Merge commit 'b6031810da58df96413bf76e068638fcab1f228a'

[unleashed.git] / arch / x86 / kernel / os / sysi86.c

/*
 * 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 (c) 1992, 2010, Oracle and/or its affiliates. All rights reserved.
 * Copyright 2018 Joyent, Inc.
 */

/*      Copyright (c) 1990, 1991 UNIX System Laboratories, Inc. */
/*      Copyright (c) 1984, 1986, 1987, 1988, 1989, 1990 AT&T   */
/*        All Rights Reserved   */

/*      Copyright (c) 1987, 1988 Microsoft Corporation  */
/*        All Rights Reserved   */

#include <sys/param.h>
#include <sys/types.h>
#include <sys/sysmacros.h>
#include <sys/systm.h>
#include <sys/signal.h>
#include <sys/errno.h>
#include <sys/fault.h>
#include <sys/syscall.h>
#include <sys/cpuvar.h>
#include <sys/sysi86.h>
#include <sys/psw.h>
#include <sys/cred.h>
#include <sys/policy.h>
#include <sys/thread.h>
#include <sys/debug.h>
#include <sys/ontrap.h>
#include <sys/privregs.h>
#include <sys/x86_archext.h>
#include <sys/vmem.h>
#include <sys/kmem.h>
#include <sys/mman.h>
#include <sys/archsystm.h>
#include <vm/hat.h>
#include <vm/as.h>
#include <vm/seg.h>
#include <vm/seg_kmem.h>
#include <vm/faultcode.h>
#include <sys/fp.h>
#include <sys/cmn_err.h>
#include <sys/segments.h>
#include <sys/clock.h>
#include <vm/hat_i86.h>

static void ldt_alloc(proc_t *, uint_t);
static void ldt_free(proc_t *);
static void ldt_dup(proc_t *, proc_t *);
static void ldt_grow(proc_t *, uint_t);

/*
 * sysi86 System Call
 */

/* ARGSUSED */
int
sysi86(short cmd, uintptr_t arg1, uintptr_t arg2, uintptr_t arg3)
{
        struct ssd ssd;
        int error = 0;
        int c;
        proc_t *pp = curproc;

        switch (cmd) {

        /*
         * The SI86V86 subsystem call of the SYSI86 system call
         * supports only one subcode -- V86SC_IOPL.
         */
        case SI86V86:
                if (arg1 == V86SC_IOPL) {
                        struct regs *rp = lwptoregs(ttolwp(curthread));
                        greg_t oldpl = rp->r_ps & PS_IOPL;
                        greg_t newpl = arg2 & PS_IOPL;

                        /*
                         * Must be privileged to run this system call
                         * if giving more io privilege.
                         */
                        if (newpl > oldpl && (error =
                            secpolicy_sys_config(CRED(), B_FALSE)) != 0)
                                return (set_errno(error));
                        rp->r_ps ^= oldpl ^ newpl;
                } else
                        error = EINVAL;
                break;

        /*
         * Set a segment descriptor
         */
        case SI86DSCR:
                /*
                 * There are considerable problems here manipulating
                 * resources shared by many running lwps.  Get everyone
                 * into a safe state before changing the LDT.
                 */
                if (curthread != pp->p_agenttp && !holdlwps(SHOLDFORK1)) {
                        error = EINTR;
                        break;
                }

                if (get_udatamodel() == DATAMODEL_LP64) {
                        error = EINVAL;
                        break;
                }

                if (copyin((caddr_t)arg1, &ssd, sizeof (ssd)) < 0) {
                        error = EFAULT;
                        break;
                }

                error = setdscr(&ssd);

                mutex_enter(&pp->p_lock);
                if (curthread != pp->p_agenttp)
                        continuelwps(pp);
                mutex_exit(&pp->p_lock);
                break;

        case SI86FPHW:
                c = fp_kind & 0xff;
                if (suword32((void *)arg1, c) == -1)
                        error = EFAULT;
                break;

        case SI86FPSTART:
                /*
                 * arg1 is the address of _fp_hw
                 * arg2 is the desired x87 FCW value
                 * arg3 is the desired SSE MXCSR value
                 * a return value of one means SSE hardware, else none.
                 */
                c = fp_kind & 0xff;
                if (suword32((void *)arg1, c) == -1) {
                        error = EFAULT;
                        break;
                }
                fpsetcw((uint16_t)arg2, (uint32_t)arg3);
                return ((fp_kind & __FP_SSE) ? 1 : 0);

        /* real time clock management commands */

        case WTODC:
                if ((error = secpolicy_settime(CRED())) == 0) {
                        timestruc_t ts;
                        mutex_enter(&tod_lock);
                        gethrestime(&ts);
                        tod_set(ts);
                        mutex_exit(&tod_lock);
                }
                break;

/* Give some timezone playing room */
#define ONEWEEK (7 * 24 * 60 * 60)

        case SGMTL:
                /*
                 * Called from 32 bit land, negative values
                 * are not sign extended, so we do that here
                 * by casting it to an int and back.  We also
                 * clamp the value to within reason and detect
                 * when a 64 bit call overflows an int.
                 */
                if ((error = secpolicy_settime(CRED())) == 0) {
                        int newlag = (int)arg1;

#ifdef _SYSCALL32_IMPL
                        if (get_udatamodel() == DATAMODEL_NATIVE &&
                            (long)newlag != (long)arg1) {
                                error = EOVERFLOW;
                        } else
#endif
                        if (newlag >= -ONEWEEK && newlag <= ONEWEEK)
                                sgmtl(newlag);
                        else
                                error = EOVERFLOW;
                }
                break;

        case GGMTL:
                if (get_udatamodel() == DATAMODEL_NATIVE) {
                        if (sulword((void *)arg1, ggmtl()) == -1)
                                error = EFAULT;
#ifdef _SYSCALL32_IMPL
                } else {
                        time_t gmtl;

                        if ((gmtl = ggmtl()) > INT32_MAX) {
                                /*
                                 * Since gmt_lag can at most be
                                 * +/- 12 hours, something is
                                 * *seriously* messed up here.
                                 */
                                error = EOVERFLOW;
                        } else if (suword32((void *)arg1, (int32_t)gmtl) == -1)
                                error = EFAULT;
#endif
                }
                break;

        case RTCSYNC:
                if ((error = secpolicy_settime(CRED())) == 0)
                        rtcsync();
                break;

        /* END OF real time clock management commands */

        default:
                error = EINVAL;
                break;
        }
        return (error == 0 ? 0 : set_errno(error));
}

void
usd_to_ssd(user_desc_t *usd, struct ssd *ssd, selector_t sel)
{
        ssd->bo = USEGD_GETBASE(usd);
        ssd->ls = USEGD_GETLIMIT(usd);
        ssd->sel = sel;

        /*
         * set type, dpl and present bits.
         */
        ssd->acc1 = usd->usd_type;
        ssd->acc1 |= usd->usd_dpl << 5;
        ssd->acc1 |= usd->usd_p << (5 + 2);

        /*
         * set avl, DB and granularity bits.
         */
        ssd->acc2 = usd->usd_avl;

#if defined(__amd64)
        ssd->acc2 |= usd->usd_long << 1;
#else
        ssd->acc2 |= usd->usd_reserved << 1;
#endif

        ssd->acc2 |= usd->usd_def32 << (1 + 1);
        ssd->acc2 |= usd->usd_gran << (1 + 1 + 1);
}

static void
ssd_to_usd(struct ssd *ssd, user_desc_t *usd)
{

        ASSERT(bcmp(usd, &null_udesc, sizeof (*usd)) == 0);

        USEGD_SETBASE(usd, ssd->bo);
        USEGD_SETLIMIT(usd, ssd->ls);

        /*
         * Set type, dpl and present bits.
         *
         * Force the "accessed" bit to on so that we don't run afoul of
         * KPTI.
         */
        usd->usd_type = ssd->acc1 | SDT_A;
        usd->usd_dpl = ssd->acc1 >> 5;
        usd->usd_p = ssd->acc1 >> (5 + 2);

        ASSERT(usd->usd_type >= SDT_MEMRO);
        ASSERT(usd->usd_dpl == SEL_UPL);

        /*
         * 64-bit code selectors are never allowed in the LDT.
         * Reserved bit is always 0 on 32-bit systems.
         */
#if defined(__amd64)
        usd->usd_long = 0;
#else
        usd->usd_reserved = 0;
#endif

        /*
         * set avl, DB and granularity bits.
         */
        usd->usd_avl = ssd->acc2;
        usd->usd_def32 = ssd->acc2 >> (1 + 1);
        usd->usd_gran = ssd->acc2 >> (1 + 1 + 1);
}


#if defined(__i386)

static void
ssd_to_sgd(struct ssd *ssd, gate_desc_t *sgd)
{

        ASSERT(bcmp(sgd, &null_sdesc, sizeof (*sgd)) == 0);

        sgd->sgd_looffset = ssd->bo;
        sgd->sgd_hioffset = ssd->bo >> 16;

        sgd->sgd_selector = ssd->ls;

        /*
         * set type, dpl and present bits.
         */
        sgd->sgd_type = ssd->acc1;
        sgd->sgd_dpl = ssd->acc1 >> 5;
        sgd->sgd_p = ssd->acc1 >> 7;
        ASSERT(sgd->sgd_type == SDT_SYSCGT);
        ASSERT(sgd->sgd_dpl == SEL_UPL);
        sgd->sgd_stkcpy = 0;
}

#endif  /* __i386 */

/*
 * Load LDT register with the current process's LDT.
 */
static void
ldt_load(void)
{
        size_t len;
        system_desc_t desc;

        /*
         * Before we can use the LDT on this CPU, we must install the LDT in the
         * user mapping table.
         */
        len = (curproc->p_ldtlimit + 1) * sizeof (user_desc_t);
        bcopy(curproc->p_ldt, CPU->cpu_m.mcpu_ldt, len);
        CPU->cpu_m.mcpu_ldt_len = len;
        set_syssegd(&desc, CPU->cpu_m.mcpu_ldt, len - 1, SDT_SYSLDT, SEL_KPL);
        *((system_desc_t *)&CPU->cpu_gdt[GDT_LDT]) = desc;

        wr_ldtr(ULDT_SEL);
}

/*
 * Store a NULL selector in the LDTR. All subsequent illegal references to
 * the LDT will result in a #gp.
 */
void
ldt_unload(void)
{
        *((system_desc_t *)&CPU->cpu_gdt[GDT_LDT]) = null_sdesc;
        wr_ldtr(0);

        bzero(CPU->cpu_m.mcpu_ldt, CPU->cpu_m.mcpu_ldt_len);
        CPU->cpu_m.mcpu_ldt_len = 0;
}

/*ARGSUSED*/
static void
ldt_savectx(proc_t *p)
{
        ASSERT(p->p_ldt != NULL);
        ASSERT(p == curproc);

#if defined(__amd64)
        /*
         * The 64-bit kernel must be sure to clear any stale ldt
         * selectors when context switching away from a process that
         * has a private ldt. Consider the following example:
         *
         *      Wine creats a ldt descriptor and points a segment register
         *      to it.
         *
         *      We then context switch away from wine lwp to kernel
         *      thread and hit breakpoint in kernel with kmdb
         *
         *      When we continue and resume from kmdb we will #gp
         *      fault since kmdb will have saved the stale ldt selector
         *      from wine and will try to restore it but we are no longer in
         *      the context of the wine process and do not have our
         *      ldtr register pointing to the private ldt.
         */
        reset_sregs();
#endif

        ldt_unload();
        cpu_fast_syscall_enable();
}

static void
ldt_restorectx(proc_t *p)
{
        ASSERT(p->p_ldt != NULL);
        ASSERT(p == curproc);

        ldt_load();
        cpu_fast_syscall_disable();
}

/*
 * At exec time, we need to clear up our LDT context and re-enable fast syscalls
 * for the new process image.
 *
 * The same is true for the other case, where we have:
 *
 * proc_exit()
 *  ->exitpctx()->ldt_savectx()
 *  ->freepctx()->ldt_freectx()
 *
 * Because pre-emption is not prevented between the two callbacks, we could have
 * come off CPU, and brought back LDT context when coming back on CPU via
 * ldt_restorectx().
 */
/* ARGSUSED */
static void
ldt_freectx(proc_t *p, int isexec)
{
        ASSERT(p->p_ldt != NULL);
        ASSERT(p == curproc);

        kpreempt_disable();
        ldt_free(p);
        cpu_fast_syscall_enable();
        kpreempt_enable();
}

/*
 * Install ctx op that ensures syscall/sysenter are disabled.
 * See comments below.
 *
 * When a thread with a private LDT forks, the new process
 * must have the LDT context ops installed.
 */
/* ARGSUSED */
static void
ldt_installctx(proc_t *p, proc_t *cp)
{
        proc_t          *targ = p;
        kthread_t       *t;

        /*
         * If this is a fork, operate on the child process.
         */
        if (cp != NULL) {
                targ = cp;
                ldt_dup(p, cp);
        }

        /*
         * The process context ops expect the target process as their argument.
         */
        ASSERT(removepctx(targ, targ, ldt_savectx, ldt_restorectx,
            ldt_installctx, ldt_savectx, ldt_freectx) == 0);

        installpctx(targ, targ, ldt_savectx, ldt_restorectx,
            ldt_installctx, ldt_savectx, ldt_freectx);

        /*
         * We've just disabled fast system call and return instructions; take
         * the slow path out to make sure we don't try to use one to return
         * back to user. We must set t_post_sys for every thread in the
         * process to make sure none of them escape out via fast return.
         */

        mutex_enter(&targ->p_lock);
        t = targ->p_tlist;
        do {
                t->t_post_sys = 1;
        } while ((t = t->t_forw) != targ->p_tlist);
        mutex_exit(&targ->p_lock);
}

int
setdscr(struct ssd *ssd)
{
        ushort_t seli;          /* selector index */
        user_desc_t *ldp;       /* descriptor pointer */
        user_desc_t ndesc;      /* new descriptor */
        proc_t  *pp = curproc;
        int     rc = 0;

        /*
         * LDT segments: executable and data at DPL 3 only.
         */
        if (!SELISLDT(ssd->sel) || !SELISUPL(ssd->sel))
                return (EINVAL);

        /*
         * check the selector index.
         */
        seli = SELTOIDX(ssd->sel);
        if (seli >= MAXNLDT || seli < LDT_UDBASE)
                return (EINVAL);

        ndesc = null_udesc;
        mutex_enter(&pp->p_ldtlock);

        /*
         * If this is the first time for this process then setup a
         * private LDT for it.
         */
        if (pp->p_ldt == NULL) {
                ldt_alloc(pp, seli);

                /*
                 * Now that this process has a private LDT, the use of
                 * the syscall/sysret and sysenter/sysexit instructions
                 * is forbidden for this processes because they destroy
                 * the contents of %cs and %ss segment registers.
                 *
                 * Explicity disable them here and add a context handler
                 * to the process. Note that disabling
                 * them here means we can't use sysret or sysexit on
                 * the way out of this system call - so we force this
                 * thread to take the slow path (which doesn't make use
                 * of sysenter or sysexit) back out.
                 */
                kpreempt_disable();
                ldt_installctx(pp, NULL);
                cpu_fast_syscall_disable();
                ASSERT(curthread->t_post_sys != 0);
                kpreempt_enable();

        } else if (seli > pp->p_ldtlimit) {
                ASSERT(pp->p_pctx != NULL);

                /*
                 * Increase size of ldt to include seli.
                 */
                ldt_grow(pp, seli);
        }

        ASSERT(seli <= pp->p_ldtlimit);
        ldp = &pp->p_ldt[seli];

        /*
         * On the 64-bit kernel, this is where things get more subtle.
         * Recall that in the 64-bit kernel, when we enter the kernel we
         * deliberately -don't- reload the segment selectors we came in on
         * for %ds, %es, %fs or %gs. Messing with selectors is expensive,
         * and the underlying descriptors are essentially ignored by the
         * hardware in long mode - except for the base that we override with
         * the gsbase MSRs.
         *
         * However, there's one unfortunate issue with this rosy picture --
         * a descriptor that's not marked as 'present' will still generate
         * an #np when loading a segment register.
         *
         * Consider this case.  An lwp creates a harmless LDT entry, points
         * one of it's segment registers at it, then tells the kernel (here)
         * to delete it.  In the 32-bit kernel, the #np will happen on the
         * way back to userland where we reload the segment registers, and be
         * handled in kern_gpfault().  In the 64-bit kernel, the same thing
         * will happen in the normal case too.  However, if we're trying to
         * use a debugger that wants to save and restore the segment registers,
         * and the debugger things that we have valid segment registers, we
         * have the problem that the debugger will try and restore the
         * segment register that points at the now 'not present' descriptor
         * and will take a #np right there.
         *
         * We should obviously fix the debugger to be paranoid about
         * -not- restoring segment registers that point to bad descriptors;
         * however we can prevent the problem here if we check to see if any
         * of the segment registers are still pointing at the thing we're
         * destroying; if they are, return an error instead. (That also seems
         * a lot better failure mode than SIGKILL and a core file
         * from kern_gpfault() too.)
         */
        if (SI86SSD_PRES(ssd) == 0) {
                kthread_t *t;
                int bad = 0;

                /*
                 * Look carefully at the segment registers of every lwp
                 * in the process (they're all stopped by our caller).
                 * If we're about to invalidate a descriptor that's still
                 * being referenced by *any* of them, return an error,
                 * rather than having them #gp on their way out of the kernel.
                 */
                ASSERT(pp->p_lwprcnt == 1);

                mutex_enter(&pp->p_lock);
                t = pp->p_tlist;
                do {
                        klwp_t *lwp = ttolwp(t);
                        struct regs *rp = lwp->lwp_regs;
#if defined(__amd64)
                        pcb_t *pcb = &lwp->lwp_pcb;
#endif

                        if (ssd->sel == rp->r_cs || ssd->sel == rp->r_ss) {
                                bad = 1;
                                break;
                        }

#if defined(__amd64)
                        if (pcb->pcb_rupdate == 1) {
                                if (ssd->sel == pcb->pcb_ds ||
                                    ssd->sel == pcb->pcb_es ||
                                    ssd->sel == pcb->pcb_fs ||
                                    ssd->sel == pcb->pcb_gs) {
                                        bad = 1;
                                        break;
                                }
                        } else
#endif
                        {
                                if (ssd->sel == rp->r_ds ||
                                    ssd->sel == rp->r_es ||
                                    ssd->sel == rp->r_fs ||
                                    ssd->sel == rp->r_gs) {
                                        bad = 1;
                                        break;
                                }
                        }

                } while ((t = t->t_forw) != pp->p_tlist);
                mutex_exit(&pp->p_lock);

                if (bad) {
                        mutex_exit(&pp->p_ldtlock);
                        return (EBUSY);
                }
        }

        /*
         * If acc1 is zero, clear the descriptor (including the 'present' bit).
         * Make sure we update the CPU-private copy of the LDT.
         */
        if (ssd->acc1 == 0) {
                rc  = ldt_update_segd(ldp, &null_udesc);
                kpreempt_disable();
                ldt_load();
                kpreempt_enable();
                mutex_exit(&pp->p_ldtlock);
                return (rc);
        }

        /*
         * Check segment type, allow segment not present and
         * only user DPL (3).
         */
        if (SI86SSD_DPL(ssd) != SEL_UPL) {
                mutex_exit(&pp->p_ldtlock);
                return (EINVAL);
        }

        /*
         * Do not allow 32-bit applications to create 64-bit mode code
         * segments.
         */
        if (SI86SSD_ISUSEG(ssd) && ((SI86SSD_TYPE(ssd) >> 3) & 1) == 1 &&
            SI86SSD_ISLONG(ssd)) {
                mutex_exit(&pp->p_ldtlock);
                return (EINVAL);
        }

        /*
         * Set up a code or data user segment descriptor, making sure to update
         * the CPU-private copy of the LDT.
         */
        if (SI86SSD_ISUSEG(ssd)) {
                ssd_to_usd(ssd, &ndesc);
                rc = ldt_update_segd(ldp, &ndesc);
                kpreempt_disable();
                ldt_load();
                kpreempt_enable();
                mutex_exit(&pp->p_ldtlock);
                return (rc);
        }

        mutex_exit(&pp->p_ldtlock);
        return (EINVAL);
}

/*
 * Allocate new LDT for process just large enough to contain seli.  Note we
 * allocate and grow LDT in PAGESIZE chunks. We do this to simplify the
 * implementation and because on the hypervisor it's required, since the LDT
 * must live on pages that have PROT_WRITE removed and which are given to the
 * hypervisor.
 *
 * Note that we don't actually load the LDT into the current CPU here: it's done
 * later by our caller.
 */
static void
ldt_alloc(proc_t *pp, uint_t seli)
{
        user_desc_t     *ldt;
        size_t          ldtsz;
        uint_t          nsels;

        ASSERT(MUTEX_HELD(&pp->p_ldtlock));
        ASSERT(pp->p_ldt == NULL);
        ASSERT(pp->p_ldtlimit == 0);

        /*
         * Allocate new LDT just large enough to contain seli. The LDT must
         * always be allocated in units of pages for KPTI.
         */
        ldtsz = P2ROUNDUP((seli + 1) * sizeof (user_desc_t), PAGESIZE);
        nsels = ldtsz / sizeof (user_desc_t);
        ASSERT(nsels >= MINNLDT && nsels <= MAXNLDT);

        ldt = kmem_zalloc(ldtsz, KM_SLEEP);
        ASSERT(IS_P2ALIGNED(ldt, PAGESIZE));


        pp->p_ldt = ldt;
        pp->p_ldtlimit = nsels - 1;
}

static void
ldt_free(proc_t *pp)
{
        user_desc_t     *ldt;
        size_t          ldtsz;

        ASSERT(pp->p_ldt != NULL);

        mutex_enter(&pp->p_ldtlock);
        ldt = pp->p_ldt;
        ldtsz = (pp->p_ldtlimit + 1) * sizeof (user_desc_t);

        ASSERT(IS_P2ALIGNED(ldtsz, PAGESIZE));

        pp->p_ldt = NULL;
        pp->p_ldtlimit = 0;
        mutex_exit(&pp->p_ldtlock);

        if (pp == curproc) {
                kpreempt_disable();
                ldt_unload();
                kpreempt_enable();
        }


        kmem_free(ldt, ldtsz);
}

/*
 * On fork copy new ldt for child.
 */
static void
ldt_dup(proc_t *pp, proc_t *cp)
{
        size_t  ldtsz;

        ASSERT(pp->p_ldt != NULL);
        ASSERT(cp != curproc);

        /*
         * I assume the parent's ldt can't increase since we're in a fork.
         */
        mutex_enter(&pp->p_ldtlock);
        mutex_enter(&cp->p_ldtlock);

        ldtsz = (pp->p_ldtlimit + 1) * sizeof (user_desc_t);

        ldt_alloc(cp, pp->p_ldtlimit);


        bcopy(pp->p_ldt, cp->p_ldt, ldtsz);

        mutex_exit(&cp->p_ldtlock);
        mutex_exit(&pp->p_ldtlock);

}

/*
 * Note that we don't actually load the LDT into the current CPU here: it's done
 * later by our caller - unless we take an error.  This works out because
 * ldt_load() does a copy of ->p_ldt instead of directly loading it into the GDT
 * (and therefore can't be using the freed old LDT), and by definition if the
 * new entry didn't pass validation, then the proc shouldn't be referencing an
 * entry in the extended region.
 */
static void
ldt_grow(proc_t *pp, uint_t seli)
{
        user_desc_t     *oldt, *nldt;
        uint_t          nsels;
        size_t          oldtsz, nldtsz;

        ASSERT(MUTEX_HELD(&pp->p_ldtlock));
        ASSERT(pp->p_ldt != NULL);
        ASSERT(pp->p_ldtlimit != 0);

        /*
         * Allocate larger LDT just large enough to contain seli. The LDT must
         * always be allocated in units of pages for KPTI.
         */
        nldtsz = P2ROUNDUP((seli + 1) * sizeof (user_desc_t), PAGESIZE);
        nsels = nldtsz / sizeof (user_desc_t);
        ASSERT(nsels >= MINNLDT && nsels <= MAXNLDT);
        ASSERT(nsels > pp->p_ldtlimit);

        oldt = pp->p_ldt;
        oldtsz = (pp->p_ldtlimit + 1) * sizeof (user_desc_t);

        nldt = kmem_zalloc(nldtsz, KM_SLEEP);
        ASSERT(IS_P2ALIGNED(nldt, PAGESIZE));

        bcopy(oldt, nldt, oldtsz);

        /*
         * unload old ldt.
         */
        kpreempt_disable();
        ldt_unload();
        kpreempt_enable();


        pp->p_ldt = nldt;
        pp->p_ldtlimit = nsels - 1;

        kmem_free(oldt, oldtsz);
}