A Deep dive into (implicit) Thread Local Storage


Thread Local Storage (henceforth TLS) is pretty cool, it may appear to be simple at first glance but a good and efficient TLS implementation requires concerted effort from compiler, linker, dynamic linker, kernel, and language runtime.

On Linux, an excellent treatment to this topic is Ulrich Drepper’s ELF Handling For Thread-Local Storage, this blog post is my take on the same topic but with a different emphasis on how the details are presented.

We’ll limit the scope to x86-64, with a primary focus on Linux ELF, as it’s most well documented, but will also touch on other hardware and platforms. The same principle applies to the other architectures anyway.

Buckle up, it’s a long ride!

Introduction

Ordinary explicit TLS (with the help of pthread_key_create, pthread_setspecific, and pthread_getspecific) is pretty easy to reason about, it can be thought of as some kind of two dimensional hash map: tls_map[tid][key], where tid is thread ID, and key is the hash map key to a particular thread local variable.

However, implicit TLS1, like the one shown in the following C code, seems a bit more magical, its usage is a little too easy, makes you wonder if there’s some sleight of hand going on to make this happen.

#include <stdio.h>

__thread int main_tls_var; int main() { return main_tls_var;
}

The Assembly

Like everything in C, the easiest way to see through the “magic” is to disassemble them.

000000000000063a <main>: 63a: 55 push %rbp 63b: 48 89 e5 mov %rsp,%rbp 63e: 64 8b 04 25 fc ff ff mov %fs:0xfffffffffffffffc,%eax 645: ff 646: 5d pop %rbp 647: c3 retq

The magical line is mov %fs:0xfffffffffffffffc,%eax, for people who don’t read assembly, this essentially means moving the 32-bit value stored at the memory address %fs:0xfffffffffffffffc to the register %eax.

Actually even if one is well versed with assembly, this line is still somewhat unusual. It’s somewhat uncommon to see instructions involving segment register FS/GS in modern x86 assembly.

It looks like, at least in this case, disassembling the code gives us more mysteries than it solves. The strange use of segment register %fs and a large constant 0xfffffffffffffffc seem particularly ominous and mysterious.

The Thread Register

While some other CPU architectures have dedicated register to hold thread specific context (whatever that means), x86 enjoys no such luxury. Infamously only having small number of general purpose registers, x86 requires programmers to be very prudent when it comes to planning register uses.

Intel 80386 introduced FS and GS as user defined segment registers without stipulating what they were for. However later with the rise of multithreaded programming, people saw an opportunity in repurposing FS and GS and started to use them as thread register.

Note that this is x86 specific, and we won’t go into the details of memory segmentation, but suffice to say that on x86-64, non-kernel programs make use of FS or GS segment register to access per thread context including TLS (GS for Windows and macOS, FS for Linux and possibly the other Unix derivatives).

Now armed with the knowledge that %fs is most likely pointing to some sort of thread specific context, let’s try again to decode the instruction mov %fs:0xfffffffffffffffc,%eax. This mov instruction uses segmentation addressing, the large constant 0xfffffffffffffffc is just -4 in two’s complement form. %fs:0xfffffffffffffffc just means:

// Pointer arithmetics to get the address of `main_tls_var`
int *main_tls_var_ptr = (int *) thread_context_ptr - 4;
// Dereference tls_var_ptr and put its value in register EAX
EAX = *main_tls_var_ptr;

Where thread_context_ptr is whatever the address %fs points to. For other architectures, one can substitute %fs with their equivalent thread register.

After the execution, the register EAX contains the value of main_tls_var.

On x86-64, user land programs (Ring 3) can retrieve FS and GS, but they are not allowed to change the addresses stored in FS and GS. It’s typically the operating systems’ job (Ring 1) to provide facility to manipulate them indirectly.

On the x86-64 kernel, the address stored at FS is managed by a Model Specific Register (MSR) called MSR_FS_BASE2. The kernel provides a syscall arch_prctl with which user land programs can use to change FS and GS for the currently running thread.

Digging around a bit more, it can also be seen that when the kernel does context switch, the switch code loads the next task’s FS to the CPU’s MSR. Confirming that whatever FS points to is on a per thread/task basis.

Based on what we’ve known so far, we may have this hypothesis that the runtime must be using some form of FS manipulation routine (such as arch_prctl) to bind the TLS to the current thread. And the kernel keeps track of this binding by swap in the right FS value when doing context switch.

Thread Specific Context

So far we’ve been hand waving the term thread specific context, so what really is this context thing?

Different platforms have different names for it, for Linux and glibc, this is called Thread Control Block (TCB), on Windows it’s called Thread Information Block (TIB) or Thread Environment Block (TEB).

We will focus on TCB in this post, for more information about TIB/TEB, see the first part of Ken Johnson’s excellent blog series on Windows TLS.

Regardless how it’s called, thread specific context is a data structure that contains information and metadata to facilitate the management of a thread and its local storage (i.e. TLS).

So what is this TCB data structure?

On x86-64 Linux + glibc, for all intents and purposes, TCB is struct pthread (some times called thread descriptor), it’s a glibc internal data structure related but not equivalent to POSIX Threads.

At this point we now know what the FS points to and vaguely what TCB is, but there are still a lot of questions to be answered. First and foremost, who sets up and allocates the TCB? We the code author certainly did no such thing anywhere in the original source.

The Initialisation of TCB or TLS

As it turns out, TCB or TLS setup is done somewhat differently for statically linked executables and dynamically linked executables. The reason will become apparent later. For now let’s focus on how dynamically linked executables, as they are by far the most common way people distribute software.

Adding more complication to the mix, TLS is also initialised differently for the main thread and the other threads that begin later in the execution. It makes sense if you stop and think about it, as the main thread is spawned by the kernel, whereas the start of non-main threads are triggered later by the programs.

When it comes to dynamically linked ELF programs, it’s useful to know that once they are loaded and mapped into memory, the kernel would then take its hands off and pass the execution baton to the dynamic linker (ld.so on Linux, dyld on macOS).

Since ld.so is part of glibc, the next step is to open up the guts of glibc and fish around to see if we can dig up anything juicy, but where to look?

Remembering the hypothesis we made earlier about the runtime would use arch_prctl to change FS register, let’s grep the glibc source for arch_prctl then.

After wading through several false positive hits, we arrive at a macro definition TLS_INIT_TP, which uses inline assembly to trigger arch_prctl syscall directly and is responsible for updating the FS register to point to TCB.

With this macro as an anchor and generous use of grep, it starts to become clear that the main thread’s TLS is setup in the function init_tls, which calls _dl_allocate_tls_storage to allocate the TCB or struct pthread, eventually invokes the aforementioned macro TLS_INIT_TP to bind the pthread TCB to the main thread.

This finding confirms our previous hypothesis that the dynamic linker runtime allocates and sets up the TCB or struct pthread and then uses arch_prctl to bind the TLS to at least the main thread.

We will look at how TLS is setup in the other threads later, before that there’s still this glaring question that we’ve been ignoring till now: Where does that mysterious -4 come from?

The Internals of ELF TLS

Depending on the platforms and architectures, there are two variants of TLS/TCB structure, but for x86-64, we will only consider variant 2:

TLS structure courtesy of ELF Handling For Thread-Local Storage:

Despite its simplicity, the diagram actually packs quite a lot of information, let’s decode it:

tpt is the thread register aka thread pointer (i.e. what FS points to) for thread t, and dtvt is Dynamic Thread Vector which can be thought of as a two dimensional array that can address any TLS variable by a module ID and a TLS variable offset. A TLS variable offset is local within a module and bound to each TLS variable at compile time. A module in glibc can refer to either an executable or a dynamically shared object, a module ID therefore is an index number for a loaded ELF object in a process.

Note that for a given running process the module ID for the main executable will always be 1, whereas the shared objects don’t know their module IDs until they are loaded and assigned by the linker3.

As for dtvt,1, dtvt,2, etc, each of them contains TLS information for a loaded shared object (with module ID 1, 2, etc…) and points to a TLS block that contains the TLS variables within a particular dynamic object. A specific TLS variable can be extracted out of this TLS block if its offset is given.

You might have noticed that the first few elements for the DTV array point to the white portion before the TCB, whereas the last couple point to the other TLS blocks labeled for Dynamically-Loaded Modules. Dynamically-loaded modules here don’t mean any dynamic shared objects, they only refer to the shared objects that are loaded by explicitly calling dlopen.

The white portion before the shaded TCB can be subdivided into TLS blocks (one block corresponds to one loaded module, delineated with tlsoffset1, tlsoffset2, etc..), containing all the TLS variables from the executable and the shared objects that are loaded before the start of the main() function. This portion is called static TLS in glibc parlance, they are called “static” to distinguish them from the TLS for modules loaded via dlopen.

The TLS variables for these “static” modules (i.e. DT_NEEDED entries) are reachable by a negative constant offset (constant through out the lifetime of the running process) from the thread register. For example, the instruction mov %fs:0xfffffffffffffffc falls under this category.

Since the module ID for the executable itself is always 1, dtvt,1 (i.e. dtv[1]) always points to the TLS block of the executable itself (module id 1). Additionally, as it can be seen from the diagram, the linker will also always place the executable’s TLS block right next to the TCB, presumably to keep the executable’s TLS variables in a predictable location and relatively hot in the cache.

The dtvt is a member variable of tcbheader_t, which in turn is first member variable of TCB pthread.

DTV has a deceptively simple data structure, but it’s an unfortunate victim of many C trickeries4 (e.g. negative pointer arithmetics, type aliasing, badly named member variables, etc…).

The first element of the dtv array in TCB is a generation counter gent, which serves as a version number. It helps inform the runtime when a resizing or reconstruction of the DTV is needed. Every time when the program does dlopen or dlfree, the global generation counter will get incremented. Whenever the runtime detects a use of a DVT and its gent doesn’t match the global generation number, the DTV will be updated again, its gent will be set to the current global generation number.

As we touched on above, the second element of dtv (i.e. dtv[1]) is pointing to the main executable’s TLS block right next to the TCB in the static TLS.

The following is the quick overview of the dtv array:

dtv[-1].counter; /* Pro tip: The length of this dtv array */
dtv[0].counter; /* Generation counter for the DTV in this thread */
dtv[1].pointer; /* Pointer to the main executable TLS block in this thread */ /* Pointer to a TLS variable defined in a module id `ti_module` */
main_tls_var = *(dtv[tls_index.ti_module].pointer + tls_index.ti_offset);

One property of this static TLS and DTV configuration is that any variable in static TLS in a particular thread is reachable either via the thread’s DTV dtv[ti_module].pointer + ti_offset, or by a negative offset from the TCB (e.g. mov %fs:0xffffffffffxxxxxx %eax, where 0xffffffffffxxxxxx is the negative offset).

It’s important to note that using DTV dtv[ti_module].pointer + ti_offset is the most general way to access TLS variables regardless if the variables reside in static TLS or dynamic TLS. We will see why this is relevant later.

The Initialisation of DTV and static TLS

So far we’ve journeyed from assembly, to kernel, and to dynamic linker, we’ve covered a lot of ground, and it looks like we are getting closer to the understanding of the weird mov instruction, but the questions still remain: how is the static TLS constructed, and where does the -4 or 0xfffffffffffffffc come from?

We’ll explore the static TLS now and find out what the fuzz is all about the -4.

As part of its bookkeeping, the dynamic linker maintains an (intrusive) link list link_map, which keeps track of all the loaded modules and their metadata.

After the linker is done mapping the modules into the program’s address space, it would call init_tls, where it would use the link_map to fill in the information for dl_tls_dtv_slotinfo_list; another global linked list that contains the metadata for modules that make use of TLS. It then proceeds to call _dl_determine_tlsoffset, which lets each loaded module know at what offset (l_tls_offset) in the static TLS it should place the module’s TLS block. Later on the function calls _dl_allocate_tls_storage to allocate (but not initialise) the static TLS and DTV.

Further along in the dynamic linker setup process, _dl_allocate_tls_init will be called to finally initialise the static TLS and DTV for the main thread.

void *
_dl_allocate_tls_init (void *result)
{ if (result == NULL) /* The memory allocation failed. */ return NULL; dtv_t *dtv = GET_DTV (result); struct dtv_slotinfo_list *listp; size_t total = 0; size_t maxgen = 0; /* Check if the current dtv is big enough. */ if (dtv[-1].counter < GL(dl_tls_max_dtv_idx)) { /* Resize the dtv. */ dtv = _dl_resize_dtv (dtv); /* Install this new dtv in the thread data structures. */ INSTALL_DTV (result, &dtv[-1]); } /* We have to prepare the dtv for all currently loaded modules using TLS. For those which are dynamically loaded we add the values indicating deferred allocation. */ listp = GL(dl_tls_dtv_slotinfo_list); while (1) { size_t cnt; for (cnt = total == 0 ? 1 : 0; cnt < listp->len; ++cnt) { struct link_map *map; void *dest; /* Check for the total number of used slots. */ if (total + cnt > GL(dl_tls_max_dtv_idx)) break; map = listp->slotinfo[cnt].map; /* snip */ dest = (char *) result + map->l_tls_offset; /* Set up the DTV entry. The simplified __tls_get_addr that some platforms use in static programs requires it. */ dtv[map->l_tls_modid].pointer.val = dest; /* Copy the initialization image and clear the BSS part. */ memset (__mempcpy (dest, map->l_tls_initimage, map->l_tls_initimage_size), '\0', map->l_tls_blocksize - map->l_tls_initimage_size); } total += cnt; if (total >= GL(dl_tls_max_dtv_idx)) break; listp = listp->next; } /* The DTV version is up-to-date now. */ dtv[0].counter = maxgen; return result;
}

The first part of the function does a sanity check for the size of DTV array, while the second part sets up the DTV with the appropriate offsets in the static TLS and copies the modules’ TLS DATA and BSS sections (tdata and tbss) to the static TLS blocks.

Here it goes, this is how TLS and DTV are setup! But who decides the -4 or 0xfffffffffffffffc is the right offset to get to our TLS variable main_tls_var?

The answer lies in the rule we touched on in the previous section: the main executable’s TLS block is always placed right before the TCB. Essentially the compiler and the dynamic linker conspire together to come up with that number, because the executable’s TLS block and TLS offset are both known ahead of time.

The cooperation relies on the compiler knowing the following “facts” at build time, and the dynamic linker guaranteeing these “facts” are true at runtime:

  • The FS register points to the TCB
  • The TLS variables’ offsets set by the compiler won’t change during runtime
  • The executable’s TLS block sits right before the TCB

You might have noticed that so far we’ve only shown executables using TLS variables that are defined within the executables themselves, how would using them in another scenario differ?

The answer should be a little clearer by the end of the last section: It’s a No, due to the fact that TLS in main executable is treated differently by both compiler and the dynamic linker.

Depending on where a TLS variable is defined and accessed, there are four different cases to examine:

  1. TLS variable locally defined and used within an executable
  2. TLS variable externally defined in a shared object but used in a executable
  3. TLS variable locally defined in a shared object and used in the same shared object
  4. TLS variable externally defined in a shared object and used in an arbitrary shared object

Let’s investigate the four scenarios one by one.

Case 1: TLS variable locally defined and used within an executable

This case is what we’ve studied so far, quick recap: Accessing this type of TLS will be in the form of mov fs:0xffffffffffxxxxxx %xxx.

The code configuration for this scenario is as follows.

A TLS variable is defined in a library code

// libfoo.c

__thread int foo_tls = 42;

The main executable attempts to use it.

// main.c

extern __thread int foo_tls; int main() { return foo_tls;
} 

Disassembling the executable main gives us the following:

000000000000073a <main>: 73a: 55 push %rbp 73b: 48 89 e5 mov %rsp,%rbp 73e: 48 8b 05 93 08 20 00 mov 0x200893(%rip),%rax # 200fd8 <foo_tls> 745: 64 8b 00 mov %fs:(%rax),%eax 748: 5d pop %rbp 749: c3 retq

Unsurprisingly, the instruction to accessing the TLS variable foo_tls is similar but not quite like that of accessing main_tls_var defined within the executable.

It still uses the FS segment register, however instead of using a constant negative offset to reference the variable, it calls out to a value located at the address 0x200fd8, and uses that value as the offset from the TCB to reach the TLS variable foo_tls.

Here is the pseudo C code for the relevant lines, ignoring PIE and RIP relative addressing:

// Retrieve the offset of `foo_tls_offset`
int foo_tls_offset = *((int *) 0x200fd8);
// Pointer arithmetics to get the address of `foo_tls`
int *foo_tls_ptr = thread_context_ptr - foo_tls_offset;
// Dereference tls_var_ptr and put its value in register EAX
EAX = *foo_tls_ptr;

The immediate question is why we can’t use a constant offset to reference foo_tls, followed by what is that 0x200fd8??

The reason why the offset is constant for the first case is due to the fact the compiler knows ahead of time both the variable’s offset and the position of the executable’s TLS block. Whereas for case 2, the compiler only has partial knowledge.

The compiler knows that the TLS variable will be located somewhere within the static TLS. Because the compiler knows foo_tls is externally defined and it also knows it’s defined in libfoo.so which is one of the direct library dependencies of the executable main. Remember static TLS contains TLS variables defined in the executable and shared objects that get loaded before the executable’s main function is reached.

So the compiler knows foo_tls is in the static TLS, it just doesn’t know its precise offset within the static TLS (or its offset from the TCB). As unlike executables, the dynamic linker doesn’t provide any guarantee for the order (therefore their module IDs) in which shared objects get loaded.

To bridge this gap, some form of runtime mechanism is necessary. This is where the mysterious 0x200fd8 comes into play.

In order to delving deeper into this mechanism, we’ll need to know about relocation. Eli Bendersky has written some good articles about ELF relocation. It’s recommended to have basic understanding of why relocation is needed and how it works before reading further.

However if you are really brimming over with patience-defying curiosity right now and so eager to soldier on without reading about relocation, an extremely simplified explanation goes like this: Relocation is required because the compiler doesn’t know where the variables defined in shared objects are located at runtime, so a Global Offset Table (GOT) is set aside by the compiler and only gets filled in the appropriate location values at runtime by the dynamic linker.

Tying back to our discussion about the foo_tls offset, because the compiler doesn’t know what the externally defined TLS variable foo_tls’s offset in the static TLS is, at build time it creates an entry in the global offset table leaving a note (R_X86_64_TPOFF64) about how the offset should be calculated.

Running readelf on our executable to dump its relocation information:

> readelf -r main Relocation section '.rela.dyn' at offset 0x518 contains 9 entries: Offset Info Type Sym. Value Sym. Name + Addend
000000200dc8 000000000008 R_X86_64_RELATIVE 730
000000200dd0 000000000008 R_X86_64_RELATIVE 6f0
000000201008 000000000008 R_X86_64_RELATIVE 201008
000000200fd0 000100000006 R_X86_64_GLOB_DAT 0000000000000000 _ITM_deregisterTMClone + 0
000000200fd8 000200000012 R_X86_64_TPOFF64 0000000000000000 foo_tls + 0
000000200fe0 000300000006 R_X86_64_GLOB_DAT 0000000000000000 __libc_start_main@GLIBC_2.2.5 + 0
000000200fe8 000400000006 R_X86_64_GLOB_DAT 0000000000000000 __gmon_start__ + 0
000000200ff0 000500000006 R_X86_64_GLOB_DAT 0000000000000000 _ITM_registerTMCloneTa + 0
000000200ff8 000600000006 R_X86_64_GLOB_DAT 0000000000000000 __cxa_finalize@GLIBC_2.2.5 + 0

Notice the entry for foo_tls, its offset 000000200fd85, and its type R_X86_64_TPOFF64.

At runtime, after the dynamic linker sets up the TCB and static TLS, it does relocation. As part of the relocation process, the runtime would determine the variable offsets for this type of relocation.

 case R_X86_64_TPOFF64: /* The offset is negative, forward from the thread pointer. */ if (sym != NULL) { CHECK_STATIC_TLS (map, sym_map); /* We know the offset of the object the symbol is contained in. It is a negative value which will be added to the thread pointer. */ value = (sym->st_value + reloc->r_addend - sym_map->l_tls_offset); *reloc_addr = value; } break;

Let’s zoom in at where the offset’s value is calculated:

value = sym->st_value /* The offset of the value within symbol section */ + reloc->addend /* Zero, can be ignored for most cases in x86-64 */ - sym_map->l_tls_offset; /* This is the module's TLS block offset within the static TLS */

After this, the value stored at the address 0x200fd8 will be a negative offset from TCB (where the thread pointer/register points to) and used by the executable to access foo_tls!

Okay, enough executables! Let’s switch gear and take a look at how shared objects access TLS variables.

// libbar.c

static __thread int s_bar_tls; int get_bar_tls() { return s_bar_tls;
}

Make sure to compile it without any optimisation or the compiler may optimise out the TLS access. With a libbar.so in hand, let’s disassemble the function get_bar_tls:

000000000000067a <get_bar_tls>: 67a: 55 push %rbp 67b: 48 89 e5 mov %rsp,%rbp 67e: 66 48 8d 3d 4a 09 20 lea 0x20094a(%rip),%rdi # 200fd0 <.got> 686: 66 66 48 e8 f2 fe ff callq <__tls_get_addr@plt> 68e: 8b 00 mov (%rax),%eax 690: 5d pop %rbp 691: c3 retq 

This time the code to access s_bar_tls looks completely different from before. There’s a mention of .got, which means an entry in the global offset table (GOT). And most importantly there’s a call to __tls_get_addr6.

lea 0x20094a(%rip),%rdi literally just assigns 0x200fd0 (the address of the GOT entry, i.e. a pointer) into the register %rdi. According to x86-64 calling conventions, %rdi is used to pass the first argument of a function call.

See following for the body of the function __tls_get_addr with some macros expanded:

void *
__tls_get_addr (tls_index *ti)
{ dtv_t *dtv = THREAD_DTV (); if (__glibc_unlikely (dtv[0].counter != GL(dl_tls_generation))) return update_get_addr (ti); void *p = dtv[ti->ti_module].pointer.val; if (__glibc_unlikely (p == TLS_DTV_UNALLOCATED)) return tls_get_addr_tail (ti->ti_offset, dtv, NULL); return (char *) p + ti->ti_offset;
}

You may have noticed the similarity of this function and what we talked earlier about using DTV dtv[ti_module].pointer + ti_offset to access any TLS variables. Essentially __tls_get_addr does the same thing but with extra checks and code to handle generation upgrade and lazy TLS block allocation.

We know that the get_bar_tls’s assembly calls __tls_get_addr with a pointer to an entry in the GOT as its argument, but what really is stored in that entry at 0x200fd0?

From the glibc source, we know __tls_get_addr’s first argument is struct tis_index, so naturally the value at that address must be an instance of that struct?

To confirm that, let’s pull out the venerable readelf again, it shows:

> readelf -r libbar.so Relocation section '.rela.dyn' at offset 0x478 contains 8 entries: Offset Info Type Sym. Value Sym. Name + Addend
000000200e00 000000000008 R_X86_64_RELATIVE 670
000000200e08 000000000008 R_X86_64_RELATIVE 630
000000201020 000000000008 R_X86_64_RELATIVE 201020
000000200fd0 000000000010 R_X86_64_DTPMOD64 0
000000200fe0 000100000006 R_X86_64_GLOB_DAT 0000000000000000 __cxa_finalize + 0
000000200fe8 000200000006 R_X86_64_GLOB_DAT 0000000000000000 _ITM_registerTMCloneTa + 0
000000200ff0 000300000006 R_X86_64_GLOB_DAT 0000000000000000 _ITM_deregisterTMClone + 0
000000200ff8 000500000006 R_X86_64_GLOB_DAT 0000000000000000 __gmon_start__ + 0

The entry at offset 000000200fd0 is what we are after. Unlike foo_tls, it doesn’t show the symbol name, that’s because s_bar_tls is locally/statically defined, its name in the symbol table can be ignored. Another thing that’s different from foo_tls is its type, this time around the type column is R_X86_64_DTPMOD64, whereas the type for foo_tls is R_X86_64_TPOFF64.

During relocation, the entry with R_X86_64_DTPMOD64 tells the dynamic linker to find out the module ID (l_tls_modid) for libbar.so at runtime and put that ID value at the address 0x200fd0.

 case R_X86_64_DTPMOD64: /* Get the information from the link map returned by the resolve function. */ if (sym_map != NULL) *reloc_addr = sym_map->l_tls_modid; break;

Astute readers might have noticed that module ID is only the first member variable of struct tis_index.

/* Type used for the representation of TLS information in the GOT. */
typedef struct dl_tls_index
{ uint64_t ti_module; uint64_t ti_offset;
} tls_index;

So when and where does ti_offset get initialised?

Remember bar_tls is statically defined, so the compiler would know its TLS offset, the only thing it didn’t know is the shared object’s module ID at runtime. So the compiler is actually smart enough to place the constant offset directly in the GOT at build time, with the ti_module right next to the constant offset ti_offset. So at runtime only the module ID needs to be relocated.

We can verify this by introducing multiple locally defined TLS variables:

// libbar2.c

static __thread int s_bar_tls1;
static __thread int s_bar_tls2;
static __thread int s_bar_tls3; int get_bar_tls() { return s_bar_tls1 + s_bar_tls2 + s_bar_tls3;
}

Its relocation shows three entries with type R_X86_64_DTPMOD64 as we expect:

> readelf -r libbar2.so Relocation section '.rela.dyn' at offset 0x478 contains 10 entries: Offset Info Type Sym. Value Sym. Name + Addend
000000200de0 000000000008 R_X86_64_RELATIVE 6a0
000000200de8 000000000008 R_X86_64_RELATIVE 660
000000201020 000000000008 R_X86_64_RELATIVE 201020
000000200fb0 000000000010 R_X86_64_DTPMOD64 0
000000200fc0 000000000010 R_X86_64_DTPMOD64 0
000000200fd0 000000000010 R_X86_64_DTPMOD64 0
000000200fe0 000100000006 R_X86_64_GLOB_DAT 0000000000000000 __cxa_finalize + 0
000000200fe8 000200000006 R_X86_64_GLOB_DAT 0000000000000000 _ITM_registerTMCloneTa + 0
000000200ff0 000300000006 R_X86_64_GLOB_DAT 0000000000000000 _ITM_deregisterTMClone + 0
000000200ff8 000500000006 R_X86_64_GLOB_DAT 0000000000000000 __gmon_start__ + 0 Relocation section '.rela.plt' at offset 0x568 contains 1 entry: Offset Info Type Sym. Value Sym. Name + Addend
000000201018 000400000007 R_X86_64_JUMP_SLO 0000000000000000 __tls_get_addr@GLIBC_2.3 + 0

Then check its GOT section:

> readelf -x .got libbar2.so Hex dump of section '.got': 0x00200fb0 00000000 00000000 00000000 00000000 ................ 0x00200fc0 00000000 00000000 04000000 00000000 ................ 0x00200fd0 00000000 00000000 08000000 00000000 ................ 0x00200fe0 00000000 00000000 00000000 00000000 ................ 0x00200ff0 00000000 00000000 00000000 00000000 ................

Notice the fourth and fifth columns of the first three rows: 00000000 00000000, 04000000 00000000, 08000000 00000000. They are three 64-bit integers 0, 4 and 8 in little endian, corresponding three offsets for the three TLS variables (The size of type int is 4 bytes in x86-64 Linux).

This case is somewhat similar to case 3, the crucial difference is that the TLS variable is defined and used in different places7.

// libxyz.c

__thread int xyz_tls;

And xyz_tls is used in another shared object:

// libuvw.c

extern __thread int xyz_tls; int get_xyz_tls() { return xyz_tls;
} 

Compile them with -O0 to avoid over zealous optimisation. And you know the drill, disassemble it:

000000000000066a <get_xyz_tls>: 66a: 55 push %rbp 66b: 48 89 e5 mov %rsp,%rbp 66e: 66 48 8d 3d 5a 09 20 lea 0x20095a(%rip),%rdi # 200fd0 <xyz_tls> 676: 66 66 48 e8 f2 fe ff callq <__tls_get_addr@plt> 67e: 8b 00 mov (%rax),%eax 680: 5d pop %rbp 681: c3 retq 

Interestingly enough, apart from the addresses, the code looks identical to get_bar_tls!

What about readelf?

> readelf -r libuvw.so Relocation section '.rela.dyn' at offset 0x458 contains 9 entries: Offset Info Type Sym. Value Sym. Name + Addend
000000200e00 000000000008 R_X86_64_RELATIVE 660
000000200e08 000000000008 R_X86_64_RELATIVE 620
000000201020 000000000008 R_X86_64_RELATIVE 201020
000000200fd0 000100000010 R_X86_64_DTPMOD64 0000000000000000 xyz_tls + 0
000000200fd8 000100000011 R_X86_64_DTPOFF64 0000000000000000 xyz_tls + 0
000000200fe0 000200000006 R_X86_64_GLOB_DAT 0000000000000000 __cxa_finalize + 0
000000200fe8 000300000006 R_X86_64_GLOB_DAT 0000000000000000 _ITM_registerTMCloneTa + 0
000000200ff0 000400000006 R_X86_64_GLOB_DAT 0000000000000000 _ITM_deregisterTMClone + 0
000000200ff8 000600000006 R_X86_64_GLOB_DAT 0000000000000000 __gmon_start__ + 0 Relocation section '.rela.plt' at offset 0x530 contains 1 entry: Offset Info Type Sym. Value Sym. Name + Addend
000000201018 000500000007 R_X86_64_JUMP_SLO 0000000000000000 __tls_get_addr@GLIBC_2.3 + 0

This time we can see xyz_tls shows up in the symbol name column as they are externally defined. What’s more interesting is that, other than R_X86_64_DTPMOD64 we saw in the previous section, another entry with the type R_X86_64_DTPOFF64 is also in the list.

Because xyz_tls is defined elsewhere in another shared object, the compiler knows neither its module ID nor its offset in the module’s TLS block. In addition to R_X86_64_DTPMOD64, the TLS offset is no longer a constant offset known at build time, the compiler needs more help from the dynamic linker.

R_X86_64_DTPOFF64 is that extra help. This relocation type means that the runtime needs to calculate xyz_tls’s offset within its module’s TLS block.

 case R_X86_64_DTPOFF64: /* During relocation all TLS symbols are defined and used. Therefore the offset is already correct. */ if (sym != NULL) { value = sym->st_value + reloc->r_addend; *reloc_addr = value; } break;

So for this case, each TLS variable requires two relocation steps to fill in enough information to use.

This concludes the four cases of TLS variable usage. At this point you might want to ask why the complexity? The simplest and the most general way of accessing TLS variables is to use __tls_get_addr, but in order to achieve fastest speed for a given condition glibc designers decided to sacrifice simplicity for performance.

Certainly accessing TLS variables via thread register (FS) is perceived to be faster (single memory fetch and more cache friendly) than __tls_get_addr, therefore if the speed of accessing TLS is important to your applications, try to place and use them in the executables when possible. If not, at least try to use them in executables where the variables will at least be in the static TLS. But of course profile first before changing your program architecture.

The Initialisation of TCB or TLS in Non-main Threads

The TCB and TLS is setup slightly differently for non-main threads. When using pthread_create to spawn a new thread, it first needs to allocate a new stack.

The allocate_stack routine first checks if there’s an already cached stack lying around (i.e. old stacks that can be freed after pthread_exit), then allocate a new one if no appropriate stack is in cache.

One interesting tidbit is that after getting hold of a stack, the routine will initialise the static TLS, TCB in place on the stack! So unlike the main thread where TCB and static TLS are heap allocated, the non-main threads just use the stack to house the TLS, saving an extra memory allocation.

After the stack and TLS have been setup properly, pthread_create invokes the clone syscall:

 const int clone_flags = (CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SYSVSEM | CLONE_SIGHAND | CLONE_THREAD | CLONE_SETTLS | CLONE_PARENT_SETTID | CLONE_CHILD_CLEARTID | 0); TLS_DEFINE_INIT_TP (tp, pd); if (__glibc_unlikely (ARCH_CLONE (&start_thread, STACK_VARIABLES_ARGS, clone_flags, pd, &pd->tid, tp, &pd->tid) == -1)) return errno;

Where tp and pd are identical on x86-64, and pointing to the TCB (i.e. struct pthread). Take note that the thread pointer (tp pointing to TCB) is the 6th argument of the clone syscall.

Cracking open clone, we can see it really is just calling out to _do_fork and passing the 6th argument tls to it too.

Following along the parameter tls takes us from _do_fork, to copy_process, to copy_thread_tls, which eventually binds tls to the thread’s FS register.

Conclusion

Finally our journey has come to a well deserved stop, but it has certainly not reached its final destinations. There are still many details left for us to explore:

But at this point the readers should be well equipped to find out themselves!