[lsan] Common leak checking module.

Leak checking functionality which will be shared between
LSan/ASan/MSan.

llvm-svn: 182249
This commit is contained in:
Sergey Matveev 2013-05-20 11:06:50 +00:00
parent 3d97cdd140
commit b5483be858
3 changed files with 692 additions and 0 deletions

View File

@ -0,0 +1,380 @@
//=-- lsan_common.cc ------------------------------------------------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file is a part of LeakSanitizer.
// Implementation of common leak checking functionality.
//
//===----------------------------------------------------------------------===//
#include "lsan_common.h"
#include "sanitizer_common/sanitizer_common.h"
#include "sanitizer_common/sanitizer_flags.h"
#include "sanitizer_common/sanitizer_stackdepot.h"
#include "sanitizer_common/sanitizer_stacktrace.h"
#include "sanitizer_common/sanitizer_stoptheworld.h"
namespace __lsan {
Flags lsan_flags;
static void InitializeFlags() {
Flags *f = flags();
// Default values.
f->sources = kSourceAllAligned;
f->report_blocks = false;
f->resolution = 0;
f->max_leaks = 0;
f->log_pointers = false;
f->log_threads = false;
const char *options = GetEnv("LSAN_OPTIONS");
if (options) {
bool aligned = true;
ParseFlag(options, &aligned, "aligned");
if (!aligned) f->sources |= kSourceUnaligned;
ParseFlag(options, &f->report_blocks, "report_blocks");
ParseFlag(options, &f->resolution, "resolution");
CHECK_GE(&f->resolution, 0);
ParseFlag(options, &f->max_leaks, "max_leaks");
CHECK_GE(&f->max_leaks, 0);
ParseFlag(options, &f->log_pointers, "log_pointers");
ParseFlag(options, &f->log_threads, "log_threads");
}
}
void InitCommonLsan() {
InitializeFlags();
InitializePlatformSpecificModules();
}
static inline bool CanBeAHeapPointer(uptr p) {
// Since our heap is located in mmap-ed memory, we can assume a sensible lower
// boundary on heap addresses.
const uptr kMinAddress = 4 * 4096;
if (p < kMinAddress) return false;
#ifdef __x86_64__
// Accept only canonical form user-space addresses.
return ((p >> 47) == 0);
#else
return true;
#endif
}
// Scan the memory range, looking for byte patterns that point into allocator
// chunks. Mark those chunks with tag and add them to the frontier.
// There are two usage modes for this function: finding non-leaked chunks
// (tag = kReachable) and finding indirectly leaked chunks
// (tag = kIndirectlyLeaked). In the second case, there's no flood fill,
// so frontier = 0.
void ScanRangeForPointers(uptr begin, uptr end, InternalVector<uptr> *frontier,
const char *region_type, ChunkTag tag) {
const uptr alignment = flags()->pointer_alignment();
if (flags()->log_pointers)
Report("Scanning %s range %p-%p.\n", region_type, begin, end);
uptr pp = begin;
if (pp % alignment)
pp = pp + alignment - pp % alignment;
for (; pp + sizeof(uptr) <= end; pp += alignment) {
void *p = *reinterpret_cast<void**>(pp);
if (!CanBeAHeapPointer(reinterpret_cast<uptr>(p))) continue;
// FIXME: PointsIntoChunk is SLOW because GetBlockBegin() in
// LargeMmapAllocator involves a lock and a linear search.
void *chunk = PointsIntoChunk(p);
if (!chunk) continue;
LsanMetadata m(chunk);
if (m.tag() == kReachable) continue;
m.set_tag(tag);
if (flags()->log_pointers)
Report("%p: found %p pointing into chunk %p-%p of size %llu.\n", pp, p,
chunk, reinterpret_cast<uptr>(chunk) + m.requested_size(),
m.requested_size());
if (frontier)
frontier->push_back(reinterpret_cast<uptr>(chunk));
}
}
// Scan thread data (stacks and TLS) for heap pointers.
static void ProcessThreads(SuspendedThreadsList const &suspended_threads,
InternalVector<uptr> *frontier) {
InternalScopedBuffer<uptr> registers(SuspendedThreadsList::RegisterCount());
uptr registers_begin = reinterpret_cast<uptr>(registers.data());
uptr registers_end = registers_begin + registers.size();
for (uptr i = 0; i < suspended_threads.thread_count(); i++) {
uptr os_id = static_cast<uptr>(suspended_threads.GetThreadID(i));
if (flags()->log_threads) Report("Processing thread %d.\n", os_id);
uptr stack_begin, stack_end, tls_begin, tls_end, cache_begin, cache_end;
bool thread_found = GetThreadRangesLocked(os_id, &stack_begin, &stack_end,
&tls_begin, &tls_end,
&cache_begin, &cache_end);
if (!thread_found) {
// If a thread can't be found in the thread registry, it's probably in the
// process of destruction. Log this event and move on.
if (flags()->log_threads)
Report("Thread %d not found in registry.\n", os_id);
continue;
}
uptr sp;
bool have_registers =
(suspended_threads.GetRegistersAndSP(i, registers.data(), &sp) == 0);
if (!have_registers) {
Report("Unable to get registers from thread %d.\n");
// If unable to get SP, consider the entire stack to be reachable.
sp = stack_begin;
}
if (flags()->use_registers() && have_registers)
ScanRangeForPointers(registers_begin, registers_end, frontier,
"REGISTERS", kReachable);
if (flags()->use_stacks()) {
if (flags()->log_threads)
Report("Stack at %p-%p, SP = %p.\n", stack_begin, stack_end, sp);
if (sp < stack_begin || sp >= stack_end) {
// SP is outside the recorded stack range (e.g. the thread is running a
// signal handler on alternate stack). Again, consider the entire stack
// range to be reachable.
if (flags()->log_threads)
Report("WARNING: stack_pointer not in stack_range.\n");
} else {
// Shrink the stack range to ignore out-of-scope values.
stack_begin = sp;
}
ScanRangeForPointers(stack_begin, stack_end, frontier, "STACK",
kReachable);
}
if (flags()->use_tls()) {
if (flags()->log_threads) Report("TLS at %p-%p.\n", tls_begin, tls_end);
// Because LSan should not be loaded with dlopen(), we can assume
// that allocator cache will be part of static TLS image.
CHECK_LE(tls_begin, cache_begin);
CHECK_GE(tls_end, cache_end);
if (tls_begin < cache_begin)
ScanRangeForPointers(tls_begin, cache_begin, frontier, "TLS",
kReachable);
if (tls_end > cache_end)
ScanRangeForPointers(cache_end, tls_end, frontier, "TLS", kReachable);
}
}
}
static void FloodFillReachable(InternalVector<uptr> *frontier) {
while (frontier->size()) {
uptr next_chunk = frontier->back();
frontier->pop_back();
LsanMetadata m(reinterpret_cast<void *>(next_chunk));
ScanRangeForPointers(next_chunk, next_chunk + m.requested_size(), frontier,
"HEAP", kReachable);
}
}
// Mark leaked chunks which are reachable from other leaked chunks.
void MarkIndirectlyLeakedCb::operator()(void *p) const {
LsanMetadata m(p);
if (m.allocated() && m.tag() != kReachable) {
ScanRangeForPointers(reinterpret_cast<uptr>(p),
reinterpret_cast<uptr>(p) + m.requested_size(),
/* frontier */ 0, "HEAP", kIndirectlyLeaked);
}
}
// Set the appropriate tag on each chunk.
static void ClassifyAllChunks(SuspendedThreadsList const &suspended_threads) {
// Holds the flood fill frontier.
InternalVector<uptr> frontier(GetPageSizeCached());
if (flags()->use_globals())
ProcessGlobalRegions(&frontier);
ProcessThreads(suspended_threads, &frontier);
FloodFillReachable(&frontier);
ProcessPlatformSpecificAllocations(&frontier);
FloodFillReachable(&frontier);
// Now all reachable chunks are marked. Iterate over leaked chunks and mark
// those that are reachable from other leaked chunks.
if (flags()->log_pointers)
Report("Now scanning leaked blocks for pointers.\n");
ForEachChunk(MarkIndirectlyLeakedCb());
}
void ClearTagCb::operator()(void *p) const {
LsanMetadata m(p);
m.set_tag(kDirectlyLeaked);
}
static void PrintStackTraceById(u32 stack_trace_id) {
CHECK(stack_trace_id);
uptr size = 0;
const uptr *trace = StackDepotGet(stack_trace_id, &size);
StackTrace::PrintStack(trace, size, common_flags()->symbolize,
common_flags()->strip_path_prefix, 0);
}
static void LockAndSuspendThreads(StopTheWorldCallback callback, void *arg) {
LockThreadRegistry();
LockAllocator();
StopTheWorld(callback, arg);
// Allocator must be unlocked by the callback.
UnlockThreadRegistry();
}
///// Normal leak checking. /////
void CollectLeaksCb::operator()(void *p) const {
LsanMetadata m(p);
if (!m.allocated()) return;
if (m.tag() != kReachable) {
uptr resolution = flags()->resolution;
if (resolution > 0) {
uptr size = 0;
const uptr *trace = StackDepotGet(m.stack_trace_id(), &size);
size = Min(size, resolution);
leak_report_->Add(StackDepotPut(trace, size), m.requested_size(),
m.tag());
} else {
leak_report_->Add(m.stack_trace_id(), m.requested_size(), m.tag());
}
}
}
static void CollectLeaks(LeakReport *leak_report) {
ForEachChunk(CollectLeaksCb(leak_report));
}
void PrintLeakedCb::operator()(void *p) const {
LsanMetadata m(p);
if (!m.allocated()) return;
if (m.tag() != kReachable) {
CHECK(m.tag() == kDirectlyLeaked || m.tag() == kIndirectlyLeaked);
Printf("%s leaked %llu byte block at %p\n",
m.tag() == kDirectlyLeaked ? "Directly" : "Indirectly",
m.requested_size(), p);
}
}
static void PrintLeaked() {
Printf("\nReporting individual blocks:\n");
ForEachChunk(PrintLeakedCb());
}
static void DoLeakCheckCallback(const SuspendedThreadsList &suspended_threads,
void *arg) {
// Allocator must not be locked when we call GetRegionBegin().
UnlockAllocator();
bool *success = reinterpret_cast<bool *>(arg);
ClassifyAllChunks(suspended_threads);
LeakReport leak_report;
CollectLeaks(&leak_report);
if (!leak_report.IsEmpty()) {
leak_report.PrintLargest(flags()->max_leaks);
if (flags()->report_blocks)
PrintLeaked();
}
ForEachChunk(ClearTagCb());
*success = true;
}
void DoLeakCheck() {
bool success = false;
LockAndSuspendThreads(DoLeakCheckCallback, &success);
if (!success)
Report("Leak check failed!\n");
}
///// Reporting of leaked blocks' addresses (for testing). /////
void ReportLeakedCb::operator()(void *p) const {
LsanMetadata m(p);
if (m.allocated() && m.tag() != kReachable)
leaked_->push_back(p);
}
struct ReportLeakedParam {
InternalVector<void *> *leaked;
uptr sources;
bool success;
};
static void ReportLeakedCallback(const SuspendedThreadsList &suspended_threads,
void *arg) {
// Allocator must not be locked when we call GetRegionBegin().
UnlockAllocator();
ReportLeakedParam *param = reinterpret_cast<ReportLeakedParam *>(arg);
flags()->sources = param->sources;
ClassifyAllChunks(suspended_threads);
ForEachChunk(ReportLeakedCb(param->leaked));
ForEachChunk(ClearTagCb());
param->success = true;
}
void ReportLeaked(InternalVector<void *> *leaked, uptr sources) {
CHECK_EQ(0, leaked->size());
ReportLeakedParam param;
param.leaked = leaked;
param.success = false;
param.sources = sources;
LockAndSuspendThreads(ReportLeakedCallback, &param);
CHECK(param.success);
}
///// LeakReport implementation. /////
// A hard limit on the number of distinct leaks, to avoid quadratic complexity
// in LeakReport::Add(). We don't expect to ever see this many leaks in
// real-world applications.
// FIXME: Get rid of this limit by changing the implementation of LeakReport to
// use a hash table.
const uptr kMaxLeaksConsidered = 1000;
void LeakReport::Add(u32 stack_trace_id, uptr leaked_size, ChunkTag tag) {
CHECK(tag == kDirectlyLeaked || tag == kIndirectlyLeaked);
bool is_directly_leaked = (tag == kDirectlyLeaked);
for (uptr i = 0; i < leaks_.size(); i++)
if (leaks_[i].stack_trace_id == stack_trace_id &&
leaks_[i].is_directly_leaked == is_directly_leaked) {
leaks_[i].hit_count++;
leaks_[i].total_size += leaked_size;
return;
}
if (leaks_.size() == kMaxLeaksConsidered) return;
Leak leak = { /* hit_count */ 1, leaked_size, stack_trace_id,
is_directly_leaked };
leaks_.push_back(leak);
}
static bool IsLarger(const Leak &leak1, const Leak &leak2) {
return leak1.total_size > leak2.total_size;
}
void LeakReport::PrintLargest(uptr max_leaks) {
CHECK(leaks_.size() <= kMaxLeaksConsidered);
Printf("\n");
if (leaks_.size() == kMaxLeaksConsidered)
Printf("Too many leaks! Only the first %llu leaks encountered will be "
"reported.\n",
kMaxLeaksConsidered);
if (max_leaks > 0 && max_leaks < leaks_.size())
Printf("The %llu largest leak%s:\n", max_leaks, max_leaks > 1 ? "s" : "");
InternalSort(&leaks_, leaks_.size(), IsLarger);
max_leaks = max_leaks > 0 ? Min(max_leaks, leaks_.size()) : leaks_.size();
for (uptr i = 0; i < max_leaks; i++) {
Printf("\n%s leak of %llu bytes in %llu objects allocated from:\n",
leaks_[i].is_directly_leaked ? "Direct" : "Indirect",
leaks_[i].total_size, leaks_[i].hit_count);
PrintStackTraceById(leaks_[i].stack_trace_id);
}
if (max_leaks < leaks_.size()) {
uptr remaining = leaks_.size() - max_leaks;
Printf("\nOmitting %llu more leak%s.\n", remaining,
remaining > 1 ? "s" : "");
}
}
} // namespace __lsan

View File

@ -0,0 +1,190 @@
//=-- lsan_common.h -------------------------------------------------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file is a part of LeakSanitizer.
// Private LSan header.
//
//===----------------------------------------------------------------------===//
#ifndef LSAN_COMMON_H
#define LSAN_COMMON_H
#include "sanitizer_common/sanitizer_common.h"
#include "sanitizer_common/sanitizer_internal_defs.h"
#include "sanitizer_common/sanitizer_symbolizer.h"
namespace __lsan {
// Chunk tags.
enum ChunkTag {
kDirectlyLeaked = 0, // default
kIndirectlyLeaked = 1,
kReachable = 2
};
// Sources of pointers.
// Global variables (.data and .bss).
const uptr kSourceGlobals = 1 << 0;
// Thread stacks.
const uptr kSourceStacks = 1 << 1;
// TLS and thread-specific storage.
const uptr kSourceTLS = 1 << 2;
// Thread registers.
const uptr kSourceRegisters = 1 << 3;
// Unaligned pointers.
const uptr kSourceUnaligned = 1 << 4;
// Aligned pointers everywhere.
const uptr kSourceAllAligned =
kSourceGlobals | kSourceStacks | kSourceTLS | kSourceRegisters;
struct Flags {
bool use_registers() const { return sources & kSourceRegisters; }
bool use_globals() const { return sources & kSourceGlobals; }
bool use_stacks() const { return sources & kSourceStacks; }
bool use_tls() const { return sources & kSourceTLS; }
uptr pointer_alignment() const {
return (sources & kSourceUnaligned) ? 1 : sizeof(uptr);
}
uptr sources;
// Print addresses of leaked blocks after main leak report.
bool report_blocks;
// Aggregate two blocks into one leak if this many stack frames match. If
// zero, the entire stack trace must match.
int resolution;
// The number of leaks reported.
int max_leaks;
// Debug logging.
bool log_pointers;
bool log_threads;
};
extern Flags lsan_flags;
inline Flags *flags() { return &lsan_flags; }
void InitCommonLsan();
// Testing interface. Find leaked chunks and dump their addresses to vector.
void ReportLeaked(InternalVector<void *> *leaked, uptr sources);
// Normal leak check. Find leaks and print a report according to flags.
void DoLeakCheck();
struct Leak {
uptr hit_count;
uptr total_size;
u32 stack_trace_id;
bool is_directly_leaked;
};
// Aggregates leaks by stack trace prefix.
class LeakReport {
public:
LeakReport() : leaks_(1) {}
void Add(u32 stack_trace_id, uptr leaked_size, ChunkTag tag);
void PrintLargest(uptr max_leaks);
bool IsEmpty() { return leaks_.size() == 0; }
private:
InternalVector<Leak> leaks_;
};
// Platform-specific functions.
void InitializePlatformSpecificModules();
void ProcessGlobalRegions(InternalVector<uptr> *frontier);
void ProcessPlatformSpecificAllocations(InternalVector<uptr> *frontier);
void ScanRangeForPointers(uptr begin, uptr end, InternalVector<uptr> *frontier,
const char *region_type, ChunkTag tag);
// Callables for iterating over chunks. Those classes are used as template
// parameters in ForEachChunk, so we must expose them here to allow for explicit
// template instantiation.
// Identifies unreachable chunks which must be treated as reachable. Marks them
// as reachable and adds them to the frontier.
class ProcessPlatformSpecificAllocationsCb {
public:
explicit ProcessPlatformSpecificAllocationsCb(InternalVector<uptr> *frontier)
: frontier_(frontier) {}
void operator()(void *p) const;
private:
InternalVector<uptr> *frontier_;
};
// Prints addresses of unreachable chunks.
class PrintLeakedCb {
public:
void operator()(void *p) const;
};
// Aggregates unreachable chunks into a LeakReport.
class CollectLeaksCb {
public:
explicit CollectLeaksCb(LeakReport *leak_report)
: leak_report_(leak_report) {}
void operator()(void *p) const;
private:
LeakReport *leak_report_;
};
// Dumps addresses of unreachable chunks to a vector (for testing).
class ReportLeakedCb {
public:
explicit ReportLeakedCb(InternalVector<void *> *leaked) : leaked_(leaked) {}
void operator()(void *p) const;
private:
InternalVector<void *> *leaked_;
};
// Resets each chunk's tag to default (kDirectlyLeaked).
class ClearTagCb {
public:
void operator()(void *p) const;
};
// Scans each leaked chunk for pointers to other leaked chunks, and marks each
// of them as indirectly leaked.
class MarkIndirectlyLeakedCb {
public:
void operator()(void *p) const;
};
// The following must be implemented in the parent tool.
template<typename Callable> void ForEachChunk(Callable const &callback);
// The address range occupied by the global allocator object.
void GetAllocatorGlobalRange(uptr *begin, uptr *end);
// Wrappers for allocator's ForceLock()/ForceUnlock().
void LockAllocator();
void UnlockAllocator();
// Wrappers for ThreadRegistry access.
void LockThreadRegistry();
void UnlockThreadRegistry();
bool GetThreadRangesLocked(uptr os_id, uptr *stack_begin, uptr *stack_end,
uptr *tls_begin, uptr *tls_end,
uptr *cache_begin, uptr *cache_end);
// If p points into a chunk that has been allocated to the user, return its
// address. Otherwise, return 0.
void *PointsIntoChunk(void *p);
// Wrapper for chunk metadata operations.
class LsanMetadata {
public:
explicit LsanMetadata(void *chunk);
bool allocated() const;
ChunkTag tag() const;
void set_tag(ChunkTag value);
uptr requested_size() const;
u32 stack_trace_id() const;
private:
void *metadata_;
};
} // namespace __lsan
#endif // LSAN_COMMON_H

View File

@ -0,0 +1,122 @@
//=-- lsan_common_linux.cc ------------------------------------------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file is a part of LeakSanitizer.
// Implementation of common leak checking functionality. Linux-specific code.
//
//===----------------------------------------------------------------------===//
#include "sanitizer_common/sanitizer_platform.h"
#if SANITIZER_LINUX
#include "lsan_common.h"
#include <link.h>
#include "sanitizer_common/sanitizer_common.h"
#include "sanitizer_common/sanitizer_linux.h"
#include "sanitizer_common/sanitizer_stackdepot.h"
namespace __lsan {
static const char kLinkerName[] = "ld";
// We request 2 modules matching "ld", so we can print a warning if there's more
// than one match. But only the first one is actually used.
static char linker_placeholder[2 * sizeof(LoadedModule)] ALIGNED(64);
static LoadedModule *linker = 0;
static bool IsLinker(const char* full_name) {
return LibraryNameIs(full_name, kLinkerName);
}
void InitializePlatformSpecificModules() {
internal_memset(linker_placeholder, 0, sizeof(linker_placeholder));
uptr num_matches = GetListOfModules(
reinterpret_cast<LoadedModule *>(linker_placeholder), 2, IsLinker);
if (num_matches == 1) {
linker = reinterpret_cast<LoadedModule *>(linker_placeholder);
return;
}
if (num_matches == 0)
Report("%s: Dynamic linker not found. TLS will not be handled correctly.\n",
SanitizerToolName);
else if (num_matches > 1)
Report("%s: Multiple modules match \"%s\". TLS will not be handled "
"correctly.\n", SanitizerToolName, kLinkerName);
linker = 0;
}
static int ProcessGlobalRegionsCallback(struct dl_phdr_info *info, size_t size,
void *data) {
InternalVector<uptr> *frontier =
reinterpret_cast<InternalVector<uptr> *>(data);
for (uptr j = 0; j < info->dlpi_phnum; j++) {
const ElfW(Phdr) *phdr = &(info->dlpi_phdr[j]);
// We're looking for .data and .bss sections, which reside in writeable,
// loadable segments.
if (!(phdr->p_flags & PF_W) || (phdr->p_type != PT_LOAD) ||
(phdr->p_memsz == 0))
continue;
uptr begin = info->dlpi_addr + phdr->p_vaddr;
uptr end = begin + phdr->p_memsz;
uptr allocator_begin = 0, allocator_end = 0;
GetAllocatorGlobalRange(&allocator_begin, &allocator_end);
if (begin <= allocator_begin && allocator_begin < end) {
CHECK_LE(allocator_begin, allocator_end);
CHECK_LT(allocator_end, end);
if (begin < allocator_begin)
ScanRangeForPointers(begin, allocator_begin, frontier, "GLOBAL",
kReachable);
if (allocator_end < end)
ScanRangeForPointers(allocator_end, end, frontier, "GLOBAL",
kReachable);
} else {
ScanRangeForPointers(begin, end, frontier, "GLOBAL", kReachable);
}
}
return 0;
}
// Scan global variables for heap pointers.
void ProcessGlobalRegions(InternalVector<uptr> *frontier) {
// FIXME: dl_iterate_phdr acquires a linker lock, so we run a risk of
// deadlocking by running this under StopTheWorld. However, the lock is
// reentrant, so we should be able to fix this by acquiring the lock before
// suspending threads.
dl_iterate_phdr(ProcessGlobalRegionsCallback, frontier);
}
static uptr GetCallerPC(u32 stack_id) {
CHECK(stack_id);
uptr size = 0;
const uptr *trace = StackDepotGet(stack_id, &size);
// The top frame is our malloc/calloc/etc. The next frame is the caller.
CHECK_GE(size, 2);
return trace[1];
}
void ProcessPlatformSpecificAllocationsCb::operator()(void *p) const {
LsanMetadata m(p);
if (m.allocated() && m.tag() != kReachable) {
if (linker->containsAddress(GetCallerPC(m.stack_trace_id()))) {
m.set_tag(kReachable);
frontier_->push_back(reinterpret_cast<uptr>(p));
}
}
}
// Handle dynamically allocated TLS blocks by treating all chunks allocated from
// ld-linux.so as reachable.
void ProcessPlatformSpecificAllocations(InternalVector<uptr> *frontier) {
if (!flags()->use_tls()) return;
if (!linker) return;
ForEachChunk(ProcessPlatformSpecificAllocationsCb(frontier));
}
} // namespace __lsan
#endif // SANITIZER_LINUX