search

secondlife / viewer

🖥️ Second Life's official client
GNU Lesser General Public License v2.1
201 stars 51 forks source link

Sound encoding on upload is nondeterministic, preventing asset reuse #1896

Open canny[bot] opened 2 weeks ago

canny[bot] commented 2 weeks ago

Now, I'm making a presumption here that sounds, like other assets, should be checking for existing duplicates based on some kind of hashing and not create copies, much like it works for textures.

The encoding used in sound upload is nondeterministic and prevents this: it uses a random value for the Vorbis stream serial number, which results in binary differences as the metadata is different even if the audio data itself may not be.

Steps to reproduce:

  1. Prepare any SL-acceptable .wav file for upload.

  2. Upload the same file to Second Life twice.

  3. Copy and write down the asset UUID for the first file, then the second. Confirm they do not match.

  4. Browse to temp files directory and guesstimate which files are the .ogg files based on size and date. Rename the .tmp to .ogg to test audio playback.

  5. Perform a binary file compare: there will be a couple mismatches, often the same byte patterns, suggesting the metadata is the only thing that has changed.

Curtailed results from Windows FC:


fc /b 3b53c3a9-c301-b9d3-9318-47383d0b2488.ogg 79705d70-6488-60bc-5c6c-03e5882598f0.ogg

Comparing files 3b53c3a9-c301-b9d3-9318-47383d0b2488.ogg and 79705D70-6488-60BC-5C6C-03E5882598F0.OGG

0000000E: E3 97

0000000F: 4F 6D

...

00000DDC: E3 97

00000DDD: 4F 6D

...

00002F38: E3 97

00002F39: 4F 6D

...

The code in question is in llaudio\llvorbisencode.cpp, encode_vorbis_file() function: ogg_stream_init(&os, ll_rand());. Maybe it should be computing a hash of the .wav file to use as a serial instead, if asset duplication is not intentional for sounds.

https://secondlife.canny.io/admin/board/bug-reports/p/sound-encoding-on-upload-is-nondeterministic-preventing-asset-reuse

canny[bot] commented 2 weeks ago

This issue has been linked to a Canny post: Sound encoding on upload is nondeterministic, preventing asset reuse :tada:

AtlasLinden commented 2 weeks ago

See also: https://github.com/secondlife/jira-archive-internal/issues/66422

maestrolinden commented 2 weeks ago

I confirmed I can reproduce the issue with Second Life Release 7.1.8.9375512768 (64bit) on SLS 2024-06-11.9458617693. When I upload a basic .wav sound asset twice, the asset_id is different for the two sounds (the easiest way to verify this is with the 'copy asset uuid' context menu option).

It would seem that the changes for https://github.com/secondlife/jira-archive-internal/issues/66422 regressed at some point.