1//===- DependencyScanningFilesystem.cpp - clang-scan-deps fs --------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8
9#include "clang/Tooling/DependencyScanning/DependencyScanningFilesystem.h"
10#include "llvm/Support/MemoryBuffer.h"
11#include "llvm/Support/Threading.h"
12#include <optional>
13
14using namespace clang;
15using namespace tooling;
16using namespace dependencies;
17
18llvm::ErrorOr<DependencyScanningWorkerFilesystem::TentativeEntry>
19DependencyScanningWorkerFilesystem::readFile(StringRef Filename) {
20 // Load the file and its content from the file system.
21 auto MaybeFile = getUnderlyingFS().openFileForRead(Path: Filename);
22 if (!MaybeFile)
23 return MaybeFile.getError();
24 auto File = std::move(*MaybeFile);
25
26 auto MaybeStat = File->status();
27 if (!MaybeStat)
28 return MaybeStat.getError();
29 auto Stat = std::move(*MaybeStat);
30
31 auto MaybeBuffer = File->getBuffer(Name: Stat.getName());
32 if (!MaybeBuffer)
33 return MaybeBuffer.getError();
34 auto Buffer = std::move(*MaybeBuffer);
35
36 // If the file size changed between read and stat, pretend it didn't.
37 if (Stat.getSize() != Buffer->getBufferSize())
38 Stat = llvm::vfs::Status::copyWithNewSize(In: Stat, NewSize: Buffer->getBufferSize());
39
40 return TentativeEntry(Stat, std::move(Buffer));
41}
42
43bool DependencyScanningWorkerFilesystem::ensureDirectiveTokensArePopulated(
44 EntryRef Ref) {
45 auto &Entry = Ref.Entry;
46
47 if (Entry.isError() || Entry.isDirectory())
48 return false;
49
50 CachedFileContents *Contents = Entry.getCachedContents();
51 assert(Contents && "contents not initialized");
52
53 // Double-checked locking.
54 if (Contents->DepDirectives.load())
55 return true;
56
57 std::lock_guard<std::mutex> GuardLock(Contents->ValueLock);
58
59 // Double-checked locking.
60 if (Contents->DepDirectives.load())
61 return true;
62
63 SmallVector<dependency_directives_scan::Directive, 64> Directives;
64 // Scan the file for preprocessor directives that might affect the
65 // dependencies.
66 if (scanSourceForDependencyDirectives(Input: Contents->Original->getBuffer(),
67 Tokens&: Contents->DepDirectiveTokens,
68 Directives)) {
69 Contents->DepDirectiveTokens.clear();
70 // FIXME: Propagate the diagnostic if desired by the client.
71 Contents->DepDirectives.store(p: new std::optional<DependencyDirectivesTy>());
72 return false;
73 }
74
75 // This function performed double-checked locking using `DepDirectives`.
76 // Assigning it must be the last thing this function does, otherwise other
77 // threads may skip the critical section (`DepDirectives != nullptr`), leading
78 // to a data race.
79 Contents->DepDirectives.store(
80 p: new std::optional<DependencyDirectivesTy>(std::move(Directives)));
81 return true;
82}
83
84DependencyScanningFilesystemSharedCache::
85 DependencyScanningFilesystemSharedCache() {
86 // This heuristic was chosen using a empirical testing on a
87 // reasonably high core machine (iMacPro 18 cores / 36 threads). The cache
88 // sharding gives a performance edge by reducing the lock contention.
89 // FIXME: A better heuristic might also consider the OS to account for
90 // the different cost of lock contention on different OSes.
91 NumShards =
92 std::max(a: 2u, b: llvm::hardware_concurrency().compute_thread_count() / 4);
93 CacheShards = std::make_unique<CacheShard[]>(num: NumShards);
94}
95
96DependencyScanningFilesystemSharedCache::CacheShard &
97DependencyScanningFilesystemSharedCache::getShardForFilename(
98 StringRef Filename) const {
99 assert(llvm::sys::path::is_absolute_gnu(Filename));
100 return CacheShards[llvm::hash_value(S: Filename) % NumShards];
101}
102
103DependencyScanningFilesystemSharedCache::CacheShard &
104DependencyScanningFilesystemSharedCache::getShardForUID(
105 llvm::sys::fs::UniqueID UID) const {
106 auto Hash = llvm::hash_combine(args: UID.getDevice(), args: UID.getFile());
107 return CacheShards[Hash % NumShards];
108}
109
110std::vector<DependencyScanningFilesystemSharedCache::OutOfDateEntry>
111DependencyScanningFilesystemSharedCache::getOutOfDateEntries(
112 llvm::vfs::FileSystem &UnderlyingFS) const {
113 // Iterate through all shards and look for cached stat errors.
114 std::vector<OutOfDateEntry> InvalidDiagInfo;
115 for (unsigned i = 0; i < NumShards; i++) {
116 const CacheShard &Shard = CacheShards[i];
117 std::lock_guard<std::mutex> LockGuard(Shard.CacheLock);
118 for (const auto &[Path, CachedPair] : Shard.CacheByFilename) {
119 const CachedFileSystemEntry *Entry = CachedPair.first;
120 llvm::ErrorOr<llvm::vfs::Status> Status = UnderlyingFS.status(Path);
121 if (Status) {
122 if (Entry->getError()) {
123 // This is the case where we have cached the non-existence
124 // of the file at Path first, and a file at the path is created
125 // later. The cache entry is not invalidated (as we have no good
126 // way to do it now), which may lead to missing file build errors.
127 InvalidDiagInfo.emplace_back(args: Path.data());
128 } else {
129 llvm::vfs::Status CachedStatus = Entry->getStatus();
130 if (Status->getType() == llvm::sys::fs::file_type::regular_file &&
131 Status->getType() == CachedStatus.getType()) {
132 // We only check regular files. Directory files sizes could change
133 // due to content changes, and reporting directory size changes can
134 // lead to false positives.
135 // TODO: At the moment, we do not detect symlinks to files whose
136 // size may change. We need to decide if we want to detect cached
137 // symlink size changes. We can also expand this to detect file
138 // type changes.
139 uint64_t CachedSize = CachedStatus.getSize();
140 uint64_t ActualSize = Status->getSize();
141 if (CachedSize != ActualSize) {
142 // This is the case where the cached file has a different size
143 // from the actual file that comes from the underlying FS.
144 InvalidDiagInfo.emplace_back(args: Path.data(), args&: CachedSize, args&: ActualSize);
145 }
146 }
147 }
148 }
149 }
150 }
151 return InvalidDiagInfo;
152}
153
154const CachedFileSystemEntry *
155DependencyScanningFilesystemSharedCache::CacheShard::findEntryByFilename(
156 StringRef Filename) const {
157 assert(llvm::sys::path::is_absolute_gnu(Filename));
158 std::lock_guard<std::mutex> LockGuard(CacheLock);
159 auto It = CacheByFilename.find(Key: Filename);
160 return It == CacheByFilename.end() ? nullptr : It->getValue().first;
161}
162
163const CachedFileSystemEntry *
164DependencyScanningFilesystemSharedCache::CacheShard::findEntryByUID(
165 llvm::sys::fs::UniqueID UID) const {
166 std::lock_guard<std::mutex> LockGuard(CacheLock);
167 auto It = EntriesByUID.find(Val: UID);
168 return It == EntriesByUID.end() ? nullptr : It->getSecond();
169}
170
171const CachedFileSystemEntry &
172DependencyScanningFilesystemSharedCache::CacheShard::
173 getOrEmplaceEntryForFilename(StringRef Filename,
174 llvm::ErrorOr<llvm::vfs::Status> Stat) {
175 std::lock_guard<std::mutex> LockGuard(CacheLock);
176 auto [It, Inserted] = CacheByFilename.insert(KV: {Filename, {nullptr, nullptr}});
177 auto &[CachedEntry, CachedRealPath] = It->getValue();
178 if (!CachedEntry) {
179 // The entry is not present in the shared cache. Either the cache doesn't
180 // know about the file at all, or it only knows about its real path.
181 assert((Inserted || CachedRealPath) && "existing file with empty pair");
182 CachedEntry =
183 new (EntryStorage.Allocate()) CachedFileSystemEntry(std::move(Stat));
184 }
185 return *CachedEntry;
186}
187
188const CachedFileSystemEntry &
189DependencyScanningFilesystemSharedCache::CacheShard::getOrEmplaceEntryForUID(
190 llvm::sys::fs::UniqueID UID, llvm::vfs::Status Stat,
191 std::unique_ptr<llvm::MemoryBuffer> Contents) {
192 std::lock_guard<std::mutex> LockGuard(CacheLock);
193 auto [It, Inserted] = EntriesByUID.try_emplace(Key: UID);
194 auto &CachedEntry = It->getSecond();
195 if (Inserted) {
196 CachedFileContents *StoredContents = nullptr;
197 if (Contents)
198 StoredContents = new (ContentsStorage.Allocate())
199 CachedFileContents(std::move(Contents));
200 CachedEntry = new (EntryStorage.Allocate())
201 CachedFileSystemEntry(std::move(Stat), StoredContents);
202 }
203 return *CachedEntry;
204}
205
206const CachedFileSystemEntry &
207DependencyScanningFilesystemSharedCache::CacheShard::
208 getOrInsertEntryForFilename(StringRef Filename,
209 const CachedFileSystemEntry &Entry) {
210 std::lock_guard<std::mutex> LockGuard(CacheLock);
211 auto [It, Inserted] = CacheByFilename.insert(KV: {Filename, {&Entry, nullptr}});
212 auto &[CachedEntry, CachedRealPath] = It->getValue();
213 if (!Inserted || !CachedEntry)
214 CachedEntry = &Entry;
215 return *CachedEntry;
216}
217
218const CachedRealPath *
219DependencyScanningFilesystemSharedCache::CacheShard::findRealPathByFilename(
220 StringRef Filename) const {
221 assert(llvm::sys::path::is_absolute_gnu(Filename));
222 std::lock_guard<std::mutex> LockGuard(CacheLock);
223 auto It = CacheByFilename.find(Key: Filename);
224 return It == CacheByFilename.end() ? nullptr : It->getValue().second;
225}
226
227const CachedRealPath &DependencyScanningFilesystemSharedCache::CacheShard::
228 getOrEmplaceRealPathForFilename(StringRef Filename,
229 llvm::ErrorOr<llvm::StringRef> RealPath) {
230 std::lock_guard<std::mutex> LockGuard(CacheLock);
231
232 const CachedRealPath *&StoredRealPath = CacheByFilename[Filename].second;
233 if (!StoredRealPath) {
234 auto OwnedRealPath = [&]() -> CachedRealPath {
235 if (!RealPath)
236 return RealPath.getError();
237 return RealPath->str();
238 }();
239
240 StoredRealPath = new (RealPathStorage.Allocate())
241 CachedRealPath(std::move(OwnedRealPath));
242 }
243
244 return *StoredRealPath;
245}
246
247bool DependencyScanningWorkerFilesystem::shouldBypass(StringRef Path) const {
248 return BypassedPathPrefix && Path.starts_with(Prefix: *BypassedPathPrefix);
249}
250
251DependencyScanningWorkerFilesystem::DependencyScanningWorkerFilesystem(
252 DependencyScanningFilesystemSharedCache &SharedCache,
253 IntrusiveRefCntPtr<llvm::vfs::FileSystem> FS)
254 : llvm::RTTIExtends<DependencyScanningWorkerFilesystem,
255 llvm::vfs::ProxyFileSystem>(std::move(FS)),
256 SharedCache(SharedCache),
257 WorkingDirForCacheLookup(llvm::errc::invalid_argument) {
258 updateWorkingDirForCacheLookup();
259}
260
261const CachedFileSystemEntry &
262DependencyScanningWorkerFilesystem::getOrEmplaceSharedEntryForUID(
263 TentativeEntry TEntry) {
264 auto &Shard = SharedCache.getShardForUID(UID: TEntry.Status.getUniqueID());
265 return Shard.getOrEmplaceEntryForUID(UID: TEntry.Status.getUniqueID(),
266 Stat: std::move(TEntry.Status),
267 Contents: std::move(TEntry.Contents));
268}
269
270const CachedFileSystemEntry *
271DependencyScanningWorkerFilesystem::findEntryByFilenameWithWriteThrough(
272 StringRef Filename) {
273 if (const auto *Entry = LocalCache.findEntryByFilename(Filename))
274 return Entry;
275 auto &Shard = SharedCache.getShardForFilename(Filename);
276 if (const auto *Entry = Shard.findEntryByFilename(Filename))
277 return &LocalCache.insertEntryForFilename(Filename, Entry: *Entry);
278 return nullptr;
279}
280
281llvm::ErrorOr<const CachedFileSystemEntry &>
282DependencyScanningWorkerFilesystem::computeAndStoreResult(
283 StringRef OriginalFilename, StringRef FilenameForLookup) {
284 llvm::ErrorOr<llvm::vfs::Status> Stat =
285 getUnderlyingFS().status(Path: OriginalFilename);
286 if (!Stat) {
287 const auto &Entry =
288 getOrEmplaceSharedEntryForFilename(Filename: FilenameForLookup, EC: Stat.getError());
289 return insertLocalEntryForFilename(Filename: FilenameForLookup, Entry);
290 }
291
292 if (const auto *Entry = findSharedEntryByUID(Stat: *Stat))
293 return insertLocalEntryForFilename(Filename: FilenameForLookup, Entry: *Entry);
294
295 auto TEntry =
296 Stat->isDirectory() ? TentativeEntry(*Stat) : readFile(Filename: OriginalFilename);
297
298 const CachedFileSystemEntry *SharedEntry = [&]() {
299 if (TEntry) {
300 const auto &UIDEntry = getOrEmplaceSharedEntryForUID(TEntry: std::move(*TEntry));
301 return &getOrInsertSharedEntryForFilename(Filename: FilenameForLookup, Entry: UIDEntry);
302 }
303 return &getOrEmplaceSharedEntryForFilename(Filename: FilenameForLookup,
304 EC: TEntry.getError());
305 }();
306
307 return insertLocalEntryForFilename(Filename: FilenameForLookup, Entry: *SharedEntry);
308}
309
310llvm::ErrorOr<EntryRef>
311DependencyScanningWorkerFilesystem::getOrCreateFileSystemEntry(
312 StringRef OriginalFilename) {
313 SmallString<256> PathBuf;
314 auto FilenameForLookup = tryGetFilenameForLookup(OriginalFilename, PathBuf);
315 if (!FilenameForLookup)
316 return FilenameForLookup.getError();
317
318 if (const auto *Entry =
319 findEntryByFilenameWithWriteThrough(Filename: *FilenameForLookup))
320 return EntryRef(OriginalFilename, *Entry).unwrapError();
321 auto MaybeEntry = computeAndStoreResult(OriginalFilename, FilenameForLookup: *FilenameForLookup);
322 if (!MaybeEntry)
323 return MaybeEntry.getError();
324 return EntryRef(OriginalFilename, *MaybeEntry).unwrapError();
325}
326
327llvm::ErrorOr<llvm::vfs::Status>
328DependencyScanningWorkerFilesystem::status(const Twine &Path) {
329 SmallString<256> OwnedFilename;
330 StringRef Filename = Path.toStringRef(Out&: OwnedFilename);
331
332 if (shouldBypass(Path: Filename))
333 return getUnderlyingFS().status(Path);
334
335 llvm::ErrorOr<EntryRef> Result = getOrCreateFileSystemEntry(OriginalFilename: Filename);
336 if (!Result)
337 return Result.getError();
338 return Result->getStatus();
339}
340
341bool DependencyScanningWorkerFilesystem::exists(const Twine &Path) {
342 // While some VFS overlay filesystems may implement more-efficient
343 // mechanisms for `exists` queries, `DependencyScanningWorkerFilesystem`
344 // typically wraps `RealFileSystem` which does not specialize `exists`,
345 // so it is not likely to benefit from such optimizations. Instead,
346 // it is more-valuable to have this query go through the
347 // cached-`status` code-path of the `DependencyScanningWorkerFilesystem`.
348 llvm::ErrorOr<llvm::vfs::Status> Status = status(Path);
349 return Status && Status->exists();
350}
351
352namespace {
353
354/// The VFS that is used by clang consumes the \c CachedFileSystemEntry using
355/// this subclass.
356class DepScanFile final : public llvm::vfs::File {
357public:
358 DepScanFile(std::unique_ptr<llvm::MemoryBuffer> Buffer,
359 llvm::vfs::Status Stat)
360 : Buffer(std::move(Buffer)), Stat(std::move(Stat)) {}
361
362 static llvm::ErrorOr<std::unique_ptr<llvm::vfs::File>> create(EntryRef Entry);
363
364 llvm::ErrorOr<llvm::vfs::Status> status() override { return Stat; }
365
366 llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>>
367 getBuffer(const Twine &Name, int64_t FileSize, bool RequiresNullTerminator,
368 bool IsVolatile) override {
369 return std::move(Buffer);
370 }
371
372 std::error_code close() override { return {}; }
373
374private:
375 std::unique_ptr<llvm::MemoryBuffer> Buffer;
376 llvm::vfs::Status Stat;
377};
378
379} // end anonymous namespace
380
381llvm::ErrorOr<std::unique_ptr<llvm::vfs::File>>
382DepScanFile::create(EntryRef Entry) {
383 assert(!Entry.isError() && "error");
384
385 if (Entry.isDirectory())
386 return std::make_error_code(e: std::errc::is_a_directory);
387
388 auto Result = std::make_unique<DepScanFile>(
389 args: llvm::MemoryBuffer::getMemBuffer(InputData: Entry.getContents(),
390 BufferName: Entry.getStatus().getName(),
391 /*RequiresNullTerminator=*/false),
392 args: Entry.getStatus());
393
394 return llvm::ErrorOr<std::unique_ptr<llvm::vfs::File>>(
395 std::unique_ptr<llvm::vfs::File>(std::move(Result)));
396}
397
398llvm::ErrorOr<std::unique_ptr<llvm::vfs::File>>
399DependencyScanningWorkerFilesystem::openFileForRead(const Twine &Path) {
400 SmallString<256> OwnedFilename;
401 StringRef Filename = Path.toStringRef(Out&: OwnedFilename);
402
403 if (shouldBypass(Path: Filename))
404 return getUnderlyingFS().openFileForRead(Path);
405
406 llvm::ErrorOr<EntryRef> Result = getOrCreateFileSystemEntry(OriginalFilename: Filename);
407 if (!Result)
408 return Result.getError();
409 return DepScanFile::create(Entry: Result.get());
410}
411
412std::error_code
413DependencyScanningWorkerFilesystem::getRealPath(const Twine &Path,
414 SmallVectorImpl<char> &Output) {
415 SmallString<256> OwnedFilename;
416 StringRef OriginalFilename = Path.toStringRef(Out&: OwnedFilename);
417
418 if (shouldBypass(Path: OriginalFilename))
419 return getUnderlyingFS().getRealPath(Path, Output);
420
421 SmallString<256> PathBuf;
422 auto FilenameForLookup = tryGetFilenameForLookup(OriginalFilename, PathBuf);
423 if (!FilenameForLookup)
424 return FilenameForLookup.getError();
425
426 auto HandleCachedRealPath =
427 [&Output](const CachedRealPath &RealPath) -> std::error_code {
428 if (!RealPath)
429 return RealPath.getError();
430 Output.assign(in_start: RealPath->begin(), in_end: RealPath->end());
431 return {};
432 };
433
434 // If we already have the result in local cache, no work required.
435 if (const auto *RealPath =
436 LocalCache.findRealPathByFilename(Filename: *FilenameForLookup))
437 return HandleCachedRealPath(*RealPath);
438
439 // If we have the result in the shared cache, cache it locally.
440 auto &Shard = SharedCache.getShardForFilename(Filename: *FilenameForLookup);
441 if (const auto *ShardRealPath =
442 Shard.findRealPathByFilename(Filename: *FilenameForLookup)) {
443 const auto &RealPath = LocalCache.insertRealPathForFilename(
444 Filename: *FilenameForLookup, RealPath: *ShardRealPath);
445 return HandleCachedRealPath(RealPath);
446 }
447
448 // If we don't know the real path, compute it...
449 std::error_code EC = getUnderlyingFS().getRealPath(Path: OriginalFilename, Output);
450 llvm::ErrorOr<llvm::StringRef> ComputedRealPath = EC;
451 if (!EC)
452 ComputedRealPath = StringRef{Output.data(), Output.size()};
453
454 // ...and try to write it into the shared cache. In case some other thread won
455 // this race and already wrote its own result there, just adopt it. Write
456 // whatever is in the shared cache into the local one.
457 const auto &RealPath = Shard.getOrEmplaceRealPathForFilename(
458 Filename: *FilenameForLookup, RealPath: ComputedRealPath);
459 return HandleCachedRealPath(
460 LocalCache.insertRealPathForFilename(Filename: *FilenameForLookup, RealPath));
461}
462
463std::error_code DependencyScanningWorkerFilesystem::setCurrentWorkingDirectory(
464 const Twine &Path) {
465 std::error_code EC = ProxyFileSystem::setCurrentWorkingDirectory(Path);
466 updateWorkingDirForCacheLookup();
467 return EC;
468}
469
470void DependencyScanningWorkerFilesystem::updateWorkingDirForCacheLookup() {
471 llvm::ErrorOr<std::string> CWD =
472 getUnderlyingFS().getCurrentWorkingDirectory();
473 if (!CWD) {
474 WorkingDirForCacheLookup = CWD.getError();
475 } else if (!llvm::sys::path::is_absolute_gnu(path: *CWD)) {
476 WorkingDirForCacheLookup = llvm::errc::invalid_argument;
477 } else {
478 WorkingDirForCacheLookup = *CWD;
479 }
480 assert(!WorkingDirForCacheLookup ||
481 llvm::sys::path::is_absolute_gnu(*WorkingDirForCacheLookup));
482}
483
484llvm::ErrorOr<StringRef>
485DependencyScanningWorkerFilesystem::tryGetFilenameForLookup(
486 StringRef OriginalFilename, llvm::SmallVectorImpl<char> &PathBuf) const {
487 StringRef FilenameForLookup;
488 if (llvm::sys::path::is_absolute_gnu(path: OriginalFilename)) {
489 FilenameForLookup = OriginalFilename;
490 } else if (!WorkingDirForCacheLookup) {
491 return WorkingDirForCacheLookup.getError();
492 } else {
493 StringRef RelFilename = OriginalFilename;
494 RelFilename.consume_front(Prefix: "./");
495 PathBuf.assign(in_start: WorkingDirForCacheLookup->begin(),
496 in_end: WorkingDirForCacheLookup->end());
497 llvm::sys::path::append(path&: PathBuf, a: RelFilename);
498 FilenameForLookup = StringRef{PathBuf.begin(), PathBuf.size()};
499 }
500 assert(llvm::sys::path::is_absolute_gnu(FilenameForLookup));
501 return FilenameForLookup;
502}
503
504const char DependencyScanningWorkerFilesystem::ID = 0;
505

source code of clang/lib/Tooling/DependencyScanning/DependencyScanningFilesystem.cpp