parlib: slab: Use the modern ctor/dtor interface
[akaros.git] / user / parlib / x86 / vcore.c
index b3c6182..32243d4 100644 (file)
@@ -1,7 +1,62 @@
 #include <ros/syscall.h>
-#include <parlib/arch/vcore.h>
+#include <parlib/vcore.h>
 #include <parlib/stdio.h>
+#include <stdlib.h>
 
+/* Here's how the HW popping works:  It sets up the future stack pointer to
+ * have extra stuff after it, and then it pops the registers, then pops the new
+ * context's stack pointer.  Then it uses the extra stuff (the new PC is on the
+ * stack, the location of notif_disabled, and a clobbered work register) to
+ * enable notifs, make sure notif IPIs weren't pending, restore the work reg,
+ * and then "ret".
+ *
+ * However, we can't just put the extra stuff directly below the rsp.  We need
+ * to leave room for the redzone: area that is potentially being used.  (Even if
+ * you compile with -mno-red-zone, some asm code (glibc memcpy) will still use
+ * that area).
+ *
+ * This is what the target uthread's stack will look like (growing down):
+ *
+ * Target RSP -> |   u_thread's old stuff   | the future %rsp, tf->tf_rsp
+ *               |            .             | beginning of Red Zone
+ *               |            .             |
+ *               |   128 Bytes of Red Zone  |
+ *               |            .             |
+ *               |            .             | end of Red Zone
+ *               |   new rip                | 0x08 below Red (one slot is 0x08)
+ *               |   rflags space           | 0x10 below
+ *               |   rdi save space         | 0x18 below
+ *               |   *sysc ptr to syscall   | 0x20 below
+ *               |   notif_pending_loc      | 0x28 below
+ *               |   notif_disabled_loc     | 0x30 below
+ *
+ * The important thing is that it can handle a notification after it enables
+ * notifications, and when it gets resumed it can ultimately run the new
+ * context.  Enough state is saved in the running context and stack to continue
+ * running.
+ *
+ * Related to that is whether or not our stack pointer is sufficiently far down
+ * so that restarting *this* code won't clobber shit we need later.  The way we
+ * do this is that we do any "stack jumping" before we enable interrupts/notifs.
+ * These jumps are when we directly modify rsp, specifically in the down
+ * direction (subtracts).  Adds would be okay. */
+
+/* Helper for writing the info we need later to the u_tf's stack.  Also, note
+ * this goes backwards, since memory reads up the stack. */
+struct restart_helper {
+       void                                            *notif_disab_loc;
+       void                                            *notif_pend_loc;
+       struct syscall                          *sysc;
+       uint64_t                                        rdi_save;
+       uint64_t                                        rflags;
+       uint64_t                                        rip;
+};
+
+/* Static syscall, used for self-notifying.  We never wait on it, and we
+ * actually might submit it multiple times in parallel on different cores!
+ * While this may seem dangerous, the kernel needs to be able to handle this
+ * scenario.  It's also important that we never wait on this, since for all but
+ * the first call, the DONE flag will be set.  (Set once, then never reset) */
 struct syscall vc_entry = {
        .num = SYS_vc_entry,
        .err = 0,
@@ -17,58 +72,358 @@ struct syscall vc_entry = {
        .arg5 = 0,
 };
 
-void print_hw_tf(struct hw_trapframe *hw_tf)
+static void pop_hw_tf(struct hw_trapframe *tf, uint32_t vcoreid)
+{
+       #define X86_RED_ZONE_SZ                 128
+       struct restart_helper *rst;
+       struct preempt_data *vcpd = &__procdata.vcore_preempt_data[vcoreid];
+
+       /* The stuff we need to write will be below the current stack and red zone
+        * of the utf */
+       rst = (struct restart_helper*)((void*)tf->tf_rsp - X86_RED_ZONE_SZ -
+                                      sizeof(struct restart_helper));
+       /* Fill in the info we'll need later */
+       rst->notif_disab_loc = &vcpd->notif_disabled;
+       rst->notif_pend_loc = &vcpd->notif_pending;
+       rst->sysc = &vc_entry;
+       rst->rdi_save = 0;                      /* avoid bugs */
+       rst->rflags = tf->tf_rflags;
+       rst->rip = tf->tf_rip;
+
+       asm volatile ("movq %0, %%rsp;       " /* jump rsp to the utf */
+                     "popq %%rax;           " /* restore registers */
+                     "popq %%rbx;           "
+                     "popq %%rcx;           "
+                     "popq %%rdx;           "
+                     "popq %%rbp;           "
+                     "popq %%rsi;           "
+                     "popq %%rdi;           "
+                     "popq %%r8;            "
+                     "popq %%r9;            "
+                     "popq %%r10;           "
+                     "popq %%r11;           "
+                     "popq %%r12;           "
+                     "popq %%r13;           "
+                     "popq %%r14;           "
+                     "popq %%r15;           "
+                     "addq $0x28, %%rsp;    " /* move to the rsp slot in the tf */
+                     "popq %%rsp;           " /* change to the utf's %rsp */
+                     "subq %[red], %%rsp;   " /* jump over the redzone */
+                     "subq $0x10, %%rsp;    " /* move rsp to below rdi's slot */
+                     "pushq %%rdi;          " /* save rdi, will clobber soon */
+                     "subq $0x18, %%rsp;    " /* move to notif_dis_loc slot */
+                     "popq %%rdi;           " /* load notif_disabled addr */
+                     "movb $0x00, (%%rdi);  " /* enable notifications */
+                                 /* Need a wrmb() here so the write of enable_notif can't pass
+                                  * the read of notif_pending (racing with a potential
+                                  * cross-core call with proc_notify()). */
+                                 "lock addb $0, (%%rdi);" /* LOCK is a CPU mb() */
+                                 /* From here down, we can get interrupted and restarted */
+                     "popq %%rdi;           " /* get notif_pending status loc */
+                     "testb $0x01, (%%rdi); " /* test if a notif is pending */
+                     "jz 1f;                " /* if not pending, skip syscall */
+                                 /* Actual syscall.  Note we don't wait on the async call */
+                     "popq %%rdi;           " /* &sysc, trap arg0 */
+                     "pushq %%rsi;          " /* save rax, will be trap arg1 */
+                     "pushq %%rax;          " /* save rax, will be trap ret */
+                     "movq $0x1, %%rsi;     " /* sending one async syscall: arg1 */
+                     "int %1;               " /* fire the syscall */
+                     "popq %%rax;           " /* restore regs after syscall */
+                     "popq %%rsi;           "
+                     "jmp 2f;               " /* skip 1:, already popped */
+                                 "1: addq $0x08, %%rsp; " /* discard &sysc (on non-sc path) */
+                     "2: popq %%rdi;        " /* restore tf's %rdi (both paths) */
+                                 "popfq;                " /* restore utf's rflags */
+                     "ret %[red];           " /* return to the new PC, skip red */
+                     :
+                     : "g"(&tf->tf_rax), "i"(T_SYSCALL), [red]"i"(X86_RED_ZONE_SZ)
+                     : "memory");
+}
+
+static void pop_sw_tf(struct sw_trapframe *sw_tf, uint32_t vcoreid)
+{
+       struct preempt_data *vcpd = &__procdata.vcore_preempt_data[vcoreid];
+
+       /* Restore callee-saved FPU state.  We need to clear exceptions before
+        * reloading the FP CW, in case the new CW unmasks any.  We also need to
+        * reset the tag word to clear out the stack.
+        *
+        * The main issue here is that while our context was saved in an
+        * ABI-complaint manner, we may be starting up on a somewhat random FPU
+        * state.  Having gibberish in registers isn't a big deal, but some of the
+        * FP environment settings could cause trouble.  If fnclex; emms isn't
+        * enough, we could also save/restore the entire FP env with fldenv, or do
+        * an fninit before fldcw. */
+       asm volatile ("ldmxcsr %0" : : "m"(sw_tf->tf_mxcsr));
+       asm volatile ("fnclex; emms; fldcw %0" : : "m"(sw_tf->tf_fpucw));
+       /* Basic plan: restore all regs, off rcx as the sw_tf.  Switch to the new
+        * stack, save the PC so we can jump to it later.  Use clobberably
+        * registers for the locations of sysc, notif_dis, and notif_pend. Once on
+        * the new stack, we enable notifs, check if we missed one, and if so, self
+        * notify.  Note the syscall clobbers rax. */
+       asm volatile ("movq 0x00(%0), %%rbx; " /* restore regs */
+                     "movq 0x08(%0), %%rbp; "
+                     "movq 0x10(%0), %%r12; "
+                     "movq 0x18(%0), %%r13; "
+                     "movq 0x20(%0), %%r14; "
+                     "movq 0x28(%0), %%r15; "
+                     "movq 0x30(%0), %%r8;  " /* save rip in r8 */
+                     "movq 0x38(%0), %%rsp; " /* jump to future stack */
+                     "movb $0x00, (%2);     " /* enable notifications */
+                     /* Need a wrmb() here so the write of enable_notif can't pass
+                      * the read of notif_pending (racing with a potential
+                      * cross-core call with proc_notify()). */
+                     "lock addb $0, (%2);   " /* LOCK is a CPU mb() */
+                     /* From here down, we can get interrupted and restarted */
+                     "testb $0x01, (%3);    " /* test if a notif is pending */
+                     "jz 1f;                " /* if not pending, skip syscall */
+                     /* Actual syscall.  Note we don't wait on the async call.
+                      * &vc_entry is already in rdi (trap arg0). */
+                     "movq $0x1, %%rsi;     " /* sending one async syscall: arg1 */
+                     "int %4;               " /* fire the syscall */
+                     "1: jmp *%%r8;         " /* ret saved earlier */
+                     :
+                     : "c"(&sw_tf->tf_rbx),
+                       "D"(&vc_entry),
+                       "S"(&vcpd->notif_disabled),
+                       "d"(&vcpd->notif_pending),
+                       "i"(T_SYSCALL)
+                     : "memory");
+}
+
+/* Pops a user context, reanabling notifications at the same time.  A Userspace
+ * scheduler can call this when transitioning off the transition stack.
+ *
+ * pop_user_ctx will fail if we have a notif_pending; you're not allowed to
+ * leave vcore context with notif_pending set.  Some code in vcore_entry
+ * needs clear notif_pending and check whatever might have caused a notif
+ * (e.g. call handle_events()).
+ *
+ * If notif_pending is not clear, this will self_notify this core, since it
+ * should be because we missed a notification message while notifs were
+ * disabled. */
+void pop_user_ctx(struct user_context *ctx, uint32_t vcoreid)
+{
+       struct preempt_data *vcpd = vcpd_of(vcoreid);
+
+       /* We check early for notif_pending, since if we deal with it during
+        * pop_hw_tf, we grow the stack slightly.  If a thread consistently fails to
+        * restart due to notif pending, it will eventually run off the bottom of
+        * its stack.  By performing the check here, we shrink that window.  You'd
+        * have to have a notif come after this check, but also *not* before this
+        * check.  If you PF in pop_user_ctx, this all failed. */
+       if (vcpd->notif_pending) {
+               /* if pop_user_ctx fails (and resets the vcore), the ctx contents must
+                * be in VCPD (due to !UTHREAD_SAVED).  it might already be there. */
+               if (ctx != &vcpd->uthread_ctx)
+                       vcpd->uthread_ctx = *ctx;
+               /* To restart the vcore, we must have the right TLS, stack pointer, and
+                * vc_ctx = TRUE. */
+               set_tls_desc((void*)vcpd->vcore_tls_desc);
+               begin_safe_access_tls_vars()
+               __vcore_context = TRUE;
+               end_safe_access_tls_vars()
+               set_stack_pointer((void*)vcpd->vcore_stack);
+               vcore_entry();
+       }
+       switch (ctx->type) {
+       case ROS_HW_CTX:
+               pop_hw_tf(&ctx->tf.hw_tf, vcoreid);
+               break;
+       case ROS_SW_CTX:
+               pop_sw_tf(&ctx->tf.sw_tf, vcoreid);
+               break;
+       case ROS_VM_CTX:
+               ros_syscall(SYS_pop_ctx, ctx, 0, 0, 0, 0, 0);
+               break;
+       }
+       assert(0);
+}
+
+/* Like the regular pop_user_ctx, but this one doesn't check or clear
+ * notif_pending.  The only case where we use this is when an IRQ/notif
+ * interrupts a uthread that is in the process of disabling notifs.
+ *
+ * If we need to support VM_CTXs here, we'll need to tell the kernel whether or
+ * not we want to enable_notifs (flag to SYS_pop_ctx).  The only use case for
+ * this is when disabling notifs.  Currently, a VM can't do this or do things
+ * like uthread_yield.  It doesn't have access to the vcore's or uthread's TLS
+ * to bootstrap any of that stuff. */
+void pop_user_ctx_raw(struct user_context *ctx, uint32_t vcoreid)
+{
+       struct hw_trapframe *tf = &ctx->tf.hw_tf;
+       struct restart_helper *rst;
+       struct preempt_data *vcpd = &__procdata.vcore_preempt_data[vcoreid];
+
+       assert(ctx->type == ROS_HW_CTX);
+       /* The stuff we need to write will be below the current stack of the utf */
+       rst = (struct restart_helper*)((void*)tf->tf_rsp -
+                                      sizeof(struct restart_helper));
+       /* Fill in the info we'll need later */
+       rst->notif_disab_loc = &vcpd->notif_disabled;
+       rst->rdi_save = 0;                      /* avoid bugs */
+       rst->rflags = tf->tf_rflags;
+       rst->rip = tf->tf_rip;
+
+       asm volatile ("movq %0, %%rsp;       " /* jump esp to the utf */
+                     "popq %%rax;           " /* restore registers */
+                     "popq %%rbx;           "
+                     "popq %%rcx;           "
+                     "popq %%rdx;           "
+                     "popq %%rbp;           "
+                     "popq %%rsi;           "
+                     "popq %%rdi;           "
+                     "popq %%r8;            "
+                     "popq %%r9;            "
+                     "popq %%r10;           "
+                     "popq %%r11;           "
+                     "popq %%r12;           "
+                     "popq %%r13;           "
+                     "popq %%r14;           "
+                     "popq %%r15;           "
+                     "addq $0x28, %%rsp;    " /* move to the rsp slot in the tf */
+                     "popq %%rsp;           " /* change to the utf's %rsp */
+                     "subq $0x10, %%rsp;    " /* move rsp to below rdi's slot */
+                     "pushq %%rdi;          " /* save rdi, will clobber soon */
+                     "subq $0x18, %%rsp;    " /* move to notif_dis_loc slot */
+                     "popq %%rdi;           " /* load notif_disabled addr */
+                     "movb $0x00, (%%rdi);  " /* enable notifications */
+                                 /* Here's where we differ from the regular pop_user_ctx().
+                                  * We need to adjust rsp and whatnot, but don't do test,
+                                  * clear notif_pending, or call a syscall. */
+                                 /* From here down, we can get interrupted and restarted */
+                     "addq $0x10, %%rsp;    " /* move to rdi save slot */
+                     "popq %%rdi;           " /* restore tf's %rdi */
+                                 "popfq;                " /* restore utf's rflags */
+                     "ret;                  " /* return to the new PC */
+                     :
+                     : "g"(&tf->tf_rax)
+                     : "memory");
+}
+
+void fprintf_hw_tf(FILE *f, struct hw_trapframe *hw_tf)
 {
-       printf("[user] HW TRAP frame 0x%016x\n", hw_tf);
-       printf("  rax  0x%016lx\n",           hw_tf->tf_rax);
-       printf("  rbx  0x%016lx\n",           hw_tf->tf_rbx);
-       printf("  rcx  0x%016lx\n",           hw_tf->tf_rcx);
-       printf("  rdx  0x%016lx\n",           hw_tf->tf_rdx);
-       printf("  rbp  0x%016lx\n",           hw_tf->tf_rbp);
-       printf("  rsi  0x%016lx\n",           hw_tf->tf_rsi);
-       printf("  rdi  0x%016lx\n",           hw_tf->tf_rdi);
-       printf("  r8   0x%016lx\n",           hw_tf->tf_r8);
-       printf("  r9   0x%016lx\n",           hw_tf->tf_r9);
-       printf("  r10  0x%016lx\n",           hw_tf->tf_r10);
-       printf("  r11  0x%016lx\n",           hw_tf->tf_r11);
-       printf("  r12  0x%016lx\n",           hw_tf->tf_r12);
-       printf("  r13  0x%016lx\n",           hw_tf->tf_r13);
-       printf("  r14  0x%016lx\n",           hw_tf->tf_r14);
-       printf("  r15  0x%016lx\n",           hw_tf->tf_r15);
-       printf("  trap 0x%08x\n",             hw_tf->tf_trapno);
-       printf("  gsbs 0x%016lx\n",           hw_tf->tf_gsbase);
-       printf("  fsbs 0x%016lx\n",           hw_tf->tf_fsbase);
-       printf("  err  0x--------%08x\n",     hw_tf->tf_err);
-       printf("  rip  0x%016lx\n",           hw_tf->tf_rip);
-       printf("  cs   0x------------%04x\n", hw_tf->tf_cs);
-       printf("  flag 0x%016lx\n",           hw_tf->tf_rflags);
-       printf("  rsp  0x%016lx\n",           hw_tf->tf_rsp);
-       printf("  ss   0x------------%04x\n", hw_tf->tf_ss);
+       fprintf(f, "[user] HW TRAP frame 0x%016x\n", hw_tf);
+       fprintf(f, "  rax  0x%016lx\n",           hw_tf->tf_rax);
+       fprintf(f, "  rbx  0x%016lx\n",           hw_tf->tf_rbx);
+       fprintf(f, "  rcx  0x%016lx\n",           hw_tf->tf_rcx);
+       fprintf(f, "  rdx  0x%016lx\n",           hw_tf->tf_rdx);
+       fprintf(f, "  rbp  0x%016lx\n",           hw_tf->tf_rbp);
+       fprintf(f, "  rsi  0x%016lx\n",           hw_tf->tf_rsi);
+       fprintf(f, "  rdi  0x%016lx\n",           hw_tf->tf_rdi);
+       fprintf(f, "  r8   0x%016lx\n",           hw_tf->tf_r8);
+       fprintf(f, "  r9   0x%016lx\n",           hw_tf->tf_r9);
+       fprintf(f, "  r10  0x%016lx\n",           hw_tf->tf_r10);
+       fprintf(f, "  r11  0x%016lx\n",           hw_tf->tf_r11);
+       fprintf(f, "  r12  0x%016lx\n",           hw_tf->tf_r12);
+       fprintf(f, "  r13  0x%016lx\n",           hw_tf->tf_r13);
+       fprintf(f, "  r14  0x%016lx\n",           hw_tf->tf_r14);
+       fprintf(f, "  r15  0x%016lx\n",           hw_tf->tf_r15);
+       fprintf(f, "  trap 0x%08x\n",             hw_tf->tf_trapno);
+       fprintf(f, "  gsbs 0x%016lx\n",           hw_tf->tf_gsbase);
+       fprintf(f, "  fsbs 0x%016lx\n",           hw_tf->tf_fsbase);
+       fprintf(f, "  err  0x--------%08x\n",     hw_tf->tf_err);
+       fprintf(f, "  rip  0x%016lx\n",           hw_tf->tf_rip);
+       fprintf(f, "  cs   0x------------%04x\n", hw_tf->tf_cs);
+       fprintf(f, "  flag 0x%016lx\n",           hw_tf->tf_rflags);
+       fprintf(f, "  rsp  0x%016lx\n",           hw_tf->tf_rsp);
+       fprintf(f, "  ss   0x------------%04x\n", hw_tf->tf_ss);
 }
 
-void print_sw_tf(struct sw_trapframe *sw_tf)
+void fprintf_sw_tf(FILE *f, struct sw_trapframe *sw_tf)
 {
-       printf("[user] SW TRAP frame 0x%016p\n", sw_tf);
-       printf("  rbx  0x%016lx\n",           sw_tf->tf_rbx);
-       printf("  rbp  0x%016lx\n",           sw_tf->tf_rbp);
-       printf("  r12  0x%016lx\n",           sw_tf->tf_r12);
-       printf("  r13  0x%016lx\n",           sw_tf->tf_r13);
-       printf("  r14  0x%016lx\n",           sw_tf->tf_r14);
-       printf("  r15  0x%016lx\n",           sw_tf->tf_r15);
-       printf("  gsbs 0x%016lx\n",           sw_tf->tf_gsbase);
-       printf("  fsbs 0x%016lx\n",           sw_tf->tf_fsbase);
-       printf("  rip  0x%016lx\n",           sw_tf->tf_rip);
-       printf("  rsp  0x%016lx\n",           sw_tf->tf_rsp);
-       printf(" mxcsr 0x%08x\n",             sw_tf->tf_mxcsr);
-       printf(" fpucw 0x%04x\n",             sw_tf->tf_fpucw);
+       fprintf(f, "[user] SW TRAP frame 0x%016p\n", sw_tf);
+       fprintf(f, "  rbx  0x%016lx\n",           sw_tf->tf_rbx);
+       fprintf(f, "  rbp  0x%016lx\n",           sw_tf->tf_rbp);
+       fprintf(f, "  r12  0x%016lx\n",           sw_tf->tf_r12);
+       fprintf(f, "  r13  0x%016lx\n",           sw_tf->tf_r13);
+       fprintf(f, "  r14  0x%016lx\n",           sw_tf->tf_r14);
+       fprintf(f, "  r15  0x%016lx\n",           sw_tf->tf_r15);
+       fprintf(f, "  gsbs 0x%016lx\n",           sw_tf->tf_gsbase);
+       fprintf(f, "  fsbs 0x%016lx\n",           sw_tf->tf_fsbase);
+       fprintf(f, "  rip  0x%016lx\n",           sw_tf->tf_rip);
+       fprintf(f, "  rsp  0x%016lx\n",           sw_tf->tf_rsp);
+       fprintf(f, " mxcsr 0x%08x\n",             sw_tf->tf_mxcsr);
+       fprintf(f, " fpucw 0x%04x\n",             sw_tf->tf_fpucw);
+}
+
+void fprintf_vm_tf(FILE *f, struct vm_trapframe *vm_tf)
+{
+       fprintf(f, "[user] VM Trapframe 0x%016x\n", vm_tf);
+       fprintf(f, "  rax  0x%016lx\n",           vm_tf->tf_rax);
+       fprintf(f, "  rbx  0x%016lx\n",           vm_tf->tf_rbx);
+       fprintf(f, "  rcx  0x%016lx\n",           vm_tf->tf_rcx);
+       fprintf(f, "  rdx  0x%016lx\n",           vm_tf->tf_rdx);
+       fprintf(f, "  rbp  0x%016lx\n",           vm_tf->tf_rbp);
+       fprintf(f, "  rsi  0x%016lx\n",           vm_tf->tf_rsi);
+       fprintf(f, "  rdi  0x%016lx\n",           vm_tf->tf_rdi);
+       fprintf(f, "  r8   0x%016lx\n",           vm_tf->tf_r8);
+       fprintf(f, "  r9   0x%016lx\n",           vm_tf->tf_r9);
+       fprintf(f, "  r10  0x%016lx\n",           vm_tf->tf_r10);
+       fprintf(f, "  r11  0x%016lx\n",           vm_tf->tf_r11);
+       fprintf(f, "  r12  0x%016lx\n",           vm_tf->tf_r12);
+       fprintf(f, "  r13  0x%016lx\n",           vm_tf->tf_r13);
+       fprintf(f, "  r14  0x%016lx\n",           vm_tf->tf_r14);
+       fprintf(f, "  r15  0x%016lx\n",           vm_tf->tf_r15);
+       fprintf(f, "  rip  0x%016lx\n",           vm_tf->tf_rip);
+       fprintf(f, "  rflg 0x%016lx\n",           vm_tf->tf_rflags);
+       fprintf(f, "  rsp  0x%016lx\n",           vm_tf->tf_rsp);
+       fprintf(f, "  cr2  0x%016lx\n",           vm_tf->tf_cr2);
+       fprintf(f, "  cr3  0x%016lx\n",           vm_tf->tf_cr3);
+       fprintf(f, "Gpcore 0x%08x\n",             vm_tf->tf_guest_pcoreid);
+       fprintf(f, "Flags  0x%08x\n",             vm_tf->tf_flags);
+       fprintf(f, "Inject 0x%08x\n",             vm_tf->tf_trap_inject);
+       fprintf(f, "ExitRs 0x%08x\n",             vm_tf->tf_exit_reason);
+       fprintf(f, "ExitQl 0x%08x\n",             vm_tf->tf_exit_qual);
+       fprintf(f, "Intr1  0x%016lx\n",           vm_tf->tf_intrinfo1);
+       fprintf(f, "Intr2  0x%016lx\n",           vm_tf->tf_intrinfo2);
+       fprintf(f, "GIntr  0x----%04x\n",         vm_tf->tf_guest_intr_status);
+       fprintf(f, "GVA    0x%016lx\n",           vm_tf->tf_guest_va);
+       fprintf(f, "GPA    0x%016lx\n",           vm_tf->tf_guest_pa);
 }
 
 void print_user_context(struct user_context *ctx)
 {
-       if (ctx->type == ROS_HW_CTX)
-               print_hw_tf(&ctx->tf.hw_tf);
-       else if (ctx->type == ROS_SW_CTX)
-               print_sw_tf(&ctx->tf.sw_tf);
-       else
-               printf("Unknown context type %d\n", ctx->type);
+       switch (ctx->type) {
+       case ROS_HW_CTX:
+               fprintf_hw_tf(stdout, &ctx->tf.hw_tf);
+               break;
+       case ROS_SW_CTX:
+               fprintf_sw_tf(stdout, &ctx->tf.sw_tf);
+               break;
+       case ROS_VM_CTX:
+               fprintf_vm_tf(stdout, &ctx->tf.vm_tf);
+               break;
+       default:
+               fprintf(stderr, "Unknown context type %d\n", ctx->type);
+       }
+}
+
+/* The second-lowest level function jumped to by the kernel on every vcore
+ * entry.  We get called from __kernel_vcore_entry.
+ *
+ * We should consider making it mandatory to set the tls_desc in the kernel. We
+ * wouldn't even need to pass the vcore id to user space at all if we did this.
+ * It would already be set in the preinstalled TLS as __vcore_id. */
+void __attribute__((noreturn)) __kvc_entry_c(void)
+{
+       /* The kernel sets the TLS desc for us, based on whatever is in VCPD.
+        *
+        * x86 32-bit TLS is pretty jacked up, so the kernel doesn't set the TLS
+        * desc for us.  it's a little more expensive to do it here, esp for
+        * amd64.  Can remove this when/if we overhaul 32 bit TLS. */
+       int id = __vcore_id_on_entry;
+
+       #ifndef __x86_64__
+       set_tls_desc(vcpd_of(id)->vcore_tls_desc);
+       #endif
+       /* Every time the vcore comes up, it must set that it is in vcore context.
+        * uthreads may share the same TLS as their vcore (when uthreads do not have
+        * their own TLS), and if a uthread was preempted, __vcore_context == FALSE,
+        * and that will continue to be true the next time the vcore pops up. */
+       __vcore_context = TRUE;
+       vcore_entry();
+       fprintf(stderr, "vcore_entry() should never return!\n");
+       abort();
+       __builtin_unreachable();
 }