fix: corner case for x87 context restore on execve

[lunaix-os.git] / lunaix-os / kernel / process / signal.c

diff --git a/lunaix-os/kernel/process/signal.c b/lunaix-os/kernel/process/signal.c
index e8683a1b80f46ced290f989a8e60dc8c65712276..8c4322f442a9fd33680b4ebbb6ba327fb91170c5 100644 (file)
--- a/lunaix-os/kernel/process/signal.c
+++ b/lunaix-os/kernel/process/signal.c
@@ -4,102 +4,90 @@
 #include <lunaix/spike.h>
 #include <lunaix/status.h>
 #include <lunaix/syscall.h>
 #include <lunaix/spike.h>
 #include <lunaix/status.h>
 #include <lunaix/syscall.h>

+#include <lunaix/syslog.h>
+
+LOG_MODULE("SIG")

 
 #include <klibc/string.h>
 
 extern struct scheduler sched_ctx; /* kernel/sched.c */
 
 
 #include <klibc/string.h>
 
 extern struct scheduler sched_ctx; /* kernel/sched.c */
 

-void __USER__
-default_sighandler_term(int signum)
+#define UNMASKABLE (sigset(SIGKILL) | sigset(SIGTERM))
+#define TERMSIG (sigset(SIGSEGV) | sigset(SIGINT) | UNMASKABLE)
+#define CORE (sigset(SIGSEGV))
+
+static inline void
+signal_terminate(int errcode)

 {
 {

-    _exit(signum);
+    terminate_proc(errcode | PEXITSIG);

 }
 
 }
 

-void* default_handlers[_SIG_NUM] = {
-    // TODO: 添加默认handler
-    [_SIGINT] = default_sighandler_term,
-    [_SIGTERM] = default_sighandler_term,
-    [_SIGKILL] = default_sighandler_term,
-    [_SIGSEGV] = default_sighandler_term,
-};
-

 // Referenced in kernel/asm/x86/interrupt.S
 void*
 signal_dispatch()
 {
 // Referenced in kernel/asm/x86/interrupt.S
 void*
 signal_dispatch()
 {

-    if (!__current->sig_pending) {
+    if (!__current->sigctx.sig_pending) {

         // 没有待处理信号
         return 0;
     }
 
         // 没有待处理信号
         return 0;
     }
 

-    int sig_selected =
-      31 - __builtin_clz(__current->sig_pending &
-                         ~(__current->sig_mask | __current->sig_inprogress));
+    struct sighail* psig = &__current->sigctx;
+    struct sigact* prev_working = psig->inprogress;
+    sigset_t mask = psig->sig_mask | (prev_working ? prev_working->sa_mask : 0);
+
+    int sig_selected = 31 - __builtin_clz(psig->sig_pending & ~mask);

 
 

-    __SIGCLEAR(__current->sig_pending, sig_selected);
+    sigset_clear(psig->sig_pending, sig_selected);
+
+    struct sigact* action = &psig->signals[sig_selected];

 
     if (sig_selected == 0) {
         // SIG0 is reserved
         return 0;
     }
 
 
     if (sig_selected == 0) {
         // SIG0 is reserved
         return 0;
     }
 

-    // TODO: SIG{INT|TERM|KILL|SEGV} should have highest priority.
-    //       Terminate the process right here if any of unmaskable signal is
-    //       set.
-
-    if (!__current->sig_handler[sig_selected] &&
-        !default_handlers[sig_selected]) {
-        // 如果该信号没有handler，则忽略
+    if (!action->sa_actor) {
+        if (sigset_test(TERMSIG, sig_selected)) {
+            signal_terminate(sig_selected);
+            schedule();
+            // never return
+        }

         return 0;
     }
 
         return 0;
     }
 

-    uintptr_t ustack = __current->ustack_top & ~0xf;
+    ptr_t ustack = __current->ustack_top;

 
     if ((int)(ustack - USTACK_END) < (int)sizeof(struct proc_sig)) {
         // 用户栈没有空间存放信号上下文
         return 0;
     }
 
 
     if ((int)(ustack - USTACK_END) < (int)sizeof(struct proc_sig)) {
         // 用户栈没有空间存放信号上下文
         return 0;
     }
 

-    struct proc_sig* sig_ctx =
-      (struct proc_sig*)(ustack - sizeof(struct proc_sig));
+    struct proc_sig* sigframe =
+      (struct proc_sig*)((ustack - sizeof(struct proc_sig)) & ~0xf);

 
 

-    /*
-        这是一个相当恶心的坑。
-        问题是出在原本的sig_ctx->prev_context = __current->intr_ctx的上面
-        这个语句会被gcc在编译时，用更加高效的 rep movsl 来代替。
+    sigframe->sig_num = sig_selected;
+    sigframe->sigact = action->sa_actor;
+    sigframe->sighand = action->sa_handler;

 
 

-        由于我们采用按需分页，所以在很多情况下，用户栈实际被分配的空间不允许我们进行完整的
-        注入，而需要走page fault handler进行动态分页。
+    sigframe->saved_ictx = __current->intr_ctx;

 
 

-        竞态条件就出现在这里！
+    action->prev = prev_working;
+    psig->inprogress = action;

 
 

-        假若我们的__current->intr_ctx注入了一半，然后产生page-fault中断，
-        那么这就会导致我们的__current->intr_ctx被这个page-fault中断导致的
-        上下文信息覆盖。那么当page-fault handler成功分配了一个页，返回，
-        拷贝也就得以进行。遗憾的是，只不过这次拷贝的内容和前面的拷贝是没有任何的关系
-        （因为此时的intr_ctx已经不是之前的intr_ctx了！）
-        而这就会导致我们保存在信号上下文中的进程上下文信息不完整，从而在soft_iret时
-        触发#GP。
-
-        解决办法就是先吧intr_ctx拷贝到一个静态分配的区域里，然后再注入到用户栈。
-    */
-    static volatile struct proc_sigstate __temp_save;
-    __temp_save.proc_regs = __current->intr_ctx;
-    memcpy(__temp_save.fxstate, __current->fxstate, 512);
-
-    sig_ctx->prev_context = __temp_save;
-
-    sig_ctx->sig_num = sig_selected;
-    sig_ctx->signal_handler = __current->sig_handler[sig_selected];
-
-    if (!sig_ctx->signal_handler) {
-        // 如果没有用户自定义的Handler，则使用系统默认Handler。
-        sig_ctx->signal_handler = default_handlers[sig_selected];
-    }
+    return sigframe;
+}

 
 

-    __SIGSET(__current->sig_inprogress, sig_selected);
+void
+proc_clear_signal(struct proc_info* proc)
+{
+    memset(&proc->sigctx, 0, sizeof(proc->sigctx));
+}

 
 

-    return sig_ctx;
+void
+proc_setsignal(struct proc_info* proc, int signum)
+{
+    sigset_add(proc->sigctx.sig_pending, signum);
+    proc->sigctx.signals[signum].sender = __current->pid;

 }
 
 int
 }
 
 int


@@ -110,7 +98,9 @@ signal_send(pid_t pid, int signum)
         return -1;
     }
 
         return -1;
     }
 

+    pid_t sender_pid = __current->pid;

     struct proc_info* proc;
     struct proc_info* proc;

+

     if (pid > 0) {
         proc = get_process(pid);
         goto send_single;
     if (pid > 0) {
         proc = get_process(pid);
         goto send_single;


@@ -131,25 +121,57 @@ send_grp:
     struct proc_info *pos, *n;
     llist_for_each(pos, n, &proc->grp_member, grp_member)
     {
     struct proc_info *pos, *n;
     llist_for_each(pos, n, &proc->grp_member, grp_member)
     {

-        __SIGSET(pos->sig_pending, signum);
+        struct sighail* sh = &pos->sigctx;
+        sigset_add(sh->sig_pending, signum);
+        sh->signals[signum].sender = sender_pid;

     }
 
 send_single:
     }
 
 send_single:

-    if (PROC_TERMINATED(proc->state)) {
+    if (proc_terminated(proc)) {

         __current->k_status = EINVAL;
         return -1;
     }
         __current->k_status = EINVAL;
         return -1;
     }

-    __SIGSET(proc->sig_pending, signum);
+
+    sigset_add(proc->sigctx.sig_pending, signum);
+    proc->sigctx.signals[signum].sender = sender_pid;
+

     return 0;
 }
 
 __DEFINE_LXSYSCALL1(int, sigreturn, struct proc_sig, *sig_ctx)
 {
     return 0;
 }
 
 __DEFINE_LXSYSCALL1(int, sigreturn, struct proc_sig, *sig_ctx)
 {

-    memcpy(__current->fxstate, sig_ctx->prev_context.fxstate, 512);
-    __current->intr_ctx = sig_ctx->prev_context.proc_regs;
-    __current->flags &= ~PROC_FINPAUSE;
-    __SIGCLEAR(__current->sig_inprogress, sig_ctx->sig_num);
+    __current->intr_ctx = sig_ctx->saved_ictx;
+
+    struct sigact* current = __current->sigctx.inprogress;
+    if (current) {
+        __current->sigctx.inprogress = current->prev;
+        current->prev = NULL;
+    } else {
+        __current->sigctx.inprogress = NULL;
+    }
+
+    if (proc_terminated(__current)) {
+        __current->exit_code |= PEXITSIG;
+    } else if (sigset_test(CORE, sig_ctx->sig_num)) {
+        signal_terminate(sig_ctx->sig_num);
+    }
+
+    ptr_t ictx = (ptr_t)__current->intr_ctx;
+
+    /*
+        Ensure our restored context is within kernel stack
+
+        This prevent user to forge their own context such that arbitrary code
+       can be executed as supervisor level
+    */
+    if (!within_kstack(ictx)) {
+        signal_terminate(SIGSEGV);
+    }
+

     schedule();
     schedule();

+
+    // never reach!
+    return 0;

 }
 
 __DEFINE_LXSYSCALL3(int,
 }
 
 __DEFINE_LXSYSCALL3(int,


@@ -161,50 +183,48 @@ __DEFINE_LXSYSCALL3(int,
                     sigset_t,
                     *oldset)
 {
                     sigset_t,
                     *oldset)
 {

-    *oldset = __current->sig_mask;
+    struct sighail* sh = &__current->sigctx;
+    *oldset = sh->sig_mask;
+

     if (how == _SIG_BLOCK) {
     if (how == _SIG_BLOCK) {

-        __current->sig_mask |= *set;
+        sigset_union(sh->sig_mask, *set);

     } else if (how == _SIG_UNBLOCK) {
     } else if (how == _SIG_UNBLOCK) {

-        __current->sig_mask &= ~(*set);
+        sigset_intersect(sh->sig_mask, ~(*set));

     } else if (how == _SIG_SETMASK) {
     } else if (how == _SIG_SETMASK) {

-        __current->sig_mask = *set;
+        sh->sig_mask = *set;

     } else {
         return 0;
     }
     } else {
         return 0;
     }

-    __current->sig_mask &= ~_SIGNAL_UNMASKABLE;
+
+    sigset_intersect(sh->sig_mask, ~UNMASKABLE);

     return 1;
 }
 
     return 1;
 }
 

-__DEFINE_LXSYSCALL2(int, signal, int, signum, sighandler_t, handler)
+__DEFINE_LXSYSCALL2(int, sys_sigaction, int, signum, struct sigaction*, action)

 {
     if (signum <= 0 || signum >= _SIG_NUM) {
         return -1;
     }
 
 {
     if (signum <= 0 || signum >= _SIG_NUM) {
         return -1;
     }
 

-    if ((__SIGNAL(signum) & _SIGNAL_UNMASKABLE)) {
+    if (sigset_test(UNMASKABLE, signum)) {

         return -1;
     }
 
         return -1;
     }
 

-    __current->sig_handler[signum] = (void*)handler;
+    struct sigact* sa = &__current->sigctx.signals[signum];
+
+    sa->sa_actor = (void*)action->sa_sigaction;
+    sa->sa_handler = (void*)action->sa_handler;
+    sigset_union(sa->sa_mask, sigset(signum));

 
     return 0;
 }
 
 
     return 0;
 }
 

-void
-__do_pause()
+__DEFINE_LXSYSCALL(int, pause)

 {
 {

-    __current->flags |= PROC_FINPAUSE;
-
-    while ((__current->flags & PROC_FINPAUSE)) {
-        sched_yieldk();
-    }
+    pause_current();
+    sched_yieldk();

 
     __current->k_status = EINTR;
 
     __current->k_status = EINTR;

-}
-
-__DEFINE_LXSYSCALL(int, pause)
-{
-    __do_pause();

     return -1;
 }
 
     return -1;
 }
 


@@ -215,15 +235,18 @@ __DEFINE_LXSYSCALL2(int, kill, pid_t, pid, int, signum)
 
 __DEFINE_LXSYSCALL1(int, sigpending, sigset_t, *sigset)
 {
 
 __DEFINE_LXSYSCALL1(int, sigpending, sigset_t, *sigset)
 {

-    *sigset = __current->sig_pending;
+    *sigset = __current->sigctx.sig_pending;

     return 0;
 }
 
 __DEFINE_LXSYSCALL1(int, sigsuspend, sigset_t, *mask)
 {
     return 0;
 }
 
 __DEFINE_LXSYSCALL1(int, sigsuspend, sigset_t, *mask)
 {

-    sigset_t tmp = __current->sig_mask;
-    __current->sig_mask = (*mask) & ~_SIGNAL_UNMASKABLE;
-    __do_pause();
-    __current->sig_mask = tmp;
+    sigset_t tmp = __current->sigctx.sig_mask;
+    __current->sigctx.sig_mask = (*mask) & ~UNMASKABLE;
+
+    pause_current();
+    sched_yieldk();
+
+    __current->sigctx.sig_mask = tmp;

     return -1;
 }
\ No newline at end of file
     return -1;
 }
\ No newline at end of file