Fix four LP64 defects and add Oracle 19 client discovery

[swgsais]

Four LP64 defects found while bringing up a 64-bit server from 64-bit-types, plus Oracle 19 client discovery. Each is present on 64-bit-types today. All five compile and none change 32-bit behaviour.

1. Packed maps desynchronise their own stream

AutoDeltaPackedMap<NetworkId,int>, <int,NetworkId> and <int,Unicode::String> write the command count as int32 but write the baseline count, and read both counts, as size_t.

	pack writes	unpack reads
ILP32	4 + 4 = 8	4 + 4 = 8
LP64	4 + 8 = 12	8 + 8 = 16

Under LP64 the reader consumes 4 bytes more than the writer produced. Little-endian makes this worse than a clean failure: the first get reads [count:4][zero:4] and yields the correct count, so the loop starts normally; the second get then swallows the remaining baseline bytes plus the first 4 bytes of command #1, and every subsequent key and value is misaligned. These run through SwgSnapshot, so the symptom is silent object-persistence corruption rather than a crash.

Fixed by using int32_t for both counts, matching the generic internal_pack/internal_unpack path and the already-converted PlayerQuestData specialization. This also keeps the serialized form byte-identical to a 32-bit server, so data written by either architecture stays readable by the other.

2. getCallStack() overruns every buffer by 2x

DebugHelp::getCallStack() passes the caller's uint32 buffer straight to backtrace(), which writes void* entries:

void DebugHelp::getCallStack(uint32 *callStack, int sizeOfCallStack)
{
    IGNORE_RETURN(backtrace(reinterpret_cast<void **>(callStack), sizeOfCallStack));
}

Call site	Buffer	backtrace writes	Overflow
Fatal.cpp	uint32[68] = 272 B	up to 544 B	272 B (stack)
CallStack	uint32[32] = 128 B	208 B	80 B (heap)
CallStackCollector	uint32[26] = 104 B	208 B	104 B (heap)

Fatal.cpp's formatMessage() backs every WARNING and FATAL, so on a 64-bit build any warning that captures a stack corrupts memory. We first saw this as a SIGSEGV inside cfree() with si_addr 0x8 shortly after an unrelated missing-asset warning — the warning was the corruption and the crash was the aftermath, which makes it very hard to attribute.

Fixed by capturing into a native void* array and widening into a uint64 buffer, which is correct under both ILP32 and LP64, and widening the storage, lookupAddress() and the printing to match. On win32 the same code truncated StackWalk's DWORD64 address to 32 bits, so symbols were unusable there; that is fixed too.

Also corrects CallStackCollector copying a single element instead of the whole call stack (sizeof(*newCallStack) without the element count), and CRCing only half the buffer.

3. QueryPerformanceCounter is not monotonic on linux

The linux shim samples gettimeofday(), i.e. CLOCK_REALTIME, which time synchronisation can slew or step backwards. Clock::update() subtracts successive samples to derive frame time, so a backwards step yields a negative elapsed time and AlterScheduler reports:

doAlterConclude() called with elapsed time [-9.2e-05] resulting in no forward time movement

Under a hypervisor that re-syncs guest time this fires continuously. Combined with defect 2, every one of those warnings was also smashing 272 bytes of stack.

Fixed by using clock_gettime(CLOCK_MONOTONIC), keeping microsecond units so the value still matches QueryPerformanceFrequency(), with a gettimeofday() fallback if the call fails. Every getFrameStartTimeMs() consumer compares values relatively, so nothing depends on the previous epoch-based origin. This one is architecture-independent and applies equally to a 32-bit build.

4. SIGSEGV handler dereferences a 32-bit frame offset

JavaLibrary::fatalHandler() reads the faulting address from a hardcoded 0x44 byte offset into its own frame:

// the address that generated the segfault is stored at an offset of 0x44 from
// the frame address
// @todo: find documentation on the offset to verify it is stable
void *crashAddress1 = *((void**)((static_cast<char *>(frameAddress)) + 0x44));

That constant describes the 32-bit x86 frame layout. On x86-64 it is both misaligned and meaningless, and dereferencing it from inside a SIGSEGV handler risks a recursive fault that loses the crash report entirely. The existing @todo already flagged the offset as unverified.

Restricted to __i386__, relying on the portable __builtin_return_address() probe already present immediately below it. Also zero-initialises the lib/file buffers, which the strstr() checks read whenever a lookup fails.

5. Oracle 19 Instant Client discovery

FindOracle.cmake knows nnz18 and nnz21 but not nnz19, and does not consider ORACLE_HOME when searching. Adds libnnz19/nnz19 and ${ORACLE_HOME} to the search paths. Purely additive.

Verification

Built from this branch against SWG-Source/swg-main:64-bit-types with its stock Ant/CMake path and bits = 64:

[echo] Architecture is x86_64
[echo] Compiling as 64-bit application.
BUILD SUCCESSFUL

20 binaries produced; LoginServer, TaskManager, CentralServer, ConnectionServer, SwgDatabaseServer and SwgGameServer all verified ELF 64-bit LSB shared object, x86-64.

Separately, these same fixes were exercised at runtime on a downstream integration branch: Oracle schema creation and template load completed with 0 ORA- errors and 60,077 object templates, the full server process set reached a healthy state, and client login and logout worked. That run carried additional downstream deltas, so it is supporting evidence for these fixes rather than a clean-room test of this branch alone.

Not exhaustively swept: format specifiers server-wide, time_t in persisted structures, alignment and packing assumptions, and third-party ABI.

[AconiteX]

just throwing in copilot review to see how it does but otherwise LGTM

2 nits on code style but need to just write a md for it

[Copilot]

Pull request overview

This PR addresses several LP64 (64-bit) correctness issues in serialization and debugging utilities, and extends CMake-based Oracle client discovery to support Oracle 19 Instant Client layouts. The changes aim to preserve 32-bit behavior and serialized compatibility while fixing 64-bit-only corruption/overflow risks.

Changes:

• Fix packed-map serialization/deserialization count type mismatches by using int32_t for command counts.
• Correct call stack capture/storage to be pointer-width safe (uint64), including Linux backtrace() handling and Win32 address preservation, and fix CallStackCollector buffer copy/CRC sizing.
• Make Linux QueryPerformanceCounter use CLOCK_MONOTONIC, harden the SIGSEGV handler to avoid invalid frame-offset dereference on x86-64, and add Oracle 19 (nnz19) + ${ORACLE_HOME} library discovery paths.

Reviewed changes

Copilot reviewed 14 out of 14 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
external/ours/library/unicodeArchive/src/shared/UnicodeAutoDeltaPackedMap.cpp	Uses int32_t for packed-map command counts to keep LP64 streams aligned.
engine/shared/library/sharedUtility/src/shared/NetworkIdAutoDeltaPackedMap.cpp	Same packed-map count fix for AutoDeltaPackedMap<int, NetworkId>.
engine/shared/library/sharedFoundation/src/shared/AutoDeltaNetworkIdPackedMap.h	Same packed-map count fix for AutoDeltaPackedMap<NetworkId, int>.
engine/shared/library/sharedFoundation/src/shared/Fatal.cpp	Widen call stack storage/printing to 64-bit addresses.
engine/shared/library/sharedFoundation/src/linux/PlatformGlue.cpp	Make QueryPerformanceCounter monotonic via clock_gettime(CLOCK_MONOTONIC).
engine/shared/library/sharedDebug/src/win32/DebugHelp.h	Widen call stack APIs to uint64 on Win32.
engine/shared/library/sharedDebug/src/win32/DebugHelp.cpp	Preserve full instruction addresses and widen stack storage/log formatting.
engine/shared/library/sharedDebug/src/shared/CallStackCollector.cpp	Fix callstack buffer type, memcpy sizing, and CRC sizing.
engine/shared/library/sharedDebug/src/shared/CallStack.h	Store instruction addresses as uint64 to match pointer width.
engine/shared/library/sharedDebug/src/shared/CallStack.cpp	Print and resolve widened call stack addresses.
engine/shared/library/sharedDebug/src/linux/DebugHelp.h	Widen call stack API to uint64 and document pointer-width requirement.
engine/shared/library/sharedDebug/src/linux/DebugHelp.cpp	Fix Linux backtrace() usage by capturing into void*[] then widening.
engine/server/library/serverScript/src/shared/JavaLibrary.cpp	Restrict frame-offset crash-address probe to __i386__ and zero-init buffers.
cmake/linux/FindOracle.cmake	Add nnz19 and ${ORACLE_HOME} to Oracle client library discovery.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[swgsais]

just throwing in copilot review to see how it does but otherwise LGTM

2 nits on code style but need to just write a md for it

Would love any feedback on the nits so I can improve style for future PRs - Thanks!

[AconiteX]

nit: this is functioning as a magic number, so could instead be at top of file, even though it is scoped and only relevant to the single function so it's 50/50 (I also don't love use of an untyped enum here since the compiler then has to make an inference and god knows what it may do in the future); alternatively, e.g., at top of file:

namespace
{
	constexpr int cs_maximumFrameCount = 256;
}

[AconiteX]

source generally (but admittedly inconsistently) uses allman style bracing and doesn't tend to use indentation block syntax/off-side rule, so code like this would properly be (but again doesn't technically matter and would surely start a not so fun debate)

if (sizeOfCallStack <= 0)
{
    return;
}