xref: /illumos-gate/usr/src/uts/i86pc/os/mp_startup.c (revision 37d22dc0ee639b47fcbc0f16ce223299317f60d5)

/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License (the "License").
 * You may not use this file except in compliance with the License.
 *
 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
 * or http://www.opensolaris.org/os/licensing.
 * See the License for the specific language governing permissions
 * and limitations under the License.
 *
 * When distributing Covered Code, include this CDDL HEADER in each
 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
 * If applicable, add the following below this CDDL HEADER, with the
 * fields enclosed by brackets "[]" replaced with your own identifying
 * information: Portions Copyright [yyyy] [name of copyright owner]
 *
 * CDDL HEADER END
 */

/*
 * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */

#include <sys/types.h>
#include <sys/thread.h>
#include <sys/cpuvar.h>
#include <sys/t_lock.h>
#include <sys/param.h>
#include <sys/proc.h>
#include <sys/disp.h>
#include <sys/class.h>
#include <sys/cmn_err.h>
#include <sys/debug.h>
#include <sys/asm_linkage.h>
#include <sys/x_call.h>
#include <sys/systm.h>
#include <sys/var.h>
#include <sys/vtrace.h>
#include <vm/hat.h>
#include <vm/as.h>
#include <vm/seg_kmem.h>
#include <vm/seg_kp.h>
#include <sys/segments.h>
#include <sys/kmem.h>
#include <sys/stack.h>
#include <sys/smp_impldefs.h>
#include <sys/x86_archext.h>
#include <sys/machsystm.h>
#include <sys/traptrace.h>
#include <sys/clock.h>
#include <sys/cpc_impl.h>
#include <sys/pg.h>
#include <sys/cmt.h>
#include <sys/dtrace.h>
#include <sys/archsystm.h>
#include <sys/fp.h>
#include <sys/reboot.h>
#include <sys/kdi_machimpl.h>
#include <vm/hat_i86.h>
#include <sys/memnode.h>
#include <sys/pci_cfgspace.h>
#include <sys/mach_mmu.h>
#include <sys/sysmacros.h>
#if defined(__xpv)
#include <sys/hypervisor.h>
#endif
#include <sys/cpu_module.h>

struct cpu	cpus[1];			/* CPU data */
struct cpu	*cpu[NCPU] = {&cpus[0]};	/* pointers to all CPUs */
cpu_core_t	cpu_core[NCPU];			/* cpu_core structures */

/*
 * Useful for disabling MP bring-up on a MP capable system.
 */
int use_mp = 1;

/*
 * to be set by a PSM to indicate what cpus
 * are sitting around on the system.
 */
cpuset_t mp_cpus;

/*
 * This variable is used by the hat layer to decide whether or not
 * critical sections are needed to prevent race conditions.  For sun4m,
 * this variable is set once enough MP initialization has been done in
 * order to allow cross calls.
 */
int flushes_require_xcalls;

cpuset_t cpu_ready_set;		/* initialized in startup() */

static 	void	mp_startup(void);

static void cpu_sep_enable(void);
static void cpu_sep_disable(void);
static void cpu_asysc_enable(void);
static void cpu_asysc_disable(void);

/*
 * Init CPU info - get CPU type info for processor_info system call.
 */
void
init_cpu_info(struct cpu *cp)
{
	processor_info_t *pi = &cp->cpu_type_info;
	char buf[CPU_IDSTRLEN];

	/*
	 * Get clock-frequency property for the CPU.
	 */
	pi->pi_clock = cpu_freq;

	/*
	 * Current frequency in Hz.
	 */
	cp->cpu_curr_clock = cpu_freq_hz;

	/*
	 * Supported frequencies.
	 */
	if (cp->cpu_supp_freqs == NULL) {
		cpu_set_supp_freqs(cp, NULL);
	}

	(void) strcpy(pi->pi_processor_type, "i386");
	if (fpu_exists)
		(void) strcpy(pi->pi_fputypes, "i387 compatible");

	(void) cpuid_getidstr(cp, buf, sizeof (buf));

	cp->cpu_idstr = kmem_alloc(strlen(buf) + 1, KM_SLEEP);
	(void) strcpy(cp->cpu_idstr, buf);

	(void) cpuid_getbrandstr(cp, buf, sizeof (buf));
	cp->cpu_brandstr = kmem_alloc(strlen(buf) + 1, KM_SLEEP);
	(void) strcpy(cp->cpu_brandstr, buf);
}

/*
 * Configure syscall support on this CPU.
 */
/*ARGSUSED*/
void
init_cpu_syscall(struct cpu *cp)
{
	kpreempt_disable();

#if defined(__amd64)
	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC)) {

#if !defined(__lint)
		/*
		 * The syscall instruction imposes a certain ordering on
		 * segment selectors, so we double-check that ordering
		 * here.
		 */
		ASSERT(KDS_SEL == KCS_SEL + 8);
		ASSERT(UDS_SEL == U32CS_SEL + 8);
		ASSERT(UCS_SEL == U32CS_SEL + 16);
#endif
		/*
		 * Turn syscall/sysret extensions on.
		 */
		cpu_asysc_enable();

		/*
		 * Program the magic registers ..
		 */
		wrmsr(MSR_AMD_STAR,
		    ((uint64_t)(U32CS_SEL << 16 | KCS_SEL)) << 32);
		wrmsr(MSR_AMD_LSTAR, (uint64_t)(uintptr_t)sys_syscall);
		wrmsr(MSR_AMD_CSTAR, (uint64_t)(uintptr_t)sys_syscall32);

		/*
		 * This list of flags is masked off the incoming
		 * %rfl when we enter the kernel.
		 */
		wrmsr(MSR_AMD_SFMASK, (uint64_t)(uintptr_t)(PS_IE | PS_T));
	}
#endif

	/*
	 * On 32-bit kernels, we use sysenter/sysexit because it's too
	 * hard to use syscall/sysret, and it is more portable anyway.
	 *
	 * On 64-bit kernels on Nocona machines, the 32-bit syscall
	 * variant isn't available to 32-bit applications, but sysenter is.
	 */
	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP)) {

#if !defined(__lint)
		/*
		 * The sysenter instruction imposes a certain ordering on
		 * segment selectors, so we double-check that ordering
		 * here. See "sysenter" in Intel document 245471-012, "IA-32
		 * Intel Architecture Software Developer's Manual Volume 2:
		 * Instruction Set Reference"
		 */
		ASSERT(KDS_SEL == KCS_SEL + 8);

		ASSERT32(UCS_SEL == ((KCS_SEL + 16) | 3));
		ASSERT32(UDS_SEL == UCS_SEL + 8);

		ASSERT64(U32CS_SEL == ((KCS_SEL + 16) | 3));
		ASSERT64(UDS_SEL == U32CS_SEL + 8);
#endif

		cpu_sep_enable();

		/*
		 * resume() sets this value to the base of the threads stack
		 * via a context handler.
		 */
		wrmsr(MSR_INTC_SEP_ESP, 0);
		wrmsr(MSR_INTC_SEP_EIP, (uint64_t)(uintptr_t)sys_sysenter);
	}

	kpreempt_enable();
}

/*
 * Multiprocessor initialization.
 *
 * Allocate and initialize the cpu structure, TRAPTRACE buffer, and the
 * startup and idle threads for the specified CPU.
 */
struct cpu *
mp_startup_init(int cpun)
{
	struct cpu *cp;
	kthread_id_t tp;
	caddr_t	sp;
	proc_t *procp;
#if !defined(__xpv)
	extern int idle_cpu_prefer_mwait;
	extern void cpu_idle_mwait();
#endif
	extern void idle();
	extern void cpu_idle();

#ifdef TRAPTRACE
	trap_trace_ctl_t *ttc = &trap_trace_ctl[cpun];
#endif

	ASSERT(cpun < NCPU && cpu[cpun] == NULL);

	cp = kmem_zalloc(sizeof (*cp), KM_SLEEP);
#if !defined(__xpv)
	if ((x86_feature & X86_MWAIT) && idle_cpu_prefer_mwait) {
		cp->cpu_m.mcpu_mwait = cpuid_mwait_alloc(CPU);
		cp->cpu_m.mcpu_idle_cpu = cpu_idle_mwait;
	} else
#endif
		cp->cpu_m.mcpu_idle_cpu = cpu_idle;

	procp = curthread->t_procp;

	mutex_enter(&cpu_lock);
	/*
	 * Initialize the dispatcher first.
	 */
	disp_cpu_init(cp);
	mutex_exit(&cpu_lock);

	cpu_vm_data_init(cp);

	/*
	 * Allocate and initialize the startup thread for this CPU.
	 * Interrupt and process switch stacks get allocated later
	 * when the CPU starts running.
	 */
	tp = thread_create(NULL, 0, NULL, NULL, 0, procp,
	    TS_STOPPED, maxclsyspri);

	/*
	 * Set state to TS_ONPROC since this thread will start running
	 * as soon as the CPU comes online.
	 *
	 * All the other fields of the thread structure are setup by
	 * thread_create().
	 */
	THREAD_ONPROC(tp, cp);
	tp->t_preempt = 1;
	tp->t_bound_cpu = cp;
	tp->t_affinitycnt = 1;
	tp->t_cpu = cp;
	tp->t_disp_queue = cp->cpu_disp;

	/*
	 * Setup thread to start in mp_startup.
	 */
	sp = tp->t_stk;
	tp->t_pc = (uintptr_t)mp_startup;
	tp->t_sp = (uintptr_t)(sp - MINFRAME);
#if defined(__amd64)
	tp->t_sp -= STACK_ENTRY_ALIGN;		/* fake a call */
#endif

	cp->cpu_id = cpun;
	cp->cpu_self = cp;
	cp->cpu_thread = tp;
	cp->cpu_lwp = NULL;
	cp->cpu_dispthread = tp;
	cp->cpu_dispatch_pri = DISP_PRIO(tp);

	/*
	 * cpu_base_spl must be set explicitly here to prevent any blocking
	 * operations in mp_startup from causing the spl of the cpu to drop
	 * to 0 (allowing device interrupts before we're ready) in resume().
	 * cpu_base_spl MUST remain at LOCK_LEVEL until the cpu is CPU_READY.
	 * As an extra bit of security on DEBUG kernels, this is enforced with
	 * an assertion in mp_startup() -- before cpu_base_spl is set to its
	 * proper value.
	 */
	cp->cpu_base_spl = ipltospl(LOCK_LEVEL);

	/*
	 * Now, initialize per-CPU idle thread for this CPU.
	 */
	tp = thread_create(NULL, PAGESIZE, idle, NULL, 0, procp, TS_ONPROC, -1);

	cp->cpu_idle_thread = tp;

	tp->t_preempt = 1;
	tp->t_bound_cpu = cp;
	tp->t_affinitycnt = 1;
	tp->t_cpu = cp;
	tp->t_disp_queue = cp->cpu_disp;

	/*
	 * Bootstrap the CPU's PG data
	 */
	pg_cpu_bootstrap(cp);

	/*
	 * Perform CPC initialization on the new CPU.
	 */
	kcpc_hw_init(cp);

	/*
	 * Allocate virtual addresses for cpu_caddr1 and cpu_caddr2
	 * for each CPU.
	 */
	setup_vaddr_for_ppcopy(cp);

	/*
	 * Allocate page for new GDT and initialize from current GDT.
	 */
#if !defined(__lint)
	ASSERT((sizeof (*cp->cpu_gdt) * NGDT) <= PAGESIZE);
#endif
	cp->cpu_gdt = kmem_zalloc(PAGESIZE, KM_SLEEP);
	bcopy(CPU->cpu_gdt, cp->cpu_gdt, (sizeof (*cp->cpu_gdt) * NGDT));

#if defined(__i386)
	/*
	 * setup kernel %gs.
	 */
	set_usegd(&cp->cpu_gdt[GDT_GS], cp, sizeof (struct cpu) -1, SDT_MEMRWA,
	    SEL_KPL, 0, 1);
#endif

	/*
	 * If we have more than one node, each cpu gets a copy of IDT
	 * local to its node. If this is a Pentium box, we use cpu 0's
	 * IDT. cpu 0's IDT has been made read-only to workaround the
	 * cmpxchgl register bug
	 */
	if (system_hardware.hd_nodes && x86_type != X86_TYPE_P5) {
#if !defined(__lint)
		ASSERT((sizeof (*CPU->cpu_idt) * NIDT) <= PAGESIZE);
#endif
		cp->cpu_idt = kmem_zalloc(PAGESIZE, KM_SLEEP);
		bcopy(CPU->cpu_idt, cp->cpu_idt, PAGESIZE);
	} else {
		cp->cpu_idt = CPU->cpu_idt;
	}

	/*
	 * Get interrupt priority data from cpu 0.
	 */
	cp->cpu_pri_data = CPU->cpu_pri_data;

	/*
	 * alloc space for cpuid info
	 */
	cpuid_alloc_space(cp);

	/*
	 * alloc space for ucode_info
	 */
	ucode_alloc_space(cp);

	hat_cpu_online(cp);

#ifdef TRAPTRACE
	/*
	 * If this is a TRAPTRACE kernel, allocate TRAPTRACE buffers
	 */
	ttc->ttc_first = (uintptr_t)kmem_zalloc(trap_trace_bufsize, KM_SLEEP);
	ttc->ttc_next = ttc->ttc_first;
	ttc->ttc_limit = ttc->ttc_first + trap_trace_bufsize;
#endif
	/*
	 * Record that we have another CPU.
	 */
	mutex_enter(&cpu_lock);
	/*
	 * Initialize the interrupt threads for this CPU
	 */
	cpu_intr_alloc(cp, NINTR_THREADS);
	/*
	 * Add CPU to list of available CPUs.  It'll be on the active list
	 * after mp_startup().
	 */
	cpu_add_unit(cp);
	mutex_exit(&cpu_lock);

	return (cp);
}

/*
 * Undo what was done in mp_startup_init
 */
static void
mp_startup_fini(struct cpu *cp, int error)
{
	mutex_enter(&cpu_lock);

	/*
	 * Remove the CPU from the list of available CPUs.
	 */
	cpu_del_unit(cp->cpu_id);

	if (error == ETIMEDOUT) {
		/*
		 * The cpu was started, but never *seemed* to run any
		 * code in the kernel; it's probably off spinning in its
		 * own private world, though with potential references to
		 * our kmem-allocated IDTs and GDTs (for example).
		 *
		 * Worse still, it may actually wake up some time later,
		 * so rather than guess what it might or might not do, we
		 * leave the fundamental data structures intact.
		 */
		cp->cpu_flags = 0;
		mutex_exit(&cpu_lock);
		return;
	}

	/*
	 * At this point, the only threads bound to this CPU should
	 * special per-cpu threads: it's idle thread, it's pause threads,
	 * and it's interrupt threads.  Clean these up.
	 */
	cpu_destroy_bound_threads(cp);
	cp->cpu_idle_thread = NULL;

	/*
	 * Free the interrupt stack.
	 */
	segkp_release(segkp,
	    cp->cpu_intr_stack - (INTR_STACK_SIZE - SA(MINFRAME)));

	mutex_exit(&cpu_lock);

#ifdef TRAPTRACE
	/*
	 * Discard the trap trace buffer
	 */
	{
		trap_trace_ctl_t *ttc = &trap_trace_ctl[cp->cpu_id];

		kmem_free((void *)ttc->ttc_first, trap_trace_bufsize);
		ttc->ttc_first = NULL;
	}
#endif

	hat_cpu_offline(cp);

	cpuid_free_space(cp);

	ucode_free_space(cp);

	if (cp->cpu_idt != CPU->cpu_idt)
		kmem_free(cp->cpu_idt, PAGESIZE);
	cp->cpu_idt = NULL;

	kmem_free(cp->cpu_gdt, PAGESIZE);
	cp->cpu_gdt = NULL;

	teardown_vaddr_for_ppcopy(cp);

	kcpc_hw_fini(cp);

	cp->cpu_dispthread = NULL;
	cp->cpu_thread = NULL;	/* discarded by cpu_destroy_bound_threads() */

	cpu_vm_data_destroy(cp);

	mutex_enter(&cpu_lock);
	disp_cpu_fini(cp);
	mutex_exit(&cpu_lock);

#if !defined(__xpv)
	if (cp->cpu_m.mcpu_mwait != NULL)
		cpuid_mwait_free(cp);
#endif
	kmem_free(cp, sizeof (*cp));
}

/*
 * Apply workarounds for known errata, and warn about those that are absent.
 *
 * System vendors occasionally create configurations which contain different
 * revisions of the CPUs that are almost but not exactly the same.  At the
 * time of writing, this meant that their clock rates were the same, their
 * feature sets were the same, but the required workaround were -not-
 * necessarily the same.  So, this routine is invoked on -every- CPU soon
 * after starting to make sure that the resulting system contains the most
 * pessimal set of workarounds needed to cope with *any* of the CPUs in the
 * system.
 *
 * workaround_errata is invoked early in mlsetup() for CPU 0, and in
 * mp_startup() for all slave CPUs. Slaves process workaround_errata prior
 * to acknowledging their readiness to the master, so this routine will
 * never be executed by multiple CPUs in parallel, thus making updates to
 * global data safe.
 *
 * These workarounds are based on Rev 3.57 of the Revision Guide for
 * AMD Athlon(tm) 64 and AMD Opteron(tm) Processors, August 2005.
 */

#if defined(OPTERON_ERRATUM_88)
int opteron_erratum_88;		/* if non-zero -> at least one cpu has it */
#endif

#if defined(OPTERON_ERRATUM_91)
int opteron_erratum_91;		/* if non-zero -> at least one cpu has it */
#endif

#if defined(OPTERON_ERRATUM_93)
int opteron_erratum_93;		/* if non-zero -> at least one cpu has it */
#endif

#if defined(OPTERON_ERRATUM_95)
int opteron_erratum_95;		/* if non-zero -> at least one cpu has it */
#endif

#if defined(OPTERON_ERRATUM_100)
int opteron_erratum_100;	/* if non-zero -> at least one cpu has it */
#endif

#if defined(OPTERON_ERRATUM_108)
int opteron_erratum_108;	/* if non-zero -> at least one cpu has it */
#endif

#if defined(OPTERON_ERRATUM_109)
int opteron_erratum_109;	/* if non-zero -> at least one cpu has it */
#endif

#if defined(OPTERON_ERRATUM_121)
int opteron_erratum_121;	/* if non-zero -> at least one cpu has it */
#endif

#if defined(OPTERON_ERRATUM_122)
int opteron_erratum_122;	/* if non-zero -> at least one cpu has it */
#endif

#if defined(OPTERON_ERRATUM_123)
int opteron_erratum_123;	/* if non-zero -> at least one cpu has it */
#endif

#if defined(OPTERON_ERRATUM_131)
int opteron_erratum_131;	/* if non-zero -> at least one cpu has it */
#endif

#if defined(OPTERON_WORKAROUND_6336786)
int opteron_workaround_6336786;	/* non-zero -> WA relevant and applied */
int opteron_workaround_6336786_UP = 0;	/* Not needed for UP */
#endif

#if defined(OPTERON_WORKAROUND_6323525)
int opteron_workaround_6323525;	/* if non-zero -> at least one cpu has it */
#endif

#if defined(OPTERON_ERRATUM_298)
int opteron_erratum_298;
#endif

static void
workaround_warning(cpu_t *cp, uint_t erratum)
{
	cmn_err(CE_WARN, "cpu%d: no workaround for erratum %u",
	    cp->cpu_id, erratum);
}

static void
workaround_applied(uint_t erratum)
{
	if (erratum > 1000000)
		cmn_err(CE_CONT, "?workaround applied for cpu issue #%d\n",
		    erratum);
	else
		cmn_err(CE_CONT, "?workaround applied for cpu erratum #%d\n",
		    erratum);
}

static void
msr_warning(cpu_t *cp, const char *rw, uint_t msr, int error)
{
	cmn_err(CE_WARN, "cpu%d: couldn't %smsr 0x%x, error %d",
	    cp->cpu_id, rw, msr, error);
}

/*
 * Determine the number of nodes in a Hammer / Greyhound / Griffin family
 * system.
 */
static uint_t
opteron_get_nnodes(void)
{
	static uint_t nnodes = 0;

	if (nnodes == 0) {
#ifdef	DEBUG
		uint_t family;

		/*
		 * This routine uses a PCI config space based mechanism
		 * for retrieving the number of nodes in the system.
		 * Device 24, function 0, offset 0x60 as used here is not
		 * AMD processor architectural, and may not work on processor
		 * families other than those listed below.
		 *
		 * Callers of this routine must ensure that we're running on
		 * a processor which supports this mechanism.
		 * The assertion below is meant to catch calls on unsupported
		 * processors.
		 */
		family = cpuid_getfamily(CPU);
		ASSERT(family == 0xf || family == 0x10 || family == 0x11);
#endif	/* DEBUG */

		/*
		 * Obtain the number of nodes in the system from
		 * bits [6:4] of the Node ID register on node 0.
		 *
		 * The actual node count is NodeID[6:4] + 1
		 *
		 * The Node ID register is accessed via function 0,
		 * offset 0x60. Node 0 is device 24.
		 */
		nnodes = ((pci_getl_func(0, 24, 0, 0x60) & 0x70) >> 4) + 1;
	}
	return (nnodes);
}

#if defined(__xpv)

/*
 * On dom0, we can determine the number of physical cpus on the machine.
 * This number is important when figuring out what workarounds are
 * appropriate, so compute it now.
 */
uint_t
xen_get_nphyscpus(void)
{
	static uint_t nphyscpus = 0;

	ASSERT(DOMAIN_IS_INITDOMAIN(xen_info));

	if (nphyscpus == 0) {
		xen_sysctl_t op;
		xen_sysctl_physinfo_t *pi = &op.u.physinfo;

		op.cmd = XEN_SYSCTL_physinfo;
		op.interface_version = XEN_SYSCTL_INTERFACE_VERSION;
		if (HYPERVISOR_sysctl(&op) == 0)
			nphyscpus = pi->threads_per_core *
			    pi->cores_per_socket * pi->sockets_per_node *
			    pi->nr_nodes;
	}
	return (nphyscpus);
}
#endif

uint_t
do_erratum_298(struct cpu *cpu)
{
	static int	osvwrc = -3;
	extern int	osvw_opteron_erratum(cpu_t *, uint_t);

	/*
	 * L2 Eviction May Occur During Processor Operation To Set
	 * Accessed or Dirty Bit.
	 */
	if (osvwrc == -3) {
		osvwrc = osvw_opteron_erratum(cpu, 298);
	} else {
		/* osvw return codes should be consistent for all cpus */
		ASSERT(osvwrc == osvw_opteron_erratum(cpu, 298));
	}

	switch (osvwrc) {
	case 0:		/* erratum is not present: do nothing */
		break;
	case 1:		/* erratum is present: BIOS workaround applied */
		/*
		 * check if workaround is actually in place and issue warning
		 * if not.
		 */
		if (((rdmsr(MSR_AMD_HWCR) & AMD_HWCR_TLBCACHEDIS) == 0) ||
		    ((rdmsr(MSR_AMD_BU_CFG) & AMD_BU_CFG_E298) == 0)) {
#if defined(OPTERON_ERRATUM_298)
			opteron_erratum_298++;
#else
			workaround_warning(cpu, 298);
			return (1);
#endif
		}
		break;
	case -1:	/* cannot determine via osvw: check cpuid */
		if ((cpuid_opteron_erratum(cpu, 298) > 0) &&
		    (((rdmsr(MSR_AMD_HWCR) & AMD_HWCR_TLBCACHEDIS) == 0) ||
		    ((rdmsr(MSR_AMD_BU_CFG) & AMD_BU_CFG_E298) == 0))) {
#if defined(OPTERON_ERRATUM_298)
			opteron_erratum_298++;
#else
			workaround_warning(cpu, 298);
			return (1);
#endif
		}
		break;
	}
	return (0);
}

uint_t
workaround_errata(struct cpu *cpu)
{
	uint_t missing = 0;

	ASSERT(cpu == CPU);

	/*LINTED*/
	if (cpuid_opteron_erratum(cpu, 88) > 0) {
		/*
		 * SWAPGS May Fail To Read Correct GS Base
		 */
#if defined(OPTERON_ERRATUM_88)
		/*
		 * The workaround is an mfence in the relevant assembler code
		 */
		opteron_erratum_88++;
#else
		workaround_warning(cpu, 88);
		missing++;
#endif
	}

	if (cpuid_opteron_erratum(cpu, 91) > 0) {
		/*
		 * Software Prefetches May Report A Page Fault
		 */
#if defined(OPTERON_ERRATUM_91)
		/*
		 * fix is in trap.c
		 */
		opteron_erratum_91++;
#else
		workaround_warning(cpu, 91);
		missing++;
#endif
	}

	if (cpuid_opteron_erratum(cpu, 93) > 0) {
		/*
		 * RSM Auto-Halt Restart Returns to Incorrect RIP
		 */
#if defined(OPTERON_ERRATUM_93)
		/*
		 * fix is in trap.c
		 */
		opteron_erratum_93++;
#else
		workaround_warning(cpu, 93);
		missing++;
#endif
	}

	/*LINTED*/
	if (cpuid_opteron_erratum(cpu, 95) > 0) {
		/*
		 * RET Instruction May Return to Incorrect EIP
		 */
#if defined(OPTERON_ERRATUM_95)
#if defined(_LP64)
		/*
		 * Workaround this by ensuring that 32-bit user code and
		 * 64-bit kernel code never occupy the same address
		 * range mod 4G.
		 */
		if (_userlimit32 > 0xc0000000ul)
			*(uintptr_t *)&_userlimit32 = 0xc0000000ul;

		/*LINTED*/
		ASSERT((uint32_t)COREHEAP_BASE == 0xc0000000u);
		opteron_erratum_95++;
#endif	/* _LP64 */
#else
		workaround_warning(cpu, 95);
		missing++;
#endif
	}

	if (cpuid_opteron_erratum(cpu, 100) > 0) {
		/*
		 * Compatibility Mode Branches Transfer to Illegal Address
		 */
#if defined(OPTERON_ERRATUM_100)
		/*
		 * fix is in trap.c
		 */
		opteron_erratum_100++;
#else
		workaround_warning(cpu, 100);
		missing++;
#endif
	}

	/*LINTED*/
	if (cpuid_opteron_erratum(cpu, 108) > 0) {
		/*
		 * CPUID Instruction May Return Incorrect Model Number In
		 * Some Processors
		 */
#if defined(OPTERON_ERRATUM_108)
		/*
		 * (Our cpuid-handling code corrects the model number on
		 * those processors)
		 */
#else
		workaround_warning(cpu, 108);
		missing++;
#endif
	}

	/*LINTED*/
	if (cpuid_opteron_erratum(cpu, 109) > 0) do {
		/*
		 * Certain Reverse REP MOVS May Produce Unpredictable Behaviour
		 */
#if defined(OPTERON_ERRATUM_109)
		/*
		 * The "workaround" is to print a warning to upgrade the BIOS
		 */
		uint64_t value;
		const uint_t msr = MSR_AMD_PATCHLEVEL;
		int err;

		if ((err = checked_rdmsr(msr, &value)) != 0) {
			msr_warning(cpu, "rd", msr, err);
			workaround_warning(cpu, 109);
			missing++;
		}
		if (value == 0)
			opteron_erratum_109++;
#else
		workaround_warning(cpu, 109);
		missing++;
#endif
	/*CONSTANTCONDITION*/
	} while (0);

	/*LINTED*/
	if (cpuid_opteron_erratum(cpu, 121) > 0) {
		/*
		 * Sequential Execution Across Non_Canonical Boundary Caused
		 * Processor Hang
		 */
#if defined(OPTERON_ERRATUM_121)
#if defined(_LP64)
		/*
		 * Erratum 121 is only present in long (64 bit) mode.
		 * Workaround is to include the page immediately before the
		 * va hole to eliminate the possibility of system hangs due to
		 * sequential execution across the va hole boundary.
		 */
		if (opteron_erratum_121)
			opteron_erratum_121++;
		else {
			if (hole_start) {
				hole_start -= PAGESIZE;
			} else {
				/*
				 * hole_start not yet initialized by
				 * mmu_init. Initialize hole_start
				 * with value to be subtracted.
				 */
				hole_start = PAGESIZE;
			}
			opteron_erratum_121++;
		}
#endif	/* _LP64 */
#else
		workaround_warning(cpu, 121);
		missing++;
#endif
	}

	/*LINTED*/
	if (cpuid_opteron_erratum(cpu, 122) > 0) do {
		/*
		 * TLB Flush Filter May Cause Coherency Problem in
		 * Multiprocessor Systems
		 */
#if defined(OPTERON_ERRATUM_122)
		uint64_t value;
		const uint_t msr = MSR_AMD_HWCR;
		int error;

		/*
		 * Erratum 122 is only present in MP configurations (multi-core
		 * or multi-processor).
		 */
#if defined(__xpv)
		if (!DOMAIN_IS_INITDOMAIN(xen_info))
			break;
		if (!opteron_erratum_122 && xen_get_nphyscpus() == 1)
			break;
#else
		if (!opteron_erratum_122 && opteron_get_nnodes() == 1 &&
		    cpuid_get_ncpu_per_chip(cpu) == 1)
			break;
#endif
		/* disable TLB Flush Filter */

		if ((error = checked_rdmsr(msr, &value)) != 0) {
			msr_warning(cpu, "rd", msr, error);
			workaround_warning(cpu, 122);
			missing++;
		} else {
			value |= (uint64_t)AMD_HWCR_FFDIS;
			if ((error = checked_wrmsr(msr, value)) != 0) {
				msr_warning(cpu, "wr", msr, error);
				workaround_warning(cpu, 122);
				missing++;
			}
		}
		opteron_erratum_122++;
#else
		workaround_warning(cpu, 122);
		missing++;
#endif
	/*CONSTANTCONDITION*/
	} while (0);

	/*LINTED*/
	if (cpuid_opteron_erratum(cpu, 123) > 0) do {
		/*
		 * Bypassed Reads May Cause Data Corruption of System Hang in
		 * Dual Core Processors
		 */
#if defined(OPTERON_ERRATUM_123)
		uint64_t value;
		const uint_t msr = MSR_AMD_PATCHLEVEL;
		int err;

		/*
		 * Erratum 123 applies only to multi-core cpus.
		 */
		if (cpuid_get_ncpu_per_chip(cpu) < 2)
			break;
#if defined(__xpv)
		if (!DOMAIN_IS_INITDOMAIN(xen_info))
			break;
#endif
		/*
		 * The "workaround" is to print a warning to upgrade the BIOS
		 */
		if ((err = checked_rdmsr(msr, &value)) != 0) {
			msr_warning(cpu, "rd", msr, err);
			workaround_warning(cpu, 123);
			missing++;
		}
		if (value == 0)
			opteron_erratum_123++;
#else
		workaround_warning(cpu, 123);
		missing++;

#endif
	/*CONSTANTCONDITION*/
	} while (0);

	/*LINTED*/
	if (cpuid_opteron_erratum(cpu, 131) > 0) do {
		/*
		 * Multiprocessor Systems with Four or More Cores May Deadlock
		 * Waiting for a Probe Response
		 */
#if defined(OPTERON_ERRATUM_131)
		uint64_t nbcfg;
		const uint_t msr = MSR_AMD_NB_CFG;
		const uint64_t wabits =
		    AMD_NB_CFG_SRQ_HEARTBEAT | AMD_NB_CFG_SRQ_SPR;
		int error;

		/*
		 * Erratum 131 applies to any system with four or more cores.
		 */
		if (opteron_erratum_131)
			break;
#if defined(__xpv)
		if (!DOMAIN_IS_INITDOMAIN(xen_info))
			break;
		if (xen_get_nphyscpus() < 4)
			break;
#else
		if (opteron_get_nnodes() * cpuid_get_ncpu_per_chip(cpu) < 4)
			break;
#endif
		/*
		 * Print a warning if neither of the workarounds for
		 * erratum 131 is present.
		 */
		if ((error = checked_rdmsr(msr, &nbcfg)) != 0) {
			msr_warning(cpu, "rd", msr, error);
			workaround_warning(cpu, 131);
			missing++;
		} else if ((nbcfg & wabits) == 0) {
			opteron_erratum_131++;
		} else {
			/* cannot have both workarounds set */
			ASSERT((nbcfg & wabits) != wabits);
		}
#else
		workaround_warning(cpu, 131);
		missing++;
#endif
	/*CONSTANTCONDITION*/
	} while (0);

	/*
	 * This isn't really an erratum, but for convenience the
	 * detection/workaround code lives here and in cpuid_opteron_erratum.
	 */
	if (cpuid_opteron_erratum(cpu, 6336786) > 0) {
#if defined(OPTERON_WORKAROUND_6336786)
		/*
		 * Disable C1-Clock ramping on multi-core/multi-processor
		 * K8 platforms to guard against TSC drift.
		 */
		if (opteron_workaround_6336786) {
			opteron_workaround_6336786++;
#if defined(__xpv)
		} else if ((DOMAIN_IS_INITDOMAIN(xen_info) &&
		    xen_get_nphyscpus() > 1) ||
		    opteron_workaround_6336786_UP) {
			/*
			 * XXPV	Hmm.  We can't walk the Northbridges on
			 *	the hypervisor; so just complain and drive
			 *	on.  This probably needs to be fixed in
			 *	the hypervisor itself.
			 */
			opteron_workaround_6336786++;
			workaround_warning(cpu, 6336786);
#else	/* __xpv */
		} else if ((opteron_get_nnodes() *
		    cpuid_get_ncpu_per_chip(cpu) > 1) ||
		    opteron_workaround_6336786_UP) {

			uint_t	node, nnodes;
			uint8_t data;

			nnodes = opteron_get_nnodes();
			for (node = 0; node < nnodes; node++) {
				/*
				 * Clear PMM7[1:0] (function 3, offset 0x87)
				 * Northbridge device is the node id + 24.
				 */
				data = pci_getb_func(0, node + 24, 3, 0x87);
				data &= 0xFC;
				pci_putb_func(0, node + 24, 3, 0x87, data);
			}
			opteron_workaround_6336786++;
#endif	/* __xpv */
		}
#else
		workaround_warning(cpu, 6336786);
		missing++;
#endif
	}

	/*LINTED*/
	/*
	 * Mutex primitives don't work as expected.
	 */
	if (cpuid_opteron_erratum(cpu, 6323525) > 0) {
#if defined(OPTERON_WORKAROUND_6323525)
		/*
		 * This problem only occurs with 2 or more cores. If bit in
		 * MSR_AMD_BU_CFG set, then not applicable. The workaround
		 * is to patch the semaphone routines with the lfence
		 * instruction to provide necessary load memory barrier with
		 * possible subsequent read-modify-write ops.
		 *
		 * It is too early in boot to call the patch routine so
		 * set erratum variable to be done in startup_end().
		 */
		if (opteron_workaround_6323525) {
			opteron_workaround_6323525++;
#if defined(__xpv)
		} else if (x86_feature & X86_SSE2) {
			if (DOMAIN_IS_INITDOMAIN(xen_info)) {
				/*
				 * XXPV	Use dom0_msr here when extended
				 *	operations are supported?
				 */
				if (xen_get_nphyscpus() > 1)
					opteron_workaround_6323525++;
			} else {
				/*
				 * We have no way to tell how many physical
				 * cpus there are, or even if this processor
				 * has the problem, so enable the workaround
				 * unconditionally (at some performance cost).
				 */
				opteron_workaround_6323525++;
			}
#else	/* __xpv */
		} else if ((x86_feature & X86_SSE2) && ((opteron_get_nnodes() *
		    cpuid_get_ncpu_per_chip(cpu)) > 1)) {
			if ((xrdmsr(MSR_AMD_BU_CFG) & (UINT64_C(1) << 33)) == 0)
				opteron_workaround_6323525++;
#endif	/* __xpv */
		}
#else
		workaround_warning(cpu, 6323525);
		missing++;
#endif
	}

	missing += do_erratum_298(cpu);

#ifdef __xpv
	return (0);
#else
	return (missing);
#endif
}

void
workaround_errata_end()
{
#if defined(OPTERON_ERRATUM_88)
	if (opteron_erratum_88)
		workaround_applied(88);
#endif
#if defined(OPTERON_ERRATUM_91)
	if (opteron_erratum_91)
		workaround_applied(91);
#endif
#if defined(OPTERON_ERRATUM_93)
	if (opteron_erratum_93)
		workaround_applied(93);
#endif
#if defined(OPTERON_ERRATUM_95)
	if (opteron_erratum_95)
		workaround_applied(95);
#endif
#if defined(OPTERON_ERRATUM_100)
	if (opteron_erratum_100)
		workaround_applied(100);
#endif
#if defined(OPTERON_ERRATUM_108)
	if (opteron_erratum_108)
		workaround_applied(108);
#endif
#if defined(OPTERON_ERRATUM_109)
	if (opteron_erratum_109) {
		cmn_err(CE_WARN,
		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
		    " processor\nerratum 109 was not detected; updating your"
		    " system's BIOS to a version\ncontaining this"
		    " microcode patch is HIGHLY recommended or erroneous"
		    " system\noperation may occur.\n");
	}
#endif
#if defined(OPTERON_ERRATUM_121)
	if (opteron_erratum_121)
		workaround_applied(121);
#endif
#if defined(OPTERON_ERRATUM_122)
	if (opteron_erratum_122)
		workaround_applied(122);
#endif
#if defined(OPTERON_ERRATUM_123)
	if (opteron_erratum_123) {
		cmn_err(CE_WARN,
		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
		    " processor\nerratum 123 was not detected; updating your"
		    " system's BIOS to a version\ncontaining this"
		    " microcode patch is HIGHLY recommended or erroneous"
		    " system\noperation may occur.\n");
	}
#endif
#if defined(OPTERON_ERRATUM_131)
	if (opteron_erratum_131) {
		cmn_err(CE_WARN,
		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
		    " processor\nerratum 131 was not detected; updating your"
		    " system's BIOS to a version\ncontaining this"
		    " microcode patch is HIGHLY recommended or erroneous"
		    " system\noperation may occur.\n");
	}
#endif
#if defined(OPTERON_WORKAROUND_6336786)
	if (opteron_workaround_6336786)
		workaround_applied(6336786);
#endif
#if defined(OPTERON_WORKAROUND_6323525)
	if (opteron_workaround_6323525)
		workaround_applied(6323525);
#endif
#if defined(OPTERON_ERRATUM_298)
	if (opteron_erratum_298) {
		cmn_err(CE_WARN,
		    "BIOS microcode patch for AMD 64/Opteron(tm)"
		    " processor\nerratum 298 was not detected; updating your"
		    " system's BIOS to a version\ncontaining this"
		    " microcode patch is HIGHLY recommended or erroneous"
		    " system\noperation may occur.\n");
	}
#endif
}

static cpuset_t procset;

/*
 * Start a single cpu, assuming that the kernel context is available
 * to successfully start another cpu.
 *
 * (For example, real mode code is mapped into the right place
 * in memory and is ready to be run.)
 */
int
start_cpu(processorid_t who)
{
	void *ctx;
	cpu_t *cp;
	int delays;
	int error = 0;

	ASSERT(who != 0);

	/*
	 * Check if there's at least a Mbyte of kmem available
	 * before attempting to start the cpu.
	 */
	if (kmem_avail() < 1024 * 1024) {
		/*
		 * Kick off a reap in case that helps us with
		 * later attempts ..
		 */
		kmem_reap();
		return (ENOMEM);
	}

	cp = mp_startup_init(who);
	if ((ctx = mach_cpucontext_alloc(cp)) == NULL ||
	    (error = mach_cpu_start(cp, ctx)) != 0) {

		/*
		 * Something went wrong before we even started it
		 */
		if (ctx)
			cmn_err(CE_WARN,
			    "cpu%d: failed to start error %d",
			    cp->cpu_id, error);
		else
			cmn_err(CE_WARN,
			    "cpu%d: failed to allocate context", cp->cpu_id);

		if (ctx)
			mach_cpucontext_free(cp, ctx, error);
		else
			error = EAGAIN;		/* hmm. */
		mp_startup_fini(cp, error);
		return (error);
	}

	for (delays = 0; !CPU_IN_SET(procset, who); delays++) {
		if (delays == 500) {
			/*
			 * After five seconds, things are probably looking
			 * a bit bleak - explain the hang.
			 */
			cmn_err(CE_NOTE, "cpu%d: started, "
			    "but not running in the kernel yet", who);
		} else if (delays > 2000) {
			/*
			 * We waited at least 20 seconds, bail ..
			 */
			error = ETIMEDOUT;
			cmn_err(CE_WARN, "cpu%d: timed out", who);
			mach_cpucontext_free(cp, ctx, error);
			mp_startup_fini(cp, error);
			return (error);
		}

		/*
		 * wait at least 10ms, then check again..
		 */
		delay(USEC_TO_TICK_ROUNDUP(10000));
	}

	mach_cpucontext_free(cp, ctx, 0);

#ifndef __xpv
	if (tsc_gethrtime_enable)
		tsc_sync_master(who);
#endif

	if (dtrace_cpu_init != NULL) {
		/*
		 * DTrace CPU initialization expects cpu_lock to be held.
		 */
		mutex_enter(&cpu_lock);
		(*dtrace_cpu_init)(who);
		mutex_exit(&cpu_lock);
	}

	while (!CPU_IN_SET(cpu_ready_set, who))
		delay(1);

	return (0);
}


/*ARGSUSED*/
void
start_other_cpus(int cprboot)
{
	uint_t who;
	uint_t skipped = 0;
	uint_t bootcpuid = 0;

	/*
	 * Initialize our own cpu_info.
	 */
	init_cpu_info(CPU);

	cmn_err(CE_CONT, "?cpu%d: %s\n", CPU->cpu_id, CPU->cpu_idstr);
	cmn_err(CE_CONT, "?cpu%d: %s\n", CPU->cpu_id, CPU->cpu_brandstr);

	/*
	 * Initialize our syscall handlers
	 */
	init_cpu_syscall(CPU);

	/*
	 * Take the boot cpu out of the mp_cpus set because we know
	 * it's already running.  Add it to the cpu_ready_set for
	 * precisely the same reason.
	 */
	CPUSET_DEL(mp_cpus, bootcpuid);
	CPUSET_ADD(cpu_ready_set, bootcpuid);

	/*
	 * if only 1 cpu or not using MP, skip the rest of this
	 */
	if (CPUSET_ISNULL(mp_cpus) || use_mp == 0) {
		if (use_mp == 0)
			cmn_err(CE_CONT, "?***** Not in MP mode\n");
		goto done;
	}

	/*
	 * perform such initialization as is needed
	 * to be able to take CPUs on- and off-line.
	 */
	cpu_pause_init();

	xc_init();		/* initialize processor crosscalls */

	if (mach_cpucontext_init() != 0)
		goto done;

	flushes_require_xcalls = 1;

	/*
	 * We lock our affinity to the master CPU to ensure that all slave CPUs
	 * do their TSC syncs with the same CPU.
	 */
	affinity_set(CPU_CURRENT);

	for (who = 0; who < NCPU; who++) {

		if (!CPU_IN_SET(mp_cpus, who))
			continue;
		ASSERT(who != bootcpuid);
		if (ncpus >= max_ncpus) {
			skipped = who;
			continue;
		}
		if (start_cpu(who) != 0)
			CPUSET_DEL(mp_cpus, who);

		mutex_enter(&cpu_lock);
		cpu_state_change_notify(who, CPU_SETUP);
		mutex_exit(&cpu_lock);
	}

	/* Free the space allocated to hold the microcode file */
	ucode_cleanup();

	affinity_clear();

	if (skipped) {
		cmn_err(CE_NOTE,
		    "System detected %d cpus, but "
		    "only %d cpu(s) were enabled during boot.",
		    skipped + 1, ncpus);
		cmn_err(CE_NOTE,
		    "Use \"boot-ncpus\" parameter to enable more CPU(s). "
		    "See eeprom(1M).");
	}

done:
	workaround_errata_end();
	mach_cpucontext_fini();

	cmi_post_mpstartup();
}

/*
 * Dummy functions - no i86pc platforms support dynamic cpu allocation.
 */
/*ARGSUSED*/
int
mp_cpu_configure(int cpuid)
{
	return (ENOTSUP);		/* not supported */
}

/*ARGSUSED*/
int
mp_cpu_unconfigure(int cpuid)
{
	return (ENOTSUP);		/* not supported */
}

/*
 * Startup function for 'other' CPUs (besides boot cpu).
 * Called from real_mode_start.
 *
 * WARNING: until CPU_READY is set, mp_startup and routines called by
 * mp_startup should not call routines (e.g. kmem_free) that could call
 * hat_unload which requires CPU_READY to be set.
 */
void
mp_startup(void)
{
	struct cpu *cp = CPU;
	uint_t new_x86_feature;
#ifndef __xpv
	extern void cpupm_init(cpu_t *);
#endif

	/*
	 * We need to get TSC on this proc synced (i.e., any delta
	 * from cpu0 accounted for) as soon as we can, because many
	 * many things use gethrtime/pc_gethrestime, including
	 * interrupts, cmn_err, etc.
	 */

	/* Let cpu0 continue into tsc_sync_master() */
	CPUSET_ATOMIC_ADD(procset, cp->cpu_id);

#ifndef __xpv
	if (tsc_gethrtime_enable)
		tsc_sync_slave();
#endif

	/*
	 * Once this was done from assembly, but it's safer here; if
	 * it blocks, we need to be able to swtch() to and from, and
	 * since we get here by calling t_pc, we need to do that call
	 * before swtch() overwrites it.
	 */
	(void) (*ap_mlsetup)();

	new_x86_feature = cpuid_pass1(cp);

#ifndef __xpv
	/*
	 * Program this cpu's PAT
	 */
	if (x86_feature & X86_PAT)
		pat_sync();
#endif

	/*
	 * Set up TSC_AUX to contain the cpuid for this processor
	 * for the rdtscp instruction.
	 */
	if (x86_feature & X86_TSCP)
		(void) wrmsr(MSR_AMD_TSCAUX, cp->cpu_id);

	/*
	 * Initialize this CPU's syscall handlers
	 */
	init_cpu_syscall(cp);

	/*
	 * Enable interrupts with spl set to LOCK_LEVEL. LOCK_LEVEL is the
	 * highest level at which a routine is permitted to block on
	 * an adaptive mutex (allows for cpu poke interrupt in case
	 * the cpu is blocked on a mutex and halts). Setting LOCK_LEVEL blocks
	 * device interrupts that may end up in the hat layer issuing cross
	 * calls before CPU_READY is set.
	 */
	splx(ipltospl(LOCK_LEVEL));
	sti();

	/*
	 * Do a sanity check to make sure this new CPU is a sane thing
	 * to add to the collection of processors running this system.
	 *
	 * XXX	Clearly this needs to get more sophisticated, if x86
	 * systems start to get built out of heterogenous CPUs; as is
	 * likely to happen once the number of processors in a configuration
	 * gets large enough.
	 */
	if ((x86_feature & new_x86_feature) != x86_feature) {
		cmn_err(CE_CONT, "?cpu%d: %b\n",
		    cp->cpu_id, new_x86_feature, FMT_X86_FEATURE);
		cmn_err(CE_WARN, "cpu%d feature mismatch", cp->cpu_id);
	}

	/*
	 * We do not support cpus with mixed monitor/mwait support if the
	 * boot cpu supports monitor/mwait.
	 */
	if ((x86_feature & ~new_x86_feature) & X86_MWAIT)
		panic("unsupported mixed cpu monitor/mwait support detected");

	/*
	 * We could be more sophisticated here, and just mark the CPU
	 * as "faulted" but at this point we'll opt for the easier
	 * answer of dieing horribly.  Provided the boot cpu is ok,
	 * the system can be recovered by booting with use_mp set to zero.
	 */
	if (workaround_errata(cp) != 0)
		panic("critical workaround(s) missing for cpu%d", cp->cpu_id);

	cpuid_pass2(cp);
	cpuid_pass3(cp);
	(void) cpuid_pass4(cp);

	init_cpu_info(cp);

	mutex_enter(&cpu_lock);

	cp->cpu_flags |= CPU_RUNNING | CPU_READY | CPU_EXISTS;

	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_idstr);
	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_brandstr);

	if (dtrace_cpu_init != NULL) {
		(*dtrace_cpu_init)(cp->cpu_id);
	}

	/*
	 * Fill out cpu_ucode_info.  Update microcode if necessary.
	 */
	ucode_check(cp);

	mutex_exit(&cpu_lock);

	post_startup_cpu_fixups();

	/*
	 * Enable preemption here so that contention for any locks acquired
	 * later in mp_startup may be preempted if the thread owning those
	 * locks is continously executing on other CPUs (for example, this
	 * CPU must be preemptible to allow other CPUs to pause it during their
	 * startup phases).  It's safe to enable preemption here because the
	 * CPU state is pretty-much fully constructed.
	 */
	curthread->t_preempt = 0;

	/* The base spl should still be at LOCK LEVEL here */
	ASSERT(cp->cpu_base_spl == ipltospl(LOCK_LEVEL));
	set_base_spl();		/* Restore the spl to its proper value */

#ifndef __xpv
	cpupm_init(cp);
#endif
	add_cpunode2devtree(cp->cpu_id, cp->cpu_m.mcpu_cpi);

	/*
	 * Processor group initialization for this CPU is dependent on the
	 * cpuid probing, which must be done in the context of the current
	 * CPU, as well as the CPU's device node initialization (for ACPI).
	 */
	mutex_enter(&cpu_lock);
	pghw_physid_create(cp);
	pg_cpu_init(cp);
	pg_cmt_cpu_startup(cp);
	mutex_exit(&cpu_lock);

	/* Enable interrupts */
	(void) spl0();

	mutex_enter(&cpu_lock);
	cpu_enable_intr(cp);
	cpu_add_active(cp);
	mutex_exit(&cpu_lock);

#ifndef __xpv
	{
		/*
		 * Set up the CPU module for this CPU.  This can't be done
		 * before this CPU is made CPU_READY, because we may (in
		 * heterogeneous systems) need to go load another CPU module.
		 * The act of attempting to load a module may trigger a
		 * cross-call, which will ASSERT unless this cpu is CPU_READY.
		 */
		cmi_hdl_t hdl;

		if ((hdl = cmi_init(CMI_HDL_NATIVE, cmi_ntv_hwchipid(CPU),
		    cmi_ntv_hwcoreid(CPU), cmi_ntv_hwstrandid(CPU))) != NULL) {
			if (x86_feature & X86_MCA)
				cmi_mca_init(hdl);
		}
	}
#endif /* __xpv */

	if (boothowto & RB_DEBUG)
		kdi_cpu_init();

	/*
	 * Setting the bit in cpu_ready_set must be the last operation in
	 * processor initialization; the boot CPU will continue to boot once
	 * it sees this bit set for all active CPUs.
	 */
	CPUSET_ATOMIC_ADD(cpu_ready_set, cp->cpu_id);

	/*
	 * Because mp_startup() gets fired off after init() starts, we
	 * can't use the '?' trick to do 'boot -v' printing - so we
	 * always direct the 'cpu .. online' messages to the log.
	 */
	cmn_err(CE_CONT, "!cpu%d initialization complete - online\n",
	    cp->cpu_id);

	/*
	 * Now we are done with the startup thread, so free it up.
	 */
	thread_exit();
	panic("mp_startup: cannot return");
	/*NOTREACHED*/
}


/*
 * Start CPU on user request.
 */
/* ARGSUSED */
int
mp_cpu_start(struct cpu *cp)
{
	ASSERT(MUTEX_HELD(&cpu_lock));
	return (0);
}

/*
 * Stop CPU on user request.
 */
/* ARGSUSED */
int
mp_cpu_stop(struct cpu *cp)
{
	extern int cbe_psm_timer_mode;
	ASSERT(MUTEX_HELD(&cpu_lock));

#ifdef __xpv
	/*
	 * We can't offline vcpu0.
	 */
	if (cp->cpu_id == 0)
		return (EBUSY);
#endif

	/*
	 * If TIMER_PERIODIC mode is used, CPU0 is the one running it;
	 * can't stop it.  (This is true only for machines with no TSC.)
	 */

	if ((cbe_psm_timer_mode == TIMER_PERIODIC) && (cp->cpu_id == 0))
		return (EBUSY);

	return (0);
}

/*
 * Take the specified CPU out of participation in interrupts.
 */
int
cpu_disable_intr(struct cpu *cp)
{
	if (psm_disable_intr(cp->cpu_id) != DDI_SUCCESS)
		return (EBUSY);

	cp->cpu_flags &= ~CPU_ENABLE;
	return (0);
}

/*
 * Allow the specified CPU to participate in interrupts.
 */
void
cpu_enable_intr(struct cpu *cp)
{
	ASSERT(MUTEX_HELD(&cpu_lock));
	cp->cpu_flags |= CPU_ENABLE;
	psm_enable_intr(cp->cpu_id);
}


/*ARGSUSED*/
void
mp_cpu_faulted_enter(struct cpu *cp)
{
#ifndef __xpv
	cmi_hdl_t hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp),
	    cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp));

	if (hdl != NULL) {
		cmi_faulted_enter(hdl);
		cmi_hdl_rele(hdl);
	}
#endif
}

/*ARGSUSED*/
void
mp_cpu_faulted_exit(struct cpu *cp)
{
#ifndef __xpv
	cmi_hdl_t hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp),
	    cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp));

	if (hdl != NULL) {
		cmi_faulted_exit(hdl);
		cmi_hdl_rele(hdl);
	}
#endif
}

/*
 * The following two routines are used as context operators on threads belonging
 * to processes with a private LDT (see sysi86).  Due to the rarity of such
 * processes, these routines are currently written for best code readability and
 * organization rather than speed.  We could avoid checking x86_feature at every
 * context switch by installing different context ops, depending on the
 * x86_feature flags, at LDT creation time -- one for each combination of fast
 * syscall feature flags.
 */

/*ARGSUSED*/
void
cpu_fast_syscall_disable(void *arg)
{
	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP))
		cpu_sep_disable();
	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC))
		cpu_asysc_disable();
}

/*ARGSUSED*/
void
cpu_fast_syscall_enable(void *arg)
{
	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP))
		cpu_sep_enable();
	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC))
		cpu_asysc_enable();
}

static void
cpu_sep_enable(void)
{
	ASSERT(x86_feature & X86_SEP);
	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);

	wrmsr(MSR_INTC_SEP_CS, (uint64_t)(uintptr_t)KCS_SEL);
}

static void
cpu_sep_disable(void)
{
	ASSERT(x86_feature & X86_SEP);
	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);

	/*
	 * Setting the SYSENTER_CS_MSR register to 0 causes software executing
	 * the sysenter or sysexit instruction to trigger a #gp fault.
	 */
	wrmsr(MSR_INTC_SEP_CS, 0);
}

static void
cpu_asysc_enable(void)
{
	ASSERT(x86_feature & X86_ASYSC);
	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);

	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) |
	    (uint64_t)(uintptr_t)AMD_EFER_SCE);
}

static void
cpu_asysc_disable(void)
{
	ASSERT(x86_feature & X86_ASYSC);
	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);

	/*
	 * Turn off the SCE (syscall enable) bit in the EFER register. Software
	 * executing syscall or sysret with this bit off will incur a #ud trap.
	 */
	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) &
	    ~((uint64_t)(uintptr_t)AMD_EFER_SCE));
}