xref: /freebsd/sys/i386/i386/vm_machdep.c (revision ba682f8b9b58f6b497a5015ef70b2120b8d96663)
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_npx.h"
46 #include "opt_reset.h"
47 #include "opt_cpu.h"
48 
49 #include <sys/param.h>
50 #include <sys/systm.h>
51 #include <sys/bio.h>
52 #include <sys/buf.h>
53 #include <sys/kernel.h>
54 #include <sys/ktr.h>
55 #include <sys/lock.h>
56 #include <sys/malloc.h>
57 #include <sys/mbuf.h>
58 #include <sys/mutex.h>
59 #include <sys/proc.h>
60 #include <sys/sysent.h>
61 #include <sys/sf_buf.h>
62 #include <sys/smp.h>
63 #include <sys/sched.h>
64 #include <sys/sysctl.h>
65 #include <sys/unistd.h>
66 #include <sys/vnode.h>
67 #include <sys/vmmeter.h>
68 
69 #include <machine/cpu.h>
70 #include <machine/cputypes.h>
71 #include <machine/md_var.h>
72 #include <machine/pcb.h>
73 #include <machine/pcb_ext.h>
74 #include <machine/smp.h>
75 #include <machine/vm86.h>
76 
77 #include <vm/vm.h>
78 #include <vm/vm_extern.h>
79 #include <vm/vm_kern.h>
80 #include <vm/vm_page.h>
81 #include <vm/vm_map.h>
82 #include <vm/vm_param.h>
83 
84 _Static_assert(__OFFSETOF_MONITORBUF == offsetof(struct pcpu, pc_monitorbuf),
85     "__OFFSETOF_MONITORBUF does not correspond with offset of pc_monitorbuf.");
86 
87 union savefpu *
88 get_pcb_user_save_td(struct thread *td)
89 {
90 	vm_offset_t p;
91 
92 	p = td->td_kstack + td->td_kstack_pages * PAGE_SIZE -
93 	    roundup2(cpu_max_ext_state_size, XSAVE_AREA_ALIGN);
94 	KASSERT((p % XSAVE_AREA_ALIGN) == 0, ("Unaligned pcb_user_save area"));
95 	return ((union savefpu *)p);
96 }
97 
98 union savefpu *
99 get_pcb_user_save_pcb(struct pcb *pcb)
100 {
101 	vm_offset_t p;
102 
103 	p = (vm_offset_t)(pcb + 1);
104 	return ((union savefpu *)p);
105 }
106 
107 struct pcb *
108 get_pcb_td(struct thread *td)
109 {
110 	vm_offset_t p;
111 
112 	p = td->td_kstack + td->td_kstack_pages * PAGE_SIZE -
113 	    roundup2(cpu_max_ext_state_size, XSAVE_AREA_ALIGN) -
114 	    sizeof(struct pcb);
115 	return ((struct pcb *)p);
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 		td1->td_pcb->pcb_gs = rgs();
148 		critical_enter();
149 		if (PCPU_GET(fpcurthread) == td1)
150 			npxsave(td1->td_pcb->pcb_save);
151 		critical_exit();
152 	}
153 
154 	/* Copy td1's pcb */
155 	bcopy(td1->td_pcb, pcb2, sizeof(*pcb2));
156 
157 	/* Properly initialize pcb_save */
158 	pcb2->pcb_save = get_pcb_user_save_pcb(pcb2);
159 
160 	/* Kernel threads start with clean NPX and segment bases. */
161 	if ((td2->td_pflags & TDP_KTHREAD) != 0) {
162 		pcb2->pcb_gs = _udatasel;
163 		set_fsbase(td2, 0);
164 		set_gsbase(td2, 0);
165 		pcb2->pcb_flags &= ~(PCB_NPXINITDONE | PCB_NPXUSERINITDONE |
166 		    PCB_KERNNPX | PCB_KERNNPX_THR);
167 	} else {
168 		MPASS((pcb2->pcb_flags & (PCB_KERNNPX | PCB_KERNNPX_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_edi = 0;
178 	pcb2->pcb_esi = (int)fork_return;		    /* trampoline arg */
179 	pcb2->pcb_ebp = 0;
180 	pcb2->pcb_esp = (int)td2->td_frame - sizeof(void *); /* trampoline arg */
181 	pcb2->pcb_ebx = (int)td2;			    /* trampoline arg */
182 	pcb2->pcb_eip = (int)fork_trampoline + setidt_disp;
183 	/*
184 	 * If we didn't copy the pcb, we'd need to do the following registers:
185 	 * pcb2->pcb_cr3:	cloned above.
186 	 * pcb2->pcb_dr*:	cloned above.
187 	 * pcb2->pcb_savefpu:	cloned above.
188 	 * pcb2->pcb_flags:	cloned above.
189 	 * pcb2->pcb_onfault:	cloned above (always NULL here?).
190 	 * pcb2->pcb_gs:	cloned above.
191 	 * pcb2->pcb_ext:	cleared below.
192 	 */
193 	pcb2->pcb_ext = NULL;
194 
195 	/* Setup to release spin count in fork_exit(). */
196 	td2->td_md.md_spinlock_count = 1;
197 	td2->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
198 }
199 
200 /*
201  * Finish a fork operation, with process p2 nearly set up.
202  * Copy and update the pcb, set up the stack so that the child
203  * ready to run and return to user mode.
204  */
205 void
206 cpu_fork(struct thread *td1, struct proc *p2, struct thread *td2, int flags)
207 {
208 	struct proc *p1;
209 	struct pcb *pcb2;
210 	struct mdproc *mdp2;
211 
212 	p1 = td1->td_proc;
213 	if ((flags & RFPROC) == 0) {
214 		if ((flags & RFMEM) == 0) {
215 			/* unshare user LDT */
216 			struct mdproc *mdp1 = &p1->p_md;
217 			struct proc_ldt *pldt, *pldt1;
218 
219 			mtx_lock_spin(&dt_lock);
220 			if ((pldt1 = mdp1->md_ldt) != NULL &&
221 			    pldt1->ldt_refcnt > 1) {
222 				pldt = user_ldt_alloc(mdp1, pldt1->ldt_len);
223 				if (pldt == NULL)
224 					panic("could not copy LDT");
225 				mdp1->md_ldt = pldt;
226 				set_user_ldt(mdp1);
227 				user_ldt_deref(pldt1);
228 			} else
229 				mtx_unlock_spin(&dt_lock);
230 		}
231 		return;
232 	}
233 
234 	/* Point the pcb to the top of the stack */
235 	pcb2 = get_pcb_td(td2);
236 	td2->td_pcb = pcb2;
237 
238 	copy_thread(td1, td2);
239 
240 	/* Reset debug registers in the new process */
241 	x86_clear_dbregs(pcb2);
242 
243 	/* Point mdproc and then copy over td1's contents */
244 	mdp2 = &p2->p_md;
245 	bcopy(&p1->p_md, mdp2, sizeof(*mdp2));
246 
247 	/*
248 	 * Copy the trap frame for the return to user mode as if from a
249 	 * syscall.  This copies most of the user mode register values.
250 	 * The -VM86_STACK_SPACE (-16) is so we can expand the trapframe
251 	 * if we go to vm86.
252 	 */
253 	td2->td_frame = (struct trapframe *)((caddr_t)td2->td_pcb -
254 	    VM86_STACK_SPACE) - 1;
255 	bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe));
256 
257 	/* Set child return values. */
258 	p2->p_sysent->sv_set_fork_retval(td2);
259 
260 	/*
261 	 * If the parent process has the trap bit set (i.e. a debugger
262 	 * had single stepped the process to the system call), we need
263 	 * to clear the trap flag from the new frame.
264 	 */
265 	td2->td_frame->tf_eflags &= ~PSL_T;
266 
267 	/* Set cr3 for the new process. */
268 	pcb2->pcb_cr3 = pmap_get_cr3(vmspace_pmap(p2->p_vmspace));
269 
270 	/*
271 	 * XXX don't copy the i/o pages.  this should probably be fixed.
272 	 */
273 	pcb2->pcb_ext = NULL;
274 
275 	/* Copy the LDT, if necessary. */
276 	mtx_lock_spin(&dt_lock);
277 	if (mdp2->md_ldt != NULL) {
278 		if (flags & RFMEM) {
279 			mdp2->md_ldt->ldt_refcnt++;
280 		} else {
281 			mdp2->md_ldt = user_ldt_alloc(mdp2,
282 			    mdp2->md_ldt->ldt_len);
283 			if (mdp2->md_ldt == NULL)
284 				panic("could not copy LDT");
285 		}
286 	}
287 	mtx_unlock_spin(&dt_lock);
288 
289 	/*
290 	 * Now, cpu_switch() can schedule the new process.
291 	 * pcb_esp is loaded pointing to the cpu_switch() stack frame
292 	 * containing the return address when exiting cpu_switch.
293 	 * This will normally be to fork_trampoline(), which will have
294 	 * %ebx loaded with the new proc's pointer.  fork_trampoline()
295 	 * will set up a stack to call fork_return(p, frame); to complete
296 	 * the return to user-mode.
297 	 */
298 }
299 
300 void
301 x86_set_fork_retval(struct thread *td)
302 {
303 	struct trapframe * frame = td->td_frame;
304 
305 	frame->tf_eax = 0;		/* Child returns zero */
306 	frame->tf_eflags &= ~PSL_C;	/* success */
307 	frame->tf_edx = 1;		/* System V emulation */
308 }
309 
310 /*
311  * Intercept the return address from a freshly forked process that has NOT
312  * been scheduled yet.
313  *
314  * This is needed to make kernel threads stay in kernel mode.
315  */
316 void
317 cpu_fork_kthread_handler(struct thread *td, void (*func)(void *), void *arg)
318 {
319 	/*
320 	 * Note that the trap frame follows the args, so the function
321 	 * is really called like this:  func(arg, frame);
322 	 */
323 	td->td_pcb->pcb_esi = (int) func;	/* function */
324 	td->td_pcb->pcb_ebx = (int) arg;	/* first arg */
325 }
326 
327 void
328 cpu_exit(struct thread *td)
329 {
330 
331 	/*
332 	 * If this process has a custom LDT, release it.  Reset pc->pcb_gs
333 	 * and %gs before we free it in case they refer to an LDT entry.
334 	 */
335 	mtx_lock_spin(&dt_lock);
336 	if (td->td_proc->p_md.md_ldt) {
337 		td->td_pcb->pcb_gs = _udatasel;
338 		load_gs(_udatasel);
339 		user_ldt_free(td);
340 	} else
341 		mtx_unlock_spin(&dt_lock);
342 }
343 
344 void
345 cpu_thread_exit(struct thread *td)
346 {
347 
348 	critical_enter();
349 	if (td == PCPU_GET(fpcurthread))
350 		npxdrop();
351 	critical_exit();
352 
353 	/* Disable any hardware breakpoints. */
354 	if (td->td_pcb->pcb_flags & PCB_DBREGS) {
355 		reset_dbregs();
356 		td->td_pcb->pcb_flags &= ~PCB_DBREGS;
357 	}
358 }
359 
360 void
361 cpu_thread_clean(struct thread *td)
362 {
363 	struct pcb *pcb;
364 
365 	pcb = td->td_pcb;
366 	if (pcb->pcb_ext != NULL) {
367 		/* if (pcb->pcb_ext->ext_refcount-- == 1) ?? */
368 		/*
369 		 * XXX do we need to move the TSS off the allocated pages
370 		 * before freeing them?  (not done here)
371 		 */
372 		pmap_trm_free(pcb->pcb_ext, ctob(IOPAGES + 1));
373 		pcb->pcb_ext = NULL;
374 	}
375 }
376 
377 void
378 cpu_thread_alloc(struct thread *td)
379 {
380 	struct pcb *pcb;
381 	struct xstate_hdr *xhdr;
382 
383 	td->td_pcb = pcb = get_pcb_td(td);
384 	td->td_frame = (struct trapframe *)((caddr_t)pcb -
385 	    VM86_STACK_SPACE) - 1;
386 	pcb->pcb_ext = NULL;
387 	pcb->pcb_save = get_pcb_user_save_pcb(pcb);
388 	if (use_xsave) {
389 		xhdr = (struct xstate_hdr *)(pcb->pcb_save + 1);
390 		bzero(xhdr, sizeof(*xhdr));
391 		xhdr->xstate_bv = xsave_mask;
392 	}
393 }
394 
395 void
396 cpu_thread_free(struct thread *td)
397 {
398 
399 	cpu_thread_clean(td);
400 }
401 
402 bool
403 cpu_exec_vmspace_reuse(struct proc *p __unused, vm_map_t map __unused)
404 {
405 
406 	return (true);
407 }
408 
409 int
410 cpu_procctl(struct thread *td __unused, int idtype __unused, id_t id __unused,
411     int com __unused, void *data __unused)
412 {
413 
414 	return (EINVAL);
415 }
416 
417 void
418 cpu_set_syscall_retval(struct thread *td, int error)
419 {
420 
421 	switch (error) {
422 	case 0:
423 		td->td_frame->tf_eax = td->td_retval[0];
424 		td->td_frame->tf_edx = td->td_retval[1];
425 		td->td_frame->tf_eflags &= ~PSL_C;
426 		break;
427 
428 	case ERESTART:
429 		/*
430 		 * Reconstruct pc, assuming lcall $X,y is 7 bytes, int
431 		 * 0x80 is 2 bytes. We saved this in tf_err.
432 		 */
433 		td->td_frame->tf_eip -= td->td_frame->tf_err;
434 		break;
435 
436 	case EJUSTRETURN:
437 		break;
438 
439 	default:
440 		td->td_frame->tf_eax = error;
441 		td->td_frame->tf_eflags |= PSL_C;
442 		break;
443 	}
444 }
445 
446 /*
447  * Initialize machine state, mostly pcb and trap frame for a new
448  * thread, about to return to userspace.  Put enough state in the new
449  * thread's PCB to get it to go back to the fork_return(), which
450  * finalizes the thread state and handles peculiarities of the first
451  * return to userspace for the new thread.
452  */
453 void
454 cpu_copy_thread(struct thread *td, struct thread *td0)
455 {
456 	copy_thread(td0, td);
457 
458 	/*
459 	 * Copy user general-purpose registers.
460 	 *
461 	 * Some of these registers are rewritten by cpu_set_upcall()
462 	 * and linux_set_upcall().
463 	 */
464 	bcopy(td0->td_frame, td->td_frame, sizeof(struct trapframe));
465 
466 	/* If the current thread has the trap bit set (i.e. a debugger had
467 	 * single stepped the process to the system call), we need to clear
468 	 * the trap flag from the new frame. Otherwise, the new thread will
469 	 * receive a (likely unexpected) SIGTRAP when it executes the first
470 	 * instruction after returning to userland.
471 	 */
472 	td->td_frame->tf_eflags &= ~PSL_T;
473 }
474 
475 /*
476  * Set that machine state for performing an upcall that starts
477  * the entry function with the given argument.
478  */
479 int
480 cpu_set_upcall(struct thread *td, void (*entry)(void *), void *arg,
481     stack_t *stack)
482 {
483 
484 	/*
485 	 * Do any extra cleaning that needs to be done.
486 	 * The thread may have optional components
487 	 * that are not present in a fresh thread.
488 	 * This may be a recycled thread so make it look
489 	 * as though it's newly allocated.
490 	 */
491 	cpu_thread_clean(td);
492 
493 	/*
494 	 * Set the trap frame to point at the beginning of the entry
495 	 * function.
496 	 */
497 	td->td_frame->tf_ebp = 0;
498 	td->td_frame->tf_esp =
499 	    (((int)stack->ss_sp + stack->ss_size - 4) & ~0x0f) - 4;
500 	td->td_frame->tf_eip = (int)entry;
501 
502 	/* Return address sentinel value to stop stack unwinding. */
503 	if (suword((void *)td->td_frame->tf_esp, 0) != 0)
504 		return (EFAULT);
505 
506 	/* Pass the argument to the entry point. */
507 	if (suword((void *)(td->td_frame->tf_esp + sizeof(void *)),
508 	    (int)arg) != 0)
509 		return (EFAULT);
510 	return (0);
511 }
512 
513 int
514 cpu_set_user_tls(struct thread *td, void *tls_base)
515 {
516 	struct segment_descriptor sd;
517 	uint32_t base;
518 
519 	/*
520 	 * Construct a descriptor and store it in the pcb for
521 	 * the next context switch.  Also store it in the gdt
522 	 * so that the load of tf_fs into %fs will activate it
523 	 * at return to userland.
524 	 */
525 	base = (uint32_t)tls_base;
526 	sd.sd_lobase = base & 0xffffff;
527 	sd.sd_hibase = (base >> 24) & 0xff;
528 	sd.sd_lolimit = 0xffff;	/* 4GB limit, wraps around */
529 	sd.sd_hilimit = 0xf;
530 	sd.sd_type  = SDT_MEMRWA;
531 	sd.sd_dpl   = SEL_UPL;
532 	sd.sd_p     = 1;
533 	sd.sd_xx    = 0;
534 	sd.sd_def32 = 1;
535 	sd.sd_gran  = 1;
536 	critical_enter();
537 	/* set %gs */
538 	td->td_pcb->pcb_gsd = sd;
539 	if (td == curthread) {
540 		PCPU_GET(fsgs_gdt)[1] = sd;
541 		load_gs(GSEL(GUGS_SEL, SEL_UPL));
542 	}
543 	critical_exit();
544 	return (0);
545 }
546 
547 /*
548  * Convert kernel VA to physical address
549  */
550 vm_paddr_t
551 kvtop(void *addr)
552 {
553 	vm_paddr_t pa;
554 
555 	pa = pmap_kextract((vm_offset_t)addr);
556 	if (pa == 0)
557 		panic("kvtop: zero page frame");
558 	return (pa);
559 }
560 
561 /*
562  * Get an sf_buf from the freelist.  May block if none are available.
563  */
564 void
565 sf_buf_map(struct sf_buf *sf, int flags)
566 {
567 
568 	pmap_sf_buf_map(sf);
569 #ifdef SMP
570 	sf_buf_shootdown(sf, flags);
571 #endif
572 }
573 
574 #ifdef SMP
575 static void
576 sf_buf_shootdown_curcpu_cb(pmap_t pmap __unused,
577     vm_offset_t addr1 __unused, vm_offset_t addr2 __unused)
578 {
579 }
580 
581 void
582 sf_buf_shootdown(struct sf_buf *sf, int flags)
583 {
584 	cpuset_t other_cpus;
585 	u_int cpuid;
586 
587 	sched_pin();
588 	cpuid = PCPU_GET(cpuid);
589 	if (!CPU_ISSET(cpuid, &sf->cpumask)) {
590 		CPU_SET(cpuid, &sf->cpumask);
591 		invlpg(sf->kva);
592 	}
593 	if ((flags & SFB_CPUPRIVATE) == 0) {
594 		other_cpus = all_cpus;
595 		CPU_CLR(cpuid, &other_cpus);
596 		CPU_ANDNOT(&other_cpus, &other_cpus, &sf->cpumask);
597 		if (!CPU_EMPTY(&other_cpus)) {
598 			CPU_OR(&sf->cpumask, &sf->cpumask, &other_cpus);
599 			smp_masked_invlpg(other_cpus, sf->kva, kernel_pmap,
600 			    sf_buf_shootdown_curcpu_cb);
601 		}
602 	}
603 	sched_unpin();
604 }
605 #endif
606 
607 /*
608  * MD part of sf_buf_free().
609  */
610 int
611 sf_buf_unmap(struct sf_buf *sf)
612 {
613 
614 	return (0);
615 }
616 
617 static void
618 sf_buf_invalidate(struct sf_buf *sf)
619 {
620 	vm_page_t m = sf->m;
621 
622 	/*
623 	 * Use pmap_qenter to update the pte for
624 	 * existing mapping, in particular, the PAT
625 	 * settings are recalculated.
626 	 */
627 	pmap_qenter(sf->kva, &m, 1);
628 	pmap_invalidate_cache_range(sf->kva, sf->kva + PAGE_SIZE);
629 }
630 
631 /*
632  * Invalidate the cache lines that may belong to the page, if
633  * (possibly old) mapping of the page by sf buffer exists.  Returns
634  * TRUE when mapping was found and cache invalidated.
635  */
636 boolean_t
637 sf_buf_invalidate_cache(vm_page_t m)
638 {
639 
640 	return (sf_buf_process_page(m, sf_buf_invalidate));
641 }
642