xref: /freebsd/sys/amd64/amd64/vm_machdep.c (revision fe2c9f83)
1 /*-
2  * SPDX-License-Identifier: BSD-4-Clause
3  *
4  * Copyright (c) 1982, 1986 The Regents of the University of California.
5  * Copyright (c) 1989, 1990 William Jolitz
6  * Copyright (c) 1994 John Dyson
7  * All rights reserved.
8  *
9  * This code is derived from software contributed to Berkeley by
10  * the Systems Programming Group of the University of Utah Computer
11  * Science Department, and William Jolitz.
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  * 3. All advertising materials mentioning features or use of this software
22  *    must display the following acknowledgement:
23  *	This product includes software developed by the University of
24  *	California, Berkeley and its contributors.
25  * 4. Neither the name of the University nor the names of its contributors
26  *    may be used to endorse or promote products derived from this software
27  *    without specific prior written permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39  * SUCH DAMAGE.
40  *
41  *	from: @(#)vm_machdep.c	7.3 (Berkeley) 5/13/91
42  *	Utah $Hdr: vm_machdep.c 1.16.1.1 89/06/23$
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include "opt_isa.h"
49 #include "opt_cpu.h"
50 
51 #include <sys/param.h>
52 #include <sys/systm.h>
53 #include <sys/bio.h>
54 #include <sys/buf.h>
55 #include <sys/kernel.h>
56 #include <sys/ktr.h>
57 #include <sys/lock.h>
58 #include <sys/malloc.h>
59 #include <sys/mbuf.h>
60 #include <sys/mutex.h>
61 #include <sys/priv.h>
62 #include <sys/proc.h>
63 #include <sys/procctl.h>
64 #include <sys/smp.h>
65 #include <sys/sysctl.h>
66 #include <sys/sysent.h>
67 #include <sys/unistd.h>
68 #include <sys/vnode.h>
69 #include <sys/vmmeter.h>
70 #include <sys/wait.h>
71 
72 #include <machine/cpu.h>
73 #include <machine/md_var.h>
74 #include <machine/pcb.h>
75 #include <machine/smp.h>
76 #include <machine/specialreg.h>
77 #include <machine/tss.h>
78 
79 #include <vm/vm.h>
80 #include <vm/vm_extern.h>
81 #include <vm/vm_kern.h>
82 #include <vm/vm_page.h>
83 #include <vm/vm_map.h>
84 #include <vm/vm_param.h>
85 
86 _Static_assert(OFFSETOF_MONITORBUF == offsetof(struct pcpu, pc_monitorbuf),
87     "OFFSETOF_MONITORBUF does not correspond with offset of pc_monitorbuf.");
88 
89 void
90 set_top_of_stack_td(struct thread *td)
91 {
92 	td->td_md.md_stack_base = td->td_kstack +
93 	    td->td_kstack_pages * PAGE_SIZE;
94 }
95 
96 struct savefpu *
97 get_pcb_user_save_td(struct thread *td)
98 {
99 	KASSERT(((vm_offset_t)td->td_md.md_usr_fpu_save %
100 	    XSAVE_AREA_ALIGN) == 0,
101 	    ("Unaligned pcb_user_save area ptr %p td %p",
102 	    td->td_md.md_usr_fpu_save, td));
103 	return (td->td_md.md_usr_fpu_save);
104 }
105 
106 struct pcb *
107 get_pcb_td(struct thread *td)
108 {
109 
110 	return (&td->td_md.md_pcb);
111 }
112 
113 struct savefpu *
114 get_pcb_user_save_pcb(struct pcb *pcb)
115 {
116 	struct thread *td;
117 
118 	td = __containerof(pcb, struct thread, td_md.md_pcb);
119 	return (get_pcb_user_save_td(td));
120 }
121 
122 void *
123 alloc_fpusave(int flags)
124 {
125 	void *res;
126 	struct savefpu_ymm *sf;
127 
128 	res = malloc(cpu_max_ext_state_size, M_DEVBUF, flags);
129 	if (use_xsave) {
130 		sf = (struct savefpu_ymm *)res;
131 		bzero(&sf->sv_xstate.sx_hd, sizeof(sf->sv_xstate.sx_hd));
132 		sf->sv_xstate.sx_hd.xstate_bv = xsave_mask;
133 	}
134 	return (res);
135 }
136 
137 /*
138  * Common code shared between cpu_fork() and cpu_copy_thread() for
139  * initializing a thread.
140  */
141 static void
142 copy_thread(struct thread *td1, struct thread *td2)
143 {
144 	struct pcb *pcb2;
145 
146 	pcb2 = td2->td_pcb;
147 
148 	/* Ensure that td1's pcb is up to date for user threads. */
149 	if ((td2->td_pflags & TDP_KTHREAD) == 0) {
150 		MPASS(td1 == curthread);
151 		fpuexit(td1);
152 		update_pcb_bases(td1->td_pcb);
153 	}
154 
155 	/* Copy td1's pcb */
156 	bcopy(td1->td_pcb, pcb2, sizeof(*pcb2));
157 
158 	/* Properly initialize pcb_save */
159 	pcb2->pcb_save = get_pcb_user_save_pcb(pcb2);
160 
161 	/* Kernel threads start with clean FPU and segment bases. */
162 	if ((td2->td_pflags & TDP_KTHREAD) != 0) {
163 		pcb2->pcb_fsbase = 0;
164 		pcb2->pcb_gsbase = 0;
165 		clear_pcb_flags(pcb2, PCB_FPUINITDONE | PCB_USERFPUINITDONE |
166 		    PCB_KERNFPU | PCB_KERNFPU_THR);
167 	} else {
168 		MPASS((pcb2->pcb_flags & (PCB_KERNFPU | PCB_KERNFPU_THR)) == 0);
169 		bcopy(get_pcb_user_save_td(td1), get_pcb_user_save_pcb(pcb2),
170 		    cpu_max_ext_state_size);
171 	}
172 
173 	/*
174 	 * Set registers for trampoline to user mode.  Leave space for the
175 	 * return address on stack.  These are the kernel mode register values.
176 	 */
177 	pcb2->pcb_r12 = (register_t)fork_return;	/* fork_trampoline argument */
178 	pcb2->pcb_rbp = 0;
179 	pcb2->pcb_rsp = (register_t)td2->td_frame - sizeof(void *);
180 	pcb2->pcb_rbx = (register_t)td2;		/* fork_trampoline argument */
181 	pcb2->pcb_rip = (register_t)fork_trampoline;
182 	/*-
183 	 * pcb2->pcb_dr*:	cloned above.
184 	 * pcb2->pcb_savefpu:	cloned above.
185 	 * pcb2->pcb_flags:	cloned above.
186 	 * pcb2->pcb_onfault:	cloned above (always NULL here?).
187 	 * pcb2->pcb_[fg]sbase:	cloned above
188 	 */
189 
190 	pcb2->pcb_tssp = NULL;
191 
192 	/* Setup to release spin count in fork_exit(). */
193 	td2->td_md.md_spinlock_count = 1;
194 	td2->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
195 	pmap_thread_init_invl_gen(td2);
196 
197 	/*
198 	 * Copy the trap frame for the return to user mode as if from a syscall.
199 	 * This copies most of the user mode register values.  Some of these
200 	 * registers are rewritten by cpu_set_upcall() and linux_set_upcall().
201 	 */
202 	if ((td1->td_proc->p_flag & P_KPROC) == 0) {
203 		bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe));
204 
205 		/*
206 		 * If the current thread has the trap bit set (i.e. a debugger
207 		 * had single stepped the process to the system call), we need
208 		 * to clear the trap flag from the new frame. Otherwise, the new
209 		 * thread will receive a (likely unexpected) SIGTRAP when it
210 		 * executes the first instruction after returning to userland.
211 		 */
212 		td2->td_frame->tf_rflags &= ~PSL_T;
213 	}
214 }
215 
216 /*
217  * Finish a fork operation, with process p2 nearly set up.
218  * Copy and update the pcb, set up the stack so that the child
219  * ready to run and return to user mode.
220  */
221 void
222 cpu_fork(struct thread *td1, struct proc *p2, struct thread *td2, int flags)
223 {
224 	struct proc *p1;
225 	struct pcb *pcb2;
226 	struct mdproc *mdp1, *mdp2;
227 	struct proc_ldt *pldt;
228 
229 	p1 = td1->td_proc;
230 	if ((flags & RFPROC) == 0) {
231 		if ((flags & RFMEM) == 0) {
232 			/* unshare user LDT */
233 			mdp1 = &p1->p_md;
234 			mtx_lock(&dt_lock);
235 			if ((pldt = mdp1->md_ldt) != NULL &&
236 			    pldt->ldt_refcnt > 1 &&
237 			    user_ldt_alloc(p1, 1) == NULL)
238 				panic("could not copy LDT");
239 			mtx_unlock(&dt_lock);
240 		}
241 		return;
242 	}
243 
244 	/* Point the stack and pcb to the actual location */
245 	set_top_of_stack_td(td2);
246 	td2->td_pcb = pcb2 = get_pcb_td(td2);
247 
248 	copy_thread(td1, td2);
249 
250 	/* Reset debug registers in the new process */
251 	x86_clear_dbregs(pcb2);
252 
253 	/* Point mdproc and then copy over p1's contents */
254 	mdp2 = &p2->p_md;
255 	bcopy(&p1->p_md, mdp2, sizeof(*mdp2));
256 
257 	/* Set child return values. */
258 	p2->p_sysent->sv_set_fork_retval(td2);
259 
260 	/* As on i386, do not copy io permission bitmap. */
261 	pcb2->pcb_tssp = NULL;
262 
263 	/* New segment registers. */
264 	set_pcb_flags_raw(pcb2, PCB_FULL_IRET);
265 
266 	/* Copy the LDT, if necessary. */
267 	mdp1 = &td1->td_proc->p_md;
268 	mdp2 = &p2->p_md;
269 	if (mdp1->md_ldt == NULL) {
270 		mdp2->md_ldt = NULL;
271 		return;
272 	}
273 	mtx_lock(&dt_lock);
274 	if (mdp1->md_ldt != NULL) {
275 		if (flags & RFMEM) {
276 			mdp1->md_ldt->ldt_refcnt++;
277 			mdp2->md_ldt = mdp1->md_ldt;
278 			bcopy(&mdp1->md_ldt_sd, &mdp2->md_ldt_sd, sizeof(struct
279 			    system_segment_descriptor));
280 		} else {
281 			mdp2->md_ldt = NULL;
282 			mdp2->md_ldt = user_ldt_alloc(p2, 0);
283 			if (mdp2->md_ldt == NULL)
284 				panic("could not copy LDT");
285 			amd64_set_ldt_data(td2, 0, max_ldt_segment,
286 			    (struct user_segment_descriptor *)
287 			    mdp1->md_ldt->ldt_base);
288 		}
289 	} else
290 		mdp2->md_ldt = NULL;
291 	mtx_unlock(&dt_lock);
292 
293 	/*
294 	 * Now, cpu_switch() can schedule the new process.
295 	 * pcb_rsp is loaded pointing to the cpu_switch() stack frame
296 	 * containing the return address when exiting cpu_switch.
297 	 * This will normally be to fork_trampoline(), which will have
298 	 * %rbx loaded with the new proc's pointer.  fork_trampoline()
299 	 * will set up a stack to call fork_return(p, frame); to complete
300 	 * the return to user-mode.
301 	 */
302 }
303 
304 void
305 x86_set_fork_retval(struct thread *td)
306 {
307 	struct trapframe *frame = td->td_frame;
308 
309 	frame->tf_rax = 0;		/* Child returns zero */
310 	frame->tf_rflags &= ~PSL_C;	/* success */
311 	frame->tf_rdx = 1;		/* System V emulation */
312 }
313 
314 /*
315  * Intercept the return address from a freshly forked process that has NOT
316  * been scheduled yet.
317  *
318  * This is needed to make kernel threads stay in kernel mode.
319  */
320 void
321 cpu_fork_kthread_handler(struct thread *td, void (*func)(void *), void *arg)
322 {
323 	/*
324 	 * Note that the trap frame follows the args, so the function
325 	 * is really called like this:  func(arg, frame);
326 	 */
327 	td->td_pcb->pcb_r12 = (long) func;	/* function */
328 	td->td_pcb->pcb_rbx = (long) arg;	/* first arg */
329 }
330 
331 void
332 cpu_exit(struct thread *td)
333 {
334 
335 	/*
336 	 * If this process has a custom LDT, release it.
337 	 */
338 	if (td->td_proc->p_md.md_ldt != NULL)
339 		user_ldt_free(td);
340 }
341 
342 void
343 cpu_thread_exit(struct thread *td)
344 {
345 	struct pcb *pcb;
346 
347 	critical_enter();
348 	if (td == PCPU_GET(fpcurthread))
349 		fpudrop();
350 	critical_exit();
351 
352 	pcb = td->td_pcb;
353 
354 	/* Disable any hardware breakpoints. */
355 	if (pcb->pcb_flags & PCB_DBREGS) {
356 		reset_dbregs();
357 		clear_pcb_flags(pcb, PCB_DBREGS);
358 	}
359 }
360 
361 void
362 cpu_thread_clean(struct thread *td)
363 {
364 	struct pcb *pcb;
365 
366 	pcb = td->td_pcb;
367 
368 	/*
369 	 * Clean TSS/iomap
370 	 */
371 	if (pcb->pcb_tssp != NULL) {
372 		pmap_pti_remove_kva((vm_offset_t)pcb->pcb_tssp,
373 		    (vm_offset_t)pcb->pcb_tssp + ctob(IOPAGES + 1));
374 		kmem_free((vm_offset_t)pcb->pcb_tssp, ctob(IOPAGES + 1));
375 		pcb->pcb_tssp = NULL;
376 	}
377 }
378 
379 void
380 cpu_thread_swapin(struct thread *td)
381 {
382 }
383 
384 void
385 cpu_thread_swapout(struct thread *td)
386 {
387 }
388 
389 void
390 cpu_thread_alloc(struct thread *td)
391 {
392 	struct pcb *pcb;
393 	struct xstate_hdr *xhdr;
394 
395 	set_top_of_stack_td(td);
396 	td->td_pcb = pcb = get_pcb_td(td);
397 	td->td_frame = (struct trapframe *)td->td_md.md_stack_base - 1;
398 	td->td_md.md_usr_fpu_save = fpu_save_area_alloc();
399 	pcb->pcb_save = get_pcb_user_save_pcb(pcb);
400 	if (use_xsave) {
401 		xhdr = (struct xstate_hdr *)(pcb->pcb_save + 1);
402 		bzero(xhdr, sizeof(*xhdr));
403 		xhdr->xstate_bv = xsave_mask;
404 	}
405 }
406 
407 void
408 cpu_thread_free(struct thread *td)
409 {
410 	cpu_thread_clean(td);
411 
412 	fpu_save_area_free(td->td_md.md_usr_fpu_save);
413 	td->td_md.md_usr_fpu_save = NULL;
414 }
415 
416 bool
417 cpu_exec_vmspace_reuse(struct proc *p, vm_map_t map)
418 {
419 
420 	return (((curproc->p_md.md_flags & P_MD_KPTI) != 0) ==
421 	    (vm_map_pmap(map)->pm_ucr3 != PMAP_NO_CR3));
422 }
423 
424 static void
425 cpu_procctl_kpti_ctl(struct proc *p, int val)
426 {
427 
428 	if (pti && val == PROC_KPTI_CTL_ENABLE_ON_EXEC)
429 		p->p_md.md_flags |= P_MD_KPTI;
430 	if (val == PROC_KPTI_CTL_DISABLE_ON_EXEC)
431 		p->p_md.md_flags &= ~P_MD_KPTI;
432 }
433 
434 static void
435 cpu_procctl_kpti_status(struct proc *p, int *val)
436 {
437 	*val = (p->p_md.md_flags & P_MD_KPTI) != 0 ?
438 	    PROC_KPTI_CTL_ENABLE_ON_EXEC:
439 	    PROC_KPTI_CTL_DISABLE_ON_EXEC;
440 	if (vmspace_pmap(p->p_vmspace)->pm_ucr3 != PMAP_NO_CR3)
441 		*val |= PROC_KPTI_STATUS_ACTIVE;
442 }
443 
444 static int
445 cpu_procctl_la_ctl(struct proc *p, int val)
446 {
447 	int error;
448 
449 	error = 0;
450 	switch (val) {
451 	case PROC_LA_CTL_LA48_ON_EXEC:
452 		p->p_md.md_flags |= P_MD_LA48;
453 		p->p_md.md_flags &= ~P_MD_LA57;
454 		break;
455 	case PROC_LA_CTL_LA57_ON_EXEC:
456 		if (la57) {
457 			p->p_md.md_flags &= ~P_MD_LA48;
458 			p->p_md.md_flags |= P_MD_LA57;
459 		} else {
460 			error = ENOTSUP;
461 		}
462 		break;
463 	case PROC_LA_CTL_DEFAULT_ON_EXEC:
464 		p->p_md.md_flags &= ~(P_MD_LA48 | P_MD_LA57);
465 		break;
466 	}
467 	return (error);
468 }
469 
470 static void
471 cpu_procctl_la_status(struct proc *p, int *val)
472 {
473 	int res;
474 
475 	if ((p->p_md.md_flags & P_MD_LA48) != 0)
476 		res = PROC_LA_CTL_LA48_ON_EXEC;
477 	else if ((p->p_md.md_flags & P_MD_LA57) != 0)
478 		res = PROC_LA_CTL_LA57_ON_EXEC;
479 	else
480 		res = PROC_LA_CTL_DEFAULT_ON_EXEC;
481 	if (p->p_sysent->sv_maxuser == VM_MAXUSER_ADDRESS_LA48)
482 		res |= PROC_LA_STATUS_LA48;
483 	else
484 		res |= PROC_LA_STATUS_LA57;
485 	*val = res;
486 }
487 
488 int
489 cpu_procctl(struct thread *td, int idtype, id_t id, int com, void *data)
490 {
491 	struct proc *p;
492 	int error, val;
493 
494 	switch (com) {
495 	case PROC_KPTI_CTL:
496 	case PROC_KPTI_STATUS:
497 	case PROC_LA_CTL:
498 	case PROC_LA_STATUS:
499 		if (idtype != P_PID) {
500 			error = EINVAL;
501 			break;
502 		}
503 		if (com == PROC_KPTI_CTL) {
504 			/* sad but true and not a joke */
505 			error = priv_check(td, PRIV_IO);
506 			if (error != 0)
507 				break;
508 		}
509 		if (com == PROC_KPTI_CTL || com == PROC_LA_CTL) {
510 			error = copyin(data, &val, sizeof(val));
511 			if (error != 0)
512 				break;
513 		}
514 		if (com == PROC_KPTI_CTL &&
515 		    val != PROC_KPTI_CTL_ENABLE_ON_EXEC &&
516 		    val != PROC_KPTI_CTL_DISABLE_ON_EXEC) {
517 			error = EINVAL;
518 			break;
519 		}
520 		if (com == PROC_LA_CTL &&
521 		    val != PROC_LA_CTL_LA48_ON_EXEC &&
522 		    val != PROC_LA_CTL_LA57_ON_EXEC &&
523 		    val != PROC_LA_CTL_DEFAULT_ON_EXEC) {
524 			error = EINVAL;
525 			break;
526 		}
527 		error = pget(id, PGET_CANSEE | PGET_NOTWEXIT | PGET_NOTID, &p);
528 		if (error != 0)
529 			break;
530 		switch (com) {
531 		case PROC_KPTI_CTL:
532 			cpu_procctl_kpti_ctl(p, val);
533 			break;
534 		case PROC_KPTI_STATUS:
535 			cpu_procctl_kpti_status(p, &val);
536 			break;
537 		case PROC_LA_CTL:
538 			error = cpu_procctl_la_ctl(p, val);
539 			break;
540 		case PROC_LA_STATUS:
541 			cpu_procctl_la_status(p, &val);
542 			break;
543 		}
544 		PROC_UNLOCK(p);
545 		if (com == PROC_KPTI_STATUS || com == PROC_LA_STATUS)
546 			error = copyout(&val, data, sizeof(val));
547 		break;
548 	default:
549 		error = EINVAL;
550 		break;
551 	}
552 	return (error);
553 }
554 
555 void
556 cpu_set_syscall_retval(struct thread *td, int error)
557 {
558 	struct trapframe *frame;
559 
560 	frame = td->td_frame;
561 	if (__predict_true(error == 0)) {
562 		frame->tf_rax = td->td_retval[0];
563 		frame->tf_rdx = td->td_retval[1];
564 		frame->tf_rflags &= ~PSL_C;
565 		return;
566 	}
567 
568 	switch (error) {
569 	case ERESTART:
570 		/*
571 		 * Reconstruct pc, we know that 'syscall' is 2 bytes,
572 		 * lcall $X,y is 7 bytes, int 0x80 is 2 bytes.
573 		 * We saved this in tf_err.
574 		 * %r10 (which was holding the value of %rcx) is restored
575 		 * for the next iteration.
576 		 * %r10 restore is only required for freebsd/amd64 processes,
577 		 * but shall be innocent for any ia32 ABI.
578 		 *
579 		 * Require full context restore to get the arguments
580 		 * in the registers reloaded at return to usermode.
581 		 */
582 		frame->tf_rip -= frame->tf_err;
583 		frame->tf_r10 = frame->tf_rcx;
584 		set_pcb_flags(td->td_pcb, PCB_FULL_IRET);
585 		break;
586 
587 	case EJUSTRETURN:
588 		break;
589 
590 	default:
591 		frame->tf_rax = error;
592 		frame->tf_rflags |= PSL_C;
593 		break;
594 	}
595 }
596 
597 /*
598  * Initialize machine state, mostly pcb and trap frame for a new
599  * thread, about to return to userspace.  Put enough state in the new
600  * thread's PCB to get it to go back to the fork_return(), which
601  * finalizes the thread state and handles peculiarities of the first
602  * return to userspace for the new thread.
603  */
604 void
605 cpu_copy_thread(struct thread *td, struct thread *td0)
606 {
607 	copy_thread(td0, td);
608 
609 	set_pcb_flags_raw(td->td_pcb, PCB_FULL_IRET);
610 }
611 
612 /*
613  * Set that machine state for performing an upcall that starts
614  * the entry function with the given argument.
615  */
616 void
617 cpu_set_upcall(struct thread *td, void (*entry)(void *), void *arg,
618     stack_t *stack)
619 {
620 
621 	/*
622 	 * Do any extra cleaning that needs to be done.
623 	 * The thread may have optional components
624 	 * that are not present in a fresh thread.
625 	 * This may be a recycled thread so make it look
626 	 * as though it's newly allocated.
627 	 */
628 	cpu_thread_clean(td);
629 
630 #ifdef COMPAT_FREEBSD32
631 	if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
632 		/*
633 		 * Set the trap frame to point at the beginning of the entry
634 		 * function.
635 		 */
636 		td->td_frame->tf_rbp = 0;
637 		td->td_frame->tf_rsp =
638 		   (((uintptr_t)stack->ss_sp + stack->ss_size - 4) & ~0x0f) - 4;
639 		td->td_frame->tf_rip = (uintptr_t)entry;
640 
641 		/* Return address sentinel value to stop stack unwinding. */
642 		suword32((void *)td->td_frame->tf_rsp, 0);
643 
644 		/* Pass the argument to the entry point. */
645 		suword32((void *)(td->td_frame->tf_rsp + sizeof(int32_t)),
646 		    (uint32_t)(uintptr_t)arg);
647 
648 		return;
649 	}
650 #endif
651 
652 	/*
653 	 * Set the trap frame to point at the beginning of the uts
654 	 * function.
655 	 */
656 	td->td_frame->tf_rbp = 0;
657 	td->td_frame->tf_rsp =
658 	    ((register_t)stack->ss_sp + stack->ss_size) & ~0x0f;
659 	td->td_frame->tf_rsp -= 8;
660 	td->td_frame->tf_rip = (register_t)entry;
661 	td->td_frame->tf_ds = _udatasel;
662 	td->td_frame->tf_es = _udatasel;
663 	td->td_frame->tf_fs = _ufssel;
664 	td->td_frame->tf_gs = _ugssel;
665 	td->td_frame->tf_flags = TF_HASSEGS;
666 
667 	/* Return address sentinel value to stop stack unwinding. */
668 	suword((void *)td->td_frame->tf_rsp, 0);
669 
670 	/* Pass the argument to the entry point. */
671 	td->td_frame->tf_rdi = (register_t)arg;
672 }
673 
674 int
675 cpu_set_user_tls(struct thread *td, void *tls_base)
676 {
677 	struct pcb *pcb;
678 
679 	if ((u_int64_t)tls_base >= VM_MAXUSER_ADDRESS)
680 		return (EINVAL);
681 
682 	pcb = td->td_pcb;
683 	set_pcb_flags(pcb, PCB_FULL_IRET);
684 #ifdef COMPAT_FREEBSD32
685 	if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
686 		pcb->pcb_gsbase = (register_t)tls_base;
687 		return (0);
688 	}
689 #endif
690 	pcb->pcb_fsbase = (register_t)tls_base;
691 	return (0);
692 }
693