• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (C) 2018 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/tracing/core/trace_buffer.h"
18 
19 #include <limits>
20 
21 #include "perfetto/base/logging.h"
22 #include "perfetto/protozero/proto_utils.h"
23 #include "perfetto/tracing/core/shared_memory_abi.h"
24 #include "perfetto/tracing/core/trace_packet.h"
25 
26 #define TRACE_BUFFER_VERBOSE_LOGGING() 0  // Set to 1 when debugging unittests.
27 #if TRACE_BUFFER_VERBOSE_LOGGING()
28 #define TRACE_BUFFER_DLOG PERFETTO_DLOG
29 namespace {
HexDump(const uint8_t * src,size_t size)30 std::string HexDump(const uint8_t* src, size_t size) {
31   std::string buf;
32   buf.reserve(4096 * 4);
33   char line[64];
34   char* c = line;
35   for (size_t i = 0; i < size; i++) {
36     c += sprintf(c, "%02x ", src[i]);
37     if (i % 16 == 15) {
38       buf.append("\n");
39       buf.append(line);
40       c = line;
41     }
42   }
43   return buf;
44 }
45 }  // namespace
46 #else
47 #define TRACE_BUFFER_DLOG(...) void()
48 #endif
49 
50 namespace perfetto {
51 
52 namespace {
53 constexpr uint8_t kFirstPacketContinuesFromPrevChunk =
54     SharedMemoryABI::ChunkHeader::kFirstPacketContinuesFromPrevChunk;
55 constexpr uint8_t kLastPacketContinuesOnNextChunk =
56     SharedMemoryABI::ChunkHeader::kLastPacketContinuesOnNextChunk;
57 constexpr uint8_t kChunkNeedsPatching =
58     SharedMemoryABI::ChunkHeader::kChunkNeedsPatching;
59 }  // namespace.
60 
61 constexpr size_t TraceBuffer::ChunkRecord::kMaxSize;
62 constexpr size_t TraceBuffer::InlineChunkHeaderSize = sizeof(ChunkRecord);
63 
64 // static
Create(size_t size_in_bytes,OverwritePolicy pol)65 std::unique_ptr<TraceBuffer> TraceBuffer::Create(size_t size_in_bytes,
66                                                  OverwritePolicy pol) {
67   std::unique_ptr<TraceBuffer> trace_buffer(new TraceBuffer(pol));
68   if (!trace_buffer->Initialize(size_in_bytes))
69     return nullptr;
70   return trace_buffer;
71 }
72 
TraceBuffer(OverwritePolicy pol)73 TraceBuffer::TraceBuffer(OverwritePolicy pol) : overwrite_policy_(pol) {
74   // See comments in ChunkRecord for the rationale of this.
75   static_assert(sizeof(ChunkRecord) == sizeof(SharedMemoryABI::PageHeader) +
76                                            sizeof(SharedMemoryABI::ChunkHeader),
77                 "ChunkRecord out of sync with the layout of SharedMemoryABI");
78 }
79 
80 TraceBuffer::~TraceBuffer() = default;
81 
Initialize(size_t size)82 bool TraceBuffer::Initialize(size_t size) {
83   static_assert(
84       base::kPageSize % sizeof(ChunkRecord) == 0,
85       "sizeof(ChunkRecord) must be an integer divider of a page size");
86   PERFETTO_CHECK(size % base::kPageSize == 0);
87   data_ = base::PagedMemory::Allocate(
88       size, base::PagedMemory::kMayFail | base::PagedMemory::kDontCommit);
89   if (!data_.IsValid()) {
90     PERFETTO_ELOG("Trace buffer allocation failed (size: %zu)", size);
91     return false;
92   }
93   size_ = size;
94   stats_.set_buffer_size(size);
95   max_chunk_size_ = std::min(size, ChunkRecord::kMaxSize);
96   wptr_ = begin();
97   index_.clear();
98   last_chunk_id_written_.clear();
99   read_iter_ = GetReadIterForSequence(index_.end());
100   return true;
101 }
102 
103 // Note: |src| points to a shmem region that is shared with the producer. Assume
104 // that the producer is malicious and will change the content of |src|
105 // while we execute here. Don't do any processing on it other than memcpy().
CopyChunkUntrusted(ProducerID producer_id_trusted,uid_t producer_uid_trusted,WriterID writer_id,ChunkID chunk_id,uint16_t num_fragments,uint8_t chunk_flags,bool chunk_complete,const uint8_t * src,size_t size)106 void TraceBuffer::CopyChunkUntrusted(ProducerID producer_id_trusted,
107                                      uid_t producer_uid_trusted,
108                                      WriterID writer_id,
109                                      ChunkID chunk_id,
110                                      uint16_t num_fragments,
111                                      uint8_t chunk_flags,
112                                      bool chunk_complete,
113                                      const uint8_t* src,
114                                      size_t size) {
115   // |record_size| = |size| + sizeof(ChunkRecord), rounded up to avoid to end
116   // up in a fragmented state where size_to_end() < sizeof(ChunkRecord).
117   const size_t record_size =
118       base::AlignUp<sizeof(ChunkRecord)>(size + sizeof(ChunkRecord));
119   if (PERFETTO_UNLIKELY(record_size > max_chunk_size_)) {
120     stats_.set_abi_violations(stats_.abi_violations() + 1);
121     PERFETTO_DCHECK(suppress_sanity_dchecks_for_testing_);
122     return;
123   }
124 
125   TRACE_BUFFER_DLOG("CopyChunk @ %lu, size=%zu", wptr_ - begin(), record_size);
126 
127 #if PERFETTO_DCHECK_IS_ON()
128   changed_since_last_read_ = true;
129 #endif
130 
131   // If the chunk hasn't been completed, we should only consider the first
132   // |num_fragments - 1| packets complete. For simplicity, we simply disregard
133   // the last one when we copy the chunk.
134   if (PERFETTO_UNLIKELY(!chunk_complete)) {
135     if (num_fragments > 0) {
136       num_fragments--;
137       chunk_flags &= ~kLastPacketContinuesOnNextChunk;
138     }
139   }
140 
141   ChunkRecord record(record_size);
142   record.producer_id = producer_id_trusted;
143   record.chunk_id = chunk_id;
144   record.writer_id = writer_id;
145   record.num_fragments = num_fragments;
146   record.flags = chunk_flags;
147   ChunkMeta::Key key(record);
148 
149   // Check whether we have already copied the same chunk previously. This may
150   // happen if the service scrapes chunks in a potentially incomplete state
151   // before receiving commit requests for them from the producer. Note that the
152   // service may scrape and thus override chunks in arbitrary order since the
153   // chunks aren't ordered in the SMB.
154   const auto it = index_.find(key);
155   if (PERFETTO_UNLIKELY(it != index_.end())) {
156     ChunkMeta* record_meta = &it->second;
157     ChunkRecord* prev = record_meta->chunk_record;
158 
159     // Verify that the old chunk's metadata corresponds to the new one.
160     // Overridden chunks should never change size, since the page layout is
161     // fixed per writer. The number of fragments should also never decrease and
162     // flags should not be removed.
163     if (PERFETTO_UNLIKELY(ChunkMeta::Key(*prev) != key ||
164                           prev->size != record_size ||
165                           prev->num_fragments > num_fragments ||
166                           (prev->flags & chunk_flags) != prev->flags)) {
167       stats_.set_abi_violations(stats_.abi_violations() + 1);
168       PERFETTO_DCHECK(suppress_sanity_dchecks_for_testing_);
169       return;
170     }
171 
172     // If we've already started reading from chunk N+1 following this chunk N,
173     // don't override chunk N. Otherwise we may end up reading a packet from
174     // chunk N after having read from chunk N+1, thereby violating sequential
175     // read of packets. This shouldn't happen if the producer is well-behaved,
176     // because it shouldn't start chunk N+1 before completing chunk N.
177     ChunkMeta::Key subsequent_key = key;
178     static_assert(std::numeric_limits<ChunkID>::max() == kMaxChunkID,
179                   "ChunkID wraps");
180     subsequent_key.chunk_id++;
181     const auto subsequent_it = index_.find(subsequent_key);
182     if (subsequent_it != index_.end() &&
183         subsequent_it->second.num_fragments_read > 0) {
184       stats_.set_abi_violations(stats_.abi_violations() + 1);
185       PERFETTO_DCHECK(suppress_sanity_dchecks_for_testing_);
186       return;
187     }
188 
189     // If this chunk was previously copied with the same number of fragments and
190     // the number didn't change, there's no need to copy it again. If the
191     // previous chunk was complete already, this should always be the case.
192     PERFETTO_DCHECK(suppress_sanity_dchecks_for_testing_ ||
193                     !record_meta->is_complete() ||
194                     (chunk_complete && prev->num_fragments == num_fragments));
195     if (prev->num_fragments == num_fragments) {
196       TRACE_BUFFER_DLOG("  skipping recommit of identical chunk");
197       return;
198     }
199 
200     // We should not have read past the last packet.
201     if (record_meta->num_fragments_read > prev->num_fragments) {
202       PERFETTO_ELOG(
203           "TraceBuffer read too many fragments from an incomplete chunk");
204       PERFETTO_DCHECK(suppress_sanity_dchecks_for_testing_);
205       return;
206     }
207 
208     uint8_t* wptr = reinterpret_cast<uint8_t*>(prev);
209     TRACE_BUFFER_DLOG("  overriding chunk @ %lu, size=%zu", wptr - begin(),
210                       record_size);
211 
212     // Update chunk meta data stored in the index, as it may have changed.
213     record_meta->num_fragments = num_fragments;
214     record_meta->flags = chunk_flags;
215     record_meta->set_complete(chunk_complete);
216 
217     // Override the ChunkRecord contents at the original |wptr|.
218     TRACE_BUFFER_DLOG("  copying @ [%lu - %lu] %zu", wptr - begin(),
219                       uintptr_t(wptr - begin()) + record_size, record_size);
220     WriteChunkRecord(wptr, record, src, size);
221     TRACE_BUFFER_DLOG("Chunk raw: %s", HexDump(wptr, record_size).c_str());
222     stats_.set_chunks_rewritten(stats_.chunks_rewritten() + 1);
223     return;
224   }
225 
226   if (PERFETTO_UNLIKELY(discard_writes_))
227     return DiscardWrite();
228 
229   // If there isn't enough room from the given write position. Write a padding
230   // record to clear the end of the buffer and wrap back.
231   const size_t cached_size_to_end = size_to_end();
232   if (PERFETTO_UNLIKELY(record_size > cached_size_to_end)) {
233     ssize_t res = DeleteNextChunksFor(cached_size_to_end);
234     if (res == -1)
235       return DiscardWrite();
236     PERFETTO_DCHECK(static_cast<size_t>(res) <= cached_size_to_end);
237     AddPaddingRecord(cached_size_to_end);
238     wptr_ = begin();
239     stats_.set_write_wrap_count(stats_.write_wrap_count() + 1);
240     PERFETTO_DCHECK(size_to_end() >= record_size);
241   }
242 
243   // At this point either |wptr_| points to an untouched part of the buffer
244   // (i.e. *wptr_ == 0) or we are about to overwrite one or more ChunkRecord(s).
245   // In the latter case we need to first figure out where the next valid
246   // ChunkRecord is (if it exists) and add padding between the new record.
247   // Example ((w) == write cursor):
248   //
249   // Initial state (wtpr_ == 0):
250   // |0 (w)    |10               |30                  |50
251   // +---------+-----------------+--------------------+--------------------+
252   // | Chunk 1 | Chunk 2         | Chunk 3            | Chunk 4            |
253   // +---------+-----------------+--------------------+--------------------+
254   //
255   // Let's assume we now want now write a 5th Chunk of size == 35. The final
256   // state should look like this:
257   // |0                                |35 (w)         |50
258   // +---------------------------------+---------------+--------------------+
259   // | Chunk 5                         | Padding Chunk | Chunk 4            |
260   // +---------------------------------+---------------+--------------------+
261 
262   // Deletes all chunks from |wptr_| to |wptr_| + |record_size|.
263   ssize_t del_res = DeleteNextChunksFor(record_size);
264   if (del_res == -1)
265     return DiscardWrite();
266   size_t padding_size = static_cast<size_t>(del_res);
267 
268   // Now first insert the new chunk. At the end, if necessary, add the padding.
269   stats_.set_chunks_written(stats_.chunks_written() + 1);
270   stats_.set_bytes_written(stats_.bytes_written() + record_size);
271   auto it_and_inserted = index_.emplace(
272       key, ChunkMeta(GetChunkRecordAt(wptr_), num_fragments, chunk_complete,
273                      chunk_flags, producer_uid_trusted));
274   PERFETTO_DCHECK(it_and_inserted.second);
275   TRACE_BUFFER_DLOG("  copying @ [%lu - %lu] %zu", wptr_ - begin(),
276                     uintptr_t(wptr_ - begin()) + record_size, record_size);
277   WriteChunkRecord(wptr_, record, src, size);
278   TRACE_BUFFER_DLOG("Chunk raw: %s", HexDump(wptr_, record_size).c_str());
279   wptr_ += record_size;
280   if (wptr_ >= end()) {
281     PERFETTO_DCHECK(padding_size == 0);
282     wptr_ = begin();
283     stats_.set_write_wrap_count(stats_.write_wrap_count() + 1);
284   }
285   DcheckIsAlignedAndWithinBounds(wptr_);
286 
287   // Chunks may be received out of order, so only update last_chunk_id if the
288   // new chunk_id is larger. But take into account overflows by only selecting
289   // the new ID if its distance to the latest ID is smaller than half the number
290   // space.
291   //
292   // This accounts for both the case where the new ID has just overflown and
293   // last_chunk_id be updated even though it's smaller (e.g. |chunk_id| = 1 and
294   // |last_chunk_id| = kMaxChunkId; chunk_id - last_chunk_id = 0) and the case
295   // where the new ID is an out-of-order ID right after an overflow and
296   // last_chunk_id shouldn't be updated even though it's larger (e.g. |chunk_id|
297   // = kMaxChunkId and |last_chunk_id| = 1; chunk_id - last_chunk_id =
298   // kMaxChunkId - 1).
299   auto producer_and_writer_id = std::make_pair(producer_id_trusted, writer_id);
300   ChunkID& last_chunk_id = last_chunk_id_written_[producer_and_writer_id];
301   static_assert(std::numeric_limits<ChunkID>::max() == kMaxChunkID,
302                 "This code assumes that ChunkID wraps at kMaxChunkID");
303   if (chunk_id - last_chunk_id < kMaxChunkID / 2) {
304     last_chunk_id = chunk_id;
305   } else {
306     stats_.set_chunks_committed_out_of_order(
307         stats_.chunks_committed_out_of_order() + 1);
308   }
309 
310   if (padding_size)
311     AddPaddingRecord(padding_size);
312 }
313 
DeleteNextChunksFor(size_t bytes_to_clear)314 ssize_t TraceBuffer::DeleteNextChunksFor(size_t bytes_to_clear) {
315   PERFETTO_CHECK(!discard_writes_);
316 
317   // Find the position of the first chunk which begins at or after
318   // (|wptr_| + |bytes|). Note that such a chunk might not exist and we might
319   // either reach the end of the buffer or a zeroed region of the buffer.
320   uint8_t* next_chunk_ptr = wptr_;
321   uint8_t* search_end = wptr_ + bytes_to_clear;
322   TRACE_BUFFER_DLOG("Delete [%zu %zu]", wptr_ - begin(), search_end - begin());
323   DcheckIsAlignedAndWithinBounds(wptr_);
324   PERFETTO_DCHECK(search_end <= end());
325   std::vector<ChunkMap::iterator> index_delete;
326   uint64_t chunks_overwritten = stats_.chunks_overwritten();
327   uint64_t bytes_overwritten = stats_.bytes_overwritten();
328   uint64_t padding_bytes_cleared = stats_.padding_bytes_cleared();
329   while (next_chunk_ptr < search_end) {
330     const ChunkRecord& next_chunk = *GetChunkRecordAt(next_chunk_ptr);
331     TRACE_BUFFER_DLOG(
332         "  scanning chunk [%zu %zu] (valid=%d)", next_chunk_ptr - begin(),
333         next_chunk_ptr - begin() + next_chunk.size, next_chunk.is_valid());
334 
335     // We just reached the untouched part of the buffer, it's going to be all
336     // zeroes from here to end().
337     // Optimization: if during Initialize() we fill the buffer with padding
338     // records we could get rid of this branch.
339     if (PERFETTO_UNLIKELY(!next_chunk.is_valid())) {
340       // This should happen only at the first iteration. The zeroed area can
341       // only begin precisely at the |wptr_|, not after. Otherwise it means that
342       // we wrapped but screwed up the ChunkRecord chain.
343       PERFETTO_DCHECK(next_chunk_ptr == wptr_);
344       return 0;
345     }
346 
347     // Remove |next_chunk| from the index, unless it's a padding record (padding
348     // records are not part of the index).
349     if (PERFETTO_LIKELY(!next_chunk.is_padding)) {
350       ChunkMeta::Key key(next_chunk);
351       auto it = index_.find(key);
352       bool will_remove = false;
353       if (PERFETTO_LIKELY(it != index_.end())) {
354         const ChunkMeta& meta = it->second;
355         if (PERFETTO_UNLIKELY(meta.num_fragments_read < meta.num_fragments)) {
356           if (overwrite_policy_ == kDiscard)
357             return -1;
358           chunks_overwritten++;
359           bytes_overwritten += next_chunk.size;
360         }
361         index_delete.push_back(it);
362         will_remove = true;
363       }
364       TRACE_BUFFER_DLOG("  del index {%" PRIu32 ",%" PRIu32
365                         ",%u} @ [%lu - %lu] %d",
366                         key.producer_id, key.writer_id, key.chunk_id,
367                         next_chunk_ptr - begin(),
368                         next_chunk_ptr - begin() + next_chunk.size, removed);
369       PERFETTO_DCHECK(will_remove);
370     } else {
371       padding_bytes_cleared += next_chunk.size;
372     }
373 
374     next_chunk_ptr += next_chunk.size;
375 
376     // We should never hit this, unless we managed to screw up while writing
377     // to the buffer and breaking the ChunkRecord(s) chain.
378     // TODO(primiano): Write more meaningful logging with the status of the
379     // buffer, to get more actionable bugs in case we hit this.
380     PERFETTO_CHECK(next_chunk_ptr <= end());
381   }
382 
383   // Remove from the index.
384   for (auto it : index_delete) {
385     index_.erase(it);
386   }
387   stats_.set_chunks_overwritten(chunks_overwritten);
388   stats_.set_bytes_overwritten(bytes_overwritten);
389   stats_.set_padding_bytes_cleared(padding_bytes_cleared);
390 
391   PERFETTO_DCHECK(next_chunk_ptr >= search_end && next_chunk_ptr <= end());
392   return static_cast<ssize_t>(next_chunk_ptr - search_end);
393 }
394 
AddPaddingRecord(size_t size)395 void TraceBuffer::AddPaddingRecord(size_t size) {
396   PERFETTO_DCHECK(size >= sizeof(ChunkRecord) && size <= ChunkRecord::kMaxSize);
397   ChunkRecord record(size);
398   record.is_padding = 1;
399   TRACE_BUFFER_DLOG("AddPaddingRecord @ [%lu - %lu] %zu", wptr_ - begin(),
400                     uintptr_t(wptr_ - begin()) + size, size);
401   WriteChunkRecord(wptr_, record, nullptr, size - sizeof(ChunkRecord));
402   stats_.set_padding_bytes_written(stats_.padding_bytes_written() + size);
403   // |wptr_| is deliberately not advanced when writing a padding record.
404 }
405 
TryPatchChunkContents(ProducerID producer_id,WriterID writer_id,ChunkID chunk_id,const Patch * patches,size_t patches_size,bool other_patches_pending)406 bool TraceBuffer::TryPatchChunkContents(ProducerID producer_id,
407                                         WriterID writer_id,
408                                         ChunkID chunk_id,
409                                         const Patch* patches,
410                                         size_t patches_size,
411                                         bool other_patches_pending) {
412   ChunkMeta::Key key(producer_id, writer_id, chunk_id);
413   auto it = index_.find(key);
414   if (it == index_.end()) {
415     stats_.set_patches_failed(stats_.patches_failed() + 1);
416     return false;
417   }
418   ChunkMeta& chunk_meta = it->second;
419 
420   // Check that the index is consistent with the actual ProducerID/WriterID
421   // stored in the ChunkRecord.
422   PERFETTO_DCHECK(ChunkMeta::Key(*chunk_meta.chunk_record) == key);
423   uint8_t* chunk_begin = reinterpret_cast<uint8_t*>(chunk_meta.chunk_record);
424   PERFETTO_DCHECK(chunk_begin >= begin());
425   uint8_t* chunk_end = chunk_begin + chunk_meta.chunk_record->size;
426   PERFETTO_DCHECK(chunk_end <= end());
427 
428   static_assert(Patch::kSize == SharedMemoryABI::kPacketHeaderSize,
429                 "Patch::kSize out of sync with SharedMemoryABI");
430 
431   for (size_t i = 0; i < patches_size; i++) {
432     uint8_t* ptr =
433         chunk_begin + sizeof(ChunkRecord) + patches[i].offset_untrusted;
434     TRACE_BUFFER_DLOG("PatchChunk {%" PRIu32 ",%" PRIu32
435                       ",%u} size=%zu @ %zu with {%02x %02x %02x %02x} cur "
436                       "{%02x %02x %02x %02x}",
437                       producer_id, writer_id, chunk_id, chunk_end - chunk_begin,
438                       patches[i].offset_untrusted, patches[i].data[0],
439                       patches[i].data[1], patches[i].data[2],
440                       patches[i].data[3], ptr[0], ptr[1], ptr[2], ptr[3]);
441     if (ptr < chunk_begin + sizeof(ChunkRecord) ||
442         ptr > chunk_end - Patch::kSize) {
443       // Either the IPC was so slow and in the meantime the writer managed to
444       // wrap over |chunk_id| or the producer sent a malicious IPC.
445       stats_.set_patches_failed(stats_.patches_failed() + 1);
446       return false;
447     }
448 
449     // DCHECK that we are writing into a zero-filled size field and not into
450     // valid data. It relies on ScatteredStreamWriter::ReserveBytes() to
451     // zero-fill reservations in debug builds.
452     char zero[Patch::kSize]{};
453     PERFETTO_DCHECK(memcmp(ptr, &zero, Patch::kSize) == 0);
454 
455     memcpy(ptr, &patches[i].data[0], Patch::kSize);
456   }
457   TRACE_BUFFER_DLOG(
458       "Chunk raw (after patch): %s",
459       HexDump(chunk_begin, chunk_meta.chunk_record->size).c_str());
460 
461   stats_.set_patches_succeeded(stats_.patches_succeeded() + patches_size);
462   if (!other_patches_pending) {
463     chunk_meta.flags &= ~kChunkNeedsPatching;
464     chunk_meta.chunk_record->flags = chunk_meta.flags;
465   }
466   return true;
467 }
468 
BeginRead()469 void TraceBuffer::BeginRead() {
470   read_iter_ = GetReadIterForSequence(index_.begin());
471 #if PERFETTO_DCHECK_IS_ON()
472   changed_since_last_read_ = false;
473 #endif
474 }
475 
GetReadIterForSequence(ChunkMap::iterator seq_begin)476 TraceBuffer::SequenceIterator TraceBuffer::GetReadIterForSequence(
477     ChunkMap::iterator seq_begin) {
478   SequenceIterator iter;
479   iter.seq_begin = seq_begin;
480   if (seq_begin == index_.end()) {
481     iter.cur = iter.seq_end = index_.end();
482     return iter;
483   }
484 
485 #if PERFETTO_DCHECK_IS_ON()
486   // Either |seq_begin| is == index_.begin() or the item immediately before must
487   // belong to a different {ProducerID, WriterID} sequence.
488   if (seq_begin != index_.begin() && seq_begin != index_.end()) {
489     auto prev_it = seq_begin;
490     prev_it--;
491     PERFETTO_DCHECK(
492         seq_begin == index_.begin() ||
493         std::tie(prev_it->first.producer_id, prev_it->first.writer_id) <
494             std::tie(seq_begin->first.producer_id, seq_begin->first.writer_id));
495   }
496 #endif
497 
498   // Find the first entry that has a greater {ProducerID, WriterID} (or just
499   // index_.end() if we reached the end).
500   ChunkMeta::Key key = seq_begin->first;  // Deliberate copy.
501   key.chunk_id = kMaxChunkID;
502   iter.seq_end = index_.upper_bound(key);
503   PERFETTO_DCHECK(iter.seq_begin != iter.seq_end);
504 
505   // Now find the first entry between [seq_begin, seq_end) that is
506   // > last_chunk_id_written_. This is where we the sequence will start (see
507   // notes about wrapping of IDs in the header).
508   auto producer_and_writer_id = std::make_pair(key.producer_id, key.writer_id);
509   PERFETTO_DCHECK(last_chunk_id_written_.count(producer_and_writer_id));
510   iter.wrapping_id = last_chunk_id_written_[producer_and_writer_id];
511   key.chunk_id = iter.wrapping_id;
512   iter.cur = index_.upper_bound(key);
513   if (iter.cur == iter.seq_end)
514     iter.cur = iter.seq_begin;
515   return iter;
516 }
517 
MoveNext()518 void TraceBuffer::SequenceIterator::MoveNext() {
519   // Stop iterating when we reach the end of the sequence.
520   // Note: |seq_begin| might be == |seq_end|.
521   if (cur == seq_end || cur->first.chunk_id == wrapping_id) {
522     cur = seq_end;
523     return;
524   }
525 
526   // If the current chunk wasn't completed yet, we shouldn't advance past it as
527   // it may be rewritten with additional packets.
528   if (!cur->second.is_complete()) {
529     cur = seq_end;
530     return;
531   }
532 
533   ChunkID last_chunk_id = cur->first.chunk_id;
534   if (++cur == seq_end)
535     cur = seq_begin;
536 
537   // There may be a missing chunk in the sequence of chunks, in which case the
538   // next chunk's ID won't follow the last one's. If so, skip the rest of the
539   // sequence. We'll return to it later once the hole is filled.
540   if (last_chunk_id + 1 != cur->first.chunk_id)
541     cur = seq_end;
542 }
543 
ReadNextTracePacket(TracePacket * packet,PacketSequenceProperties * sequence_properties,bool * previous_packet_on_sequence_dropped)544 bool TraceBuffer::ReadNextTracePacket(
545     TracePacket* packet,
546     PacketSequenceProperties* sequence_properties,
547     bool* previous_packet_on_sequence_dropped) {
548   // Note: MoveNext() moves only within the next chunk within the same
549   // {ProducerID, WriterID} sequence. Here we want to:
550   // - return the next patched+complete packet in the current sequence, if any.
551   // - return the first patched+complete packet in the next sequence, if any.
552   // - return false if none of the above is found.
553   TRACE_BUFFER_DLOG("ReadNextTracePacket()");
554 
555   // Just in case we forget to initialize these below.
556   *sequence_properties = {0, kInvalidUid, 0};
557   *previous_packet_on_sequence_dropped = false;
558 
559   // At the start of each sequence iteration, we consider the last read packet
560   // dropped. While iterating over the chunks in the sequence, we update this
561   // flag based on our knowledge about the last packet that was read from each
562   // chunk (|last_read_packet_skipped| in ChunkMeta).
563   bool previous_packet_dropped = true;
564 
565 #if PERFETTO_DCHECK_IS_ON()
566   PERFETTO_DCHECK(!changed_since_last_read_);
567 #endif
568   for (;; read_iter_.MoveNext()) {
569     if (PERFETTO_UNLIKELY(!read_iter_.is_valid())) {
570       // We ran out of chunks in the current {ProducerID, WriterID} sequence or
571       // we just reached the index_.end().
572 
573       if (PERFETTO_UNLIKELY(read_iter_.seq_end == index_.end()))
574         return false;
575 
576       // We reached the end of sequence, move to the next one.
577       // Note: ++read_iter_.seq_end might become index_.end(), but
578       // GetReadIterForSequence() knows how to deal with that.
579       read_iter_ = GetReadIterForSequence(read_iter_.seq_end);
580       PERFETTO_DCHECK(read_iter_.is_valid() && read_iter_.cur != index_.end());
581       previous_packet_dropped = true;
582     }
583 
584     ChunkMeta* chunk_meta = &*read_iter_;
585 
586     // If the chunk has holes that are awaiting to be patched out-of-band,
587     // skip the current sequence and move to the next one.
588     if (chunk_meta->flags & kChunkNeedsPatching) {
589       read_iter_.MoveToEnd();
590       continue;
591     }
592 
593     const ProducerID trusted_producer_id = read_iter_.producer_id();
594     const WriterID writer_id = read_iter_.writer_id();
595     const uid_t trusted_uid = chunk_meta->trusted_uid;
596 
597     // At this point we have a chunk in |chunk_meta| that has not been fully
598     // read. We don't know yet whether we have enough data to read the full
599     // packet (in the case it's fragmented over several chunks) and we are about
600     // to find that out. Specifically:
601     // A) If the first fragment is unread and is a fragment continuing from a
602     //    previous chunk, it means we have missed the previous ChunkID. In
603     //    fact, if this wasn't the case, a previous call to ReadNext() shouldn't
604     //    have moved the cursor to this chunk.
605     // B) Any fragment > 0 && < last is always readable. By definition an inner
606     //    packet is never fragmented and hence doesn't require neither stitching
607     //    nor any out-of-band patching. The same applies to the last packet
608     //    iff it doesn't continue on the next chunk.
609     // C) If the last packet (which might be also the only packet in the chunk)
610     //    is a fragment and continues on the next chunk, we peek at the next
611     //    chunks and, if we have all of them, mark as read and move the cursor.
612     //
613     // +---------------+   +-------------------+  +---------------+
614     // | ChunkID: 1    |   | ChunkID: 2        |  | ChunkID: 3    |
615     // |---------------+   +-------------------+  +---------------+
616     // | Packet 1      |   |                   |  | ... Packet 3  |
617     // | Packet 2      |   | ... Packet 3  ... |  | Packet 4      |
618     // | Packet 3  ... |   |                   |  | Packet 5 ...  |
619     // +---------------+   +-------------------+  +---------------+
620 
621     PERFETTO_DCHECK(chunk_meta->num_fragments_read <=
622                     chunk_meta->num_fragments);
623 
624     // If we didn't read any packets from this chunk, the last packet was from
625     // the previous chunk we iterated over; so don't update
626     // |previous_packet_dropped| in this case.
627     if (chunk_meta->num_fragments_read > 0)
628       previous_packet_dropped = chunk_meta->last_read_packet_skipped();
629 
630     while (chunk_meta->num_fragments_read < chunk_meta->num_fragments) {
631       enum { kSkip = 0, kReadOnePacket, kTryReadAhead } action;
632       if (chunk_meta->num_fragments_read == 0) {
633         if (chunk_meta->flags & kFirstPacketContinuesFromPrevChunk) {
634           action = kSkip;  // Case A.
635         } else if (chunk_meta->num_fragments == 1 &&
636                    (chunk_meta->flags & kLastPacketContinuesOnNextChunk)) {
637           action = kTryReadAhead;  // Case C.
638         } else {
639           action = kReadOnePacket;  // Case B.
640         }
641       } else if (chunk_meta->num_fragments_read <
642                      chunk_meta->num_fragments - 1 ||
643                  !(chunk_meta->flags & kLastPacketContinuesOnNextChunk)) {
644         action = kReadOnePacket;  // Case B.
645       } else {
646         action = kTryReadAhead;  // Case C.
647       }
648 
649       TRACE_BUFFER_DLOG("  chunk %u, packet %hu of %hu, action=%d",
650                         read_iter_.chunk_id(), chunk_meta->num_fragments_read,
651                         chunk_meta->num_fragments, action);
652 
653       if (action == kSkip) {
654         // This fragment will be skipped forever, not just in this ReadPacket()
655         // iteration. This happens by virtue of ReadNextPacketInChunk()
656         // incrementing the |num_fragments_read| and marking the fragment as
657         // read even if we didn't really.
658         ReadNextPacketInChunk(chunk_meta, nullptr);
659         chunk_meta->set_last_read_packet_skipped(true);
660         previous_packet_dropped = true;
661         continue;
662       }
663 
664       if (action == kReadOnePacket) {
665         // The easy peasy case B.
666         ReadPacketResult result = ReadNextPacketInChunk(chunk_meta, packet);
667 
668         if (PERFETTO_LIKELY(result == ReadPacketResult::kSucceeded)) {
669           *sequence_properties = {trusted_producer_id, trusted_uid, writer_id};
670           *previous_packet_on_sequence_dropped = previous_packet_dropped;
671           return true;
672         } else if (result == ReadPacketResult::kFailedEmptyPacket) {
673           // We can ignore and skip empty packets.
674           PERFETTO_DCHECK(packet->slices().empty());
675           continue;
676         }
677 
678         // In extremely rare cases (producer bugged / malicious) the chunk might
679         // contain an invalid fragment. In such case we don't want to stall the
680         // sequence but just skip the chunk and move on. ReadNextPacketInChunk()
681         // marks the chunk as fully read, so we don't attempt to read from it
682         // again in a future call to ReadBuffers(). It also already records an
683         // abi violation for this.
684         PERFETTO_DCHECK(result == ReadPacketResult::kFailedInvalidPacket);
685         chunk_meta->set_last_read_packet_skipped(true);
686         previous_packet_dropped = true;
687         break;
688       }
689 
690       PERFETTO_DCHECK(action == kTryReadAhead);
691       ReadAheadResult ra_res = ReadAhead(packet);
692       if (ra_res == ReadAheadResult::kSucceededReturnSlices) {
693         stats_.set_readaheads_succeeded(stats_.readaheads_succeeded() + 1);
694         *sequence_properties = {trusted_producer_id, trusted_uid, writer_id};
695         *previous_packet_on_sequence_dropped = previous_packet_dropped;
696         return true;
697       }
698 
699       if (ra_res == ReadAheadResult::kFailedMoveToNextSequence) {
700         // readahead didn't find a contiguous packet sequence. We'll try again
701         // on the next ReadPacket() call.
702         stats_.set_readaheads_failed(stats_.readaheads_failed() + 1);
703 
704         // TODO(primiano): optimization: this MoveToEnd() is the reason why
705         // MoveNext() (that is called in the outer for(;;MoveNext)) needs to
706         // deal gracefully with the case of |cur|==|seq_end|. Maybe we can do
707         // something to avoid that check by reshuffling the code here?
708         read_iter_.MoveToEnd();
709 
710         // This break will go back to beginning of the for(;;MoveNext()). That
711         // will move to the next sequence because we set the read iterator to
712         // its end.
713         break;
714       }
715 
716       PERFETTO_DCHECK(ra_res == ReadAheadResult::kFailedStayOnSameSequence);
717 
718       // In this case ReadAhead() might advance |read_iter_|, so we need to
719       // re-cache the |chunk_meta| pointer to point to the current chunk.
720       chunk_meta = &*read_iter_;
721       chunk_meta->set_last_read_packet_skipped(true);
722       previous_packet_dropped = true;
723     }  // while(...)  [iterate over packet fragments for the current chunk].
724   }    // for(;;MoveNext()) [iterate over chunks].
725 }
726 
ReadAhead(TracePacket * packet)727 TraceBuffer::ReadAheadResult TraceBuffer::ReadAhead(TracePacket* packet) {
728   static_assert(static_cast<ChunkID>(kMaxChunkID + 1) == 0,
729                 "relying on kMaxChunkID to wrap naturally");
730   TRACE_BUFFER_DLOG(" readahead start @ chunk %u", read_iter_.chunk_id());
731   ChunkID next_chunk_id = read_iter_.chunk_id() + 1;
732   SequenceIterator it = read_iter_;
733   for (it.MoveNext(); it.is_valid(); it.MoveNext(), next_chunk_id++) {
734     // We should stay within the same sequence while iterating here.
735     PERFETTO_DCHECK(it.producer_id() == read_iter_.producer_id() &&
736                     it.writer_id() == read_iter_.writer_id());
737 
738     TRACE_BUFFER_DLOG("   expected chunk ID: %u, actual ID: %u", next_chunk_id,
739                       it.chunk_id());
740 
741     if (PERFETTO_UNLIKELY((*it).num_fragments == 0))
742       continue;
743 
744     // If we miss the next chunk, stop looking in the current sequence and
745     // try another sequence. This chunk might come in the near future.
746     // The second condition is the edge case of a buggy/malicious
747     // producer. The ChunkID is contiguous but its flags don't make sense.
748     if (it.chunk_id() != next_chunk_id ||
749         PERFETTO_UNLIKELY(
750             !((*it).flags & kFirstPacketContinuesFromPrevChunk))) {
751       return ReadAheadResult::kFailedMoveToNextSequence;
752     }
753 
754     // If the chunk is contiguous but has not been patched yet move to the next
755     // sequence and try coming back here on the next ReadNextTracePacket() call.
756     // TODO(primiano): add a test to cover this, it's a subtle case.
757     if ((*it).flags & kChunkNeedsPatching)
758       return ReadAheadResult::kFailedMoveToNextSequence;
759 
760     // This is the case of an intermediate chunk which contains only one
761     // fragment which continues on the next chunk. This is the case for large
762     // packets, e.g.: [Packet0, Packet1(0)] [Packet1(1)] [Packet1(2), ...]
763     // (Packet1(X) := fragment X of Packet1).
764     if ((*it).num_fragments == 1 &&
765         ((*it).flags & kLastPacketContinuesOnNextChunk)) {
766       continue;
767     }
768 
769     // We made it! We got all fragments for the packet without holes.
770     TRACE_BUFFER_DLOG("  readahead success @ chunk %u", it.chunk_id());
771     PERFETTO_DCHECK(((*it).num_fragments == 1 &&
772                      !((*it).flags & kLastPacketContinuesOnNextChunk)) ||
773                     (*it).num_fragments > 1);
774 
775     // Now let's re-iterate over the [read_iter_, it] sequence and mark
776     // all the fragments as read.
777     bool packet_corruption = false;
778     for (;;) {
779       PERFETTO_DCHECK(read_iter_.is_valid());
780       TRACE_BUFFER_DLOG("    commit chunk %u", read_iter_.chunk_id());
781       if (PERFETTO_LIKELY((*read_iter_).num_fragments > 0)) {
782         // In the unlikely case of a corrupted packet (corrupted or empty
783         // fragment), invalidate the all stitching and move on to the next chunk
784         // in the same sequence, if any.
785         packet_corruption |= ReadNextPacketInChunk(&*read_iter_, packet) ==
786                              ReadPacketResult::kFailedInvalidPacket;
787       }
788       if (read_iter_.cur == it.cur)
789         break;
790       read_iter_.MoveNext();
791     }  // for(;;)
792     PERFETTO_DCHECK(read_iter_.cur == it.cur);
793 
794     if (PERFETTO_UNLIKELY(packet_corruption)) {
795       // ReadNextPacketInChunk() already records an abi violation for this case.
796       *packet = TracePacket();  // clear.
797       return ReadAheadResult::kFailedStayOnSameSequence;
798     }
799 
800     return ReadAheadResult::kSucceededReturnSlices;
801   }  // for(it...)  [readahead loop]
802   return ReadAheadResult::kFailedMoveToNextSequence;
803 }
804 
ReadNextPacketInChunk(ChunkMeta * chunk_meta,TracePacket * packet)805 TraceBuffer::ReadPacketResult TraceBuffer::ReadNextPacketInChunk(
806     ChunkMeta* chunk_meta,
807     TracePacket* packet) {
808   PERFETTO_DCHECK(chunk_meta->num_fragments_read < chunk_meta->num_fragments);
809   PERFETTO_DCHECK(!(chunk_meta->flags & kChunkNeedsPatching));
810 
811   const uint8_t* record_begin =
812       reinterpret_cast<const uint8_t*>(chunk_meta->chunk_record);
813   const uint8_t* record_end = record_begin + chunk_meta->chunk_record->size;
814   const uint8_t* packets_begin = record_begin + sizeof(ChunkRecord);
815   const uint8_t* packet_begin = packets_begin + chunk_meta->cur_fragment_offset;
816 
817   if (PERFETTO_UNLIKELY(packet_begin < packets_begin ||
818                         packet_begin >= record_end)) {
819     // The producer has a bug or is malicious and did declare that the chunk
820     // contains more packets beyond its boundaries.
821     stats_.set_abi_violations(stats_.abi_violations() + 1);
822     PERFETTO_DCHECK(suppress_sanity_dchecks_for_testing_);
823     chunk_meta->cur_fragment_offset = 0;
824     chunk_meta->num_fragments_read = chunk_meta->num_fragments;
825     if (PERFETTO_LIKELY(chunk_meta->is_complete())) {
826       stats_.set_chunks_read(stats_.chunks_read() + 1);
827       stats_.set_bytes_read(stats_.bytes_read() +
828                             chunk_meta->chunk_record->size);
829     }
830     return ReadPacketResult::kFailedInvalidPacket;
831   }
832 
833   // A packet (or a fragment) starts with a varint stating its size, followed
834   // by its content. The varint shouldn't be larger than 4 bytes (just in case
835   // the producer is using a redundant encoding)
836   uint64_t packet_size = 0;
837   const uint8_t* header_end =
838       std::min(packet_begin + protozero::proto_utils::kMessageLengthFieldSize,
839                record_end);
840   const uint8_t* packet_data = protozero::proto_utils::ParseVarInt(
841       packet_begin, header_end, &packet_size);
842 
843   const uint8_t* next_packet = packet_data + packet_size;
844   if (PERFETTO_UNLIKELY(next_packet <= packet_begin ||
845                         next_packet > record_end)) {
846     stats_.set_abi_violations(stats_.abi_violations() + 1);
847     PERFETTO_DCHECK(suppress_sanity_dchecks_for_testing_);
848     chunk_meta->cur_fragment_offset = 0;
849     chunk_meta->num_fragments_read = chunk_meta->num_fragments;
850     if (PERFETTO_LIKELY(chunk_meta->is_complete())) {
851       stats_.set_chunks_read(stats_.chunks_read() + 1);
852       stats_.set_bytes_read(stats_.bytes_read() +
853                             chunk_meta->chunk_record->size);
854     }
855     return ReadPacketResult::kFailedInvalidPacket;
856   }
857 
858   chunk_meta->cur_fragment_offset =
859       static_cast<uint16_t>(next_packet - packets_begin);
860   chunk_meta->num_fragments_read++;
861 
862   if (PERFETTO_UNLIKELY(chunk_meta->num_fragments_read ==
863                             chunk_meta->num_fragments &&
864                         chunk_meta->is_complete())) {
865     stats_.set_chunks_read(stats_.chunks_read() + 1);
866     stats_.set_bytes_read(stats_.bytes_read() + chunk_meta->chunk_record->size);
867   }
868 
869   chunk_meta->set_last_read_packet_skipped(false);
870 
871   if (PERFETTO_UNLIKELY(packet_size == 0))
872     return ReadPacketResult::kFailedEmptyPacket;
873 
874   if (PERFETTO_LIKELY(packet))
875     packet->AddSlice(packet_data, static_cast<size_t>(packet_size));
876 
877   return ReadPacketResult::kSucceeded;
878 }
879 
DiscardWrite()880 void TraceBuffer::DiscardWrite() {
881   PERFETTO_DCHECK(overwrite_policy_ == kDiscard);
882   discard_writes_ = true;
883   stats_.set_chunks_discarded(stats_.chunks_discarded() + 1);
884   TRACE_BUFFER_DLOG("  discarding write");
885 }
886 
887 }  // namespace perfetto
888