CASSANDRA-21085: Fix offset calculation in NativeLibrary.trySkipCache for files > 2GB

[cheeeee]

Jira Issue

• CASSANDRA-21085

Motivation

In NativeLibrary.trySkipCache(int fd, long offset, long len, String path), the chunking loop across 2GB boundaries decremented offset (offset -= sublen) instead of incrementing it (offset += sublen). On SSTables larger than 2GB, the second and subsequent iterations passed negative offsets to posix_fadvise(..., POSIX_FADV_DONTNEED).

This caused:

1. posix_fadvise returning -EINVAL (Invalid argument), flooding server logs with:
   WARN Failed trySkipCache on file: ... Error: Invalid argument
2. Failure to evict page cache pages beyond the first 2GB of large SSTables during compaction and sstableupgrade.

Additionally, when len == 0, trySkipCache(fd, offset, 0, path) was hardcoding 0 as the offset, ignoring the requested starting offset.

Changes

1. Fixed offset arithmetic in NativeLibrary.trySkipCache: offset += sublen;.
2. When len == 0, pass offset to correctly drop from offset to EOF.
3. Added unit test testSkipCacheLargeFile in NativeLibraryTest covering multi-chunk execution across the 2GB boundary.

Testing & Kernel eBPF Verification

• NativeLibraryTest: 3/3 tests pass.
• Linux Kernel eBPF Trace (tracepoint:syscalls:sys_enter_fadvise64):
  • Before fix:
    • Chunk 1: offset = 0, len = 2147483647
    • Chunk 2: offset = -2147483647, len = 1073741825 (negative offset, returns EINVAL).
  • After fix:
    • Chunk 1: offset = 0, len = 2147483647
    • Chunk 2: offset = 2147483647, len = 1073741825 (contiguous positive offset, returns 0).