xref: /freebsd/sys/amd64/amd64/vm_machdep.c (revision 069ac184)
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  *	Utah $Hdr: vm_machdep.c 1.16.1.1 89/06/23$
41  */
42 
43 #include <sys/cdefs.h>
44 #include "opt_isa.h"
45 #include "opt_cpu.h"
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/bio.h>
50 #include <sys/buf.h>
51 #include <sys/kernel.h>
52 #include <sys/ktr.h>
53 #include <sys/lock.h>
54 #include <sys/malloc.h>
55 #include <sys/mbuf.h>
56 #include <sys/mutex.h>
57 #include <sys/priv.h>
58 #include <sys/proc.h>
59 #include <sys/procctl.h>
60 #include <sys/smp.h>
61 #include <sys/sysctl.h>
62 #include <sys/sysent.h>
63 #include <sys/unistd.h>
64 #include <sys/vnode.h>
65 #include <sys/vmmeter.h>
66 #include <sys/wait.h>
67 
68 #include <machine/cpu.h>
69 #include <machine/md_var.h>
70 #include <machine/pcb.h>
71 #include <machine/smp.h>
72 #include <machine/specialreg.h>
73 #include <machine/tss.h>
74 
75 #include <vm/vm.h>
76 #include <vm/vm_extern.h>
77 #include <vm/vm_kern.h>
78 #include <vm/vm_page.h>
79 #include <vm/vm_map.h>
80 #include <vm/vm_param.h>
81 
82 _Static_assert(OFFSETOF_MONITORBUF == offsetof(struct pcpu, pc_monitorbuf),
83     "OFFSETOF_MONITORBUF does not correspond with offset of pc_monitorbuf.");
84 
85 void
86 set_top_of_stack_td(struct thread *td)
87 {
88 	td->td_md.md_stack_base = td->td_kstack +
89 	    td->td_kstack_pages * PAGE_SIZE;
90 }
91 
92 struct savefpu *
93 get_pcb_user_save_td(struct thread *td)
94 {
95 	KASSERT(((vm_offset_t)td->td_md.md_usr_fpu_save %
96 	    XSAVE_AREA_ALIGN) == 0,
97 	    ("Unaligned pcb_user_save area ptr %p td %p",
98 	    td->td_md.md_usr_fpu_save, td));
99 	return (td->td_md.md_usr_fpu_save);
100 }
101 
102 struct pcb *
103 get_pcb_td(struct thread *td)
104 {
105 
106 	return (&td->td_md.md_pcb);
107 }
108 
109 struct savefpu *
110 get_pcb_user_save_pcb(struct pcb *pcb)
111 {
112 	struct thread *td;
113 
114 	td = __containerof(pcb, struct thread, td_md.md_pcb);
115 	return (get_pcb_user_save_td(td));
116 }
117 
118 void *
119 alloc_fpusave(int flags)
120 {
121 	void *res;
122 	struct savefpu_ymm *sf;
123 
124 	res = malloc(cpu_max_ext_state_size, M_DEVBUF, flags);
125 	if (use_xsave) {
126 		sf = (struct savefpu_ymm *)res;
127 		bzero(&sf->sv_xstate.sx_hd, sizeof(sf->sv_xstate.sx_hd));
128 		sf->sv_xstate.sx_hd.xstate_bv = xsave_mask;
129 	}
130 	return (res);
131 }
132 
133 /*
134  * Common code shared between cpu_fork() and cpu_copy_thread() for
135  * initializing a thread.
136  */
137 static void
138 copy_thread(struct thread *td1, struct thread *td2)
139 {
140 	struct pcb *pcb2;
141 
142 	pcb2 = td2->td_pcb;
143 
144 	/* Ensure that td1's pcb is up to date for user threads. */
145 	if ((td2->td_pflags & TDP_KTHREAD) == 0) {
146 		MPASS(td1 == curthread);
147 		fpuexit(td1);
148 		update_pcb_bases(td1->td_pcb);
149 	}
150 
151 	/* Copy td1's pcb */
152 	bcopy(td1->td_pcb, pcb2, sizeof(*pcb2));
153 
154 	/* Properly initialize pcb_save */
155 	pcb2->pcb_save = get_pcb_user_save_pcb(pcb2);
156 
157 	/* Kernel threads start with clean FPU and segment bases. */
158 	if ((td2->td_pflags & TDP_KTHREAD) != 0) {
159 		pcb2->pcb_fsbase = 0;
160 		pcb2->pcb_gsbase = 0;
161 		clear_pcb_flags(pcb2, PCB_FPUINITDONE | PCB_USERFPUINITDONE |
162 		    PCB_KERNFPU | PCB_KERNFPU_THR);
163 	} else {
164 		MPASS((pcb2->pcb_flags & (PCB_KERNFPU | PCB_KERNFPU_THR)) == 0);
165 		bcopy(get_pcb_user_save_td(td1), get_pcb_user_save_pcb(pcb2),
166 		    cpu_max_ext_state_size);
167 	}
168 
169 	td2->td_frame = (struct trapframe *)td2->td_md.md_stack_base - 1;
170 
171 	/*
172 	 * Set registers for trampoline to user mode.  Leave space for the
173 	 * return address on stack.  These are the kernel mode register values.
174 	 */
175 	pcb2->pcb_r12 = (register_t)fork_return;	/* fork_trampoline argument */
176 	pcb2->pcb_rbp = 0;
177 	pcb2->pcb_rsp = (register_t)td2->td_frame - sizeof(void *);
178 	pcb2->pcb_rbx = (register_t)td2;		/* fork_trampoline argument */
179 	pcb2->pcb_rip = (register_t)fork_trampoline;
180 	/*-
181 	 * pcb2->pcb_dr*:	cloned above.
182 	 * pcb2->pcb_savefpu:	cloned above.
183 	 * pcb2->pcb_flags:	cloned above.
184 	 * pcb2->pcb_onfault:	cloned above (always NULL here?).
185 	 * pcb2->pcb_[fg]sbase:	cloned above
186 	 */
187 
188 	pcb2->pcb_tssp = NULL;
189 
190 	/* Setup to release spin count in fork_exit(). */
191 	td2->td_md.md_spinlock_count = 1;
192 	td2->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
193 	pmap_thread_init_invl_gen(td2);
194 
195 	/*
196 	 * Copy the trap frame for the return to user mode as if from a syscall.
197 	 * This copies most of the user mode register values.  Some of these
198 	 * registers are rewritten by cpu_set_upcall() and linux_set_upcall().
199 	 */
200 	if ((td1->td_proc->p_flag & P_KPROC) == 0) {
201 		bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe));
202 
203 		/*
204 		 * If the current thread has the trap bit set (i.e. a debugger
205 		 * had single stepped the process to the system call), we need
206 		 * to clear the trap flag from the new frame. Otherwise, the new
207 		 * thread will receive a (likely unexpected) SIGTRAP when it
208 		 * executes the first instruction after returning to userland.
209 		 */
210 		td2->td_frame->tf_rflags &= ~PSL_T;
211 	}
212 }
213 
214 /*
215  * Finish a fork operation, with process p2 nearly set up.
216  * Copy and update the pcb, set up the stack so that the child
217  * ready to run and return to user mode.
218  */
219 void
220 cpu_fork(struct thread *td1, struct proc *p2, struct thread *td2, int flags)
221 {
222 	struct proc *p1;
223 	struct pcb *pcb2;
224 	struct mdproc *mdp1, *mdp2;
225 	struct proc_ldt *pldt;
226 
227 	p1 = td1->td_proc;
228 	if ((flags & RFPROC) == 0) {
229 		if ((flags & RFMEM) == 0) {
230 			/* unshare user LDT */
231 			mdp1 = &p1->p_md;
232 			mtx_lock(&dt_lock);
233 			if ((pldt = mdp1->md_ldt) != NULL &&
234 			    pldt->ldt_refcnt > 1 &&
235 			    user_ldt_alloc(p1, 1) == NULL)
236 				panic("could not copy LDT");
237 			mtx_unlock(&dt_lock);
238 		}
239 		return;
240 	}
241 
242 	/* Point the stack and pcb to the actual location */
243 	set_top_of_stack_td(td2);
244 	td2->td_pcb = pcb2 = get_pcb_td(td2);
245 
246 	copy_thread(td1, td2);
247 
248 	/* Reset debug registers in the new process */
249 	x86_clear_dbregs(pcb2);
250 
251 	/* Point mdproc and then copy over p1's contents */
252 	mdp2 = &p2->p_md;
253 	bcopy(&p1->p_md, mdp2, sizeof(*mdp2));
254 
255 	/* Set child return values. */
256 	p2->p_sysent->sv_set_fork_retval(td2);
257 
258 	/* As on i386, do not copy io permission bitmap. */
259 	pcb2->pcb_tssp = NULL;
260 
261 	/* New segment registers. */
262 	set_pcb_flags_raw(pcb2, PCB_FULL_IRET);
263 
264 	/* Copy the LDT, if necessary. */
265 	mdp1 = &td1->td_proc->p_md;
266 	mdp2 = &p2->p_md;
267 	if (mdp1->md_ldt == NULL) {
268 		mdp2->md_ldt = NULL;
269 		return;
270 	}
271 	mtx_lock(&dt_lock);
272 	if (mdp1->md_ldt != NULL) {
273 		if (flags & RFMEM) {
274 			mdp1->md_ldt->ldt_refcnt++;
275 			mdp2->md_ldt = mdp1->md_ldt;
276 			bcopy(&mdp1->md_ldt_sd, &mdp2->md_ldt_sd, sizeof(struct
277 			    system_segment_descriptor));
278 		} else {
279 			mdp2->md_ldt = NULL;
280 			mdp2->md_ldt = user_ldt_alloc(p2, 0);
281 			if (mdp2->md_ldt == NULL)
282 				panic("could not copy LDT");
283 			amd64_set_ldt_data(td2, 0, max_ldt_segment,
284 			    (struct user_segment_descriptor *)
285 			    mdp1->md_ldt->ldt_base);
286 		}
287 	} else
288 		mdp2->md_ldt = NULL;
289 	mtx_unlock(&dt_lock);
290 
291 	/*
292 	 * Now, cpu_switch() can schedule the new process.
293 	 * pcb_rsp is loaded pointing to the cpu_switch() stack frame
294 	 * containing the return address when exiting cpu_switch.
295 	 * This will normally be to fork_trampoline(), which will have
296 	 * %rbx loaded with the new proc's pointer.  fork_trampoline()
297 	 * will set up a stack to call fork_return(p, frame); to complete
298 	 * the return to user-mode.
299 	 */
300 }
301 
302 void
303 x86_set_fork_retval(struct thread *td)
304 {
305 	struct trapframe *frame = td->td_frame;
306 
307 	frame->tf_rax = 0;		/* Child returns zero */
308 	frame->tf_rflags &= ~PSL_C;	/* success */
309 	frame->tf_rdx = 1;		/* System V emulation */
310 }
311 
312 /*
313  * Intercept the return address from a freshly forked process that has NOT
314  * been scheduled yet.
315  *
316  * This is needed to make kernel threads stay in kernel mode.
317  */
318 void
319 cpu_fork_kthread_handler(struct thread *td, void (*func)(void *), void *arg)
320 {
321 	/*
322 	 * Note that the trap frame follows the args, so the function
323 	 * is really called like this:  func(arg, frame);
324 	 */
325 	td->td_pcb->pcb_r12 = (long) func;	/* function */
326 	td->td_pcb->pcb_rbx = (long) arg;	/* first arg */
327 }
328 
329 void
330 cpu_exit(struct thread *td)
331 {
332 
333 	/*
334 	 * If this process has a custom LDT, release it.
335 	 */
336 	if (td->td_proc->p_md.md_ldt != NULL)
337 		user_ldt_free(td);
338 }
339 
340 void
341 cpu_thread_exit(struct thread *td)
342 {
343 	struct pcb *pcb;
344 
345 	critical_enter();
346 	if (td == PCPU_GET(fpcurthread))
347 		fpudrop();
348 	critical_exit();
349 
350 	pcb = td->td_pcb;
351 
352 	/* Disable any hardware breakpoints. */
353 	if (pcb->pcb_flags & PCB_DBREGS) {
354 		reset_dbregs();
355 		clear_pcb_flags(pcb, PCB_DBREGS);
356 	}
357 }
358 
359 void
360 cpu_thread_clean(struct thread *td)
361 {
362 	struct pcb *pcb;
363 
364 	pcb = td->td_pcb;
365 
366 	/*
367 	 * Clean TSS/iomap
368 	 */
369 	if (pcb->pcb_tssp != NULL) {
370 		pmap_pti_remove_kva((vm_offset_t)pcb->pcb_tssp,
371 		    (vm_offset_t)pcb->pcb_tssp + ctob(IOPAGES + 1));
372 		kmem_free(pcb->pcb_tssp, ctob(IOPAGES + 1));
373 		pcb->pcb_tssp = NULL;
374 	}
375 }
376 
377 void
378 cpu_thread_swapin(struct thread *td)
379 {
380 }
381 
382 void
383 cpu_thread_swapout(struct thread *td)
384 {
385 }
386 
387 void
388 cpu_thread_alloc(struct thread *td)
389 {
390 	struct pcb *pcb;
391 	struct xstate_hdr *xhdr;
392 
393 	set_top_of_stack_td(td);
394 	td->td_pcb = pcb = get_pcb_td(td);
395 	td->td_frame = (struct trapframe *)td->td_md.md_stack_base - 1;
396 	td->td_md.md_usr_fpu_save = fpu_save_area_alloc();
397 	pcb->pcb_save = get_pcb_user_save_pcb(pcb);
398 	if (use_xsave) {
399 		xhdr = (struct xstate_hdr *)(pcb->pcb_save + 1);
400 		bzero(xhdr, sizeof(*xhdr));
401 		xhdr->xstate_bv = xsave_mask;
402 	}
403 }
404 
405 void
406 cpu_thread_free(struct thread *td)
407 {
408 	cpu_thread_clean(td);
409 
410 	fpu_save_area_free(td->td_md.md_usr_fpu_save);
411 	td->td_md.md_usr_fpu_save = NULL;
412 }
413 
414 bool
415 cpu_exec_vmspace_reuse(struct proc *p, vm_map_t map)
416 {
417 
418 	return (((curproc->p_md.md_flags & P_MD_KPTI) != 0) ==
419 	    (vm_map_pmap(map)->pm_ucr3 != PMAP_NO_CR3));
420 }
421 
422 static void
423 cpu_procctl_kpti_ctl(struct proc *p, int val)
424 {
425 
426 	if (pti && val == PROC_KPTI_CTL_ENABLE_ON_EXEC)
427 		p->p_md.md_flags |= P_MD_KPTI;
428 	if (val == PROC_KPTI_CTL_DISABLE_ON_EXEC)
429 		p->p_md.md_flags &= ~P_MD_KPTI;
430 }
431 
432 static void
433 cpu_procctl_kpti_status(struct proc *p, int *val)
434 {
435 	*val = (p->p_md.md_flags & P_MD_KPTI) != 0 ?
436 	    PROC_KPTI_CTL_ENABLE_ON_EXEC:
437 	    PROC_KPTI_CTL_DISABLE_ON_EXEC;
438 	if (vmspace_pmap(p->p_vmspace)->pm_ucr3 != PMAP_NO_CR3)
439 		*val |= PROC_KPTI_STATUS_ACTIVE;
440 }
441 
442 static int
443 cpu_procctl_la_ctl(struct proc *p, int val)
444 {
445 	int error;
446 
447 	error = 0;
448 	switch (val) {
449 	case PROC_LA_CTL_LA48_ON_EXEC:
450 		p->p_md.md_flags |= P_MD_LA48;
451 		p->p_md.md_flags &= ~P_MD_LA57;
452 		break;
453 	case PROC_LA_CTL_LA57_ON_EXEC:
454 		if (la57) {
455 			p->p_md.md_flags &= ~P_MD_LA48;
456 			p->p_md.md_flags |= P_MD_LA57;
457 		} else {
458 			error = ENOTSUP;
459 		}
460 		break;
461 	case PROC_LA_CTL_DEFAULT_ON_EXEC:
462 		p->p_md.md_flags &= ~(P_MD_LA48 | P_MD_LA57);
463 		break;
464 	}
465 	return (error);
466 }
467 
468 static void
469 cpu_procctl_la_status(struct proc *p, int *val)
470 {
471 	int res;
472 
473 	if ((p->p_md.md_flags & P_MD_LA48) != 0)
474 		res = PROC_LA_CTL_LA48_ON_EXEC;
475 	else if ((p->p_md.md_flags & P_MD_LA57) != 0)
476 		res = PROC_LA_CTL_LA57_ON_EXEC;
477 	else
478 		res = PROC_LA_CTL_DEFAULT_ON_EXEC;
479 	if (p->p_sysent->sv_maxuser == VM_MAXUSER_ADDRESS_LA48)
480 		res |= PROC_LA_STATUS_LA48;
481 	else
482 		res |= PROC_LA_STATUS_LA57;
483 	*val = res;
484 }
485 
486 int
487 cpu_procctl(struct thread *td, int idtype, id_t id, int com, void *data)
488 {
489 	struct proc *p;
490 	int error, val;
491 
492 	switch (com) {
493 	case PROC_KPTI_CTL:
494 	case PROC_KPTI_STATUS:
495 	case PROC_LA_CTL:
496 	case PROC_LA_STATUS:
497 		if (idtype != P_PID) {
498 			error = EINVAL;
499 			break;
500 		}
501 		if (com == PROC_KPTI_CTL) {
502 			/* sad but true and not a joke */
503 			error = priv_check(td, PRIV_IO);
504 			if (error != 0)
505 				break;
506 		}
507 		if (com == PROC_KPTI_CTL || com == PROC_LA_CTL) {
508 			error = copyin(data, &val, sizeof(val));
509 			if (error != 0)
510 				break;
511 		}
512 		if (com == PROC_KPTI_CTL &&
513 		    val != PROC_KPTI_CTL_ENABLE_ON_EXEC &&
514 		    val != PROC_KPTI_CTL_DISABLE_ON_EXEC) {
515 			error = EINVAL;
516 			break;
517 		}
518 		if (com == PROC_LA_CTL &&
519 		    val != PROC_LA_CTL_LA48_ON_EXEC &&
520 		    val != PROC_LA_CTL_LA57_ON_EXEC &&
521 		    val != PROC_LA_CTL_DEFAULT_ON_EXEC) {
522 			error = EINVAL;
523 			break;
524 		}
525 		error = pget(id, PGET_CANSEE | PGET_NOTWEXIT | PGET_NOTID, &p);
526 		if (error != 0)
527 			break;
528 		switch (com) {
529 		case PROC_KPTI_CTL:
530 			cpu_procctl_kpti_ctl(p, val);
531 			break;
532 		case PROC_KPTI_STATUS:
533 			cpu_procctl_kpti_status(p, &val);
534 			break;
535 		case PROC_LA_CTL:
536 			error = cpu_procctl_la_ctl(p, val);
537 			break;
538 		case PROC_LA_STATUS:
539 			cpu_procctl_la_status(p, &val);
540 			break;
541 		}
542 		PROC_UNLOCK(p);
543 		if (com == PROC_KPTI_STATUS || com == PROC_LA_STATUS)
544 			error = copyout(&val, data, sizeof(val));
545 		break;
546 	default:
547 		error = EINVAL;
548 		break;
549 	}
550 	return (error);
551 }
552 
553 void
554 cpu_set_syscall_retval(struct thread *td, int error)
555 {
556 	struct trapframe *frame;
557 
558 	frame = td->td_frame;
559 	if (__predict_true(error == 0)) {
560 		frame->tf_rax = td->td_retval[0];
561 		frame->tf_rdx = td->td_retval[1];
562 		frame->tf_rflags &= ~PSL_C;
563 		return;
564 	}
565 
566 	switch (error) {
567 	case ERESTART:
568 		/*
569 		 * Reconstruct pc, we know that 'syscall' is 2 bytes,
570 		 * lcall $X,y is 7 bytes, int 0x80 is 2 bytes.
571 		 * We saved this in tf_err.
572 		 * %r10 (which was holding the value of %rcx) is restored
573 		 * for the next iteration.
574 		 * %r10 restore is only required for freebsd/amd64 processes,
575 		 * but shall be innocent for any ia32 ABI.
576 		 *
577 		 * Require full context restore to get the arguments
578 		 * in the registers reloaded at return to usermode.
579 		 */
580 		frame->tf_rip -= frame->tf_err;
581 		frame->tf_r10 = frame->tf_rcx;
582 		set_pcb_flags(td->td_pcb, PCB_FULL_IRET);
583 		break;
584 
585 	case EJUSTRETURN:
586 		break;
587 
588 	default:
589 		frame->tf_rax = error;
590 		frame->tf_rflags |= PSL_C;
591 		break;
592 	}
593 }
594 
595 /*
596  * Initialize machine state, mostly pcb and trap frame for a new
597  * thread, about to return to userspace.  Put enough state in the new
598  * thread's PCB to get it to go back to the fork_return(), which
599  * finalizes the thread state and handles peculiarities of the first
600  * return to userspace for the new thread.
601  */
602 void
603 cpu_copy_thread(struct thread *td, struct thread *td0)
604 {
605 	copy_thread(td0, td);
606 
607 	set_pcb_flags_raw(td->td_pcb, PCB_FULL_IRET);
608 }
609 
610 /*
611  * Set that machine state for performing an upcall that starts
612  * the entry function with the given argument.
613  */
614 int
615 cpu_set_upcall(struct thread *td, void (*entry)(void *), void *arg,
616     stack_t *stack)
617 {
618 
619 	/*
620 	 * Do any extra cleaning that needs to be done.
621 	 * The thread may have optional components
622 	 * that are not present in a fresh thread.
623 	 * This may be a recycled thread so make it look
624 	 * as though it's newly allocated.
625 	 */
626 	cpu_thread_clean(td);
627 
628 #ifdef COMPAT_FREEBSD32
629 	if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
630 		/*
631 		 * Set the trap frame to point at the beginning of the entry
632 		 * function.
633 		 */
634 		td->td_frame->tf_rbp = 0;
635 		td->td_frame->tf_rsp =
636 		   (((uintptr_t)stack->ss_sp + stack->ss_size - 4) & ~0x0f) - 4;
637 		td->td_frame->tf_rip = (uintptr_t)entry;
638 
639 		/* Return address sentinel value to stop stack unwinding. */
640 		if (suword32((void *)td->td_frame->tf_rsp, 0) != 0)
641 			return (EFAULT);
642 
643 		/* Pass the argument to the entry point. */
644 		if (suword32(
645 		    (void *)(td->td_frame->tf_rsp + sizeof(int32_t)),
646 		    (uint32_t)(uintptr_t)arg) != 0)
647 			return (EFAULT);
648 		return (0);
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 	if (suword((void *)td->td_frame->tf_rsp, 0) != 0)
669 		return (EFAULT);
670 
671 	/* Pass the argument to the entry point. */
672 	td->td_frame->tf_rdi = (register_t)arg;
673 
674 	return (0);
675 }
676 
677 int
678 cpu_set_user_tls(struct thread *td, void *tls_base)
679 {
680 	struct pcb *pcb;
681 
682 	if ((u_int64_t)tls_base >= VM_MAXUSER_ADDRESS)
683 		return (EINVAL);
684 
685 	pcb = td->td_pcb;
686 	set_pcb_flags(pcb, PCB_FULL_IRET);
687 #ifdef COMPAT_FREEBSD32
688 	if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
689 		pcb->pcb_gsbase = (register_t)tls_base;
690 		return (0);
691 	}
692 #endif
693 	pcb->pcb_fsbase = (register_t)tls_base;
694 	return (0);
695 }
696