Allow one reusable proxy URL per ESPHome device (#125845)

* Allow one reusable URL per device

* Move process to convert info

* Stop previous process

* Change to 404

* Better error handling

homeassistant/components/esphome/ffmpeg_proxy.py

@@ -1,7 +1,6 @@
 """HTTP view that converts audio from a URL to a preferred format."""
 
 import asyncio
-from collections import defaultdict
 from dataclasses import dataclass, field
 from http import HTTPStatus
 import logging
@@ -28,7 +27,7 @@ def async_create_proxy_url(
     channels: int | None = None,
     width: int | None = None,
 ) -> str:
-    """Create a one-time use proxy URL that automatically converts the media."""
+    """Create a use proxy URL that automatically converts the media."""
     data: FFmpegProxyData = hass.data[DATA_FFMPEG_PROXY]
     return data.async_create_proxy_url(
         device_id, media_url, media_format, rate, channels, width
@@ -39,7 +38,10 @@ def async_create_proxy_url(
 class FFmpegConversionInfo:
     """Information for ffmpeg conversion."""
 
-    url: str
+    convert_id: str
+    """Unique id for media conversion."""
+
+    media_url: str
     """Source URL of media to convert."""
 
     media_format: str
@@ -54,18 +56,16 @@ class FFmpegConversionInfo:
     width: int | None
     """Target sample width in bytes (None to keep source width)."""
 
+    proc: asyncio.subprocess.Process | None = None
+    """Subprocess doing ffmpeg conversion."""
+
 
 @dataclass
 class FFmpegProxyData:
     """Data for ffmpeg proxy conversion."""
 
-    # device_id -> convert_id -> info
+    # device_id -> info
-    conversions: dict[str, dict[str, FFmpegConversionInfo]] = field(
+    conversions: dict[str, FFmpegConversionInfo] = field(default_factory=dict)
-        default_factory=lambda: defaultdict(dict)
-    )
-
-    # device_id -> process
-    processes: dict[str, asyncio.subprocess.Process] = field(default_factory=dict)
 
     def async_create_proxy_url(
         self,
@@ -77,9 +77,19 @@ class FFmpegProxyData:
         width: int | None,
     ) -> str:
         """Create a one-time use proxy URL that automatically converts the media."""
+        if (convert_info := self.conversions.pop(device_id, None)) is not None:
+            # Stop existing conversion before overwriting info
+            if (convert_info.proc is not None) and (
+                convert_info.proc.returncode is None
+            ):
+                _LOGGER.debug(
+                    "Stopping existing ffmpeg process for device: %s", device_id
+                )
+                convert_info.proc.kill()
+
         convert_id = secrets.token_urlsafe(16)
-        self.conversions[device_id][convert_id] = FFmpegConversionInfo(
+        self.conversions[device_id] = FFmpegConversionInfo(
-            media_url, media_format, rate, channels, width
+            convert_id, media_url, media_format, rate, channels, width
         )
         _LOGGER.debug("Media URL allowed by proxy: %s", media_url)
 
@@ -128,7 +138,7 @@ class FFmpegConvertResponse(web.StreamResponse):
 
         command_args = [
             "-i",
-            self.convert_info.url,
+            self.convert_info.media_url,
             "-f",
             self.convert_info.media_format,
         ]
@@ -156,12 +166,12 @@ class FFmpegConvertResponse(web.StreamResponse):
             stderr=asyncio.subprocess.PIPE,
         )
 
+        # Only one conversion process per device is allowed
+        self.convert_info.proc = proc
+
         assert proc.stdout is not None
         assert proc.stderr is not None
 
-        # Only one conversion process per device is allowed
-        self.proxy_data.processes[self.device_id] = proc
-
         try:
             # Pull audio chunks from ffmpeg and pass them to the HTTP client
             while (
@@ -173,22 +183,26 @@ class FFmpegConvertResponse(web.StreamResponse):
             ):
                 await writer.write(chunk)
                 await writer.drain()
+        except asyncio.CancelledError:
+            raise  # don't log error
+        except:
+            _LOGGER.exception("Unexpected error during ffmpeg conversion")
+
+            # Process did not exit successfully
+            stderr_text = ""
+            while line := await proc.stderr.readline():
+                stderr_text += line.decode()
+            _LOGGER.error("FFmpeg output: %s", stderr_text)
+
+            raise
         finally:
+            # Terminate hangs, so kill is used
+            if proc.returncode is None:
+                proc.kill()
+
             # Close connection
             await writer.write_eof()
 
-            # Terminate hangs, so kill is used
-            proc.kill()
-
-            if proc.returncode != 0:
-                # Process did not exit successfully
-                stderr_text = ""
-                while line := await proc.stderr.readline():
-                    stderr_text += line.decode()
-                _LOGGER.error("Error shutting down ffmpeg: %s", stderr_text)
-            else:
-                _LOGGER.debug("Conversion completed: %s", self.convert_info)
-
         return writer
 
 
@@ -208,27 +222,25 @@ class FFmpegProxyView(HomeAssistantView):
         self, request: web.Request, device_id: str, filename: str
     ) -> web.StreamResponse:
         """Start a get request."""
-
+        if (convert_info := self.proxy_data.conversions.get(device_id)) is None:
-        # {id}.mp3 -> id
-        convert_id = filename.rsplit(".")[0]
-
-        try:
-            convert_info = self.proxy_data.conversions[device_id].pop(convert_id)
-        except KeyError:
-            _LOGGER.error(
-                "Unrecognized convert id %s for device: %s", convert_id, device_id
-            )
             return web.Response(
-                body="Convert id not recognized", status=HTTPStatus.BAD_REQUEST
+                body="No proxy URL for device", status=HTTPStatus.NOT_FOUND
             )
 
-        # Stop any existing process
+        # {id}.mp3 -> id, mp3
-        proc = self.proxy_data.processes.pop(device_id, None)
+        convert_id, media_format = filename.rsplit(".")
-        if (proc is not None) and (proc.returncode is None):
-            _LOGGER.debug("Stopping existing ffmpeg process for device: %s", device_id)
 
-            # Terminate hangs, so kill is used
+        if (convert_info.convert_id != convert_id) or (
-            proc.kill()
+            convert_info.media_format != media_format
+        ):
+            return web.Response(body="Invalid proxy URL", status=HTTPStatus.BAD_REQUEST)
+
+        # Stop previous process if the URL is being reused.
+        # We could continue from where the previous connection left off, but
+        # there would be no media header.
+        if (convert_info.proc is not None) and (convert_info.proc.returncode is None):
+            convert_info.proc.kill()
+            convert_info.proc = None
 
         # Stream converted audio back to client
         return FFmpegConvertResponse(

tests/components/esphome/test_ffmpeg_proxy.py

@@ -61,7 +61,7 @@ async def test_proxy_view(
 
         # Should fail because we haven't allowed the URL yet
         req = await client.get(url)
-        assert req.status == HTTPStatus.BAD_REQUEST
+        assert req.status == HTTPStatus.NOT_FOUND
 
         # Allow the URL
         with patch(
@@ -75,6 +75,12 @@ async def test_proxy_view(
                 == url
             )
 
+        # Requesting the wrong media format should fail
+        wrong_url = f"/api/esphome/ffmpeg_proxy/{device_id}/{convert_id}.flac"
+        req = await client.get(wrong_url)
+        assert req.status == HTTPStatus.BAD_REQUEST
+
+        # Correct URL
         req = await client.get(url)
         assert req.status == HTTPStatus.OK
 
@@ -90,11 +96,11 @@ async def test_proxy_view(
         assert round(mp3_file.info.length, 0) == 1
 
 
-async def test_ffmpeg_error(
+async def test_ffmpeg_file_doesnt_exist(
     hass: HomeAssistant,
     hass_client: ClientSessionGenerator,
 ) -> None:
-    """Test proxy HTTP view with an ffmpeg error."""
+    """Test ffmpeg conversion with a file that doesn't exist."""
     device_id = "1234"
 
     await async_setup_component(hass, esphome.DOMAIN, {esphome.DOMAIN: {}})
@@ -109,3 +115,120 @@ async def test_ffmpeg_error(
     assert req.status == HTTPStatus.OK
     mp3_data = await req.content.read()
     assert not mp3_data
+
+
+async def test_lingering_process(
+    hass: HomeAssistant,
+    hass_client: ClientSessionGenerator,
+) -> None:
+    """Test that a new request stops the old ffmpeg process."""
+    device_id = "1234"
+
+    await async_setup_component(hass, esphome.DOMAIN, {esphome.DOMAIN: {}})
+    client = await hass_client()
+
+    with tempfile.NamedTemporaryFile(mode="wb+", suffix=".wav") as temp_file:
+        with wave.open(temp_file.name, "wb") as wav_file:
+            wav_file.setframerate(16000)
+            wav_file.setsampwidth(2)
+            wav_file.setnchannels(1)
+            wav_file.writeframes(bytes(16000 * 2))  # 1s
+
+        temp_file.seek(0)
+        wav_url = pathname2url(temp_file.name)
+        url1 = async_create_proxy_url(
+            hass,
+            device_id,
+            wav_url,
+            media_format="wav",
+            rate=22050,
+            channels=2,
+            width=2,
+        )
+
+        # First request will start ffmpeg
+        req1 = await client.get(url1)
+        assert req1.status == HTTPStatus.OK
+
+        # Only read part of the data
+        await req1.content.readexactly(100)
+
+        # Allow another URL
+        url2 = async_create_proxy_url(
+            hass,
+            device_id,
+            wav_url,
+            media_format="wav",
+            rate=22050,
+            channels=2,
+            width=2,
+        )
+
+        req2 = await client.get(url2)
+        assert req2.status == HTTPStatus.OK
+
+        wav_data = await req2.content.read()
+
+    # All of the data should be there because this is a new ffmpeg process
+    with io.BytesIO(wav_data) as wav_io, wave.open(wav_io, "rb") as wav_file:
+        # We can't use getnframes() here because the WAV header will be incorrect.
+        # WAV encoders usually go back and update the WAV header after all of
+        # the frames are written, but ffmpeg can't do that because we're
+        # streaming the data.
+        # So instead, we just read and count frames until we run out.
+        num_frames = 0
+        while chunk := wav_file.readframes(1024):
+            num_frames += len(chunk) // (2 * 2)  # 2 channels, 16-bit samples
+
+        assert num_frames == 22050  # 1s
+
+
+async def test_request_same_url_multiple_times(
+    hass: HomeAssistant,
+    hass_client: ClientSessionGenerator,
+) -> None:
+    """Test that the ffmpeg process is restarted if the same URL is requested multiple times."""
+    device_id = "1234"
+
+    await async_setup_component(hass, esphome.DOMAIN, {esphome.DOMAIN: {}})
+    client = await hass_client()
+
+    with tempfile.NamedTemporaryFile(mode="wb+", suffix=".wav") as temp_file:
+        with wave.open(temp_file.name, "wb") as wav_file:
+            wav_file.setframerate(16000)
+            wav_file.setsampwidth(2)
+            wav_file.setnchannels(1)
+            wav_file.writeframes(bytes(16000 * 2 * 10))  # 10s
+
+        temp_file.seek(0)
+        wav_url = pathname2url(temp_file.name)
+        url = async_create_proxy_url(
+            hass,
+            device_id,
+            wav_url,
+            media_format="wav",
+            rate=22050,
+            channels=2,
+            width=2,
+        )
+
+        # First request will start ffmpeg
+        req1 = await client.get(url)
+        assert req1.status == HTTPStatus.OK
+
+        # Only read part of the data
+        await req1.content.readexactly(100)
+
+        # Second request should restart ffmpeg
+        req2 = await client.get(url)
+        assert req2.status == HTTPStatus.OK
+
+        wav_data = await req2.content.read()
+
+    # All of the data should be there because this is a new ffmpeg process
+    with io.BytesIO(wav_data) as wav_io, wave.open(wav_io, "rb") as wav_file:
+        num_frames = 0
+        while chunk := wav_file.readframes(1024):
+            num_frames += len(chunk) // (2 * 2)  # 2 channels, 16-bit samples
+
+        assert num_frames == 22050 * 10  # 10s