protos/perfetto/config/profiling/perf_event_config.proto

/*
 * Copyright (C) 2019 The Android Open Source Project
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *      http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

syntax = "proto2";

import "protos/perfetto/common/perf_events.proto";

package perfetto.protos;

// Configuration for the traced_perf profiler.
//
// Example config for basic cpu profiling:
//   perf_event_config {
//     timebase {
//       frequency: 80
//     }
//     callstack_sampling {
//       scope {
//         target_cmdline: "surfaceflinger"
//         target_cmdline: "system_server"
//       }
//       kernel_frames: true
//     }
//   }
//
// Next id: 20
message PerfEventConfig {
  // What event to sample on, and how often.
  // Defined in common/perf_events.proto.
  optional PerfEvents.Timebase timebase = 15;

  // Other events associated with the leader described in the timebase.
  repeated FollowerEvent followers = 19;

  // If set, the profiler will sample userspace processes' callstacks at the
  // interval specified by the |timebase|.
  // If unset, the profiler will record only the event counts.
  optional CallstackSampling callstack_sampling = 16;

  //
  // Kernel <-> userspace ring buffer options:
  //

  // How often the per-cpu ring buffers are read by the producer.
  // If unset, an implementation-defined default is used.
  optional uint32 ring_buffer_read_period_ms = 8;

  // Size (in 4k pages) of each per-cpu ring buffer that is filled by the
  // kernel. If set, must be a power of two.
  // If unset, an implementation-defined default is used.
  optional uint32 ring_buffer_pages = 3;

  //
  // Daemon's resource usage limits:
  //

  // Drop samples if the heap memory held by the samples in the unwinder queue
  // is above the given limit. This counts the memory across all concurrent data
  // sources (not just this one's), and there is no fairness guarantee - the
  // whole quota might be used up by a concurrent source.
  optional uint64 max_enqueued_footprint_kb = 17;

  // Stop the data source if traced_perf's combined {RssAnon + Swap} memory
  // footprint exceeds this value.
  optional uint32 max_daemon_memory_kb = 13;

  //
  // Uncommon options:
  //

  // Timeout for the remote /proc/<pid>/{maps,mem} file descriptors for a
  // sampled process. This is primarily for Android, where this lookup is
  // asynchronous. As long as the producer is waiting, the associated samples
  // will be kept enqueued (putting pressure on the capacity of the shared
  // unwinding queue). Once a lookup for a process expires, all associated
  // samples are discarded. However, if the lookup still succeeds after the
  // timeout, future samples will be handled normally.
  // If unset, an implementation-defined default is used.
  optional uint32 remote_descriptor_timeout_ms = 9;

  // Optional period for clearing state cached by the unwinder. This is a heavy
  // operation that is only necessary for traces that target a wide set of
  // processes, and require the memory footprint to be reset periodically.
  // If unset, the cached state will not be cleared.
  optional uint32 unwind_state_clear_period_ms = 10;

  // If set, only profile target if it was installed by a package with one of
  // these names. Special values:
  // * "@system": installed on the system partition
  // * "@product": installed on the product partition
  // * "@null": sideloaded
  // Supported on Android 12+.
  repeated string target_installed_by = 18;

  //
  // Deprecated (superseded by options above):
  //
  // Do not set *any* of these fields in new configs.
  //

  // Note: legacy configs had to set |all_cpus| to true to pass parsing.
  // We rely on this to detect such configs.
  optional bool all_cpus = 1;
  optional uint32 sampling_frequency = 2;
  optional bool kernel_frames = 12;
  repeated int32 target_pid = 4;
  repeated string target_cmdline = 5;
  repeated int32 exclude_pid = 6;
  repeated string exclude_cmdline = 7;
  optional uint32 additional_cmdline_count = 11;
  // previously |tracepoint|
  reserved 14;

  //
  // Sub-messages (nested for generated code namespacing).
  //

  message CallstackSampling {
    // Defines a set of processes for which samples are retained/skipped. If
    // unset, all samples are kept, but beware that it will be very heavy on the
    // stack unwinder, which might start dropping samples due to overload.
    optional Scope scope = 1;

    // If true, callstacks will include the kernel-space frames. Such frames can
    // be identified by a magical "kernel" string as their mapping name.
    // Requires traced_perf to be running as root, or kptr_restrict to have been
    // manually unrestricted. On Android, the platform should do the right thing
    // on debug builds.
    // This does *not* disclose KASLR, as only the function names are emitted.
    optional bool kernel_frames = 2;

    // Whether to record and unwind userspace callstacks. If unset, defaults to
    // including userspace (UNWIND_DWARF) both for backwards compatibility and
    // as the most common default (this defaulting is only applicable if the
    // outer CallstackSampling message is explicitly set).
    optional UnwindMode user_frames = 3;
  }

  message Scope {
    // Process ID (TGID) allowlist. If this list is not empty, only matching
    // samples will be retained. If multiple allow/deny-lists are
    // specified by the config, then all of them are evaluated for each sampled
    // process.
    repeated int32 target_pid = 1;

    // Command line allowlist, matched against the /proc/<pid>/cmdline (not the
    // comm string). The semantics of this field were changed since its original
    // introduction.
    //
    // On Android T+ (13+), this field can specify a single wildcard (*), and
    // the profiler will attempt to match it in two possible ways:
    // * if the pattern starts with a '/', then it is matched against the first
    //   segment of the cmdline (i.e. argv0). For example "/bin/e*" would match
    //   "/bin/echo".
    // * otherwise the pattern is matched against the part of argv0
    //   corresponding to the binary name (this is unrelated to /proc/pid/exe).
    //   For example "echo" would match "/bin/echo".
    //
    // On Android S (12) and below, both this pattern and /proc/pid/cmdline get
    // normalized prior to an exact string comparison. Normalization is as
    // follows: (1) trim everything beyond the first null or "@" byte; (2) if
    // the string contains forward slashes, trim everything up to and including
    // the last one.
    //
    // Implementation note: in either case, at most 511 characters of cmdline
    // are considered.
    repeated string target_cmdline = 2;

    // List of excluded pids.
    repeated int32 exclude_pid = 3;

    // List of excluded cmdlines. See description of |target_cmdline| for how
    // this is handled.
    repeated string exclude_cmdline = 4;

    // Niche features for systemwide callstacks:

    // Number of additional command lines to sample. Only those which are
    // neither explicitly included nor excluded will be considered. Processes
    // are accepted on a first come, first served basis.
    optional uint32 additional_cmdline_count = 5;

    // If set to N, all encountered processes will be put into one of the N
    // possible bins, and only one randomly-chosen bin will be selected for
    // unwinding. The binning is simply "pid % N", under the assumption that
    // low-order bits of pids are roughly uniformly distributed. Other explicit
    // inclusions/exclusions in this |Scope| message are still respected.
    //
    // The profiler will report the chosen shard in PerfSampleDefaults, and the
    // values will be queryable in trace processor under the "stats" table as
    // "perf_process_shard_count" and "perf_chosen_process_shard".
    //
    // NB: all data sources in a config that set |process_shard_count| must set
    // it to the same value. The profiler will choose one bin for all those data
    // sources.
    optional uint32 process_shard_count = 6;
  }

  // Userspace unwinding mode. A possible future addition is kernel-unwound
  // callchains for frame pointer based systems.
  enum UnwindMode {
    UNWIND_UNKNOWN = 0;
    // Do not unwind userspace:
    UNWIND_SKIP = 1;
    // Use libunwindstack (default):
    UNWIND_DWARF = 2;
  }
}