1 /*
2 * Copyright (C) 2020 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include "src/trace_processor/importers/ninja/ninja_log_parser.h"
18 #include "perfetto/ext/base/string_splitter.h"
19 #include "perfetto/ext/base/string_utils.h"
20 #include "src/trace_processor/importers/common/process_tracker.h"
21 #include "src/trace_processor/importers/common/slice_tracker.h"
22 #include "src/trace_processor/importers/common/track_tracker.h"
23 #include "src/trace_processor/storage/trace_storage.h"
24 #include "src/trace_processor/trace_sorter.h"
25
26 namespace perfetto {
27 namespace trace_processor {
28
29 using base::StringSplitter;
30
NinjaLogParser(TraceProcessorContext * ctx)31 NinjaLogParser::NinjaLogParser(TraceProcessorContext* ctx) : ctx_(ctx) {}
32 NinjaLogParser::~NinjaLogParser() = default;
33
Parse(std::unique_ptr<uint8_t[]> buf,size_t len)34 util::Status NinjaLogParser::Parse(std::unique_ptr<uint8_t[]> buf, size_t len) {
35 // A trace is read in chunks of arbitrary size (for http fetch() pipeliniing),
36 // not necessarily aligned on a line boundary.
37 // Here we push everything into a vector and, on each call, consume only
38 // the leading part until the last \n, keeping the rest for the next call.
39 const char* src = reinterpret_cast<const char*>(&buf[0]);
40 log_.insert(log_.end(), src, src + len);
41
42 // Find the last \n.
43 size_t valid_size = log_.size();
44 for (; valid_size > 0 && log_[valid_size - 1] != '\n'; --valid_size) {
45 }
46
47 for (StringSplitter line(log_.data(), valid_size, '\n'); line.Next();) {
48 static const char kHeader[] = "# ninja log v";
49 if (!header_parsed_) {
50 if (!base::StartsWith(line.cur_token(), kHeader))
51 return util::ErrStatus("Failed to parse ninja log header");
52 header_parsed_ = true;
53 auto version = base::CStringToUInt32(line.cur_token() + strlen(kHeader));
54 if (!version || *version != 5)
55 return util::ErrStatus("Unsupported ninja log version");
56 continue;
57 }
58
59 // Each line in the ninja log looks like this:
60 // 4 12 1579224178 ui/assets/modal.scss 832a958a9e234dfa
61 // Where:
62 // - [4, 12] are the timestamps in ms of [start, end] of the job, measured
63 // from the beginning of the build.
64 // - 1579224178 is the "restat" (ignored).
65 // - ui/assets/modal.scss is the name of the output file being built.
66 // - 832a958a9e234dfa is a hash of the compiler invocation.
67 // In most cases, each hash should be unique per ninja invocation (because
68 // two rules shouln't generate the same output). However, in rare
69 // circumstances the same hash can show up more than once. Examples:
70 // - A GN action generates > 1 output per invocation (e.g., protos). In this
71 // case all items will have the same [start, end] timestamp. In this case
72 // we want to merge all the output names into one build step, because from
73 // the build system viewpoint, that was the same compiler/tool invocation.
74 // - A subtle script that generates different outputs without taking a
75 // --output=filename argument (e.g. via env vars or similar). Note that
76 // this happens in the perfetto codebase itself (goto.google.com/nigew).
77 // In this case we want to treat the two entries as two distinct jobs.
78 //
79 // In summary the deduping logic here is: if both the hash and the
80 // timestamps match -> merge, if not, keep distinct.
81 StringSplitter tok(&line, '\t');
82 auto t_start = base::CStringToInt64(tok.Next() ? tok.cur_token() : "");
83 auto t_end = base::CStringToInt64(tok.Next() ? tok.cur_token() : "");
84 tok.Next(); // Ignore restat.
85 const char* name = tok.Next() ? tok.cur_token() : nullptr;
86 auto cmdhash = base::CStringToUInt64(tok.Next() ? tok.cur_token() : "", 16);
87
88 if (!t_start || !t_end || !name || !cmdhash) {
89 ctx_->storage->IncrementStats(stats::ninja_parse_errors);
90 continue;
91 }
92
93 // The same log file can contain timestamps for different builds. The only
94 // way we can tell when a new build starts is by detecting the end timestamp
95 // breaking monotonicity.
96 if (last_end_seen_ == 0 || *t_end < last_end_seen_) {
97 // Create a new "process" for each build. In the UI this causes each build
98 // to be nested under a track group. |cur_build_id_| is the fake pid
99 // of the synthesized process.
100 ++cur_build_id_;
101 StringId name_id = ctx_->storage->InternString("Build");
102 ctx_->process_tracker->SetProcessNameIfUnset(
103 ctx_->process_tracker->GetOrCreateProcess(cur_build_id_), name_id);
104 }
105 last_end_seen_ = *t_end;
106
107 // If more hashes show up back-to-back with the same timestamps, merge them
108 // together as they identify multiple outputs for the same build rule.
109 if (!jobs_.empty() && *cmdhash == jobs_.back().hash &&
110 *t_start == jobs_.back().start_ms && *t_end == jobs_.back().end_ms) {
111 jobs_.back().names.append(" ");
112 jobs_.back().names.append(name);
113 continue;
114 }
115
116 jobs_.emplace_back(cur_build_id_, *t_start, *t_end, *cmdhash, name);
117 }
118 log_.erase(log_.begin(), log_.begin() + static_cast<ssize_t>(valid_size));
119 return util::OkStatus();
120 }
121
122 // This is called after the last Parase() call. At this point all |jobs_| have
123 // been populated.
NotifyEndOfFile()124 void NinjaLogParser::NotifyEndOfFile() {
125 std::sort(jobs_.begin(), jobs_.end(),
126 [](const Job& x, const Job& y) { return x.start_ms < y.start_ms; });
127
128 // Now we need to work out the job parallelism. There's no direct indication
129 // of that in the ninja logs, so it must be inferred by observing overlapping
130 // of timestamps. In this context a "Worker" is an inferred sequence of jobs
131 // that happened concurrently with other sequences.
132 // Here we pack jobs according the following heuristic, for the sake of making
133 // the graph nicer to read to humans. Consider the initial situation:
134 // 1: [ job 1 ]
135 // 2: [ job 2 ]
136 // 3: [ job 3 ]
137 // T=0 | T=6
138 // Assume that a new job starts at T=6. It's very likely that job4 was started
139 // as a consequence of job2 completion (othewise it could have been started
140 // earlier, soon after job 1 or Job 3). It seems to make more sense to draw
141 // it next in the 2nd worker, i.e. next to job 2.
142 struct Worker {
143 int64_t busy_until;
144 TrackId track_id;
145 };
146 std::map<uint32_t /*build_id*/, std::vector<Worker>> workers_by_build;
147
148 // Assign thread ids to worker without conflicting with builds' process ids
149 // (to avoid main-thread auto-mapping).s
150 uint32_t last_worker_id = cur_build_id_;
151
152 for (const auto& job : jobs_) {
153 Worker* worker = nullptr;
154 auto& workers = workers_by_build[job.build_id];
155 for (Worker& cur : workers) {
156 // Pick the worker which has the greatest end time (busy_until) <= the
157 // job's start time.
158 if (cur.busy_until <= job.start_ms) {
159 if (!worker || cur.busy_until > worker->busy_until)
160 worker = &cur;
161 }
162 }
163 if (worker) {
164 // Update the worker's end time with the newly assigned job.
165 worker->busy_until = job.end_ms;
166 } else {
167 // All workers are busy, allocate a new one.
168 uint32_t worker_id = ++last_worker_id;
169 char name[32];
170 snprintf(name, sizeof(name), "Worker %zu", workers.size() + 1);
171 StringId name_id = ctx_->storage->InternString(name);
172 auto utid = ctx_->process_tracker->UpdateThread(worker_id, job.build_id);
173 ctx_->process_tracker->SetThreadNameIfUnset(utid, name_id);
174 TrackId track_id = ctx_->track_tracker->InternThreadTrack(utid);
175 workers.emplace_back(Worker{/*busy_until=*/job.end_ms, track_id});
176 worker = &workers.back();
177 }
178
179 static constexpr int64_t kMsToNs = 1000 * 1000;
180 const int64_t start_ns = job.start_ms * kMsToNs;
181 const int64_t dur_ns = (job.end_ms - job.start_ms) * kMsToNs;
182 StringId name_id = ctx_->storage->InternString(base::StringView(job.names));
183 ctx_->slice_tracker->Scoped(start_ns, worker->track_id, StringId::Null(),
184 name_id, dur_ns);
185 }
186 }
187
188 } // namespace trace_processor
189 } // namespace perfetto
190