[dfsan] Add a DFSan allocator

[stephan.yichao.zhao]

Summary

This is a part of https://reviews.llvm.org/D101204

[stephan.yichao.zhao]

added missing files

[stephan.yichao.zhao]

fixed test/build failures

[stephan.yichao.zhao]

fixed failed test cases

[stephan.yichao.zhao]

fixed failed test cases

[morehouse]

compiler-rt/lib/dfsan/dfsan.cpp

24

We still need dfsan_flags.h.

543

The start + size implementation seems cleaner to me. Can we use that for WriteShadowWithSize and make WriteShadowInRange call that instead?

626

This might be redundant since on Linux releasing the memory zeroes it out on the next load.

637

Should we just call ReleaseOrigins?

670–671

What's the reason to change this from ReleaseMemoryPagesToOS?

709

Can we simplify by moving all this logic into WriteShadowWithSize? It's confusing to have multiple layers of functions to set the shadow, with some repeated checks in each layer.

711–712

Does this do anything since we already called MmapFixedSuperNoReserve?

965

What's going on here? intercept_tld_get_addr appears unused, so why do we need to override it?

1080

dfsan_init is now unused. Why do we need this?

compiler-rt/lib/dfsan/dfsan.h

24

This is unused.

70

Why int instead of bool?

71

Since we currently only call dfsan_init from preinit_array, which happens once in a single thread, do we actually need to guard against multiple initialization?

122

This looks unused.

compiler-rt/lib/dfsan/dfsan_allocator.cpp

32

Maybe we don't need this. We release (zero) shadow on unmap, so it should still be zero if we mmap it again.

43

1. This appears unused except for kSpaceBeg below.
2. The name sounds like the max size of the heap, rather than the start address of it.

[stephan.yichao.zhao]

compiler-rt/lib/dfsan/dfsan.cpp

24

dfsan_flags.h is included in dfsan.h

543

If we use "WriteShadowWithSize(dfsan_label label, uptr beg_shadow_addr, uptr size)" with the base function,
here the size is preferred to be the number of labels. If not, we may have to assert size % sizeof(dfsan_label) == 0.

But if size is # fo labels, we would have

WriteShadowInRange(dfsan_label label, uptr beg_shadow_addr,
                               uptr end_shadow_addr)  {
  assert(beg_shadow_addr <= end_shadow_addr);
  assert((end_shadow_addr - beg_shadow_addr) % 2 == 0);
  WriteShadowWithSize(label, (end_shadow_addr - beg_shadow_addr) / 2);
}

I feel after the fast16label -> fast8label change, it is possible to do this because then #bytes == #labels.

626

This mainly follows how MSan does when allocator does unmmap:

https://github.com/llvm/llvm-project/blob/main/compiler-rt/lib/msan/msan_allocator.cpp#L33

dfsan_set_label only does mmap, at unmmap event, we call ReleaseMemoryPagesToOS explicitly.

637

ReleaseOrigins does not call ReleaseMemoryPagesToOS.

670–671

This mainly follows MSan's approach.

Although MSan does not have a similar function to release origin, it has SetShadow

https://github.com/llvm/llvm-project/blob/main/compiler-rt/lib/msan/msan_poisoning.cpp#L216

to zero out Shadow, it calls MmapFixedSuperNoReserve w/o ReleaseMemoryPagesToOS,

To be consistent with this, ReleaseOrigins replaces ReleaseMemoryPagesToOS by MmapFixedSuperNoReserve.

709

WriteShadowInRange works for both 0 and non-zero labels.
Refactored this code into ReleaseOrClearShadows.

711–712

This should have been removed to be consistent with
https://github.com/llvm/llvm-project/blob/main/compiler-rt/lib/msan/msan_poisoning.cpp#L216

965

This will be used by D101204 when interceptors are defined.

1080

This will be used by D101204.

compiler-rt/lib/dfsan/dfsan.h

24

This would be used by those new/delete injections dfsan_new_delete.cpp in D101204.

71

I found dfsan_init_is_running is mainly to ensure interceptors use real calls w/o wrappers before dfsan_init is done.
See dfsan_interceptors.cpp from D101204.
Otherwise mmap called by dfsan_init can be intercepted too.

122

This will be called by dfsan_interceptors.cpp in D101204.
It makes sure that when interceptors are called, dfsan_init (shadow/origin allocation) must be done because wrappers code run.

I also feel like dfsan_init/dfsan_is_running/dfsan_inited are not straightforward. But from MSan's code blame, they seem useful to some corner cases.
For example, calloc, realloc and malloc can be called by DL before preinit_array runs, etc.

compiler-rt/lib/dfsan/dfsan_allocator.cpp

32

Thank you. Removed.

43

inlined the constant.

[stephan.yichao.zhao]

update

[morehouse]

compiler-rt/lib/dfsan/dfsan.cpp

24

Yes. It's just a nit, but I usually try to explicitly include the things used in each file, so that we can change header files without touching all the places where they're included.

546–553

637

Shouldn't mmap(MAP_NORESERVE) have a similar effect on resident memory to madvise(MADV_DONTNEED)?

Besides, we already have ReleaseOrClearShadow and ReleaseOrigins. Why do we need to reimplement that functionality in this function?

670–671

Don't they have a similar effect? It's confusing to use two different ways of releasing memory without at least a comment explaining why.

compiler-rt/lib/dfsan/dfsan.h

24

In that case, maybe we should define it in the file it will be used in, instead of here.

[stephan.yichao.zhao]

compiler-rt/lib/dfsan/dfsan.cpp

24

moved #include "dfsan/dfsan_flags.h" to each cc file.

546–553

Thank you.

637

removed madvise(MADV_DONTNEED), and replaced dfsan_release_meta_memory by dfsan_set_label(0).

670–671

removed those madvise calls.

compiler-rt/lib/dfsan/dfsan.h

24

moved to dfsan_new_delete.cpp

[stephan.yichao.zhao]

update

[stephan.yichao.zhao]

update

[stephan.yichao.zhao]

update

[morehouse]

compiler-rt/lib/dfsan/dfsan_allocator.cpp

32

Looks like you added it back. I think we only need to set 0 label on unmap.

[stephan.yichao.zhao]

compiler-rt/lib/dfsan/dfsan_allocator.cpp

32

I found those mmap interceptors in dfsan_intercerptors.cp also do dfsan_set_label(0).

https://github.com/llvm/llvm-project/blob/main/compiler-rt/lib/dfsan/dfsan_interceptors.cpp#L41

For the same reason they are not needed either.
In the next change, I will remove all mmap related dfsan_set_label(0) together.

[stephan.yichao.zhao]

update

[stephan.yichao.zhao]

compiler-rt/lib/dfsan/dfsan_allocator.cpp

32

Thought about the problem a bit more.

I think our discussion is based on an assumption that all unmmap will be precisely intercepted.
This may not be true since if some code calls unmmap by system call number directly like what this sanitizer allocator does, there is no way to intercept that kind of unmmap.
So in general, we may not be able to remove intercepting those mmap and mmap64.

But for this OnMap, it should be safe because its corresponding OnUmmap is injected by our code.

So in the following change, I will only remove dfsan_set_label(0) for OnMap.

[morehouse]

compiler-rt/lib/dfsan/dfsan_allocator.cpp

32

Thought about the problem a bit more.

I think our discussion is based on an assumption that all unmmap will be precisely intercepted.
This may not be true since if some code calls unmmap by system call number directly like what this sanitizer allocator does, there is no way to intercept that kind of unmmap.
So in general, we may not be able to remove intercepting those mmap and mmap64.

Ah yes, good point.

But for this OnMap, it should be safe because its corresponding OnUmmap is injected by our code.

Well, I'm not 100% sure it's safe. Something else could do syscall(SYS_mmap), then taint the memory, then syscall_(SYS_munmap). Then we could end up mapping that same region for our heap and calling this OnMap callback.

So in the following change, I will only remove dfsan_set_label(0) for OnMap.

compiler-rt/lib/dfsan/CMakeLists.txt

 include_directories(..)
 
 # Runtime library sources and build flags.
 set(DFSAN_RTL_SOURCES
   dfsan.cpp
+  dfsan_allocator.cpp
   dfsan_chained_origin_depot.cpp
   dfsan_custom.cpp
   dfsan_interceptors.cpp
   dfsan_thread.cpp
   )
 
 set(DFSAN_RTL_HEADERS
   dfsan.h
+  dfsan_allocator.h
   dfsan_chained_origin_depot.h
   dfsan_flags.inc
   dfsan_flags.h
   dfsan_platform.h
   dfsan_thread.h
   )
 
 set(DFSAN_COMMON_CFLAGS ${SANITIZER_COMMON_CFLAGS})

compiler-rt/lib/dfsan/dfsan.h

 #ifndef DFSAN_H
 #define DFSAN_H
 
 #include "sanitizer_common/sanitizer_internal_defs.h"
 
 #include "dfsan_flags.h"
 #include "dfsan_platform.h"
 
+#ifndef DFSAN_REPLACE_OPERATORS_NEW_AND_DELETE
+#define DFSAN_REPLACE_OPERATORS_NEW_AND_DELETE 1
+#endif
+
 using __sanitizer::u16;
 using __sanitizer::u32;
 using __sanitizer::uptr;
 
 // Copy declarations from public sanitizer/dfsan_interface.h header here.
 typedef u16 dfsan_label;
 typedef u32 dfsan_origin;
 

 
 template <typename T>
 void dfsan_set_label(dfsan_label label, T &data) {  // NOLINT
   dfsan_set_label(label, (void *)&data, sizeof(T));
 }
 
 namespace __dfsan {
 
-void InitializeInterceptors();
+extern int dfsan_inited;
+extern bool dfsan_init_is_running;
+
+void initialize_interceptors();
 
 inline dfsan_label *shadow_for(void *ptr) {
   return (dfsan_label *) ((((uptr) ptr) & ShadowMask()) << 1);
 }
 
 inline const dfsan_label *shadow_for(const void *ptr) {
   return shadow_for(const_cast<void *>(ptr));
 }
inline bool is_shadow_addr_valid(uptr shadow_addr) {
   return (uptr)shadow_addr >= ShadowAddr() && (uptr)shadow_addr < OriginAddr();
 }
 
 inline bool has_valid_shadow_addr(const void *ptr) {
   const dfsan_label *ptr_s = shadow_for(ptr);
   return is_shadow_addr_valid((uptr)ptr_s);
 }
 
+void dfsan_copy_memory(void *dst, const void *src, uptr size);
+void dfsan_release_meta_memory(const void *addr, uptr size);
+
+void dfsan_allocator_init();
+void dfsan_deallocate(void *ptr);
+
+void *dfsan_malloc(uptr size);
+void *dfsan_calloc(uptr nmemb, uptr size);
+void *dfsan_realloc(void *ptr, uptr size);
+void *dfsan_reallocarray(void *ptr, uptr nmemb, uptr size);
+void *dfsan_valloc(uptr size);
+void *dfsan_pvalloc(uptr size);
+void *dfsan_aligned_alloc(uptr alignment, uptr size);
+void *dfsan_memalign(uptr alignment, uptr size);
+int dfsan_posix_memalign(void **memptr, uptr alignment, uptr size);
+
+void dfsan_init();
+
 }  // namespace __dfsan
 
 #endif  // DFSAN_H

compiler-rt/lib/dfsan/dfsan.cpp

 // The public interface is defined in include/sanitizer/dfsan_interface.h whose
 // functions are prefixed dfsan_ while the compiler interface functions are
 // prefixed __dfsan_.
 //===----------------------------------------------------------------------===//
 
 #include "dfsan/dfsan.h"
 
 #include "dfsan/dfsan_chained_origin_depot.h"
-#include "dfsan/dfsan_flags.h"
 #include "dfsan/dfsan_origin.h"
 #include "dfsan/dfsan_thread.h"
 #include "sanitizer_common/sanitizer_atomic.h"
 #include "sanitizer_common/sanitizer_common.h"
 #include "sanitizer_common/sanitizer_file.h"
 #include "sanitizer_common/sanitizer_flag_parser.h"
 #include "sanitizer_common/sanitizer_flags.h"
 #include "sanitizer_common/sanitizer_internal_defs.h"
if (*(u64 *)addr == origin64)
       continue;
     *(u64 *)addr = origin64;
   }
   if (end & 7ULL)
     if (*(u32 *)(end - kOriginAlign) != origin)
       *(u32 *)(end - kOriginAlign) = origin;
 }
 
-static void WriteShadowIfDifferent(dfsan_label label, uptr shadow_addr,
-                                   uptr size) {
-  dfsan_label *labelp = (dfsan_label *)shadow_addr;
-  for (; size != 0; --size, ++labelp) {
+static void WriteShadowInRange(dfsan_label label, uptr beg_shadow_addr,
+                               uptr end_shadow_addr) {
+  // TODO: After changing dfsan_label to 8bit, use internal_memset when label
+  // is not 0.
+  if (label) {
+    dfsan_label *labelp = (dfsan_label *)beg_shadow_addr;
+    for (; (uptr)labelp < end_shadow_addr; ++labelp) *labelp = label;
+    return;
+  }
+
+  dfsan_label *labelp = (dfsan_label *)beg_shadow_addr;
+  for (; (uptr)labelp < end_shadow_addr; ++labelp) {
     // Don't write the label if it is already the value we need it to be.
     // In a program where most addresses are not labeled, it is common that
     // a page of shadow memory is entirely zeroed.  The Linux copy-on-write
     // implementation will share all of the zeroed pages, making a copy of a
     // page when any value is written.  The un-sharing will happen even if
     // the value written does not change the value in memory.  Avoiding the
     // write when both |label| and |*labelp| are zero dramatically reduces
     // the amount of real memory used by large programs.
-    if (label == *labelp)
+    if (!*labelp)
       continue;
 
-    *labelp = label;
+    *labelp = 0;
   }
 }
 
+static void WriteShadowWithSize(dfsan_label label, uptr shadow_addr,
+                                uptr size) {
+  WriteShadowInRange(label, shadow_addr, shadow_addr + size * sizeof(label));
+}
+
 #define RET_CHAIN_ORIGIN(id)           \
   GET_CALLER_PC_BP_SP;                 \
   (void)sp;                            \
   GET_STORE_STACK_TRACE_PC_BP(pc, bp); \
   return ChainOrigin(id, &stack);
 
 // Return a new origin chain with the previous ID id and the current stack
 // trace.

 }
 
 SANITIZER_INTERFACE_ATTRIBUTE void dfsan_mem_origin_transfer(const void *dst,
                                                              const void *src,
                                                              uptr len) {
   __dfsan_mem_origin_transfer(dst, src, len);
 }
 
+namespace __dfsan {
+
+int dfsan_inited = 0;
+bool dfsan_init_is_running = false;
+
+void dfsan_copy_memory(void *dst, const void *src, uptr size) {
+  internal_memcpy(dst, src, size);
+  internal_memcpy((void *)shadow_for(dst), (const void *)shadow_for(src),
+                  size * sizeof(dfsan_label));
+  if (__dfsan_get_track_origins())
+    dfsan_mem_origin_transfer(dst, src, size);
+}
+
+void dfsan_release_meta_memory(const void *addr, uptr size) {
+  dfsan_set_label(0, (void *)addr, size);
+
+  // We are about to unmap a chunk of user memory.
+  // Mark the corresponding shadow memory as not needed.
+  const uptr beg_shadow_addr = (uptr)__dfsan::shadow_for(addr);
+  const void *end_addr = (void *)((uptr)addr + size);
+  const uptr end_shadow_addr = (uptr)__dfsan::shadow_for(end_addr);
+  ReleaseMemoryPagesToOS(beg_shadow_addr, end_shadow_addr);
+  if (__dfsan_get_track_origins()) {
+    const uptr beg_origin_addr = (uptr)__dfsan::origin_for(addr);
+    const uptr end_origin_addr = (uptr)__dfsan::origin_for(end_addr);
+    ReleaseMemoryPagesToOS(beg_origin_addr, end_origin_addr);
+  }
+}
+
+}  // namespace __dfsan
+
 // If the label s is tainted, set the size bytes from the address p to be a new
 // origin chain with the previous ID o and the current stack trace. This is
 // used by instrumentation to reduce code size when too much code is inserted.
 extern "C" SANITIZER_INTERFACE_ATTRIBUTE void __dfsan_maybe_store_origin(
     u16 s, void *p, uptr size, dfsan_origin o) {
   if (UNLIKELY(s)) {
     GET_CALLER_PC_BP_SP;
     (void)sp;
     GET_STORE_STACK_TRACE_PC_BP(pc, bp);
     SetOrigin(p, size, ChainOrigin(o, &stack));
   }
 }
 
-// Releases the pages within the origin address range, and sets the origin
-// addresses not on the pages to be 0.
-static void ReleaseOrClearOrigins(void *addr, uptr size) {
+// Releases the pages within the origin address range.
+static void ReleaseOrigins(void *addr, uptr size) {
   const uptr beg_origin_addr = (uptr)__dfsan::origin_for(addr);
   const void *end_addr = (void *)((uptr)addr + size);
   const uptr end_origin_addr = (uptr)__dfsan::origin_for(end_addr);
+
+  if (end_origin_addr - beg_origin_addr <
+      common_flags()->clear_shadow_mmap_threshold)
+    return;
+
   const uptr page_size = GetPageSizeCached();
   const uptr beg_aligned = RoundUpTo(beg_origin_addr, page_size);
   const uptr end_aligned = RoundDownTo(end_origin_addr, page_size);
 
-  // dfsan_set_label can be called from the following cases
-  // 1) mapped ranges by new/delete and malloc/free. This case has origin memory
-  // size > 50k, and happens less frequently.
-  // 2) zero-filling internal data structures by utility libraries. This case
-  // has origin memory size < 16k, and happens more often.
-  // Set kNumPagesThreshold to be 4 to avoid releasing small pages.
-  const int kNumPagesThreshold = 4;
-  if (beg_aligned + kNumPagesThreshold * page_size >= end_aligned)
-    return;
-
-  ReleaseMemoryPagesToOS(beg_aligned, end_aligned);
+  if (!MmapFixedSuperNoReserve(beg_aligned, end_aligned - beg_aligned))
+    Die();
 }
 
 void SetShadow(dfsan_label label, void *addr, uptr size, dfsan_origin origin) {
   const uptr beg_shadow_addr = (uptr)__dfsan::shadow_for(addr);
 
   if (0 != label) {
-    WriteShadowIfDifferent(label, beg_shadow_addr, size);
+    WriteShadowWithSize(label, beg_shadow_addr, size);
     if (__dfsan_get_track_origins())
       SetOrigin(addr, size, origin);
     return;
   }
 
   if (__dfsan_get_track_origins())
-    ReleaseOrClearOrigins(addr, size);
+    ReleaseOrigins(addr, size);
 
   // If label is 0, releases the pages within the shadow address range, and sets
   // the shadow addresses not on the pages to be 0.
   const void *end_addr = (void *)((uptr)addr + size);
   const uptr end_shadow_addr = (uptr)__dfsan::shadow_for(end_addr);
+
+  if (end_shadow_addr - beg_shadow_addr <
+      common_flags()->clear_shadow_mmap_threshold)
+    return WriteShadowWithSize(label, beg_shadow_addr, size);
+
   const uptr page_size = GetPageSizeCached();
   const uptr beg_aligned = RoundUpTo(beg_shadow_addr, page_size);
   const uptr end_aligned = RoundDownTo(end_shadow_addr, page_size);
 
-  // dfsan_set_label can be called from the following cases
-  // 1) mapped ranges by new/delete and malloc/free. This case has shadow memory
-  // size > 100k, and happens less frequently.
-  // 2) zero-filling internal data structures by utility libraries. This case
-  // has shadow memory size < 32k, and happens more often.
-  // Set kNumPagesThreshold to be 8 to avoid releasing small pages.
-  const int kNumPagesThreshold = 8;
-  if (beg_aligned + kNumPagesThreshold * page_size >= end_aligned)
-    return WriteShadowIfDifferent(label, beg_shadow_addr, size);
+  if (beg_aligned >= end_aligned) {
+    WriteShadowWithSize(0, beg_shadow_addr, size);
+  } else {
+    if (beg_aligned != beg_shadow_addr)
+      WriteShadowInRange(0, beg_shadow_addr, beg_aligned);
+    if (end_aligned != end_shadow_addr)
+      WriteShadowInRange(0, end_aligned, end_shadow_addr);
+    if (!MmapFixedSuperNoReserve(beg_aligned, end_aligned - beg_aligned))
+      Die();
+  }
 
-  WriteShadowIfDifferent(label, beg_shadow_addr, beg_aligned - beg_shadow_addr);
   ReleaseMemoryPagesToOS(beg_aligned, end_aligned);
-  WriteShadowIfDifferent(label, end_aligned, end_shadow_addr - end_aligned);
 }
 
 extern "C" SANITIZER_INTERFACE_ATTRIBUTE void __dfsan_set_label(
     dfsan_label label, dfsan_origin origin, void *addr, uptr size) {
   SetShadow(label, addr, size, origin);
 }
 
 SANITIZER_INTERFACE_ATTRIBUTE
 void dfsan_set_label(dfsan_label label, void *addr, uptr size) {

 #define DFSAN_FLAG(Type, Name, DefaultValue, Description) \
   RegisterFlag(parser, #Name, Description, &f->Name);
 #include "dfsan_flags.inc"
 #undef DFSAN_FLAG
 }
 
 static void InitializeFlags() {
   SetCommonFlagsDefaults();
+  {
+    CommonFlags cf;
+    cf.CopyFrom(*common_flags());
+    cf.intercept_tls_get_addr = true;
+    OverrideCommonFlags(cf);
+  }
   flags().SetDefaults();
 
   FlagParser parser;
   RegisterCommonFlags(&parser);
   RegisterDfsanFlags(&parser, &flags());
   parser.ParseStringFromEnv("DFSAN_OPTIONS");
   InitializeCommonFlags();
   if (Verbosity()) ReportUnrecognizedFlags();
static void dfsan_fini() {
   }
 }
 
 extern "C" void dfsan_flush() {
   if (!MmapFixedSuperNoReserve(ShadowAddr(), UnusedAddr() - ShadowAddr()))
     Die();
 }
 
-static void dfsan_init(int argc, char **argv, char **envp) {
+static void DFsanInit(int argc, char **argv, char **envp) {
+  CHECK(!dfsan_init_is_running);
+  if (dfsan_inited)
+    return;
+  dfsan_init_is_running = true;
+  SanitizerToolName = "DataflowSanitizer";
+
   InitializeFlags();
 
   ::InitializePlatformEarly();
 
   dfsan_flush();
   if (common_flags()->use_madv_dontdump)
     DontDumpShadowMemory(ShadowAddr(), UnusedAddr() - ShadowAddr());
 
   // Protect the region of memory we don't use, to preserve the one-to-one
   // mapping from application to shadow memory. But if ASLR is disabled, Linux
   // will load our executable in the middle of our unused region. This mostly
   // works so long as the program doesn't use too much memory. We support this
   // case by disabling memory protection when ASLR is disabled.
-  uptr init_addr = (uptr)&dfsan_init;
+  uptr init_addr = (uptr)&DFsanInit;
   if (!(init_addr >= UnusedAddr() && init_addr < AppAddr()))
     MmapFixedNoAccess(UnusedAddr(), AppAddr() - UnusedAddr());
 
-  InitializeInterceptors();
+  initialize_interceptors();
 
   // Register the fini callback to run when the program terminates successfully
   // or it is killed by the runtime.
   Atexit(dfsan_fini);
   AddDieCallback(dfsan_fini);
 
   // Set up threads
   DFsanTSDInit(DFsanTSDDtor);
+
+  dfsan_allocator_init();
+
   DFsanThread *main_thread = DFsanThread::Create(nullptr, nullptr, nullptr);
   SetCurrentThread(main_thread);
   main_thread->ThreadStart();
 
   __dfsan_label_info[kInitializingLabel].desc = "<init label>";
+
+  dfsan_init_is_running = true;
+  dfsan_inited = 1;
 }
 
+namespace __dfsan {
+
+void dfsan_init() { DFsanInit(0, nullptr, nullptr); }
+
+}  // namespace __dfsan
+
 #if SANITIZER_CAN_USE_PREINIT_ARRAY
-__attribute__((section(".preinit_array"), used))
-static void (*dfsan_init_ptr)(int, char **, char **) = dfsan_init;
+__attribute__((section(".preinit_array"),
+               used)) static void (*dfsan_init_ptr)(int, char **,
+                                                    char **) = DFsanInit;
 #endif

compiler-rt/lib/dfsan/dfsan_allocator.h

+//===-- dfsan_allocator.h ---------------------------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of DataflowSanitizer.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef DFSAN_ALLOCATOR_H
+#define DFSAN_ALLOCATOR_H
+
+#include "sanitizer_common/sanitizer_common.h"
+
+namespace __dfsan {
+
+struct DFsanThreadLocalMallocStorage {
+  ALIGNED(8) uptr allocator_cache[96 * (512 * 8 + 16)];  // Opaque.
+  void CommitBack();
+
+ private:
+  // These objects are allocated via mmap() and are zero-initialized.
+  DFsanThreadLocalMallocStorage() {}
+};
+
+}  // namespace __dfsan
+#endif  // DFSAN_ALLOCATOR_H

compiler-rt/lib/dfsan/dfsan_allocator.cpp

+//===-- dfsan_allocator.cpp -------------------------- --------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of DataflowSanitizer.
+//
+// DataflowSanitizer allocator.
+//===----------------------------------------------------------------------===//
+
+#include "dfsan_allocator.h"
+
+#include "dfsan.h"
+#include "dfsan_thread.h"
+#include "sanitizer_common/sanitizer_allocator.h"
+#include "sanitizer_common/sanitizer_allocator_checks.h"
+#include "sanitizer_common/sanitizer_allocator_interface.h"
+#include "sanitizer_common/sanitizer_allocator_report.h"
+#include "sanitizer_common/sanitizer_errno.h"
+
+namespace __dfsan {
+
+struct Metadata {
+  uptr requested_size;
+};
+
+struct DFsanMapUnmapCallback {
+  void OnMap(uptr p, uptr size) const { dfsan_set_label(0, (void *)p, size); }
+  void OnUnmap(uptr p, uptr size) const {
+    __dfsan::dfsan_release_meta_memory((const void *)p, size);
+  }
+};
+
+// TODO: DFSan assumes application memory starts from 0x700000008000. For
+// unknown reason, the sanitizer allocator does not support any start address
+// between 0x701000000000 and 0x700000008000. After switching to fast8labels
+// mode, DFSan memory layout will be changed to the same to MSan's. Then we
+// set the start address to 0x700000000000 as MSan.
+static const uptr kAllocatorSpace = 0x701000000000ULL;
+static const uptr kMaxAllowedMallocSize = 8UL << 30;
+
+struct AP64 {  // Allocator64 parameters. Deliberately using a short name.
+  static const uptr kSpaceBeg = kAllocatorSpace;
+  static const uptr kSpaceSize = 0x40000000000;  // 4T.
+  static const uptr kMetadataSize = sizeof(Metadata);
+  typedef DefaultSizeClassMap SizeClassMap;
+  typedef DFsanMapUnmapCallback MapUnmapCallback;
+  static const uptr kFlags = 0;
+  using AddressSpaceView = LocalAddressSpaceView;
+};
+
+typedef SizeClassAllocator64<AP64> PrimaryAllocator;
+
+typedef CombinedAllocator<PrimaryAllocator> Allocator;
+typedef Allocator::AllocatorCache AllocatorCache;
+
+static Allocator allocator;
+static AllocatorCache fallback_allocator_cache;
+static StaticSpinMutex fallback_mutex;
+
+static uptr max_malloc_size;
+
+void dfsan_allocator_init() {
+  SetAllocatorMayReturnNull(common_flags()->allocator_may_return_null);
+  allocator.Init(common_flags()->allocator_release_to_os_interval_ms);
+  if (common_flags()->max_allocation_size_mb)
+    max_malloc_size = Min(common_flags()->max_allocation_size_mb << 20,
+                          kMaxAllowedMallocSize);
+  else
+    max_malloc_size = kMaxAllowedMallocSize;
+}
+
+AllocatorCache *GetAllocatorCache(DFsanThreadLocalMallocStorage *ms) {
+  CHECK(ms);
+  CHECK_LE(sizeof(AllocatorCache), sizeof(ms->allocator_cache));
+  return reinterpret_cast<AllocatorCache *>(ms->allocator_cache);
+}
+
+void DFsanThreadLocalMallocStorage::CommitBack() {
+  allocator.SwallowCache(GetAllocatorCache(this));
+}
+
+static void *DFsanAllocate(uptr size, uptr alignment, bool zeroise) {
+  if (size > max_malloc_size) {
+    if (AllocatorMayReturnNull()) {
+      Report("WARNING: DataflowSanitizer failed to allocate 0x%zx bytes\n",
+             size);
+      return nullptr;
+    }
+    BufferedStackTrace stack;
+    ReportAllocationSizeTooBig(size, max_malloc_size, &stack);
+  }
+  DFsanThread *t = GetCurrentThread();
+  void *allocated;
+  if (t) {
+    AllocatorCache *cache = GetAllocatorCache(&t->malloc_storage());
+    allocated = allocator.Allocate(cache, size, alignment);
+  } else {
+    SpinMutexLock l(&fallback_mutex);
+    AllocatorCache *cache = &fallback_allocator_cache;
+    allocated = allocator.Allocate(cache, size, alignment);
+  }
+  if (UNLIKELY(!allocated)) {
+    SetAllocatorOutOfMemory();
+    if (AllocatorMayReturnNull())
+      return nullptr;
+    BufferedStackTrace stack;
+    ReportOutOfMemory(size, &stack);
+  }
+  Metadata *meta =
+      reinterpret_cast<Metadata *>(allocator.GetMetaData(allocated));
+  meta->requested_size = size;
+  if (zeroise) {
+    internal_memset(allocated, 0, size);
+    dfsan_set_label(0, allocated, size);
+  } else if (flags().zero_in_malloc) {
+    dfsan_set_label(0, allocated, size);
+  }
+  return allocated;
+}
+
+void dfsan_deallocate(void *p) {
+  CHECK(p);
+  Metadata *meta = reinterpret_cast<Metadata *>(allocator.GetMetaData(p));
+  uptr size = meta->requested_size;
+  meta->requested_size = 0;
+  if (flags().zero_in_free)
+    dfsan_set_label(0, p, size);
+  DFsanThread *t = GetCurrentThread();
+  if (t) {
+    AllocatorCache *cache = GetAllocatorCache(&t->malloc_storage());
+    allocator.Deallocate(cache, p);
+  } else {
+    SpinMutexLock l(&fallback_mutex);
+    AllocatorCache *cache = &fallback_allocator_cache;
+    allocator.Deallocate(cache, p);
+  }
+}
+
+void *DFsanReallocate(void *old_p, uptr new_size, uptr alignment) {
+  Metadata *meta = reinterpret_cast<Metadata *>(allocator.GetMetaData(old_p));
+  uptr old_size = meta->requested_size;
+  uptr actually_allocated_size = allocator.GetActuallyAllocatedSize(old_p);
+  if (new_size <= actually_allocated_size) {
+    // We are not reallocating here.
+    meta->requested_size = new_size;
+    if (new_size > old_size && flags().zero_in_malloc)
+      dfsan_set_label(0, (char *)old_p + old_size, new_size - old_size);
+    return old_p;
+  }
+  uptr memcpy_size = Min(new_size, old_size);
+  void *new_p = DFsanAllocate(new_size, alignment, false /*zeroise*/);
+  if (new_p) {
+    dfsan_copy_memory(new_p, old_p, memcpy_size);
+    dfsan_deallocate(old_p);
+  }
+  return new_p;
+}
+
+void *DFsanCalloc(uptr nmemb, uptr size) {
+  if (UNLIKELY(CheckForCallocOverflow(size, nmemb))) {
+    if (AllocatorMayReturnNull())
+      return nullptr;
+    BufferedStackTrace stack;
+    ReportCallocOverflow(nmemb, size, &stack);
+  }
+  return DFsanAllocate(nmemb * size, sizeof(u64), true /*zeroise*/);
+}
+
+static uptr AllocationSize(const void *p) {
+  if (!p)
+    return 0;
+  const void *beg = allocator.GetBlockBegin(p);
+  if (beg != p)
+    return 0;
+  Metadata *b = (Metadata *)allocator.GetMetaData(p);
+  return b->requested_size;
+}
+
+void *dfsan_malloc(uptr size) {
+  return SetErrnoOnNull(DFsanAllocate(size, sizeof(u64), false /*zeroise*/));
+}
+
+void *dfsan_calloc(uptr nmemb, uptr size) {
+  return SetErrnoOnNull(DFsanCalloc(nmemb, size));
+}
+
+void *dfsan_realloc(void *ptr, uptr size) {
+  if (!ptr)
+    return SetErrnoOnNull(DFsanAllocate(size, sizeof(u64), false /*zeroise*/));
+  if (size == 0) {
+    dfsan_deallocate(ptr);
+    return nullptr;
+  }
+  return SetErrnoOnNull(DFsanReallocate(ptr, size, sizeof(u64)));
+}
+
+void *dfsan_reallocarray(void *ptr, uptr nmemb, uptr size) {
+  if (UNLIKELY(CheckForCallocOverflow(size, nmemb))) {
+    errno = errno_ENOMEM;
+    if (AllocatorMayReturnNull())
+      return nullptr;
+    BufferedStackTrace stack;
+    ReportReallocArrayOverflow(nmemb, size, &stack);
+  }
+  return dfsan_realloc(ptr, nmemb * size);
+}
+
+void *dfsan_valloc(uptr size) {
+  return SetErrnoOnNull(
+      DFsanAllocate(size, GetPageSizeCached(), false /*zeroise*/));
+}
+
+void *dfsan_pvalloc(uptr size) {
+  uptr PageSize = GetPageSizeCached();
+  if (UNLIKELY(CheckForPvallocOverflow(size, PageSize))) {
+    errno = errno_ENOMEM;
+    if (AllocatorMayReturnNull())
+      return nullptr;
+    BufferedStackTrace stack;
+    ReportPvallocOverflow(size, &stack);
+  }
+  // pvalloc(0) should allocate one page.
+  size = size ? RoundUpTo(size, PageSize) : PageSize;
+  return SetErrnoOnNull(DFsanAllocate(size, PageSize, false /*zeroise*/));
+}
+
+void *dfsan_aligned_alloc(uptr alignment, uptr size) {
+  if (UNLIKELY(!CheckAlignedAllocAlignmentAndSize(alignment, size))) {
+    errno = errno_EINVAL;
+    if (AllocatorMayReturnNull())
+      return nullptr;
+    BufferedStackTrace stack;
+    ReportInvalidAlignedAllocAlignment(size, alignment, &stack);
+  }
+  return SetErrnoOnNull(DFsanAllocate(size, alignment, false /*zeroise*/));
+}
+
+void *dfsan_memalign(uptr alignment, uptr size) {
+  if (UNLIKELY(!IsPowerOfTwo(alignment))) {
+    errno = errno_EINVAL;
+    if (AllocatorMayReturnNull())
+      return nullptr;
+    BufferedStackTrace stack;
+    ReportInvalidAllocationAlignment(alignment, &stack);
+  }
+  return SetErrnoOnNull(DFsanAllocate(size, alignment, false /*zeroise*/));
+}
+
+int dfsan_posix_memalign(void **memptr, uptr alignment, uptr size) {
+  if (UNLIKELY(!CheckPosixMemalignAlignment(alignment))) {
+    if (AllocatorMayReturnNull())
+      return errno_EINVAL;
+    BufferedStackTrace stack;
+    ReportInvalidPosixMemalignAlignment(alignment, &stack);
+  }
+  void *ptr = DFsanAllocate(size, alignment, false /*zeroise*/);
+  if (UNLIKELY(!ptr))
+    // OOM error is already taken care of by DFsanAllocate.
+    return errno_ENOMEM;
+  CHECK(IsAligned((uptr)ptr, alignment));
+  *memptr = ptr;
+  return 0;
+}
+
+}  // namespace __dfsan
+
+using namespace __dfsan;
+
+uptr __sanitizer_get_current_allocated_bytes() {
+  uptr stats[AllocatorStatCount];
+  allocator.GetStats(stats);
+  return stats[AllocatorStatAllocated];
+}
+
+uptr __sanitizer_get_heap_size() {
+  uptr stats[AllocatorStatCount];
+  allocator.GetStats(stats);
+  return stats[AllocatorStatMapped];
+}
+
+uptr __sanitizer_get_free_bytes() { return 1; }
+
+uptr __sanitizer_get_unmapped_bytes() { return 1; }
+
+uptr __sanitizer_get_estimated_allocated_size(uptr size) { return size; }
+
+int __sanitizer_get_ownership(const void *p) { return AllocationSize(p) != 0; }
+
+uptr __sanitizer_get_allocated_size(const void *p) { return AllocationSize(p); }

compiler-rt/lib/dfsan/dfsan_flags.inc

 //===-- dfsan_flags.inc -----------------------------------------*- C++ -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
 // DFSan runtime flags.
 //
 //===----------------------------------------------------------------------===//
 #ifndef DFSAN_FLAG
 # error "Define DFSAN_FLAG prior to including this file!"
 #endif
 
 // DFSAN_FLAG(Type, Name, DefaultValue, Description)
 // See COMMON_FLAG in sanitizer_flags.inc for more details.
 
 DFSAN_FLAG(bool, warn_unimplemented, true,
            "Whether to warn on unimplemented functions.")
 DFSAN_FLAG(bool, warn_nonzero_labels, false,
            "Whether to warn on unimplemented functions.")
 DFSAN_FLAG(
     bool, strict_data_dependencies, true,
     "Whether to propagate labels only when there is an obvious data dependency"
     "(e.g., when comparing strings, ignore the fact that the output of the"
     "comparison might be data-dependent on the content of the strings). This"
     "applies only to the custom functions defined in 'custom.c'.")
 DFSAN_FLAG(const char *, dump_labels_at_exit, "", "The path of the file where "
                                                   "to dump the labels when the "
                                                   "program terminates.")
 DFSAN_FLAG(
      int, origin_history_size, Origin::kMaxDepth,
     "The limit of origin chain length. Non-positive values mean unlimited.")
 DFSAN_FLAG(
      int, origin_history_per_stack_limit, 20000,
     "The limit of origin node's references count. "
     "Non-positive values mean unlimited.")
 DFSAN_FLAG(int, store_context_size, 20,
            "The depth limit of origin tracking stack traces.")
 DFSAN_FLAG(bool, check_origin_invariant, false,
            "Whether to check if the origin invariant holds.")
+DFSAN_FLAG(bool, zero_in_malloc, true,
+           "Whether to zero shadow space of new allocated memory.")
+DFSAN_FLAG(bool, zero_in_free, true,
+           "Whether to zero shadow space of deallocated memory.")

compiler-rt/lib/dfsan/dfsan_thread.h

 //===-- dfsan_thread.h -------------------------------------------*- C++
 //-*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
 // This file is a part of DataFlowSanitizer.
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef DFSAN_THREAD_H
 #define DFSAN_THREAD_H
 
+#include "dfsan_allocator.h"
 #include "sanitizer_common/sanitizer_common.h"
 
 namespace __dfsan {
 
 class DFsanThread {
  public:
   // NOTE: There is no DFsanThread constructor. It is allocated
   // via mmap() and *must* be valid in zero-initialized state.
 
   static DFsanThread *Create(void *start_routine_trampoline,
                              thread_callback_t start_routine, void *arg,
                              bool track_origins = false);
   static void TSDDtor(void *tsd);
   void Destroy();
 
   void Init();  // Should be called from the thread itself.
   thread_return_t ThreadStart();
 
   uptr stack_top();
   uptr stack_bottom();
+  uptr tls_begin() { return tls_begin_; }
+  uptr tls_end() { return tls_end_; }
   bool IsMainThread() { return start_routine_ == nullptr; }
 
   bool InSignalHandler() { return in_signal_handler_; }
   void EnterSignalHandler() { in_signal_handler_++; }
   void LeaveSignalHandler() { in_signal_handler_--; }
 
+  DFsanThreadLocalMallocStorage &malloc_storage() { return malloc_storage_; }
+
   int destructor_iterations_;
 
  private:
   void SetThreadStackAndTls();
+  void ClearShadowForThreadStackAndTLS();
   struct StackBounds {
     uptr bottom;
     uptr top;
   };
   StackBounds GetStackBounds() const;
 
   bool AddrIsInStack(uptr addr);
 
   void *start_routine_trampoline_;
   thread_callback_t start_routine_;
   void *arg_;
   bool track_origins_;
 
   StackBounds stack_;
 
+  uptr tls_begin_;
+  uptr tls_end_;
+
   unsigned in_signal_handler_;
+
+  DFsanThreadLocalMallocStorage malloc_storage_;
 };
 
 DFsanThread *GetCurrentThread();
 void SetCurrentThread(DFsanThread *t);
 void DFsanTSDInit(void (*destructor)(void *tsd));
 void DFsanTSDDtor(void *tsd);
 
 }  // namespace __dfsan
 
 #endif  // DFSAN_THREAD_H

compiler-rt/lib/dfsan/dfsan_thread.cpp

 #include "dfsan_thread.h"
 
 #include <pthread.h>
 
 #include "dfsan.h"
+#include "sanitizer_common/sanitizer_tls_get_addr.h"
 
 namespace __dfsan {
 
 DFsanThread *DFsanThread::Create(void *start_routine_trampoline,
                                  thread_callback_t start_routine, void *arg,
                                  bool track_origins) {
   uptr PageSize = GetPageSizeCached();
   uptr size = RoundUpTo(sizeof(DFsanThread), PageSize);
   DFsanThread *thread = (DFsanThread *)MmapOrDie(size, __func__);
   thread->start_routine_trampoline_ = start_routine_trampoline;
   thread->start_routine_ = start_routine;
   thread->arg_ = arg;
   thread->track_origins_ = track_origins;
   thread->destructor_iterations_ = GetPthreadDestructorIterations();
 
   return thread;
 }
 
 void DFsanThread::SetThreadStackAndTls() {
   uptr tls_size = 0;
   uptr stack_size = 0;
-  uptr tls_begin;
-  GetThreadStackAndTls(IsMainThread(), &stack_.bottom, &stack_size, &tls_begin,
+  GetThreadStackAndTls(IsMainThread(), &stack_.bottom, &stack_size, &tls_begin_,
                        &tls_size);
   stack_.top = stack_.bottom + stack_size;
+  tls_end_ = tls_begin_ + tls_size;
 
   int local;
   CHECK(AddrIsInStack((uptr)&local));
 }
 
-void DFsanThread::Init() { SetThreadStackAndTls(); }
+void DFsanThread::ClearShadowForThreadStackAndTLS() {
+  dfsan_set_label(0, (void *)stack_.bottom, stack_.top - stack_.bottom);
+  if (tls_begin_ != tls_end_)
+    dfsan_set_label(0, (void *)tls_begin_, tls_end_ - tls_begin_);
+  DTLS *dtls = DTLS_Get();
+  CHECK_NE(dtls, 0);
+  ForEachDVT(dtls, [](const DTLS::DTV &dtv, int id) {
+    dfsan_set_label(0, (void *)(dtv.beg), dtv.size);
+  });
+}
+
+void DFsanThread::Init() {
+  SetThreadStackAndTls();
+  ClearShadowForThreadStackAndTLS();
+}
 
 void DFsanThread::TSDDtor(void *tsd) {
   DFsanThread *t = (DFsanThread *)tsd;
   t->Destroy();
 }
 
 void DFsanThread::Destroy() {
+  malloc_storage().CommitBack();
+  // We also clear the shadow on thread destruction because
+  // some code may still be executing in later TSD destructors
+  // and we don't want it to have any poisoned stack.
+  ClearShadowForThreadStackAndTLS();
   uptr size = RoundUpTo(sizeof(DFsanThread), GetPageSizeCached());
   UnmapOrDie(this, size);
+  DTLS_Destroy();
 }
 
 thread_return_t DFsanThread::ThreadStart() {
   Init();
 
   if (!start_routine_) {
     // start_routine_ == 0 if we're on the main thread or on one of the
     // OS X libdispatch worker threads. But nobody is supposed to call