types import ProcessToStackSampleCounters
 from gprofiler.log import get_logger_adapter
@@ -66,7 +68,9 @@ def perf_extra_args(self) -> List[str]:
         return ["-e", self.value]
 
 
-def discover_appropriate_perf_event(tmp_dir: Path, stop_event: Event) -> SupportedPerfEvent:
+def discover_appropriate_perf_event(
+    tmp_dir: Path, stop_event: Event, pids: Optional[List[Process]] = None
+) -> SupportedPerfEvent:
     """
     Get the appropriate event should be used by `perf record`.
 
@@ -95,7 +99,7 @@ def discover_appropriate_perf_event(tmp_dir: Path, stop_event: Event) -> Support
                 is_dwarf=False,
                 inject_jit=False,
                 extra_args=current_extra_args,
-                processes_to_profile=None,
+                processes_to_profile=pids,
                 switch_timeout_s=15,
             )
             perf_process.start()

tests/test_preconditions.py

@@ -102,7 +102,7 @@ def test_not_root(
         wait_for_container(gprofiler)
 
     assert e.value.exit_status == 1
-    assert e.value.stderr == b"Must run gprofiler as root, please re-run.\n"
+    assert e.value.stderr == b"Not running as root, rerun with --rootless or as root.\n"
 
 
 def test_not_host_pid(