buffer: segfault writing values with noAssert=true

[bnoordhuis]

Reported by @guidovranken. Test case:

$ gdb --args ./out/Release/node -e 'new Buffer(10).writeFloatBE(1, 0xFFFFFFFF-1000, 1);'
Reading symbols from ./out/Release/node...done.
(gdb) run
# <elided>
Thread 1 "node" received signal SIGSEGV, Segmentation fault.
0x00007ffff6be36be in __memcpy_sse2_unaligned () from /lib64/libc.so.6
Missing separate debuginfos, use: dnf debuginfo-install libgcc-6.1.1-3.fc24.x86_64 libstdc++-6.1.1-3.fc24.x86_64
(gdb) backtrace 5
#0  0x00007ffff6be36be in __memcpy_sse2_unaligned () from /lib64/libc.so.6
#1  0x00000000012cf00c in void node::Buffer::WriteFloatGeneric<float, (node::Endianness)1>(v8::FunctionCallbackInfo<v8::Value> const&) ()
#2  0x0000000000a129dd in v8::internal::FunctionCallbackArguments::Call(void (*)(v8::FunctionCallbackInfo<v8::Value> const&)) ()
#3  0x0000000000a882b8 in v8::internal::MaybeHandle<v8::internal::Object> v8::internal::(anonymous namespace)::HandleApiCallHelper<false>(v8::internal::Isolate*, v8::internal::Ha
ndle<v8::internal::HeapObject>, v8::internal::Handle<v8::internal::HeapObject>, v8::internal::Handle<v8::internal::FunctionTemplateInfo>, v8::internal::Handle<v8::internal::Object>, v8::internal::BuiltinArguments) ()
#4  0x0000000000a88fcd in v8::internal::Builtin_HandleApiCall(int, v8::internal::Object**, v8::internal::Isolate*) ()

The documentation says this:

`offset` {Integer} Where to start writing. Must satisfy: `0 <= offset <= buf.length - 4`
`noAssert` {Boolean} Skip `value` and `offset` validation? **Default:** `false`

IOW, it's technically allowed for node.js to crash but whether that's actually a good idea is something reasonable people can disagree on. Anyone have opinions on either:

1. Removing noAssert; i.e., always checking the inputs, or
2. Skipping out-of-bounds reads and writes?

[addaleax]

Skipping out-of-bounds reads and writes?

If it doesn’t benchmark horribly, I think we should do that.

[rvagg]

any idea what the cost of permanently turning on noAssert are? that seems like reasonable behaviour to me but I don't really know the implications.

[bnoordhuis]

It performs about the same. This is master with noAssert=true:

 Performance counter stats for 'out/Release/node -e var b = Buffer.alloc(8); for (var i = 0; i < 5e7; ++i) b.writeDoubleLE(1.0, 0, true)' (5 runs):

       2744.897728      task-clock:u (msec)       #    1.000 CPUs utilized            ( +-  1.04% )
                 0      context-switches:u        #    0.000 K/sec                  
                 0      cpu-migrations:u          #    0.000 K/sec                  
              2503      page-faults:u             #    0.912 K/sec                    ( +-  0.01% )

       2.745411566 seconds time elapsed                                          ( +-  1.05% )

Master with noAssert=false. Omitting noAssert benchmarks the same:

 Performance counter stats for 'out/Release/node -e var b = Buffer.alloc(8); for (var i = 0; i < 5e7; ++i) b.writeDoubleLE(1.0, 0, false)' (5 runs):

       3019.762714      task-clock:u (msec)       #    1.000 CPUs utilized            ( +-  1.01% )
                 0      context-switches:u        #    0.000 K/sec                  
                 0      cpu-migrations:u          #    0.000 K/sec                  
              2504      page-faults:u             #    0.829 K/sec                    ( +-  0.03% )

       3.020504516 seconds time elapsed                                          ( +-  1.02% )

And this is with noAssert removed, see patch below:

 Performance counter stats for 'out/Release/node -e var b = Buffer.alloc(8); for (var i = 0; i < 5e7; ++i) b.writeDoubleLE(1.0, 0)' (5 runs):

       2750.038828      task-clock:u (msec)       #    1.000 CPUs utilized            ( +-  0.84% )
                 0      context-switches:u        #    0.000 K/sec                  
                 0      cpu-migrations:u          #    0.000 K/sec                  
              2506      page-faults:u             #    0.911 K/sec                    ( +-  0.02% )

       2.751283658 seconds time elapsed                                          ( +-  0.85% )

Patch:

diff --git a/lib/buffer.js b/lib/buffer.js
index 876bdbe..d3e48a1 100644
--- a/lib/buffer.js
+++ b/lib/buffer.js
@@ -1241,13 +1241,8 @@ Buffer.prototype.writeInt32BE = function(value, offset, noAssert) {
 };


-Buffer.prototype.writeFloatLE = function writeFloatLE(val, offset, noAssert) {
-  val = +val;
-  offset = offset >>> 0;
-  if (!noAssert)
-    binding.writeFloatLE(this, val, offset);
-  else
-    binding.writeFloatLE(this, val, offset, true);
+Buffer.prototype.writeFloatLE = function writeFloatLE(val, offset) {
+  binding.writeFloatLE(this, +val, offset >>> 0);
   return offset + 4;
 };

diff --git a/src/node_buffer.cc b/src/node_buffer.cc
index 4baa8d9..027b963 100644
--- a/src/node_buffer.cc
+++ b/src/node_buffer.cc
@@ -817,12 +817,7 @@ template <typename T, enum Endianness endianness>
 void WriteFloatGeneric(const FunctionCallbackInfo<Value>& args) {
   Environment* env = Environment::GetCurrent(args);

-  bool should_assert = args.Length() < 4;
-
-  if (should_assert) {
-    THROW_AND_RETURN_UNLESS_BUFFER(env, args[0]);
-  }
-
+  THROW_AND_RETURN_UNLESS_BUFFER(env, args[0]);
   Local<Uint8Array> ts_obj = args[0].As<Uint8Array>();
   ArrayBuffer::Contents ts_obj_c = ts_obj->Buffer()->GetContents();
   const size_t ts_obj_offset = ts_obj->ByteOffset();
@@ -837,10 +832,8 @@ void WriteFloatGeneric(const FunctionCallbackInfo<Value>& args) {

   size_t memcpy_num = sizeof(T);

-  if (should_assert) {
-    CHECK_NOT_OOB(offset + memcpy_num >= memcpy_num);
-    CHECK_NOT_OOB(offset + memcpy_num <= ts_obj_length);
-  }
+  CHECK_NOT_OOB(offset + memcpy_num >= memcpy_num);
+  CHECK_NOT_OOB(offset + memcpy_num <= ts_obj_length);

   if (offset + memcpy_num > ts_obj_length)
     memcpy_num = ts_obj_length - offset;

[seishun]

This is just about writing floats and doubles, right? noAssert=false causes a significant performance drop for ints: #11245

[bnoordhuis]

I'm reopening this. I can't detect any measurable difference between noAssert=true and removing noAssert altogether, neither for integers nor floating-point numbers, so I think we should just get rid of the flag.

[seishun]

@bnoordhuis could you post your benchmark results for ints? I still get a significant difference with noAssert=false, as seen in the last chart in #11245 (comment).

[bnoordhuis]

Okay, I'm able to reproduce with benchmark/buffers/buffer-write.js when I crank up the number of iterations to 100 million. The default of 1 million iterations finishes too quickly on my machine for the optimizer to really kick in, with the result that the numbers are pretty close.

So, I see about a 3.5x performance improvement with noAssert=true. With some clever open coding I can bring that back to 2.7x.

Something is fishy though: the open-coded version is only 5% slower than a direct this[offset] = value or this[offset >>> 0] = value >>> 0 so why the performance gap with noAssert=true?

I also can't reproduce the gap outside the benchmark with e.g. the test case below; in fact, it's a little slower with the current implementation. I'm starting to think the benchmark is suspect.

Current version:

Buffer.prototype.writeUInt8 = function(value, offset, noAssert) {
  value = +value;
  offset = offset >>> 0;
  if (!noAssert)
    checkInt(this, value, offset, 1, 0xff, 0);
  this[offset] = value;
  return offset + 1;
};

Open-coded version:

function failUInt8(that, value, offset) {
  if (value >= 256)
    throw new TypeError('"value" argument is out of bounds');
  if (offset >= that.length)
    throw new RangeError('Index out of range');
}

Buffer.prototype.writeUInt8 = function(value, offset) {
  value = value >>> 0;
  offset = offset >>> 0;
  if (value >= 256 || offset >= this.length)
    failUInt8(this, value, offset);
  this[offset] = value;
  return offset + 1;
};

Direct version:

Buffer.prototype.writeUInt8 = function(value, offset) {
  this[offset] = value;
};

Simple test:

$ time -p out/Release/node -e '(function() {
  "use strict";
  for (var b = Buffer.alloc(1), i = 0; i < 1e8; ++i)
    b.writeUInt8(i & 255, 0, true);
})()'

[Trott]

Does TurboFan/Ignition change anything here in either direction as far as performance of our code that handles noAssert?