kern/arch/x86/trap.c

#include <arch/mmu.h>
#include <arch/x86.h>
#include <arch/arch.h>
#include <arch/console.h>
#include <arch/apic.h>
#include <arch/perfmon.h>
#include <ros/common.h>
#include <smp.h>
#include <assert.h>
#include <pmap.h>
#include <trap.h>
#include <monitor.h>
#include <process.h>
#include <mm.h>
#include <stdio.h>
#include <slab.h>
#include <syscall.h>
#include <kdebug.h>
#include <kmalloc.h>
#include <ex_table.h>
#include <arch/mptables.h>
#include <ros/procinfo.h>

enum {
	NMI_NORMAL_OPN = 0,
	NMI_IN_PROGRESS,
	NMI_HANDLE_ANOTHER,
};

taskstate_t ts;

/* Interrupt descriptor table.  64 bit needs 16 byte alignment (i think). */
gatedesc_t __attribute__((aligned (16))) idt[256] = { { 0 } };
pseudodesc_t idt_pd;

/* interrupt handler table, each element is a linked list of handlers for a
 * given IRQ.  Modification requires holding the lock (TODO: RCU) */
struct irq_handler *irq_handlers[NUM_IRQS];
spinlock_t irq_handler_wlock = SPINLOCK_INITIALIZER_IRQSAVE;

static bool try_handle_exception_fixup(struct hw_trapframe *hw_tf)
{
	if (in_kernel(hw_tf)) {
		uintptr_t fixup_ip = get_fixup_ip(hw_tf->tf_rip);

		if (fixup_ip != 0) {
			hw_tf->tf_rip = fixup_ip;
			return true;
		}
	}

	return false;
}

const char *x86_trapname(int trapno)
{
	static const char *const excnames[] = {
		"Divide error",
		"Debug",
		"Non-Maskable Interrupt",
		"Breakpoint",
		"Overflow",
		"BOUND Range Exceeded",
		"Invalid Opcode",
		"Device Not Available",
		"Double Fault",
		"Coprocessor Segment Overrun",
		"Invalid TSS",
		"Segment Not Present",
		"Stack Fault",
		"General Protection",
		"Page Fault",
		"(unknown trap)",
		"x87 FPU Floating-Point Error",
		"Alignment Check",
		"Machine-Check",
		"SIMD Floating-Point Exception"
	};

	if (trapno < sizeof(excnames)/sizeof(excnames[0]))
		return excnames[trapno];
	if (trapno == T_SYSCALL)
		return "System call";
	return "(unknown trap)";
}

/* Set stacktop for the current core to be the stack the kernel will start on
 * when trapping/interrupting from userspace. */
void set_stack_top(uintptr_t stacktop)
{
	struct per_cpu_info *pcpui = &per_cpu_info[core_id()];

	/* No need to reload the task register, this takes effect immediately */
	x86_set_stacktop_tss(pcpui->tss, stacktop);
	/* Also need to make sure sysenters come in correctly */
	x86_set_sysenter_stacktop(stacktop);
}

/* Note the check implies we only are on a one page stack (or the first page) */
uintptr_t get_stack_top(void)
{
	struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
	uintptr_t stacktop;

	stacktop = x86_get_stacktop_tss(pcpui->tss);
	if (stacktop != ROUNDUP(read_sp(), PGSIZE))
		panic("Bad stacktop: %p esp one is %p\n", stacktop,
		      ROUNDUP(read_sp(), PGSIZE));
	return stacktop;
}

/* Sends a non-maskable interrupt; the handler will print a trapframe. */
void send_nmi(uint32_t os_coreid)
{
	/* NMI / IPI for x86 are limited to 8 bits */
	uint8_t hw_core = (uint8_t)get_hw_coreid(os_coreid);

	__send_nmi(hw_core);
}

void idt_init(void)
{
	/* This table is made in trapentry$BITS.S by each macro in that file.
	 * It is layed out such that the ith entry is the ith's traphandler's
	 * (uintptr_t) trap addr, then (uint32_t) trap number. */
	struct trapinfo { uintptr_t trapaddr; uint32_t trapnumber; }
	       __attribute__((packed));
	extern struct trapinfo trap_tbl[];
	extern struct trapinfo trap_tbl_end[];
	int i, trap_tbl_size = trap_tbl_end - trap_tbl;
	extern void ISR_default(void);
	extern void ISR_syscall(void);

	/* set all to default, to catch everything */
	for (i = 0; i < 256; i++)
		SETGATE(idt[i], 0, GD_KT, &ISR_default, 0);

	/* set all entries that have real trap handlers
	 * we need to stop short of the last one, since the last is the default
	 * handler with a fake interrupt number (500) that is out of bounds of
	 * the idt[] */
	for (i = 0; i < trap_tbl_size - 1; i++)
		SETGATE(idt[trap_tbl[i].trapnumber], 0, GD_KT,
			trap_tbl[i].trapaddr, 0);
	/* Sanity check */
	assert((uintptr_t)ISR_syscall ==
	       ((uintptr_t)idt[T_SYSCALL].gd_off_63_32 << 32 |
	        (uintptr_t)idt[T_SYSCALL].gd_off_31_16 << 16 |
	        (uintptr_t)idt[T_SYSCALL].gd_off_15_0));
	/* turn on trap-based syscall handling and other user-accessible ints
	 * DPL 3 means this can be triggered by the int instruction */
	idt[T_SYSCALL].gd_dpl = 3;
	idt[T_BRKPT].gd_dpl = 3;
	/* Send NMIs to their own stack (IST1 in every core's TSS) */
	idt[T_NMI].gd_ist = 1;
	/* Send double faults to their own stack (IST2 in every core's TSS) */
	idt[T_DBLFLT].gd_ist = 2;

	/* The sooner we set this, the sooner we can use set/get_stack_top. */
	per_cpu_info[0].tss = &ts;
	per_cpu_info[0].gdt = gdt;

	/* Set up our kernel stack when changing rings */
	/* Note: we want 16 byte aligned kernel stack frames (AMD 2:8.9.3) */
	x86_sysenter_init();
	/* We will set this properly once we have a kstack from the slab. */
	set_stack_top(0xdeadbeef);

	/* Initialize the TSS field of the gdt.  The size of the TSS desc
	 * differs between 64 and 32 bit, hence the pointer acrobatics */
	syssegdesc_t *ts_slot = (syssegdesc_t*)&gdt[GD_TSS >> 3];
	*ts_slot = (syssegdesc_t)SEG_SYS_SMALL(STS_T32A, (uintptr_t)&ts,
	                                       sizeof(taskstate_t), 0);

	/* Init the IDT PD.  Need to do this before ltr for some reason.  (Doing
	 * this between ltr and lidt causes the machine to reboot... */
	idt_pd.pd_lim = sizeof(idt) - 1;
	idt_pd.pd_base = (uintptr_t)idt;

	ltr(GD_TSS);

	asm volatile("lidt %0" : : "m"(idt_pd));

	pic_remap();
	pic_mask_all();

	int ncleft = MAX_NUM_CORES;
	int num_cores_mpacpi;

	ncleft = mpsinit(ncleft);
	ncleft = mpacpi(ncleft);
	num_cores_mpacpi = MAX_NUM_CORES - ncleft;
	printk("MP and ACPI found %d cores\n", num_cores_mpacpi);
	if (num_cores != num_cores_mpacpi)
		warn("Topology (%d) and MP/ACPI (%d) differ on num_cores!",
		     num_cores, num_cores_mpacpi);

	apiconline();
	ioapiconline();

	/* the lapic IRQs need to be unmasked on a per-core basis */
	register_irq(IdtLAPIC_TIMER, timer_interrupt, NULL,
	             MKBUS(BusLAPIC, 0, 0, 0));
	register_irq(IdtLAPIC_ERROR, handle_lapic_error, NULL,
	             MKBUS(BusLAPIC, 0, 0, 0));
	register_irq(IdtLAPIC_PCINT, perfmon_interrupt, NULL,
	             MKBUS(BusLAPIC, 0, 0, 0));
	register_irq(I_KERNEL_MSG, handle_kmsg_ipi, NULL,
		     MKBUS(BusIPI, 0, 0, 0));
}

static void print_fperr(struct hw_trapframe *hw_tf)
{
	uint16_t fpcw, fpsw;
	uint32_t mxcsr;

	asm volatile ("fnstcw %0" : "=m"(fpcw));
	asm volatile ("fnstsw %0" : "=m"(fpsw));
	asm volatile ("stmxcsr %0" : "=m"(mxcsr));
	print_lock();
	print_trapframe(hw_tf);
	printk("Core %d: FP ERR, CW: 0x%04x, SW: 0x%04x, MXCSR 0x%08x\n",
	       core_id(), fpcw, fpsw, mxcsr);
	printk("Core %d: The following faults are unmasked:\n", core_id());
	if (fpsw & ~fpcw & FP_EXCP_IE) {
		printk("\tInvalid Operation: ");
		if (fpsw & FP_SW_SF) {
			if (fpsw & FP_SW_C1)
				printk("Stack overflow\n");
			else
				printk("Stack underflow\n");
		} else {
			printk("invalid arithmetic operand\n");
		}
	}
	if (fpsw & ~fpcw & FP_EXCP_DE)
		printk("\tDenormalized operand\n");
	if (fpsw & ~fpcw & FP_EXCP_ZE)
		printk("\tDivide by zero\n");
	if (fpsw & ~fpcw & FP_EXCP_OE)
		printk("\tNumeric Overflow\n");
	if (fpsw & ~fpcw & FP_EXCP_UE)
		printk("\tNumeric Underflow\n");
	if (fpsw & ~fpcw & FP_EXCP_PE)
		printk("\tInexact result (precision)\n");
	print_unlock();
}

static bool __handler_user_page_fault(struct hw_trapframe *hw_tf,
                                      uintptr_t fault_va, int prot)
{
	struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
	int err;

	assert(pcpui->owning_proc == pcpui->cur_proc);
	enable_irq();
	err = handle_page_fault(pcpui->owning_proc, fault_va, prot);
	disable_irq();
	if (err) {
		if (err == -EAGAIN)
			hw_tf->tf_err |= PF_VMR_BACKED;
		return FALSE;
	}
	return TRUE;
}

static bool __handler_kernel_page_fault(struct hw_trapframe *hw_tf,
                                        uintptr_t fault_va, int prot)
{
	struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
	int err;

	/* The only thing an NMI handler that faults can do is a fixup */
	if (pcpui->nmi_status != NMI_NORMAL_OPN) {
		assert(in_kernel(hw_tf));
		return try_handle_exception_fixup(hw_tf);
	}
	/* In general, if there's no cur_proc, a KPF is a bug. */
	if (!pcpui->cur_proc) {
		/* This only runs from test_uaccess(), where it is expected to
		 * fail. */
		if (try_handle_exception_fixup(hw_tf))
			return TRUE;
		panic_hwtf(hw_tf, "Proc-less Page Fault in the Kernel at %p!",
		           fault_va);
	}
	/* TODO - handle kernel page faults.  This is dangerous, since we might
	 * be holding locks in the kernel and could deadlock when we HPF.  For
	 * now, I'm just disabling the lock checker, since it'll flip out when
	 * it sees there is a kernel trap.  Will need to think about this a bit,
	 * esp when we properly handle bad addrs and whatnot. */
	pcpui->__lock_checking_enabled--;
	/* It is a bug for the kernel to access user memory while holding locks
	 * that are used by handle_page_fault.  At a minimum, this includes
	 * p->vmr_lock and memory allocation locks.
	 *
	 * In an effort to reduce the number of locks (both now and in the
	 * future), the kernel will not attempt to handle faults on file-back
	 * VMRs.  We probably can turn that on in the future, but I'd rather
	 * keep things safe for now.  (We'll probably need to change this when
	 * we stop MAP_POPULATE | MAP_LOCKED entire binaries).
	 *
	 * Note that we do not enable IRQs here, unlike in the user case.
	 * Again, this is to limit the locks we could be grabbing. */
	err = handle_page_fault_nofile(pcpui->cur_proc, fault_va, prot);
	pcpui->__lock_checking_enabled++;
	if (err) {
		if (try_handle_exception_fixup(hw_tf))
			return TRUE;
		/* Turn this on to help debug bad function pointers */
		printd("rsp %p\n\t 0(rsp): %p\n\t 8(rsp): %p\n\t 16(rsp): %p\n"
		       "\t24(rsp): %p\n", hw_tf->tf_rsp,
		       *(uintptr_t*)(hw_tf->tf_rsp +  0),
		       *(uintptr_t*)(hw_tf->tf_rsp +  8),
		       *(uintptr_t*)(hw_tf->tf_rsp + 16),
		       *(uintptr_t*)(hw_tf->tf_rsp + 24));
		panic_hwtf(hw_tf, "Proc-ful Page Fault in the Kernel at %p!",
			   fault_va);
		/* if we want to do something like kill a process or other code,
		 * be aware we are in a sort of irq-like context, meaning the
		 * main kernel code we 'interrupted' could be holding locks -
		 * even irqsave locks. */
	}
	return TRUE;
}

static bool __handle_page_fault(struct hw_trapframe *hw_tf, unsigned long *aux)
{
	uintptr_t fault_va = rcr2();
	int prot = hw_tf->tf_err & PF_ERROR_WRITE ? PROT_WRITE : PROT_READ;

	*aux = fault_va;
	if (in_kernel(hw_tf))
		return __handler_kernel_page_fault(hw_tf, fault_va, prot);
	else
		return __handler_user_page_fault(hw_tf, fault_va, prot);
}

/* Actual body of work done when an NMI arrives */
static void do_nmi_work(struct hw_trapframe *hw_tf)
{
	assert(!irq_is_enabled());
	/* It's mostly harmless to snapshot the TF, and we can send a spurious
	 * PCINT interrupt.  perfmon.c just uses the interrupt to tell it to
	 * check its counters for overflow.  Note that the PCINT interrupt is
	 * just a regular IRQ.  The backtrace was recorded during the NMI and
	 * emitted during IRQ.
	 *
	 * That being said, it's OK if the monitor triggers debugging NMIs while
	 * perf is running.  If perf triggers an NMI when the monitor wants to
	 * print, the monitor will debug *that* NMI, and not the one that gets
	 * sent moments later.  That's fine. */
	emit_monitor_backtrace(ROS_HW_CTX, hw_tf);
	perfmon_snapshot_hwtf(hw_tf);
	send_self_ipi(IdtLAPIC_PCINT);
}

/* NMI HW_TF hacking involves four symbols:
 *
 * [__nmi_pop_ok_start, __nmi_pop_ok_end) mark the beginning and end of the
 * code for an nmi popping routine that will actually pop at the end.
 *
 * [__nmi_pop_fail_start, __nmi_pop_fail_end) mark the beginning and end of the
 * shadow code for an nmi popping routine that will fail at the end.
 *
 * If we see a TF in the OK section, we'll move it to the FAIL section.  If it's
 * already in the FAIL section, we'll report that as a success. */
extern char __nmi_pop_ok_start[], __nmi_pop_ok_end[];
extern char __nmi_pop_fail_start[], __nmi_pop_fail_end[];

static bool nmi_hw_tf_needs_hacked(struct hw_trapframe *hw_tf)
{
	return ((uintptr_t)__nmi_pop_ok_start <= hw_tf->tf_rip) &&
	       (hw_tf->tf_rip < (uintptr_t)__nmi_pop_ok_end);
}

static bool nmi_hw_tf_was_hacked(struct hw_trapframe *hw_tf)
{
	return ((uintptr_t)__nmi_pop_fail_start <= hw_tf->tf_rip) &&
	       (hw_tf->tf_rip < (uintptr_t)__nmi_pop_fail_end);
}

/* Helper.  Hacks the TF if it was in the OK section so that it is at the same
 * spot in the FAIL section.  Returns TRUE if the TF is hacked, meaning the NMI
 * handler can just return. */
static bool nmi_check_and_hack_tf(struct hw_trapframe *hw_tf)
{
	uintptr_t offset;

	if (!nmi_hw_tf_needs_hacked(hw_tf))
		return FALSE;
	if (nmi_hw_tf_was_hacked(hw_tf))
		return TRUE;
	offset = hw_tf->tf_rip - (uintptr_t)__nmi_pop_ok_start;
	hw_tf->tf_rip = (uintptr_t)__nmi_pop_fail_start + offset;
	return TRUE;
}

/* Bottom half of the NMI handler.  This can be interrupted under some
 * circumstances by NMIs.  It exits by popping the hw_tf in assembly. */
void noinline __attribute__((noreturn))
__nmi_bottom_half(struct hw_trapframe *hw_tf)
{
	struct per_cpu_info *pcpui = &per_cpu_info[core_id()];

	while (1) {
		/* Signal that we're doing work.  A concurrent NMI will set this
		 * to NMI_HANDLE_ANOTHER if we should continue, which we'll
		 * catch later. */
		pcpui->nmi_status = NMI_IN_PROGRESS;
		do_nmi_work(hw_tf);
		/* We need to check nmi_status to see if it is
		 * NMI_HANDLE_ANOTHER (if so, run again), write NMI_NORMAL_OPN,
		 * leave this stack, and return to the original context.  We
		 * need to do that in such a manner that an NMI can come in at
		 * any time.  There are two concerns.
		 *
		 * First, we need to not "miss the signal" telling us to re-run
		 * the NMI handler.  To do that, we'll do the actual checking in
		 * asm.  Being in the asm code block is a signal to the real NMI
		 * handler that we need to abort and do_nmi_work() again.
		 *
		 * Second, we need to atomically leave the stack and return.  By
		 * being in asm, the NMI handler knows to just hack our PC to
		 * make us return, instead of starting up a fresh
		 * __nmi_bottom_half().
		 *
		 * The NMI handler works together with the following function
		 * such that if that race occurs while we're in the function,
		 * it'll fail and return.  Then we'll just do_nmi_work() and try
		 * again. */
		extern void nmi_try_to_pop(struct hw_trapframe *tf, int *status,
		                           int old_val, int new_val);

		nmi_try_to_pop(hw_tf, &pcpui->nmi_status, NMI_IN_PROGRESS,
		               NMI_NORMAL_OPN);
		/* Either we returned on our own, since we lost a race with
		 * nmi_status and didn't write (status = ANOTHER), or we won the
		 * race, but an NMI handler set the status to ANOTHER and
		 * restarted us. */
		assert(pcpui->nmi_status != NMI_NORMAL_OPN);
	}
}

/* Separate handler from traps, since there's too many rules for NMI ctx.
 *
 * The general rule is that any writes from NMI context must be very careful.
 * When talking about reads and writes to per-core data:
 * - If NMIs write things written by normal kernel contexts, including IRQs and
 *   traps with IRQs disabled, then you must use atomics on both sides.
 * - If NMIs write things read by normal contexts, then readers must be careful,
 *   since the data can change at will.
 * - If NMIs read things written by normal contexts, don't worry: you're running
 *   uninterrupted (given x86 NMI caveats).
 * - We cannot block.  The current kthread thinks its stacktop is different than
 *   the one we're on.  Just get in and get out.
 * - If we interrupted a user TF, then we don't need to worry any more than for
 *   normal traps/IRQs.
 * - However, we cannot call proc_restartcore.  That could trigger all sorts of
 *   things, like kthreads blocking.
 * - Parallel accesses (from other cores) are the same as always.  You just
 *   can't lock easily.
 *
 * Normally, once you're in NMI, other NMIs are blocked until we return.
 * However, if our NMI handler faults (PF, GPF, breakpoint) due to something
 * like tracing, the iret from that fault will cancel our NMI protections.  Thus
 * we need another layer of code to make sure we don't run the NMI handler
 * concurrently on the same core.  See https://lwn.net/Articles/484932/ for more
 * info.
 *
 * We'll get around the problem by running on yet another NMI stack.  All NMIs
 * come in on the nmi entry stack (tss->ist1).  While we're on that stack, we
 * will not be interrupted.  We jump to another stack to do_nmi_work.  That code
 * can be interrupted, but we are careful to only have one 'thread' running on
 * that stack at a time.  We do this by carefully hopping off the stack in
 * assembly, similar to popping user TFs. */
void handle_nmi(struct hw_trapframe *hw_tf)
{
	struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
	struct hw_trapframe *hw_tf_copy;
	uintptr_t worker_stacktop;

	/* At this point, we're an NMI and other NMIs are blocked.  Only once we
	 * hop to the bottom half could that be no longer true.  NMI with NMIs
	 * fully blocked will run without interruption.  For that reason, we
	 * don't have to be careful about any memory accesses or compiler
	 * tricks. */
	if (pcpui->nmi_status == NMI_HANDLE_ANOTHER)
		return;
	if (pcpui->nmi_status == NMI_IN_PROGRESS) {
		/* Force the handler to run again.  We don't need to worry about
		 * concurrent access here.  We're running, they are not.  We
		 * cannot 'PAUSE' since NMIs are fully blocked.
		 *
		 * The asm routine, for its part, does a compare-and-swap, so if
		 * we happened to interrupt it before it wrote NMI_NORMAL_OPN,
		 * it'll notice, abort, and not write the status. */
		pcpui->nmi_status = NMI_HANDLE_ANOTHER;
		return;
	}
	assert(pcpui->nmi_status == NMI_NORMAL_OPN);
	pcpui->nmi_status = NMI_HANDLE_ANOTHER;
	/* We could be interrupting an NMI that is trying to pop back to a
	 * normal context.  We can tell by looking at its PC.  If it is within
	 * the popping routine, then we interrupted it at this bad time.  We'll
	 * hack the TF such that it will return instead of succeeding. */
	if (nmi_check_and_hack_tf(hw_tf))
		return;
	/* OK, so we didn't interrupt an NMI that was trying to return.  So we
	 * need to run the bottom half.  We're going to jump stacks, but we also
	 * need to copy the hw_tf.  The existing one will be clobbered by any
	 * interrupting NMIs.
	 *
	 * We also need to save some space on the top of that stack for a
	 * pointer to pcpui and a scratch register, which nmi_try_to_pop() will
	 * use.  The target stack will look like this:
	 *
	 *           +--------------------------+ Page boundary (e.g. 0x6000)
	 *           |   scratch space (rsp)    |
	 *           |       pcpui pointer      |
	 *           |      tf_ss + padding     | HW_TF end
	 *           |          tf_rsp          |
	 *           |            .             |
	 *           |            .             |
	 * RSP ->    |         tf_gsbase        | HW_TF start, hw_tf_copy
	 *           +--------------------------+
	 *           |            .             |
	 *           |            .             |
	 *           |            .             |
	 *           +--------------------------+ Page boundary (e.g. 0x5000)
	 *
	 * __nmi_bottom_half() just picks up using the stack below tf_gsbase.
	 * It'll push as needed, growing down.  Basically we're just using the
	 * space 'above' the stack as storage. */
	worker_stacktop = pcpui->nmi_worker_stacktop - 2 * sizeof(uintptr_t);
	*(uintptr_t*)worker_stacktop = (uintptr_t)pcpui;
	worker_stacktop = worker_stacktop - sizeof(struct hw_trapframe);
	hw_tf_copy = (struct hw_trapframe*)worker_stacktop;
	*hw_tf_copy = *hw_tf;
	/* Once we head to the bottom half, consider ourselves interruptible
	 * (though it's not until the first time we do_nmi_work()).  We'll never
	 * come back to this stack.  Doing this in asm so we can easily pass an
	 * argument.  We don't need to call (vs jmp), but it helps keep the
	 * stack aligned. */
	asm volatile("mov $0x0, %%rbp;"
	             "mov %0, %%rsp;"
	             "call __nmi_bottom_half;"
	             : : "r"(worker_stacktop), "D"(hw_tf_copy));
	assert(0);
}

void handle_double_fault(struct hw_trapframe *hw_tf)
{
	panic_hwtf(hw_tf, "Double fault!  Check the kernel stack pointer; you likely ran off the end of the stack.");
}

/* Certain traps want IRQs enabled, such as the syscall.  Others can't handle
 * it, like the page fault handler.  Turn them on on a case-by-case basis. */
static void trap_dispatch(struct hw_trapframe *hw_tf)
{
	struct per_cpu_info *pcpui;
	bool handled = FALSE;
	unsigned long aux = 0;
	uintptr_t fixup_ip;

	// Handle processor exceptions.
	switch(hw_tf->tf_trapno) {
	case T_BRKPT:
		if (!in_kernel(hw_tf))
			backtrace_user_ctx(current, current_ctx);
		else
			monitor(hw_tf);
		handled = TRUE;
		break;
	case T_ILLOP:
	{
		/* TODO: this can PF if there is a concurrent unmap/PM removal.
		 * */
		uintptr_t ip = get_hwtf_pc(hw_tf);

		pcpui = &per_cpu_info[core_id()];
		pcpui->__lock_checking_enabled--; /* for print debugging */
		/* We will muck with the actual TF.  If we're dealing with
		 * userspace, we need to make sure we edit the actual TF that
		 * will get restarted (pcpui), and not the TF on the kstack
		 * (which aren't the same).  See set_current_ctx() for more
		 * info. */
		if (!in_kernel(hw_tf))
			hw_tf = &pcpui->cur_ctx->tf.hw_tf;
		printd("bad opcode, eip: %p, next 3 bytes: %x %x %x\n", ip,
		       *(uint8_t*)(ip + 0),
		       *(uint8_t*)(ip + 1),
		       *(uint8_t*)(ip + 2));
		/* rdtscp: 0f 01 f9 */
		if (*(uint8_t*)(ip + 0) == 0x0f,
		    *(uint8_t*)(ip + 1) == 0x01,
		    *(uint8_t*)(ip + 2) == 0xf9) {
			x86_fake_rdtscp(hw_tf);
			handled = TRUE;
		}
		pcpui->__lock_checking_enabled++; /* for print debugging */
		break;
	}
	case T_PGFLT:
		handled = __handle_page_fault(hw_tf, &aux);
		break;
	case T_GPFLT:
	case T_FPERR:
		handled = try_handle_exception_fixup(hw_tf);
		break;
	case T_SYSCALL:
		enable_irq();
		// check for userspace, for now
		assert(hw_tf->tf_cs != GD_KT);
		/* Set up and run the async calls */
		/* TODO: this is using the wrong reg1 for traps for 32 bit */
		prep_syscalls(current,
			      (struct syscall*)x86_get_systrap_arg0(hw_tf),
			      (unsigned int)x86_get_systrap_arg1(hw_tf));
		disable_irq();
		handled = TRUE;
		break;
	}

	if (!handled) {
		if (in_kernel(hw_tf))
			panic_hwtf(hw_tf,
				   "Damn Damn!  Unhandled trap in the kernel!");
		reflect_unhandled_trap(hw_tf->tf_trapno, hw_tf->tf_err, aux);
	}
}

/* Helper.  For now, this copies out the TF to pcpui.  Eventually, we should
 * consider doing this in trapentry.S
 *
 * TODO: consider having this return the tf used, so we can set tf in trap and
 * irq handlers to edit the TF that will get restarted.  Right now, the kernel
 * uses and restarts tf, but userspace restarts the old pcpui tf.  It is
 * tempting to do this, but note that tf stays on the stack of the kthread,
 * while pcpui->cur_ctx is for the core we trapped in on.  Meaning if we ever
 * block, suddenly cur_ctx is pointing to some old clobbered state that was
 * already returned to and can't be trusted.  Meanwhile tf can always be trusted
 * (like with an in_kernel() check).  The only types of traps from the user that
 * can be expected to have editable trapframes are ones that don't block. */
static void set_current_ctx_hw(struct per_cpu_info *pcpui,
                               struct hw_trapframe *hw_tf)
{
	assert(!irq_is_enabled());
	pcpui->actual_ctx.type = ROS_HW_CTX;
	pcpui->actual_ctx.tf.hw_tf = *hw_tf;
	pcpui->cur_ctx = &pcpui->actual_ctx;
}

static void set_current_ctx_sw(struct per_cpu_info *pcpui,
                               struct sw_trapframe *sw_tf)
{
	assert(!irq_is_enabled());
	pcpui->actual_ctx.type = ROS_SW_CTX;
	pcpui->actual_ctx.tf.sw_tf = *sw_tf;
	pcpui->cur_ctx = &pcpui->actual_ctx;
}

static void set_current_ctx_vm(struct per_cpu_info *pcpui,
                               struct vm_trapframe *vm_tf)
{
	assert(!irq_is_enabled());
	pcpui->actual_ctx.type = ROS_VM_CTX;
	pcpui->actual_ctx.tf.vm_tf = *vm_tf;
	pcpui->cur_ctx = &pcpui->actual_ctx;
}

void trap(struct hw_trapframe *hw_tf)
{
	struct per_cpu_info *pcpui = &per_cpu_info[core_id()];

	/* Copy out the TF for now */
	if (!in_kernel(hw_tf)) {
		set_current_ctx_hw(pcpui, hw_tf);
		/* ignoring state for nested kernel traps.  should be rare. */
		__set_cpu_state(pcpui, CPU_STATE_KERNEL);
	} else {
		inc_ktrap_depth(pcpui);
	}
	printd("Incoming TRAP %d on core %d, TF at %p\n", hw_tf->tf_trapno,
	       core_id(), hw_tf);
	if ((hw_tf->tf_cs & ~3) != GD_UT && (hw_tf->tf_cs & ~3) != GD_KT)
		panic_hwtf(hw_tf, "Trapframe with invalid CS!");
	trap_dispatch(hw_tf);
	/* Return to the current process, which should be runnable.  If we're
	 * the kernel, we should just return naturally.  Note that current and
	 * tf need to still be okay (might not be after blocking) */
	if (in_kernel(hw_tf)) {
		dec_ktrap_depth(pcpui);
		return;
	}
	proc_restartcore();
	assert(0);
}

static bool vector_is_irq(int apic_vec)
{
	/* arguably, we could limit them to MaxIdtIOAPIC */
	return (IdtPIC <= apic_vec) && (apic_vec <= IdtMAX);
}

static void irq_dispatch(struct hw_trapframe *hw_tf)
{
	struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
	struct irq_handler *irq_h;

	if (!in_irq_ctx(pcpui))
		__set_cpu_state(pcpui, CPU_STATE_IRQ);
	inc_irq_depth(pcpui);
	//if (core_id())
	if (hw_tf->tf_trapno != IdtLAPIC_TIMER)	/* timer irq */
	if (hw_tf->tf_trapno != I_KERNEL_MSG)
	if (hw_tf->tf_trapno != 65)	/* qemu serial tends to get this one */
		printd("Incoming IRQ, ISR: %d on core %d\n", hw_tf->tf_trapno,
		       core_id());
	/* TODO: RCU read lock */
	irq_h = irq_handlers[hw_tf->tf_trapno];
	if (!irq_h) {
		warn_once("Received IRQ %d, had no handler registered!",
		          hw_tf->tf_trapno);
		/* If we don't have an IRQ handler, we don't know how to EOI.
		 * Odds are, it's a LAPIC IRQ, such as I_TESTING */
		if (!lapic_check_spurious(hw_tf->tf_trapno))
			lapic_send_eoi(hw_tf->tf_trapno);
		goto out_no_eoi;
	}
	if (irq_h->check_spurious(hw_tf->tf_trapno))
		goto out_no_eoi;
	/* Can now be interrupted/nested by higher priority IRQs, but not by our
	 * current IRQ vector, til we EOI. */
	enable_irq();
	while (irq_h) {
		irq_h->isr(hw_tf, irq_h->data);
		irq_h = irq_h->next;
	}
	// if we're a general purpose IPI function call, down the cpu_list
	extern handler_wrapper_t handler_wrappers[NUM_HANDLER_WRAPPERS];
	if ((I_SMP_CALL0 <= hw_tf->tf_trapno) &&
	    (hw_tf->tf_trapno <= I_SMP_CALL_LAST))
		down_checklist(handler_wrappers[hw_tf->tf_trapno & 0x0f]
			       .cpu_list);
	disable_irq();
	/* Keep in sync with ipi_is_pending */
	irq_handlers[hw_tf->tf_trapno]->eoi(hw_tf->tf_trapno);
	/* Fall-through */
out_no_eoi:
	dec_irq_depth(pcpui);
	if (!in_irq_ctx(pcpui))
		__set_cpu_state(pcpui, CPU_STATE_KERNEL);
}

/* Note IRQs are disabled unless explicitly turned on.
 *
 * In general, we should only get trapno's >= PIC1_OFFSET (32).  Anything else
 * should be a trap.  Even if we don't use the PIC, that should be the standard.
 * It is possible to get a spurious LAPIC IRQ with vector 15 (or similar), but
 * the spurious check should catch that.
 *
 * Note that from hardware's perspective (PIC, etc), IRQs start from 0, but they
 * are all mapped up at PIC1_OFFSET for the cpu / irq_handler. */
void handle_irq(struct hw_trapframe *hw_tf)
{
	struct per_cpu_info *pcpui = &per_cpu_info[core_id()];

	/* Copy out the TF for now */
	if (!in_kernel(hw_tf))
		set_current_ctx_hw(pcpui, hw_tf);
	irq_dispatch(hw_tf);
	/* Return to the current process, which should be runnable.  If we're
	 * the kernel, we should just return naturally.  Note that current and
	 * tf need to still be okay (might not be after blocking) */
	if (in_kernel(hw_tf))
		return;
	proc_restartcore();
	assert(0);
}

/* The irq field may be ignored based on the type of Bus. */
int register_irq(int irq, isr_t handler, void *irq_arg, uint32_t tbdf)
{
	struct irq_handler *irq_h;
	int vector;

	irq_h = kzmalloc(sizeof(struct irq_handler), 0);
	assert(irq_h);
	irq_h->dev_irq = irq;
	irq_h->tbdf = tbdf;
	vector = bus_irq_setup(irq_h);
	if (vector == -1) {
		kfree(irq_h);
		return -1;
	}
	printk("IRQ %d, vector %d (0x%x), type %s\n", irq, vector, vector,
	       irq_h->type);
	assert(irq_h->check_spurious && irq_h->eoi);
	irq_h->isr = handler;
	irq_h->data = irq_arg;
	irq_h->apic_vector = vector;
	/* RCU write lock */
	spin_lock_irqsave(&irq_handler_wlock);
	irq_h->next = irq_handlers[vector];
	wmb();	/* make sure irq_h is done before publishing to readers */
	irq_handlers[vector] = irq_h;
	spin_unlock_irqsave(&irq_handler_wlock);
	/* Most IRQs other than the BusIPI should need their irq unmasked.
	 * Might need to pass the irq_h, in case unmask needs more info.
	 * The lapic IRQs need to be unmasked on a per-core basis */
	if (irq_h->unmask && strcmp(irq_h->type, "lapic"))
		irq_h->unmask(irq_h, vector);
	return 0;
}

/* These routing functions only allow the routing of an irq to a single core.
 * If we want to route to multiple cores, we'll probably need to set up logical
 * groups or something and take some additional parameters. */
static int route_irq_h(struct irq_handler *irq_h, int os_coreid)
{
	int hw_coreid;
	if (!irq_h->route_irq) {
		printk("[kernel] apic_vec %d, type %s cannot be routed\n",
		       irq_h->apic_vector, irq_h->type);
		return -1;
	}
	if (os_coreid >= MAX_NUM_CORES) {
		printk("[kernel] os_coreid %d out of range!\n", os_coreid);
		return -1;
	}
	hw_coreid = get_hw_coreid(os_coreid);
	if (hw_coreid == -1) {
		printk("[kernel] os_coreid %d not a valid hw core!\n",
		       os_coreid);
		return -1;
	}
	irq_h->route_irq(irq_h, irq_h->apic_vector, hw_coreid);
	return 0;
}

/* Routes all irqs for a given apic_vector to os_coreid.  Returns 0 if all of
 * them succeeded.  -1 if there were none or if any of them failed.  We don't
 * share IRQs often (if ever anymore), so this shouldn't be an issue. */
int route_irqs(int apic_vec, int os_coreid)
{
	struct irq_handler *irq_h;
	int ret = -1;

	if (!vector_is_irq(apic_vec)) {
		printk("[kernel] vector %d is not an IRQ vector!\n", apic_vec);
		return -1;
	}
	irq_h = irq_handlers[apic_vec];
	while (irq_h) {
		assert(irq_h->apic_vector == apic_vec);
		ret = route_irq_h(irq_h, os_coreid);
		irq_h = irq_h->next;
	}
	return ret;
}

/* It's a moderate pain in the ass to put these in bit-specific files (header
 * hell with the set_current_ helpers) */
void sysenter_callwrapper(struct syscall *sysc, unsigned long count,
                          struct sw_trapframe *sw_tf)
{
	struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
	set_current_ctx_sw(pcpui, sw_tf);
	__set_cpu_state(pcpui, CPU_STATE_KERNEL);
	/* Once we've set_current_ctx, we can enable interrupts.  This used to
	 * be mandatory (we had immediate KMSGs that would muck with cur_ctx).
	 * Now it should only help for sanity/debugging. */
	enable_irq();
	/* Set up and run the async calls.  This may block, and we could migrate
	 * to another core.  If you use pcpui again, you need to reread it. */
	prep_syscalls(current, sysc, count);
	disable_irq();
	proc_restartcore();
}

/* Declared in x86/arch.h */
void send_ipi(uint32_t os_coreid, uint8_t vector)
{
	int hw_coreid = get_hw_coreid(os_coreid);

	if (hw_coreid == -1) {
		panic("Unmapped OS coreid (OS %d)!\n", os_coreid);
		return;
	}
	assert(vector != T_NMI);
	__send_ipi(hw_coreid, vector);
}

/****************** VM exit handling ******************/

static bool handle_vmexit_cpuid(struct vm_trapframe *tf)
{
	uint32_t eax, ebx, ecx, edx;
	const uint32_t *sigptr;
	const char kvm_sig[] = "KVMKVMKVM\0\0\0";
	const char akaros_sig[] = "AKAROSINSIDE";

	if (tf->tf_rax == 0x0B)
		return FALSE;	// Handle in userspace.

	cpuid(tf->tf_rax, tf->tf_rcx, &eax, &ebx, &ecx, &edx);
	switch (tf->tf_rax) {
	/* TODO: If we can move this to userspace, vmrunkernel can make GPCS on
	 * the fly. */
	case 0x01:
		/* Set the hypervisor bit to let the guest know it is
		 * virtualized */
		ecx |= 1 << 31;
		/* Unset the monitor capability bit so that the guest does not
		 * try to use monitor/mwait. */
		ecx &= ~(1 << 3);
		/* Unset the vmx capability bit so that the guest does not try
		 * to turn it on. */
		ecx &= ~(1 << 5);
		/* Unset the perf capability bit so that the guest does not try
		 * to turn it on. */
		ecx &= ~(1 << 15);

		/* Set the guest pcore id into the apic ID field in CPUID. */
		ebx &= 0x0000ffff;
		ebx |= (current->vmm.nr_guest_pcores & 0xff) << 16;
		ebx |= (tf->tf_guest_pcoreid & 0xff) << 24;
		break;
	case 0x0A:
		eax = 0;
		ebx = 0;
		ecx = 0;
		edx = 0;
		break;
	/* Signal the use of KVM. */
	case 0x40000000:
		sigptr = (const uint32_t *)kvm_sig;
		eax = 0;
		ebx = sigptr[0];
		ecx = sigptr[1];
		edx = sigptr[2];
		break;
	/* Hypervisor Features. */
	case 0x40000003:
		/* Unset the monitor capability bit so that the guest does not
		 * try to use monitor/mwait. */
		edx &= ~(1 << 0);
		break;
	/* Signal the use of AKAROS. */
	case 0x40000100:
		sigptr = (const uint32_t *)akaros_sig;
		eax = 0;
		ebx = sigptr[0];
		ecx = sigptr[1];
		edx = sigptr[2];
		break;
	/* Hypervisor Features. */
	case 0x40000103:
		/* Unset the monitor capability bit so that the guest does not
		 * try to use monitor/mwait. */
		edx &= ~(1 << 0);
		break;
	default:
		break;
	}
	tf->tf_rax = eax;
	tf->tf_rbx = ebx;
	tf->tf_rcx = ecx;
	tf->tf_rdx = edx;
	tf->tf_rip += 2;
	return TRUE;
}

static bool handle_vmexit_ept_fault(struct vm_trapframe *tf)
{
	int prot = 0;
	int ret;

	prot |= tf->tf_exit_qual & VMX_EPT_FAULT_READ ? PROT_READ : 0;
	prot |= tf->tf_exit_qual & VMX_EPT_FAULT_WRITE ? PROT_WRITE : 0;
	prot |= tf->tf_exit_qual & VMX_EPT_FAULT_INS ? PROT_EXEC : 0;
	ret = handle_page_fault(current, tf->tf_guest_pa, prot);
	if (ret == 0)
		return TRUE;

	//Mirror behavior in uthreads, tell userspace to try again.
	if (ret == -EAGAIN)
		tf->tf_flags |= VMCTX_FL_EPT_VMR_BACKED;

	return FALSE;
}

/* Regarding NMI blocking,
 * 	"An NMI causes subsequent NMIs to be blocked, but only after the VM exit
 * 	completes." (SDM)
 *
 * Like handle_nmi(), this function and anything it calls directly cannot fault,
 * or else we lose our NMI protections. */
static bool handle_vmexit_nmi(struct vm_trapframe *tf)
{
	/* Sanity checks, make sure we really got an NMI.  Feel free to remove.
	 */
	assert((tf->tf_intrinfo2 & INTR_INFO_INTR_TYPE_MASK)
	       == INTR_TYPE_NMI_INTR);
	assert((tf->tf_intrinfo2 & INTR_INFO_VECTOR_MASK) == T_NMI);
	assert(!irq_is_enabled());

	emit_monitor_backtrace(ROS_VM_CTX, tf);
	perfmon_snapshot_vmtf(tf);
	send_self_ipi(IdtLAPIC_PCINT);
	return TRUE;
}

bool handle_vmexit_msr(struct vm_trapframe *tf)
{
	bool ret;

	ret = vmm_emulate_msr(tf, (tf->tf_exit_reason == EXIT_REASON_MSR_READ
				   ? VMM_MSR_EMU_READ : VMM_MSR_EMU_WRITE));
	if (ret)
		tf->tf_rip += 2;
	return ret;
}

bool handle_vmexit_extirq(struct vm_trapframe *tf)
{
	struct hw_trapframe hw_tf;
	uint32_t trap_nr;

	/* For now, we just handle external IRQs.  I think guest traps should go
	 * to the guest, based on our vmctls */
	assert((tf->tf_intrinfo2 & INTR_INFO_INTR_TYPE_MASK)
	       == INTR_TYPE_EXT_INTR);
	/* The POKE_HANDLER doesn't run for an ExtINT that triggers a vmexit */
	trap_nr = tf->tf_intrinfo2 & INTR_INFO_VECTOR_MASK;
	if (trap_nr == I_POKE_CORE) {
		lapic_send_eoi(trap_nr);
		return TRUE;
	}
	/* TODO: Our IRQ handlers all expect TFs.  Let's fake one.  A bunch of
	 * handlers (e.g. backtrace/perf) will probably be unhappy about a user
	 * TF that is really a VM, so this all needs work. */
	hw_tf.tf_gsbase = 0;
	hw_tf.tf_fsbase = 0;
	hw_tf.tf_rax = tf->tf_rax;
	hw_tf.tf_rbx = tf->tf_rbx;
	hw_tf.tf_rcx = tf->tf_rcx;
	hw_tf.tf_rdx = tf->tf_rdx;
	hw_tf.tf_rbp = tf->tf_rbp;
	hw_tf.tf_rsi = tf->tf_rsi;
	hw_tf.tf_rdi = tf->tf_rdi;
	hw_tf.tf_r8 = tf->tf_r8;
	hw_tf.tf_r9 = tf->tf_r9;
	hw_tf.tf_r10 = tf->tf_r10;
	hw_tf.tf_r11 = tf->tf_r11;
	hw_tf.tf_r12 = tf->tf_r12;
	hw_tf.tf_r13 = tf->tf_r13;
	hw_tf.tf_r14 = tf->tf_r14;
	hw_tf.tf_r15 = tf->tf_r15;
	hw_tf.tf_trapno = trap_nr;
	hw_tf.tf_err = 0;
	hw_tf.tf_rip = tf->tf_rip;
	hw_tf.tf_cs = GD_UT;	/* faking a user TF, even though it's a VM */
	hw_tf.tf_rflags = tf->tf_rflags;
	hw_tf.tf_rsp = tf->tf_rsp;
	hw_tf.tf_ss = GD_UD;

	irq_dispatch(&hw_tf);
	/* Consider returning whether or not there was a handler registered */
	return TRUE;
}

static bool handle_vmexit_xsetbv(struct vm_trapframe *tf)
{
	// The VM's requested-feature bitmap is represented by edx:eax
	uint64_t vm_rfbm = (tf->tf_rdx << 32) | tf->tf_rax;

	// If the VM tries to set xcr0 to a superset
	// of Akaros's default value, kill the VM.

	// Bit in vm_rfbm and x86_default_xcr0: Ok. Requested and allowed.
	// Bit in vm_rfbm but not x86_default_xcr0: Bad! Requested, not allowed.
	// Bit not in vm_rfbm but in x86_default_xcr0: Ok. Not requested.

	// vm_rfbm & (~x86_default_xcr0) is nonzero if any bits
	// are set in vm_rfbm but not x86_default_xcr0

	if (vm_rfbm & (~__proc_global_info.x86_default_xcr0))
		return FALSE;


	// If attempting to use vm_rfbm for xsetbv
	// causes a fault, we reflect to the VMM.
	if (safe_lxcr0(vm_rfbm))
		return FALSE;


	// If no fault, advance the instruction pointer
	// and return TRUE to make the VM resume.
	tf->tf_rip += 3; // XSETBV is a 3-byte instruction
	return TRUE;
}

static void vmexit_dispatch(struct vm_trapframe *tf)
{
	bool handled = FALSE;

	/* Do not block in any of these functions.
	 *
	 * If we block, we'll probably need to finalize the context.  If we do,
	 * then there's a chance the guest pcore can start somewhere else, and
	 * then we can't get the GPC loaded again.  Plus, they could be running
	 * a GPC with an unresolved vmexit.  It's just mess.
	 *
	 * If we want to enable IRQs, we can do so on a case-by-case basis.
	 * Don't do it for external IRQs - the irq_dispatch code will handle it.
	 * */
	switch (tf->tf_exit_reason) {
	case EXIT_REASON_VMCALL:
		if (current->vmm.flags & VMM_CTL_FL_KERN_PRINTC &&
		    tf->tf_rax == VMCALL_PRINTC) {
			printk("%c", tf->tf_rdi);
			tf->tf_rip += 3;
			handled = TRUE;
		}
		break;
	case EXIT_REASON_CPUID:
		handled = handle_vmexit_cpuid(tf);
		break;
	case EXIT_REASON_EPT_VIOLATION:
		handled = handle_vmexit_ept_fault(tf);
		break;
	case EXIT_REASON_EXCEPTION_NMI:
		handled = handle_vmexit_nmi(tf);
		break;
	case EXIT_REASON_MSR_READ:
	case EXIT_REASON_MSR_WRITE:
		handled = handle_vmexit_msr(tf);
		break;
	case EXIT_REASON_EXTERNAL_INTERRUPT:
		handled = handle_vmexit_extirq(tf);
		break;
	case EXIT_REASON_XSETBV:
		handled = handle_vmexit_xsetbv(tf);
		break;
	default:
		printd("Unhandled vmexit: reason 0x%x, exit qual 0x%x\n",
		       tf->tf_exit_reason, tf->tf_exit_qual);
	}
	if (!handled) {
		tf->tf_flags |= VMCTX_FL_HAS_FAULT;
		if (reflect_current_context()) {
			/* VM contexts shouldn't be in vcore context, so this
			 * should be pretty rare (unlike SCPs or VC ctx page
			 * faults). */
			printk("[kernel] Unable to reflect VM Exit\n");
			print_vmtrapframe(tf);
			proc_destroy(current);
		}
	}
}

void handle_vmexit(struct vm_trapframe *tf)
{
	struct per_cpu_info *pcpui = &per_cpu_info[core_id()];

	tf->tf_rip = vmcs_read(GUEST_RIP);
	tf->tf_rflags = vmcs_read(GUEST_RFLAGS);
	tf->tf_rsp = vmcs_read(GUEST_RSP);
	tf->tf_cr2 = rcr2();
	tf->tf_cr3 = vmcs_read(GUEST_CR3);
	tf->tf_guest_pcoreid = pcpui->guest_pcoreid;
	tf->tf_flags |= VMCTX_FL_PARTIAL;
	tf->tf_guest_intr_status = vmcs_read(GUEST_INTR_STATUS);
	tf->tf_exit_reason = vmcs_read(VM_EXIT_REASON);
	tf->tf_exit_qual = vmcs_read(EXIT_QUALIFICATION);
	tf->tf_intrinfo1 = vmcs_read(GUEST_INTERRUPTIBILITY_INFO);
	tf->tf_intrinfo2 = vmcs_read(VM_EXIT_INTR_INFO);
	tf->tf_guest_va = vmcs_read(GUEST_LINEAR_ADDRESS);
	tf->tf_guest_pa = vmcs_read(GUEST_PHYSICAL_ADDRESS);

	set_current_ctx_vm(pcpui, tf);
	__set_cpu_state(pcpui, CPU_STATE_KERNEL);
	tf = &pcpui->cur_ctx->tf.vm_tf;
	vmexit_dispatch(tf);
	/* We're either restarting a partial VM ctx (vmcs was launched, loaded
	 * on the core, etc) or a SW vc ctx for the reflected trap.  Or the proc
	 * is dying and we'll handle a __death KMSG shortly. */
	proc_restartcore();
}

/* Partial contexts for HW and SW TFs have the user's gs in MSR_KERNEL_GS_BASE.
 * The kernel's gs is loaded into gs.  We need to put the kernel's gs into
 * KERNEL_GS_BASE so the core is ready to run another full context, save the
 * user's {GS,FS}_BASE into their TF so it can run on another core, and keep GS
 * loaded with the current GS (the kernel's). */
static void x86_finalize_hwtf(struct hw_trapframe *tf)
{
	tf->tf_gsbase = read_kern_gsbase();
	write_kern_gsbase(read_gsbase());
	tf->tf_fsbase = read_fsbase();
	x86_hwtf_clear_partial(tf);
}

static void x86_finalize_swtf(struct sw_trapframe *tf)
{
	tf->tf_gsbase = read_kern_gsbase();
	write_kern_gsbase(read_gsbase());
	tf->tf_fsbase = read_fsbase();
	x86_swtf_clear_partial(tf);
}

static void x86_finalize_vmtf(struct vm_trapframe *tf)
{
	struct per_cpu_info *pcpui = &per_cpu_info[core_id()];

	x86_vmtf_clear_partial(tf);
	unload_guest_pcore(pcpui->owning_proc, pcpui->guest_pcoreid);
}

/* Makes sure that the user context is fully saved into ctx and not split across
 * the struct and HW, meaning it is not a "partial context".
 *
 * Be careful to zero out any part of the ctx struct not in use, to avoid
 * leaking information from other processes. */
void arch_finalize_ctx(struct user_context *ctx)
{
	if (!arch_ctx_is_partial(ctx))
		return;
	switch (ctx->type) {
	case ROS_HW_CTX:
		x86_finalize_hwtf(&ctx->tf.hw_tf);
		memset((uint8_t*)&ctx->tf + sizeof(struct hw_trapframe), 0,
			   sizeof(ctx->tf) - sizeof(struct hw_trapframe));
		break;
	case ROS_SW_CTX:
		x86_finalize_swtf(&ctx->tf.sw_tf);
		memset((uint8_t*)&ctx->tf + sizeof(struct sw_trapframe), 0,
			   sizeof(ctx->tf) - sizeof(struct sw_trapframe));
		break;
	case ROS_VM_CTX:
		x86_finalize_vmtf(&ctx->tf.vm_tf);
		memset((uint8_t*)&ctx->tf + sizeof(struct vm_trapframe), 0,
			   sizeof(ctx->tf) - sizeof(struct vm_trapframe));
		break;
	}
}