ref: 847cd0b5262ebdf9deb3c294bedffc162b4a80d2
parent: e4df0925260ce4f14ba6db4025d1ed4913b0f9ec
author: halfwit <michaelmisch1985@gmail.com>
date: Thu Sep 5 13:58:00 PDT 2024
Some movement of some things
--- a/include/lib.h
+++ b/include/lib.h
@@ -315,7 +315,7 @@
extern char* utfecpy(char*, char*, char*);
extern int tas(int*);
extern int trampoline(void*);
-extern void run(uintptr_t, Tos *, int, char **);
+extern void start(uintptr_t, Tos *, int, char **);
extern long sysintercept(void*, void*, void*, void*, void*, void*, void*);
extern int patch(void*, int);
extern void quotefmtinstall(void);
--- a/kern/posix.c
+++ b/kern/posix.c
@@ -147,7 +147,7 @@
argc = nelem((char**)p->arg);
if(pthread_setspecific(prdakey, p))
panic("cannot setspecific");- run(up->bin->text, &tos, argc, p->arg);
+ start(up->bin->text, &tos, argc, p->arg);
pexit("", 0);return 0;
}
@@ -193,7 +193,7 @@
void
ospatchtext(void)
{- int n;
+ int n, flag;
void *text, *final;
/* Set up trampoline. Mach dependent */
@@ -202,9 +202,14 @@
if(n != TRAMPSIZE)
error("building trampoline failed");- final = mmap(0, (up->bin->ts)+n, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
+ flag = MAP_PRIVATE|MAP_ANONYMOUS;
+
+ #ifdef __APPLE__
+ flag |= MAP_JIT;
+ #endif
+ final = mmap(0, up->bin->ts+n, PROT_READ|PROT_WRITE, flag, -1, 0);
memmove(final, text, n);
- memmove(final+n+1, (void*)up->bin->text, up->bin->ts);
+ memmove(final+n, (void*)up->bin->text, up->bin->ts);
if(!final)
error("unable to set up text segment with trampoline");@@ -211,8 +216,9 @@
/* Patch. Mach dependent */
if(patch(final+n+1, up->bin->ts) < 0)
error("unable to patch syscalls");- mprotect(final, n, PROT_EXEC);
- up->bin->text = (uintptr)final+n+1;
+
+ mprotect(final, n + up->bin->ts, PROT_READ|PROT_EXEC);
+ up->bin->text = (uintptr)final+n;
poperror();
}
@@ -226,7 +232,7 @@
n = devtab[tc->type]->read(tc, data, up->bin->ds, up->bin->ts);
if(!data || n == 0)
error("unable to set up data segment");- //up->bin->data = (uintptr)data;
+ up->bin->data = (uintptr)data;
}
void
@@ -237,7 +243,7 @@
bss = mmap((void*)up->bin->bss, up->bin->bssz, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
if(!bss)
error("unable to set up bss segment");- //up->bin->bss = (uintptr)bss;
+ up->bin->bss = (uintptr)bss;
}
void
--- a/kern/sysproc.c
+++ b/kern/sysproc.c
@@ -261,8 +261,8 @@
error(Ebadexec);
up->bin->text += beswal(u.ehdr.ex.text);
// BUG: This errors, errors seem broken currently
-// if(text <= entry || text >= (uintptr)(USTKTOP-USTKSIZE))
-// error(Ebadexec);
+ if(up->bin->text <= up->bin->entry || up->bin->text >= (uintptr)(USTKTOP-USTKSIZE))
+ error(Ebadexec);
switch(magic){case S_MAGIC: /* 2MB segment alignment for amd64 */
align = 0x1fffff;
@@ -299,16 +299,14 @@
// Check the actual sysproc, there's some missing bits here still
up->arg = progarg;
}
-
up->bin->ts = ((up->bin->text+align) & ~align) / BY2WD;
- up->bin->text -= UTZERO;
up->bin->data = beswal(u.ehdr.ex.data);
up->bin->bss = beswal(u.ehdr.ex.bss);
+ up->bin->text -= UTZERO;
align = BY2PG-1;
up->bin->ds = ((up->bin->ts + up->bin->data + align) & ~align) / BY2WD;
bssend = up->bin->ts + up->bin->data + up->bin->bss;
up->bin->bssz = ((bssend + align) & ~align) / BY2WD;
-
if(up->bin->ts >= (ulong)(USTKTOP-USTKSIZE) || up->bin->ds >= (ulong)(USTKTOP-USTKSIZE) || up->bin->bss >= (ulong)(USTKTOP-USTKSIZE))
error(Ebadexec);
@@ -315,21 +313,25 @@
/* Set up text, data, and bss. Patch syscalls. OS dependent. */
osbuildtext(tc);
if(waserror()){+ print("%s\n", up->errstr);osclrmem();
nexterror();
}
ospatchtext();
if(waserror()){+ print("%s\n", up->errstr);osclrmem();
nexterror();
}
osbuilddata(tc);
if(waserror()){+ print("%s\n", up->errstr);osclrmem();
nexterror();
}
osbuildbss(tc);
if(waserror()){+ print("%s\n", up->errstr);osclrmem();
nexterror();
}
--- a/librc/exec.c
+++ b/librc/exec.c
@@ -16,7 +16,7 @@
* Start executing the given code at the given pc with the given redirection
*/
void
-start(code *c, int pc, var *local, redir *redir)
+startexec(code *c, int pc, var *local, redir *redir)
{thread *p = new(thread);
p->code = codecopy(c);
@@ -36,7 +36,7 @@
void
startfunc(var *func, word *starval, var *local, redir *redir)
{- start(func->fn, func->pc, local, redir);
+ startexec(func->fn, func->pc, local, redir);
runq->local = newvar("*", runq->local);runq->local->val = starval;
runq->local->changed = 1;
@@ -272,7 +272,7 @@
bootstrap[17].f = Xsimple;
bootstrap[18].f = Xexit;
bootstrap[19].f = 0;
- start(bootstrap, 2, (var*)0, (redir*)0);
+ startexec(bootstrap, 2, (var*)0, (redir*)0);
/* prime bootstrap argv */
pushlist();
for(i = argc-1;i!=0;--i) pushword(argv[i]);
@@ -1038,7 +1038,7 @@
p->lex = 0;
} else
--p->pc; /* re-execute Xrdcmds after codebuf runs */
- start(codebuf, 2, p->local, p->redir);
+ startexec(codebuf, 2, p->local, p->redir);
}
lex = 0;
freenodes();
--- a/librc/fns.h
+++ b/librc/fns.h
@@ -60,7 +60,7 @@
void heredoc(tree*);
void setstatus(char*);
void skipnl(void);
-void start(code*, int, var*, redir*);
+void startexec(code*, int, var*, redir*);
int truestatus(void);
void usage(char*);
int wordchr(int);
--- a/librc/havefork.c
+++ b/librc/havefork.c
@@ -54,7 +54,7 @@
break;
case 0:
clearwaitpids();
- start(runq->code, runq->pc+1, runq->local, runq->redir);
+ startexec(runq->code, runq->pc+1, runq->local, runq->redir);
runq->ret = 0;
break;
default:
@@ -86,7 +86,7 @@
case 0:
clearwaitpids();
Close(pfd[PRD]);
- start(p->code, pc+2, runq->local, runq->redir);
+ startexec(p->code, pc+2, runq->local, runq->redir);
runq->ret = 0;
pushredir(ROPEN, pfd[PWR], lfd);
break;
@@ -93,7 +93,7 @@
default:
addwaitpid(pid);
Close(pfd[PWR]);
- start(p->code, p->code[pc].i, runq->local, runq->redir);
+ startexec(p->code, p->code[pc].i, runq->local, runq->redir);
pushredir(ROPEN, pfd[PRD], rfd);
p->pc = p->code[pc+1].i;
p->pid = pid;
@@ -126,7 +126,7 @@
case 0:
clearwaitpids();
Close(pfd[PRD]);
- start(runq->code, runq->pc+1, runq->local, runq->redir);
+ startexec(runq->code, runq->pc+1, runq->local, runq->redir);
pushredir(ROPEN, pfd[PWR], 1);
return;
default:
@@ -181,7 +181,7 @@
case 0:
clearwaitpids();
Close(mainfd);
- start(p->code, pc+2, runq->local, runq->redir);
+ startexec(p->code, pc+2, runq->local, runq->redir);
pushredir(ROPEN, sidefd, p->code[pc].i==READ?1:0);
runq->ret = 0;
break;
@@ -209,7 +209,7 @@
break;
case 0:
clearwaitpids();
- start(runq->code, runq->pc+1, runq->local, runq->redir);
+ startexec(runq->code, runq->pc+1, runq->local, runq->redir);
runq->ret = 0;
break;
default:
--- a/librc/simple.c
+++ b/librc/simple.c
@@ -324,7 +324,7 @@
if(exitnext()) turfstack(local);
- start(rdcmds, 2, local, redir);
+ startexec(rdcmds, 2, local, redir);
runq->lex = newlexer(input, file);
}
--- a/posix-amd64/Makefile
+++ b/posix-amd64/Makefile
@@ -7,7 +7,7 @@
OFILES=\
getcallerpc.$O\
tas.$O\
- run.$O\
+ start.$O\
trampoline.$O\
default: $(LIB)
--- a/posix-amd64/run.c
+++ /dev/null
@@ -1,38 +1,0 @@
-#include "u.h"
-#include "libc.h"
-
-void run(uintptr_t entry, Tos *_tos, int argc, char *argv[]) {- // entry point
- register uintptr_t rdi asm("rdi") = entry;- register Tos *rsi asm("rsi") = _tos;- register int edx asm("edx") = argc;- register char **rcx asm("rcx") = argv;-
- __asm__ __volatile__ (
- // Load values into registers
- "mov rdi, %0\n\t"
- "mov rsi, %1\n\t"
- "mov edx, %2\n\t"
- "mov rcx, %3\n\t"
-
- // push argv onto stack
- "mov r8, rdx\n\t"
- "add r8, r8, 1\n\t"
- "shl r8, r8, 3\n\t"
- "sub rsp, r8\n\t"
- "mov r9, rsp\n\t"
- "rep movsb\n\t"
-
- // push argc onto stack
- "mov [rsp], edx\n\t"
- "sub rsp, 8\n\t"
-
- // jump to entry point
- "jmp rdi\n\t"
- "nop\n\t"
-
- :
- : "r" (rdi), "r" (rsi), "r" (edx), "r" (rcx)
- : "r8", "r9"
- );
-}
\ No newline at end of file
--- /dev/null
+++ b/posix-amd64/start.s
@@ -1,0 +1,22 @@
+.text
+.globl start
+start:
+ mov %rdi, %rbp /* entry */
+ mov %rsi, %rax /* _tos */
+ mov %rdx, %rbx /* argc */
+ mov %rcx, %rsi /* argv */
+
+ /* push argv onto stack */
+ mov %rbx, %rcx
+ add $1, %rcx
+ sal $3, %rcx
+ sub %rcx, %rsp
+ mov %rsp, %rdi
+ rep movsb
+
+ /* push argc onto stack */
+ push %rbx
+
+ jmp *%rbp
+
+.section .note.GNU-stack,"",@progbits
\ No newline at end of file
--- a/posix-arm64/Makefile
+++ b/posix-arm64/Makefile
@@ -7,7 +7,7 @@
OFILES=\
getcallerpc.$O\
tas.$O\
- run.$O\
+ start.$O\
trampoline.$O\
patch.$O\
--- a/posix-arm64/patch.c
+++ b/posix-arm64/patch.c
@@ -11,10 +11,11 @@
// MOV X0, #immediate (could be D2800000 | syscall_number)
// BL X0 (could be D4000000 | offset)
// 0xD63F0000 is our BLR X0
+ // This sets up our jump to the trampoline code
if ((*(ulong*)&text[i] & 0xFFFF0000) == 0xD2800000 && (*(ulong*)&text[i+BY2SE] & 0xFFFF0000) == 0xD4000000) {- long *ptr = (ulong*)&text[i];
- ptr[1] = 0xD6;
- ptr[0] = 0x3F;
+ ulong *ptr = (ulong*)&text[i];
+ ptr[1] = 0x3F;
+ ptr[0] = 0xD6;
ret++;
}
}
--- a/posix-arm64/run.s
+++ /dev/null
@@ -1,24 +1,0 @@
-/* SPDX-License-Identifier: Unlicense */
-.text
-.global _run
-_run:
- mov x29, x0 // entry (x0 is the first argument, equivalent to rdi)
- mov x9, x1 // _tos (x1 is the second argument, equivalent to rsi)
- mov x19, x2 // argc (x2 is the third argument, equivalent to rdx)
- mov x20, x3 // argv (x3 is the fourth argument, equivalent to rcx)
-
- // Push argv onto stack
- mov x10, x19 // x10 = argc
- add x10, x10, #1 // x10 = argc + 1
- lsl x10, x10, #3 // x10 = (argc + 1) * 8 (shifting left by 3 is multiplying by 8)
- sub sp, sp, x10 // Allocate space on stack
- mov x0, sp // x0 = new stack pointer (destination for memcpy)
- mov x1, x20 // x1 = argv (source for memcpy)
- mov x2, x10 // x2 = number of bytes to copy
- bl _memcpy // Call memcpy to copy argv to stack
-
- // Push argc onto stack
- str x19, [sp, #-16]! // Push argc onto stack and update stack pointer
-
- // Jump to entry point
- br x29
\ No newline at end of file
--- /dev/null
+++ b/posix-arm64/start.s
@@ -1,0 +1,48 @@
+.text
+.global _start
+.align 4
+_start:
+ // Function prologue
+ stp x29, x30, [sp, #-16]! // Save frame pointer and link register
+ mov x29, sp // Set up frame pointer
+
+ // Save callee-saved registers
+ stp x19, x20, [sp, #-16]!
+
+ // Your original code starts here
+ mov x19, x0 // entry (x0 is the first argument)
+ mov x9, x1 // _tos (x1 is the second argument)
+ mov x20, x2 // argc (x2 is the third argument)
+ mov x21, x3 // argv (x3 is the fourth argument)
+
+ // Push argv onto stack
+ mov x10, x20 // x10 = argc
+ add x10, x10, #1 // x10 = argc + 1
+ lsl x10, x10, #3 // x10 = (argc + 1) * 8
+ sub sp, sp, x10 // Allocate space on stack
+
+ // Ensure 16-byte stack alignment
+ and x11, x10, #15 // Get the misalignment amount
+ sub sp, sp, x11 // Adjust stack to ensure alignment
+
+ mov x0, sp // x0 = new stack pointer (destination for memcpy)
+ mov x1, x21 // x1 = argv (source for memcpy)
+ mov x2, x10 // x2 = number of bytes to copy
+ bl _memcpy
+
+ // Push argc onto stack (after the argv data)
+ str x20, [sp, x10]
+
+ // Call the entry point
+ blr x19
+
+ // Restore stack pointer
+ mov sp, x29
+
+ // Restore callee-saved registers
+ ldp x19, x20, [sp], #16
+
+ // Function epilogue
+ ldp x29, x30, [sp], #16
+
+ ret
\ No newline at end of file