Fix XNN PreLU use after free (#21480) - #21480
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21480
Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
@JakeStevens has exported this pull request. If you are a Meta employee, you can view the originating Diff in D114141346. |
meta-cla
Bot
added
the
CLA Signed
This PR needs a
|
Summary: Most XNNPACK parameter / constant data is packed and XNNPACK takes ownership of it. This means that we can free the buffer after load. However, PreLU doesn't pack and thus expects the buffer to stay alive. There is a reproducible segfault when running PreLU on ET XNNPACK when running from a file instead of in-memory buffer data. This change fixes that by only freeing buffers for ops we know take ownership of the data. Practically speaking this should be a no-op for most cases like linear and conv, but it plays it safe for other constant tensors. Fixes pytorch#17559. Differential Revision: D114141346
meta-codesync
Bot
changed the title
d7a1b0c to
d5dcc83
Compare
digantdesai
left a comment
digantdesai
left a comment
There was a problem hiding this comment.
Review automatically exported from Phabricator review in Meta.
2 participants
Summary:
Most XNNPACK parameter / constant data is packed and XNNPACK takes ownership of it. This means that we can free the buffer after load. However, PreLU doesn't pack and thus expects the buffer to stay alive. There is a reproducible segfault when running PreLU on ET XNNPACK when running from a file instead of in-memory buffer data.
This change fixes that by only freeing buffers for ops we know take ownership of the data. Practically speaking this should be a no-op for most cases like linear and conv, but it plays it safe for other constant tensors.
Fixes #17559.
Differential Revision: D114141346