windows-vs-linux-loader-architecture
windows-vs-linux-loader-architecture copied to clipboard
Side-by-side comparison of the Windows and Linux (GNU dl) Loaders
Windows vs Linux Loader Architecture
This repo was released in conjunction with an article on "What is Loader Lock?".
The loader is a vital part of any operating system. It's responsible for loading programs and libraries into a process' address space and is the first component to run code when a process starts. Starting a process includes tasks like initializing critical data structures, loading dependencies, and running the program.
The intentions of this document are to:
- Compare the Windows, Linux, and sometimes MacOS loaders
- Provide perspective on architectural and ecosystem differences as well as how they coincide with the loader
- Including experiments on how flexible or rigid they are with what can safely be done during module initialization (with the loader's internal locks held)
- Formally document how a modern Windows loader supports concurrency
- Current open source Windows implementations, including Wine and ReactOS, perform locking similar to the legacy Windows loader (they currently don't support the "parallel loading" ability present in a modern Windows loader)
- Educate, satisfy curiosity, and help fellow reverse engineers
- If you're looking for information on anything in particular, just give this document a
CTRL+F
- If you're looking for information on anything in particular, just give this document a
All of the information contained here covers Windows 10 22H2 and glibc 2.38. Some sections of this document also touch on the MacOS loader.
Highlights
Table of Contents
-
Windows vs Linux Loader Architecture
- Highlights
- Table of Contents
- Module Information Data Structures
- Locks
- Atomic state
- State
- Concurrency Experiments
-
The Root of
DllMain
Problems - Comparing OS Library Loading Locations
-
DllMain
vs Global Constructors and Destructors - Lazy Loading and Lazy Linking OS Comparison
-
LDR_DDAG_NODE.State
Analysis -
Procedure/Symbol Lookup Comparison (Windows
GetProcAddress
vs POSIXdlsym
GNU Implementation) -
How Does
GetProcAddress
/dlsym
Handle Concurrent Library Unload? - Lazy Linking Synchronization
- GNU Loader Lock Hierarchy
-
Loader Enclaves
-
The Windows
LdrpObtainLockedEnclave
Function
-
The Windows
-
Modern Windows Loader Synchronization: State > Locks
- Case Study
- Windows Loader Initialization Locking Requirements
-
ELF Flat Symbol Namespace (GNU Namespaces and
STB_GNU_UNIQUE
) - What's the Difference Between Dynamic Linking, Static Linking, Dynamic Loading, Static Loading?
-
LoadLibrary
vsdlopen
Return Type -
Investigating COM Server Deadlock from
DllMain
-
On Making COM from
DllMain
Safe- Avoiding ABBA Deadlock
- Other Deadlock Possibilities
- Conclusion
- Case Study on ABBA Deadlock Between the Loader Lock and GDI+ Lock
-
LdrpDrainWorkQueue
andLdrpProcessWork
-
LdrpLoadCompleteEvent
andLdrpWorkCompleteEvent
- When a Process Would Rather Terminate Then Wait on a Critical Section
- ABBA Deadlock
- ABA Problem
- Dining Philosophers Problem
-
Reverse Engineered Windows Loader Functions
-
LdrpDrainWorkQueue
-
LdrpDecrementModuleLoadCountEx
-
-
Analysis Commands
-
LDR_DATA_TABLE_ENTRY
Analysis -
LDR_DDAG_NODE
Analysis -
Trace
TEB.SameTebFlags
- Searching Assembly for Structure Offsets
- Monitor a Critical Section Lock
- Track Loader Events
- Disable Loader Worker Threads
-
List All
LdrpCriticalLoaderFunctions
- Loader Debug Logging
- Get TCB and Set GSCOPE Watchpoint (GNU Loader)
-
- Compiling & Running
- Debugging Notes
- Big thanks to Microsoft for successfully nerd sniping me!
Module Information Data Structures
A Windows module list is a circular doubly linked list of type LDR_DATA_TABLE_ENTRY
. The loader maintains multiple LDR_DATA_TABLE_ENTRY
lists containing the same entries but in different link orders. These lists include InLoadOrderModuleList
, InMemoryOrderModuleList
, and InInitializationOrderModuleList
. These list heads are stored in ntdll!PEB_LDR_DATA
. Each LDR_DATA_TABLE_ENTRY
structure houses a LIST_ENTRY
structure (containing both Flink
and Blink
pointers) for each module list. See the "What is Loader Lock?" article for information on all the Windows loader data structures.
The Linux (glibc) module list is a non-circular doubly linked list of type link_map
. link_map
contains both the next
and prev
pointers used to link the module information structures together into a list.
Locks
OS-level synchronization mechanisms of all kinds, including thread synchronization mechanisms (i.e. Windows critical section or POSIX mutex), exclusive/write and shared/read locks, and other synchronization objects (Windows). OS-level, meaning these locks may make a system call (a syscall
instruction on x86-64) to perform a non-busy wait if the synchronization mechanism is owned/contended/waiting.
Some lock varieties are a mix of an OS lock and a spinlock (i.e. busy loop). For example, both a Windows critical section and a GNU mutex (not POSIX; this is a GNU extension) support specifying a spin count. A spin count is a potential performance optimization for avoiding expensive context switches between user mode and kernel mode.
Most OS locks use atomic flags to implement locking primitives when there's no contention (e.g. implemented with the lock cmpxchg
instruction on x86). An exception to this would be Win32 event synchronization objects, which rely entirely on system calls (i.e. NtSetEvent
and NtResetEvent
functions are just stubs containing a syscall
instruction).
Windows LdrpModuleDatatableLock
- Performs full blocking (exclusive/write) access to its respective module information data structures
- This includes two linked lists (
InLoadOrderModuleList
andInMemoryOrderModuleList
), a hash table, two red-black trees, and two directed acyclic graphs-
Note: The DAGs only require
LdrpModuleDatatableLock
protection to ensure synchronization/consistency with other module information data structures during write operations (e.g. adding/deleting nodes)- In addition to acquiring the
LdrpModuleDatatableLock
lock, safely modifying the DAGs requires that your thread be the loader owner (LoadOwner
inTEB.SameTebFlags
and incrementingntdll!LdrpWorkInProgress
) - Read operations (e.g. walking between nodes) are naturally protected depending on where the load owner is in the loading process (see:
LDR_DDAG_NODE.State
Analysis)
- In addition to acquiring the
-
Note: The DAGs only require
- This lock also protects some structure members contained within these data structures (e.g. the
LDR_DDAG_NODE.LoadCount
reference count)
- This includes two linked lists (
- Windows shortly acquires
LdrpModuleDatatableLock
many times (I counted 17 exactly) for everyLoadLibrary
(tested with a full path to an empty DLL; a completely fresh Visual Studio DLL project)- Acquiring this lock so many times could create contention on
LdrpModuleDatatableLock
, even if the lock is only held for short sprints - Monitor changes to
LdrpModuleDatatableLock
by setting a watchpoint:ba w8 ntdll!LdrpModuleDatatableLock
(ensure you don't count unlocks)-
Note: There are a few occurrences of this lock's data being modified directly for unlocking instead of calling
RtlReleaseSRWLockExclusive
(this is likely done as a performance optimization on hot paths)
-
Note: There are a few occurrences of this lock's data being modified directly for unlocking instead of calling
- Acquiring this lock so many times could create contention on
- Implemented as a slim read/write (SRW) lock, typically acquired exclusively
Linux (GNU loader) _rtld_global._dl_load_write_lock
- Performs full blocking (exclusive/write) access to its respective module data structures
- On Linux, this is only a linked list (the
link_map
list) - Linux shortly acquires
_dl_load_write_lock
once on everydlopen
from the_dl_add_to_namespace_list
internal function (see GDB log for evidence of this)- Other functions that acquire
_dl_load_write_lock
(not called duringdlopen
) include thedl_iterate_phdr
function which is for iterating over the module list- According to glibc source code, this lock is acquired to: "keep
__dl_iterate_phdr
from inspecting the list of loaded objects while an object is added to or removed from that list." - On Windows, acquiring the equivalent
LdrpModuleDatatableLock
is required to iterate the module list safely (e.g. when calling theLdrpFindLoadedDllByNameLockHeld
function)
- According to glibc source code, this lock is acquired to: "keep
- Other functions that acquire
Windows LdrpLoaderLock
- Blocks concurrent module initialization/deinitialization and protects the
InInitializationOrderModuleList
linked list- Safely running a module's
DLL_PROCESS_ATTACH
requires holding loader lock- At process exit,
RtlExitUserProcess
acquires loader lock before running the library deinitialization case (DLL_PROCESS_DETACH
) ofDllMain
- At process exit,
- This protection includes DLL thread initialization/deinitialization during
DLL_THREAD_ATTACH
/DLL_THREAD_DETACH
- Safely running a module's
- On the modern Windows loader at
DLL_PROCESS_ATTACH
, the loader lock remains locked as each full dependency chain of a loading DLL, including the loading DLL itself, is initialized (i.e. the loader locks loader lock once before startingLdrpInitializeGraphRecurse
and unlocks after returning from that function) - Loader lock protects against concurrent module initialization because the initialization routine of one module may depend on another module already being initialized, hence why library initialization must be serialized (i.e. done in series; not in parallel)
- Implemented as a critical section
Linux (GNU loader) _dl_load_lock
- This lock is acquired right at the start of a
_dl_open
,dlclose
, and other loader functions-
dlopen
eventually calls_dlopen
after some preparation work (which shows in the call stack) like setting up an exception handler, at which point the loader is committed to doing some loader work
-
- According to glibc source code, this lock's purpose is to: "Protect against concurrent loads and unloads."
- This protection includes concurrent module initialization similar to how a modern Windows
ntdll!LdrpLoaderLock
does - For example,
_dl_load_lock
protects a concurrentdlclose
from running a library's global destructors before that library's global constructors has finished running
- This protection includes concurrent module initialization similar to how a modern Windows
-
_dl_load_lock
is at the top of the loader's lock hierarchy - Since
_dl_load_lock
protcts the entire library loading/unloading process from beginning to end, the closest modern Windows loader equivalent synchronization mechanism would be theLdrpLoadCompleteEvent
loader event (this is when is the soon-to-be load owner thread increments ntdll!LdrpWorkInProgress from 0 to 1) combined with holding thentdll!LdrpLoaderLock
lock
Linux (GNU loader) _ns_unique_sym_table.lock
- This is a per-namespace lock for protecting that namespace's unique (
GNU_STB_UNIQUE
) symbol hash table -
GNU_STB_UNIQUE
symbols are a type of symbol that a module can expose; they are considered a misfeature of the GNU loader - As standardized by the ELF executable format, the GNU loader uses a per-module statically allocated (at compile time) symbol table for locating symbols within a module (
.so
shared object file); however, the_ns_unique_sym_table.lock
lock protects a separate dynamically allocated hash table specially forGNU_STB_UNIQUE
symbols- See the Procedure/Symbol Lookup Comparison (Windows
GetProcAddress
vs POSIXdlsym
GNU Implementation) section for more info on how the GNU loader typically locates symbols - In Windows terminology, the closest approximation to a symbol would be a DLL's function exports (there's no mention of the word "export" in the
objdump
manual)
- See the Procedure/Symbol Lookup Comparison (Windows
- You can use
objdump -t
to dump all the symbols, including unique symbols, of an ELF file-
objdump -t
displays unique global symbols (as the manual references them) with theu
flag character
-
- Internally, the call chain for looking up a
GNU_STB_UNIQUE
symbol starting withdlsym
goesdlsym
➜dl_lookup_symbol_x
➜do_lookup_x
➜do_lookup_unique
where finally,_ns_unique_sym_table.lock
is acquired - For more info on
GNU_STB_UNIQUE
symbols, see theGNU_STB_UNIQUE
section
This is where the loader's locks end for the GNU loader on Linux. Other than that, there's only a lock specific to thread-local storage (TLS) if your module uses that, as well as global scope (GSCOPE) locking.
However, Windows has more synchronization mechanisms that control the loader, including:
- Loader events (Win32 event objects), including:
-
ntdll!LdrpInitCompleteEvent
- This event being set indicates loader initialization is complete
- Loader initialization includes process initialization
- This event is only set (
NtSetEvent
) by theLdrpProcessInitializationComplete
function soon afterLdrpInitializeProcess
returns, at which point it's never set/unset again
- Thread startup waits on this
- This event is not an auto-reset event is and created in the nonsignaled (i.e. waiting) state
- The
LdrpInitialize
(_LdrpInitialize
) function creates this event- This event is created before loader initialization begins (early at process creation)
- This event being set indicates loader initialization is complete
-
ntdll!LdrpLoadCompleteEvent
- This event being set indicates an entire load/unload has completed.
- This event is set (
NtSetEvent
) in theLdrpDropLastInProgressCount
function before relinquishing control as the load owner (LoadOwner
flag inTEB.SameTebFlags
)
- This event is set (
- Thread startup waits on this
- This event is an auto-reset event and is created in the nonsignaled (i.e. waiting) state
- Created by
LdrpInitParallelLoadingSupport
callingLdrpCreateLoaderEvents
- This event being set indicates an entire load/unload has completed.
-
LdrpWorkCompleteEvent
- This event being set indicates the loader has completed processing (i.e. mapping and snapping) the entire work queue and all loader work threads are finished processing work
- This event is set (
NtSetEvent
) immediately before theLdrpProcessWork
function returns if the work queue is now empty or all current loader worker threads are finished processing work
- This event is set (
- This event is an auto-reset event and is created in the nonsignaled (i.e. waiting) state
- Created by
LdrpInitParallelLoadingSupport
callingLdrpCreateLoaderEvents
- This event being set indicates the loader has completed processing (i.e. mapping and snapping) the entire work queue and all loader work threads are finished processing work
- All of these events are created by NTDLL at loader/process startup using
NtCreateEvent
- For in-depth information on the latter two events, see the section on
LdrpLoadCompleteEvent
andLdrpWorkCompleteEvent
-
-
ntdll!LdrpWorkQueueLock
- Implemented as a critical section
- Used in
LdrpDrainWorkQueue
so only one thread can access theLdrpWorkQueue
queue at a time
-
ntdll!LdrpDllNotificationLock
- This is a critical section; it's typically locked (
RtlEnterCriticalSection
) in thentdll!LdrpSendPostSnapNotifications
function (called byntdll!PrepareModuleForExecution
➜ntdll!LdrpNotifyLoadOfGraph
; this happens in advance ofntdll!PrepareModuleForExecution
callingLdrpInitializeGraphRecurse
to do module initialization - The
LdrpSendPostSnapNotifications
function acquires this lock to ensure consistency with other functions that run DLL notification callbacks before fetching app compatibility data (SbUpdateSwitchContextBasedOnDll
function) and potentially running post snap DLL notification callbacks (look forcall qword ptr [ntdll!__guard_dispatch_icall_fptr (<ADDRESS>)]
in disassembly)- Typically,
LdrpSendPostSnapNotifications
doesn't run any callback functions - Other functions that directly send DLL notifications:
LdrpSendShimEngineInitialNotifications
(called byLdrpLoadShimEngine
andLdrpDynamicShimModule
)
- Typically,
- The
ntdll!LdrpSendNotifications
function (called byLdrpSendPostSnapNotifications
) function recursively acquires this lock to safely access theLdrpDllNotificationList
list so it can call the callback functions stored inside- By default, the
LdrpDllNotificationList
list is empty (so theLdrpSendDllNotifications
function doesn't send any callbacks) - Notifications callbacks are registered with
LdrRegisterDllNotification
and are then sent withLdrpSendDllNotifications
(it runs the callback function) - Functions calling DLL notification callbacks must hold the
LdrpDllNotificationLock
lock during callback execution. This is similar to how executing module initialization/deinitialization code (DllMain
) requires holding theLdrpLoaderLock
lock. - Other functions that may call
LdrpSendDllNotifications
include:LdrpUnloadNode
andLdrpCorProcessImports
(may be called byLdrpMapDllWithSectionHandle
) -
LdrpSendDllNotifications
takes a pointer to aLDR_DATA_TABLE_ENTRY
structure as its first argument - In ReactOS,
LdrpSendDllNotifications
is referenced inLdrUnloadDll
sending a shutdown notification with aFIXME
comment (not implemented yet)://LdrpSendDllNotifications(CurrentEntry, 2, LdrpShutdownInProgress);
- ReactOS code analysis: If
LdrpShutdownInProgress
is set,LdrUnloadDll
skips deinitialization and releases loader lock early, presumably so shutdown happens faster (don't need to deinitialize a module if the whole process is about to not exist) -
LdrpShutdownInProgress
is referenced inPEB_LDR_DATA
- ReactOS code analysis: If
- By default, the
- Reading loader disassembly, you may see quite a few places where loader functions check the
LdrpDllNotificationLock
lock like so:RtlIsCriticalSectionLockedByThread(&LdrpDllNotificationLock)
- For instance, in the
ntdll!LdrpAllocateModuleEntry
,ntdll!LdrGetProcedureAddressForCaller
,ntdll!LdrpPrepareModuleForExecution
, andntdll!LdrpMapDllWithSectionHandle
functions - These checks detect if the current thread is executing a DLL notification callback and then implement special logic for that edge case (for this reason, they can generally be ignored)
- The exact differing logic varies widely depending on the function checking
- By putting Google Chrome under WinDbg, I found an instance where the loader ran a post-snap DLL notification callback. The callback ran the
apphelp!SE_DllLoaded
function. - Lastly, ignore any
AVrf*
prefixed functions as this only ties in with debugging functionality provided by Application Verifier
- For instance, in the
- This is a critical section; it's typically locked (
-
LDR_DATA_TABLE_ENTRY.Lock
- Starting with Windows 10, each
LDR_DATA_TABLE_ENTRY
has aPVOID
Lock
member (it replaced aSpare
slot) -
LdrpWriteBackProtectedDelayLoad
uses this per-node lock to implement some level of protection during delay loading- This function calls
NtProtectVirtualMemory
, so it's likely something to do with the setting memory protection on a module's Import Address Table (IAT)
- This function calls
- Starting with Windows 10, each
-
PEB.TppWorkerpListLock
- This SRW lock (typically acquired exclusively) exists in the PEB to control access to the member immediately below it, which is the
TppWorkerpList
doubly linked list - This list keeps track of all the threads belonging to any thread pool in the process
- These threads show up as
ntdll!TppWorkerThread
threads in WinDbg - There's a list head, after which each
LIST_ENTRY
points into the stack memory a thread within thentdll!TppWorkerThread
function's stack frame- The
TpInitializePackage
function (called byLdrpInitializeProcess
) initializes the list head, then the mainntdll!TppWorkerThread
function of each new thread belonging to a thread pool adds itself to the list
- The
- The threads in this list include threads belonging to the loader worker thread pool (
LoaderWorker
inTEB.SameTebFlags
)
- These threads show up as
- This is a bit out of scope since it relates to thread pool internals not loader internals (however, the loader relies on thread pool internals to implement parallelism for loader workers)
- This SRW lock (typically acquired exclusively) exists in the PEB to control access to the member immediately below it, which is the
- Searching symbols reveals more of the loader's locks:
x ntdll!Ldr*Lock
-
LdrpDllDirectoryLock
,LdrpTlsLock
(this is an SRW lock typically acquired in the shared mode),LdrpEnclaveListLock
,LdrpPathLock
,LdrpInvertedFunctionTableSRWLock
,LdrpVehLock
,LdrpForkActiveLock
,LdrpCODScenarioLock
,LdrpMrdataLock
, andLdrpVchLock
- A lot of these locks seem to be for controlling access to list data structures
-
Atomic state
An atomic state value is modified using a single assembly instruction. On an SMP operating system (e.g. Windows and very likely your build of Linux; check with uname -a
command) with a multi-core processor, this must include a lock
prefix (on x86) so the processor knows to synchronize that memory access across CPU cores. The x86 ISA requires that a single read/write operation is atomic by default. It's only when atomically combining multiple reads or writes (e.g. to atomically increment a reference count) that the lock
prefix is necessary. Only a few key pieces of the Windows loader atomic state I came across are listed here.
-
ntdll!LdrpProcessInitialized
- This value is modified atomically with a
lock cmpxchg
instruction - As the name implies, it indicates whether process initialization has been completed (
LdrpInitializeProcess
) - This is an enum ranging from zero to three; here are the state transitions:
- NTDLL compile time initialization starts
LdrpProcessInitialized
with a value of zero (process is uninitialized) -
LdrpInitialize
incrementsLdrpProcessInitialized
to one zero early on (initialization event created)- If the process is still initializing, newly spawned threads jump to calling
NtWaitForSingleObject
, waiting on theLdrpInitCompleteEvent
loader event before proceeding - Before the loader calls
NtCreateEvent
to createLdrpInitCompleteEvent
at process startup, spawning a thread into a process causes it to useLdrpProcessInitialized
as a spinlock (i.e. a busy loop) - For example, if a remote process calls
CreateRemoteThread
and the thread spawns before the creation ofLdrpInitCompleteEvent
(an unlikely but possible race condition)
- If the process is still initializing, newly spawned threads jump to calling
-
LdrpProcessInitializationComplete
incrementsLdrpProcessInitialized
to two (process initialization is done)- This happens immediately before setting the
LdrpInitCompleteEvent
loader event so other threads can run - After
LdrpProcessInitializationComplete
returns,NtTestAlert
processes the asynchronous procedure call (APC) queue, and finally,NtContinue
yields code execution of the current thread toKERNEL32!BaseThreadInitThunk
, which eventually runs our program'smain
function
- This happens immediately before setting the
- NTDLL compile time initialization starts
- This value is modified atomically with a
-
LDR_DATA_TABLE_ENTRY.ReferenceCount
- This is a reference count for
LDR_DATA_TABLE_ENTRY
structures - On load/unload, this state is modified by
lock
inc
/dec
instructions, respectively (although, on the initial allocation of aLDR_DATA_TABLE_ENTRY
before linking it into any shared data structures, of course, no locking is necessary) - The
LdrpDereferenceModule
function atomically decrementsLDR_DATA_TABLE_ENTRY.ReferenceCount
by passing0xffffffff
to alock xadd
assembly instruction, causing the 32-bit integer to overflow to one less than what it was (x86 assembly doesn't have anxsub
instruction so this is the standard way of doing this)- Note that the
xadd
instruction isn't the same as theadd
instruction because the former also atomically exchanges (hence the "x") the previous memory value into the source operand- This is at the assembly level; in code, Microsoft is likely using the
InterlockedExchangeSubtract
macro to do this
- This is at the assembly level; in code, Microsoft is likely using the
- The
LdrpDereferenceModule
function tests (among other things) if the previous memory value was 1 (meaning post-decrement, it's now zero; i.e. nobody is referencing thisLDR_DATA_TABLE_ENTRY
anymore) and takes that as a cue to unmap the entire module from memory (calling theLdrpUnmapModule
function, deallocating memory structures, etc.)
- Note that the
- This is a reference count for
-
ntdll!LdrpLoaderLockAcquisitionCount
- This value is modified atomically with the
lock xadd
prefixed instructions - It was only ever used as part of cookie generation in the
LdrLockLoaderLock
function- On both older/modern loaders,
LdrLockLoaderLock
adds toLdrpLoaderLockAcquisitionCount
every time it acquires the loader lock (it's never decremented) - In a legacy (Windows Server 2003) Windows loader, the
LdrLockLoaderLock
function (an NTDLL export) was often used internally by NTDLL even inLdr
prefixed functions. However, in a modern Windows loader, it's mostly phased out in favor of theLdrpAcquireLoaderLock
function - In a modern loader, the only places where I see
LdrLockLoaderLock
called are from non-Ldr
prefixed functions, specifically:TppWorkCallbackPrologRelease
andTppIopExecuteCallback
(thread pool internals, still in NTDLL)
- On both older/modern loaders,
- This value is modified atomically with the
State
A state value may either be shared state (also known as global state) or local state (i.e. whether separate threads may access the state). Shared state has access to it protected by one of the aforementioned locks. Local state doesn't require protection. Only a few key pieces of Windows loader state I came across are listed here.
-
LDR_DDAG_NODE.State
(pointed to byLDR_DATA_TABLE_ENTRY.DdagNode
)- Each module has a
LDR_DDAG_NODE
structure with aState
member containing 15 possible states -5 through 9 -
LDR_DDAG_NODE.State
tracks a module's entire lifetime from allocating module information data structures (LdrpAllocatePlaceHolder
) and loading to unload and subsequent deallocation of its module information structures- In my opinion, this makes the combined
LDR_DDAG_NODE.State
values of all modules the most important piece of loader state
- In my opinion, this makes the combined
- Performing each state change may necessitate acquiring the
LdrpModuleDatatableLock
lock to ensure consistency between module information data structures- The specific state changes requiring
LdrpModuleDatatableLock
protection (i.e. consistency between all module information data structures) are documented in the link below
- The specific state changes requiring
- Please see the
LDR_DDAG_NODE.State
Analysis for more info
- Each module has a
-
ntdll!LdrpWorkInProgress
- This reference counter is a key piece of loader state (zero meaning work is not in progress and up from that meaning work is in progress)
- It's not modified atomically with
lock
prefixed instructions
- It's not modified atomically with
- Acquiring the
LdrpWorkQueueLock
lock is a requirement for safely modifying theLdrpWorkInProgress
state andLdrpWorkQueue
linked list- I verified this by setting a watchpoint on
LdrpWorkInProgress
and noticing thatLdrpWorkQueueLock
is always locked while checking/modifying theLdrpWorkInProgress
state (also, I searched disassembly code)- The
LdrpDropLastInProgressCount
function makes this clear because it briefly acquiresLdrpWorkQueueLock
around only the single assembly instruction that setsLdrpWorkInProgress
to zero
- The
- I verified this by setting a watchpoint on
- Please see the
LDR_DDAG_NODE.State
Analysis for more info
- This reference counter is a key piece of loader state (zero meaning work is not in progress and up from that meaning work is in progress)
-
ntdll!LdrInitState
- This value is not modified atomically with
lock
prefixed instructions- Loader initialization is a procedural process only occurring once and on one thread, so this value doesn't require protection
- In ReactOS code, the equivalent value is
LdrpInLdrInit
, which the code declares as aBOOLEAN
value - In a modern Windows loader, this a 32-bit integer (likely an
enum
) ranging from zero to three; here are the state transitions:-
LdrpInitialize
initializesLdrInitState
to zero (loader is uninitialized) -
LdrpInitializeProcess
callsLdrpEnableParallelLoading
and immediately after setsLdrInitState
to one (import loading in progress; thank you, Windows Internals book authors) -
LdrpInitializeProcess
callsLdrpPrepareModuleForExecution
on the EXE (first argument isLDR_DATA_TABLE_ENTRY
of EXE) and immediately after setsLdrInitState
to two (imports have loaded)-
Note:
LdrpPrepareModuleForExecution
does not callLdrpInitializeGraphRecurse
in this scenario because our EXE isn't a DLL with aDllMain
initialization function but still calls other functions likeLdrpNotifyLoadOfGraph
andLdrpDynamicShimModule
-
Note:
-
LdrpInitialize
(LdrpInitializeProcess
returned), shortly before callingLdrpProcessInitializationComplete
, setsLdrInitState
to three (loader initialization is done)
-
- This value is not modified atomically with
-
LDR_DDAG_NODE.LoadCount
- This is the reference count for a
LDR_DDAG_NODE
structure; safely modifying it requires acquiring theLdrpModuleDataTableLock
lock
- This is the reference count for a
-
TEB.WaitingOnLoaderLock
is thread-specific data set when a thread is waiting for loader lock-
RtlpWaitOnCriticalSection
(RtlEnterCriticalSection
callsRtlpEnterCriticalSectionContended
, which calls this function) checks if the contended critical section isLdrpLoaderLock
and if so, setsTEB.WaitingOnLoaderLock
equal to one- This branch condition runs every time any contended critical section gets waited on, which is interesting (monolithic much?)
-
RtlpNotOwnerCriticalSection
(called fromRtlLeaveCriticalSection
) also checksLdrpLoaderLock
(and some other info fromPEB_LDR_DATA
) for special handling- However, this is only for error handling and debugging because a thread that doesn't own a critical section should have never attempted to leave it in the first place
-
- Flags in
TEB.SameTebFlags
, including:LoadOwner
,LoaderWorker
, andSkipLoaderInit
- All of these were introduced in Windows 10 (
SkipLoaderInit
only in 1703 and later) -
LoadOwner
(flag mask0x1000
) is state that a thread uses to inform itself that it's the one responsible for completing the work in progress (ntdll!LdrpWorkInProgress
)- The
LdrpDrainWorkQueue
function sets theLoadOwner
flag on the current thread immediately after settingntdll!LdrpWorkInProgress
to1
, thus directly connecting these two pieces of state - The
LdrpDropLastInProgressCount
function unsets this flag along withntdll!LdrpWorkInProgress
- Any thread doing loader work (e.g.
LoadLibrary
) will temporarily receive this TEB flag - This state is local to the thread (in the TEB), so it doesn't require the protection of a lock
- The
-
LoaderWorker
(flag mask0x2000
) identifies loader worker threads- These show up as
ntdll!TppWorkerThread
in WinDbg - On thread creation,
LdrpInitialize
checks if the thread is a loader worker and, if so, handles it specially (for more info, see the "What is Loader Lock?" article) - Flag can be set on a new thread using
NtCreateThreadEx
- These show up as
-
SkipLoaderInit
(flag mask0x4000
) tells the spawning thread to skip all loader initialization- As per the Windows Internals book and confirmed by myself
- In
LdrpInitialize
IDA decompilation, you can seeSameTebFlags
being tested for0x4000
, and if present, loader initialization is completely skipped (_LdrpInitialize
is never called)
- In
- This could be useful for creating new threads without being blocked by loader events
- Threads with the
SkipLoaderInit
flag can be created usingNtCreateThreadEx
- As per the Windows Internals book and confirmed by myself
- All of these were introduced in Windows 10 (
-
ntdll!LdrpMapAndSnapWork
- An undocumented, global structure that loader worker (
LoaderWorker
flag inTEB.SameTebFlags
) threads read from to get mapping and snapping work - The first member of this undocumented structure is an atomic reference counter that gets incremented whenever work is enqueued (
ntdll!LdrpQueueWork
function) to the global work queue (ntdll!LdrpWorkQueue
linked list) and decremented whenever a loader worker thread consumes work - The undocumented structure is incremented by the
ntdll!TppWorkPost
function (called byntdll!LdrpQueueWork
) and decremented by thentdll!TppIopCallbackEpilog
function (called by thread pool internals in a loader worker thread), which means this undocumented structure belongs to thread pool internals and so is a bit out of scope here
- An undocumented, global structure that loader worker (
Indeed, the Windows loader is an intricate and monolithic state machine. Its complexity stands out when put side-by-side with the simple glibc loader on Linux. This complexity helps explain why process creation on Linux is faster than on Windows.
Concurrency Experiments
- Load libraries
- We set hardware (write) watchpoints on the
_dl_load_lock
and_dl_load_write_lock
data, then load two libraries to check how often and where in the code these locks are acquired - See the GDB log
- We set hardware (write) watchpoints on the
- Loading a library from a global constructor (recursive loading)
- Windows: ✔️
- Linux: ✔️
- Lazy/delay loading of a library import from a global constructor
- Windows: ✘ (deadlock)
- Lazy loading may fail (even within the same thread), but lazy linking works
- This hurdle came up in "Perfect DLL Hijacking"; look for the default delay load helper function:
__delayLoadHelper2
in the deadlocked call stack -
Microsoft documentation: "A DLL project that delays the loading of one or more DLLs itself shouldn't call a delay-loaded entry point in
DllMain
."
- Linux: ✔️
- Both dynamic loading (
dlopen
) and dynamic linking (at load time) tests were done in the most equivalent way
- Both dynamic loading (
- As mentioned in the Lazy Loading and Lazy Linking OS Comparison section, Linux doesn't natively support lazy loading
- However, effectively accomplishing lazy loading through various techniques comes down to recursive loading, which is already known to work
- What I tested in this experiment is that lazy linking works under
_dl_load_lock
- Regardless, the GNU loader is immune to lazy loading deadlocks, so it gets a pass
- Windows: ✘ (deadlock)
- Spawning a thread and then waiting for its creation/termination from a global constructor
- Windows: ✘ (deadlock)
-
CreateThread
thenWaitForSingleObject
on thread handle fromDllMain
- Deadlock on
LdrpInitCompleteEvent
inLdrpInitialize
function during process startup (see Windows Loader Initialization Locking Requirements for more info) or deadlock onLdrpLoadCompleteEvent
inLdrpInitializeThread
(runsDLL_THREAD_ATTACH
; this is essentially useless by the way and libraries often disable it, among other things) ➜LdrpDrainWorkQueue
function during process run-time, if it was possible to continue still then thread creation would further block when acquiring loader lock to runDLL_THREAD_ATTACH
- This represents an overextension of the loader's locks because the threading implementation should have zero knowledge of the loader (or at least its locking) and vice-versa
- Specifying
THREAD_CREATE_FLAGS_SKIP_LOADER_INIT
(new in Windows 10) onNtCreateThreadEx
can bypass these thread creation blockers- I haven't seen an occurrence of Windows internally spawning a thread with this flag, so these deadlocks remain prevalent
-
- Linux: ✔️
- Windows: ✘ (deadlock)
- Registering an
atexit
handler from a library- Windows: ✘ (either loader lock or CRT exit lock)
- In a DLL, VC++ compiles
atexit
in a DLL down running in theDLL_PROCESS_DETACH
of that module, so there's loader lock - In an EXE, an
atexit
handler runs under the CRT "exit" critical section lock, which isn't thread safe
- In a DLL, VC++ compiles
- Linux: ✔️
- At process exit, there's no loader lock (
_dl_load_lock
); if a programmer closes the library (dlclose
) then there's loader lock- The POSIX
atexit
specification doesn't mention libraries and what should happen if theatexit
handler gets unloaded from memory, but this seems like the best possible handling of the scenario
- The POSIX
- Releases the more specialized
__exit_funcs_lock
lock before calling into anatexit
handler, then reacquires when the lock after - The
__run_exit_handlers
function was designed to be reentrant and thread-safe
- At process exit, there's no loader lock (
- Windows: ✘ (either loader lock or CRT exit lock)
- Spawning a thread from a global constructor, waiting on the thread's termination, then loading another library in the new thread
- Windows: ✘ (deadlock, failed the prerequisite test)
- Linux: ✘ (deadlock)
-
See the log, we deadlock while trying to acquire
_dl_load_lock
on the new thread
-
See the log, we deadlock while trying to acquire
- This failing makes sense because the thread loading
lib1
already holds the loader lock. Consequently, waiting on the termination of a new thread that tries to acquire the loader lock so it can loadlib2
will deadlock. Recursive acquisition of the loader lock can't happen becauselib2
is being loaded on a separate thread. We created a deadlock in the loader.
The Root of DllMain
Problems
The Windows loader, in contrast to the GNU loader (or Unix-like loaders in general), is more vulnerable to deadlock or crash scenarios for a variety of architectural reasons. "The Root of DllMain
Problems" (or more casually, "DllMain
Rules Rewritten") provides a fundamental understanding of DllMain
hurdles and why they exist. It improves on Microsoft's "DLL Best Practices", often referred to as the "DllMain
rules" informally. "DLL Best Practices" originally goes back to an ancient (in technology years) 2006 Microsoft document (the document has undergone minimal changes on Microsoft's documentation website since then). The architectural reasons for DllMain
issues in Windows include:
- Windows is a monolith
- The monolithic architecture of the Windows API may cause separate lock hierarchies to interleave between threads resulting in an ABBA deadlock
- Countless Windows API operations may unexpectedly perform COM operations (e.g. WinHTTP AutoProxy)
- The Windows threading implementation meshes with the loader at thread startup and exit
- Windows kernel mode and user mode closely integrate (NT and NTDLL), whereas, on GNU/Linux, these two vital OS components aren't even made by the same people (glibc userland and Linux kernel all held together by, at minimum, the POSIX, System V ABI, and C standards)
- Generally, monolithic systems where everything is touching makes designing concurrent systems that don't deadlock harder
- Contrast this with the Unix philosophy, which encourages modularity
- The monolithic architecture of the Windows API may cause separate lock hierarchies to interleave between threads resulting in an ABBA deadlock
- Windows architecture prioritizes libraries that provide functionality, processes housing multiple threads, and local/remote procedure calls (LPC/RPC) over separate processes
- Splitting the Windows API, C runtime, and lots of core system functionality across many DLLs (there's 3000+ DLLs in
C:\Windows\System32
) leaves more opportunities for actions done during library loading/unloading or initialization/deinitialization to go wrong- Library loader initialization reentrancy issues
- Microsoft mostly built the legacy loader to be reentrant, however, how it performed module initialization was subject to crashes or correctness issues due to the loader's poor ability to enforce the correct order of operations when initializing modules or even initialzing modules at all upon being reentered (this issue was at the heart of the previous "
DllMain
Best Practices") - Microsoft introducing the (dynamic) dependency graph starting with Windows 8 significantly helps solve the module initialization problems such as when loading a library from
DllMain
, but there could potentially still be issues due to circular dependencies ("dependency loops") - The legacy loader effectively had a "dependency graph" (immediately collapsed into a linked list), however it was only formed by walking the import address tables (IATs) at the start of a load and wasn't able to readjust dynamically to the reentrant case of someone calling
LoadLibrary
fromDllMain
- Microsoft mostly built the legacy loader to be reentrant, however, how it performed module initialization was subject to crashes or correctness issues due to the loader's poor ability to enforce the correct order of operations when initializing modules or even initialzing modules at all upon being reentered (this issue was at the heart of the previous "
- A module you depend on or call a function from may have already deinitialized (
DLL_PROCESS_DETACH
) by the time your module deinitializes (the loader correctly deinitializes modules in the opposite order it initialized them but this issue can still happen due to circular dependencies or aDLL_PROCESS_DETACH
doing some dynamic loader operations) - Problematic
GetModuleHandle
thenGetProcAddress
fromDllMain
assuming a DLL is already loaded when it may not be yet or has only partially loaded (also generally, because the dependency graph won't pick up on a dynamic operation such asGetModuleHandle
orLoadLibrary
)- Nor POSIX or GNU loader extensions define a function like
GetModuleHandle
, instead only havingdlopen
due to the former function's inherent weaknesses
- Nor POSIX or GNU loader extensions define a function like
- Poses a greater risk of an application performing an action that assumes
DLL_PROCESS_ATTACH
andDLL_PROCESS_DETACH
always run on the same thread when they may not (this is true of the GNU loader too, but Windows has a higher risk due to its architecture)
- Library loader initialization reentrancy issues
- In contrast, Unix-like systems prioritize processes and the idea that everything is a file
- GNU/Linux exposes all its core system functionality through a single
libc
implementation library (also including a separateld
loader library but GNU is working on merging these into a single library similar to what musl has already done to improve performance)
- GNU/Linux exposes all its core system functionality through a single
- Splitting the Windows API, C runtime, and lots of core system functionality across many DLLs (there's 3000+ DLLs in
- Non-reentrant or thread-unsafe design
- Notable Windows components like CRT/DLL exit (
atexit
) and FLS callbacks either weren't designed with reentrancy or thread safety in mind and may also run under loader lock (Note: These aren't part of the loader but they may integrate with it so the point stands) - Delay loading is non-reentrant in some cases involving
DllMain
, potentially leading to deadlock- Microsoft likely does this on purpose to guard against a partially initialized library accidentally loading another library which then (perhaps indirectly) depends on the currently partially uninitialized library
- Notable Windows components like CRT/DLL exit (
- Windows overrelies on dynamic initialization and dynamic operations in general
- It's always best practice for robustness and performance to initialize statically (i.e. at compile time) over dynamically (using global consturctors/destructors or Windows
DllMain
) - It's generally more common for even simple operations on Windows to require dynamic function calls such as getting a pseudo-handle to the current thread (
GetCurrentThread
is a stub that returns0xfffffffffffffffe
at least on Windows 10) - In contrast, POSIX data structures provide a static initialization option when possible
- It's always best practice for robustness and performance to initialize statically (i.e. at compile time) over dynamically (using global consturctors/destructors or Windows
- Unexpectedly loading libraries
- Inherently, delay loading may unexpectedly cause library loading when the programmer didn't intend, thus leading to deadlock or other issues
- MacOS previously supported lazy loading until Apple removed it likely due to scenarios where it becomes an anti-feature and any performance gains not being worth the trade-off
- Inherently, delay loading may unexpectedly cause library loading when the programmer didn't intend, thus leading to deadlock or other issues
- Backward compatibility
- Running arbitrary Windows API functions under a DLL's activation context (
LDR_DATA_TABLE_ENTRY.EntryPointActivationContext
; the loader activates this context before running that module'sDllMain
code) for application compatibility, if it has one, may cause problems - Microsoft's intense commitment to staying backward compatible means you have to remain bug-compatible with all your mistakes once they (whether you want them to or not) become part of a public API/ABI
- Historically, people have been less motivated to stick to the public Windows API because there was only the Microsoft Windows implementation (also, Wine tries to remain bug-compatible) to support; unlike standardized Unix-like systems where going beyond the public API will certainly break across distros using different libc/loader implementations
- With concurrency, this means you must be incredibly careful not to break any assumptions applications may make about thread synchronization, the order of operations, and related state
- Also, if you want to completely redesign a system or data structure to unlock certain operations/functions and increase overall concurrency (perhaps switching to a careful design utilizing more atomic loads and stores), well, there's a good chance you can't due to backward compatibility
- This fact rings especially true when DLLs like
ntdll.dll
exposes its locks, such as loader lock, through the exportedLdrLockLoaderLock
/LdrUnlockLoaderLock
functions ormsvcrt.dll
makes the_lock
and_unlock
functions a part of its exported API (the new UCRTucrtbase.dll
DLL introduced with Windows 10 doesn't expose this functionality) - This fact also doesn't bode well with the Redmond-based business' reputation of "moving fast and breaking things" to achieve market dominance in a new space (although this strategy with Windows has subsided now, the technical debt remains)
- Running arbitrary Windows API functions under a DLL's activation context (
The outcomes of these architectural facts and faults are far-reaching. Additionally, they often tie into each other, thus complicating matters.
With a newfound understanding of problems affecting DllMain
and the greater perspective gained from how Unix-like operating systems do it right, you can reason about the safety of performing a given action from Dllmain
.
Comparing OS Library Loading Locations
The Windows loader is searching for DLLs to load in a vast (and growing) number of places. Strangely, Windows uses the PATH
environment variable for locating programs (similar to Unix-like systems) as well as DLLs. This Microsoft documentation still doesn't cover all the possible locations, though, because while debugging the loader during a LoadLibrary
, I saw LdrpSendPostSnapNotifications
eventually calls through to SbpRetrieveCompatibilityManifest
(this isn't part of a notification callback). This Sbp
-prefixed function searches for application compatibility shims in SDB files which may result in a compat DLL loading. Then there's also WinSxS and activation contexts (DLLs in C:\Windows\WinSxS
) for loading versioned DLLs typically based on the application's manifest (these are usually embedded in the binary). The plethora of possible search locations may contribute to DLL Hell and DLL hijacking (also known as DLL preloading or DLL sideloading) problems in Windows, the latter of which makes vulnerabilities due to a privileged process loading an attacker-controlled library more likely.
The GNU/Linux ecosystem differs due to system package managers (e.g. apt
or dnf
). All programs are built against the same system libraries (this is possible because all the packages are open source). Proprietary apps are generally statically linked or come with all their necessary libraries. The trusted directories for loading libraries and configuration file path for setting additional directories can be found in the ldconfig
manual. Beyond that, you can set the LD_LIBRARY_PATH
environment variable to choose other places the loader should search for libraries and LD_PRELOAD
to specify a libraries to load before any other library (including libc
). Loading libraries based on environment variables is a default feature that may optionally be disabled (and is always disabled for setuid
binaries). Binaries can include an rpath
to specify additional run-time library search paths.
DllMain
vs Global Constructors and Destructors
A library's global constructors or destructors are the cross-platform equivalent to the Windows DllMain
when it's called at DLL_PROCESS_ATTACH
and DLL_PROCESS_DETACH
times. Global constructors/destructors are umbrella terms which include the Windows DllMain
only present in libraries.
Additonally, the Windows loader may call each module's DllMain
at DLL_THREAD_ATTACH
and DLL_THREAD_DETACH
times. This only happens at thread start/exit time, meaning, for example, that a DLL loaded after thread start won't preempt that thread to run its DllMain
with DLL_THREAD_ATTACH
. These calls can be disabled as a performance optimization by calling DisableThreadLibraryCalls
at DLL_PROCESS_ATTACH
time.
At a language level, a constructor is global if the code creates an instance of that class (speaking in C++ terms) in the global scope; otherwise, the constructor will run upon creating an instance of that class instead of during library initialization. Compilers also commonly provide access to module initialization/deinitialization functions with compiler-specific syntax. In GCC, a programmer can create global constructors/destructors using the __attribute__((constructor))
and __attribute__((destructor))
functions or the _init
and _fini
functions historically.
Constructors and destructors originate from object-oriented programming (OOP) standardized in C++ (C++ got initial standardization in 1998). They don't exist in the C standard. Global constructors/destructors are what module initialization/deinitialization routines are known as in OOP languages. However, the concept of code that runs when a module loads/unloads goes at least as far back as System V Application Binary Interface Version 4. However, even before this time, the inherent need for custom code that runs prior to a program calling a library's exported routines was clear because some resources require dynamic initialization (i.e. during program load/run-time). For example, you may want dynamic initialization to communicate with another process or to initialize an implementation dependent data structure (rather than storing the internal POD directly in your module, which would necessitate that ABI remain backward compatible forever or be versioned). In the common case where default mutex attributes are appropriate, POSIX mutexes can be statically initialized with the PTHREAD_MUTEX_INITIALIZER
macro; otherwise, dyanmic initialization with the pthread_mutex_init
function is necessary.
In the ELF executable format, global constructors and destructors are standardized in the System V Application Binary Interface as the .init
and .fini
sections. Note that this doesn't cover the newer, non-standard but common and generally agreed-upon .init_array
/.fini_array
sections or the deprecated .ctors
/.dtors
sections, which modern binaries don't include (verified with objdump -h
).
The PE (Windows) executable format standard doesn't define any sections specific to module initialization; instead, a DllMain
or any global constructors/destructors are included with the rest of the program code in the .text
section.
The CLR loader uses a module's .cctor
section to initialize .NET assemblies. Each module's .cctor
section is the "managed module initializer". However, this carries no relation to the Windows loader and native libraries.
The Windows loader calls a module's LDR_DATA_TABLE_ENTRY.EntryPoint
at module initialization; it has no knowledge of DllMain
or C++ global constructors/destructors. Merging these into one callable EntryPoint
is the job of a compiler. For instance, MSVC, which compiles a stub into your DLL that calls any global constructors/destructors followed by DllMain
.
Lazy Loading and Lazy Linking OS Comparison
It's important to differentiate between loading and linking. Loading involves mapping into memory. Linking (sometimes also known as "binding") refers to resolving symbols. Doing either of these operations lazily means that it's done on an as-needed basis instead of all at process startup or all at library load in the case of lazy linking.
Windows collectively refers to lazy loading and lazy linking as "delay loading". However, I use the distinguished terms throughout this section.
Lazy loading exists as a first-class citizen on Windows but not on Unix systems. While it's a potentially useful optimization, architectural differences make it less needed on Unix systems than Windows. Windows prefers fewer programs with more threads, which get their functionality from libraries. In contrast, Unix favors more programs with fewer threads, leading to, on average, fewer libraries per process. Looking at the lengthy dependency chains for many common DLLs ties this in with the more monolithic architecture of Windows.
One can quickly check the dependency chain of a DLL by using dumpbin
, which is a tool that ships with Visual Studio: dumpbin.exe /imports C:\Windows\System32\shell32.dll
Using dumpbin
on kernel32.dll
reveals it contains a lazy load for rpcrt4.dll
when any of these also lazy linking imports are called:
Section contains the following delay load imports:
...
RPCRT4.dll
00000001 Characteristics
6B8B06D0 Address of HMODULE
6B8C0000 Import Address Table
6B892B2C Import Name Table
6B892C70 Bound Import Name Table
00000000 Unload Import Name Table
0 time date stamp
6B824205 167 RpcAsyncCompleteCall
6B8241F1 20C RpcStringBindingComposeW
6B8241E7 171 RpcBindingFromStringBindingW
6B8241CC 169 RpcAsyncInitializeHandle
6B8241FB 30 I_RpcExceptionFilter
6B82420F 181 RpcBindingSetAuthInfoExW
6B824237 99 NdrAsyncClientCall
6B824223 166 RpcAsyncCancelCall
6B82422D 16F RpcBindingFree
6B824219 210 RpcStringFreeW
On Windows, lazy loading/linking is both a feature of the MSVC linker and the Windows loader with the NTDLL exported functions LdrResolveDelayLoadsFromDll
, LdrResolveDelayLoadedAPI
, and LdrQueryOptionalDelayLoadedAPI
(more functions internally, not exposed as exports).
On Unix systems, lazy loading can be effectively achieved by doing dlopen
and then dlsym
at run-time. There are techniques to get more seamless lazy loading on Unix systems, but only with caveats. One could also employ a proxy design pattern in their code to achieve the effect. Mac used to support lazy loading until Apple removed this feature from the linker used by Xcode. Apple likely removed lazy loading because it's inherently vulnerable to deadlocking on loader lock. This vulnerability is due to lazy loading's ability to cause an application to load a library unexpectedly when calling a lazy-loaded function.
On the other hand, lazy linking (as it's referred to in the GNU ld
manual) is a first-class citizen on both POSIX and Windows systems (with an asterisk for Windows). On POSIX systems, this is simply calling dlopen
with the RTLD_LAZY
flag or providing the -z lazy
flag to the linker. Windows doesn't seem to expose this functionality as readily (e.g. through LoadLibrary
), but if nothing else, you always call the aforementioned NTDLL exports directly.
A key difference between lazy linking as done by the POSIX loaders (or at least the GNU loader) compared to the Windows loader is that Windows only supports lazy linking as a part of lazy loading (hence Microsoft collectively refers to it as "delay loading"). There's no way to make functions a program may import from a DLL link lazily without having the entire DLL be lazily loaded.
While upstream GNU sets lazy linking (also known as lazy binding) as the default (see ld
manual), some GNU/Linux distributions may include patches to remove lazy linking as the default for security reasons. On my Fedora 38 machine, only some packaged binaries are compiled without lazy linking (i.e. full RELRO) and GCC compiles with lazy linking by default.
LDR_DDAG_NODE.State
Analysis
LDR_DDAG_NODE.State
or LDR_DDAG_STATE
tracks a module's entire lifetime from beginning to end. With this analysis, I intend to extrapolate information based on the known types given to us by Microsoft (dt _LDR_DDAG_STATE
command in WinDbg).
Findings were gathered by tracing all LDR_DDAG_STATE.State
values at load-time and tracing a library unload, as well as searching disassembly.
Each state represents a stage of loader work. This table comprehensively documents where these state changes occur throughout the loader and which locks are present.
Loader work requires first incrementing the ntdll!LdrpWorkInProgress
reference counter. ntdll!LdrpWorkInProgress
can only safely be decremented when the loader has completed all stages of loader work on the module(s) it's operating on. Only one thread, the caller of LdrpDrainWorkQueue
(i.e. the future load owner), can do loader work at a time. For instance, a typical load ranging from LdrModulesPlaceHolder
to LdrModulesReadyToRun
(may also include LdrModulesMerged
) or a typical unload ranging from LdrModulesUnloading
to LdrModulesUnloaded
.
When loading a library, loader work begins with the requesting thread calling LdrpDrainWorkQueue
to perform mapping and snapping (i.e. both together is "processing"). By calling LdrpDrainWorkQueue
, the thread requesting a library load agrees to become the load owner (LoadOwner
flag in TEB.SameTebFlags
). LdrpDrainWorkQueue
increments ntdll!LdrpWorkInProgress
and does the necessary mapping and snapping work, offloading this work to the parallel loader worker threads (LoaderWorker
flag in TEB.SameTebFlags
) whenever possible. Every increment to the ntdll!LdrpWorkInProgress
reference counter past the initial increment from 0
to 1
by LdrpDrainWorkQueue
indicates another loader worker thread processing a work item in parallel. While ntdll!LdrpWorkInProgress
equals 1
, the last thing LdrpDrainWorkQueue
does before returning is set the current thread as the load owner. With mapping and snapping complete, it's the load owner's job to fulfill the remainder of the loading process. The loader only uses the ntdll!LdrpWorkQueue
linked list data structure for enqueuing mapping and snapping work to the work queue. For all other operations, the loader primarily relies on the Dependencies
and IncomingDependencies
module information DAG data structures to navigate modules. From LdrpDrainWorkQueue
to LdrpDropLastInProgressCount
(where ntdll!LdrpWorkInProgress
is finally set to 0
and the current thread is decommissioned as the load owner) ties the entire loading process together. For more information, see the full reverse engineering of the LdrpDrainWorkQueue
function.
LDR_DDAG_STATE States | State Changing Function(s) | Remarks |
---|---|---|
LdrModulesMerged (-5) | LdrpMergeNodes |
LdrpModuleDatatableLock is held during this state change. See LdrModulesCondensed state for more info. |
LdrModulesInitError (-4) | LdrpInitializeGraphRecurse |
During DLL_PROCESS_ATTACH , if a module's DllMain returns FALSE for failure then this module state is set (any other return value counts as success). LdrpLoaderLock is held here. |
LdrModulesSnapError (-3) | LdrpCondenseGraphRecurse |
This function may set this state on a module if a snap error occurs. See LdrModulesCondensed state for more info. |
LdrModulesUnloaded (-2) | LdrpUnloadNode |
Before setting this state, LdrpUnloadNode may walk LDR_DDAG_NODE.Dependencies , holding LdrpModuleDataTableLock to call LdrpDecrementNodeLoadCountLockHeld thus decrementing the LDR_DDAG_NODE.LoadCount of dependencies and recursively calling LdrpUnloadNode to unload dependencies. Loader lock (LdrpLoaderLock ) is held here. |
LdrModulesUnloading (-1) | LdrpUnloadNode |
Set near the start of this function. This function checks for LdrModulesInitError , LdrModulesReadyToInit , and LdrModulesReadyToRun states before setting this new state. After setting state, this function calls LdrpProcessDetachNode . Loader lock (LdrpLoaderLock ) is held here. |
LdrModulesPlaceHolder (0) | LdrpAllocateModuleEntry |
The loader directly calls LdrpAllocateModuleEntry until parallel loader initialization (LdrpInitParallelLoadingSupport ) occurs at process startup. At which point (with exception to directly calling LdrpAllocateModuleEntry once more soon after parallel loader initialization to allocate a module entry for the EXE), the loader calls LdrpAllocatePlaceHolder (this function first allocates a LDRP_LOAD_CONTEXT structure), which calls through to LdrpAllocateModuleEntry (this function places a pointer to this module entry's LDRP_LOAD_CONTEXT at LDR_DATA_TABLE_ENTRY.LoadContext ). LdrpAllocateModuleEntry , along other like creating the module's LDR_DATA_TABLE_ENTRY , allocates a LDR_DDAG_NODE with zero-initialized heap memory. |
LdrModulesMapping (1) | LdrpMapCleanModuleView |
I've never seen this function get called; state typically jumps from 0 to 2. Only the LdrpGetImportDescriptorForSnap function may call this function which itself may only be called by LdrpMapAndSnapDependency (according to a disassembly search). LdrpMapAndSnapDependency typically calls LdrpGetImportDescriptorForSnap ; however, LdrpGetImportDescriptorForSnap doesn't typically call LdrpMapCleanModuleView . This state is set before mapping a memory section (NtMapViewOfSection ). Mapping is the process of loading a file from disk into memory. |
LdrModulesMapped (2) | LdrpProcessMappedModule |
LdrpModuleDatatableLock is held during this state change. |
LdrModulesWaitingForDependencies (3) | LdrpLoadDependentModule |
This state isn't typically set but, during a trace, I was able to observe the loader set it by launching a web browser (Google Chrome) under WinDbg, which triggered the watchpoint in this function when loading app compatibility DLL C:\Windows\System32\ACLayers.dll . Interstingly, the LDR_DDAG_STATE decreases by one here from LdrModulesSnapping to LdrModulesWaitingForDependencies ; the only time I've observed this. LdrpModuleDatatableLock is held during this state change. |
LdrModulesSnapping (4) | LdrpSignalModuleMapped or LdrpMapAndSnapDependency |
In the LdrpMapAndSnapDependency case, a jump from LdrModulesMapped to LdrModulesSnapping may happen. LdrpModuleDatatableLock is held during state change in LdrpSignalModuleMapped , but not in LdrpMapAndSnapDependency . Snapping is the process of resolving the library’s import address table (module imports and exports). |
LdrModulesSnapped (5) | LdrpSnapModule or LdrpMapAndSnapDependency |
In the LdrpMapAndSnapDependency case, a jump from LdrModulesMapped to LdrModulesSnapped may happen. LdrpModuleDatatableLock isn't held here in either case. |
LdrModulesCondensed (6) | LdrpCondenseGraphRecurse |
This function recevies a LDR_DDAG_NODE as its first argument and recursively calls itself to walk LDR_DDAG_NODE.Dependencies . On every recursion, this function checks whether it can remove the passed LDR_DDAG_NODE from the graph. If so, this function acquires LdrpModuleDataTableLock to call the LdrpMergeNodes function, which receives the same first argument, then releasing LdrpModuleDataTableLock after it returns. LdrpMergeNodes discards the uneeded node from the LDR_DDAG_NODE.Dependencies and LDR_DDAG_NODE.IncomingDependencies DAG adjacency lists of any modules starting from the given parent node (first function argument), decrements LDR_DDAG_NODE.LoadCount to zero, and calls RtlFreeHeap to deallocate LDR_DDAG_NODE DAG nodes. After LdrpMergeNodes returns, LdrpCondenseGraphRecurse calls LdrpDestroyNode to deallocate any DAG nodes in the LDR_DDAG_NODE.ServiceTagList list of the parent LDR_DDAG_NODE then deallocate the parent LDR_DDAG_NODE itself. LdrpCondenseGraphRecurse sets the state to LdrModulesCondensed before returning. Note: The LdrpCondenseGraphRecurse function and its callees rely heavily on all members of the LDR_DDAG_NODE structure, which needs further reverse engineering to fully understand the inner workings and "whys" of what's occurring here. Condensing is the process of discarding unnecessary nodes from the dependency graph. |
LdrModulesReadyToInit (7) | LdrpNotifyLoadOfGraph |
This state is set immediately before this function calls LdrpSendPostSnapNotifications to run post-snap DLL notification callbacks. The module is mapped and snapped but pending initialization. As the loader initializes nodes (i.e. modules) in the dependency graph (loader lock is held), each node's state will transition to LdrModulesInitializing then LdrModulesReadyToRun (or LdrModulesInitError if initialization fails). |
LdrModulesInitializing (8) | LdrpInitializeNode |
Set at the start of this function, immediately before linking a module into the InInitializationOrderModuleList list. Loader lock (LdrpLoaderLock ) is held here. Initializing is the process of running a module's initialization function (i.e. DllMain ). |
LdrModulesReadyToRun (9) | LdrpInitializeNode |
Set at the end of this function, before it returns. Loader lock (LdrpLoaderLock ) is held here. |
See what a LDR_DDAG_STATE trace log looks like (be aware of the warnings).
Procedure/Symbol Lookup Comparison (Windows GetProcAddress
vs POSIX dlsym
GNU Implementation)
The Windows GetProcAddress
and POSIX dlsym
functions are platform equivalents because they resolve a procedure/symbol name to its address. They differ because GetProcAddress
can only resolve function export symbols, whereas dlsym
can resolve any global (RTLD_GLOBAL
) or local (RTLD_LOCAL
) symbol. There's also this difference: The first argument of GetProcAddress
requires passing in a module handle. In contrast, the first argument of dlsym
can take a module handle, but it also accepts one of the RTLD_DEFAULT
or RTLD_NEXT
flags (or "pseudo handles" in more Windows terminology). Let's discuss how GetProcAddress
functions first.
GetProcAddress
receives an HMODULE
(a module's base address) as its first argument. The loader maintains a red-black tree sorted by each module's base address called ntdll!LdrpModuleBaseAddressIndex
. GetProcAddress
➜ LdrGetProcedureAddressForCaller
➜ LdrpFindLoadedDllByAddress
(this is a call chain) searches this red-black tree for the matching module base address to ensure a valid DLL handle. Searching the LdrpModuleBaseAddressIndex
red-black tree mandates acquiring the ntdll!LdrpModuleDataTableLock
lock. If locating the module fails, GetProcAddress
sets the thread error code (retrieved with the GetLastError
function) to ERROR_MOD_NOT_FOUND
. GetProcAddress
receives a procedure name as a string for its second argument. GetProcAddress
➜ LdrGetProcedureAddressForCaller
➜ LdrpResolveProcedureAddress
➜ LdrpGetProcedureAddress
calls RtlImageNtHeaderEx
to get the NT header (IMAGE_NT_HEADERS
) of the PE image. IMAGE_NT_HEADERS
contains optional headers (IMAGE_OPTIONAL_HEADER
), including the image data directory (IMAGE_DATA_DIRECTORY
, this is in the .rdata
section). The data directory (IMAGE_DATA_DIRECTORY
) includes multiple directory entries, including IMAGE_DIRECTORY_ENTRY_EXPORT
, IMAGE_DIRECTORY_ENTRY_IMPORT
, IMAGE_DIRECTORY_ENTRY_RESOURCE
, and more. LdrpGetProcedureAddress
gets the PE's export directory entry. The compiler sorted the PE export directory entries alphabetically by procedure name ahead of time. LdrpGetProcedureAddress
performs a binary search over the sorted export entry names, looking for a name matching the procedure name passed into GetProcAddress
. If locating the procedure fails, GetProcAddress
sets the thread error code to ERROR_PROC_NOT_FOUND
. Locking isn't required while searching for an export because PE image exports are resolved once during library load and remain unchanged (this doesn't cover delay loading). In classic Windows monolithic fashion, GetProcAddress
may do much more than what's stated here on edge cases. Still, for the sake of our comparison, we only need to know how GetProcAddress
works at its core.
GNU's POSIX-compliant dlopen
implementation firstly differs from GetProcAddress
because the former will not validate a correct module handle before searching for a symbol. Pass in an invalid module handle, and the program will crash; to be fair, you deserve to crash if you do that. Also, not validating the module handle provides a great performance boost. Depending on the handle
passed to dlopen
, the GNU loader searches for symbols in a few ways. Here, we will cover the most straightforward case when calling dlsym
with a handle to a library (i.e. dlsym("mylib.so", "myfunc")
). The ELF standard specifies the use of a hash table for searching symbols. do_sym
(called by _dl_sym
) calls _dl_lookup_symbol_x
to find the symbol in our specified library (also called an object). _dl_lookup_symbol_x
calls _dl_new_hash
to hash our symbol name with the djb2 hash function (look for the magic numbers). Recent versions of the GNU loader use this djb2 based hash function, which differs from the old standard ELF hash function based on the PJW hash function. At its introduction, this new hash function improved dynamic linking time by 50%. Though commonly used across Unix systems, this hash function is only a GNU extension and requires standardization. It's also worth noting that, in 2023, someone caught an overflow bug in the original hash function described by the System V Application Binary Interface. _dl_lookup_symbol_x
calls do_lookup_x
, where the real searching begins. do_lookup_x
filters on our library's link_map
to see if it should disregard searching it for any reason. Passing that check, do_lookup_x
gets pointers into our library's DT_SYMTAB
and DT_STRTAB
ELF tables (the latter for use later while searching for matching symbol names). Based on our symbol's hash (calculated in _dl_new_hash
), do_lookup_x
selects a bucket from the hash table to search for symbols from. l_gnu_buckets
is an array of buckets in our hash table to chose from. At build time during the linking phase, the linker builds each ELF image's hash table with the number of buckets, which adjusts depending on how many symbols are in the binary. With a bucket selected, do_lookup_x
fetches the l_gnu_chain_zero
chain for the given bucket and puts a reference to the chain in the hasharr
pointer variable for easy access. l_gnu_chain_zero
is an array of GNU hash entries containing standard ELF symbol indexes. The ELF symbol indexes inside are what's relevant to us right now. Each of these indexes points to a symbol table entry in the DT_SYMTAB
table. do_lookup_x
iterates through the symbol table entries in the chain until the selected symbol table entry holds the desired name or hits STN_UNDEF
(i.e. 0
), marking the end of the array. l_gnu_buckets
and l_gnu_chain_zero
inherit similarities in structure from the original ELF standardized l_buckets
and l_chain
arrays which the GNU loader still implements for backward compatibility. For the memory layout of the symbol hash table, see this diagram. There appears to be a fast path to quickly eliminate any entries that we can know early on won't match—passing the fast path check, do_lookup_x
calls ELF_MACHINE_HASH_SYMIDX
to extract the offset to the standard ELF symbol table index from within the GNU hash entry. The GNU hash entry is a layer on top of the standard ELF symbol table entry; in the old but standard hash table implementation, you can see that the chain is directly an array of symbol table indexes. Having obtained the symbol table index, dl_lookup_x
passes the address to DT_SYMTAB
at the offset of our symbol table index to the check_match
function. The check_match
function then examines the symbol name cross-referencing with DT_STRTAB
where the ELF binary stores strings to see if we've found a matching symbol. Upon finding a symbol, check_match
looks if the symbol requires a certain version (i.e. dlvsym
GNU extension). In the unversioned case with dlsym
, check_match
determines if this is a hidden symbol (e.g. -fvisibility=hidden
compiler option in GCC/Clang), and if so not returning the symbol. This loop restarts at the fast path check in dl_lookup_x
until check_match
finds a matching symbol or there are no more chain entries to search. Having found a matching symbol, do_lookup_symbol_x
determines the symbol type; in this case, it's STB_GLOBAL
and returns the successfully found symbol address. Finally, the loader internals pass this value back through the call chain until dlsym
returns the symbol address to the user!
Note that the glibc source code loosely follows a Hungarian notation that prefixes structure members with l_
and structures with r_
(e.g. throughout the link_map
structure).
Also, note that, unlike the Windows ntdll!LdrpHashTable
(which serves an entirely different purpose), the hash table in each ELF DT_SYMTAB
is made up of arrays instead of linked lists for each chain (each bucket has a chain). Using arrays (size determined during binary compilation) is possible because the ELF images aren't dynamically allocated structures like the LDR_DATA_TABLE_ENTRY
structures ntdll!LdrpHashTable
keeps track of. Arrays are significantly faster than linked lists because following links is a relatively expensive operation (e.g. you lose locality of reference).
Due to the increased locality of reference and a hash table being O(1) average and amortized time complexity vs a binary search being O(log n) time complexity, I believe that searching a hash table (bucket count optimized at compile time) and then iterating through the also optimally sized array is faster than the binary search approach employed by LdrpGetProcedureAddress
in Windows for finding symbol addresses.
How Does GetProcAddress
/dlsym
Handle Concurrent Library Unload?
The dlsym
function can locate RTLD_LOCAL
symbols (i.e. libraries that have been dlopen
ed with the RTLD_LOCAL
flag). Internally, the dlsym_implementation
function acquires _dl_load_lock
at its entry. Since dlclose
must also acquire _dl_load_lock
to unload a library, this prevents a library from unloading while another thread searches for a symbol in that same (or any) library. dlsym
eventually calls into _dl_lookup_symbol_x
to perform the symbol lookup.
Answering this question on Windows requires a more in-depth two-part investigation of GetProcAddress
and FreeLibrary
internals:
For GetProcAddress
, LdrGetProcedureAddressForCaller
(this is the NTDLL function GetProcAddress
calls through to) acquires the LdrpModuleDatatableLock
lock, searches for our module LDR_DATA_TABLE_ENTRY
structure in the ntdll!LdrpModuleBaseAddressIndex
red-black tree, checks if our DLL was loaded by dynamically loaded DLL (dynamically linked DLLs are already pinned by default, so they will never unload) if so atomically incrementing LDR_DATA_TABLE_ENTRY.ReferenceCount
. LdrGetProcedureAddressForCaller
releases the LdrpModuleDatatableLock
lock. Before acquiring the LdrpModuleDatatableLock
lock, note there's a special path for NTDLL whereby LdrGetProcedureAddressForCaller
checks if the passed base address matches ntdll!LdrpSystemDllBase
(this holds NTDLL's base address) and lets it skip close to the meat of the function where LdrpFindLoadedDllByAddress
and LdrpResolveProcedureAddress
occurs. Towards the end of LdrGetProcedureAddressForCaller
, it calls LdrpDereferenceModule
, passing the LDR_DATA_TABLE_ENTRY
of the module it was searching for a procedure in. LdrpDereferenceModule
, assuming the module isn't pinned (LDR_ADDREF_DLL_PIN
) or a static import (ProcessStaticImport
in LDR_DATA_TABLE_ENTRY.Flags
), atomically decrements the same LDR_DATA_TABLE_ENTRY.ReferenceCount
of the searched module. If LdrpDerferenceModule
senses the LDR_DATA_TABLE_ENTRY.ReferenceCount
reference counter has dropped to zero (this could only occur due to a concurrent thread decrementing the reference count), it will delete the necessary module information data structures and unmap the module. By reference counting, GetProcAddress
ensures a module isn't unmapped midway through its search.
The Windows loader maintains LDR_DATA_TABLE_ENTRY.ReferenceCount
and LDR_DDAG_NODE.LoadCount
reference counters for each module. The loader ensures there are no references to a module's LDR_DDAG_NODE
(LDR_DDAG_NODE.LoadCount
= 0) before there are no references to the same module's LDR_DATA_TABLE_ENTRY
(LDR_DATA_TABLE_ENTRY.ReferenceCount
= 0). This sequence is the correct order of operations for when the loader is dropping the final references to a module. The LdrUnloadDll
NTDLL function (public FreeLibrary
calls this) calls LdrpDecrementModuleLoadCountEx
which typically decrements a module's LDR_DDAG_NODE.LoadCount
then, if it hit zero, runs DLL_PROCESS_DETACH
. Lastly, LdrUnloadDll
calls LdrpDereferenceModule
which decrements the module's LDR_DATA_TABLE_ENTRY.ReferenceCount
. Unloading a library (when LoadCount
decrements for the last time from 1
to 0
) requires becoming the load owner (LdrUnloadDll
calls LdrpDrainWorkQueue
). Once the thread is appointed as the load owner (only one thread can be a load owner at a time), LdrUnloadDll
calls LdrpDecrementModuleLoadCountEx
with the DontCompleteUnload
argument set to FALSE
thus allowing actual module unload to occur instead of just decrementing the LDR_DDAG_NODE.LoadCount
reference counter. With LDR_DDAG_NODE.LoadCount
now at zero but the thread still being the load owner, LdrpDecrementModuleLoadCountEx
calls LdrpUnloadNode
to run the module's DLL_PROCESS_DETACH
event (LdrpUnloadNode
can also walk the dependency graph to unload other now unused libraries). LdrUnloadDll
then calls LdrpDropLastInProgressCount
to decomission the current thread as the load owner followed by calling LdrpDereferenceModule
to remove the reamining module information data structures and unmap the module.
The coarse-grained locking approach of GNU dlsym
is the only place the Windows loader approach is superior for maximizing concurrency and preventing deadlocks. I recommend that glibc switches use a more fine-grained locking approach like they already do with RTLD_GLOBAL
symbols and their global scope locking system (GSCOPE). Note that acquiring _dl_load_lock
also allows dlsym
to safely search link maps in the RTLD_DEFAULT
and RTLD_NEXT
pseudo-handle scenarios without acquiring _dl_load_write_lock
. However, that goal could also be accomplished by shortly acquiring the _dl_load_write_lock
lock in the _dl_find_dso_for_object
function, so this is only a helpful side effect.
Something interesting to note is that the GNU dlsym
approach to supporting concurrency ensures a module isn't initializing/uninitializing midway through its search while Windows GetProcAddress
provides no such guarantee. However, module initialization isn't a requirement for safely retrieving a symbol from a library. By splitting module mapping and snapping as well as the maintenance of module information data structures (not including LDR_DDAG_NODE
) from module initialization/deinitialization, the Windows loader achieves greater concurrency.
Keep in mind, that if your program frees libraries whose exports/symbols are still in use (after locating them with GetProcAddress
or dlsym
), then you can expect your application to crash due to your own negligence. In other words, GetProcAddress
/dlsym
only protects from internal data races (within the loader). However, you the programmer are responsible for guarding against external data races. Said in another way again: The loader doesn't know what your program might do. So, the loader can only maintain consistency with itself.
Lazy Linking Synchronization
The Procedure/Symbol Lookup Comparison (Windows GetProcAddress
vs POSIX dlsym
GNU Implementation) and How Does GetProcAddress
/dlsym
Handle Concurrent Library Unload? sections covers, including synchronization, how the GetProcAddress
/dlsym
functions, resolve a symbol name to an address. Lazy linking also does this and so they typically share many internals. However, given that the program may lazily resolve a library symbol at any time in execution, one must be careful to design a system that is flexible to concurrent scenarios.
On the GNU loader, dlsym
internally calls into the same internals as lazy linking. However, the GNU loader's approach to supporting the additional functionality POSIX specifies dlsym
to support creates some notable differences that you can read about in the aforementioned sections.
Windows lazy linking, which is merely a part of Window delay loading, is a different beast which I've only barely researched and there's not much public information on. See Lazy Loading and Lazy Linking OS Comparison for some context on Windows delay loading.
Stepping into a lazy linked function on the GNU loader reveals that it eventually calls the familiar _dl_lookup_symbol_x
function to locate a symbol. During dynamic linking, _dl_lookup_symbol_x
can resolve global symbols. The GNU loader uses GSCOPE, the global scope system, to ensure consistent access to STB_GLOBAL
symbols (as it's known in the ELF standard; this maps to the RTLD_GLOBAL
flag of POSIX dlopen
). The global scope is the searchlist
in the main link map. The main link map refers to the program's link map (not one of the libraries). In the TCB (this is the generic term for the Windows TEB) of each thread is a piece an atomic state flag (this is not a reference count), which can hold one of three states known as the gscope_flag
that keeps track of which threads are currently depending on the global scope for their operations. A thread uses the THREAD_GSCOPE_SET_FLAG
macro (internally calls THREAD_SETMEM
) to atomically set this flag and the THREAD_GSCOPE_RESET_FLAG
macro to atomically unset this flag. When the GNU loader requires synchronization of the global scope, it uses the THREAD_GSCOPE_WAIT
macro to call __thread_gscope_wait
. Note there are two implementations of __thread_gscope_wait
, one for the Native POSIX Threads Library (NPTL) used on Linux systems and the other for the Hurd Threads Library (HTL) which was previously used on GNU Hurd systems (with the GNU Mach microkernel). GNU Hurd has since switched to using NPTL. For our purposes, only NPTL is relevant. The __thread_gscope_wait
function iterates through the gscope_flag
of all (user and system) threads, signalling to them that it's waiting (THREAD_GSCOPE_FLAG_WAIT
) to synchronize. GSCOPE is a custom approach to locking that works by the GNU loader creating its own synchronization mechanisms based on low-level locking primitives. Creating your own synchronization mechanisms can similarly be done on Windows using the WaitOnAddress
and WakeByAddressSingle
/WakeByAddressAll
functions. Note that the THREAD_SETMEM
and THREAD_GSCOPE_RESET_FLAG
macros don't prepend a lock
prefix to the assembly instruction when atomically modifying a thread's gscope_flag
. These modifications are still atomic because xchg
is atomic by default on x86, and a single aligned load or store (e.g. mov
) operation is also atomic by default on x86 up to 64 bits. If gscope_flag
were a reference count, the assembly instruction would require a lock
prefix (e.g. lock inc
) because incrementing internally requires two memory operations, one load and one store. The GNU loader must still use a locking assembly instruction/prefix on architectures where memory consistency doesn't automatically guarantee this (e.g. AArch64/ARM). Also, note that all assembly in the glibc source code is in AT&T syntax, not Intel syntax.
The Windows loader resolves lazy linkage by calling LdrResolveDelayLoadedAPI
(an NTDLL export). This function makes heavy use of LDR_DATA_TABLE_ENTRY.LoadContext
. Symbols for the LoadContext
structure aren't publicly available, and some members require reverse engineering. In the protected load case, LdrpHandleProtectedDelayload
➜ LdrpWriteBackProtectedDelayLoad
(LdrResolveDelayLoadedAPI
may call this or LdrpHandleUnprotectedDelayLoad
) acquires the LDR_DATA_TABLE_ENTRY.Lock
lock and never calls through to LdrpResolveProcedureAddress
. The unprotected case eventually calls through to the same LdrpResolveProcedureAddress
function that GetProcAddress
uses; this is not true for the protected case. From glazing over some of the functions LdrResolveDelayLoadedAPI
calls, a module's LDR_DATA_TABLE_ENTRY.ReferenceCount
is atomically modified quite a bit. Setting a breakpoint on both LdrpHandleProtectedDelayload
and LdrpHandleUnprotectedDelayLoad
shows that a typical process will only ever call the former function (experimented with a ShellExecute
).
The above paragraph is only an informal glance; I or someone else must thoroughly look into delay loading in the modern Windows loader.
GNU Loader Lock Hierarchy
_dl_load_lock
is higher than _dl_load_write_lock
in the loader's lock hierarchy. There are occurrences where the GNU loader walks the list of link maps (a read operation) while only holding the _dl_load_lock
lock. For instance, walking the link map list while only holding _dl_load_lock
but not dl_load_write_lock
occurs in _dl_sym_find_caller_link_map
➜ _dl_find_dso_for_object
. This action is safe because the only occurrence where the loader modifies the list of link maps (following early loader startup) is while loading a library which means acquiring _dl_load_lock
. Linking a new link map into the list of link maps (a write operation) requires acquiring _dl_load_lock
then dl_load_write_lock
(e.g. see see the GDB log from the load-library
experiment). It's unsafe to modify the link map list without also acquiring _dl_load_write_lock
because, for instance, the __dl_iterate_phdr
function acquires _dl_load_write_lock
without first acquiring _dl_load_lock
to ensure the list of link maps remains consistent while the function walks between nodes (a read operation).
This lock hierarchy differs from the Windows closest equiavelnts of LdrpLoaderLock
and LdrpModuleDataTableLock
locks, where no hierarchy exists between them.
The _dl_load_lock
lock is the highest in the loader's lock hierarchy and is also above GSCOPE locking.
Loader Enclaves
An enclave is a security feature that isolates a region of data or code within an application's address space. Enclaves utilize one of three backing technologies to provide this security feature: Intel Software Guard Extensions (SGX), AMD Secure Encrypted Virtualization (SEV), or Virtualization-based Security (VBS). The Intel and AMD solutions are memory encryptors; they safeguard sensitive memory by encrypting it at the hardware level. VBS securely isolates sensitive memory by containing it in virtual secure mode where even the NT kernel cannot access it.
Within SGX, there is SGX1 and SGX2. SGX1 only allows using statically allocated memory with a set size before enclave initialization. SGX2 adds support for an enclave to allocate memory dynamically.
Due to this limitation, putting a library into an enclave on SGX1 requires that the library be statically linked. On the other hand, SGX2 supports dynamically loading/linking libraries.
A Windows VBS-based enclave requires Microsoft's signature as the root in the chain of trust or as a countersignature on a third party's certificate of an Authenticode-signed DLL. The signature must contain a specific extended key usage (EKU) value that permits running as an enclave. Enabling test signing on your system can get an unsigned enclave running. VBS enclaves may call enclave-compatible versions of Windows API functions from the enclave version of the same library.
Both Windows and Linux support loading libraries into enclaves. An enclave library requires special preparation; a programmer cannot load any generic library into an enclave.
Windows integrates enclaves as part of their loader. One may CreateEnclave
to make an enclave and then call LoadEnclaveImage to load a library into an enclave. Internally, CreateEnclave
(public API) calls ntdll!LdrCreateEnclave
, which then calls ntdll!NtCreateEnclave
; this function only does the system call to create an enclave, and then calls LdrpCreateSoftwareEnclave
to initialize and link the new enclave entry into the ntdll!LdrpEnclaveList
list. ntdll!LdrpEnclaveList
is the list head, its list entries of type LDR_SOFTWARE_ENCLAVE
are allocated onto the heap and linked into the list. Compiling an enclave library requires special compilation steps (e.g. compiling for Intel SGX).
The GNU loader has no knowledge of enclaves. Intel provides the Intel SGX SDK and the Intel SGX Platform Software (PSW) necessary for using SGX on Linux. The SGX driver awaits upstreaming into the Linux source tree. Linux currently has no equivalent to Windows Virtualization Based Security. Here's some sample code for loading an SGX module. However, Hypervisor-Enforced Kernel Integrity (Heki) is on its way, including patches to the kernel. Developers at Microsoft are introducing this new feature into Linux.
If you're interested, here's a comparison of Intel SGX and AMD SEV.
Open Enclave is a cross-platform and hardware-agnostic open source library for utilizing enclaves. Here's some sample code calling into an enclave library.
The Windows LdrpObtainLockedEnclave
Function
The Windows loader uses ntdll!LdrpObtainLockedEnclave
to obtain the enclave lock for a module, doing per-node locking. The LdrpObtainLockedEnclave
function acquires the ntdll!LdrpEnclaveListLock
lock and then searches from the ntdll!LdrpEnclaveList
list head to find an enclave for the DLL image base address this function receives as its first argument. If LdrpObtainLockedEnclave
finds a matching enclave, it atomically increments a reference count and enters a critical section stored in the enclave structure before returning that enclave's address to the caller. Typically (unless your process uses enclaves), the list at ntdll!LdrpEnclaveList
will be empty, effectively making LdrpObtainLockedEnclave
a no-op.
The LdrpObtainLockedEnclave
function came up in LdrGetProcedureAddressForCaller
(called by GetProcAddress
), so I wanted to give it a look.
Modern Windows Loader Synchronization: State > Locks
Me from the future: This section needs to be split up and revised. All of the facts and analysis are correct, but the thesis is weak. For instance, ntdll!LdrpWorkInProgress
is protected by ntdll!LdrpWorkQueueLock
; however, you can't know the state of ntdll!LdrpWorkInProgress
by testing that lock because the loader also uses that lock to protect the work queue. It's the same thing with loader lock in the following case study due to process exit.
In the modern Windows loader, there's a clear trend regarding how the Windows loader performs synchronization: State is more important than locks. Specifically, this state is more globalized than it is in a legacy Windows loader, which is likely, in large part, to support the parallel loader.
The loader primarily uses locks to protect the state. State, kept consistent through the use of locks, is what the loader predominantly uses to decide what actions it should perform. So, there's a level of indirection here.
Some good examples of such state include (see the relevant locks): LDR_DDAG_NODE.LDR_DDAG_STATE
, ntdll!LdrpWorkInProgress
, and LoadOwner
in TEB.SameTebFlags
)
There's one minor exception to this rule where the loader regularly depends on the locking status of a lock itself (minor because this lock isn't central to the loader state machine, and a typical process won't even run any DLL notification callbacks): ntdll!LdrpDllNotificationLock
In contrast, a legacy Windows loader relies solely on locks (predominantly just loader lock), locking and unlocking, and some local state (often passed as a function argument). Sometimes, also checking the state of loader lock directly to determine control flow.
Case Study
Let me show you what I mean because comparing the modern LdrGetProcedureAddressForCaller
with the old (ReactOS) equivalent function of LdrpGetProcedureAddress
is a textbook example of this. GetProcAddress
is the public Windows API function that calls through to these NTDLL exports.
Internally, when GetProcAddress
resolves an exported procedure to an address, it checks if the loader has performed initialization on the containing module. If the module requires initialization, then GetProcAddress
initializes the module (i.e. calling its DllMain
) before returning a procedure address to the caller.
UPDATE: GetProcAddress
may perform module initialization to fix the issue of someone doing GetModuleHandle
(not LoadLibrary
) ➜ GetProcAddress
on a partially loaded library. The GetModuleHandle
function shouldn't exist (as it doesn't in POSIX) but it does and so Microsoft had to implement their best hack to workaround this pitfall. A better solution might have been to make GetModuleHandle
not return the module's handle if the module isn't fully loaded. However, Microsoft clearly didn't choose that solution because I tested doing GetModuleHandle
on a partially loaded library from DllMain
(I verified the module I was getting a handle of was uninitialized in WinDbg) and it successfully gave me a handle to the uninitialzied library.
In the ReactOS code for LdrpGetProcedureAddress
, we see this happen:
/* Acquire lock unless we are initing */
/* MY COMMENT: This refers to the loader initing; ignore this for now */
if (!LdrpInLdrInit) RtlEnterCriticalSection(&LdrpLoaderLock);
...
/* Finally, see if we're supposed to run the init routines */
/* MY COMMENT: ExecuteInit is a function argument which GetProcAddress always passes as true */
if ((NT_SUCCESS(Status)) && (ExecuteInit))
{
/*
* It's possible a forwarded entry had us load the DLL. In that case,
* then we will call its DllMain. Use the last loaded DLL for this.
*/
Entry = NtCurrentPeb()->Ldr->InInitializationOrderModuleList.Blink;
LdrEntry = CONTAINING_RECORD(Entry,
LDR_DATA_TABLE_ENTRY,
InInitializationOrderLinks);
/* Make sure we didn't process it yet*/
/* MY COMMENT: If module initialization hasn't already run... */
if (!(LdrEntry->Flags & LDRP_ENTRY_PROCESSED))
{
/* Call the init routine */
_SEH2_TRY
{
Status = LdrpRunInitializeRoutines(NULL);
}
_SEH2_EXCEPT(EXCEPTION_EXECUTE_HANDLER)
{
/* Get the exception code */
Status = _SEH2_GetExceptionCode();
}
_SEH2_END;
}
}
...
This code is perfectly safe, as we know that even the legacy loader uses loader lock to protect against concurrent initialization/deinitialization or load/unload (amongst other things).
Now, let's see how the modern Windows loader handles this. In LdrGetProcedureAddressForCaller
, there's an instance where module initialization may occur without LdrGetProcedureAddressForCaller
itself acquiring loader lock (what follows is a marked-up IDA decompilation):
...
ReturnValue = LdrpResolveProcedureAddress(
(unsigned int)v24,
(unsigned int)Current_LDR_DATA_TABLE_ENTRY,
(unsigned int)Heap,
v38,
v30,
(char **)&v34);
...
// If LdrpResolveProcedureAddress succeeds
if ( ReturnValue >= 0 )
{
// Test for the searched module having a LDR_DDAG_STATE of LdrModulesReadyToInit
if ( Current_Module_LDR_DDAG_STATE == 7
// Test whether module module initialization should occur
// LdrGetProcedureAddressForCaller receives this as its fifth argument, KERNELBASE!GetProcAddresForCaller (called by KERNELBASE!GetProcAddress) always sets it to TRUE
&& ExecuteInit
// Test for the current thread having the LoadOwner flag in TEB.SameTebFlags
&& (NtCurrentTeb()->SameTebFlags & 0x1000) != 0
// Test for the current thread not holding the LdrpDllNotificationLock lock (i.e. we're not executing a DLL notification callback)
&& !(unsigned int)RtlIsCriticalSectionLockedByThread(&LdrpDllNotificationLock) )
{
DdagNode = *(_QWORD *)(Current_LDR_DATA_TABLE_ENTRY + 0x98);
v33[0] = 0;
// Perform module initialization
ReturnValue = LdrpInitializeGraphRecurse(DdagNode, 0i64, v33);
}
...
Huh? LdrGetProcedureAddressForCaller
didn't acquire loader lock, yet it is performing module initialization! How could that be safe?
It's a layer of abstraction. By testing the module's LDR_DDAG_STATE
, LoadOwner
in TEB.SameTebFlags
, ntdll!LdrpDllNotificationLock
, and ExecuteInit
states, LdrGetProcedureAddressForCaller
knows that the only way the loader would have let it reach this point where all these conditions pass is if it were safe to do module initialization. LdrGetProcedureAddressForCaller
doesn't have to be aware of loader lock because it trusts the loader state machine to have done the right thing before handing off execution to LdrGetProcedureAddressForCaller
.
This premise holds up when considering where the loader must be for all conditions to pass. Two of the things the LdrpPrepareModuleForExecution
function does is run DLL notification callbacks (calls LdrpNotifyLoadOfGraph
) and then perform module initialization (calls LdrpInitializeGraphRecurse
). We know a module's LDR_DDAG_STATE
is set to LdrModulesReadyToInit
immediately before running any DLL notification callbacks. DLL notification callbacks may call into third-party code (anyone can register for a callback), and a thread executing a DLL notification callback holds ntdll!LdrpDllNotificationLock
as it runs. LoadOwner
in TEB.SameTebFlags
correlates with ntdll!LdrpWorkInProgress
; it's what a thread uses to inform itself that it's responsible for the work in progress. Lastly, ExecuteInit
is the fifth argument passed to LdrGetProcedureAddressForCaller
, so should NTDLL want to get the address of a procedure within a module while the conditions up to this point pass without doing module initialization, it could simply pass the ExecuteInit
argument as FALSE
. By process of elimination, this logically leaves module initialization (i.e. DllMain
) as the only place this code path of LdrGetProcedureAddressForCaller
could execute. And everybody knows that DllMain
is run under loader lock.
Therefore, by testing all these conditions, LdrGetProcedureAddressForCaller
knows that the loader will already have acquired loader lock if need be. Having passed these conditions, LdrGetProcedureAddressForCaller
could also recursively acquire loader lock itself. However, there's no point in expending the extra resources to do that when it's already known to be safe to perform module initialization now. Additionally, other than module initialization, a process acquires loader lock at its exit (RtlExitUserProcess
), so testing these conditions instead of loader lock ensures LdrGetProcedureAddressForCaller
doesn't initialize a module during this phase of the process.
Continue to the next section for information regarding that "if need be" part in the context of the wider loader.
Windows Loader Initialization Locking Requirements
On Windows, loader initialization includes initialization of the process, as well as loading and initializing DLLs. Once loader initialization is complete, the process can proceed with running the application.
Reading the ReactOS LdrGetProcedureAddressForCaller
code, you saw this line of code:
/* Acquire lock unless we are initing */
/* MY COMMENT: This refers to the loader initing */
if (!LdrpInLdrInit) RtlEnterCriticalSection(&LdrpLoaderLock);
The loader won't acquire loader lock during module initialization when the loader is initializing (at process startup). What's up with that? It's a performance optimization to forego locking during process startup.
Not acquiring loader lock here is safe because the legacy loader, like the modern loader, includes a mechanism for blocking new threads spawned into the process until loader initialization is complete. The legacy loader waits in the LdrpInit
function using a ntdll!LdrpProcessInitialized
spinlock and sleeping with ZwDelayExecution
/NtDelayExecution
. The LdrpInit
function sets LdrpInLdrInit
to FALSE
, initializes the process by calling LdrpInitializeProcess
, then upon returning, LdrpInit
sets LdrpInLdrInit
to TRUE
. Hence, during LdrpInLdrInit
, one can safely forgo acquiring loader lock.
The modern loader optimizes waiting by using the LdrpInitCompleteEvent
Win32 event instead of sleeping for a set time. The modern loader also includes a ntdll!LdrpProcessInitialized
spinlock. However, the loader may only spin on it until event creation (NtCreateEvent
), at which point the loader waits solely using that synchronization object. While ntdll!LdrInitState
is 0
, it's safe not to acquire any locks. This includes accessing shared module information data structures without acquiring ntdll!LdrpModuleDataTableLock
and performing module initialization/deinitialization. ntdll!LdrInitState
changes to 1
immediately after LdrpInitializeProcess
calls LdrpEnableParallelLoading
which creates the loader work threads. However, these loader work threads won't have any work yet, so it should still be safe not to acquire locks during this time. In practice, one can see that the loader may call the LdrpInitShimEngine
function before LdrpEnableParallelLoading
(not always; it happens when running Google Chrome under WinDbg). The LdrpInitShimEngine
function calls LdrpLoadShimEngine
, which does a whole bunch of typically unsafe actions like module initialization (calls LdrpInitalizeNode
directly and calls LdrpInitializeShimDllDependencies
, which in turn calls LdrpInitializeGraphRecurse
) without loader lock and walking the PEB_LDR_DATA.InLoadOrderModuleList
without acquiring ntdll!LdrpModuleDataTableLock
. Of course, all these actions are safe due to the unique circumstances of process initialization. Note that the shim initialization function may still acquire the LdrpDllNotificationLock
lock not for safety but because the loader branches on its state using the RtlIsCriticalSectionLockedByThread
function.
The modern loader explicitly checks ntdll!LdrInitState
to optionally perform locking as an optimization in a few places. Notably, the ntdll!RtlpxLookupFunctionTable
function (this is in connection to the ntdll!LdrpInvertedFunctionTable
data structure, which is something I haven't looked into yet) opts to perform no locking (ntdll!LdrpInvertedFunctionTableSRWLock
) before accessing shared data structures if ntdll!LdrInitState
does not equal 3 (i.e. loader initialization is done). Similarly, the ntdll!LdrLockLoaderLock
function only acquires loader lock if loader initialization is done.
There's a notable difference between how the legacy loader vs the modern loader performs library loading and initialization at process startup (and beyond). In LdrpInitializeProcess
, the legacy loader calls LdrpWalkImportDescriptor
to "walk the IAT and load all the DLLs" (mapping and snapping) in one big step then later calls LdrpRunInitializeRoutines
to initialize all the DLLs in one big step. In contrast, the modern LdrpInitializeProcess
loads (including mapping, snapping, and initializing) DLLs in small steps based on how libraries rely on each other according to the dependency graph. The new approach significantly decreases the risk that one DLL uses another DLL that has only been partially loaded (mapped and snapped but pending initialization). Even when the loader reenters itself, the modern loader initializes partially loaded dependencies left over from the previous load if the now newly loading module also depends on them (I've tested this). Therefore, the modern loader's attention to order of operations when loading libraries makes DllMain
much "safer" than the legacy loader.
Windows must wait on loader initialization due to its architecture of libraries providing core system functionality. If other threads are able to launch (e.g. from DllMain
) before finishing loading and initializing library dependencies then problems could arise. Since the process is already limited to a single thread, Microsoft takes advantage of this opportunity to also implement some performance optimizations that forego locking during loader initialization.
The GNU loader doesn't implement any such performance hack to forego locking on process startup. The absence of any such mechanism by the GNU loader enables threads to start and exit from within a global constructor at process startup. Therefore, the GNU loader is more flexible in this respect.
ELF Flat Symbol Namespace (GNU Namespaces and STB_GNU_UNIQUE
)
Windows EXE and MacOS Mach-O starting with OS X v10.1 executable formats store library symbols in a two-dimensional namespace; thus effectively making every library its own namespace.
On the other hand, the Linux ELF executable format specifies a flat namespace such that two libraries having the same symbol can collide. For example, if two libraries expose a malloc
function, then the dynamic linker won't be able to differentiate between them. As a result, the dynamic linker recognizes the first malloc
symbol definition it sees as the malloc
, ignoring any malloc
definitions that come later. This refers to cases of loading a library (dlopen
) with RTLD_GLOBAL
; with RTLD_LOCAL
, the newly loaded library's symbols aren't made available to other libraries in the process.
These namespace collisions have been the source of some bugs, and as a result, there have been workarounds to fix them. The most straightforward being: dlsym("MyLibrary.so", "malloc")
. However, there are plenty of cases where that doesn't cut it, so GNU devised some solutions of their own.
Let's talk about GNU loader namespaces. Since 2004, Glibc loader has supported a feature known as loader namespaces for separating symbols contained within a loaded library into a separate namespace. Creating a new namespace for a loading library requires calling dlmopen
(this is a GNU extension). Loader namespaces allow a programmer to isolate a module to its own namespace. There are various reasons GNU gives for why a developer might want to do this and why RTLD_LOCAL
isn't a substitute for this, mainly regarding loading an untrusted library with an unwieldy number of generic symbol names polluting your namespace. It's worth noting that there's a hard limit of 16 on the number of GNU loader namespaces a process can have. Some might say this is just a bandage patch around the core issue of flat ELF symbol namespaces, but it doesn't hurt to exist.
In 2009, the GNU loader received a new symbol type calledSTB_GNU_UNIQUE
. In the dl_lookup_x
function, determining the symbol type is the final step after successfully locating a symbol. We previously saw this occur when our symbol was determined to be type STB_GLOBAL
. STB_GNU_UNIQUE
is another one of those symbol types, and as the name implies, it's a GNU extension. The purpose of this new symbol type was to fix symbol collision problems in C++ by giving precedence to STB_GNU_UNIQUE
symbols even above RTLD_LOCAL
symbols within the same module. Binding symbols with the STB_GNU_UNIQUE
symbol type became how GCC compiles C++ binaries. Due to the global nature of STB_GNU_UNIQUE
symbols and their lack of a reference counting feature (this could impact performance), their usage in a library makes unloading impossible by design. On top of this, STB_GNU_UNIQUE
introduced significant bloat to the GNU loader by adding a special, dynamically allocated hash table known as the _ns_unique_sym_table
just for handling this one new symbol type, which, along with a lock for controlling access to this new data structure, is included in the global _rtld_global
. rtld_global
(run-time dynamic linker) being the structure that defines all variables global to the loader. There's also a performance hit because locating a STB_GNU_UNIQUE
symbol requires a lookup in two separate hash tables.
These workarounds prove that ELF symbol namespaces should have been two-dimensional long ago. Standards organizations could still modify the ELF standard to support two-dimensional namespaces (although it would likely require a new version or an ABI compatibility hack); that's what Apple did with Mac OS a long time ago.
What's the Difference Between Dynamic Linking, Static Linking, Dynamic Loading, Static Loading?
Loading and linking are two operations essential for program execution. A loader is responsible for placing executable modules into the program's memory for execution. Linking resolves dependencies between executable modules or stitches executable modules together to create a runnable program.
Dynamic linking resolves dependencies between executable modules. Dynamic linking typically occurs when a program load/run time; however, it can also occur later due to a lazy linking optimization.
Static linking is when a compiler (e.g. GCC, Clang, and MSVC) uses a linker (e.g. the ld
program on GNU/Linux or link.exe
program on MSVC) to stitch object files together into single executables. Linkers can also write information into a program about where a dynamic linker or loader can find its dependencies at process load/run time.
Dynamic loading refers to loading a library at run-time with dlopen
/LoadLibrary
.
Static loading refers to loading that occurs due to dependencies found during dynamic linking.
LoadLibrary
vs dlopen
Return Type
On Windows, LoadLibrary
returns an officially opaque HMODULE
, which is implemented as the base address of the loaded module.
In POSIX, dlopen
returns a symbol table handle. On my GNU/Linux system, this handle points to the object's own link_map
structure located in the heap. This handle is opaque, meaning you must not access the contents behind it directly (they could change between versions, and it's implementation dependent); instead, only pass this handle to other dl*
functions.
How cool is that? That's like if Windows serviced your LoadLibrary
request by handing you back the module's LDR_DATA_TABLE_ENTRY
!
Investigating COM Server Deadlock from DllMain
Trying to create an in-process COM server under loader lock fails deterministically. For instance, this code run from the DllMain
on DLL_PROCESS_ATTACH
will deadlock:
HRESULT hres;
HWND hwnd = GetDesktopWindow();
// Ensure valid LNK file with this CMD command:
// explorer "C:\ProgramData\Microsoft\Windows\Start Menu\Programs\Accessories\Notepad.lnk"
LPCSTR linkFilePath = "C:\\ProgramData\\Microsoft\\Windows\\Start Menu\\Programs\\Accessories\\Notepad.lnk";
WCHAR resolvedPath[MAX_PATH];
hres = CoInitializeEx(NULL, 0);
if (SUCCEEDED(hres)) {
// Implementation: https://learn.microsoft.com/en-us/windows/win32/shell/links#resolving-a-shortcut
ResolveIt(hwnd, linkFilePath, resolvedPath, MAX_PATH);
}
CoUninitialize();
Here's the deadlocked call stack showing that NtAlpcSendWaitReceivePort
is waiting for something (this function only exists to make the NtAlpcSendWaitReceivePort
system call):
0:000> k
# Child-SP RetAddr Call Site
00 00000080`f96fcde8 00007ffd`26c93f8f ntdll!NtAlpcSendWaitReceivePort+0x14
01 00000080`f96fcdf0 00007ffd`26ca94d7 RPCRT4!LRPC_BASE_CCALL::SendReceive+0x12f
02 00000080`f96fcec0 00007ffd`26c517c0 RPCRT4!NdrpSendReceive+0x97
03 00000080`f96fcef0 00007ffd`26c524bf RPCRT4!NdrpClientCall2+0x5d0
04 00000080`f96fd510 00007ffd`28491ce5 RPCRT4!NdrClientCall2+0x1f
05 (Inline Function) --------`-------- combase!ServerAllocateOXIDAndOIDs+0x73 [onecore\com\combase\idl\internal\daytona\objfre\amd64\lclor_c.c @ 313]
06 00000080`f96fd540 00007ffd`28491acd combase!CRpcResolver::ServerRegisterOXID+0xd5 [onecore\com\combase\dcomrem\resolver.cxx @ 1056]
07 00000080`f96fd600 00007ffd`28494531 combase!OXIDEntry::RegisterOXIDAndOIDs+0x71 [onecore\com\combase\dcomrem\ipidtbl.cxx @ 1642]
08 (Inline Function) --------`-------- combase!OXIDEntry::AllocOIDs+0xc2 [onecore\com\combase\dcomrem\ipidtbl.cxx @ 1696]
09 00000080`f96fd710 00007ffd`2849438f combase!CComApartment::CallTheResolver+0x14d [onecore\com\combase\dcomrem\aprtmnt.cxx @ 693]
0a 00000080`f96fd8c0 00007ffd`284abc2f combase!CComApartment::InitRemoting+0x25b [onecore\com\combase\dcomrem\aprtmnt.cxx @ 991]
0b (Inline Function) --------`-------- combase!CComApartment::StartServer+0x52 [onecore\com\combase\dcomrem\aprtmnt.cxx @ 1214]
0c 00000080`f96fd930 00007ffd`2849c285 combase!InitChannelIfNecessary+0xbf [onecore\com\combase\dcomrem\channelb.cxx @ 1028]
0d 00000080`f96fd960 00007ffd`2849a644 combase!CGIPTable::RegisterInterfaceInGlobalHlp+0x61 [onecore\com\combase\dcomrem\giptbl.cxx @ 815]
0e 00000080`f96fda10 00007ffd`21b86399 combase!CGIPTable::RegisterInterfaceInGlobal+0x14 [onecore\com\combase\dcomrem\giptbl.cxx @ 776]
0f 00000080`f96fda50 00007ffd`21b5adb3 PROPSYS!CApartmentLocalObject::_RegisterInterfaceInGIT+0x81
10 00000080`f96fda90 00007ffd`21b842e6 PROPSYS!CApartmentLocalObject::_SetApartmentObject+0x7b
11 00000080`f96fdac0 00007ffd`21b5c1fc PROPSYS!CApartmentLocalObject::TrySetApartmentObject+0x4e
12 00000080`f96fdaf0 00007ffd`21b5bde6 PROPSYS!CreateObjectWithCachedFactory+0x2bc
13 00000080`f96fdbd0 00007ffd`21b5d16c PROPSYS!CreateMultiplexPropertyStore+0x46
14 00000080`f96fdc30 00007ffd`241d3235 PROPSYS!PSCreateItemStoresFromDelegate+0xbfc
15 00000080`f96fde90 00007ffd`2422892f windows_storage!CShellItem::_GetPropertyStoreWorker+0x2d5
16 00000080`f96fe3d0 00007ffd`2422b7e7 windows_storage!CShellItem::GetPropertyStoreForKeys+0x14f
17 00000080`f96fe6a0 00007ffd`2415f2b6 windows_storage!CShellItem::GetCLSID+0x67
18 00000080`f96fe760 00007ffd`2415eb0b windows_storage!GetParentNamespaceCLSID+0xde
19 00000080`f96fe7c0 00007ffd`241772fb windows_storage!CShellLink::_LoadFromStream+0x2d3
1a 00000080`f96feaf0 00007ffd`2417709c windows_storage!CShellLink::LoadFromPathHelper+0x97
1b 00000080`f96feb40 00007ffd`24177039 windows_storage!CShellLink::_LoadFromFile+0x48
1c 00000080`f96febd0 00007ffd`21aa10e2 windows_storage!CShellLink::Load+0x29
1d (Inline Function) --------`-------- TestDLL!ResolveIt+0x8c [C:\Users\user\source\repos\TestDLL\TestDLL\dllmain.cpp @ 110]
1e 00000080`f96fec00 00007ffd`21aa143b TestDLL!DllMain+0xd2 [C:\Users\user\source\repos\TestDLL\TestDLL\dllmain.cpp @ 170]
1f 00000080`f96ff4f0 00007ffd`28929a1d TestDLL!dllmain_dispatch+0x8f [d:\a01\_work\20\s\src\vctools\crt\vcstartup\src\startup\dll_dllmain.cpp @ 281]
20 00000080`f96ff550 00007ffd`2897c2c7 ntdll!LdrpCallInitRoutine+0x61
21 00000080`f96ff5c0 00007ffd`2897c05a ntdll!LdrpInitializeNode+0x1d3
22 00000080`f96ff710 00007ffd`2894d947 ntdll!LdrpInitializeGraphRecurse+0x42
23 00000080`f96ff750 00007ffd`2892fbae ntdll!LdrpPrepareModuleForExecution+0xbf
24 00000080`f96ff790 00007ffd`289273e4 ntdll!LdrpLoadDllInternal+0x19a
25 00000080`f96ff810 00007ffd`28926af4 ntdll!LdrpLoadDll+0xa8
26 00000080`f96ff9c0 00007ffd`260156b2 ntdll!LdrLoadDll+0xe4
27 00000080`f96ffab0 00007ff7`8fda1022 KERNELBASE!LoadLibraryExW+0x162
28 00000080`f96ffb20 00007ff7`8fda1260 EmptyProject!main+0x12 [C:\Users\user\source\repos\EmptyProject\EmptyProject\source.c @ 82]
29 (Inline Function) --------`-------- EmptyProject!invoke_main+0x22 [d:\a01\_work\20\s\src\vctools\crt\vcstartup\src\startup\exe_common.inl @ 78]
2a 00000080`f96ffb50 00007ffd`26f37344 EmptyProject!__scrt_common_main_seh+0x10c [d:\a01\_work\20\s\src\vctools\crt\vcstartup\src\startup\exe_common.inl @ 288]
2b 00000080`f96ffb90 00007ffd`289626b1 KERNEL32!BaseThreadInitThunk+0x14
2c 00000080`f96ffbc0 00000000`00000000 ntdll!RtlUserThreadStart+0x21
The NtAlpcSendWaitReceivePort
function is indeed waiting for something, thus the reason for our deadlock.
Note that the deadlock occurs later on the first call to a COM method (hres = ppf->Load(wsz, STGM_READ)
), not when we initially call CoCreateInstance
to create an in-process COM server (CLSCTX_INPROC_SERVER
). The deadlock occurs here because a COM server's startup is lazy loading. Lazy loading occurs for the most common CLSCTX_INPROC_SERVER
execution context, but not necessarily for other execution contexts such as CLSCTX_LOCAL_SERVER
(the deadlock occurs in the CoCreateInstance
function itself in the latter execution contect).
Notice this call stack is similar to the deadlock we receive from ShellExecute
under DllMain
. As we know from "Perfect DLL Hijacking", loader lock is the root cause of this deadlock and releasing this lock allows execution to continue (also, see the DEBUG NOTICE
in the LdrLockLiberator project for further potential blockers). However, setting a read watchpoint on loader lock reveals that user-mode COM internals never check the state of the loader lock. This finding leads me to believe that the local procedure call by ntdll!NtAlpcSendWaitReceivePort
causes a remote process (csrss.exe
?) to introspect on the state of loader lock within our process, potentially explaining why WinDbg never hits our watchpoint. Starting a COM server (combase!CComApartment::StartServer
) involves calling NdrClientCall2
to perform a local procedure call (technically "LRPC" which is RPC done locally). COM must support real RPC (to another machine) in the case of DCOM, which was eventually integrated into COM (this is likely what combase!CComApartment::InitRemoting
in the call stack refers to). In particular, the combase!ServerAllocateOXIDAndOID
function does RPC according to the lclor
interface.
Microsoft explains this deadlock (emphasis mine):
To do it, the Windows loader uses a process-global critical section (often called the "loader lock") that prevents unsafe access during module initialization.
The CLR will attempt to automatically load that assembly, which may require the Windows loader to block on the loader lock. A deadlock occurs since the loader lock is already held by code earlier in the call sequence.
Note that Microsoft wrote this documentation to resolve loader lock issues high-level developers may come across with the CLR (the .NET runtime environment); however, it appears that it also roughly applies to COM (where the deadlocked in the call stack looks similar). Microsoft likely takes this step to mitigate the danger of a partially initialized module or partially loaded (fully uninitialized) modules. Given Microsoft's architecture of providing functionality through DLLs and the fact COM may specially integrate with a DLL (a COM DLL typically exports DLLRegisterServer
and DLLUnregisterServer
functions), this answer appears sensible. Additionally, I reason that initializing a COM object, like a CLR assembly, may take "actions that are invalid under loader" (e.g. spawning and waiting on a thread deadlocks on Windows). So, instead of allowing for "non-determinism", Microsoft took steps to make starting an in-process COM server from DllMain
deadlock deterministically.
Emphasis on the word "block" because it indicates that, in this case, Windows treats loader lock as a readers-writer lock (SRW lock in Windows) that's been acquired in exclusive/write mode instead of a critical section (a thread synchronization mechanism), the latter of which allows recursive acquisition (reacquiring a lock on the same thread; doing this action safely requires the surrounding code to have a reentrant design). This fact aligns with what we see when starting a COM server from DllMain
.
While Microsoft has designed the loader to be reentrant (e.g. to support loading a library from another library's DllMain
; although it's not best practice), there are still other unfavourable properties, which we have covered, about performing ceratin actions under loader lock. Thus, explaining why Microsoft would prefer to deadlock deterministically here (even when loading a non-COM DLL) rather than potentially get a hard to disagnose and debug deadlock/crash later.
When building a .NET project, compiling a binary with the /clr
option causes the MSVC compiler to put global constructors in the .cctor
section. Now, the high-level CLR loader will perform initialization of the global consturctors/destructors instead of the native loader, thus working around "loader lock issues" that are symptomatic of Windows architecture. Note that DllMain
is still run by the native loader under loader lock.
Please note that the above analysis is a strong theory based on the surrounding evidence and my work. However, verifying it with full certainty would require checking whether the remote process is checking the loader lock of our process. I want to figure out how to debug this; please do contribute if you know.
On Making COM from DllMain
Safe
In this section will analyze why doing COM initialization or start a COM server from DllMain
is currently unsafe and more interestingly, how these actions could be made safe.
Avoiding ABBA Deadlock
In the documentation for CoInitializeEx
lies this statement:
Because there is no way to control the order in which in-process servers are loaded or unloaded, do not call CoInitialize, CoInitializeEx, or CoUninitialize from the DllMain function.
This warning concerns the risk of nesting the native (e.g. ntdll!LdrpLoaderLock
) and COM (e.g. combase!gComLock
, combase!gChannelInitLock
, and combase!g<a href="https://devblogs.microsoft.com/oldnewthing/20210108-00/?p=104684" target="_blank">OXID</a>Lock
) lock hierarchies within a thread. Realizing this risk would require at least two threads acquiring the locks in different orders to interleave, thus causing an ABBA deadlock. Ideally, Microsoft would support this by ensuring a clean lock hierarchy between these subsystems or controlling load/unload order (it's easier said than done given the monolithic nature of it all). Microsoft has realized this risk in their code (note that starting with OLE 2, OLE is built on top of COM, so they share internals). The simplest method for fixing this ABBA deadlock would be to acquire the loader lock (it's in the PEB) before the COM lock, thereby keeping a clean lock hierarchy. However, this method would decrease concurrency, thus reducing performance. Another solution would be to specifically target COM functions that interact with the native loader while already holding the COM lock, such as CoFreeUnusedLibraries
. CoFreeUnusedLibraries
would leave the COM shared data structures/components in a consistent state before unlocking the COM lock, acquire loader lock, and then perform consistency checks after reacquiring the COM lock. COM architecture might precisely track each component's state to support its reentrant design. A state tracking mechanism could work like LDR_DDAG_STATE
in the native loader. CoFreeUnusedLibraries
will acquire loader lock followed by the COM lock(s) and then perform its task of freeing unused libraries. My inspiration for the second approach partially came from the GNU loader. In both solutions, the process maintains a single coherent lock hierarchy. Thus, either of these solutions would solve the issue potential ABBA deadlock issue upon COM performing a CoFreeUnusedLibraries
when combined with some module's DllMain
interacting with COM at the same time.
Given how monolithic Windows tends to be, the second solution for ABBA deadlock would be challenging; however, concurrency is hardly ever easy, and Microsoft engineers have taken on more complex problems. It's possible to make it work. The monolithic architecture of Windows is often what unexpectedly causes problems between these separate lock hierarchies to arise in the first place (e.g. due to a Windows API function internally doing COM without the programmer's explicit knowledge).
Other Deadlock Possibilities
COM may perform other operations that cause a deadlock:
If some COM operation requires spawning and waiting on a new thread, then spawn that thread with the THREAD_CREATE_FLAGS_SKIP_LOADER_INIT
flag (or remove loader thread blockers). I think skipping/removing thread blockers would be the only way to avoid deadlock here when accounting for both DLL_THREAD_ATTACH
and DLL_THREAD_DETACH
(i.e. we can't temporarily offload some loading operations or otherwise, just for thread startup).
If some COM operation further requires that waited on new thread to load additional libraries (Is this something that can happen? As noted in our concurrency experiments, it's a naturally tricky situation that even the GNU loader will deadlock under.) then things get hard (if not impossible). As a thought experiment though, we will give it our best shot. We are currently on "thread A" and the thread we just spawned is "thread B". Thread A is already the load owner. So, our approach to avoiding deadlock should be to attempt offloading library loading work from thread B to thread A (thread A would have to wait to see if it picks up any work) or just have thread A wait so thread B can act as the load owner for a bit. We will go for the second option. After thread A does the system call to spawn thread B, thread A (in NtCreateThreadEx
) should check if it's under loader lock. If so, thread A must wait until it receives a signal from thread B that it's completed its loader work. Thread B should assign its thread as the load owner (LoadOwner
flag in TEB.SameTebFlags
; the global LdrpWorkInProgress
state is already 1
) at the same time as thread A is already the load owner. With thread A waiting, thread B can now safely load libraries as normal and this works because the loader is reentrant and due to all the loader data structures being global. When thread B finishes loader work, it can signal thread A to proceed.
As with any approach, there are glaring issues with this method to avoiding deadlock. The first issue is that thread B cannot know when it will be done with loader work (if it is to do any; assuming we already removed loader thread blockers) because the loader cannot forsee dynamic loading operations (LoadLibrary
/FreeLibrary
or delay loading) ahead of time. Some Unix-like loader implementations don't support dynamic loading due to the deadlock problems that can occur when libraries load unexpectedly and the performance benefits you can achieve from not having to program for that scenario. The ideal situation is always to load all a program's dependencies once at process startup then never load/unload libraries again for the lifetime of the process. This issue is already a showstopper, but in the name of education, we will continue on. The second issue comes from the fact that thread A is in a DLL's DllMain
performing its initialization. So, complex operations in general could be problematic because a DLL spawning a thread from its DllMain
may itself be partially initialized or the DLL spawning a thread could depend on other DLLs which are still completely uninitialized (the latter could only be due to circular dependencies on a modern loader). The third issue is that waiting on a thread for any amount of time when not explicitly told to (WaitForSingleObject(hThread)
) leads to poor performance. There are probably more issues.
Having a dedicated load owner thread at all times (similar to the current loader worker threads), provided suitable optimizations are in place, is likely the most realistic and workable solution. Although, Microsoft would have to create a mechanism for threaded operations (e.g. FlsAlloc
, FlsGetValue
, and FlsSetValue
functions) done at DLL_THREAD_ATTACH
and DLL_THREAD_DETACH
to affect and access the thread requesting the DLL_THREAD_ATTACH
/DLL_THREAD_DETACH
instead of the dedicated load owner thread itself (also, the GetCurrentThread
and GetCurrentThreadId
functions may have to lie to maintain application compatibility).
Conclusion
DllMain
can now safely run COM code (in theory).
Note that applying this solution to the CLR may mean removing the CLR loader because its continued presence means foreign (third-party) code could violate the loader lock ➜ CLR assembly initialization lock (I'm assuming the CLR loader has an assembly initialization lock similar to the native loader's loader lock) lock hierarchy without our knowledge.
The other solution is what most Microsoft engineers currently advise, which is to write a rule saying that nobody should do COM (e.g. CoInitializeEx
) from DllMain
, which works great until people, such as Microsoft employees, do COM from DllMain
(or one of many other places the loader unexpectedly holds loader lock). Additionally, in the case of CLR, requiring Microsoft creates a loader on top of a loader (the CLR loader) so "unmanaged and managed initialization is done in two separate and distinct stages" over the native loader to workaround "loader lock issues" that are symptomatic of Windows architecture.
Note that the solutions I provide above are mostly thought experiments (especially when it comes to launching an in-process COM server). Due to Windows' architectural issues, I can only offer bandage patches.
Case Study on ABBA Deadlock Between the Loader Lock and GDI+ Lock
This case study is not regarding code that Microsoft engineers wrote but of a third-party shell extension that this Old New Thing article reports on. Here we have the ABBA deadlock between loader lock and GDI+ lock. The ABBA deadlock occurs more inadvertently than in the previously shown COM case study because of the GdiPlus!GdiplusShutdown
thread waiting for a worker thread that holds the loader lock (so, there's a layer of indirection). Microsoft documentation and this Old New Thing article says the root cause of this problem is running GdiPlus!GdiplusStartup
from DllMain
. However, my opinion on the root cause of this problem leans more on the architectural side of things, and it's that each library's DllMain
receives DLL_THREAD_ATTACH
and DLL_THREAD_DETACH
notifications in the first place. As noted elsewhere in this document, these notifications are effectively useless, and well-written libraries often disable them to improve performance. So, it would appear that DLL thread notifications come with more disadvantages than advantages. I'm also unaware of any other operating system requiring such notifications.
In this scenario, if Windows could refrain from the practice of "DLLs creating worker threads behind your back", then that would also nullify the issue (although in the GDI+ case, GUI programs typically require at least one spare thread for the message/event loop, the Windows API shouldn't create these threads without the programmer's knowledge).
LdrpDrainWorkQueue
and LdrpProcessWork
The high-level LdrpDrainWorkQueue
and LdrpProcessWork
parallel loader functions work hand-in-hand to perform module mapping and snapping.
The LdrpDrainWorkQueue
function empties the work queue by processing all work in the work queue. LdrpDrainWorkQueue
iterates over the ntdll!LdrpWorkQueue
linked list, calling LdrpProcessWork
on each work item to achieve its goal.
Processing a work item means doing mapping or snapping work on it, or both. LdrpProcessWork
checks a module's LDR_DDAG_STATE
. If it's zero (i.e. LdrModulesPlaceHolder
), the module is unmapped, so LdrpProcessWork
maps it; otherwise, LdrpProcessWork
snaps it. In the mapping and snapping scenario, LdrpProcessWork
calls LdrpMapDllFullPath
or LdrpMapDllSearchPath
. These mapping functions internally call through to the bloated (in my opinion, because it's doing things beyond what its name suggests) LdrpMapDllWithSectionHandle
function. After mapping, LdrpMapDllWithSectionHandle
decides whether also to do snapping by checking what LdrpFindLoadedDllByMappingLockHeld
outputted for its fourth (out
) argument (more research required). If this fourth argument outputs a non-zero value, LdrpMapDllWithSectionHandle
calls LdrpLoadContextReplaceModule
to enqueue the work (LdrpQueueWork
) to the work queue; otherwise, LdrpMapDllWithSectionHandle
adds the mapped module into the module information data structures (LdrpInsertDataTableEntry
and LdrpInsertModuleToIndexLockHeld
) before snapping. The loader can add a module's LDRP_LOAD_CONTEXT
(undocumented structure; allocated by the LdrpAllocatePlaceHolder
function) to the work queue before that module's LDR_DATA_TABLE_ENTRY
has even been added to any of the module information data structures.
A work item refers to one LDRP_LOAD_CONTEXT
entry (undocumented structure) in the loader work queue (ntdll!LdrpWorkQueue
) linked list. Each work item relates directly to one module because they're allocated at the same time with the LdrpAllocatePlaceHolder
function and each LDRP_LOAD_CONTEXT
contains that a UNICODE_STRING
of its own DLL's BaseDllName
as its first member. However, processing a single work item for one module can still indirectly cause other modules to process within the same call to LdrpProcessWork
because LdrpMapDllWithSectionHandle
may call LdrpMapAndSnapDependency
to process dependencies. Depending on the LDRP_LOAD_CONTEXT
structure contents (undocumented), LdrpMapAndSnapDependency
may process the further work itself or outsource that work by enqueuing it (LdrpQueueWork
) to the work queue (ntdll!LdrpWorkQueue
) where loader worker (LoaderWorker
flag in TEB.SameTebFlags
) threads can process the work. A parallel loader worker threads receive a callback (thread pool internals) to its LdrpWorkCallback
entrypoint which then calls LdrpProcessWork
to pick up some mapping and snapping operations from the work queue as a performance optimization.
LdrpLoadCompleteEvent
and LdrpWorkCompleteEvent
When the loader sets the LdrpLoadCompleteEvent
event, it's signalling that the loader is reemerging from completing all mapping and snapping work for the entire work queue. When this occurs, it directly correlates with LdrpWorkInProgress
equalling zero and the decommissioning of the current thread as the load owner (LoadOwner
flag in TEB.SameTebFlags
). Here's a minimal reverse engineering of the ntdll!LdrpDropLastInProgressCount
function showing this:
NTSTATUS LdrpDropLastInProgressCount()
{
// Remove thread's load owner flag
PTEB CurrentTeb = NtCurrentTeb();
CurrentTeb->SameTebFlags &= ~LoadOwner; // 0x1000
// Load/unload is now complete
RtlEnterCriticalSection(&LdrpWorkQueueLock);
LdrpWorkInProgress = 0;
RtlLeaveCriticalSection(&LdrpWorkQueueLock);
// Signal completion of load/unload to any waiting threads
return NtSetEvent(LdrpLoadCompleteEvent, NULL);
}
When the loader sets the LdrpWorkCompleteEvent
event, it's signalling that the loader has completed the mapping and snapping work on the entire work queue including across all of the currently processing loader worker (LoaderWorker
flag in TEB.SameTebFlags
) threads. Here's a minimal reverse enginering of where the ntdll!LdrpProcessWork
function returns showing this:
// Second argument of LdrpProcessWork: isCurrentThreadLoadOwner
// If the current thread is a loader worker (i.e. not a load owner)
if (!isCurrentThreadLoadOwner)
{
RtlEnterCriticalSection(&LdrpWorkQueueLock);
// If the work queue is empty AND we we are the last loader worker thread processing work
// There were some double negatives I had to sort out here in the reverse engineering
BOOL doSetEvent = &LdrpWorkQueue == LdrpWorkQueue.Flink && --LdrpWorkInProgress == 1
Status = RtlLeaveCriticalSection(&LdrpWorkQueueLock);
if ( doSetEvent )
return NtSetEvent(LdrpWorkCompleteEvent, NULL);
}
return Status;
Here are all the loader's usages of LdrpLoadCompleteEvent
and LdrpWorkCompleteEvent
:
0:000> # "ntdll!LdrpLoadCompleteEvent" <NTDLL ADDRESS> L9999999
ntdll!LdrpDropLastInProgressCount+0x38:
00007ffd`2896d9c4 488b0db5e91000 mov rcx,qword ptr [ntdll!LdrpLoadCompleteEvent (00007ffd`28a7c380)]
ntdll!LdrpDrainWorkQueue+0x2d:
00007ffd`2896ea01 4c0f443577d91000 cmove r14,qword ptr [ntdll!LdrpLoadCompleteEvent (00007ffd`28a7c380)]
ntdll!LdrpCreateLoaderEvents+0x12:
00007ffd`2898e182 488d0df7e10e00 lea rcx,[ntdll!LdrpLoadCompleteEvent (00007ffd`28a7c380)]
0:000> # "ntdll!LdrpWorkCompleteEvent" <NTDLL ADDRESS> L9999999
ntdll!LdrpDrainWorkQueue+0x18:
00007ffd`2896e9ec 4c8b35bdd91000 mov r14,qword ptr [ntdll!LdrpWorkCompleteEvent (00007ffd`28a7c3b0)]
ntdll!LdrpProcessWork+0x1e4:
00007ffd`2896ede0 488b0dc9d51000 mov rcx,qword ptr [ntdll!LdrpWorkCompleteEvent (00007ffd`28a7c3b0)]
ntdll!LdrpCreateLoaderEvents+0x35:
00007ffd`2898e1a5 488d0d04e20e00 lea rcx,[ntdll!LdrpWorkCompleteEvent (00007ffd`28a7c3b0)]
ntdll!LdrpProcessWork$fin$0+0x7c:
00007ffd`289b5ad7 488b0dd2680c00 mov rcx,qword ptr [ntdll!LdrpWorkCompleteEvent (00007ffd`28a7c3b0)]
The LdrpCreateLoaderEvents
function creates both events. The LdrpDrainWorkQueue
function may wait on either LdrpLoadCompleteEvent
or LdrpWorkCompleteEvent
. Only the LdrpDropLastInProgressCount
function sets LdrpLoadCompleteEvent
. Only the LdrpProcessWork
function sets LdrpWorkCompleteEvent
.
At event creation (NtCreateEvent
), LdrpLoadCompleteEvent
and LdrpWorkCompleteEvent
are configured to be auto-reset events:
Auto-reset event objects automatically change from signaled to nonsignaled after a single waiting thread is released.
The LdrpDrainWorkQueue
function calls NtWaitForSingleObject
on these auto-reset events. The loader never manually resets the LdrpLoadCompleteEvent
and LdrpWorkCompleteEvent
events (NtResetEvent
).
The LdrpDrainWorkQueue
function takes one argument. This argument is a boolean, indicating whether the function should wait on LdrpLoadCompleteEvent
or LdrpWorkCompleteEvent
loader event before draining the work queue. Please see my reverse engineering of the LdrpDrainWorkQueue
function.
The loader LdrpLoadCompleteEvent
or LdrpWorkCompleteEvent
events synchronize when the LdrpDrainWorkQueue
function should do a loop around (the inner infinite while ( TRUE )
loop) to try draining work. If the LdrpLoadCompleteEvent
and LdrpWorkCompleteEvent
were always set, the loader would never crash. However, it would effectively turn the entire inner while ( TRUE )
loop of the LdrpDrainWorkQueue
function into a spinlock that's constantly acquiring the LdrpWorkQueueLock
critical section lock (this lock protects the relevant shared state) to see if there's new work to drain while testing on CompleteRetryOrReturn
to know if the code should break out of the loop. I've tested this myself by permanently setting both events (just close their handles and then recreate them as events which don't auto-reset) thereby ensuring the loader never crashes without these events; your CPU usage just goes very high due to busy waiting instead of waiting.
What follows documents what parts of the loader call LdrpDrainWorkQueue
while synchronizing on either the LdrpLoadCompleteEvent
or LdrpWorkCompleteEvent
loader event (data gathered by searching disassembly for calls to the LdrpDrainWorkQueue
function):
ntdll!LdrUnloadDll+0x80: FALSE
ntdll!RtlQueryInformationActivationContext+0x43c: FALSE
ntdll!LdrShutdownThread+0x98: FALSE
ntdll!LdrpInitializeThread+0x86: FALSE
ntdll!LdrpLoadDllInternal+0xbe: FALSE
ntdll!LdrpLoadDllInternal+0x144: TRUE
ntdll!LdrpLoadDllInternal$fin$0+0x38: TRUE
ntdll!LdrGetProcedureAddressForCaller+0x270: FALSE
ntdll!LdrEnumerateLoadedModules+0xa7: FALSE
ntdll!RtlExitUserProcess+0x23: TRUE OR FALSE
- Depends on `TEB.SameTebFlags`, typically FALSE if `LoadOwner` or `LoadWorker` flags are absent, TRUE if either of these flags is present
ntdll!RtlPrepareForProcessCloning+0x23: FALSE
ntdll!LdrpFindLoadedDll+0x9127a: FALSE
ntdll!LdrpFastpthReloadedDll+0x9033a: FALSE
ntdll!LdrpInitializeImportRedirection+0x46d44: FALSE
ntdll!LdrInitShimEngineDynamic+0x3c: FALSE
ntdll!LdrpInitializeProcess+0x130a: FALSE
ntdll!LdrpInitializeProcess+0x1d0d: FALSE
ntdll!LdrpInitializeProcess+0x1e22: TRUE
ntdll!LdrpInitializeProcess+0x1f33: FALSE
ntdll!RtlCloneUserProcess+0x71: FALSE
Notably, there are many more instances of synchronizing on the entire load's completion rather than just the completion of mapping and snapping work. For example, thread initialization (LdrpInitializeThread
) synchronizes on the LdrpLoadCompleteEvent
loader event. The only functions that may synchronize on LdrpWorkCompleteEvent
are LdrpLoadDllInternal
(the public LoadLibrary
function internally calls this), LdrpInitializeProcess
, and RtlExitUserProcess
.
Comparing the loader's number of calls to LdrpDrainWorkQueue
in contrast to directly calling LdrpProcessWork
, we see the former is considerably more prevalent:
0:000> # "call ntdll!LdrpProcessWork" <NTDLL ADDRESS> L9999999
ntdll!LdrpLoadDependentModule+0x184c:
00007ffb`1ca8942c e8cb570400 call ntdll!LdrpProcessWork (00007ffb`1cacebfc)
ntdll!LdrpLoadDllInternal+0x13a:
00007ffb`1ca8fb4e e8a9f00300 call ntdll!LdrpProcessWork (00007ffb`1cacebfc)
ntdll!LdrpDrainWorkQueue+0x17f:
00007ffb`1caceb53 e8a4000000 call ntdll!LdrpProcessWork (00007ffb`1cacebfc)
ntdll!LdrpWorkCallback+0x6e:
00007ffb`1cacebde e819000000 call ntdll!LdrpProcessWork (00007ffb`1cacebfc)
There are only three occurrences of the loader directly calling LdrpProcessWork
(not including the call from LdrpDrainWorkQueue
) versus the twenty times it calls LdrpDrainWorkQueue
.
When a Process Would Rather Terminate Then Wait on a Critical Section
In the documentation for EnterCriticalSection
is this interesting snippet:
While a process is exiting, if a call to EnterCriticalSection would block, it will instead terminate the process immediately. This may cause global destructors to not be called.
Curious to see how Microsoft implemented this, I found that RtlpWaitOnCriticalSection
checks if PEB_LDR_DATA.ShutdownInProgress
is true. If so, the function jumps to calling NtTerminateProcess
passing -1
as the first argument, thereby forcefully terminating the process.
PEB_LDR_DATA.ShutdownInProgress
impacts several things, including FLS callbacks at process exit, TLS destructors at process exit, and DLL_PROCESS_DETACH
notifications to DllMain
at process exit.
The LdrShutdownProcess
function (called by RtlExitUserProcess
) sets PEB_LDR_DATA.ShutdownInProgress
to true. Since RtlExitUserProcess
acquires loader lock before setting PEB_LDR_DATA.ShutdownInProgress
, code run past this point also runs under loader lock.
ABBA Deadlock
An ABBA deadlock is a deadlock due to lock order inversion. An ABBA deadlock may occur due to a lock order inversion ("lock hierarchy violation") in a single lock hierarchy. Or, an ABBA deadlock can occur due to the more complex case of nesting separate lock hierarchies within a thread (i.e. lock order inversion between lock hierarchies). ABBA deadlock is a general term that isn't tied to lock hierarchies. The latter case can be tricky because there may not necessarily be a defined order for correctly acquiring the locks between separate lock hierarchies.
Whether lock order inversion results in an ABBA deadlock is probabilistic because it depends on whether at least two threads interleave while acquiring the locks in a different order.
Microsoft refers to an ABBA deadlock as a "deadlock caused by lock order inversion". These are synonyms. However, I prefer the more concise term common throughout the Linux ecosystem.
Although the ABBA deadlock and the ABA problem names may sound similar, they have entirely different meanings, so be sure to distinguish between them.
ABA Problem
The ABA problem ABA problem is a lower-level concept that can cause data structure inconsistency in lock-free code (i.e. code that relies entirely on atomic assembly instructions to ensure data structure consistency):
In multithreaded computing, the ABA problem occurs during synchronization, when a location is read twice, has the same value for both reads, and the read value being the same twice is used to conclude that nothing has happened in the interim; however, another thread can execute between the two reads and change the value, do other work, then change the value back, thus fooling the first thread into thinking nothing has changed even though the second thread did work that violates that assumption.
A common case of the ABA problem is encountered when implementing a lock-free data structure. If an item is removed from the list, deleted, and then a new item is allocated and added to the list, it is common for the allocated object to be at the same location as the deleted object due to MRU memory allocation. A pointer to the new item is thus often equal to a pointer to the old item, causing an ABA problem.
This example shows how atomic compare-and-swap instructions (lock cmpxchg
on x86) could mix up list items on concurrent deletion and creation because a new list item allocation at the same place in memory as the just deleted list item (e.g., in a linked list) could interleave, thus causing the ABA problem.
As a result, the risk of naively programmed lock-free code realizing the ABA problem is probabilistic. It's highly probabilistic with dynamically allocated memory, particularly when considering that modern heap allocators avoid returning the same block of memory in too close succession as a form of exploit mitigation (i.e. modern heap allocators may not do MRU memory allocation as the ABA problem Wikipedia page suggests).
Modern instruction set architectures (ISAs), such as AArch64 (ARM), support load-link/store-conditional (LL/SC) instructions, which provide a stronger synchronization primitive than compare-and-swap (CAS). LL/SC solves the ABA problem on supporting architectures. On older architechtures (e.g. x86), one must use a workaround to create correct lock-free code that avoids the ABA problem. However, these workarounds can be complex and are often difficult to verify the correctness of.
Although the ABA problem and the ABBA deadlock names may sound similar, they have entirely different meanings, so be sure to distinguish between them.
Dining Philosophers Problem
The dining philosophers' problem is a classic scenario originally devised by Dijkstra to illustrate synchronization issues that occur in concurrent algorithms and how to resolve them. Here's the problem statement.
The simplest solution is that philosophers pick a side (e.g., the left side) and then agree always to take a fork from that side first. If a philosopher picks up the left fork and then fails to pick up the right fork because someone else is holding it, he puts the left fork back down. Otherwise, now holding both forks, he eats some spaghetti and then puts both forks back down. By controlling the order in which philosophers pick up forks, you effectively create lock hierarchies between the left and right forks, thus preventing deadlock.
Anyone who's learned about concurrency throughout their computer science program would be familiar with this problem.
Reverse Engineered Windows Loader Functions
LdrpDrainWorkQueue
Status: Fully reverse engineered! ✔️
LdrpDrainWorkQueue
is the high-level mapping and snapping function. It's frequently called all throughout the loader, so I made reversing it my priority. See the LdrpDrainWorkQueue and LdrpProcessWork and LdrpLoadCompleteEvent
and LdrpWorkCompleteEvent
sections for more information on this function. For context on how the LdrpDrainWorkQueue
function fits in with the rest of the loader, see the LDR_DDAG_NODE.State
Analysis section.
// Variable names are my own
// The caller decides whether it wants LdrpDrainWorkQueue to synchronize on an entire load completing (i.e. setting ntdll!LdrpWorkInProgress to 0 and decommissioning of the current thread as the load owner) or mapping and snapping work completing (i.e. processing all the work in the work queue)
// Typically, this decision depends on the whether or not the current thread is already the load owner (see: ntdll!LdrpLoadDllInternal function)
// For instance, if this is a recursive load then the current thread will already be the load owner
// In that scenario, LdrpLoadDllInternal will synchronize LdrpDrainWorkQueue on an item from the work queue completing its mapping and snapping work
struct PTEB __fastcall LdrpDrainWorkQueue(BOOL SynchronizeOnWorkCompletion)
{
HANDLE EventHandle;
BOOL CompleteRetryOrReturn;
BOOL LdrpDetourExistAtStart;
PLIST_ENTRY LdrpWorkQueueEntry;
//PLIST_ENTRY LinkHolderCheckCorruptionTemp; // This variable is inlined by the call to RtlpCheckListEntry
PTEB CurrentTeb;
PLIST_ENTRY LdrpRetryQueueEntry;
PLIST_ENTRY LdrpRetryQueueBlink;
EventHandle = (HANDLE)LdrpWorkCompleteEvent;
CompleteRetryOrReturn = FALSE
if ( !SynchronizeOnWorkCompletion )
EventHandle = LdrpLoadCompleteEvent;
while ( TRUE )
{
while ( TRUE )
{
RtlEnterCriticalSection(&LdrpWorkQueueLock);
// LdrpDetourExists relates to LdrpCriticalLoaderFunctions; see this document's section on that
LdrpDetourExistAtStart = LdrpDetourExist;
if ( !LdrpDetourExist || SynchronizeOnWorkCompletion == TRUE )
{
LdrpWorkQueueEntry = &LdrpWorkQueue;
// Corruption check on LdrpWorkQueue list: https://www.alex-ionescu.com/new-security-assertions-in-windows-8/
RtlpCheckListEntry(LdrpWorkQueueEntry);
// Get LdrpWorkQueue list head blink to get the last entry in the circular list
LdrpWorkQueueEntry = LdrpWorkQueue.Blink;
if (&LdrpWorkQueue == LdrpWorkQueueEntry) {
if (LdrpWorkInProgress == SynchronizeOnWorkCompletion) {
LdrpWorkInProgress = 1;
CompleteRetryOrReturn = 1;
}
} else {
if (!LdrpDetourExistAtStart)
++LdrpWorkInProgress;
// LdrpUpdateStatistics is a very small function with one branch on whether we're a loader worker thread
LdrpUpdateStatistics();
}
}
else
{
if ( LdrpWorkInProgress == SynchronizeOnWorkCompletion ) {
LdrpWorkInProgress = 1;
CompleteRetryOrReturn = TRUE;
}
// NOTE: This path always sets LdrpWorkQueueEntry = LdrpWorkQueue so the following check on whether LdrpWorkQueue is empty will always pass
LdrpWorkQueueEntry = &LdrpWorkQueue;
}
RtlLeaveCriticalSection(&LdrpWorkQueueLock);
if ( CompleteRetryOrReturn )
break;
// Test if LdrpWorkQueue is empty
// LdrpWorkQueueEntry may point to the last LDRP_LOAD_CONTEXT entry in the LdrpWorkQueue list
// If the LdrpWorkQueue list head equals the last entry in the circular linked list, the list must be empty
if ( &LdrpWorkQueue == LdrpWorkQueueEntry )
{
NtWaitForSingleObject(EventHandle, 0, NULL);
}
else
{
// LdrpProcessWork processes (i.e. mapping and snapping) the specified work item
LdrpProcessWork(LdrpWorkQueueEntry - 8, LdrpDetourExistAtStart);
}
}
// Test if not SynchronizeOnWorkCompletion OR LdrpRetryQueue is empty
//
// WinDbg disassembly (IDA disassembly with "cs:" and decompilation is poor here):
// lea rbx, [ntdll!LdrpRetryQueue (7ffb5bebc3a0)]
// cmp qword ptr [ntdll!LdrpRetryQueue (7ffb5bebc3a0)], rbx
// je ntdll!LdrpDrainWorkQueue+0xb1 (7ffb5bdaea85)
// https://stackoverflow.com/a/68702967
if ( !SynchronizeOnWorkCompletion || &LdrpRetryQueue == LdrpRetryQueue.Flink )
break;
RtlEnterCriticalSection(&LdrpWorkQueueLock);
// Complete a retried mapping and snapping operation
// Remove work item from LdrpRetryQueue/LdrpWorkQueue
// Reverse engineered based on WinDbg disassembly due to the IDA issue described above
// This is likely some macro (please contribute if you know)
// r12 is ntdll!LdrpWorkQueue
// rbx is ntdll!LdrpRetryQueue
LdrpRetryQueueEntry = LdrpRetryQueue; // mov rax, qword ptr [ntdll!LdrpRetryQueue (7ffb5bebc3a0)]
// lea rcx, [ntdll!LdrpWorkQueueLock (7ffb5bebc3c0)]
// xorps xmm0, xmm0 (any value xor'd by itself is zero)
LdrpRetryQueueEntry.Blink = LdrpWorkQueue.Flink; // mov qword ptr [rax+8], r12
LdrpWorkQueue.Flink = LdrpRetryQueueEntry.Flink; // mov qword ptr [ntdll!LdrpWorkQueue (7ffb5bebc3f0)], rax
// mov rax, qword ptr [ntdll!LdrpRetryQueue+0x8 (7ffb5bebc3a8)] (not a memory write)
LdrpRetryQueue.Blink = LdrpWorkQueue.Flink; // mov qword ptr [rax], r12
LdrpWorkQueue.Blink = LdrpRetryQueue.Blink; // mov qword ptr [ntdll!LdrpWorkQueue+0x8 (7ffb5bebc3f8)], rax
LdrpRetryQueue.Blink = LdrpRetryQueue.Flink; // mov qword ptr [ntdll!LdrpRetryQueue+0x8 (7ffb5bebc3a8)], rbx
LdrpRetryQueue.Flink = LdrpRetryQueue.Flink; // mov qword ptr [ntdll!LdrpRetryQueue (7ffb5bebc3a0)], rbx
// Clear ntdll!LdrpRetryingModuleIndex
// Global used by ntdll!LdrpCheckForRetryLoading function (may be called during file mapping by ntdll!LdrpMinimalMapModule or ntdll!LdrpMapDllNtFileName functions)
// LdrpRetryingModuleIndex is a red-black tree (LdrpCheckForRetryLoading modifies it by calling RtlRbInsertNodeEx)
// LdrpCheckForRetryLoading entries point into the undocumented LDRP_LOAD_CONTEXT structure
LdrpRetryingModuleIndex = NULL; // movdqu xmmword ptr [ntdll!LdrpRetryingModuleIndex (7ffb5bebd350)], xmm0
RtlLeaveCriticalSection(&LdrpWorkQueueLock);
CompleteRetryOrReturn = FALSE;
}
// Set current thread as the load owner so the rest of the loading process can take place upon returning
// This happens *after* processing (i.e. mapping and snapping) work
// Note that parallel loaders threads never call LdrpDrainWorkQueue (they do LdrpWorkCallback -> LdrpProcessWork)
// Likely retuns a pointer to the TEB because that's faster than having the caller retrieve the TEB again
CurrentTeb = NtCurrentTeb();
CurrentTeb->SameTebFlags |= LoadOwner; // 0x1000
return CurrentTeb;
}
LdrpDecrementModuleLoadCountEx
Status: Fully reverse engineered! ✔️
Reversing LdrpDecrementModuleLoadCountEx
was necessary for my investigation on how GetProcAddress
handles concurrent library unload.
NTSTATUS LdrpDecrementModuleLoadCountEx(LDR_DATA_TABLE_ENTRY Entry, BOOL DontCompleteUnload)
{
LDR_DDAG_NODE Node;
NTSTATUS Status;
BOOL CanUnloadNode;
DWORD_PTR LdrpReleaseLoaderLockReserved; // Not used or even initialized, so I still consider this function fully reverse engineered (also, LdrpReleaseLoaderLock never touches this parameter)
// If the next reference counter decrement will drop the LDR_DDAG_NODE into having zero references then we may want to retry later
// Specifying DontCompleteUnload = FALSE requires that the caller has made this thread the load owner
if ( DontCompleteUnload && Entry->DdagNode->LoadCount == 1 )
{
// Retry when we're the load owner
// NTSTATUS code 0xC000022D: https://learn.microsoft.com/en-us/openspecs/windows_protocols/ms-erref/596a1078-e883-4972-9bbc-49e60bebca55
return STATUS_RETRY;
}
RtlAcquireSRWLockExclusive(&LdrpModuleDatatableLock);
Node = Entry->DdagNode;
Status = LdrpDecrementNodeLoadCountLockHeld(Node, DontCompleteUnload, &CanUnloadNode);
RtlReleaseSRWLockExclusive(&LdrpModuleDatatableLock);
if ( CanUnloadNode )
{
LdrpAcquireLoaderLock();
// LdrpUnloadNode runs a module's DLL_PROCESS_DETACH
// It also walks the dependency graph potentially to unload other now unused modules
LdrpUnloadNode(Node);
LdrpReleaseLoaderLock(LdrpReleaseLoaderLockReserved, 8); // Second parameter is an ID for use in log messages
}
}
Analysis Commands
These are some helpful tips and commands to aid analysis.
LDR_DATA_TABLE_ENTRY
Analysis
List all module LDR_DATA_TABLE_ENTRY
structures:
!list -x "dt ntdll!_LDR_DATA_TABLE_ENTRY" @@C++(&@$peb->Ldr->InLoadOrderModuleList)
LDR_DDAG_NODE
Analysis
List all module DdagNode
structures:
!list -x "r $t0 =" -a "; dt ntdll!_LDR_DATA_TABLE_ENTRY @$t0 -t BaseDllName; dt ntdll!_LDR_DDAG_NODE @@C++(((ntdll!_LDR_DATA_TABLE_ENTRY *)@$t0)->DdagNode)" @@C++(&@$peb->Ldr->InLoadOrderModuleList)
List all module DdagNode.State
values:
!list -x "r $t0 =" -a "; dt ntdll!_LDR_DATA_TABLE_ENTRY @$t0 -cio -t BaseDllName; dt ntdll!_LDR_DDAG_NODE @@C++(((ntdll!_LDR_DATA_TABLE_ENTRY *)@$t0)->DdagNode) -cio -t State" @@C++(&@$peb->Ldr->InLoadOrderModuleList)
Trace all DdagNode.State
values starting from its initial allocation to the next module load for every module:
bp ntdll!LdrpAllocateModuleEntry "bp /1 @$ra \"bc 999; ba999 w8 @@C++(&((ntdll!_LDR_DATA_TABLE_ENTRY *)@$retreg)->DdagNode->State) \\\"ub . L1; g\\\"; g\"; g"
- This command breaks on
ntdll!LdrpAllocateModuleEntry
, sets a temporary breakpoint on its return address (@$ra
), then continuing execution, sets a watchpoint onDdagNode.State
in the returnedLDR_DATA_TABLE_ENTRY
, and log the previous disassembly line on watchpoint hit - We must clear the previous watchpoint (
bc 999
) to set a new one every time a new module load occurs (ModLoad
message in WinDbg debug output). Deleting hardware breakpoints is necessary because they are a finite resource.
Throw this in with the trace command after ub . L1;
to check some locks: !critsec LdrpLoaderLock; dp LdrpModuleDataTableLock L1;
Warning: Due to how module loads work out, some state changes will never appear in a trace due to deleting watchpoints. For example, this trace command won't uncover a LdrModulesSnapping
state change in the LdrpSignalModuleMapped
function or a LdrModulesWaitingForDependencies
state change. For this reason, it's necessary to sample modules randomly over a few separate executions. Run sxe ld:<DLL_TO_BEGIN_SAMPLE_AT>
and remove the watchpoint deletion command (bc 999
), then run the trace command starting from that point.
Warning: This command will continue tracing an address after a module unloads, which causes LdrpUnloadNode
➜ RtlFreeHeap
of memory. Disregard junk results from that point forward due to potential reallocation of the underlying memory until the next ModLoad
WinDbg message, meaning a fresh watchpoint.
Trace TEB.SameTebFlags
ba w8 @@C++(&@$teb->SameTebFlags-3)
TEB.SameTebFlags
isn't memory-aligned, so we need to subtract 3
to set a hardware breakpoint. This watchpoint still captures the full TEB.SameTebFlags
but now TEB.MuiImpersonation
, TEB.CrossTebFlags
, and TEB.SpareCrossTebBits
, in front of TEB.SameTebFlags
in the TEB
structure, will also fire our watchpoint. However, these other members are rarely used, so it's only a minor inconvenience.
ASLR randomizes the TEB's location on every execution, so you must set the watchpoint again when restarting the program in WinDbg.
Read the TEB: dt @$teb ntdll!_TEB
Searching Assembly for Structure Offsets
These commands contain the offset/register values specific to a 64-bit process. Please adjust them if you're working with a 32-bit process.
WinDbg command: # "*\\+38h\\]*" <NTDLL ADDRESS> L9999999
-
Search NTDLL disassembly for occurrences of offset
+0x38
(useful to search for potentialLDR_DDAG_NODE.State
references) - Put the output of this command into an
offset-search.txt
file
Filter command pipeline (POSIX sh): sed '/^ntdll!Ldr/!d;N;' offset-search.txt | sed 'N;/\[rsp+/d;'
- The WinDbg
#
command outputs two lines for each finding, hence thesed
commands using itsN
command to read the next line into pattern space - First command: Filter for findings beginning with
ntdll!Ldr
because we're interested in the loader- Use this
sed
as agrep -v -A1
substitute because the latter prints a group separator, which we don't want (GNUgrep
supports--no-group-separator
to remove this, but I prefer to keep POSIX compliant)
- Use this
- Second command: Filter out offsets to the stack pointer, which is just noise operating on local variables
Depending on what you're looking for, you should filter further—for example, filtering down to only mov
or lea
assembly instructions. Be aware that the results won't be comprehensive because the code could also add/subtract an offset from a member in the structure it's already at to reach another member in the current structure (i.e. use of the CONTAINING_RECORD
macro). It's still generally helpful, though.
Due to two's complement, negative numbers will, of course, show up like: 0FFFFFFFFh
(e.g. for -1
, WinDbg disassembly prints it with a leading zero)
Monitor a Critical Section Lock
Watch for reads/writes to a critical section's locking status.
ba r4 @@C++(&((ntdll!_RTL_CRITICAL_SECTION *)@@(ntdll!LdrpDllNotificationLock))->LockCount)
I tested placing a watchpoint on loader lock (LdrpLoaderLock
). Doing this on a modern Windows loader won't tell you much because, as previously mentioned, it's the state, such as the LDR_DDAG_NODE.State
or LoadOwner
in TEB.SameTebFlags
that the loader internally tests to make decisions. Setting this watchpoint myself during complex operations, I've never seen an occurrence of Windows using the ntdll!RtlIsCriticalSectionLocked
or ntdll!RtlIsCriticalSectionLockedByThread
functions to branch on the state of loader lock itself.
Track Loader Events
This WinDbg command tracks load and work completions.
ba e1 ntdll!NtSetEvent ".if ($argreg == poi(ntdll!LdrpLoadCompleteEvent)) { .echo Set LoadComplete Event; k }; .elsif ($argreg == poi(ntdll!LdrpWorkCompleteEvent)) { .echo Set WorkComplete Event; k }; gc"
We use a hardware execution breakpoint (ba e1
) instead of a software breakpoint (bp
) because otherwise, there's some strange bug where WinDbg may never hit the breakpoint for LdrpWorkCompleteEvent
(root cause requires investigation).
Note that this tracer is slow because Windows often uses events (calling NtSetEvent
) even outside the loader.
Swap out ntdll!NtSetEvent
for ntdll!NtWaitForSingleObject
and change echo
messages to get more info. The LdrpLoadCompleteEvent
and LdrpWorkCompleteEvent
events are never manually reset (ntdll!NtResetEvent
). They're auto-reset events, so passing them into a wait function causes them to reset (even if no actual waiting occurs).
Be aware that anti-malware software hooking on ntdll!NtWaitForSingleObject
could cause the breakpoint command to run twice (I ran into this issue).
View state of Win32 events WinDbg command: !handle 0 ff Event
Disable Loader Worker Threads
This CMD command helps analyze what locks are acquired in what parts of the code without the chance of a loader thread muddying your analysis (among other things). These show up as ntdll!TppWorkerThread
threads in WinDbg at process startup (note that any threads belonging to other thread pools also take this thread name).
reg add "HKLM\SOFTWARE\Microsoft\Windows NT\CurrentVersion\Image File Execution Options\<YOUR_EXE_FILENAME>" /v MaxLoaderThreads /t REG_DWORD /d 1 /f
Yes, the correct value for disabling loader threads from spawning is 1
. Set 0
and loader threads still spawn (0
appears to be the same as unset). Setting 2
spawns one loader work thread, and that pattern continues...
Internally, at process initialization (LdrpInitializeProcess
), the loader retrieves MaxLoaderThreads
in LdrpInitializeExecutionOptions
. LdrpInitializeExecutionOptions
queries this registry value (along with others) and saves the result. Later, the MaxLoaderThreads
registry value becomes the first argument to the LdrpEnableParallelLoading
function. LdrpEnableParallelLoading
validates the value before passing it as the second argument to the TpSetPoolMaxThreads
function. The TpSetPoolMaxThreads
function does a NtSetInformationWorkerFactory
system call with the second argument being the enum value WorkerFactoryThreadMaximum
(5
) to set the maximum thread count and the desired count pointed to by the third argument within the WorkerFactoryInformation
structure (undocumented).
The maximum number of threads the parallel loader allows is 4
. Specifying a higher thread count will top out at 4
.
Creating threads happens with ntdll!NtCreateWorkerFactory
. The ntdll!NtCreateWorkerFactory
function only does the NtCreateWorkerFactory
system call; this is a kernel wrapper for creating multiple threads at once, improving performance because it avoids extra user-mode ⬌ kernel-mode context switches.
Some operations, such as ShellExecute
, may spawn threads named ntdll!TppWorkerThread
. However, these worker threads belong to an unrelated thread pool, not the loader work thread pool, and thus, never call the ntdll!LdrpWorkCallback
function entry point. Make sure to distinguish loader work thread pools: dt @$teb ntdll!_TEB -t LoaderWorker
List All LdrpCriticalLoaderFunctions
Whether a critical loader function (ntdll!LdrpCriticalLoaderFunctions
) is detoured (i.e. hooked) determines whether ntdll!LdrpDetourExist
is TRUE
or FALSE
. BlackBerry's Jeffrey Tang covered this in "Windows 10 Parallel Loading Breakdown"; however, this article only lists the first few critical loader functions, so here we show all of them. Below are the commands for determining all critical loader functions should they change in the future:
0:000> ln ntdll!LdrpCriticalLoaderFunctions
(00007ffb`1cb8df20) ntdll!LdrpCriticalLoaderFunctions | (00007ffb`1cb8e020) ntdll!RtlpMemoryZoneCriticalRoutines
0:000> ? (00007ffb`1cb8e020 - 00007ffb`1cb8df20) / @@(sizeof(void*))
Evaluate expression: 32 = 00000000`00000020
0:000> .for (r $t0=0; $t0 < <OUTPUT_OF_PREVIOUS_EXPRESSION_HERE>; r $t0=$t0+1) { u poi(ntdll!LdrpCriticalLoaderFunctions + $t0 * @@(sizeof(void*))) L1 }
ntdll!NtOpenFile:
00007ffb`1cb0d630 4c8bd1 mov r10,rcx
ntdll!NtCreateSection:
00007ffb`1cb0d910 4c8bd1 mov r10,rcx
ntdll!NtQueryAttributesFile:
00007ffb`1cb0d770 4c8bd1 mov r10,rcx
... more output ...
Clean up the output by filtering it through this POSIX sed
pipeline: sed 'n;d;' <OUTPUT_FILE> | sed 's/.$//'
Finally, we have a comprehensive list of critical loader functions:
ntdll!NtOpenFile
ntdll!NtCreateSection
ntdll!NtQueryAttributesFile
ntdll!NtOpenSection
ntdll!NtMapViewOfSection
ntdll!NtWriteVirtualMemory
ntdll!NtResumeThread
ntdll!NtOpenSemaphore
ntdll!NtOpenMutant
ntdll!NtOpenEvent
ntdll!NtCreateUserProcess
ntdll!NtCreateSemaphore
ntdll!NtCreateMutant
ntdll!NtCreateEvent
ntdll!NtSetDriverEntryOrder
ntdll!NtQueryDriverEntryOrder
ntdll!NtAccessCheck
ntdll!NtCreateThread
ntdll!NtCreateThreadEx
ntdll!NtCreateProcess
ntdll!NtCreateProcessEx
ntdll!NtCreateJobObject
ntdll!NtCreateEnclave
ntdll!NtOpenThread
ntdll!NtOpenProcess
ntdll!NtOpenJobObject
ntdll!NtSetInformationThread
ntdll!NtSetInformationProcess
ntdll!NtDeleteFile
ntdll!NtCreateFile
ntdll!NtMapViewOfSectionEx
ntdll!NtExtendSection
Make sure to prepend 0n
to OUTPUT_OF_PREVIOUS_EXPRESSION_HERE
if you use the base ten representation (0n32
); otherwise, WinDbg interprets that number as hex (base 16) by default.
Microsoft's public symbols only come with names that typically do not include any information on type or size. Hence, we must find the ntdll!LdrpCriticalLoaderFunctions
size with the ln
command instead of sizeof
.
Loader Debug Logging
The Windows loader supports printing debug messages. These are helpful for determining what actions the loader is doing at a high-level. However, these messages won't help you reverse engineer how the loader is works internally (e.g. they certainly don't help with figuring out how the loader supports concurrency).
The loader performs bitwise and
operations on LdrpDebugFlags
to determine debugging behavior (the individual flag values should be reasonably clear based on the commands given below). Flags values are in hexadecimal.
In WinDbg, run sxe ld:ntdll
(run sxe ld:-
to undo this command) to break on the first assembly instruction of NTDLL, restart the program, and then run one of the following commands to enable loader debug logging.
Info, warning, and error messages: eb ntdll!LdrpDebugFlags 1
Warning and error messages: eb ntdll!LdrpDebugFlags 2
Info messages: eb ntdll!LdrpDebugFlags 4
No messages + debug break on errors: eb ntdll!LdrpDebugFlags 10
No messages + debug break on warnings: eb ntdll!LdrpDebugFlags 40
No messages + debug break on warnings or errors: eb ntdll!LdrpDebugFlags 50
Info, warning and error messages + debug break on warnings or errors: eb ntdll!LdrpDebugFlags 51
Warning and error messages + debug break on warnings or errors: eb ntdll!LdrpDebugFlags 52
Info messages + debug break on warnings or errors: eb ntdll!LdrpDebugFlags 54
A warning would be something like LdrpGetProcedureAddress
(exposed as GetProcAddress
in the public API) failing to locate a proceedure. An error would typically be something fatal to the entire loading/unloading process such as a DLL failing to initialize (i.e. DllMain
returning FALSE
), or raising an exception during DLL initialization (the loader will catch this if you don't).
Get TCB and Set GSCOPE Watchpoint (GNU Loader)
Read the thread control block on GNU/Linux (equivalent to TEB on Windows):
set print pretty
print *(tcbhead_t*)($fs_base)
Set a watchpoint on this thread's GSCOPE flag:
print &((tcbhead_t*)($fs_base))->gscope_flag
watch *<OUTPUT_OF_LAST_COMMAND>
Compiling & Running
- Run
make
where there is aMakefile
- Run
source ../set-ld-lib-path.sh
- Run
./main
orgdb ./main
to debug
Debugging Notes
Ensure you have glibc debug symbols and, preferably, source code downloaded. Fedora's GDB automatically downloads symbols and source code. On Debian, you must install the libc6-dbg
(you may need to enable the debug source in your /etc/apt/sources.list
) and glibc-source
packages.
When printing a backtrace
, you may see <optimized out>
in the output for some function parameter values. Seeing these requires compiling a debug build of glibc. Not debug symbols, a debug build. You could also set a breakpoint where the value is optimized out to see it. Generally, you don't need to see these values, though, so you can ignore them.
In backtraces contained within the gdb-log.txt
files of this repo, you may entries to functions like dlerror_run
, _dl_catch_error
and __GI__dl_catch_exception
(_dl_catch_exception
in the code). These aren't indicative of an error occurring. Rather, these functions merely set up an error handler and perform handling to catch exceptions if an error/exception occurs in the dynamic linker. The error codes are conformant to errno. The most common error code is an ENOMEM
(out of memory), but a file access error code (ENOENT
or EACCES
) can also occur. Dynamic linker functions may also pass 0
as the errno
to indicate an error specific to the dynamic linker's functionality. In either case, an exception occurs, thus stopping the dynamic linker from proceeding and continuing execution in _dl_catch_exception
(called by _dl_catch_error
). A programmer can determine if dlopen
failed by checking for a NULL
return value and then get an error message with dlerror
.
Big thanks to Microsoft for successfully nerd sniping me!
The document you just read is under a CC BY-SA License.
This repo's code is dual licensed under the MIT License or GPL 2+ at your choice.
Copyright (C) 2023-2024 Elliot Killick [email protected]
EOF