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<StringRef>
111DependencyScanningFilesystemSharedCache::getInvalidNegativeStatCachedPaths(
112 llvm::vfs::FileSystem &UnderlyingFS) const {
113 // Iterate through all shards and look for cached stat errors.
114 std::vector<StringRef> InvalidPaths;
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
121 if (Entry->getError()) {
122 // Only examine cached errors.
123 llvm::ErrorOr<llvm::vfs::Status> Stat = UnderlyingFS.status(Path);
124 if (Stat) {
125 // This is the case where we have cached the non-existence
126 // of the file at Path first, and a a file at the path is created
127 // later. The cache entry is not invalidated (as we have no good
128 // way to do it now), which may lead to missing file build errors.
129 InvalidPaths.push_back(x: Path);
130 }
131 }
132 }
133 }
134 return InvalidPaths;
135}
136
137const CachedFileSystemEntry *
138DependencyScanningFilesystemSharedCache::CacheShard::findEntryByFilename(
139 StringRef Filename) const {
140 assert(llvm::sys::path::is_absolute_gnu(Filename));
141 std::lock_guard<std::mutex> LockGuard(CacheLock);
142 auto It = CacheByFilename.find(Key: Filename);
143 return It == CacheByFilename.end() ? nullptr : It->getValue().first;
144}
145
146const CachedFileSystemEntry *
147DependencyScanningFilesystemSharedCache::CacheShard::findEntryByUID(
148 llvm::sys::fs::UniqueID UID) const {
149 std::lock_guard<std::mutex> LockGuard(CacheLock);
150 auto It = EntriesByUID.find(Val: UID);
151 return It == EntriesByUID.end() ? nullptr : It->getSecond();
152}
153
154const CachedFileSystemEntry &
155DependencyScanningFilesystemSharedCache::CacheShard::
156 getOrEmplaceEntryForFilename(StringRef Filename,
157 llvm::ErrorOr<llvm::vfs::Status> Stat) {
158 std::lock_guard<std::mutex> LockGuard(CacheLock);
159 auto [It, Inserted] = CacheByFilename.insert(KV: {Filename, {nullptr, nullptr}});
160 auto &[CachedEntry, CachedRealPath] = It->getValue();
161 if (!CachedEntry) {
162 // The entry is not present in the shared cache. Either the cache doesn't
163 // know about the file at all, or it only knows about its real path.
164 assert((Inserted || CachedRealPath) && "existing file with empty pair");
165 CachedEntry =
166 new (EntryStorage.Allocate()) CachedFileSystemEntry(std::move(Stat));
167 }
168 return *CachedEntry;
169}
170
171const CachedFileSystemEntry &
172DependencyScanningFilesystemSharedCache::CacheShard::getOrEmplaceEntryForUID(
173 llvm::sys::fs::UniqueID UID, llvm::vfs::Status Stat,
174 std::unique_ptr<llvm::MemoryBuffer> Contents) {
175 std::lock_guard<std::mutex> LockGuard(CacheLock);
176 auto [It, Inserted] = EntriesByUID.try_emplace(Key: UID);
177 auto &CachedEntry = It->getSecond();
178 if (Inserted) {
179 CachedFileContents *StoredContents = nullptr;
180 if (Contents)
181 StoredContents = new (ContentsStorage.Allocate())
182 CachedFileContents(std::move(Contents));
183 CachedEntry = new (EntryStorage.Allocate())
184 CachedFileSystemEntry(std::move(Stat), StoredContents);
185 }
186 return *CachedEntry;
187}
188
189const CachedFileSystemEntry &
190DependencyScanningFilesystemSharedCache::CacheShard::
191 getOrInsertEntryForFilename(StringRef Filename,
192 const CachedFileSystemEntry &Entry) {
193 std::lock_guard<std::mutex> LockGuard(CacheLock);
194 auto [It, Inserted] = CacheByFilename.insert(KV: {Filename, {&Entry, nullptr}});
195 auto &[CachedEntry, CachedRealPath] = It->getValue();
196 if (!Inserted || !CachedEntry)
197 CachedEntry = &Entry;
198 return *CachedEntry;
199}
200
201const CachedRealPath *
202DependencyScanningFilesystemSharedCache::CacheShard::findRealPathByFilename(
203 StringRef Filename) const {
204 assert(llvm::sys::path::is_absolute_gnu(Filename));
205 std::lock_guard<std::mutex> LockGuard(CacheLock);
206 auto It = CacheByFilename.find(Key: Filename);
207 return It == CacheByFilename.end() ? nullptr : It->getValue().second;
208}
209
210const CachedRealPath &DependencyScanningFilesystemSharedCache::CacheShard::
211 getOrEmplaceRealPathForFilename(StringRef Filename,
212 llvm::ErrorOr<llvm::StringRef> RealPath) {
213 std::lock_guard<std::mutex> LockGuard(CacheLock);
214
215 const CachedRealPath *&StoredRealPath = CacheByFilename[Filename].second;
216 if (!StoredRealPath) {
217 auto OwnedRealPath = [&]() -> CachedRealPath {
218 if (!RealPath)
219 return RealPath.getError();
220 return RealPath->str();
221 }();
222
223 StoredRealPath = new (RealPathStorage.Allocate())
224 CachedRealPath(std::move(OwnedRealPath));
225 }
226
227 return *StoredRealPath;
228}
229
230bool DependencyScanningWorkerFilesystem::shouldBypass(StringRef Path) const {
231 return BypassedPathPrefix && Path.starts_with(Prefix: *BypassedPathPrefix);
232}
233
234DependencyScanningWorkerFilesystem::DependencyScanningWorkerFilesystem(
235 DependencyScanningFilesystemSharedCache &SharedCache,
236 IntrusiveRefCntPtr<llvm::vfs::FileSystem> FS)
237 : llvm::RTTIExtends<DependencyScanningWorkerFilesystem,
238 llvm::vfs::ProxyFileSystem>(std::move(FS)),
239 SharedCache(SharedCache),
240 WorkingDirForCacheLookup(llvm::errc::invalid_argument) {
241 updateWorkingDirForCacheLookup();
242}
243
244const CachedFileSystemEntry &
245DependencyScanningWorkerFilesystem::getOrEmplaceSharedEntryForUID(
246 TentativeEntry TEntry) {
247 auto &Shard = SharedCache.getShardForUID(UID: TEntry.Status.getUniqueID());
248 return Shard.getOrEmplaceEntryForUID(UID: TEntry.Status.getUniqueID(),
249 Stat: std::move(TEntry.Status),
250 Contents: std::move(TEntry.Contents));
251}
252
253const CachedFileSystemEntry *
254DependencyScanningWorkerFilesystem::findEntryByFilenameWithWriteThrough(
255 StringRef Filename) {
256 if (const auto *Entry = LocalCache.findEntryByFilename(Filename))
257 return Entry;
258 auto &Shard = SharedCache.getShardForFilename(Filename);
259 if (const auto *Entry = Shard.findEntryByFilename(Filename))
260 return &LocalCache.insertEntryForFilename(Filename, Entry: *Entry);
261 return nullptr;
262}
263
264llvm::ErrorOr<const CachedFileSystemEntry &>
265DependencyScanningWorkerFilesystem::computeAndStoreResult(
266 StringRef OriginalFilename, StringRef FilenameForLookup) {
267 llvm::ErrorOr<llvm::vfs::Status> Stat =
268 getUnderlyingFS().status(Path: OriginalFilename);
269 if (!Stat) {
270 const auto &Entry =
271 getOrEmplaceSharedEntryForFilename(Filename: FilenameForLookup, EC: Stat.getError());
272 return insertLocalEntryForFilename(Filename: FilenameForLookup, Entry);
273 }
274
275 if (const auto *Entry = findSharedEntryByUID(Stat: *Stat))
276 return insertLocalEntryForFilename(Filename: FilenameForLookup, Entry: *Entry);
277
278 auto TEntry =
279 Stat->isDirectory() ? TentativeEntry(*Stat) : readFile(Filename: OriginalFilename);
280
281 const CachedFileSystemEntry *SharedEntry = [&]() {
282 if (TEntry) {
283 const auto &UIDEntry = getOrEmplaceSharedEntryForUID(TEntry: std::move(*TEntry));
284 return &getOrInsertSharedEntryForFilename(Filename: FilenameForLookup, Entry: UIDEntry);
285 }
286 return &getOrEmplaceSharedEntryForFilename(Filename: FilenameForLookup,
287 EC: TEntry.getError());
288 }();
289
290 return insertLocalEntryForFilename(Filename: FilenameForLookup, Entry: *SharedEntry);
291}
292
293llvm::ErrorOr<EntryRef>
294DependencyScanningWorkerFilesystem::getOrCreateFileSystemEntry(
295 StringRef OriginalFilename) {
296 SmallString<256> PathBuf;
297 auto FilenameForLookup = tryGetFilenameForLookup(OriginalFilename, PathBuf);
298 if (!FilenameForLookup)
299 return FilenameForLookup.getError();
300
301 if (const auto *Entry =
302 findEntryByFilenameWithWriteThrough(Filename: *FilenameForLookup))
303 return EntryRef(OriginalFilename, *Entry).unwrapError();
304 auto MaybeEntry = computeAndStoreResult(OriginalFilename, FilenameForLookup: *FilenameForLookup);
305 if (!MaybeEntry)
306 return MaybeEntry.getError();
307 return EntryRef(OriginalFilename, *MaybeEntry).unwrapError();
308}
309
310llvm::ErrorOr<llvm::vfs::Status>
311DependencyScanningWorkerFilesystem::status(const Twine &Path) {
312 SmallString<256> OwnedFilename;
313 StringRef Filename = Path.toStringRef(Out&: OwnedFilename);
314
315 if (shouldBypass(Path: Filename))
316 return getUnderlyingFS().status(Path);
317
318 llvm::ErrorOr<EntryRef> Result = getOrCreateFileSystemEntry(OriginalFilename: Filename);
319 if (!Result)
320 return Result.getError();
321 return Result->getStatus();
322}
323
324bool DependencyScanningWorkerFilesystem::exists(const Twine &Path) {
325 // While some VFS overlay filesystems may implement more-efficient
326 // mechanisms for `exists` queries, `DependencyScanningWorkerFilesystem`
327 // typically wraps `RealFileSystem` which does not specialize `exists`,
328 // so it is not likely to benefit from such optimizations. Instead,
329 // it is more-valuable to have this query go through the
330 // cached-`status` code-path of the `DependencyScanningWorkerFilesystem`.
331 llvm::ErrorOr<llvm::vfs::Status> Status = status(Path);
332 return Status && Status->exists();
333}
334
335namespace {
336
337/// The VFS that is used by clang consumes the \c CachedFileSystemEntry using
338/// this subclass.
339class DepScanFile final : public llvm::vfs::File {
340public:
341 DepScanFile(std::unique_ptr<llvm::MemoryBuffer> Buffer,
342 llvm::vfs::Status Stat)
343 : Buffer(std::move(Buffer)), Stat(std::move(Stat)) {}
344
345 static llvm::ErrorOr<std::unique_ptr<llvm::vfs::File>> create(EntryRef Entry);
346
347 llvm::ErrorOr<llvm::vfs::Status> status() override { return Stat; }
348
349 llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>>
350 getBuffer(const Twine &Name, int64_t FileSize, bool RequiresNullTerminator,
351 bool IsVolatile) override {
352 return std::move(Buffer);
353 }
354
355 std::error_code close() override { return {}; }
356
357private:
358 std::unique_ptr<llvm::MemoryBuffer> Buffer;
359 llvm::vfs::Status Stat;
360};
361
362} // end anonymous namespace
363
364llvm::ErrorOr<std::unique_ptr<llvm::vfs::File>>
365DepScanFile::create(EntryRef Entry) {
366 assert(!Entry.isError() && "error");
367
368 if (Entry.isDirectory())
369 return std::make_error_code(e: std::errc::is_a_directory);
370
371 auto Result = std::make_unique<DepScanFile>(
372 args: llvm::MemoryBuffer::getMemBuffer(InputData: Entry.getContents(),
373 BufferName: Entry.getStatus().getName(),
374 /*RequiresNullTerminator=*/false),
375 args: Entry.getStatus());
376
377 return llvm::ErrorOr<std::unique_ptr<llvm::vfs::File>>(
378 std::unique_ptr<llvm::vfs::File>(std::move(Result)));
379}
380
381llvm::ErrorOr<std::unique_ptr<llvm::vfs::File>>
382DependencyScanningWorkerFilesystem::openFileForRead(const Twine &Path) {
383 SmallString<256> OwnedFilename;
384 StringRef Filename = Path.toStringRef(Out&: OwnedFilename);
385
386 if (shouldBypass(Path: Filename))
387 return getUnderlyingFS().openFileForRead(Path);
388
389 llvm::ErrorOr<EntryRef> Result = getOrCreateFileSystemEntry(OriginalFilename: Filename);
390 if (!Result)
391 return Result.getError();
392 return DepScanFile::create(Entry: Result.get());
393}
394
395std::error_code
396DependencyScanningWorkerFilesystem::getRealPath(const Twine &Path,
397 SmallVectorImpl<char> &Output) {
398 SmallString<256> OwnedFilename;
399 StringRef OriginalFilename = Path.toStringRef(Out&: OwnedFilename);
400
401 if (shouldBypass(Path: OriginalFilename))
402 return getUnderlyingFS().getRealPath(Path, Output);
403
404 SmallString<256> PathBuf;
405 auto FilenameForLookup = tryGetFilenameForLookup(OriginalFilename, PathBuf);
406 if (!FilenameForLookup)
407 return FilenameForLookup.getError();
408
409 auto HandleCachedRealPath =
410 [&Output](const CachedRealPath &RealPath) -> std::error_code {
411 if (!RealPath)
412 return RealPath.getError();
413 Output.assign(in_start: RealPath->begin(), in_end: RealPath->end());
414 return {};
415 };
416
417 // If we already have the result in local cache, no work required.
418 if (const auto *RealPath =
419 LocalCache.findRealPathByFilename(Filename: *FilenameForLookup))
420 return HandleCachedRealPath(*RealPath);
421
422 // If we have the result in the shared cache, cache it locally.
423 auto &Shard = SharedCache.getShardForFilename(Filename: *FilenameForLookup);
424 if (const auto *ShardRealPath =
425 Shard.findRealPathByFilename(Filename: *FilenameForLookup)) {
426 const auto &RealPath = LocalCache.insertRealPathForFilename(
427 Filename: *FilenameForLookup, RealPath: *ShardRealPath);
428 return HandleCachedRealPath(RealPath);
429 }
430
431 // If we don't know the real path, compute it...
432 std::error_code EC = getUnderlyingFS().getRealPath(Path: OriginalFilename, Output);
433 llvm::ErrorOr<llvm::StringRef> ComputedRealPath = EC;
434 if (!EC)
435 ComputedRealPath = StringRef{Output.data(), Output.size()};
436
437 // ...and try to write it into the shared cache. In case some other thread won
438 // this race and already wrote its own result there, just adopt it. Write
439 // whatever is in the shared cache into the local one.
440 const auto &RealPath = Shard.getOrEmplaceRealPathForFilename(
441 Filename: *FilenameForLookup, RealPath: ComputedRealPath);
442 return HandleCachedRealPath(
443 LocalCache.insertRealPathForFilename(Filename: *FilenameForLookup, RealPath));
444}
445
446std::error_code DependencyScanningWorkerFilesystem::setCurrentWorkingDirectory(
447 const Twine &Path) {
448 std::error_code EC = ProxyFileSystem::setCurrentWorkingDirectory(Path);
449 updateWorkingDirForCacheLookup();
450 return EC;
451}
452
453void DependencyScanningWorkerFilesystem::updateWorkingDirForCacheLookup() {
454 llvm::ErrorOr<std::string> CWD =
455 getUnderlyingFS().getCurrentWorkingDirectory();
456 if (!CWD) {
457 WorkingDirForCacheLookup = CWD.getError();
458 } else if (!llvm::sys::path::is_absolute_gnu(path: *CWD)) {
459 WorkingDirForCacheLookup = llvm::errc::invalid_argument;
460 } else {
461 WorkingDirForCacheLookup = *CWD;
462 }
463 assert(!WorkingDirForCacheLookup ||
464 llvm::sys::path::is_absolute_gnu(*WorkingDirForCacheLookup));
465}
466
467llvm::ErrorOr<StringRef>
468DependencyScanningWorkerFilesystem::tryGetFilenameForLookup(
469 StringRef OriginalFilename, llvm::SmallVectorImpl<char> &PathBuf) const {
470 StringRef FilenameForLookup;
471 if (llvm::sys::path::is_absolute_gnu(path: OriginalFilename)) {
472 FilenameForLookup = OriginalFilename;
473 } else if (!WorkingDirForCacheLookup) {
474 return WorkingDirForCacheLookup.getError();
475 } else {
476 StringRef RelFilename = OriginalFilename;
477 RelFilename.consume_front(Prefix: "./");
478 PathBuf.assign(in_start: WorkingDirForCacheLookup->begin(),
479 in_end: WorkingDirForCacheLookup->end());
480 llvm::sys::path::append(path&: PathBuf, a: RelFilename);
481 FilenameForLookup = StringRef{PathBuf.begin(), PathBuf.size()};
482 }
483 assert(llvm::sys::path::is_absolute_gnu(FilenameForLookup));
484 return FilenameForLookup;
485}
486
487const char DependencyScanningWorkerFilesystem::ID = 0;
488

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