Hi All,
it turns out I jumped the gun on this. We have been hunting a buffer overrun bug and I mistakenly thought this was it. In fact our bug was in one of the media codec support libraries we use from a third-party.
I still think the code is confusing, and it looks to me as if it is broken on Mac OS and correct, but confusing, on win32. Anyone who wants to review this /please/ do. Here's the original code:
SoundPlugin>>primitiveSoundRecordSamplesInto: buf startingAt: startWordIndex "Record a buffer's worth of 16-bit sound samples." | bufSizeInBytes samplesRecorded | self primitive: 'primitiveSoundRecordSamples' parameters: #(WordArray SmallInteger ).
interpreterProxy failed ifFalse: [bufSizeInBytes := (interpreterProxy slotSizeOf: buf cPtrAsOop) * 4. interpreterProxy success: (startWordIndex >= 1 and: [startWordIndex - 1 * 2 < bufSizeInBytes])].
interpreterProxy failed ifFalse: [samplesRecorded := self cCode: 'snd_RecordSamplesIntoAtLength((int)buf, startWordIndex - 1, bufSizeInBytes)']. ^ samplesRecorded asPositiveIntegerObj
My objection in the above is that we do the bounds check above but leave it up to the support code snd_RecordSamplesIntoAtLength to derive the start pointer buf + ((startWordIndex - 1) * 2). IMO these two belong together.
Now in the Mac code there is no code to subtract (startWordIndex - 1) * 2 from bufSizeInBytes, and so without my fix there could be a buffer overrun:
int snd_RecordSamplesIntoAtLength(int buf, int startSliceIndex, int bufferSizeInBytes) { /* if data is available, copy as many sample slices as possible into the given buffer starting at the given slice index. do not write past the end of the buffer, which is buf + bufferSizeInBytes. return the number of slices (not bytes) copied. a slice is one 16-bit sample in mono or two 16-bit samples in stereo. */ int bytesPerSlice = (recordBuffer1.brokenOSXWasMono) ? 2 : ((recordBuffer1.stereo) ? 4 : 2); char *nextBuf = (char *) buf + (startSliceIndex * bytesPerSlice); char *bufEnd = (char *) buf + bufferSizeInBytes; char *src, *srcEnd; RecordBuffer recBuf = nil; int bytesCopied;
On win32 there is code to do the subtraction. e.g. looking at the Direct Sound side of the code:
int dx_snd_RecordSamplesIntoAtLength(int buf, int startSliceIndex, int bufferSizeInBytes) { /* if data is available, copy as many sample slices as possible into the given buffer starting at the given slice index. do not write past the end of the buffer, which is buf + bufferSizeInBytes. return the number of slices (not bytes) copied. a slice is one 16-bit sample in mono or two 16-bit samples in stereo. */ int bytesPerSlice = (waveInFormat.nChannels * 2); int bytesCopied; char *srcPtr, *srcPtr2, *dstPtr; HRESULT hRes; DWORD srcLen, srcLen2;
... /* Figure out how much data we want to copy (don't overflow either the source or destination buffers. */ bytesCopied = bufferSizeInBytes - (startSliceIndex * bytesPerSlice); if(bytesCopied > recBufferAvailable) bytesCopied = recBufferAvailable;
/* Lock the portion of the DirectSound buffer that we will copy data from. */ hRes = IDirectSoundCaptureBuffer_Lock(lpdRecBuffer, recBufferReadPosition, bytesCopied, (void*)&srcPtr, &srcLen, (void*)&srcPtr2, &srcLen2, 0);
So this is an example of a really confusing API. If instead the API were snd_RecordSamplesIntoAtLength(int buf, int bufferSizeInBytes) and the primitive read
SoundPlugin>>primitiveSoundRecordSamplesInto: buf startingAt: startWordIndex "Record a buffer's worth of 16-bit sound samples." | bufSizeInBytes samplesRecorded startByteIndex | self primitive: 'primitiveSoundRecordSamples' parameters: #(WordArray SmallInteger ).
interpreterProxy failed ifFalse: [bufSizeInBytes := (interpreterProxy slotSizeOf: buf cPtrAsOop) * 4. startByteIndex := startWordIndex - 1 * . interpreterProxy success: (startWordIndex >= 1 and: [startByteIndex < bufSizeInBytes])].
interpreterProxy failed ifFalse: [samplesRecorded := self cCode: 'snd_RecordSamplesIntoAtLength((int)buf + startByteIndex, bufSizeInBytes - startByteIndex']. ^ samplesRecorded asPositiveIntegerObj
there would be much less chance for (as the Fat Controller says) confusion and delay.
cheers Eliot On Mon, Mar 22, 2010 at 5:18 PM, Eliot Miranda eliot.miranda@gmail.comwrote:
Use of primitiveSoundRecordSamples with a start index other than 1 could corrupt the heap if enough samples were recorded.
Name: VMMaker-eem.163 Author: eem Time: 22 March 2010, 5:15:57 pm UUID: 2fa86dda-e18c-48cd-bcac-856504aba8dc Ancestors: VMMaker-ar.162
SoundPlugin: fix bounds bug in primitiveSoundRecordSamples. The call to snd_RecordSamplesIntoAtLength neglected to subtract the start index from the size of the buffer to be written into.
best Eliot