SchOS/kernel/sys/syscall.c

#include "core/types.h"
#include "core/defs.h"
#include "core/param.h"
#include "arch/memlayout.h"
#include "arch/mmu.h"
#include "arch/traps.h"
#include "core/proc.h"
#include "arch/x86.h"
#include "core/syscall.h"
#include "core/uaccess.h"
#include "uapi/errno.h"
#include "uapi/posix.h"

// -----------------------------------------------------------------------
// Argument helpers — read syscall args from the saved register frame.
// x86-64 SysV ABI: rdi, rsi, rdx, r10, r8, r9
// -----------------------------------------------------------------------

// Note: address validation goes through uaccess.h (page-table presence +
// user bit on every page of the range). There is deliberately no
// `addr < myproc()->sz` check: that would assume every valid user pointer
// lives in the heap — false for stack addresses and mmap()'d buffers
// (both live above myproc()->sz), which made syscalls fail whenever a
// Linux-ABI program passed a string/pointer from either region.
//
// Validated pointers are returned as raw user VAs — the process page
// table is live in kernel mode, so they dereference correctly across
// page boundaries (unlike a uva2ka() alias, which is only good within
// the one page it translated).

int
fetchint(addr_t addr, int *ip)
{
    return copy_from_user(ip, addr, sizeof(*ip)) ? -1 : 0;
}

int
fetchaddr(addr_t addr, addr_t *ip)
{
    return copy_from_user(ip, addr, sizeof(*ip)) ? -1 : 0;
}

int
fetchstr(addr_t addr, char **pp)
{
    long len = strlen_user(addr);
    if (len < 0)
        return -1;
    *pp = (char*)addr;
    return len;
}

// /proc/strace. Which argument of a Linux syscall is a path, by number, so
// the trace can show the name rather than an address: that is most of what
// one wants from a trace of a program that will not start.
static int
path_arg(uint64 nr)
{
    switch (nr) {
    case 2: case 4: case 6: case 21: case 59: case 76: case 80: case 82:
    case 83: case 84: case 85: case 86: case 87: case 88: case 89: case 90:
    case 92: case 137: case 161:
        return 0;
    case 257: case 258: case 260: case 262: case 263: case 264: case 267:
    case 268: case 269: case 280: case 316: case 332: case 437: case 439:
        return 1;
    }
    return -1;
}

static void strace_call(struct trapframe *tf);

static addr_t
fetcharg(int n)
{
    switch (n) {
    case 0: return myproc()->tf->rdi;
    case 1: return myproc()->tf->rsi;
    case 2: return myproc()->tf->rdx;
    case 3: return myproc()->tf->r10;
    case 4: return myproc()->tf->r8;
    case 5: return myproc()->tf->r9;
    }
    panic("fetcharg: bad n");
    return 0;
}

int argint(int n, int *ip)    { *ip = (int)fetcharg(n); return 0; }
int argaddr(int n, addr_t *ip) { *ip = fetcharg(n);      return 0; }

int
argptr(int n, char **pp, int size)
{
    addr_t i;
    if (argaddr(n, &i) < 0)
        return -1;
    if (size < 0)
        return -1;
    char *p = user_ptr(i, (uint64)size);
    if (!p)
        return -1;
    *pp = p;
    return 0;
}

int
argstr(int n, char **pp)
{
    addr_t addr;
    if (argaddr(n, &addr) < 0)
        return -1;
    return fetchstr(addr, pp);
}

// -----------------------------------------------------------------------
// Dispatch table — auto-generated from table.inc
// -----------------------------------------------------------------------

#define SYSCALL(nr, lnr, name) extern addr_t sys_##name(void);
#define SYSCALL_ALIAS(lnr, name)
#include "table.inc"
#undef SYSCALL
#undef SYSCALL_ALIAS

static addr_t (*syscalls[NSYSCALLS])(void) = {
#define SYSCALL(nr, lnr, name) [SYS_##name] = sys_##name,
#define SYSCALL_ALIAS(lnr, name)
#include "table.inc"
#undef SYSCALL
#undef SYSCALL_ALIAS
};

// -----------------------------------------------------------------------
// Main entry point called from trap.c
// -----------------------------------------------------------------------

void
syscall(struct trapframe *tf)
{
    if (myproc()->strace) {
        strace_call(tf);
        return;
    }
    kstack_check_canary("syscall entry");
    myproc()->tf = tf;
    uint64 orig = myproc()->tf->rax;
    uint64 num  = orig;

    if (myproc()->syscall_map) {
        num = (num < (uint64)myproc()->map_size) ? myproc()->syscall_map[num] : 0;
    }

    if (num > 0 && num < NSYSCALLS && syscalls[num]) {
        tf->rax = syscalls[num]();
    } else {
        tf->rax = -ENOSYS;
    }

    if (myproc()->killed)
        exit(W_EXITCODE(0, myproc()->killsig ? myproc()->killsig : SIGKILL));

    deliver_signals(tf);

    kstack_check_canary("syscall exit");
}

// syscall() for a traced process: the same dispatch, with a line in the log
// for each call -- the Linux number, four arguments, a path where there is
// one, and what came back. The path is copied before the call runs, since
// execve replaces the memory it is in.
static void
strace_call(struct trapframe *tf)
{
    struct proc *p = myproc();
    uint64 orig = tf->rax, num = orig;
    uint64 a0 = tf->rdi, a1 = tf->rsi, a2 = tf->rdx, a3 = tf->r10;
    char path[64];
    int pa = p->syscall_map ? path_arg(orig) : -1;

    path[0] = 0;
    if (pa >= 0) {
        addr_t u = pa == 0 ? a0 : a1;
        int i;
        for (i = 0; i < (int)sizeof(path) - 1; i++) {
            char c;
            if (copy_from_user(&c, u + i, 1) < 0 || c == 0)
                break;
            path[i] = c;
        }
        path[i] = 0;
    }

    // The result line is written when the call returns, so a call that never
    // does would leave no trace at all. Everything that can wait gets a line
    // on the way in as well.
    switch (orig) {
    case 0: case 2: case 7: case 23: case 35: case 43: case 45: case 47:
    case 59: case 61: case 202: case 232: case 257: case 270: case 271:
    case 281: case 288:
        if (p->syscall_map)
            kinfo("strace %d %s: %d(%lx, %lx, %lx, %lx)%s%s%s ...\n",
                 p->pid, p->name, (int)orig, a0, a1, a2, a3,
                 pa >= 0 ? " \"" : "", path, pa >= 0 ? "\"" : "");
    }

    kstack_check_canary("syscall entry");
    p->tf = tf;
    if (p->syscall_map)
        num = (num < (uint64)p->map_size) ? p->syscall_map[num] : 0;
    if (num > 0 && num < NSYSCALLS && syscalls[num])
        tf->rax = syscalls[num]();
    else
        tf->rax = -ENOSYS;

    // A successful execve has replaced the program: its name is the new one.
    kinfo("strace %d %s: %s%d(%lx, %lx, %lx, %lx)%s%s%s = %ld\n",
         p->pid, p->name, p->syscall_map ? "" : "native ", (int)orig,
         a0, a1, a2, a3, pa >= 0 ? " \"" : "", path, pa >= 0 ? "\"" : "",
         (long)tf->rax);

    if (p->killed)
        exit(W_EXITCODE(0, p->killsig ? p->killsig : SIGKILL));
    deliver_signals(tf);
    kstack_check_canary("syscall exit");
}