void TableTicker::doNativeBacktrace(ThreadProfile &aProfile, TickSample* aSample)
{
void *pc_array[1000];
void *sp_array[1000];
PCArray array = {
pc_array,
sp_array,
mozilla::ArrayLength(pc_array),
0
};
ucontext_t *ucontext = reinterpret_cast<ucontext_t *>(aSample->context);
array.count = EHABIStackWalk(ucontext->uc_mcontext, aProfile.GetStackTop(),
sp_array, pc_array, array.size);
mergeNativeBacktrace(aProfile, array);
}
void TableTicker::doNativeBacktrace(ThreadProfile &aProfile, TickSample* aSample)
{
void *pc_array[1000];
void *sp_array[1000];
PCArray array = {
pc_array,
sp_array,
mozilla::ArrayLength(pc_array),
0
};
const mcontext_t *mcontext = &reinterpret_cast<ucontext_t *>(aSample->context)->uc_mcontext;
mcontext_t savedContext;
PseudoStack *pseudoStack = aProfile.GetPseudoStack();
array.count = 0;
// The pseudostack contains an "EnterJIT" frame whenever we enter
// JIT code with profiling enabled; the stack pointer value points
// the saved registers. We use this to unwind resume unwinding
// after encounting JIT code.
for (uint32_t i = pseudoStack->stackSize(); i > 0; --i) {
// The pseudostack grows towards higher indices, so we iterate
// backwards (from callee to caller).
volatile StackEntry &entry = pseudoStack->mStack[i - 1];
if (!entry.js() && strcmp(entry.label(), "EnterJIT") == 0) {
// Found JIT entry frame. Unwind up to that point (i.e., force
// the stack walk to stop before the block of saved registers;
// note that it yields nondecreasing stack pointers), then restore
// the saved state.
uint32_t *vSP = reinterpret_cast<uint32_t*>(entry.stackAddress());
array.count += EHABIStackWalk(*mcontext,
/* stackBase = */ vSP,
sp_array + array.count,
pc_array + array.count,
array.size - array.count);
memset(&savedContext, 0, sizeof(savedContext));
// See also: struct EnterJITStack in js/src/jit/arm/Trampoline-arm.cpp
savedContext.arm_r4 = *vSP++;
savedContext.arm_r5 = *vSP++;
savedContext.arm_r6 = *vSP++;
savedContext.arm_r7 = *vSP++;
savedContext.arm_r8 = *vSP++;
savedContext.arm_r9 = *vSP++;
savedContext.arm_r10 = *vSP++;
savedContext.arm_fp = *vSP++;
savedContext.arm_lr = *vSP++;
savedContext.arm_sp = reinterpret_cast<uint32_t>(vSP);
savedContext.arm_pc = savedContext.arm_lr;
mcontext = &savedContext;
}
}
// Now unwind whatever's left (starting from either the last EnterJIT
// frame or, if no EnterJIT was found, the original registers).
array.count += EHABIStackWalk(*mcontext,
aProfile.GetStackTop(),
sp_array + array.count,
pc_array + array.count,
array.size - array.count);
mergeNativeBacktrace(aProfile, array);
}
void GeckoSampler::doNativeBacktrace(ThreadProfile &aProfile, TickSample* aSample)
{
const mcontext_t* mc
= &reinterpret_cast<ucontext_t *>(aSample->context)->uc_mcontext;
lul::UnwindRegs startRegs;
memset(&startRegs, 0, sizeof(startRegs));
# if defined(SPS_PLAT_amd64_linux)
startRegs.xip = lul::TaggedUWord(mc->gregs[REG_RIP]);
startRegs.xsp = lul::TaggedUWord(mc->gregs[REG_RSP]);
startRegs.xbp = lul::TaggedUWord(mc->gregs[REG_RBP]);
# elif defined(SPS_PLAT_arm_android)
startRegs.r15 = lul::TaggedUWord(mc->arm_pc);
startRegs.r14 = lul::TaggedUWord(mc->arm_lr);
startRegs.r13 = lul::TaggedUWord(mc->arm_sp);
startRegs.r12 = lul::TaggedUWord(mc->arm_ip);
startRegs.r11 = lul::TaggedUWord(mc->arm_fp);
startRegs.r7 = lul::TaggedUWord(mc->arm_r7);
# elif defined(SPS_PLAT_x86_linux) || defined(SPS_PLAT_x86_android)
startRegs.xip = lul::TaggedUWord(mc->gregs[REG_EIP]);
startRegs.xsp = lul::TaggedUWord(mc->gregs[REG_ESP]);
startRegs.xbp = lul::TaggedUWord(mc->gregs[REG_EBP]);
# else
# error "Unknown plat"
# endif
/* Copy up to N_STACK_BYTES from rsp-REDZONE upwards, but not
going past the stack's registered top point. Do some basic
sanity checks too. This assumes that the TaggedUWord holding
the stack pointer value is valid, but it should be, since it
was constructed that way in the code just above. */
lul::StackImage stackImg;
{
# if defined(SPS_PLAT_amd64_linux)
uintptr_t rEDZONE_SIZE = 128;
uintptr_t start = startRegs.xsp.Value() - rEDZONE_SIZE;
# elif defined(SPS_PLAT_arm_android)
uintptr_t rEDZONE_SIZE = 0;
uintptr_t start = startRegs.r13.Value() - rEDZONE_SIZE;
# elif defined(SPS_PLAT_x86_linux) || defined(SPS_PLAT_x86_android)
uintptr_t rEDZONE_SIZE = 0;
uintptr_t start = startRegs.xsp.Value() - rEDZONE_SIZE;
# else
# error "Unknown plat"
# endif
uintptr_t end = reinterpret_cast<uintptr_t>(aProfile.GetStackTop());
uintptr_t ws = sizeof(void*);
start &= ~(ws-1);
end &= ~(ws-1);
uintptr_t nToCopy = 0;
if (start < end) {
nToCopy = end - start;
if (nToCopy > lul::N_STACK_BYTES)
nToCopy = lul::N_STACK_BYTES;
}
MOZ_ASSERT(nToCopy <= lul::N_STACK_BYTES);
stackImg.mLen = nToCopy;
stackImg.mStartAvma = start;
if (nToCopy > 0) {
memcpy(&stackImg.mContents[0], (void*)start, nToCopy);
(void)VALGRIND_MAKE_MEM_DEFINED(&stackImg.mContents[0], nToCopy);
}
}
// The maximum number of frames that LUL will produce. Setting it
// too high gives a risk of it wasting a lot of time looping on
// corrupted stacks.
const int MAX_NATIVE_FRAMES = 256;
size_t scannedFramesAllowed = 0;
uintptr_t framePCs[MAX_NATIVE_FRAMES];
uintptr_t frameSPs[MAX_NATIVE_FRAMES];
size_t framesAvail = mozilla::ArrayLength(framePCs);
size_t framesUsed = 0;
size_t scannedFramesAcquired = 0;
sLUL->Unwind( &framePCs[0], &frameSPs[0],
&framesUsed, &scannedFramesAcquired,
framesAvail, scannedFramesAllowed,
&startRegs, &stackImg );
NativeStack nativeStack = {
reinterpret_cast<void**>(framePCs),
reinterpret_cast<void**>(frameSPs),
mozilla::ArrayLength(framePCs),
0
};
nativeStack.count = framesUsed;
mergeStacksIntoProfile(aProfile, aSample, nativeStack);
// Update stats in the LUL stats object. Unfortunately this requires
// three global memory operations.
sLUL->mStats.mContext += 1;
sLUL->mStats.mCFI += framesUsed - 1 - scannedFramesAcquired;
sLUL->mStats.mScanned += scannedFramesAcquired;
}
