1
0
mirror of https://github.com/golang/go synced 2024-11-19 23:24:45 -07:00
go/src/pkg/runtime/asm_386.s

636 lines
14 KiB
ArmAsm
Raw Normal View History

// Copyright 2009 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
#include "zasm_GOOS_GOARCH.h"
TEXT _rt0_386(SB),7,$0
// copy arguments forward on an even stack
MOVL 0(SP), AX // argc
LEAL 4(SP), BX // argv
SUBL $128, SP // plenty of scratch
ANDL $~15, SP
MOVL AX, 120(SP) // save argc, argv away
MOVL BX, 124(SP)
// set default stack bounds.
// initcgo may update stackguard.
MOVL $runtime·g0(SB), BP
LEAL (-64*1024+104)(SP), BX
MOVL BX, g_stackguard(BP)
MOVL SP, g_stackbase(BP)
// if there is an initcgo, call it to let it
// initialize and to set up GS. if not,
// we set up GS ourselves.
MOVL initcgo(SB), AX
TESTL AX, AX
JZ needtls
PUSHL BP
CALL AX
POPL BP
// skip runtime·ldt0setup(SB) and tls test after initcgo for non-windows
CMPL runtime·iswindows(SB), $0
JEQ ok
needtls:
// skip runtime·ldt0setup(SB) and tls test on Plan 9 in all cases
CMPL runtime·isplan9(SB), $1
JEQ ok
// set up %gs
CALL runtime·ldt0setup(SB)
// store through it, to make sure it works
get_tls(BX)
MOVL $0x123, g(BX)
MOVL runtime·tls0(SB), AX
CMPL AX, $0x123
JEQ ok
MOVL AX, 0 // abort
ok:
// set up m and g "registers"
get_tls(BX)
LEAL runtime·g0(SB), CX
MOVL CX, g(BX)
LEAL runtime·m0(SB), AX
MOVL AX, m(BX)
// save m->g0 = g0
MOVL CX, m_g0(AX)
CALL runtime·emptyfunc(SB) // fault if stack check is wrong
// convention is D is always cleared
CLD
CALL runtime·check(SB)
// saved argc, argv
MOVL 120(SP), AX
MOVL AX, 0(SP)
MOVL 124(SP), AX
MOVL AX, 4(SP)
CALL runtime·args(SB)
CALL runtime·osinit(SB)
CALL runtime·schedinit(SB)
// create a new goroutine to start program
PUSHL $runtime·main(SB) // entry
PUSHL $0 // arg size
CALL runtime·newproc(SB)
POPL AX
POPL AX
// start this M
CALL runtime·mstart(SB)
INT $3
RET
TEXT runtime·breakpoint(SB),7,$0
INT $3
RET
TEXT runtime·asminit(SB),7,$0
// Linux, Windows start the FPU in extended double precision.
// Other operating systems use double precision.
// Change to double precision to match them,
// and to match other hardware that only has double.
PUSHL $0x27F
FLDCW 0(SP)
POPL AX
RET
/*
* go-routine
*/
runtime: scheduler, cgo reorganization * Change use of m->g0 stack (aka scheduler stack). * Provide runtime.mcall(f) to invoke f() on m->g0 stack. * Replace scheduler loop entry with runtime.mcall(schedule). Runtime.mcall eliminates the need for fake scheduler states that exist just to run a bit of code on the m->g0 stack (Grecovery, Gstackalloc). The elimination of the scheduler as a loop that stops and starts using gosave and gogo fixes a bad interaction with the way cgo uses the m->g0 stack. Cgo runs external (gcc-compiled) C functions on that stack, and then when calling back into Go, it sets m->g0->sched.sp below the added call frames, so that other uses of m->g0's stack will not interfere with those frames. Unfortunately, gogo (longjmp) back to the scheduler loop at this point would end up running scheduler with the lower sp, which no longer points at a valid stack frame for a call to scheduler. If scheduler then wrote any function call arguments or local variables to where it expected the stack frame to be, it would overwrite other data on the stack. I realized this possibility while debugging a problem with calling complex Go code in a Go -> C -> Go cgo callback. This wasn't the bug I was looking for, it turns out, but I believe it is a real bug nonetheless. Switching to runtime.mcall, which only adds new frames to the stack and never jumps into functions running in existing ones, fixes this bug. * Move cgo-related code out of proc.c into cgocall.c. * Add very large comment describing cgo call sequences. * Simpilify, regularize cgo function implementations and names. * Add test suite as misc/cgo/test. Now the Go -> C path calls cgocall, which calls asmcgocall, and the C -> Go path calls cgocallback, which calls cgocallbackg. The shuffling, which affects mainly the callback case, moves most of the callback implementation to cgocallback running on the m->curg stack (not the m->g0 scheduler stack) and only while accounted for with $GOMAXPROCS (between calls to exitsyscall and entersyscall). The previous callback code did not block in startcgocallback's approximation to exitsyscall, so if, say, the garbage collector were running, it would still barge in and start doing things like call malloc. Similarly endcgocallback's approximation of entersyscall did not call matchmg to kick off new OS threads when necessary, which caused the bug in issue 1560. Fixes #1560. R=iant CC=golang-dev https://golang.org/cl/4253054
2011-03-07 08:37:42 -07:00
// void gosave(Gobuf*)
// save state in Gobuf; setjmp
TEXT runtime·gosave(SB), 7, $0
MOVL 4(SP), AX // gobuf
LEAL 4(SP), BX // caller's SP
MOVL BX, gobuf_sp(AX)
MOVL 0(SP), BX // caller's PC
MOVL BX, gobuf_pc(AX)
get_tls(CX)
MOVL g(CX), BX
MOVL BX, gobuf_g(AX)
RET
// void gogo(Gobuf*, uintptr)
// restore state from Gobuf; longjmp
TEXT runtime·gogo(SB), 7, $0
MOVL 8(SP), AX // return 2nd arg
MOVL 4(SP), BX // gobuf
MOVL gobuf_g(BX), DX
MOVL 0(DX), CX // make sure g != nil
get_tls(CX)
MOVL DX, g(CX)
MOVL gobuf_sp(BX), SP // restore SP
MOVL gobuf_pc(BX), BX
JMP BX
// void gogocall(Gobuf*, void (*fn)(void))
// restore state from Gobuf but then call fn.
// (call fn, returning to state in Gobuf)
TEXT runtime·gogocall(SB), 7, $0
MOVL 8(SP), AX // fn
MOVL 4(SP), BX // gobuf
MOVL gobuf_g(BX), DX
get_tls(CX)
MOVL DX, g(CX)
MOVL 0(DX), CX // make sure g != nil
MOVL gobuf_sp(BX), SP // restore SP
MOVL gobuf_pc(BX), BX
PUSHL BX
JMP AX
POPL BX // not reached
runtime: scheduler, cgo reorganization * Change use of m->g0 stack (aka scheduler stack). * Provide runtime.mcall(f) to invoke f() on m->g0 stack. * Replace scheduler loop entry with runtime.mcall(schedule). Runtime.mcall eliminates the need for fake scheduler states that exist just to run a bit of code on the m->g0 stack (Grecovery, Gstackalloc). The elimination of the scheduler as a loop that stops and starts using gosave and gogo fixes a bad interaction with the way cgo uses the m->g0 stack. Cgo runs external (gcc-compiled) C functions on that stack, and then when calling back into Go, it sets m->g0->sched.sp below the added call frames, so that other uses of m->g0's stack will not interfere with those frames. Unfortunately, gogo (longjmp) back to the scheduler loop at this point would end up running scheduler with the lower sp, which no longer points at a valid stack frame for a call to scheduler. If scheduler then wrote any function call arguments or local variables to where it expected the stack frame to be, it would overwrite other data on the stack. I realized this possibility while debugging a problem with calling complex Go code in a Go -> C -> Go cgo callback. This wasn't the bug I was looking for, it turns out, but I believe it is a real bug nonetheless. Switching to runtime.mcall, which only adds new frames to the stack and never jumps into functions running in existing ones, fixes this bug. * Move cgo-related code out of proc.c into cgocall.c. * Add very large comment describing cgo call sequences. * Simpilify, regularize cgo function implementations and names. * Add test suite as misc/cgo/test. Now the Go -> C path calls cgocall, which calls asmcgocall, and the C -> Go path calls cgocallback, which calls cgocallbackg. The shuffling, which affects mainly the callback case, moves most of the callback implementation to cgocallback running on the m->curg stack (not the m->g0 scheduler stack) and only while accounted for with $GOMAXPROCS (between calls to exitsyscall and entersyscall). The previous callback code did not block in startcgocallback's approximation to exitsyscall, so if, say, the garbage collector were running, it would still barge in and start doing things like call malloc. Similarly endcgocallback's approximation of entersyscall did not call matchmg to kick off new OS threads when necessary, which caused the bug in issue 1560. Fixes #1560. R=iant CC=golang-dev https://golang.org/cl/4253054
2011-03-07 08:37:42 -07:00
// void mcall(void (*fn)(G*))
// Switch to m->g0's stack, call fn(g).
runtime: stack split + garbage collection bug The g->sched.sp saved stack pointer and the g->stackbase and g->stackguard stack bounds can change even while "the world is stopped", because a goroutine has to call functions (and therefore might split its stack) when exiting a system call to check whether the world is stopped (and if so, wait until the world continues). That means the garbage collector cannot access those values safely (without a race) for goroutines executing system calls. Instead, save a consistent triple in g->gcsp, g->gcstack, g->gcguard during entersyscall and have the garbage collector refer to those. The old code was occasionally seeing (because of the race) an sp and stk that did not correspond to each other, so that stk - sp was not the number of stack bytes following sp. In that case, if sp < stk then the call scanblock(sp, stk - sp) scanned too many bytes (anything between the two pointers, which pointed into different allocation blocks). If sp > stk then stk - sp wrapped around. On 32-bit, stk - sp is a uintptr (uint32) converted to int64 in the call to scanblock, so a large (~4G) but positive number. Scanblock would try to scan that many bytes and eventually fault accessing unmapped memory. On 64-bit, stk - sp is a uintptr (uint64) promoted to int64 in the call to scanblock, so a negative number. Scanblock would not scan anything, possibly causing in-use blocks to be freed. In short, 32-bit platforms would have seen either ineffective garbage collection or crashes during garbage collection, while 64-bit platforms would have seen either ineffective or incorrect garbage collection. You can see the invalid arguments to scanblock in the stack traces in issue 1620. Fixes #1620. Fixes #1746. R=iant, r CC=golang-dev https://golang.org/cl/4437075
2011-04-27 21:21:12 -06:00
// Fn must never return. It should gogo(&g->sched)
runtime: scheduler, cgo reorganization * Change use of m->g0 stack (aka scheduler stack). * Provide runtime.mcall(f) to invoke f() on m->g0 stack. * Replace scheduler loop entry with runtime.mcall(schedule). Runtime.mcall eliminates the need for fake scheduler states that exist just to run a bit of code on the m->g0 stack (Grecovery, Gstackalloc). The elimination of the scheduler as a loop that stops and starts using gosave and gogo fixes a bad interaction with the way cgo uses the m->g0 stack. Cgo runs external (gcc-compiled) C functions on that stack, and then when calling back into Go, it sets m->g0->sched.sp below the added call frames, so that other uses of m->g0's stack will not interfere with those frames. Unfortunately, gogo (longjmp) back to the scheduler loop at this point would end up running scheduler with the lower sp, which no longer points at a valid stack frame for a call to scheduler. If scheduler then wrote any function call arguments or local variables to where it expected the stack frame to be, it would overwrite other data on the stack. I realized this possibility while debugging a problem with calling complex Go code in a Go -> C -> Go cgo callback. This wasn't the bug I was looking for, it turns out, but I believe it is a real bug nonetheless. Switching to runtime.mcall, which only adds new frames to the stack and never jumps into functions running in existing ones, fixes this bug. * Move cgo-related code out of proc.c into cgocall.c. * Add very large comment describing cgo call sequences. * Simpilify, regularize cgo function implementations and names. * Add test suite as misc/cgo/test. Now the Go -> C path calls cgocall, which calls asmcgocall, and the C -> Go path calls cgocallback, which calls cgocallbackg. The shuffling, which affects mainly the callback case, moves most of the callback implementation to cgocallback running on the m->curg stack (not the m->g0 scheduler stack) and only while accounted for with $GOMAXPROCS (between calls to exitsyscall and entersyscall). The previous callback code did not block in startcgocallback's approximation to exitsyscall, so if, say, the garbage collector were running, it would still barge in and start doing things like call malloc. Similarly endcgocallback's approximation of entersyscall did not call matchmg to kick off new OS threads when necessary, which caused the bug in issue 1560. Fixes #1560. R=iant CC=golang-dev https://golang.org/cl/4253054
2011-03-07 08:37:42 -07:00
// to keep running g.
TEXT runtime·mcall(SB), 7, $0
MOVL fn+0(FP), DI
get_tls(CX)
MOVL g(CX), AX // save state in g->gobuf
MOVL 0(SP), BX // caller's PC
MOVL BX, (g_sched+gobuf_pc)(AX)
LEAL 4(SP), BX // caller's SP
MOVL BX, (g_sched+gobuf_sp)(AX)
MOVL AX, (g_sched+gobuf_g)(AX)
// switch to m->g0 & its stack, call fn
MOVL m(CX), BX
MOVL m_g0(BX), SI
CMPL SI, AX // if g == m->g0 call badmcall
JNE 2(PC)
CALL runtime·badmcall(SB)
MOVL SI, g(CX) // g = m->g0
MOVL (g_sched+gobuf_sp)(SI), SP // sp = m->g0->gobuf.sp
PUSHL AX
CALL DI
POPL AX
CALL runtime·badmcall2(SB)
RET
/*
* support for morestack
*/
// Called during function prolog when more stack is needed.
TEXT runtime·morestack(SB),7,$0
// Cannot grow scheduler stack (m->g0).
get_tls(CX)
MOVL m(CX), BX
MOVL m_g0(BX), SI
CMPL g(CX), SI
JNE 2(PC)
INT $3
// frame size in DX
// arg size in AX
// Save in m.
MOVL DX, m_moreframesize(BX)
MOVL AX, m_moreargsize(BX)
// Called from f.
// Set m->morebuf to f's caller.
MOVL 4(SP), DI // f's caller's PC
MOVL DI, (m_morebuf+gobuf_pc)(BX)
LEAL 8(SP), CX // f's caller's SP
MOVL CX, (m_morebuf+gobuf_sp)(BX)
MOVL CX, m_moreargp(BX)
get_tls(CX)
MOVL g(CX), SI
MOVL SI, (m_morebuf+gobuf_g)(BX)
// Set m->morepc to f's PC.
MOVL 0(SP), AX
MOVL AX, m_morepc(BX)
runtime: scheduler, cgo reorganization * Change use of m->g0 stack (aka scheduler stack). * Provide runtime.mcall(f) to invoke f() on m->g0 stack. * Replace scheduler loop entry with runtime.mcall(schedule). Runtime.mcall eliminates the need for fake scheduler states that exist just to run a bit of code on the m->g0 stack (Grecovery, Gstackalloc). The elimination of the scheduler as a loop that stops and starts using gosave and gogo fixes a bad interaction with the way cgo uses the m->g0 stack. Cgo runs external (gcc-compiled) C functions on that stack, and then when calling back into Go, it sets m->g0->sched.sp below the added call frames, so that other uses of m->g0's stack will not interfere with those frames. Unfortunately, gogo (longjmp) back to the scheduler loop at this point would end up running scheduler with the lower sp, which no longer points at a valid stack frame for a call to scheduler. If scheduler then wrote any function call arguments or local variables to where it expected the stack frame to be, it would overwrite other data on the stack. I realized this possibility while debugging a problem with calling complex Go code in a Go -> C -> Go cgo callback. This wasn't the bug I was looking for, it turns out, but I believe it is a real bug nonetheless. Switching to runtime.mcall, which only adds new frames to the stack and never jumps into functions running in existing ones, fixes this bug. * Move cgo-related code out of proc.c into cgocall.c. * Add very large comment describing cgo call sequences. * Simpilify, regularize cgo function implementations and names. * Add test suite as misc/cgo/test. Now the Go -> C path calls cgocall, which calls asmcgocall, and the C -> Go path calls cgocallback, which calls cgocallbackg. The shuffling, which affects mainly the callback case, moves most of the callback implementation to cgocallback running on the m->curg stack (not the m->g0 scheduler stack) and only while accounted for with $GOMAXPROCS (between calls to exitsyscall and entersyscall). The previous callback code did not block in startcgocallback's approximation to exitsyscall, so if, say, the garbage collector were running, it would still barge in and start doing things like call malloc. Similarly endcgocallback's approximation of entersyscall did not call matchmg to kick off new OS threads when necessary, which caused the bug in issue 1560. Fixes #1560. R=iant CC=golang-dev https://golang.org/cl/4253054
2011-03-07 08:37:42 -07:00
// Call newstack on m->g0's stack.
MOVL m_g0(BX), BP
MOVL BP, g(CX)
runtime: scheduler, cgo reorganization * Change use of m->g0 stack (aka scheduler stack). * Provide runtime.mcall(f) to invoke f() on m->g0 stack. * Replace scheduler loop entry with runtime.mcall(schedule). Runtime.mcall eliminates the need for fake scheduler states that exist just to run a bit of code on the m->g0 stack (Grecovery, Gstackalloc). The elimination of the scheduler as a loop that stops and starts using gosave and gogo fixes a bad interaction with the way cgo uses the m->g0 stack. Cgo runs external (gcc-compiled) C functions on that stack, and then when calling back into Go, it sets m->g0->sched.sp below the added call frames, so that other uses of m->g0's stack will not interfere with those frames. Unfortunately, gogo (longjmp) back to the scheduler loop at this point would end up running scheduler with the lower sp, which no longer points at a valid stack frame for a call to scheduler. If scheduler then wrote any function call arguments or local variables to where it expected the stack frame to be, it would overwrite other data on the stack. I realized this possibility while debugging a problem with calling complex Go code in a Go -> C -> Go cgo callback. This wasn't the bug I was looking for, it turns out, but I believe it is a real bug nonetheless. Switching to runtime.mcall, which only adds new frames to the stack and never jumps into functions running in existing ones, fixes this bug. * Move cgo-related code out of proc.c into cgocall.c. * Add very large comment describing cgo call sequences. * Simpilify, regularize cgo function implementations and names. * Add test suite as misc/cgo/test. Now the Go -> C path calls cgocall, which calls asmcgocall, and the C -> Go path calls cgocallback, which calls cgocallbackg. The shuffling, which affects mainly the callback case, moves most of the callback implementation to cgocallback running on the m->curg stack (not the m->g0 scheduler stack) and only while accounted for with $GOMAXPROCS (between calls to exitsyscall and entersyscall). The previous callback code did not block in startcgocallback's approximation to exitsyscall, so if, say, the garbage collector were running, it would still barge in and start doing things like call malloc. Similarly endcgocallback's approximation of entersyscall did not call matchmg to kick off new OS threads when necessary, which caused the bug in issue 1560. Fixes #1560. R=iant CC=golang-dev https://golang.org/cl/4253054
2011-03-07 08:37:42 -07:00
MOVL (g_sched+gobuf_sp)(BP), AX
MOVL -4(AX), BX // fault if CALL would, before smashing SP
MOVL AX, SP
CALL runtime·newstack(SB)
MOVL $0, 0x1003 // crash if newstack returns
RET
// Called from reflection library. Mimics morestack,
// reuses stack growth code to create a frame
// with the desired args running the desired function.
//
// func call(fn *byte, arg *byte, argsize uint32).
TEXT reflect·call(SB), 7, $0
get_tls(CX)
MOVL m(CX), BX
// Save our caller's state as the PC and SP to
// restore when returning from f.
MOVL 0(SP), AX // our caller's PC
MOVL AX, (m_morebuf+gobuf_pc)(BX)
LEAL 4(SP), AX // our caller's SP
MOVL AX, (m_morebuf+gobuf_sp)(BX)
MOVL g(CX), AX
MOVL AX, (m_morebuf+gobuf_g)(BX)
// Set up morestack arguments to call f on a new stack.
// We set f's frame size to 1, as a hint to newstack
// that this is a call from reflect·call.
// If it turns out that f needs a larger frame than
// the default stack, f's usual stack growth prolog will
// allocate a new segment (and recopy the arguments).
MOVL 4(SP), AX // fn
MOVL 8(SP), DX // arg frame
MOVL 12(SP), CX // arg size
MOVL AX, m_morepc(BX) // f's PC
MOVL DX, m_moreargp(BX) // f's argument pointer
MOVL CX, m_moreargsize(BX) // f's argument size
MOVL $1, m_moreframesize(BX) // f's frame size
runtime: scheduler, cgo reorganization * Change use of m->g0 stack (aka scheduler stack). * Provide runtime.mcall(f) to invoke f() on m->g0 stack. * Replace scheduler loop entry with runtime.mcall(schedule). Runtime.mcall eliminates the need for fake scheduler states that exist just to run a bit of code on the m->g0 stack (Grecovery, Gstackalloc). The elimination of the scheduler as a loop that stops and starts using gosave and gogo fixes a bad interaction with the way cgo uses the m->g0 stack. Cgo runs external (gcc-compiled) C functions on that stack, and then when calling back into Go, it sets m->g0->sched.sp below the added call frames, so that other uses of m->g0's stack will not interfere with those frames. Unfortunately, gogo (longjmp) back to the scheduler loop at this point would end up running scheduler with the lower sp, which no longer points at a valid stack frame for a call to scheduler. If scheduler then wrote any function call arguments or local variables to where it expected the stack frame to be, it would overwrite other data on the stack. I realized this possibility while debugging a problem with calling complex Go code in a Go -> C -> Go cgo callback. This wasn't the bug I was looking for, it turns out, but I believe it is a real bug nonetheless. Switching to runtime.mcall, which only adds new frames to the stack and never jumps into functions running in existing ones, fixes this bug. * Move cgo-related code out of proc.c into cgocall.c. * Add very large comment describing cgo call sequences. * Simpilify, regularize cgo function implementations and names. * Add test suite as misc/cgo/test. Now the Go -> C path calls cgocall, which calls asmcgocall, and the C -> Go path calls cgocallback, which calls cgocallbackg. The shuffling, which affects mainly the callback case, moves most of the callback implementation to cgocallback running on the m->curg stack (not the m->g0 scheduler stack) and only while accounted for with $GOMAXPROCS (between calls to exitsyscall and entersyscall). The previous callback code did not block in startcgocallback's approximation to exitsyscall, so if, say, the garbage collector were running, it would still barge in and start doing things like call malloc. Similarly endcgocallback's approximation of entersyscall did not call matchmg to kick off new OS threads when necessary, which caused the bug in issue 1560. Fixes #1560. R=iant CC=golang-dev https://golang.org/cl/4253054
2011-03-07 08:37:42 -07:00
// Call newstack on m->g0's stack.
MOVL m_g0(BX), BP
get_tls(CX)
MOVL BP, g(CX)
runtime: scheduler, cgo reorganization * Change use of m->g0 stack (aka scheduler stack). * Provide runtime.mcall(f) to invoke f() on m->g0 stack. * Replace scheduler loop entry with runtime.mcall(schedule). Runtime.mcall eliminates the need for fake scheduler states that exist just to run a bit of code on the m->g0 stack (Grecovery, Gstackalloc). The elimination of the scheduler as a loop that stops and starts using gosave and gogo fixes a bad interaction with the way cgo uses the m->g0 stack. Cgo runs external (gcc-compiled) C functions on that stack, and then when calling back into Go, it sets m->g0->sched.sp below the added call frames, so that other uses of m->g0's stack will not interfere with those frames. Unfortunately, gogo (longjmp) back to the scheduler loop at this point would end up running scheduler with the lower sp, which no longer points at a valid stack frame for a call to scheduler. If scheduler then wrote any function call arguments or local variables to where it expected the stack frame to be, it would overwrite other data on the stack. I realized this possibility while debugging a problem with calling complex Go code in a Go -> C -> Go cgo callback. This wasn't the bug I was looking for, it turns out, but I believe it is a real bug nonetheless. Switching to runtime.mcall, which only adds new frames to the stack and never jumps into functions running in existing ones, fixes this bug. * Move cgo-related code out of proc.c into cgocall.c. * Add very large comment describing cgo call sequences. * Simpilify, regularize cgo function implementations and names. * Add test suite as misc/cgo/test. Now the Go -> C path calls cgocall, which calls asmcgocall, and the C -> Go path calls cgocallback, which calls cgocallbackg. The shuffling, which affects mainly the callback case, moves most of the callback implementation to cgocallback running on the m->curg stack (not the m->g0 scheduler stack) and only while accounted for with $GOMAXPROCS (between calls to exitsyscall and entersyscall). The previous callback code did not block in startcgocallback's approximation to exitsyscall, so if, say, the garbage collector were running, it would still barge in and start doing things like call malloc. Similarly endcgocallback's approximation of entersyscall did not call matchmg to kick off new OS threads when necessary, which caused the bug in issue 1560. Fixes #1560. R=iant CC=golang-dev https://golang.org/cl/4253054
2011-03-07 08:37:42 -07:00
MOVL (g_sched+gobuf_sp)(BP), SP
CALL runtime·newstack(SB)
MOVL $0, 0x1103 // crash if newstack returns
RET
// Return point when leaving stack.
TEXT runtime·lessstack(SB), 7, $0
// Save return value in m->cret
get_tls(CX)
MOVL m(CX), BX
MOVL AX, m_cret(BX)
runtime: scheduler, cgo reorganization * Change use of m->g0 stack (aka scheduler stack). * Provide runtime.mcall(f) to invoke f() on m->g0 stack. * Replace scheduler loop entry with runtime.mcall(schedule). Runtime.mcall eliminates the need for fake scheduler states that exist just to run a bit of code on the m->g0 stack (Grecovery, Gstackalloc). The elimination of the scheduler as a loop that stops and starts using gosave and gogo fixes a bad interaction with the way cgo uses the m->g0 stack. Cgo runs external (gcc-compiled) C functions on that stack, and then when calling back into Go, it sets m->g0->sched.sp below the added call frames, so that other uses of m->g0's stack will not interfere with those frames. Unfortunately, gogo (longjmp) back to the scheduler loop at this point would end up running scheduler with the lower sp, which no longer points at a valid stack frame for a call to scheduler. If scheduler then wrote any function call arguments or local variables to where it expected the stack frame to be, it would overwrite other data on the stack. I realized this possibility while debugging a problem with calling complex Go code in a Go -> C -> Go cgo callback. This wasn't the bug I was looking for, it turns out, but I believe it is a real bug nonetheless. Switching to runtime.mcall, which only adds new frames to the stack and never jumps into functions running in existing ones, fixes this bug. * Move cgo-related code out of proc.c into cgocall.c. * Add very large comment describing cgo call sequences. * Simpilify, regularize cgo function implementations and names. * Add test suite as misc/cgo/test. Now the Go -> C path calls cgocall, which calls asmcgocall, and the C -> Go path calls cgocallback, which calls cgocallbackg. The shuffling, which affects mainly the callback case, moves most of the callback implementation to cgocallback running on the m->curg stack (not the m->g0 scheduler stack) and only while accounted for with $GOMAXPROCS (between calls to exitsyscall and entersyscall). The previous callback code did not block in startcgocallback's approximation to exitsyscall, so if, say, the garbage collector were running, it would still barge in and start doing things like call malloc. Similarly endcgocallback's approximation of entersyscall did not call matchmg to kick off new OS threads when necessary, which caused the bug in issue 1560. Fixes #1560. R=iant CC=golang-dev https://golang.org/cl/4253054
2011-03-07 08:37:42 -07:00
// Call oldstack on m->g0's stack.
MOVL m_g0(BX), BP
MOVL BP, g(CX)
MOVL (g_sched+gobuf_sp)(BP), SP
CALL runtime·oldstack(SB)
MOVL $0, 0x1004 // crash if oldstack returns
RET
// bool cas(int32 *val, int32 old, int32 new)
// Atomically:
// if(*val == old){
// *val = new;
// return 1;
// }else
// return 0;
TEXT runtime·cas(SB), 7, $0
MOVL 4(SP), BX
MOVL 8(SP), AX
MOVL 12(SP), CX
LOCK
CMPXCHGL CX, 0(BX)
JZ 3(PC)
MOVL $0, AX
RET
MOVL $1, AX
RET
// bool runtime·cas64(uint64 *val, uint64 *old, uint64 new)
// Atomically:
// if(*val == *old){
// *val = new;
// return 1;
// } else {
// *old = *val
// return 0;
// }
TEXT runtime·cas64(SB), 7, $0
MOVL 4(SP), BP
MOVL 8(SP), SI
MOVL 0(SI), AX
MOVL 4(SI), DX
MOVL 12(SP), BX
MOVL 16(SP), CX
LOCK
CMPXCHG8B 0(BP)
JNZ cas64_fail
MOVL $1, AX
RET
cas64_fail:
MOVL AX, 0(SI)
MOVL DX, 4(SI)
MOVL $0, AX
RET
// bool casp(void **p, void *old, void *new)
// Atomically:
// if(*p == old){
// *p = new;
// return 1;
// }else
// return 0;
TEXT runtime·casp(SB), 7, $0
MOVL 4(SP), BX
MOVL 8(SP), AX
MOVL 12(SP), CX
LOCK
CMPXCHGL CX, 0(BX)
JZ 3(PC)
MOVL $0, AX
RET
MOVL $1, AX
RET
// uint32 xadd(uint32 volatile *val, int32 delta)
// Atomically:
// *val += delta;
// return *val;
TEXT runtime·xadd(SB), 7, $0
MOVL 4(SP), BX
MOVL 8(SP), AX
MOVL AX, CX
LOCK
XADDL AX, 0(BX)
ADDL CX, AX
RET
runtime: improve Linux mutex The implementation is hybrid active/passive spin/blocking mutex. The design minimizes amount of context switches and futex calls. The idea is that all critical sections in runtime are intentially small, so pure blocking mutex behaves badly causing a lot of context switches, thread parking/unparking and kernel calls. Note that some synthetic benchmarks become somewhat slower, that's due to increased contention on other data structures, it should not affect programs that do any real work. On 2 x Intel E5620, 8 HT cores, 2.4GHz benchmark old ns/op new ns/op delta BenchmarkSelectContended 521.00 503.00 -3.45% BenchmarkSelectContended-2 661.00 320.00 -51.59% BenchmarkSelectContended-4 1139.00 629.00 -44.78% BenchmarkSelectContended-8 2870.00 878.00 -69.41% BenchmarkSelectContended-16 5276.00 818.00 -84.50% BenchmarkChanContended 112.00 103.00 -8.04% BenchmarkChanContended-2 631.00 174.00 -72.42% BenchmarkChanContended-4 682.00 272.00 -60.12% BenchmarkChanContended-8 1601.00 520.00 -67.52% BenchmarkChanContended-16 3100.00 372.00 -88.00% BenchmarkChanSync 253.00 239.00 -5.53% BenchmarkChanSync-2 5030.00 4648.00 -7.59% BenchmarkChanSync-4 4826.00 4694.00 -2.74% BenchmarkChanSync-8 4778.00 4713.00 -1.36% BenchmarkChanSync-16 5289.00 4710.00 -10.95% BenchmarkChanProdCons0 273.00 254.00 -6.96% BenchmarkChanProdCons0-2 599.00 400.00 -33.22% BenchmarkChanProdCons0-4 1168.00 659.00 -43.58% BenchmarkChanProdCons0-8 2831.00 1057.00 -62.66% BenchmarkChanProdCons0-16 4197.00 1037.00 -75.29% BenchmarkChanProdCons10 150.00 140.00 -6.67% BenchmarkChanProdCons10-2 607.00 268.00 -55.85% BenchmarkChanProdCons10-4 1137.00 404.00 -64.47% BenchmarkChanProdCons10-8 2115.00 828.00 -60.85% BenchmarkChanProdCons10-16 4283.00 855.00 -80.04% BenchmarkChanProdCons100 117.00 110.00 -5.98% BenchmarkChanProdCons100-2 558.00 218.00 -60.93% BenchmarkChanProdCons100-4 722.00 287.00 -60.25% BenchmarkChanProdCons100-8 1840.00 431.00 -76.58% BenchmarkChanProdCons100-16 3394.00 448.00 -86.80% BenchmarkChanProdConsWork0 2014.00 1996.00 -0.89% BenchmarkChanProdConsWork0-2 1207.00 1127.00 -6.63% BenchmarkChanProdConsWork0-4 1913.00 611.00 -68.06% BenchmarkChanProdConsWork0-8 3016.00 949.00 -68.53% BenchmarkChanProdConsWork0-16 4320.00 1154.00 -73.29% BenchmarkChanProdConsWork10 1906.00 1897.00 -0.47% BenchmarkChanProdConsWork10-2 1123.00 1033.00 -8.01% BenchmarkChanProdConsWork10-4 1076.00 571.00 -46.93% BenchmarkChanProdConsWork10-8 2748.00 1096.00 -60.12% BenchmarkChanProdConsWork10-16 4600.00 1105.00 -75.98% BenchmarkChanProdConsWork100 1884.00 1852.00 -1.70% BenchmarkChanProdConsWork100-2 1235.00 1146.00 -7.21% BenchmarkChanProdConsWork100-4 1217.00 619.00 -49.14% BenchmarkChanProdConsWork100-8 1534.00 509.00 -66.82% BenchmarkChanProdConsWork100-16 4126.00 918.00 -77.75% BenchmarkSyscall 34.40 33.30 -3.20% BenchmarkSyscall-2 160.00 121.00 -24.38% BenchmarkSyscall-4 131.00 136.00 +3.82% BenchmarkSyscall-8 139.00 131.00 -5.76% BenchmarkSyscall-16 161.00 168.00 +4.35% BenchmarkSyscallWork 950.00 950.00 +0.00% BenchmarkSyscallWork-2 481.00 480.00 -0.21% BenchmarkSyscallWork-4 268.00 270.00 +0.75% BenchmarkSyscallWork-8 156.00 169.00 +8.33% BenchmarkSyscallWork-16 188.00 184.00 -2.13% BenchmarkSemaSyntNonblock 36.40 35.60 -2.20% BenchmarkSemaSyntNonblock-2 81.40 45.10 -44.59% BenchmarkSemaSyntNonblock-4 126.00 108.00 -14.29% BenchmarkSemaSyntNonblock-8 112.00 112.00 +0.00% BenchmarkSemaSyntNonblock-16 110.00 112.00 +1.82% BenchmarkSemaSyntBlock 35.30 35.30 +0.00% BenchmarkSemaSyntBlock-2 118.00 124.00 +5.08% BenchmarkSemaSyntBlock-4 105.00 108.00 +2.86% BenchmarkSemaSyntBlock-8 101.00 111.00 +9.90% BenchmarkSemaSyntBlock-16 112.00 118.00 +5.36% BenchmarkSemaWorkNonblock 810.00 811.00 +0.12% BenchmarkSemaWorkNonblock-2 476.00 414.00 -13.03% BenchmarkSemaWorkNonblock-4 238.00 228.00 -4.20% BenchmarkSemaWorkNonblock-8 140.00 126.00 -10.00% BenchmarkSemaWorkNonblock-16 117.00 116.00 -0.85% BenchmarkSemaWorkBlock 810.00 811.00 +0.12% BenchmarkSemaWorkBlock-2 454.00 466.00 +2.64% BenchmarkSemaWorkBlock-4 243.00 241.00 -0.82% BenchmarkSemaWorkBlock-8 145.00 137.00 -5.52% BenchmarkSemaWorkBlock-16 132.00 123.00 -6.82% BenchmarkContendedSemaphore 123.00 102.00 -17.07% BenchmarkContendedSemaphore-2 34.80 34.90 +0.29% BenchmarkContendedSemaphore-4 34.70 34.80 +0.29% BenchmarkContendedSemaphore-8 34.70 34.70 +0.00% BenchmarkContendedSemaphore-16 34.80 34.70 -0.29% BenchmarkMutex 26.80 26.00 -2.99% BenchmarkMutex-2 108.00 45.20 -58.15% BenchmarkMutex-4 103.00 127.00 +23.30% BenchmarkMutex-8 109.00 147.00 +34.86% BenchmarkMutex-16 102.00 152.00 +49.02% BenchmarkMutexSlack 27.00 26.90 -0.37% BenchmarkMutexSlack-2 149.00 165.00 +10.74% BenchmarkMutexSlack-4 121.00 209.00 +72.73% BenchmarkMutexSlack-8 101.00 158.00 +56.44% BenchmarkMutexSlack-16 97.00 129.00 +32.99% BenchmarkMutexWork 792.00 794.00 +0.25% BenchmarkMutexWork-2 407.00 409.00 +0.49% BenchmarkMutexWork-4 220.00 209.00 -5.00% BenchmarkMutexWork-8 267.00 160.00 -40.07% BenchmarkMutexWork-16 315.00 300.00 -4.76% BenchmarkMutexWorkSlack 792.00 793.00 +0.13% BenchmarkMutexWorkSlack-2 406.00 404.00 -0.49% BenchmarkMutexWorkSlack-4 225.00 212.00 -5.78% BenchmarkMutexWorkSlack-8 268.00 136.00 -49.25% BenchmarkMutexWorkSlack-16 300.00 300.00 +0.00% BenchmarkRWMutexWrite100 27.10 27.00 -0.37% BenchmarkRWMutexWrite100-2 33.10 40.80 +23.26% BenchmarkRWMutexWrite100-4 113.00 88.10 -22.04% BenchmarkRWMutexWrite100-8 119.00 95.30 -19.92% BenchmarkRWMutexWrite100-16 148.00 109.00 -26.35% BenchmarkRWMutexWrite10 29.60 29.40 -0.68% BenchmarkRWMutexWrite10-2 111.00 61.40 -44.68% BenchmarkRWMutexWrite10-4 270.00 208.00 -22.96% BenchmarkRWMutexWrite10-8 204.00 185.00 -9.31% BenchmarkRWMutexWrite10-16 261.00 190.00 -27.20% BenchmarkRWMutexWorkWrite100 1040.00 1036.00 -0.38% BenchmarkRWMutexWorkWrite100-2 593.00 580.00 -2.19% BenchmarkRWMutexWorkWrite100-4 470.00 365.00 -22.34% BenchmarkRWMutexWorkWrite100-8 468.00 289.00 -38.25% BenchmarkRWMutexWorkWrite100-16 604.00 374.00 -38.08% BenchmarkRWMutexWorkWrite10 951.00 951.00 +0.00% BenchmarkRWMutexWorkWrite10-2 1001.00 928.00 -7.29% BenchmarkRWMutexWorkWrite10-4 1555.00 1006.00 -35.31% BenchmarkRWMutexWorkWrite10-8 2085.00 1171.00 -43.84% BenchmarkRWMutexWorkWrite10-16 2082.00 1614.00 -22.48% R=rsc, iant, msolo, fw, iant CC=golang-dev https://golang.org/cl/4711045
2011-07-29 10:44:06 -06:00
TEXT runtime·xchg(SB), 7, $0
MOVL 4(SP), BX
MOVL 8(SP), AX
XCHGL AX, 0(BX)
RET
TEXT runtime·procyield(SB),7,$0
MOVL 4(SP), AX
again:
PAUSE
SUBL $1, AX
JNZ again
RET
TEXT runtime·atomicstorep(SB), 7, $0
MOVL 4(SP), BX
MOVL 8(SP), AX
XCHGL AX, 0(BX)
RET
TEXT runtime·atomicstore(SB), 7, $0
MOVL 4(SP), BX
MOVL 8(SP), AX
XCHGL AX, 0(BX)
RET
// uint64 atomicload64(uint64 volatile* addr);
// so actually
// void atomicload64(uint64 *res, uint64 volatile *addr);
TEXT runtime·atomicload64(SB), 7, $0
MOVL 4(SP), BX
MOVL 8(SP), AX
// MOVQ (%EAX), %MM0
BYTE $0x0f; BYTE $0x6f; BYTE $0x00
// MOVQ %MM0, 0(%EBX)
BYTE $0x0f; BYTE $0x7f; BYTE $0x03
// EMMS
BYTE $0x0F; BYTE $0x77
RET
// void runtime·atomicstore64(uint64 volatile* addr, uint64 v);
TEXT runtime·atomicstore64(SB), 7, $0
MOVL 4(SP), AX
// MOVQ and EMMS were introduced on the Pentium MMX.
// MOVQ 0x8(%ESP), %MM0
BYTE $0x0f; BYTE $0x6f; BYTE $0x44; BYTE $0x24; BYTE $0x08
// MOVQ %MM0, (%EAX)
BYTE $0x0f; BYTE $0x7f; BYTE $0x00
// EMMS
BYTE $0x0F; BYTE $0x77
// This is essentially a no-op, but it provides required memory fencing.
// It can be replaced with MFENCE, but MFENCE was introduced only on the Pentium4 (SSE2).
MOVL $0, AX
LOCK
XADDL AX, (SP)
RET
TEXT runtime·prefetch(SB), 7, $0
MOVL 4(SP), AX
PREFETCHNTA (AX)
RET
// void jmpdefer(fn, sp);
// called from deferreturn.
// 1. pop the caller
// 2. sub 5 bytes from the callers return
// 3. jmp to the argument
TEXT runtime·jmpdefer(SB), 7, $0
MOVL 4(SP), AX // fn
MOVL 8(SP), BX // caller sp
LEAL -4(BX), SP // caller sp after CALL
SUBL $5, (SP) // return to CALL again
JMP AX // but first run the deferred function
runtime: scheduler, cgo reorganization * Change use of m->g0 stack (aka scheduler stack). * Provide runtime.mcall(f) to invoke f() on m->g0 stack. * Replace scheduler loop entry with runtime.mcall(schedule). Runtime.mcall eliminates the need for fake scheduler states that exist just to run a bit of code on the m->g0 stack (Grecovery, Gstackalloc). The elimination of the scheduler as a loop that stops and starts using gosave and gogo fixes a bad interaction with the way cgo uses the m->g0 stack. Cgo runs external (gcc-compiled) C functions on that stack, and then when calling back into Go, it sets m->g0->sched.sp below the added call frames, so that other uses of m->g0's stack will not interfere with those frames. Unfortunately, gogo (longjmp) back to the scheduler loop at this point would end up running scheduler with the lower sp, which no longer points at a valid stack frame for a call to scheduler. If scheduler then wrote any function call arguments or local variables to where it expected the stack frame to be, it would overwrite other data on the stack. I realized this possibility while debugging a problem with calling complex Go code in a Go -> C -> Go cgo callback. This wasn't the bug I was looking for, it turns out, but I believe it is a real bug nonetheless. Switching to runtime.mcall, which only adds new frames to the stack and never jumps into functions running in existing ones, fixes this bug. * Move cgo-related code out of proc.c into cgocall.c. * Add very large comment describing cgo call sequences. * Simpilify, regularize cgo function implementations and names. * Add test suite as misc/cgo/test. Now the Go -> C path calls cgocall, which calls asmcgocall, and the C -> Go path calls cgocallback, which calls cgocallbackg. The shuffling, which affects mainly the callback case, moves most of the callback implementation to cgocallback running on the m->curg stack (not the m->g0 scheduler stack) and only while accounted for with $GOMAXPROCS (between calls to exitsyscall and entersyscall). The previous callback code did not block in startcgocallback's approximation to exitsyscall, so if, say, the garbage collector were running, it would still barge in and start doing things like call malloc. Similarly endcgocallback's approximation of entersyscall did not call matchmg to kick off new OS threads when necessary, which caused the bug in issue 1560. Fixes #1560. R=iant CC=golang-dev https://golang.org/cl/4253054
2011-03-07 08:37:42 -07:00
// Dummy function to use in saved gobuf.PC,
// to match SP pointing at a return address.
// The gobuf.PC is unused by the contortions here
// but setting it to return will make the traceback code work.
TEXT return<>(SB),7,$0
RET
// asmcgocall(void(*fn)(void*), void *arg)
// Call fn(arg) on the scheduler stack,
// aligned appropriately for the gcc ABI.
// See cgocall.c for more details.
TEXT runtime·asmcgocall(SB),7,$0
MOVL fn+0(FP), AX
MOVL arg+4(FP), BX
MOVL SP, DX
// Figure out if we need to switch to m->g0 stack.
// We get called to create new OS threads too, and those
// come in on the m->g0 stack already.
get_tls(CX)
MOVL m(CX), BP
MOVL m_g0(BP), SI
MOVL g(CX), DI
CMPL SI, DI
JEQ 6(PC)
MOVL SP, (g_sched+gobuf_sp)(DI)
MOVL $return<>(SB), (g_sched+gobuf_pc)(DI)
MOVL DI, (g_sched+gobuf_g)(DI)
MOVL SI, g(CX)
MOVL (g_sched+gobuf_sp)(SI), SP
// Now on a scheduling stack (a pthread-created stack).
SUBL $32, SP
ANDL $~15, SP // alignment, perhaps unnecessary
MOVL DI, 8(SP) // save g
MOVL DX, 4(SP) // save SP
MOVL BX, 0(SP) // first argument in x86-32 ABI
CALL AX
// Restore registers, g, stack pointer.
get_tls(CX)
MOVL 8(SP), DI
MOVL DI, g(CX)
MOVL 4(SP), SP
RET
// cgocallback(void (*fn)(void*), void *frame, uintptr framesize)
// See cgocall.c for more details.
TEXT runtime·cgocallback(SB),7,$12
MOVL fn+0(FP), AX
MOVL frame+4(FP), BX
MOVL framesize+8(FP), DX
// Save current m->g0->sched.sp on stack and then set it to SP.
get_tls(CX)
MOVL m(CX), BP
// If m is nil, it is almost certainly because we have been called
// on a thread that Go did not create. We're going to crash as
// soon as we try to use m; instead, try to print a nice error and exit.
CMPL BP, $0
JNE 2(PC)
CALL runtime·badcallback(SB)
runtime: scheduler, cgo reorganization * Change use of m->g0 stack (aka scheduler stack). * Provide runtime.mcall(f) to invoke f() on m->g0 stack. * Replace scheduler loop entry with runtime.mcall(schedule). Runtime.mcall eliminates the need for fake scheduler states that exist just to run a bit of code on the m->g0 stack (Grecovery, Gstackalloc). The elimination of the scheduler as a loop that stops and starts using gosave and gogo fixes a bad interaction with the way cgo uses the m->g0 stack. Cgo runs external (gcc-compiled) C functions on that stack, and then when calling back into Go, it sets m->g0->sched.sp below the added call frames, so that other uses of m->g0's stack will not interfere with those frames. Unfortunately, gogo (longjmp) back to the scheduler loop at this point would end up running scheduler with the lower sp, which no longer points at a valid stack frame for a call to scheduler. If scheduler then wrote any function call arguments or local variables to where it expected the stack frame to be, it would overwrite other data on the stack. I realized this possibility while debugging a problem with calling complex Go code in a Go -> C -> Go cgo callback. This wasn't the bug I was looking for, it turns out, but I believe it is a real bug nonetheless. Switching to runtime.mcall, which only adds new frames to the stack and never jumps into functions running in existing ones, fixes this bug. * Move cgo-related code out of proc.c into cgocall.c. * Add very large comment describing cgo call sequences. * Simpilify, regularize cgo function implementations and names. * Add test suite as misc/cgo/test. Now the Go -> C path calls cgocall, which calls asmcgocall, and the C -> Go path calls cgocallback, which calls cgocallbackg. The shuffling, which affects mainly the callback case, moves most of the callback implementation to cgocallback running on the m->curg stack (not the m->g0 scheduler stack) and only while accounted for with $GOMAXPROCS (between calls to exitsyscall and entersyscall). The previous callback code did not block in startcgocallback's approximation to exitsyscall, so if, say, the garbage collector were running, it would still barge in and start doing things like call malloc. Similarly endcgocallback's approximation of entersyscall did not call matchmg to kick off new OS threads when necessary, which caused the bug in issue 1560. Fixes #1560. R=iant CC=golang-dev https://golang.org/cl/4253054
2011-03-07 08:37:42 -07:00
MOVL m_g0(BP), SI
PUSHL (g_sched+gobuf_sp)(SI)
MOVL SP, (g_sched+gobuf_sp)(SI)
// Switch to m->curg stack and call runtime.cgocallbackg
runtime: scheduler, cgo reorganization * Change use of m->g0 stack (aka scheduler stack). * Provide runtime.mcall(f) to invoke f() on m->g0 stack. * Replace scheduler loop entry with runtime.mcall(schedule). Runtime.mcall eliminates the need for fake scheduler states that exist just to run a bit of code on the m->g0 stack (Grecovery, Gstackalloc). The elimination of the scheduler as a loop that stops and starts using gosave and gogo fixes a bad interaction with the way cgo uses the m->g0 stack. Cgo runs external (gcc-compiled) C functions on that stack, and then when calling back into Go, it sets m->g0->sched.sp below the added call frames, so that other uses of m->g0's stack will not interfere with those frames. Unfortunately, gogo (longjmp) back to the scheduler loop at this point would end up running scheduler with the lower sp, which no longer points at a valid stack frame for a call to scheduler. If scheduler then wrote any function call arguments or local variables to where it expected the stack frame to be, it would overwrite other data on the stack. I realized this possibility while debugging a problem with calling complex Go code in a Go -> C -> Go cgo callback. This wasn't the bug I was looking for, it turns out, but I believe it is a real bug nonetheless. Switching to runtime.mcall, which only adds new frames to the stack and never jumps into functions running in existing ones, fixes this bug. * Move cgo-related code out of proc.c into cgocall.c. * Add very large comment describing cgo call sequences. * Simpilify, regularize cgo function implementations and names. * Add test suite as misc/cgo/test. Now the Go -> C path calls cgocall, which calls asmcgocall, and the C -> Go path calls cgocallback, which calls cgocallbackg. The shuffling, which affects mainly the callback case, moves most of the callback implementation to cgocallback running on the m->curg stack (not the m->g0 scheduler stack) and only while accounted for with $GOMAXPROCS (between calls to exitsyscall and entersyscall). The previous callback code did not block in startcgocallback's approximation to exitsyscall, so if, say, the garbage collector were running, it would still barge in and start doing things like call malloc. Similarly endcgocallback's approximation of entersyscall did not call matchmg to kick off new OS threads when necessary, which caused the bug in issue 1560. Fixes #1560. R=iant CC=golang-dev https://golang.org/cl/4253054
2011-03-07 08:37:42 -07:00
// with the three arguments. Because we are taking over
// the execution of m->curg but *not* resuming what had
// been running, we need to save that information (m->curg->gobuf)
// so that we can restore it when we're done.
// We can restore m->curg->gobuf.sp easily, because calling
// runtime.cgocallbackg leaves SP unchanged upon return.
runtime: scheduler, cgo reorganization * Change use of m->g0 stack (aka scheduler stack). * Provide runtime.mcall(f) to invoke f() on m->g0 stack. * Replace scheduler loop entry with runtime.mcall(schedule). Runtime.mcall eliminates the need for fake scheduler states that exist just to run a bit of code on the m->g0 stack (Grecovery, Gstackalloc). The elimination of the scheduler as a loop that stops and starts using gosave and gogo fixes a bad interaction with the way cgo uses the m->g0 stack. Cgo runs external (gcc-compiled) C functions on that stack, and then when calling back into Go, it sets m->g0->sched.sp below the added call frames, so that other uses of m->g0's stack will not interfere with those frames. Unfortunately, gogo (longjmp) back to the scheduler loop at this point would end up running scheduler with the lower sp, which no longer points at a valid stack frame for a call to scheduler. If scheduler then wrote any function call arguments or local variables to where it expected the stack frame to be, it would overwrite other data on the stack. I realized this possibility while debugging a problem with calling complex Go code in a Go -> C -> Go cgo callback. This wasn't the bug I was looking for, it turns out, but I believe it is a real bug nonetheless. Switching to runtime.mcall, which only adds new frames to the stack and never jumps into functions running in existing ones, fixes this bug. * Move cgo-related code out of proc.c into cgocall.c. * Add very large comment describing cgo call sequences. * Simpilify, regularize cgo function implementations and names. * Add test suite as misc/cgo/test. Now the Go -> C path calls cgocall, which calls asmcgocall, and the C -> Go path calls cgocallback, which calls cgocallbackg. The shuffling, which affects mainly the callback case, moves most of the callback implementation to cgocallback running on the m->curg stack (not the m->g0 scheduler stack) and only while accounted for with $GOMAXPROCS (between calls to exitsyscall and entersyscall). The previous callback code did not block in startcgocallback's approximation to exitsyscall, so if, say, the garbage collector were running, it would still barge in and start doing things like call malloc. Similarly endcgocallback's approximation of entersyscall did not call matchmg to kick off new OS threads when necessary, which caused the bug in issue 1560. Fixes #1560. R=iant CC=golang-dev https://golang.org/cl/4253054
2011-03-07 08:37:42 -07:00
// To save m->curg->gobuf.pc, we push it onto the stack.
// This has the added benefit that it looks to the traceback
// routine like cgocallbackg is going to return to that
// PC (because we defined cgocallbackg to have
runtime: scheduler, cgo reorganization * Change use of m->g0 stack (aka scheduler stack). * Provide runtime.mcall(f) to invoke f() on m->g0 stack. * Replace scheduler loop entry with runtime.mcall(schedule). Runtime.mcall eliminates the need for fake scheduler states that exist just to run a bit of code on the m->g0 stack (Grecovery, Gstackalloc). The elimination of the scheduler as a loop that stops and starts using gosave and gogo fixes a bad interaction with the way cgo uses the m->g0 stack. Cgo runs external (gcc-compiled) C functions on that stack, and then when calling back into Go, it sets m->g0->sched.sp below the added call frames, so that other uses of m->g0's stack will not interfere with those frames. Unfortunately, gogo (longjmp) back to the scheduler loop at this point would end up running scheduler with the lower sp, which no longer points at a valid stack frame for a call to scheduler. If scheduler then wrote any function call arguments or local variables to where it expected the stack frame to be, it would overwrite other data on the stack. I realized this possibility while debugging a problem with calling complex Go code in a Go -> C -> Go cgo callback. This wasn't the bug I was looking for, it turns out, but I believe it is a real bug nonetheless. Switching to runtime.mcall, which only adds new frames to the stack and never jumps into functions running in existing ones, fixes this bug. * Move cgo-related code out of proc.c into cgocall.c. * Add very large comment describing cgo call sequences. * Simpilify, regularize cgo function implementations and names. * Add test suite as misc/cgo/test. Now the Go -> C path calls cgocall, which calls asmcgocall, and the C -> Go path calls cgocallback, which calls cgocallbackg. The shuffling, which affects mainly the callback case, moves most of the callback implementation to cgocallback running on the m->curg stack (not the m->g0 scheduler stack) and only while accounted for with $GOMAXPROCS (between calls to exitsyscall and entersyscall). The previous callback code did not block in startcgocallback's approximation to exitsyscall, so if, say, the garbage collector were running, it would still barge in and start doing things like call malloc. Similarly endcgocallback's approximation of entersyscall did not call matchmg to kick off new OS threads when necessary, which caused the bug in issue 1560. Fixes #1560. R=iant CC=golang-dev https://golang.org/cl/4253054
2011-03-07 08:37:42 -07:00
// a frame size of 12, the same amount that we use below),
// so that the traceback will seamlessly trace back into
// the earlier calls.
MOVL m_curg(BP), SI
MOVL SI, g(CX)
MOVL (g_sched+gobuf_sp)(SI), DI // prepare stack as DI
// Push gobuf.pc
MOVL (g_sched+gobuf_pc)(SI), BP
SUBL $4, DI
MOVL BP, 0(DI)
// Push arguments to cgocallbackg.
// Frame size here must match the frame size above
// to trick traceback routines into doing the right thing.
SUBL $12, DI
MOVL AX, 0(DI)
MOVL BX, 4(DI)
MOVL DX, 8(DI)
// Switch stack and make the call.
MOVL DI, SP
CALL runtime·cgocallbackg(SB)
// Restore g->gobuf (== m->curg->gobuf) from saved values.
get_tls(CX)
MOVL g(CX), SI
MOVL 12(SP), BP
MOVL BP, (g_sched+gobuf_pc)(SI)
LEAL (12+4)(SP), DI
MOVL DI, (g_sched+gobuf_sp)(SI)
// Switch back to m->g0's stack and restore m->g0->sched.sp.
// (Unlike m->curg, the g0 goroutine never uses sched.pc,
// so we do not have to restore it.)
MOVL m(CX), BP
MOVL m_g0(BP), SI
MOVL SI, g(CX)
MOVL (g_sched+gobuf_sp)(SI), SP
POPL (g_sched+gobuf_sp)(SI)
// Done!
RET
// check that SP is in range [g->stackbase, g->stackguard)
TEXT runtime·stackcheck(SB), 7, $0
get_tls(CX)
MOVL g(CX), AX
CMPL g_stackbase(AX), SP
JHI 2(PC)
INT $3
CMPL SP, g_stackguard(AX)
JHI 2(PC)
INT $3
RET
TEXT runtime·memclr(SB),7,$0
MOVL 4(SP), DI // arg 1 addr
MOVL 8(SP), CX // arg 2 count
MOVL CX, BX
ANDL $3, BX
SHRL $2, CX
MOVL $0, AX
CLD
REP
STOSL
MOVL BX, CX
REP
STOSB
RET
TEXT runtime·getcallerpc(SB),7,$0
MOVL x+0(FP),AX // addr of first arg
MOVL -4(AX),AX // get calling pc
RET
TEXT runtime·setcallerpc(SB),7,$0
MOVL x+0(FP),AX // addr of first arg
MOVL x+4(FP), BX
MOVL BX, -4(AX) // set calling pc
RET
TEXT runtime·getcallersp(SB), 7, $0
MOVL sp+0(FP), AX
RET
// int64 runtime·cputicks(void), so really
// void runtime·cputicks(int64 *ticks)
TEXT runtime·cputicks(SB),7,$0
RDTSC
MOVL ret+0(FP), DI
MOVL AX, 0(DI)
MOVL DX, 4(DI)
RET
TEXT runtime·ldt0setup(SB),7,$16
// set up ldt 7 to point at tls0
// ldt 1 would be fine on Linux, but on OS X, 7 is as low as we can go.
// the entry number is just a hint. setldt will set up GS with what it used.
MOVL $7, 0(SP)
LEAL runtime·tls0(SB), AX
MOVL AX, 4(SP)
MOVL $32, 8(SP) // sizeof(tls array)
CALL runtime·setldt(SB)
RET
TEXT runtime·emptyfunc(SB),0,$0
RET
TEXT runtime·abort(SB),7,$0
INT $0x3
TEXT runtime·stackguard(SB),7,$0
MOVL SP, DX
MOVL DX, sp+0(FP)
get_tls(CX)
MOVL g(CX), BX
MOVL g_stackguard(BX), DX
MOVL DX, guard+4(FP)
RET
GLOBL runtime·tls0(SB), $32