xref: /freebsd/sys/i386/i386/vm_machdep.c (revision 1d386b48a555f61cb7325543adbbb5c3f3407a66)
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 #include "opt_isa.h"
47 #include "opt_npx.h"
48 #include "opt_reset.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/proc.h>
62 #include <sys/sysent.h>
63 #include <sys/sf_buf.h>
64 #include <sys/smp.h>
65 #include <sys/sched.h>
66 #include <sys/sysctl.h>
67 #include <sys/unistd.h>
68 #include <sys/vnode.h>
69 #include <sys/vmmeter.h>
70 
71 #include <machine/cpu.h>
72 #include <machine/cputypes.h>
73 #include <machine/md_var.h>
74 #include <machine/pcb.h>
75 #include <machine/pcb_ext.h>
76 #include <machine/smp.h>
77 #include <machine/vm86.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 union savefpu *
90 get_pcb_user_save_td(struct thread *td)
91 {
92 	vm_offset_t p;
93 
94 	p = td->td_kstack + td->td_kstack_pages * PAGE_SIZE -
95 	    roundup2(cpu_max_ext_state_size, XSAVE_AREA_ALIGN);
96 	KASSERT((p % XSAVE_AREA_ALIGN) == 0, ("Unaligned pcb_user_save area"));
97 	return ((union savefpu *)p);
98 }
99 
100 union savefpu *
101 get_pcb_user_save_pcb(struct pcb *pcb)
102 {
103 	vm_offset_t p;
104 
105 	p = (vm_offset_t)(pcb + 1);
106 	return ((union savefpu *)p);
107 }
108 
109 struct pcb *
110 get_pcb_td(struct thread *td)
111 {
112 	vm_offset_t p;
113 
114 	p = td->td_kstack + td->td_kstack_pages * PAGE_SIZE -
115 	    roundup2(cpu_max_ext_state_size, XSAVE_AREA_ALIGN) -
116 	    sizeof(struct pcb);
117 	return ((struct pcb *)p);
118 }
119 
120 void *
121 alloc_fpusave(int flags)
122 {
123 	void *res;
124 	struct savefpu_ymm *sf;
125 
126 	res = malloc(cpu_max_ext_state_size, M_DEVBUF, flags);
127 	if (use_xsave) {
128 		sf = (struct savefpu_ymm *)res;
129 		bzero(&sf->sv_xstate.sx_hd, sizeof(sf->sv_xstate.sx_hd));
130 		sf->sv_xstate.sx_hd.xstate_bv = xsave_mask;
131 	}
132 	return (res);
133 }
134 
135 /*
136  * Common code shared between cpu_fork() and cpu_copy_thread() for
137  * initializing a thread.
138  */
139 static void
140 copy_thread(struct thread *td1, struct thread *td2)
141 {
142 	struct pcb *pcb2;
143 
144 	pcb2 = td2->td_pcb;
145 
146 	/* Ensure that td1's pcb is up to date for user threads. */
147 	if ((td2->td_pflags & TDP_KTHREAD) == 0) {
148 		MPASS(td1 == curthread);
149 		td1->td_pcb->pcb_gs = rgs();
150 		critical_enter();
151 		if (PCPU_GET(fpcurthread) == td1)
152 			npxsave(td1->td_pcb->pcb_save);
153 		critical_exit();
154 	}
155 
156 	/* Copy td1's pcb */
157 	bcopy(td1->td_pcb, pcb2, sizeof(*pcb2));
158 
159 	/* Properly initialize pcb_save */
160 	pcb2->pcb_save = get_pcb_user_save_pcb(pcb2);
161 
162 	/* Kernel threads start with clean NPX and segment bases. */
163 	if ((td2->td_pflags & TDP_KTHREAD) != 0) {
164 		pcb2->pcb_gs = _udatasel;
165 		set_fsbase(td2, 0);
166 		set_gsbase(td2, 0);
167 		pcb2->pcb_flags &= ~(PCB_NPXINITDONE | PCB_NPXUSERINITDONE |
168 		    PCB_KERNNPX | PCB_KERNNPX_THR);
169 	} else {
170 		MPASS((pcb2->pcb_flags & (PCB_KERNNPX | PCB_KERNNPX_THR)) == 0);
171 		bcopy(get_pcb_user_save_td(td1), get_pcb_user_save_pcb(pcb2),
172 		    cpu_max_ext_state_size);
173 	}
174 
175 	/*
176 	 * Set registers for trampoline to user mode.  Leave space for the
177 	 * return address on stack.  These are the kernel mode register values.
178 	 */
179 	pcb2->pcb_edi = 0;
180 	pcb2->pcb_esi = (int)fork_return;		    /* trampoline arg */
181 	pcb2->pcb_ebp = 0;
182 	pcb2->pcb_esp = (int)td2->td_frame - sizeof(void *); /* trampoline arg */
183 	pcb2->pcb_ebx = (int)td2;			    /* trampoline arg */
184 	pcb2->pcb_eip = (int)fork_trampoline + setidt_disp;
185 	/*
186 	 * If we didn't copy the pcb, we'd need to do the following registers:
187 	 * pcb2->pcb_cr3:	cloned above.
188 	 * pcb2->pcb_dr*:	cloned above.
189 	 * pcb2->pcb_savefpu:	cloned above.
190 	 * pcb2->pcb_flags:	cloned above.
191 	 * pcb2->pcb_onfault:	cloned above (always NULL here?).
192 	 * pcb2->pcb_gs:	cloned above.
193 	 * pcb2->pcb_ext:	cleared below.
194 	 */
195 	pcb2->pcb_ext = NULL;
196 
197 	/* Setup to release spin count in fork_exit(). */
198 	td2->td_md.md_spinlock_count = 1;
199 	td2->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
200 }
201 
202 /*
203  * Finish a fork operation, with process p2 nearly set up.
204  * Copy and update the pcb, set up the stack so that the child
205  * ready to run and return to user mode.
206  */
207 void
208 cpu_fork(struct thread *td1, struct proc *p2, struct thread *td2, int flags)
209 {
210 	struct proc *p1;
211 	struct pcb *pcb2;
212 	struct mdproc *mdp2;
213 
214 	p1 = td1->td_proc;
215 	if ((flags & RFPROC) == 0) {
216 		if ((flags & RFMEM) == 0) {
217 			/* unshare user LDT */
218 			struct mdproc *mdp1 = &p1->p_md;
219 			struct proc_ldt *pldt, *pldt1;
220 
221 			mtx_lock_spin(&dt_lock);
222 			if ((pldt1 = mdp1->md_ldt) != NULL &&
223 			    pldt1->ldt_refcnt > 1) {
224 				pldt = user_ldt_alloc(mdp1, pldt1->ldt_len);
225 				if (pldt == NULL)
226 					panic("could not copy LDT");
227 				mdp1->md_ldt = pldt;
228 				set_user_ldt(mdp1);
229 				user_ldt_deref(pldt1);
230 			} else
231 				mtx_unlock_spin(&dt_lock);
232 		}
233 		return;
234 	}
235 
236 	/* Point the pcb to the top of the stack */
237 	pcb2 = get_pcb_td(td2);
238 	td2->td_pcb = pcb2;
239 
240 	copy_thread(td1, td2);
241 
242 	/* Reset debug registers in the new process */
243 	x86_clear_dbregs(pcb2);
244 
245 	/* Point mdproc and then copy over td1's contents */
246 	mdp2 = &p2->p_md;
247 	bcopy(&p1->p_md, mdp2, sizeof(*mdp2));
248 
249 	/*
250 	 * Copy the trap frame for the return to user mode as if from a
251 	 * syscall.  This copies most of the user mode register values.
252 	 * The -VM86_STACK_SPACE (-16) is so we can expand the trapframe
253 	 * if we go to vm86.
254 	 */
255 	td2->td_frame = (struct trapframe *)((caddr_t)td2->td_pcb -
256 	    VM86_STACK_SPACE) - 1;
257 	bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe));
258 
259 	/* Set child return values. */
260 	p2->p_sysent->sv_set_fork_retval(td2);
261 
262 	/*
263 	 * If the parent process has the trap bit set (i.e. a debugger
264 	 * had single stepped the process to the system call), we need
265 	 * to clear the trap flag from the new frame.
266 	 */
267 	td2->td_frame->tf_eflags &= ~PSL_T;
268 
269 	/* Set cr3 for the new process. */
270 	pcb2->pcb_cr3 = pmap_get_cr3(vmspace_pmap(p2->p_vmspace));
271 
272 	/*
273 	 * XXX don't copy the i/o pages.  this should probably be fixed.
274 	 */
275 	pcb2->pcb_ext = NULL;
276 
277 	/* Copy the LDT, if necessary. */
278 	mtx_lock_spin(&dt_lock);
279 	if (mdp2->md_ldt != NULL) {
280 		if (flags & RFMEM) {
281 			mdp2->md_ldt->ldt_refcnt++;
282 		} else {
283 			mdp2->md_ldt = user_ldt_alloc(mdp2,
284 			    mdp2->md_ldt->ldt_len);
285 			if (mdp2->md_ldt == NULL)
286 				panic("could not copy LDT");
287 		}
288 	}
289 	mtx_unlock_spin(&dt_lock);
290 
291 	/*
292 	 * Now, cpu_switch() can schedule the new process.
293 	 * pcb_esp 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 	 * %ebx 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_eax = 0;		/* Child returns zero */
308 	frame->tf_eflags &= ~PSL_C;	/* success */
309 	frame->tf_edx = 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_esi = (int) func;	/* function */
326 	td->td_pcb->pcb_ebx = (int) 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.  Reset pc->pcb_gs
335 	 * and %gs before we free it in case they refer to an LDT entry.
336 	 */
337 	mtx_lock_spin(&dt_lock);
338 	if (td->td_proc->p_md.md_ldt) {
339 		td->td_pcb->pcb_gs = _udatasel;
340 		load_gs(_udatasel);
341 		user_ldt_free(td);
342 	} else
343 		mtx_unlock_spin(&dt_lock);
344 }
345 
346 void
347 cpu_thread_exit(struct thread *td)
348 {
349 
350 	critical_enter();
351 	if (td == PCPU_GET(fpcurthread))
352 		npxdrop();
353 	critical_exit();
354 
355 	/* Disable any hardware breakpoints. */
356 	if (td->td_pcb->pcb_flags & PCB_DBREGS) {
357 		reset_dbregs();
358 		td->td_pcb->pcb_flags &= ~PCB_DBREGS;
359 	}
360 }
361 
362 void
363 cpu_thread_clean(struct thread *td)
364 {
365 	struct pcb *pcb;
366 
367 	pcb = td->td_pcb;
368 	if (pcb->pcb_ext != NULL) {
369 		/* if (pcb->pcb_ext->ext_refcount-- == 1) ?? */
370 		/*
371 		 * XXX do we need to move the TSS off the allocated pages
372 		 * before freeing them?  (not done here)
373 		 */
374 		pmap_trm_free(pcb->pcb_ext, ctob(IOPAGES + 1));
375 		pcb->pcb_ext = 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 	td->td_pcb = pcb = get_pcb_td(td);
396 	td->td_frame = (struct trapframe *)((caddr_t)pcb -
397 	    VM86_STACK_SPACE) - 1;
398 	pcb->pcb_ext = NULL;
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 
411 	cpu_thread_clean(td);
412 }
413 
414 bool
415 cpu_exec_vmspace_reuse(struct proc *p __unused, vm_map_t map __unused)
416 {
417 
418 	return (true);
419 }
420 
421 int
422 cpu_procctl(struct thread *td __unused, int idtype __unused, id_t id __unused,
423     int com __unused, void *data __unused)
424 {
425 
426 	return (EINVAL);
427 }
428 
429 void
430 cpu_set_syscall_retval(struct thread *td, int error)
431 {
432 
433 	switch (error) {
434 	case 0:
435 		td->td_frame->tf_eax = td->td_retval[0];
436 		td->td_frame->tf_edx = td->td_retval[1];
437 		td->td_frame->tf_eflags &= ~PSL_C;
438 		break;
439 
440 	case ERESTART:
441 		/*
442 		 * Reconstruct pc, assuming lcall $X,y is 7 bytes, int
443 		 * 0x80 is 2 bytes. We saved this in tf_err.
444 		 */
445 		td->td_frame->tf_eip -= td->td_frame->tf_err;
446 		break;
447 
448 	case EJUSTRETURN:
449 		break;
450 
451 	default:
452 		td->td_frame->tf_eax = error;
453 		td->td_frame->tf_eflags |= PSL_C;
454 		break;
455 	}
456 }
457 
458 /*
459  * Initialize machine state, mostly pcb and trap frame for a new
460  * thread, about to return to userspace.  Put enough state in the new
461  * thread's PCB to get it to go back to the fork_return(), which
462  * finalizes the thread state and handles peculiarities of the first
463  * return to userspace for the new thread.
464  */
465 void
466 cpu_copy_thread(struct thread *td, struct thread *td0)
467 {
468 	copy_thread(td0, td);
469 
470 	/*
471 	 * Copy user general-purpose registers.
472 	 *
473 	 * Some of these registers are rewritten by cpu_set_upcall()
474 	 * and linux_set_upcall().
475 	 */
476 	bcopy(td0->td_frame, td->td_frame, sizeof(struct trapframe));
477 
478 	/* If the current thread has the trap bit set (i.e. a debugger had
479 	 * single stepped the process to the system call), we need to clear
480 	 * the trap flag from the new frame. Otherwise, the new thread will
481 	 * receive a (likely unexpected) SIGTRAP when it executes the first
482 	 * instruction after returning to userland.
483 	 */
484 	td->td_frame->tf_eflags &= ~PSL_T;
485 }
486 
487 /*
488  * Set that machine state for performing an upcall that starts
489  * the entry function with the given argument.
490  */
491 void
492 cpu_set_upcall(struct thread *td, void (*entry)(void *), void *arg,
493     stack_t *stack)
494 {
495 
496 	/*
497 	 * Do any extra cleaning that needs to be done.
498 	 * The thread may have optional components
499 	 * that are not present in a fresh thread.
500 	 * This may be a recycled thread so make it look
501 	 * as though it's newly allocated.
502 	 */
503 	cpu_thread_clean(td);
504 
505 	/*
506 	 * Set the trap frame to point at the beginning of the entry
507 	 * function.
508 	 */
509 	td->td_frame->tf_ebp = 0;
510 	td->td_frame->tf_esp =
511 	    (((int)stack->ss_sp + stack->ss_size - 4) & ~0x0f) - 4;
512 	td->td_frame->tf_eip = (int)entry;
513 
514 	/* Return address sentinel value to stop stack unwinding. */
515 	suword((void *)td->td_frame->tf_esp, 0);
516 
517 	/* Pass the argument to the entry point. */
518 	suword((void *)(td->td_frame->tf_esp + sizeof(void *)),
519 	    (int)arg);
520 }
521 
522 int
523 cpu_set_user_tls(struct thread *td, void *tls_base)
524 {
525 	struct segment_descriptor sd;
526 	uint32_t base;
527 
528 	/*
529 	 * Construct a descriptor and store it in the pcb for
530 	 * the next context switch.  Also store it in the gdt
531 	 * so that the load of tf_fs into %fs will activate it
532 	 * at return to userland.
533 	 */
534 	base = (uint32_t)tls_base;
535 	sd.sd_lobase = base & 0xffffff;
536 	sd.sd_hibase = (base >> 24) & 0xff;
537 	sd.sd_lolimit = 0xffff;	/* 4GB limit, wraps around */
538 	sd.sd_hilimit = 0xf;
539 	sd.sd_type  = SDT_MEMRWA;
540 	sd.sd_dpl   = SEL_UPL;
541 	sd.sd_p     = 1;
542 	sd.sd_xx    = 0;
543 	sd.sd_def32 = 1;
544 	sd.sd_gran  = 1;
545 	critical_enter();
546 	/* set %gs */
547 	td->td_pcb->pcb_gsd = sd;
548 	if (td == curthread) {
549 		PCPU_GET(fsgs_gdt)[1] = sd;
550 		load_gs(GSEL(GUGS_SEL, SEL_UPL));
551 	}
552 	critical_exit();
553 	return (0);
554 }
555 
556 /*
557  * Convert kernel VA to physical address
558  */
559 vm_paddr_t
560 kvtop(void *addr)
561 {
562 	vm_paddr_t pa;
563 
564 	pa = pmap_kextract((vm_offset_t)addr);
565 	if (pa == 0)
566 		panic("kvtop: zero page frame");
567 	return (pa);
568 }
569 
570 /*
571  * Get an sf_buf from the freelist.  May block if none are available.
572  */
573 void
574 sf_buf_map(struct sf_buf *sf, int flags)
575 {
576 
577 	pmap_sf_buf_map(sf);
578 #ifdef SMP
579 	sf_buf_shootdown(sf, flags);
580 #endif
581 }
582 
583 #ifdef SMP
584 static void
585 sf_buf_shootdown_curcpu_cb(pmap_t pmap __unused,
586     vm_offset_t addr1 __unused, vm_offset_t addr2 __unused)
587 {
588 }
589 
590 void
591 sf_buf_shootdown(struct sf_buf *sf, int flags)
592 {
593 	cpuset_t other_cpus;
594 	u_int cpuid;
595 
596 	sched_pin();
597 	cpuid = PCPU_GET(cpuid);
598 	if (!CPU_ISSET(cpuid, &sf->cpumask)) {
599 		CPU_SET(cpuid, &sf->cpumask);
600 		invlpg(sf->kva);
601 	}
602 	if ((flags & SFB_CPUPRIVATE) == 0) {
603 		other_cpus = all_cpus;
604 		CPU_CLR(cpuid, &other_cpus);
605 		CPU_ANDNOT(&other_cpus, &other_cpus, &sf->cpumask);
606 		if (!CPU_EMPTY(&other_cpus)) {
607 			CPU_OR(&sf->cpumask, &sf->cpumask, &other_cpus);
608 			smp_masked_invlpg(other_cpus, sf->kva, kernel_pmap,
609 			    sf_buf_shootdown_curcpu_cb);
610 		}
611 	}
612 	sched_unpin();
613 }
614 #endif
615 
616 /*
617  * MD part of sf_buf_free().
618  */
619 int
620 sf_buf_unmap(struct sf_buf *sf)
621 {
622 
623 	return (0);
624 }
625 
626 static void
627 sf_buf_invalidate(struct sf_buf *sf)
628 {
629 	vm_page_t m = sf->m;
630 
631 	/*
632 	 * Use pmap_qenter to update the pte for
633 	 * existing mapping, in particular, the PAT
634 	 * settings are recalculated.
635 	 */
636 	pmap_qenter(sf->kva, &m, 1);
637 	pmap_invalidate_cache_range(sf->kva, sf->kva + PAGE_SIZE);
638 }
639 
640 /*
641  * Invalidate the cache lines that may belong to the page, if
642  * (possibly old) mapping of the page by sf buffer exists.  Returns
643  * TRUE when mapping was found and cache invalidated.
644  */
645 boolean_t
646 sf_buf_invalidate_cache(vm_page_t m)
647 {
648 
649 	return (sf_buf_process_page(m, sf_buf_invalidate));
650 }
651