entry: Add support for TIF_NOTIFY_SIGNAL
authorJens Axboe <axboe@kernel.dk>
Mon, 26 Oct 2020 20:32:28 +0000 (14:32 -0600)
committerThomas Gleixner <tglx@linutronix.de>
Thu, 29 Oct 2020 08:37:36 +0000 (09:37 +0100)
Add TIF_NOTIFY_SIGNAL handling in the generic entry code, which if set,
will return true if signal_pending() is used in a wait loop. That causes an
exit of the loop so that notify_signal tracehooks can be run. If the wait
loop is currently inside a system call, the system call is restarted once
task_work has been processed.

In preparation for only having arch_do_signal() handle syscall restarts if
_TIF_SIGPENDING isn't set, rename it to arch_do_signal_or_restart().  Pass
in a boolean that tells the architecture specific signal handler if it
should attempt to get a signal, or just process a potential syscall
restart.

For !CONFIG_GENERIC_ENTRY archs, add the TIF_NOTIFY_SIGNAL handling to
get_signal(). This is done to minimize the needed architecture changes to
support this feature.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Oleg Nesterov <oleg@redhat.com>
Link: https://lore.kernel.org/r/20201026203230.386348-3-axboe@kernel.dk
arch/x86/kernel/signal.c
include/linux/entry-common.h
include/linux/entry-kvm.h
include/linux/sched/signal.h
include/linux/tracehook.h
kernel/entry/common.c
kernel/entry/kvm.c
kernel/signal.c

index d5fa494..ec3b9c6 100644 (file)
@@ -804,11 +804,11 @@ static inline unsigned long get_nr_restart_syscall(const struct pt_regs *regs)
  * want to handle. Thus you cannot kill init even with a SIGKILL even by
  * mistake.
  */
-void arch_do_signal(struct pt_regs *regs)
+void arch_do_signal_or_restart(struct pt_regs *regs, bool has_signal)
 {
        struct ksignal ksig;
 
-       if (get_signal(&ksig)) {
+       if (has_signal && get_signal(&ksig)) {
                /* Whee! Actually deliver the signal.  */
                handle_signal(&ksig, regs);
                return;
index efebbff..c7bfac4 100644 (file)
 # define _TIF_UPROBE                   (0)
 #endif
 
+#ifndef _TIF_NOTIFY_SIGNAL
+# define _TIF_NOTIFY_SIGNAL            (0)
+#endif
+
 /*
  * TIF flags handled in syscall_enter_from_usermode()
  */
@@ -69,7 +73,7 @@
 
 #define EXIT_TO_USER_MODE_WORK                                         \
        (_TIF_SIGPENDING | _TIF_NOTIFY_RESUME | _TIF_UPROBE |           \
-        _TIF_NEED_RESCHED | _TIF_PATCH_PENDING |                       \
+        _TIF_NEED_RESCHED | _TIF_PATCH_PENDING | _TIF_NOTIFY_SIGNAL |  \
         ARCH_EXIT_TO_USER_MODE_WORK)
 
 /**
@@ -226,12 +230,13 @@ static __always_inline void arch_exit_to_user_mode(void) { }
 #endif
 
 /**
- * arch_do_signal -  Architecture specific signal delivery function
+ * arch_do_signal_or_restart -  Architecture specific signal delivery function
  * @regs:      Pointer to currents pt_regs
+ * @has_signal:        actual signal to handle
  *
  * Invoked from exit_to_user_mode_loop().
  */
-void arch_do_signal(struct pt_regs *regs);
+void arch_do_signal_or_restart(struct pt_regs *regs, bool has_signal);
 
 /**
  * arch_syscall_exit_tracehook - Wrapper around tracehook_report_syscall_exit()
index 0cef17a..9b93f85 100644 (file)
@@ -11,8 +11,8 @@
 # define ARCH_XFER_TO_GUEST_MODE_WORK  (0)
 #endif
 
-#define XFER_TO_GUEST_MODE_WORK                                        \
-       (_TIF_NEED_RESCHED | _TIF_SIGPENDING |                  \
+#define XFER_TO_GUEST_MODE_WORK                                                \
+       (_TIF_NEED_RESCHED | _TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL |     \
         _TIF_NOTIFY_RESUME | ARCH_XFER_TO_GUEST_MODE_WORK)
 
 struct kvm_vcpu;
index 404145d..bd5afa0 100644 (file)
@@ -360,6 +360,15 @@ static inline int task_sigpending(struct task_struct *p)
 
 static inline int signal_pending(struct task_struct *p)
 {
+#if defined(TIF_NOTIFY_SIGNAL)
+       /*
+        * TIF_NOTIFY_SIGNAL isn't really a signal, but it requires the same
+        * behavior in terms of ensuring that we break out of wait loops
+        * so that notify signal callbacks can be processed.
+        */
+       if (unlikely(test_tsk_thread_flag(p, TIF_NOTIFY_SIGNAL)))
+               return 1;
+#endif
        return task_sigpending(p);
 }
 
@@ -507,7 +516,7 @@ extern int set_user_sigmask(const sigset_t __user *umask, size_t sigsetsize);
 static inline void restore_saved_sigmask_unless(bool interrupted)
 {
        if (interrupted)
-               WARN_ON(!test_thread_flag(TIF_SIGPENDING));
+               WARN_ON(!signal_pending(current));
        else
                restore_saved_sigmask();
 }
index 36fb3bb..1e8caca 100644 (file)
@@ -198,4 +198,31 @@ static inline void tracehook_notify_resume(struct pt_regs *regs)
        blkcg_maybe_throttle_current();
 }
 
+/*
+ * called by exit_to_user_mode_loop() if ti_work & _TIF_NOTIFY_SIGNAL. This
+ * is currently used by TWA_SIGNAL based task_work, which requires breaking
+ * wait loops to ensure that task_work is noticed and run.
+ */
+static inline void tracehook_notify_signal(void)
+{
+#if defined(TIF_NOTIFY_SIGNAL)
+       clear_thread_flag(TIF_NOTIFY_SIGNAL);
+       smp_mb__after_atomic();
+       if (current->task_works)
+               task_work_run();
+#endif
+}
+
+/*
+ * Called when we have work to process from exit_to_user_mode_loop()
+ */
+static inline void set_notify_signal(struct task_struct *task)
+{
+#if defined(TIF_NOTIFY_SIGNAL)
+       if (!test_and_set_tsk_thread_flag(task, TIF_NOTIFY_SIGNAL) &&
+           !wake_up_state(task, TASK_INTERRUPTIBLE))
+               kick_process(task);
+#endif
+}
+
 #endif /* <linux/tracehook.h> */
index 9852e0d..42eff11 100644 (file)
@@ -109,7 +109,15 @@ static __always_inline void exit_to_user_mode(void)
 }
 
 /* Workaround to allow gradual conversion of architecture code */
-void __weak arch_do_signal(struct pt_regs *regs) { }
+void __weak arch_do_signal_or_restart(struct pt_regs *regs, bool has_signal) { }
+
+static void handle_signal_work(struct pt_regs *regs, unsigned long ti_work)
+{
+       if (ti_work & _TIF_NOTIFY_SIGNAL)
+               tracehook_notify_signal();
+
+       arch_do_signal_or_restart(regs, ti_work & _TIF_SIGPENDING);
+}
 
 static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
                                            unsigned long ti_work)
@@ -131,8 +139,8 @@ static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
                if (ti_work & _TIF_PATCH_PENDING)
                        klp_update_patch_state(current);
 
-               if (ti_work & _TIF_SIGPENDING)
-                       arch_do_signal(regs);
+               if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
+                       handle_signal_work(regs, ti_work);
 
                if (ti_work & _TIF_NOTIFY_RESUME) {
                        clear_thread_flag(TIF_NOTIFY_RESUME);
index eb1a8a4..b828a3d 100644 (file)
@@ -8,6 +8,9 @@ static int xfer_to_guest_mode_work(struct kvm_vcpu *vcpu, unsigned long ti_work)
        do {
                int ret;
 
+               if (ti_work & _TIF_NOTIFY_SIGNAL)
+                       tracehook_notify_signal();
+
                if (ti_work & _TIF_SIGPENDING) {
                        kvm_handle_signal_exit(vcpu);
                        return -EINTR;
index b179ecc..61b377e 100644 (file)
@@ -2529,6 +2529,20 @@ bool get_signal(struct ksignal *ksig)
        struct signal_struct *signal = current->signal;
        int signr;
 
+       /*
+        * For non-generic architectures, check for TIF_NOTIFY_SIGNAL so
+        * that the arch handlers don't all have to do it. If we get here
+        * without TIF_SIGPENDING, just exit after running signal work.
+        */
+#ifdef TIF_NOTIFY_SIGNAL
+       if (!IS_ENABLED(CONFIG_GENERIC_ENTRY)) {
+               if (test_thread_flag(TIF_NOTIFY_SIGNAL))
+                       tracehook_notify_signal();
+               if (!task_sigpending(current))
+                       return false;
+       }
+#endif
+
        if (unlikely(uprobe_deny_signal()))
                return false;