Describe the bug

Garnet 2.2.1 closes the client connection when a missing 1,024-byte key collides with a disk-backed key in the same index bucket and tag. The server logs System.NullReferenceException from ObjectAllocatorImpl.VerifyRecordFromDiskCallback at line 1260.

Steps to reproduce the bug

Start Garnet:

docker run --rm -p 127.0.0.1:6379:6379 \
  ghcr.io/microsoft/garnet:2.2.1 \
  --storage-tier true --index 1m --enable-debug-command yes

In another Bash shell:

prefix="synthetic:$(printf 'x%.0s' {1..982})"
stored="${prefix}87b7bf4eda4f44ccda64f1f5c40c5811"
missing="${prefix}fe6295bee3767698743c3edcf341e7cc"

redis-cli SET "$stored" value
redis-cli DEBUG FLUSHANDEVICT
redis-cli GET "$missing"

SET and DEBUG FLUSHANDEVICT succeed. GET reports:

Error: Server closed the connection

Expected behavior

GET should return a missing-key response without closing the connection.

Release version

2.2.1, official image ghcr.io/microsoft/garnet:2.2.1, digest sha256:05cb9eef041c9151b555eccbd62ce33b0661f8f9c02272cf0d604efb0ec2a143.

IDE

N/A

OS version

Linux x86_64, Docker.

Additional context

The base callback transfers ctx.record to ctx.diskLogRecord, and TransferFrom clears ctx.record. On an overflow-key mismatch, the derived callback then dereferences the cleared buffer.

It appears the previous address should instead be read from the owned record:

prevAddressToRead = ctx.diskLogRecord.Info.PreviousAddress;