LibC+LibELF: Support loading shared libraries compiled with dynamic TLS

This is a prerequisite for upstreaming our LLVM patches, as our current
hack forcing `-ftls-model=initial-exec` in the Clang driver is not
acceptable upstream.

Currently, our kernel-managed TLS implementation limits us to only
having a single block of storage for all thread-local variables that's
initialized at load time. This PR merely implements the dynamic TLS
interface (`__tls_get_addr` and TLSDESC) on top of our static TLS
infrastructure. The current model's limitations still stand:
- a single static TLS block is reserved at load time, `dlopen()`-ing
  shared libraries that define thread-local variables might cause us to
  run out of space.
- the initial TLS image is not changeable post-load, so `dlopen()`-ing
  libraries with non-zero-initialized TLS variables is not supported.

The way we repurpose `ti_module` to mean "offset within static TLS
block" instead of "module index" is not ABI-compliant.

Userland/DynamicLoader/CMakeLists.txt

@@ -13,6 +13,8 @@ file(GLOB LIBC_SOURCES3 "../Libraries/LibC/arch/${ARCH_FOLDER}/*.S")
 set(ELF_SOURCES ${ELF_SOURCES} "../Libraries/LibELF/Arch/${ARCH_FOLDER}/entry.S" "../Libraries/LibELF/Arch/${ARCH_FOLDER}/plt_trampoline.S")
 if ("${SERENITY_ARCH}" STREQUAL "x86_64")
     set(LIBC_SOURCES3 ${LIBC_SOURCES3} "../Libraries/LibC/arch/x86_64/memset.cpp")
+elseif ("${SERENITY_ARCH}" STREQUAL "aarch64")
+    set(ELF_SOURCES ${ELF_SOURCES} "../Libraries/LibELF/Arch/aarch64/tls.S")
 endif()
 
 file(GLOB LIBSYSTEM_SOURCES "../Libraries/LibSystem/*.cpp")

Userland/Libraries/LibC/CMakeLists.txt

@@ -67,6 +67,7 @@ set(LIBC_SOURCES
     termios.cpp
     time.cpp
     times.cpp
+    tls.cpp
     ulimit.cpp
     unistd.cpp
     utime.cpp
@@ -99,7 +100,7 @@ file(GLOB ELF_SOURCES CONFIGURE_DEPENDS "../LibELF/*.cpp")
 
 if ("${SERENITY_ARCH}" STREQUAL "aarch64")
     set(ASM_SOURCES "arch/aarch64/setjmp.S")
-    set(ELF_SOURCES ${ELF_SOURCES} ../LibELF/Arch/aarch64/entry.S ../LibELF/Arch/aarch64/plt_trampoline.S)
+    set(ELF_SOURCES ${ELF_SOURCES} ../LibELF/Arch/aarch64/entry.S ../LibELF/Arch/aarch64/plt_trampoline.S ../LibELF/Arch/aarch64/tls.S)
     set(CRTI_SOURCE "arch/aarch64/crti.S")
     set(CRTN_SOURCE "arch/aarch64/crtn.S")
 elseif ("${SERENITY_ARCH}" STREQUAL "x86_64")

Userland/Libraries/LibC/sys/internals.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020, the SerenityOS developers.
+ * Copyright (c) 2020-2023, the SerenityOS developers.
  *
  * SPDX-License-Identifier: BSD-2-Clause
  */
@@ -7,6 +7,7 @@
 #pragma once
 
 #include <sys/cdefs.h>
+#include <sys/types.h>
 
 __BEGIN_DECLS
 
@@ -28,4 +29,11 @@ __attribute__((noreturn)) void __cxa_pure_virtual(void) __attribute__((weak));
 __attribute__((noreturn)) void __stack_chk_fail(void);
 __attribute__((noreturn)) void __stack_chk_fail_local(void);
 
+struct __tls_index {
+    size_t ti_module;
+    size_t ti_offset;
+};
+
+void* __tls_get_addr(__tls_index*);
+
 __END_DECLS

Userland/Libraries/LibC/tls.cpp

@@ -0,0 +1,26 @@
+/*
+ * Copyright (c) 2023, Daniel Bertalan <dani@danielbertalan.dev>
+ *
+ * SPDX-License-Identifier: BSD-2-Clause
+ */
+
+#include <AK/Types.h>
+#include <sys/internals.h>
+
+extern "C" {
+
+// This function is called to compute the address of a thread-local variable
+// which might not be stored in the static TLS block (local-dynamic and
+// global-dynamic models). Compilers default to this when creating shared
+// libraries, as they may be loaded after program startup by `dlopen()`.
+//
+// We currently only support a static TLS block, so we take a shortcut in the
+// implementation of this interface: instead of storing the module ID in
+// ti_module, we store the module's TLS block offset. This avoids the need to
+// have a per-thread module ID -> TLS block address. This will have to be
+// changed if we support dynamically allocated TLS blocks.
+void* __tls_get_addr(__tls_index* index)
+{
+    return reinterpret_cast<void*>(reinterpret_cast<FlatPtr>(__builtin_thread_pointer()) + index->ti_module + index->ti_offset);
+}
+}

Userland/Libraries/LibELF/Arch/aarch64/tls.S

@@ -0,0 +1,63 @@
+/*
+ * Copyright (c) 2023, Daniel Bertalan <dani@danielbertalan.dev>
+ *
+ * SPDX-License-Identifier: BSD-2-Clause
+ */
+
+// This file implements the runtime components of the AArch64 TLSDESC ABI,
+// which is used when accessing thread-local variables which might not be
+// stored in the static TLS block (global-dynamic and local-dynamic access
+// models). Compilers default to this when creating shared libraries, as they
+// may be loaded after program startup by `dlopen()`.
+//
+// Each referenced thread-local symbol is associated with a descriptor:
+//
+// struct TlsDescriptor {
+//     size_t (*resolver)(TlsDescriptor*);
+//     union {
+//         size_t tpoff; // for static TLS
+//         struct {
+//             size_t module_id;
+//             size_t module_offset;
+//         } *dynamic; // for dynamic TLS, not yet implemented
+//     };
+// };
+//
+// The resolver takes a pointer to the descriptor as an argument and returns
+// the symbol's offset to the thread pointer (tpidr_el1). The second field of
+// the descriptor is an implementation-defined value which the resolver uses to
+// identify the symbol.
+//
+// Thus, the address of a thread-local variable is retrieved as follows:
+//
+//   &var = thread_pointer + descriptor.resolver(&descriptor);
+//
+// The two essential types of resolver functions are:
+//
+// - `__tlsdesc_static`: If the variable is located in the static TLS block,
+//   its thread pointer offset is a load-time constant, which can be stored in
+//   the descriptor. This function simply returns that.
+//
+// - `tlsdesc_dynamic`: Looks up a variable by its module ID and module offset.
+//   This is used if the TLS block is allocated separately, so might have a
+//   different thread pointer offset for each thread. This works similarly to
+//   the traditional TLS ABI's __tls_get_addr function. Not yet implemented in
+//   SerenityOS.
+//
+// The TLSDESC format strives to make the code sequence for thread-local
+// variable access as short as possible, hence the resolver functions follow a
+// special calling convention: they must not clobber any registers. To ensure
+// that even the usually volatile registers are saved off, we need to implement
+// the resolvers in assembly.
+
+// size_t __tlsdesc_static(TlsDescriptor* desc)
+// {
+//     return desc->tpoff;
+// }
+.p2align 4
+.globl __tlsdesc_static
+.hidden __tlsdesc_static
+.type __tlsdesc_static,@function
+__tlsdesc_static:
+    ldr x0, [x0, #8]
+    ret

Userland/Libraries/LibELF/DynamicLoader.cpp

@@ -2,7 +2,7 @@
  * Copyright (c) 2019-2020, Andrew Kaster <akaster@serenityos.org>
  * Copyright (c) 2020, Itamar S. <itamar8910@gmail.com>
  * Copyright (c) 2021, Andreas Kling <kling@serenityos.org>
- * Copyright (c) 2022, Daniel Bertalan <dani@danielbertalan.dev>
+ * Copyright (c) 2022-2023, Daniel Bertalan <dani@danielbertalan.dev>
  *
  * SPDX-License-Identifier: BSD-2-Clause
  */
@@ -35,6 +35,13 @@ static void* mmap_with_name(void* addr, size_t length, int prot, int flags, int
 #    define MAP_RANDOMIZED 0
 #endif
 
+#if ARCH(AARCH64)
+#    define HAS_TLSDESC_SUPPORT
+extern "C" {
+void* __tlsdesc_static(void*);
+}
+#endif
+
 namespace ELF {
 
 Result<NonnullRefPtr<DynamicLoader>, DlErrorMessage> DynamicLoader::try_create(int fd, DeprecatedString filepath)
@@ -527,6 +534,23 @@ DynamicLoader::RelocationResult DynamicLoader::do_direct_relocation(DynamicObjec
         return VirtualAddress { reinterpret_cast<DynamicObject::IfuncResolver>(address.get())() };
     };
 
+    struct ResolvedTLSSymbol {
+        DynamicObject const& dynamic_object;
+        FlatPtr value;
+    };
+
+    auto resolve_tls_symbol = [](DynamicObject::Relocation const& relocation) -> Optional<ResolvedTLSSymbol> {
+        if (relocation.symbol_index() == 0)
+            return ResolvedTLSSymbol { relocation.dynamic_object(), 0 };
+
+        auto res = lookup_symbol(relocation.symbol());
+        if (!res.has_value())
+            return {};
+        VERIFY(relocation.symbol().type() != STT_GNU_IFUNC);
+        VERIFY(res.value().dynamic_object != nullptr);
+        return ResolvedTLSSymbol { *res.value().dynamic_object, res.value().value };
+    };
+
     switch (relocation.type()) {
 
     case R_X86_64_NONE:
@@ -601,30 +625,51 @@ DynamicLoader::RelocationResult DynamicLoader::do_direct_relocation(DynamicObjec
     }
     case R_AARCH64_TLS_TPREL:
     case R_X86_64_TPOFF64: {
-        auto symbol = relocation.symbol();
-        FlatPtr symbol_value;
-        DynamicObject const* dynamic_object_of_symbol;
-        if (relocation.symbol_index() != 0) {
-            auto res = lookup_symbol(symbol);
-            if (!res.has_value())
+        auto maybe_resolution = resolve_tls_symbol(relocation);
+        if (!maybe_resolution.has_value())
             break;
-            VERIFY(symbol.type() != STT_GNU_IFUNC);
-            symbol_value = res.value().value;
-            dynamic_object_of_symbol = res.value().dynamic_object;
-        } else {
-            symbol_value = 0;
-            dynamic_object_of_symbol = &relocation.dynamic_object();
-        }
-        VERIFY(dynamic_object_of_symbol);
-        size_t addend = relocation.addend_used() ? relocation.addend() : *patch_ptr;
+        auto [dynamic_object_of_symbol, symbol_value] = maybe_resolution.value();
 
-        *patch_ptr = addend + dynamic_object_of_symbol->tls_offset().value() + symbol_value;
+        size_t addend = relocation.addend_used() ? relocation.addend() : *patch_ptr;
+        *patch_ptr = addend + dynamic_object_of_symbol.tls_offset().value() + symbol_value;
 
         // At offset 0 there's the thread's ThreadSpecificData structure, we don't want to collide with it.
         VERIFY(static_cast<ssize_t>(*patch_ptr) < 0);
-
         break;
     }
+    case R_X86_64_DTPMOD64: {
+        auto maybe_resolution = resolve_tls_symbol(relocation);
+        if (!maybe_resolution.has_value())
+            break;
+
+        // We repurpose the module index to store the TLS block's TP offset. This is fine
+        // because we currently only support a single static TLS block.
+        *patch_ptr = maybe_resolution->dynamic_object.tls_offset().value();
+        break;
+    }
+    case R_X86_64_DTPOFF64: {
+        auto maybe_resolution = resolve_tls_symbol(relocation);
+        if (!maybe_resolution.has_value())
+            break;
+
+        size_t addend = relocation.addend_used() ? relocation.addend() : *patch_ptr;
+        *patch_ptr = addend + maybe_resolution->value;
+        break;
+    }
+#ifdef HAS_TLSDESC_SUPPORT
+    case R_AARCH64_TLSDESC: {
+        auto maybe_resolution = resolve_tls_symbol(relocation);
+        if (!maybe_resolution.has_value())
+            break;
+        auto [dynamic_object_of_symbol, symbol_value] = maybe_resolution.value();
+
+        size_t addend = relocation.addend_used() ? relocation.addend() : *patch_ptr;
+
+        patch_ptr[0] = (FlatPtr)__tlsdesc_static;
+        patch_ptr[1] = addend + dynamic_object_of_symbol.tls_offset().value() + symbol_value;
+        break;
+    }
+#endif
     case R_AARCH64_IRELATIVE:
     case R_X86_64_IRELATIVE: {
         if (should_call_ifunc_resolver == ShouldCallIfuncResolver::No)