[esan|wset] Iterate all memory to compute the total working set
Summary:
Adds iteration of all application memory in an efficient manner using
shadow faults. Shadow memory starts out inaccessible and we mark it
writable one page at a time on each fault when the instrumentation touches
it. This allows iteration over just the mapped shadow memory, saving
significant time.
Adds a process-end iteration and pretty-printing of the final result.
Adds a new test and updates the existing tests.
Reviewers: aizatsky, filcab
Subscribers: vitalybuka, zhaoqin, kcc, eugenis, llvm-commits, kubabrecka
Differential Revision: http://reviews.llvm.org/D20578
llvm-svn: 271277
diff --git a/compiler-rt/lib/esan/esan.cpp b/compiler-rt/lib/esan/esan.cpp
index c9629d6..f0a4965 100644
--- a/compiler-rt/lib/esan/esan.cpp
+++ b/compiler-rt/lib/esan/esan.cpp
@@ -149,8 +149,16 @@
VPrintf(1, "Shadow #%d: [%zx-%zx) (%zuGB)\n", i, ShadowStart, ShadowEnd,
(ShadowEnd - ShadowStart) >> 30);
- uptr Map = (uptr)MmapFixedNoReserve(ShadowStart, ShadowEnd - ShadowStart,
- "shadow");
+ uptr Map;
+ if (WhichTool == ESAN_WorkingSet) {
+ // We want to identify all shadow pages that are touched so we start
+ // out inaccessible.
+ Map = (uptr)MmapFixedNoAccess(ShadowStart, ShadowEnd- ShadowStart,
+ "shadow");
+ } else {
+ Map = (uptr)MmapFixedNoReserve(ShadowStart, ShadowEnd - ShadowStart,
+ "shadow");
+ }
if (Map != ShadowStart) {
Printf("FATAL: EfficiencySanitizer failed to map its shadow memory.\n");
Die();
diff --git a/compiler-rt/lib/esan/working_set.cpp b/compiler-rt/lib/esan/working_set.cpp
index 68ea9e7..dcd6837 100644
--- a/compiler-rt/lib/esan/working_set.cpp
+++ b/compiler-rt/lib/esan/working_set.cpp
@@ -16,6 +16,7 @@
#include "esan.h"
#include "esan_flags.h"
#include "esan_shadow.h"
+#include "sanitizer_common/sanitizer_procmaps.h"
// We shadow every cache line of app memory with one shadow byte.
// - The highest bit of each shadow byte indicates whether the corresponding
@@ -30,6 +31,9 @@
namespace __esan {
+// Our shadow memory assumes that the line size is 64.
+static const u32 CacheLineSize = 64;
+
// See the shadow byte layout description above.
static const u32 TotalWorkingSetBitIdx = 7;
static const u32 CurWorkingSetBitIdx = 0;
@@ -75,16 +79,80 @@
}
}
+// This routine will word-align ShadowStart and ShadowEnd prior to scanning.
+static u32 countAndClearShadowValues(u32 BitIdx, uptr ShadowStart,
+ uptr ShadowEnd) {
+ u32 WorkingSetSize = 0;
+ u32 ByteValue = 0x1 << BitIdx;
+ u32 WordValue = ByteValue | ByteValue << 8 | ByteValue << 16 |
+ ByteValue << 24;
+ // Get word aligned start.
+ ShadowStart = RoundDownTo(ShadowStart, sizeof(u32));
+ for (u32 *Ptr = (u32 *)ShadowStart; Ptr < (u32 *)ShadowEnd; ++Ptr) {
+ if ((*Ptr & WordValue) != 0) {
+ byte *BytePtr = (byte *)Ptr;
+ for (u32 j = 0; j < sizeof(u32); ++j) {
+ if (BytePtr[j] & ByteValue) {
+ ++WorkingSetSize;
+ // TODO: Accumulate to the lower-frequency bit to the left.
+ }
+ }
+ // Clear this bit from every shadow byte.
+ *Ptr &= ~WordValue;
+ }
+ }
+ return WorkingSetSize;
+}
+
+// Scan shadow memory to calculate the number of cache lines being accessed,
+// i.e., the number of non-zero bits indexed by BitIdx in each shadow byte.
+// We also clear the lowest bits (most recent working set snapshot).
+static u32 computeWorkingSizeAndReset(u32 BitIdx) {
+ u32 WorkingSetSize = 0;
+ MemoryMappingLayout MemIter(true/*cache*/);
+ uptr Start, End, Prot;
+ while (MemIter.Next(&Start, &End, nullptr/*offs*/, nullptr/*file*/,
+ 0/*file size*/, &Prot)) {
+ VPrintf(4, "%s: considering %p-%p app=%d shadow=%d prot=%u\n",
+ __FUNCTION__, Start, End, Prot, isAppMem(Start),
+ isShadowMem(Start));
+ if (isShadowMem(Start) && (Prot & MemoryMappingLayout::kProtectionWrite)) {
+ VPrintf(3, "%s: walking %p-%p\n", __FUNCTION__, Start, End);
+ WorkingSetSize += countAndClearShadowValues(BitIdx, Start, End);
+ }
+ }
+ return WorkingSetSize;
+}
+
void initializeWorkingSet() {
- // The shadow mapping assumes 64 so this cannot be changed.
- CHECK(getFlags()->cache_line_size == 64);
+ CHECK(getFlags()->cache_line_size == CacheLineSize);
registerMemoryFaultHandler();
}
+static u32 getSizeForPrinting(u32 NumOfCachelines, const char *&Unit) {
+ // We need a constant to avoid software divide support:
+ static const u32 KilobyteCachelines = (0x1 << 10) / CacheLineSize;
+ static const u32 MegabyteCachelines = KilobyteCachelines << 10;
+
+ if (NumOfCachelines > 10 * MegabyteCachelines) {
+ Unit = "MB";
+ return NumOfCachelines / MegabyteCachelines;
+ } else if (NumOfCachelines > 10 * KilobyteCachelines) {
+ Unit = "KB";
+ return NumOfCachelines / KilobyteCachelines;
+ } else {
+ Unit = "Bytes";
+ return NumOfCachelines * CacheLineSize;
+ }
+}
+
int finalizeWorkingSet() {
- // FIXME NYI: we need to add memory scanning to report the total lines
- // touched, and later add sampling to get intermediate values.
- Report("%s is not finished: nothing yet to report\n", SanitizerToolName);
+ // Get the working set size for the entire execution.
+ u32 NumOfCachelines = computeWorkingSizeAndReset(TotalWorkingSetBitIdx);
+ const char *Unit;
+ u32 Size = getSizeForPrinting(NumOfCachelines, Unit);
+ Report(" %s: the total working set size: %u %s (%u cache lines)\n",
+ SanitizerToolName, Size, Unit, NumOfCachelines);
return 0;
}
diff --git a/compiler-rt/lib/esan/working_set_posix.cpp b/compiler-rt/lib/esan/working_set_posix.cpp
index f019176..0f36c86 100644
--- a/compiler-rt/lib/esan/working_set_posix.cpp
+++ b/compiler-rt/lib/esan/working_set_posix.cpp
@@ -68,10 +68,16 @@
// app to handle it just as the app would do without our tool in place.
static void handleMemoryFault(int SigNum, void *Info, void *Ctx) {
if (SigNum == SIGSEGV) {
-
- // TODO: Add shadow memory fault detection and handling.
-
- if (AppSigAct.sigaction) {
+ // We rely on si_addr being filled in (thus we do not support old kernels).
+ siginfo_t *SigInfo = (siginfo_t *)Info;
+ uptr Addr = (uptr)SigInfo->si_addr;
+ if (isShadowMem(Addr)) {
+ VPrintf(3, "Shadow fault @%p\n", Addr);
+ uptr PageSize = GetPageSizeCached();
+ int Res = internal_mprotect((void *)RoundDownTo(Addr, PageSize),
+ PageSize, PROT_READ|PROT_WRITE);
+ CHECK(Res == 0);
+ } else if (AppSigAct.sigaction) {
// FIXME: For simplicity we ignore app options including its signal stack
// (we just use ours) and all the delivery flags.
AppSigAct.sigaction(SigNum, Info, Ctx);