Radically simplify stack linking.

This commit:

  * Gets rid of 2nd init trampoline.
  * Gets rid of any custom CFI, instead using the standard CFI
    for target prologue.
  * Makes backtraces accessible to tools that only track frame
    pointer chains, like perf or dtrace.
  * Keeps the performance at the exact same level.
This commit is contained in:
whitequark 2016-09-03 12:25:36 +00:00 committed by edef
parent 0ccf72fcb3
commit 491af39fc9
3 changed files with 105 additions and 207 deletions

View File

@ -14,30 +14,6 @@
// while swapping context; this is an arbitrary choice // while swapping context; this is an arbitrary choice
// (we clobber all registers and could use any of them) but this allows us // (we clobber all registers and could use any of them) but this allows us
// to reuse the swap function to perform the initial call. // to reuse the swap function to perform the initial call.
//
// To understand the DWARF CFI code in this file, keep in mind these facts:
// * CFI is "call frame information"; a set of instructions to a debugger or
// an unwinder that allow it to simulate returning from functions. This implies
// restoring every register to its pre-call state, as well as the stack pointer.
// * CFA is "call frame address"; the value of stack pointer right before the call
// instruction in the caller. Everything strictly below CFA (and inclusive until
// the next CFA) is the call frame of the callee. This implies that the return
// address is the part of callee's call frame.
// * Logically, DWARF CFI is a table where rows are instruction pointer values and
// columns describe where registers are spilled (mostly using expressions that
// compute a memory location as CFA+n). A .cfi_offset pseudoinstruction changes
// the state of a column for all IP numerically larger than the one it's placed
// after. A .cfi_def_* pseudoinstruction changes the CFA value similarly.
// * Simulating return is as easy as restoring register values from the CFI table
// and then setting stack pointer to CFA.
//
// A high-level overview of the function of the trampolines is:
// * The 2nd init trampoline puts a controlled value (written in swap to `new_cfa`)
// into r13.
// * The 1st init trampoline tells the unwinder to set r1 to r13, thus continuing
// unwinding at the swap call site instead of falling off the end of context stack.
// * The 1st init trampoline together with the swap trampoline also restore r2
// when unwinding as well as returning normally, because LLVM does not do it for us.
use stack::Stack; use stack::Stack;
pub const STACK_ALIGNMENT: usize = 4; pub const STACK_ALIGNMENT: usize = 4;
@ -47,7 +23,7 @@ pub struct StackPointer(*mut usize);
pub unsafe fn init(stack: &Stack, f: unsafe extern "C" fn(usize) -> !) -> StackPointer { pub unsafe fn init(stack: &Stack, f: unsafe extern "C" fn(usize) -> !) -> StackPointer {
#[naked] #[naked]
unsafe extern "C" fn trampoline_1() { unsafe extern "C" fn trampoline() {
asm!( asm!(
r#" r#"
# gdb has a hardcoded check that rejects backtraces where frame addresses # gdb has a hardcoded check that rejects backtraces where frame addresses
@ -58,37 +34,25 @@ pub unsafe fn init(stack: &Stack, f: unsafe extern "C" fn(usize) -> !) -> StackP
__morestack: __morestack:
.local __morestack .local __morestack
# Set up the first part of our DWARF CFI linking stacks together. # When a normal function is entered, the return address is pushed onto the stack,
# When unwinding the frame corresponding to this function, a DWARF unwinder # and the first thing it does is pushing the frame pointer. The init trampoline
# will use r13 as the next call frame address, restore return address (r9) # is not a normal function; on entry the stack pointer is one word above the place
# from CFA-4 and restore frame pointer (r2) from CFA-8. # where the return address should be, and right under it the return address as
# This mirrors what the second half of `swap_trampoline` does. # well as the stack pointer are already pre-filled. So, simply move the stack
.cfi_def_cfa r13, 0 # pointer where it belongs; and add CFI just like in any other function prologue.
l.addi r1, r1, -8
.cfi_def_cfa_offset 8
.cfi_offset r2, -8 .cfi_offset r2, -8
.cfi_offset r9, -4 l.or r2, r1, r0
# Call the next trampoline. .cfi_def_cfa_register r2
l.j ${0} # Call f.
l.lwz r9, 8(r1)
l.jr r9
l.nop l.nop
.Lend: .Lend:
.size __morestack, .Lend-__morestack .size __morestack, .Lend-__morestack
"# "#
: : "s" (trampoline_2 as usize) : : "volatile")
}
#[naked]
unsafe extern "C" fn trampoline_2() {
asm!(
r#"
# Set up the second part of our DWARF CFI.
# When unwinding the frame corresponding to this function, a DWARF unwinder
# will restore r13 (and thus CFA of the first trampoline) from the stack slot.
.cfi_offset r13, 4
# Call the provided function.
l.lwz r9, 0(r1)
l.jr r9
l.nop
"#
: : : : "volatile") : : : : "volatile")
} }
@ -98,11 +62,10 @@ pub unsafe fn init(stack: &Stack, f: unsafe extern "C" fn(usize) -> !) -> StackP
} }
let mut sp = StackPointer(stack.base() as *mut usize); let mut sp = StackPointer(stack.base() as *mut usize);
push(&mut sp, 0xdead0cfa); // CFA slot
push(&mut sp, f as usize); // function push(&mut sp, f as usize); // function
let rsp = sp.clone(); let rsp = sp;
push(&mut sp, trampoline_1 as usize); // saved r9 push(&mut sp, trampoline as usize); // trampoline / linked return address
push(&mut sp, 0xdeadbbbb); // saved r2 push(&mut sp, 0xdead0bbb); // initial %ebp / linked %ebp
rsp rsp
} }
@ -110,12 +73,17 @@ pub unsafe fn init(stack: &Stack, f: unsafe extern "C" fn(usize) -> !) -> StackP
pub unsafe fn swap(arg: usize, old_sp: *mut StackPointer, new_sp: StackPointer, pub unsafe fn swap(arg: usize, old_sp: *mut StackPointer, new_sp: StackPointer,
new_stack: &Stack) -> usize { new_stack: &Stack) -> usize {
// Address of the topmost CFA stack slot. // Address of the topmost CFA stack slot.
let new_cfa = (new_stack.base() as *mut usize).offset(-1); let new_cfa = (new_stack.base() as *mut usize).offset(-3);
#[naked] #[naked]
unsafe extern "C" fn trampoline() { unsafe extern "C" fn trampoline() {
asm!( asm!(
r#" r#"
# Remember the frame and instruction pointers in the callee, to link
# the stacks together later.
l.or r18, r2, r0
l.or r19, r9, r0
# Save instruction pointer of the old context. # Save instruction pointer of the old context.
l.sw -4(r1), r9 l.sw -4(r1), r9
@ -129,11 +97,18 @@ pub unsafe fn swap(arg: usize, old_sp: *mut StackPointer, new_sp: StackPointer,
# Load stack pointer of the new context. # Load stack pointer of the new context.
l.or r1, r0, r5 l.or r1, r0, r5
# Restore frame pointer of the new context. # Load frame and instruction pointers of the new context.
l.lwz r2, -8(r1) l.lwz r2, -8(r1)
# Return into the new context.
l.lwz r9, -4(r1) l.lwz r9, -4(r1)
# Put the frame and instruction pointers into the trampoline stack frame,
# making it appear to return right after the call instruction that invoked
# this trampoline. This is done after the loads above, since on the very first
# swap, the saved r2/r9 intentionally alias 0(r6)/4(r6).
l.sw 0(r6), r18
l.sw 4(r6), r19
# Return into new context.
l.jr r9 l.jr r9
l.nop l.nop
"# "#
@ -143,9 +118,7 @@ pub unsafe fn swap(arg: usize, old_sp: *mut StackPointer, new_sp: StackPointer,
let ret: usize; let ret: usize;
asm!( asm!(
r#" r#"
# Link the call stacks together. # Push instruction pointer of the old context and switch to
l.sw 0(r6), r1
# Put instruction pointer of the old context into r9 and switch to
# the new context. # the new context.
l.jal ${1} l.jal ${1}
l.nop l.nop

View File

@ -15,30 +15,6 @@
// * i686 SysV C ABI passes the first argument on the stack. This is // * i686 SysV C ABI passes the first argument on the stack. This is
// unfortunate, because unlike every other architecture we can't reuse // unfortunate, because unlike every other architecture we can't reuse
// `swap` for the initial call, and so we use a trampoline. // `swap` for the initial call, and so we use a trampoline.
//
// To understand the DWARF CFI code in this file, keep in mind these facts:
// * CFI is "call frame information"; a set of instructions to a debugger or
// an unwinder that allow it to simulate returning from functions. This implies
// restoring every register to its pre-call state, as well as the stack pointer.
// * CFA is "call frame address"; the value of stack pointer right before the call
// instruction in the caller. Everything strictly below CFA (and inclusive until
// the next CFA) is the call frame of the callee. This implies that the return
// address is the part of callee's call frame.
// * Logically, DWARF CFI is a table where rows are instruction pointer values and
// columns describe where registers are spilled (mostly using expressions that
// compute a memory location as CFA+n). A .cfi_offset pseudoinstruction changes
// the state of a column for all IP numerically larger than the one it's placed
// after. A .cfi_def_* pseudoinstruction changes the CFA value similarly.
// * Simulating return is as easy as restoring register values from the CFI table
// and then setting stack pointer to CFA.
//
// A high-level overview of the function of the trampolines is:
// * The 2nd init trampoline puts a controlled value (written in swap to `new_cfa`)
// into %ebx.
// * The 1st init trampoline tells the unwinder to set %esp to %ebx, thus continuing
// unwinding at the swap call site instead of falling off the end of context stack.
// * The 1st init trampoline together with the swap trampoline also restore %ebp
// when unwinding as well as returning normally, because LLVM does not do it for us.
use stack::Stack; use stack::Stack;
pub const STACK_ALIGNMENT: usize = 16; pub const STACK_ALIGNMENT: usize = 16;
@ -48,7 +24,7 @@ pub struct StackPointer(*mut usize);
pub unsafe fn init(stack: &Stack, f: unsafe extern "C" fn(usize) -> !) -> StackPointer { pub unsafe fn init(stack: &Stack, f: unsafe extern "C" fn(usize) -> !) -> StackPointer {
#[naked] #[naked]
unsafe extern "C" fn trampoline_1() { unsafe extern "C" fn trampoline() {
asm!( asm!(
r#" r#"
# gdb has a hardcoded check that rejects backtraces where frame addresses # gdb has a hardcoded check that rejects backtraces where frame addresses
@ -59,36 +35,24 @@ pub unsafe fn init(stack: &Stack, f: unsafe extern "C" fn(usize) -> !) -> StackP
__morestack: __morestack:
.local __morestack .local __morestack
# Set up the first part of our DWARF CFI linking stacks together. # When a normal function is entered, the return address is pushed onto the stack,
# When unwinding the frame corresponding to this function, a DWARF unwinder # and the first thing it does is pushing the frame pointer. The init trampoline
# will use %ebx as the next call frame address, restore return address # is not a normal function; on entry the stack pointer is one word above the place
# from CFA-4 and restore %ebp from CFA-8. This mirrors what the second half # where the return address should be, and right under it the return address as
# of `swap_trampoline` does. # well as the stack pointer are already pre-filled. So, simply move the stack
.cfi_def_cfa %ebx, 0 # pointer where it belongs; and add CFI just like in any other function prologue.
subl $$8, %esp
.cfi_def_cfa_offset 8
.cfi_offset %ebp, -8 .cfi_offset %ebp, -8
# Call the next trampoline. movl %esp, %ebp
call ${0:c} .cfi_def_cfa_register %ebp
# Call f.
pushl %eax
calll *12(%esp)
.Lend: .Lend:
.size __morestack, .Lend-__morestack .size __morestack, .Lend-__morestack
"# "#
: : "s" (trampoline_2 as usize) : : "volatile")
}
#[naked]
unsafe extern "C" fn trampoline_2() {
asm!(
r#"
# Set up the second part of our DWARF CFI.
# When unwinding the frame corresponding to this function, a DWARF unwinder
# will restore %ebx (and thus CFA of the first trampoline) from the stack slot.
.cfi_offset %ebx, 4
# Push argument.
.cfi_def_cfa_offset 8
pushl %eax
# Call the provided function.
call *8(%esp)
"#
: : : : "volatile") : : : : "volatile")
} }
@ -98,10 +62,9 @@ pub unsafe fn init(stack: &Stack, f: unsafe extern "C" fn(usize) -> !) -> StackP
} }
let mut sp = StackPointer(stack.base() as *mut usize); let mut sp = StackPointer(stack.base() as *mut usize);
push(&mut sp, 0xdead0cfa); // CFA slot
push(&mut sp, f as usize); // function push(&mut sp, f as usize); // function
push(&mut sp, trampoline_1 as usize); push(&mut sp, trampoline as usize); // trampoline / linked return address
push(&mut sp, 0xdeadbbbb); // saved %ebp push(&mut sp, 0xdead0bbb); // initial %ebp / linked %ebp
sp sp
} }
@ -109,12 +72,19 @@ pub unsafe fn init(stack: &Stack, f: unsafe extern "C" fn(usize) -> !) -> StackP
pub unsafe fn swap(arg: usize, old_sp: *mut StackPointer, new_sp: StackPointer, pub unsafe fn swap(arg: usize, old_sp: *mut StackPointer, new_sp: StackPointer,
new_stack: &Stack) -> usize { new_stack: &Stack) -> usize {
// Address of the topmost CFA stack slot. // Address of the topmost CFA stack slot.
let new_cfa = (new_stack.base() as *mut usize).offset(-1); let new_cfa = (new_stack.base() as *mut usize).offset(-3);
#[naked] #[naked]
unsafe extern "C" fn trampoline() { unsafe extern "C" fn trampoline() {
asm!( asm!(
r#" r#"
# Remember the frame and instruction pointers in the callee, to link
# the stacks together later. We put them on stack because x86 doesn't
# have enough registers.
movl %ebp, -8(%edx)
movl (%esp), %ebx
movl %ebx, -12(%edx)
# Save frame pointer explicitly; the unwinder uses it to find CFA of # Save frame pointer explicitly; the unwinder uses it to find CFA of
# the caller, and so it has to have the correct value immediately after # the caller, and so it has to have the correct value immediately after
# the call instruction that invoked the trampoline. # the call instruction that invoked the trampoline.
@ -125,12 +95,20 @@ pub unsafe fn swap(arg: usize, old_sp: *mut StackPointer, new_sp: StackPointer,
# Load stack pointer of the new context. # Load stack pointer of the new context.
movl %edx, %esp movl %edx, %esp
# Restore frame pointer of the new context. # Load frame and instruction pointers of the new context.
popl %ebp popl %ebp
# Return into the new context. Use `pop` and `jmp` instead of a `ret`
# to avoid return address mispredictions (~8ns per `ret` on Ivy Bridge).
popl %ebx popl %ebx
# Put the frame and instruction pointers into the trampoline stack frame,
# making it appear to return right after the call instruction that invoked
# this trampoline. This is done after the loads above, since on the very first
# swap, the saved %ebp/%ebx intentionally alias 0(%edi)/4(%edi).
movl -8(%edx), %esi
movl %esi, 0(%edi)
movl -12(%edx), %esi
movl %esi, 4(%edi)
# Return into new context.
jmpl *%ebx jmpl *%ebx
"# "#
: : : : "volatile") : : : : "volatile")
@ -139,8 +117,6 @@ pub unsafe fn swap(arg: usize, old_sp: *mut StackPointer, new_sp: StackPointer,
let ret: usize; let ret: usize;
asm!( asm!(
r#" r#"
# Link the call stacks together.
movl %esp, (%edi)
# Push instruction pointer of the old context and switch to # Push instruction pointer of the old context and switch to
# the new context. # the new context.
call ${1:c} call ${1:c}

View File

@ -19,30 +19,6 @@
// to pass a value while swapping context; this is an arbitrary choice // to pass a value while swapping context; this is an arbitrary choice
// (we clobber all registers and could use any of them) but this allows us // (we clobber all registers and could use any of them) but this allows us
// to reuse the swap function to perform the initial call. // to reuse the swap function to perform the initial call.
//
// To understand the DWARF CFI code in this file, keep in mind these facts:
// * CFI is "call frame information"; a set of instructions to a debugger or
// an unwinder that allow it to simulate returning from functions. This implies
// restoring every register to its pre-call state, as well as the stack pointer.
// * CFA is "call frame address"; the value of stack pointer right before the call
// instruction in the caller. Everything strictly below CFA (and inclusive until
// the next CFA) is the call frame of the callee. This implies that the return
// address is the part of callee's call frame.
// * Logically, DWARF CFI is a table where rows are instruction pointer values and
// columns describe where registers are spilled (mostly using expressions that
// compute a memory location as CFA+n). A .cfi_offset pseudoinstruction changes
// the state of a column for all IP numerically larger than the one it's placed
// after. A .cfi_def_* pseudoinstruction changes the CFA value similarly.
// * Simulating return is as easy as restoring register values from the CFI table
// and then setting stack pointer to CFA.
//
// A high-level overview of the function of the trampolines is:
// * The 2nd init trampoline puts a controlled value (written in swap to `new_cfa`)
// into %rbx.
// * The 1st init trampoline tells the unwinder to set %rsp to %rbx, thus continuing
// unwinding at the swap call site instead of falling off the end of context stack.
// * The 1st init trampoline together with the swap trampoline also restore %rbp
// when unwinding as well as returning normally, because LLVM does not do it for us.
use stack::Stack; use stack::Stack;
pub const STACK_ALIGNMENT: usize = 16; pub const STACK_ALIGNMENT: usize = 16;
@ -51,9 +27,8 @@ pub const STACK_ALIGNMENT: usize = 16;
pub struct StackPointer(*mut usize); pub struct StackPointer(*mut usize);
pub unsafe fn init(stack: &Stack, f: unsafe extern "C" fn(usize) -> !) -> StackPointer { pub unsafe fn init(stack: &Stack, f: unsafe extern "C" fn(usize) -> !) -> StackPointer {
#[cfg(not(target_vendor = "apple"))]
#[naked] #[naked]
unsafe extern "C" fn trampoline_1() { unsafe extern "C" fn trampoline() {
asm!( asm!(
r#" r#"
# gdb has a hardcoded check that rejects backtraces where frame addresses # gdb has a hardcoded check that rejects backtraces where frame addresses
@ -64,57 +39,23 @@ pub unsafe fn init(stack: &Stack, f: unsafe extern "C" fn(usize) -> !) -> StackP
__morestack: __morestack:
.local __morestack .local __morestack
# Set up the first part of our DWARF CFI linking stacks together. # When a normal function is entered, the return address is pushed onto the stack,
# When unwinding the frame corresponding to this function, a DWARF unwinder # and the first thing it does is pushing the frame pointer. The init trampoline
# will use %rbx as the next call frame address, restore return address # is not a normal function; on entry the stack pointer is one word above the place
# from CFA-8 and restore %rbp from CFA-16. This mirrors what the second half # where the return address should be, and right under it the return address as
# of `swap_trampoline` does. # well as the stack pointer are already pre-filled. So, simply move the stack
.cfi_def_cfa %rbx, 0 # pointer where it belongs; and add CFI just like in any other function prologue.
subq $$16, %rsp
.cfi_def_cfa_offset 16
.cfi_offset %rbp, -16 .cfi_offset %rbp, -16
# Call the next trampoline. movq %rsp, %rbp
call ${0:c} .cfi_def_cfa_register %rbp
# Call f.
callq *16(%rsp)
.Lend: .Lend:
.size __morestack, .Lend-__morestack .size __morestack, .Lend-__morestack
"# "#
: : "s" (trampoline_2 as usize) : : "volatile")
}
#[cfg(target_vendor = "apple")]
#[naked]
unsafe extern "C" fn trampoline_1() {
asm!(
r#"
# Same as above; however, .local and .size are not supported in Mach-O.
__morestack:
.private_extern __morestack
.cfi_def_cfa %rbx, 0
.cfi_offset %rbp, -16
call ${0:c}
"#
: : "s" (trampoline_2 as usize) : : "volatile")
}
#[naked]
unsafe extern "C" fn trampoline_2() {
asm!(
r#"
# Set up the second part of our DWARF CFI.
# When unwinding the frame corresponding to this function, a DWARF unwinder
# will restore %rbx (and thus CFA of the first trampoline) from the stack slot.
#
# The following is functionally equivalent to:
# .cfi_offset %rbx, 16
# however positive offsets in .cfi_offset translate to DW_CFA_offset_extended_sf,
# and ld64's CFI parser only supports regular DW_CFA_offset (which only supports
# negative offsets, with the sign being implicit), so to avoid crashing the linker
# on OS X, fold offset into DW_CFA_def_offset.
.cfi_def_cfa_offset 24
.cfi_offset %rip, -24
.cfi_offset %rbx, 0
# Call the provided function.
call *8(%rsp)
"#
: : : : "volatile") : : : : "volatile")
} }
@ -124,11 +65,10 @@ pub unsafe fn init(stack: &Stack, f: unsafe extern "C" fn(usize) -> !) -> StackP
} }
let mut sp = StackPointer(stack.base() as *mut usize); let mut sp = StackPointer(stack.base() as *mut usize);
push(&mut sp, 0xdeaddeaddead0cfa); // CFA slot
push(&mut sp, 0 as usize); // alignment push(&mut sp, 0 as usize); // alignment
push(&mut sp, f as usize); // function push(&mut sp, f as usize); // function
push(&mut sp, trampoline_1 as usize); push(&mut sp, trampoline as usize); // trampoline / linked return address
push(&mut sp, 0xdeaddeaddeadbbbb); // saved %rbp push(&mut sp, 0xdeaddeaddead0bbb); // initial %rbp / linked %rbp
sp sp
} }
@ -136,12 +76,17 @@ pub unsafe fn init(stack: &Stack, f: unsafe extern "C" fn(usize) -> !) -> StackP
pub unsafe fn swap(arg: usize, old_sp: *mut StackPointer, new_sp: StackPointer, pub unsafe fn swap(arg: usize, old_sp: *mut StackPointer, new_sp: StackPointer,
new_stack: &Stack) -> usize { new_stack: &Stack) -> usize {
// Address of the topmost CFA stack slot. // Address of the topmost CFA stack slot.
let new_cfa = (new_stack.base() as *mut usize).offset(-1); let new_cfa = (new_stack.base() as *mut usize).offset(-4);
#[naked] #[naked]
unsafe extern "C" fn trampoline() { unsafe extern "C" fn trampoline() {
asm!( asm!(
r#" r#"
# Remember the frame and instruction pointers in the callee, to link
# the stacks together later.
movq %rbp, %r8
movq (%rsp), %r9
# Save frame pointer explicitly; the unwinder uses it to find CFA of # Save frame pointer explicitly; the unwinder uses it to find CFA of
# the caller, and so it has to have the correct value immediately after # the caller, and so it has to have the correct value immediately after
# the call instruction that invoked the trampoline. # the call instruction that invoked the trampoline.
@ -152,12 +97,18 @@ pub unsafe fn swap(arg: usize, old_sp: *mut StackPointer, new_sp: StackPointer,
# Load stack pointer of the new context. # Load stack pointer of the new context.
movq %rdx, %rsp movq %rdx, %rsp
# Restore frame pointer of the new context. # Load frame and instruction pointers of the new context.
popq %rbp popq %rbp
# Return into the new context. Use `pop` and `jmp` instead of a `ret`
# to avoid return address mispredictions (~8ns per `ret` on Ivy Bridge).
popq %rbx popq %rbx
# Put the frame and instruction pointers into the trampoline stack frame,
# making it appear to return right after the call instruction that invoked
# this trampoline. This is done after the loads above, since on the very first
# swap, the saved %rbp/%rbx intentionally alias 0(%rcx)/8(%rcx).
movq %r8, 0(%rcx)
movq %r9, 8(%rcx)
# Return into new context.
jmpq *%rbx jmpq *%rbx
"# "#
: : : : "volatile") : : : : "volatile")
@ -166,8 +117,6 @@ pub unsafe fn swap(arg: usize, old_sp: *mut StackPointer, new_sp: StackPointer,
let ret: usize; let ret: usize;
asm!( asm!(
r#" r#"
# Link the call stacks together.
movq %rsp, (%rcx)
# Push instruction pointer of the old context and switch to # Push instruction pointer of the old context and switch to
# the new context. # the new context.
call ${1:c} call ${1:c}