6 files changed, 73 insertions, 8 deletions
diff --git a/sys/arch/amd64/amd64/machdep.c b/sys/arch/amd64/amd64/machdep.c
index 82fa97b..528326a 100644
--- a/sys/arch/amd64/amd64/machdep.c
+++ b/sys/arch/amd64/amd64/machdep.c
@@ -95,20 +95,36 @@ try_mitigate_spectre(void)
 struct cpu_info *
 this_cpu(void)
 {
-    return (void *)amd64_read_gs_base();
+    struct cpu_info *ci;
+
+    /*
+     * This might look crazy but we are just leveraging the "m"
+     * constraint to calculate the offset of the self field within
+     * cpu_info. The self field points to the cpu_info structure
+     * itself allowing us to access cpu_info through %gs.
+     */
+    __ASMV("mov %%gs:%1, %0"
+        : "=r" (ci)
+        : "m" (*&((struct cpu_info *)0)->self)
+        : "memory");
+
+    return ci;
 }
 
 void
 cpu_startup(struct cpu_info *ci)
 {
+    ci->self = ci;
     gdt_load(&bsp_gdtr);
     idt_load();
 
     setup_vectors();
     amd64_write_gs_base((uintptr_t)ci);
-    init_tss(ci);
+    __ASMV("swapgs");           /* Get kernel GS */
 
+    init_tss(ci);
     try_mitigate_spectre();
+
     __ASMV("sti");              /* Unmask interrupts */
     lapic_init();
 }
diff --git a/sys/arch/amd64/amd64/proc_machdep.c b/sys/arch/amd64/amd64/proc_machdep.c
index 2b4368b..723582b 100644
--- a/sys/arch/amd64/amd64/proc_machdep.c
+++ b/sys/arch/amd64/amd64/proc_machdep.c
@@ -34,6 +34,7 @@
 #include <sys/exec.h>
 #include <machine/frame.h>
 #include <machine/gdt.h>
+#include <machine/cpu.h>
 #include <vm/physmem.h>
 #include <vm/vm.h>
 #include <vm/map.h>
@@ -113,6 +114,35 @@ setregs(struct proc *td, struct exec_prog *prog, uintptr_t stack)
 }
 
 /*
+ * Startup a user thread.
+ *
+ * @td: Thead to start up.
+ */
+void
+md_td_kick(struct proc *td)
+{
+    struct trapframe *tfp;
+
+    tfp = &td->tf;
+    __ASMV(
+        "push %0\n"
+        "push %1\n"
+        "pushf\n"
+        "push %2\n"
+        "push %3\n"
+        "swapgs\n"
+        "iretq"
+        :
+        : "i" (USER_DS | 3),
+          "r" (tfp->rsp),
+          "i" (USER_CS | 3),
+          "r" (tfp->rip)
+    );
+
+    __builtin_unreachable();
+}
+
+/*
  * MD thread init code
  *
  * @p: New process.
diff --git a/sys/include/arch/amd64/cpu.h b/sys/include/arch/amd64/cpu.h
index 16936e9..2ce363c 100644
--- a/sys/include/arch/amd64/cpu.h
+++ b/sys/include/arch/amd64/cpu.h
@@ -42,6 +42,7 @@ struct cpu_info {
     size_t lapic_tmr_freq;
     struct tss_entry *tss;
     struct proc *curtd;
+    struct cpu_info *self;
 };
 
 void cpu_startup(struct cpu_info *ci);
diff --git a/sys/include/arch/amd64/frameasm.h b/sys/include/arch/amd64/frameasm.h
index b8791ba..8983d36 100644
--- a/sys/include/arch/amd64/frameasm.h
+++ b/sys/include/arch/amd64/frameasm.h
@@ -106,11 +106,19 @@
  */
 #define INTRENTRY(ENTLABEL, HANDLER)    \
     ENTLABEL:                           \
-        push_trapframe $0             ; \
+        testq $0x3, 8(%rsp)           ; \
+        jz 1f                         ; \
+        lfence                        ; \
+        swapgs                        ; \
+    1:  push_trapframe $0             ; \
         mov %rsp, %rdi                ; \
         call HANDLER                  ; \
         pop_trapframe                 ; \
-        iretq
+        testq $0x3, 8(%rsp)           ; \
+        jz 2f                         ; \
+        lfence                        ; \
+        swapgs                        ; \
+    2:  iretq
 
 /*
  * Trap entry where an error code is on
@@ -118,10 +126,18 @@
  */
 #define TRAPENTRY(ENTLABEL, TRAPNO)    \
     ENTLABEL:                          \
-        push_trapframe_ec TRAPNO     ; \
+        testq $0x3, 8(%rsp)          ; \
+        jz 1f                        ; \
+        lfence                       ; \
+        swapgs                       ; \
+    1:  push_trapframe_ec TRAPNO     ; \
         mov %rsp, %rdi               ; \
         call trap_handler            ; \
         pop_trapframe_ec             ; \
-        iretq
+        testq $0x3, 8(%rsp)          ; \
+        jz 2f                        ; \
+        lfence                       ; \
+        swapgs                       ; \
+    2:  iretq
 
 #endif  /* !_MACHINE_FRAMEASM_H_ */
diff --git a/sys/include/sys/proc.h b/sys/include/sys/proc.h
index ba931f3..805c4a5 100644
--- a/sys/include/sys/proc.h
+++ b/sys/include/sys/proc.h
@@ -34,6 +34,7 @@
 #include <sys/spinlock.h>
 #include <sys/queue.h>
 #include <sys/param.h>
+#include <sys/cdefs.h>
 #if defined(_KERNEL)
 #include <machine/frame.h>
 #include <machine/pcb.h>
@@ -63,6 +64,7 @@ struct proc *this_td(void);
 int md_fork(struct proc *p, struct proc *parent, uintptr_t ip);
 
 void md_td_stackinit(struct proc *td, void *stack_top, struct exec_prog *prog);
+__dead void md_td_kick(struct proc *td);
 
 int fork1(struct proc *cur, int flags, void(*ip)(void), struct proc **newprocp);
 int exit1(struct proc *td);
diff --git a/sys/kern/kern_exec.c b/sys/kern/kern_exec.c
index b21404b..7825448 100644
--- a/sys/kern/kern_exec.c
+++ b/sys/kern/kern_exec.c
@@ -98,11 +98,11 @@ execve(struct proc *td, const struct execve_args *args)
     setregs(td, &prog, stack_top);
 
     /*
-     * Done, reset flags and yield
+     * Done, reset flags and start the user thread.
      *
      * XXX: PROC_EXEC is unset by the scheduler so it
      *      can properly start the new exec'd program.
      */
     td->flags &= ~PROC_INEXEC;
-    sched_enter();
+    md_td_kick(td);
 }