std.os.linux.tls: Refactor and improve documentation.

* Elaborate on the sub-variants of Variant I.
* Clarify the use of the TCB term.
* Rename a bunch of stuff to be more accurate/descriptive.
* Follow Zig's style around namespacing more.
* Use a structure for the ABI TCB.

No functional change intended.
This commit is contained in:
Alex Rønne Petersen 2024-07-23 00:57:25 +02:00
parent 38e0f049c5
commit b52e054261
No known key found for this signature in database
3 changed files with 273 additions and 192 deletions

View File

@ -1261,9 +1261,9 @@ const LinuxThreadImpl = struct {
bytes = std.mem.alignForward(usize, bytes, page_size);
stack_offset = bytes;
bytes = std.mem.alignForward(usize, bytes, linux.tls.tls_image.alloc_align);
bytes = std.mem.alignForward(usize, bytes, linux.tls.area_desc.alignment);
tls_offset = bytes;
bytes += linux.tls.tls_image.alloc_size;
bytes += linux.tls.area_desc.size;
bytes = std.mem.alignForward(usize, bytes, @alignOf(Instance));
instance_offset = bytes;
@ -1304,12 +1304,12 @@ const LinuxThreadImpl = struct {
};
// Prepare the TLS segment and prepare a user_desc struct when needed on x86
var tls_ptr = linux.tls.prepareTLS(mapped[tls_offset..]);
var tls_ptr = linux.tls.prepareArea(mapped[tls_offset..]);
var user_desc: if (target.cpu.arch == .x86) linux.user_desc else void = undefined;
if (target.cpu.arch == .x86) {
defer tls_ptr = @intFromPtr(&user_desc);
user_desc = .{
.entry_number = linux.tls.tls_image.gdt_entry_number,
.entry_number = linux.tls.area_desc.gdt_entry_number,
.base_addr = tls_ptr,
.limit = 0xfffff,
.flags = .{

View File

@ -1,3 +1,14 @@
//! This file implements the two TLS variants [1] used by ELF-based systems. Note that, in reality,
//! Variant I has two sub-variants.
//!
//! It is important to understand that the term TCB (Thread Control Block) is overloaded here.
//! Official ABI documentation uses it simply to mean the ABI TCB, i.e. a small area of ABI-defined
//! data, usually one or two words (see the `AbiTcb` type below). People will also often use TCB to
//! refer to the libc TCB, which can be any size and contain anything. (One could even omit it!) We
//! refer to the latter as the Zig TCB; see the `ZigTcb` type below.
//!
//! [1] https://www.akkadia.org/drepper/tls.pdf
const std = @import("std");
const mem = std.mem;
const elf = std.elf;
@ -7,56 +18,58 @@ const native_arch = @import("builtin").cpu.arch;
const linux = std.os.linux;
const posix = std.posix;
// This file implements the two TLS variants [1] used by ELF-based systems.
//
// The variant I has the following layout in memory:
// -------------------------------------------------------
// | DTV | Zig | DTV | Alignment | TLS |
// | storage | thread data | pointer | | block |
// ------------------------^------------------------------
// `-- The thread pointer register points here
//
// In this case we allocate additional space for our control structure that's
// placed _before_ the DTV pointer together with the DTV.
//
// NOTE: Some systems such as power64 or mips use this variant with a twist: the
// alignment is not present and the tp and DTV addresses are offset by a
// constant.
//
// On the other hand the variant II has the following layout in memory:
// ---------------------------------------
// | TLS | TCB | Zig | DTV |
// | block | | thread data | storage |
// --------^------------------------------
// `-- The thread pointer register points here
//
// The structure of the TCB is not defined by the ABI so we reserve enough space
// for a single pointer as some architectures such as x86 and x86_64 need a
// pointer to the TCB block itself at the address pointed by the tp.
//
// In this case the control structure and DTV are placed one after another right
// after the TLS block data.
//
// At the moment the DTV is very simple since we only support static TLS, all we
// need is a two word vector to hold the number of entries (1) and the address
// of the first TLS block.
//
// [1] https://www.akkadia.org/drepper/tls.pdf
const TLSVariant = enum {
VariantI,
VariantII,
/// Represents an ELF TLS variant.
///
/// In all variants, the TP and the TLS blocks must be aligned to the `p_align` value in the
/// `PT_TLS` ELF program header. Everything else has natural alignment.
///
/// The location of the DTV does not actually matter. For simplicity, we put it in the TLS area, but
/// there is no actual ABI requirement that it reside there.
const Variant = enum {
/// The original Variant I:
///
/// ----------------------------------------
/// | DTV | Zig TCB | ABI TCB | TLS Blocks |
/// ----------------^-----------------------
/// `-- The TP register points here.
///
/// The layout in this variant necessitates separate alignment of both the TP and the TLS
/// blocks.
///
/// The first word in the ABI TCB points to the DTV. For some architectures, there may be a
/// second word with an unspecified meaning.
I_original,
/// The modified Variant I:
///
/// ---------------------------------------------------
/// | DTV | Zig TCB | ABI TCB | [Offset] | TLS Blocks |
/// -------------------------------------^-------------
/// `-- The TP register points here.
///
/// The offset (which can be zero) is applied to the TP only; there is never physical gap
/// between the ABI TCB and the TLS blocks. This implies that we only need to align the TP.
///
/// The first (and only) word in the ABI TCB points to the DTV.
I_modified,
/// Variant II:
///
/// ----------------------------------------
/// | TLS Blocks | ABI TCB | Zig TCB | DTV |
/// -------------^--------------------------
/// `-- The TP register points here.
///
/// The first (and only) word in the ABI TCB points to the ABI TCB itself.
II,
};
const tls_variant = switch (native_arch) {
const current_variant: Variant = switch (native_arch) {
.arm,
.armeb,
.thumb,
.thumbeb,
.aarch64,
.aarch64_be,
.riscv32,
.riscv64,
.thumb,
.thumbeb,
=> .I_original,
.mips,
.mipsel,
.mips64,
@ -65,73 +78,126 @@ const tls_variant = switch (native_arch) {
.powerpcle,
.powerpc64,
.powerpc64le,
=> TLSVariant.VariantI,
.x86_64, .x86, .sparc64 => TLSVariant.VariantII,
else => @compileError("undefined tls_variant for this architecture"),
.riscv32,
.riscv64,
=> .I_modified,
.sparc64,
.x86,
.x86_64,
=> .II,
else => @compileError("undefined TLS variant for this architecture"),
};
// Controls how many bytes are reserved for the Thread Control Block
const tls_tcb_size = switch (native_arch) {
// ARM EABI mandates enough space for two pointers: the first one points to
// the DTV while the second one is unspecified but reserved
.arm, .armeb, .thumb, .thumbeb, .aarch64, .aarch64_be => 2 * @sizeOf(usize),
// One pointer-sized word that points either to the DTV or the TCB itself
else => @sizeOf(usize),
};
// Controls if the TP points to the end of the TCB instead of its beginning
const tls_tp_points_past_tcb = switch (native_arch) {
.riscv32, .riscv64, .mips, .mipsel, .mips64, .mips64el, .powerpc, .powerpcle, .powerpc64, .powerpc64le => true,
else => false,
};
// Some architectures add some offset to the tp and dtv addresses in order to
// make the generated code more efficient
const tls_tp_offset = switch (native_arch) {
.mips, .mipsel, .mips64, .mips64el, .powerpc, .powerpcle, .powerpc64, .powerpc64le => 0x7000,
/// The Offset value for the modified Variant I.
const current_tp_offset = switch (native_arch) {
.mips,
.mipsel,
.mips64,
.mips64el,
.powerpc,
.powerpcle,
.powerpc64,
.powerpc64le,
=> 0x7000,
else => 0,
};
const tls_dtv_offset = switch (native_arch) {
.mips, .mipsel, .mips64, .mips64el, .powerpc, .powerpcle, .powerpc64, .powerpc64le => 0x8000,
.riscv32, .riscv64 => 0x800,
/// Usually only used by the modified Variant I.
const current_dtv_offset = switch (native_arch) {
.mips,
.mipsel,
.mips64,
.mips64el,
.powerpc,
.powerpcle,
.powerpc64,
.powerpc64le,
=> 0x8000,
.riscv32,
.riscv64,
=> 0x800,
else => 0,
};
// Per-thread storage for Zig's use
const CustomData = struct {
/// Per-thread storage for the ELF TLS ABI.
const AbiTcb = switch (current_variant) {
.I_original, .I_modified => switch (native_arch) {
// ARM EABI mandates enough space for two pointers: the first one points to the DTV as
// usual, while the second one is unspecified.
.aarch64,
.aarch64_be,
.arm,
.armeb,
.thumb,
.thumbeb,
=> extern struct {
/// This is offset by `current_dtv_offset`.
dtv: usize,
reserved: ?*anyopaque,
},
else => extern struct {
/// This is offset by `current_dtv_offset`.
dtv: usize,
},
},
.II => extern struct {
/// This is self-referential.
self: *AbiTcb,
},
};
/// Per-thread storage for Zig's use. Currently unused.
const ZigTcb = struct {
dummy: usize,
};
// Dynamic Thread Vector
const DTV = extern struct {
entries: usize,
tls_block: [1][*]u8,
/// Dynamic Thread Vector as specified in the ELF TLS ABI. Ordinarily, there is a block pointer per
/// dynamically-loaded module, but since we only support static TLS, we only need one block pointer.
const Dtv = extern struct {
len: usize = 1,
tls_block: [*]u8,
};
// Holds all the information about the process TLS image
const TLSImage = struct {
init_data: []const u8,
alloc_size: usize,
alloc_align: usize,
tcb_offset: usize,
dtv_offset: usize,
data_offset: usize,
data_size: usize,
// Only used on the x86 architecture
/// Describes a process's TLS area. The area encompasses the DTV, both TCBs, and the TLS block, with
/// the exact layout of these being dependent primarily on `current_variant`.
const AreaDesc = struct {
size: usize,
alignment: usize,
dtv: struct {
/// Offset into the TLS area.
offset: usize,
},
abi_tcb: struct {
/// Offset into the TLS area.
offset: usize,
},
block: struct {
/// The initial data to be copied into the TLS block. Note that this may be smaller than
/// `size`, in which case any remaining data in the TLS block is simply left uninitialized.
init: []const u8,
/// Offset into the TLS area.
offset: usize,
/// This is the effective size of the TLS block, which may be greater than `init.len`.
size: usize,
},
/// Only used on the 32-bit x86 architecture (not x86_64, nor x32).
gdt_entry_number: usize,
};
pub var tls_image: TLSImage = undefined;
pub var area_desc: AreaDesc = undefined;
pub fn setThreadPointer(addr: usize) void {
@setRuntimeSafety(false);
@disableInstrumentation();
switch (native_arch) {
.x86 => {
var user_desc: linux.user_desc = .{
.entry_number = tls_image.gdt_entry_number,
.entry_number = area_desc.gdt_entry_number,
.base_addr = addr,
.limit = 0xfffff,
.flags = .{
@ -148,7 +214,7 @@ pub fn setThreadPointer(addr: usize) void {
const gdt_entry_number = user_desc.entry_number;
// We have to keep track of our slot as it's also needed for clone()
tls_image.gdt_entry_number = gdt_entry_number;
area_desc.gdt_entry_number = gdt_entry_number;
// Update the %gs selector
asm volatile ("movl %[gs_val], %%gs"
:
@ -206,7 +272,7 @@ pub fn setThreadPointer(addr: usize) void {
}
}
fn initTLS(phdrs: []elf.Phdr) void {
fn computeAreaDesc(phdrs: []elf.Phdr) void {
@setRuntimeSafety(false);
@disableInstrumentation();
@ -221,72 +287,85 @@ fn initTLS(phdrs: []elf.Phdr) void {
}
}
var tls_align_factor: usize = undefined;
var tls_data: []const u8 = undefined;
var tls_data_alloc_size: usize = undefined;
var align_factor: usize = undefined;
var block_init: []const u8 = undefined;
var block_size: usize = undefined;
if (tls_phdr) |phdr| {
// The effective size in memory is represented by p_memsz, the length of
// the data stored in the PT_TLS segment is p_filesz and may be less
// than the former
tls_align_factor = phdr.p_align;
tls_data = @as([*]u8, @ptrFromInt(img_base + phdr.p_vaddr))[0..phdr.p_filesz];
tls_data_alloc_size = phdr.p_memsz;
align_factor = phdr.p_align;
// The effective size in memory is represented by `p_memsz`; the length of the data stored
// in the `PT_TLS` segment is `p_filesz` and may be less than the former.
block_init = @as([*]u8, @ptrFromInt(img_base + phdr.p_vaddr))[0..phdr.p_filesz];
block_size = phdr.p_memsz;
} else {
tls_align_factor = @alignOf(usize);
tls_data = &[_]u8{};
tls_data_alloc_size = 0;
align_factor = @alignOf(usize);
block_init = &[_]u8{};
block_size = 0;
}
// Offsets into the allocated TLS area
var tcb_offset: usize = undefined;
// Offsets into the allocated TLS area.
var dtv_offset: usize = undefined;
var data_offset: usize = undefined;
// Compute the total size of the ABI-specific data plus our own control
// structures. All the offset calculated here assume a well-aligned base
// address.
const alloc_size = switch (tls_variant) {
.VariantI => blk: {
var abi_tcb_offset: usize = undefined;
var block_offset: usize = undefined;
// Compute the total size of the ABI-specific data plus our own `ZigTcb` structure. All the
// offsets calculated here assume a well-aligned base address.
const area_size = switch (current_variant) {
.I_original, .I_modified => blk: {
var l: usize = 0;
dtv_offset = l;
l += @sizeOf(DTV);
// Add some padding here so that the thread pointer (tcb_offset) is
// aligned to p_align and the CustomData structure can be found by
// simply subtracting its @sizeOf from the tp value
const delta = (l + @sizeOf(CustomData)) & (tls_align_factor - 1);
l += @sizeOf(Dtv);
// Add some padding here so that the TP (`abi_tcb_offset`) is aligned to `align_factor`
// and the `ZigTcb` structure can be found by simply subtracting `@sizeOf(ZigTcb)` from
// the TP.
const delta = (l + @sizeOf(ZigTcb)) & (align_factor - 1);
if (delta > 0)
l += tls_align_factor - delta;
l += @sizeOf(CustomData);
tcb_offset = l;
l += alignForward(tls_tcb_size, tls_align_factor);
data_offset = l;
l += tls_data_alloc_size;
l += align_factor - delta;
l += @sizeOf(ZigTcb);
abi_tcb_offset = l;
l += alignForward(@sizeOf(AbiTcb), align_factor);
block_offset = l;
l += block_size;
break :blk l;
},
.VariantII => blk: {
.II => blk: {
var l: usize = 0;
data_offset = l;
l += alignForward(tls_data_alloc_size, tls_align_factor);
// The thread pointer is aligned to p_align
tcb_offset = l;
l += tls_tcb_size;
// The CustomData structure is right after the TCB with no padding
// in between so it can be easily found
l += @sizeOf(CustomData);
l = alignForward(l, @alignOf(DTV));
block_offset = l;
l += alignForward(block_size, align_factor);
// The TP is aligned to `align_factor`.
abi_tcb_offset = l;
l += @sizeOf(AbiTcb);
// The `ZigTcb` structure is right after the `AbiTcb` with no padding in between so it
// can be easily found.
l += @sizeOf(ZigTcb);
// It doesn't really matter where we put the DTV, so give it natural alignment.
l = alignForward(l, @alignOf(Dtv));
dtv_offset = l;
l += @sizeOf(DTV);
l += @sizeOf(Dtv);
break :blk l;
},
};
tls_image = TLSImage{
.init_data = tls_data,
.alloc_size = alloc_size,
.alloc_align = tls_align_factor,
.tcb_offset = tcb_offset,
.dtv_offset = dtv_offset,
.data_offset = data_offset,
.data_size = tls_data_alloc_size,
area_desc = .{
.size = area_size,
.alignment = align_factor,
.dtv = .{
.offset = dtv_offset,
},
.abi_tcb = .{
.offset = abi_tcb_offset,
},
.block = .{
.init = block_init,
.offset = block_offset,
.size = block_size,
},
.gdt_entry_number = @as(usize, @bitCast(@as(isize, -1))),
};
}
@ -306,78 +385,80 @@ inline fn alignPtrCast(comptime T: type, ptr: [*]u8) *T {
return @ptrCast(@alignCast(ptr));
}
/// Initializes all the fields of the static TLS area and returns the computed
/// architecture-specific value of the thread-pointer register
///
/// This function is inline because thread local storage is not set up yet.
pub fn prepareTLS(area: []u8) usize {
/// Initializes all the fields of the static TLS area and returns the computed architecture-specific
/// value of the TP register.
pub fn prepareArea(area: []u8) usize {
@setRuntimeSafety(false);
@disableInstrumentation();
// Clear the area we're going to use, just to be safe
@memset(area, 0);
// Prepare the DTV
const dtv = alignPtrCast(DTV, area.ptr + tls_image.dtv_offset);
dtv.entries = 1;
dtv.tls_block[0] = area.ptr + tls_dtv_offset + tls_image.data_offset;
// Prepare the TCB
const tcb_ptr = alignPtrCast([*]u8, area.ptr + tls_image.tcb_offset);
tcb_ptr.* = switch (tls_variant) {
.VariantI => area.ptr + tls_image.dtv_offset,
.VariantII => area.ptr + tls_image.tcb_offset,
};
// Copy the data
@memcpy(area[tls_image.data_offset..][0..tls_image.init_data.len], tls_image.init_data);
// Return the corrected value (if needed) for the tp register.
// Overflow here is not a problem, the pointer arithmetic involving the tp
// is done with wrapping semantics.
return @intFromPtr(area.ptr) +% tls_tp_offset +%
if (tls_tp_points_past_tcb) tls_image.data_offset else tls_image.tcb_offset;
// Clear the area we're going to use, just to be safe.
@memset(area, 0);
// Prepare the ABI TCB.
const abi_tcb = alignPtrCast(AbiTcb, area.ptr + area_desc.abi_tcb.offset);
switch (current_variant) {
.I_original, .I_modified => abi_tcb.dtv = @intFromPtr(area.ptr + area_desc.dtv.offset),
.II => abi_tcb.self = abi_tcb,
}
// Prepare the DTV.
const dtv = alignPtrCast(Dtv, area.ptr + area_desc.dtv.offset);
dtv.len = 1;
dtv.tls_block = area.ptr + current_dtv_offset + area_desc.block.offset;
// Copy the initial data.
@memcpy(area[area_desc.block.offset..][0..area_desc.block.init.len], area_desc.block.init);
// Return the corrected value (if needed) for the TP register. Overflow here is not a problem;
// the pointer arithmetic involving the TP is done with wrapping semantics.
return @intFromPtr(area.ptr) +% switch (current_variant) {
.I_original, .II => area_desc.abi_tcb.offset,
.I_modified => area_desc.block.offset +% current_tp_offset,
};
}
// The main motivation for the size chosen here is this is how much ends up being
// requested for the thread local variables of the std.crypto.random implementation.
// I'm not sure why it ends up being so much; the struct itself is only 64 bytes.
// I think it has to do with being page aligned and LLVM or LLD is not smart enough
// to lay out the TLS data in a space conserving way. Anyway I think it's fine
// because it's less than 3 pages of memory, and putting it in the ELF like this
// is equivalent to moving the mmap call below into the kernel, avoiding syscall
// overhead.
var main_thread_tls_buffer: [0x2100]u8 align(mem.page_size) = undefined;
// The main motivation for the size chosen here is that this is how much ends up being requested for
// the thread-local variables of the `std.crypto.random` implementation. I'm not sure why it ends up
// being so much; the struct itself is only 64 bytes. I think it has to do with being page-aligned
// and LLVM or LLD is not smart enough to lay out the TLS data in a space-conserving way. Anyway, I
// think it's fine because it's less than 3 pages of memory, and putting it in the ELF like this is
// equivalent to moving the `mmap` call below into the kernel, avoiding syscall overhead.
var main_thread_area_buffer: [0x2100]u8 align(mem.page_size) = undefined;
pub fn initStaticTLS(phdrs: []elf.Phdr) void {
/// Computes the layout of the static TLS area, allocates the area, initializes all of its fields,
/// and assigns the architecture-specific value to the TP register.
pub fn initStatic(phdrs: []elf.Phdr) void {
@setRuntimeSafety(false);
@disableInstrumentation();
initTLS(phdrs);
computeAreaDesc(phdrs);
const tls_area = blk: {
// Fast path for the common case where the TLS data is really small,
// avoid an allocation and use our local buffer.
if (tls_image.alloc_align <= mem.page_size and
tls_image.alloc_size <= main_thread_tls_buffer.len)
{
break :blk main_thread_tls_buffer[0..tls_image.alloc_size];
const area = blk: {
// Fast path for the common case where the TLS data is really small, avoid an allocation and
// use our local buffer.
if (area_desc.alignment <= mem.page_size and area_desc.size <= main_thread_area_buffer.len) {
break :blk main_thread_area_buffer[0..area_desc.size];
}
const begin_addr = mmap(
null,
tls_image.alloc_size + tls_image.alloc_align - 1,
area_desc.size + area_desc.alignment - 1,
posix.PROT.READ | posix.PROT.WRITE,
.{ .TYPE = .PRIVATE, .ANONYMOUS = true },
-1,
0,
);
if (@as(isize, @bitCast(begin_addr)) < 0) @trap();
const alloc_tls_area: [*]align(mem.page_size) u8 = @ptrFromInt(begin_addr);
const area_ptr: [*]align(mem.page_size) u8 = @ptrFromInt(begin_addr);
// Make sure the slice is correctly aligned.
const begin_aligned_addr = alignForward(begin_addr, tls_image.alloc_align);
const begin_aligned_addr = alignForward(begin_addr, area_desc.alignment);
const start = begin_aligned_addr - begin_addr;
break :blk alloc_tls_area[start..][0..tls_image.alloc_size];
break :blk area_ptr[start..][0..area_desc.size];
};
const tp_value = prepareTLS(tls_area);
const tp_value = prepareArea(area);
setThreadPointer(tp_value);
}

View File

@ -456,7 +456,7 @@ fn posixCallMainAndExit(argc_argv_ptr: [*]usize) callconv(.C) noreturn {
}
// Initialize the TLS area.
std.os.linux.tls.initStaticTLS(phdrs);
std.os.linux.tls.initStatic(phdrs);
}
// The way Linux executables represent stack size is via the PT_GNU_STACK