Spikhalskiy opened a new issue, #54916:
URL: https://github.com/apache/arrow/issues/54916
### Describe the bug, including details regarding any error messages,
version, and platform.
`ipc::Message::Open(metadata, body)` verifies and reads the Flatbuffers
metadata in place. When the metadata buffer is not 8-byte aligned, every
Flatbuffers access is a misaligned load (UB, reported by
`-fsanitize=alignment`).
Flight hits this on every read: `DeserializeFlightData` slices `data_header`
zero-copy out of the gRPC message (`ReadBytesZeroCopy`), and
`FlightData::OpenMessage()` passes it straight to `Message::Open()`.
`message.cc` already has `MaybeAlignMetadata()` for exactly this ("to avoid
potential UBSAN issues from Flatbuffers"), but only `MessageDecoder` calls it.
Reproduction (on main, built with `-fsanitize=alignment`): copy a serialized
record batch message to an odd address and call `ipc::Message::Open()` on it.
UBSan reports misaligned accesses in `VerifyMessage` and `GetKeyValueMetadata`:
```
flatbuffers/buffer.h:186:20: runtime error: load of misaligned address
0x000756c38081 for type 'uoffset_t' (aka 'unsigned int'), which requires 4 byte
alignment
#3 in arrow::ipc::internal::VerifyMessage(...) metadata_internal.h:190
#4 in arrow::ipc::Message::MessageImpl::Open() message.cc:90
```
Suggested fix: call `MaybeAlignMetadata(&metadata_)` at the start of
`MessageImpl::Open()`. Already-aligned metadata is not copied. (Unaligned
bodies are a separate matter, covered by `ensure_alignment`, #32276.)
### Component(s)
C++, FlightRPC
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
