2 * Copyright (c) 1982, 1986 The Regents of the University of California.
3 * Copyright (c) 1989, 1990 William Jolitz
4 * Copyright (c) 1994 John Dyson
7 * This code is derived from software contributed to Berkeley by
8 * the Systems Programming Group of the University of Utah Computer
9 * Science Department, and William Jolitz.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in the
18 * documentation and/or other materials provided with the distribution.
19 * 3. All advertising materials mentioning features or use of this software
20 * must display the following acknowledgement:
21 * This product includes software developed by the University of
22 * California, Berkeley and its contributors.
23 * 4. Neither the name of the University nor the names of its contributors
24 * may be used to endorse or promote products derived from this software
25 * without specific prior written permission.
27 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
28 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
29 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
30 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
31 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
32 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
33 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
34 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
35 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
36 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39 * from: @(#)vm_machdep.c 7.3 (Berkeley) 5/13/91
40 * Utah $Hdr: vm_machdep.c 1.16.1.1 89/06/23$
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
48 #include "opt_reset.h"
52 #include <sys/param.h>
53 #include <sys/systm.h>
56 #include <sys/kernel.h>
59 #include <sys/malloc.h>
61 #include <sys/mutex.h>
62 #include <sys/pioctl.h>
64 #include <sys/sysent.h>
65 #include <sys/sf_buf.h>
67 #include <sys/sched.h>
68 #include <sys/sysctl.h>
69 #include <sys/unistd.h>
70 #include <sys/vnode.h>
71 #include <sys/vmmeter.h>
73 #include <machine/cpu.h>
74 #include <machine/cputypes.h>
75 #include <machine/md_var.h>
76 #include <machine/pcb.h>
77 #include <machine/pcb_ext.h>
78 #include <machine/smp.h>
79 #include <machine/vm86.h>
82 #include <machine/elan_mmcr.h>
86 #include <vm/vm_extern.h>
87 #include <vm/vm_kern.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_param.h>
93 #include <pc98/cbus/cbus.h>
95 #include <isa/isareg.h>
99 #include <machine/xbox.h>
103 #define NSFBUFS (512 + maxusers * 16)
106 _Static_assert(OFFSETOF_CURTHREAD == offsetof(struct pcpu, pc_curthread),
107 "OFFSETOF_CURTHREAD does not correspond with offset of pc_curthread.");
108 _Static_assert(OFFSETOF_CURPCB == offsetof(struct pcpu, pc_curpcb),
109 "OFFSETOF_CURPCB does not correspond with offset of pc_curpcb.");
110 _Static_assert(__OFFSETOF_MONITORBUF == offsetof(struct pcpu, pc_monitorbuf),
111 "__OFFSETOF_MONINORBUF does not correspond with offset of pc_monitorbuf.");
113 static void cpu_reset_real(void);
115 static void cpu_reset_proxy(void);
116 static u_int cpu_reset_proxyid;
117 static volatile u_int cpu_reset_proxy_active;
121 get_pcb_user_save_td(struct thread *td)
125 p = td->td_kstack + td->td_kstack_pages * PAGE_SIZE -
126 roundup2(cpu_max_ext_state_size, XSAVE_AREA_ALIGN);
127 KASSERT((p % XSAVE_AREA_ALIGN) == 0, ("Unaligned pcb_user_save area"));
128 return ((union savefpu *)p);
132 get_pcb_user_save_pcb(struct pcb *pcb)
136 p = (vm_offset_t)(pcb + 1);
137 return ((union savefpu *)p);
141 get_pcb_td(struct thread *td)
145 p = td->td_kstack + td->td_kstack_pages * PAGE_SIZE -
146 roundup2(cpu_max_ext_state_size, XSAVE_AREA_ALIGN) -
148 return ((struct pcb *)p);
152 alloc_fpusave(int flags)
155 struct savefpu_ymm *sf;
157 res = malloc(cpu_max_ext_state_size, M_DEVBUF, flags);
159 sf = (struct savefpu_ymm *)res;
160 bzero(&sf->sv_xstate.sx_hd, sizeof(sf->sv_xstate.sx_hd));
161 sf->sv_xstate.sx_hd.xstate_bv = xsave_mask;
166 * Finish a fork operation, with process p2 nearly set up.
167 * Copy and update the pcb, set up the stack so that the child
168 * ready to run and return to user mode.
171 cpu_fork(struct thread *td1, struct proc *p2, struct thread *td2, int flags)
173 register struct proc *p1;
178 if ((flags & RFPROC) == 0) {
179 if ((flags & RFMEM) == 0) {
180 /* unshare user LDT */
181 struct mdproc *mdp1 = &p1->p_md;
182 struct proc_ldt *pldt, *pldt1;
184 mtx_lock_spin(&dt_lock);
185 if ((pldt1 = mdp1->md_ldt) != NULL &&
186 pldt1->ldt_refcnt > 1) {
187 pldt = user_ldt_alloc(mdp1, pldt1->ldt_len);
189 panic("could not copy LDT");
192 user_ldt_deref(pldt1);
194 mtx_unlock_spin(&dt_lock);
199 /* Ensure that td1's pcb is up to date. */
200 if (td1 == curthread)
201 td1->td_pcb->pcb_gs = rgs();
203 if (PCPU_GET(fpcurthread) == td1)
204 npxsave(td1->td_pcb->pcb_save);
207 /* Point the pcb to the top of the stack */
208 pcb2 = get_pcb_td(td2);
212 bcopy(td1->td_pcb, pcb2, sizeof(*pcb2));
214 /* Properly initialize pcb_save */
215 pcb2->pcb_save = get_pcb_user_save_pcb(pcb2);
216 bcopy(get_pcb_user_save_td(td1), get_pcb_user_save_pcb(pcb2),
217 cpu_max_ext_state_size);
219 /* Point mdproc and then copy over td1's contents */
221 bcopy(&p1->p_md, mdp2, sizeof(*mdp2));
224 * Create a new fresh stack for the new process.
225 * Copy the trap frame for the return to user mode as if from a
226 * syscall. This copies most of the user mode register values.
227 * The -16 is so we can expand the trapframe if we go to vm86.
229 td2->td_frame = (struct trapframe *)((caddr_t)td2->td_pcb - 16) - 1;
230 bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe));
232 td2->td_frame->tf_eax = 0; /* Child returns zero */
233 td2->td_frame->tf_eflags &= ~PSL_C; /* success */
234 td2->td_frame->tf_edx = 1;
237 * If the parent process has the trap bit set (i.e. a debugger had
238 * single stepped the process to the system call), we need to clear
239 * the trap flag from the new frame unless the debugger had set PF_FORK
240 * on the parent. Otherwise, the child will receive a (likely
241 * unexpected) SIGTRAP when it executes the first instruction after
242 * returning to userland.
244 if ((p1->p_pfsflags & PF_FORK) == 0)
245 td2->td_frame->tf_eflags &= ~PSL_T;
248 * Set registers for trampoline to user mode. Leave space for the
249 * return address on stack. These are the kernel mode register values.
251 #if defined(PAE) || defined(PAE_TABLES)
252 pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdpt);
254 pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdir);
257 pcb2->pcb_esi = (int)fork_return; /* fork_trampoline argument */
259 pcb2->pcb_esp = (int)td2->td_frame - sizeof(void *);
260 pcb2->pcb_ebx = (int)td2; /* fork_trampoline argument */
261 pcb2->pcb_eip = (int)fork_trampoline;
262 pcb2->pcb_psl = PSL_KERNEL; /* ints disabled */
264 * pcb2->pcb_dr*: cloned above.
265 * pcb2->pcb_savefpu: cloned above.
266 * pcb2->pcb_flags: cloned above.
267 * pcb2->pcb_onfault: cloned above (always NULL here?).
268 * pcb2->pcb_gs: cloned above.
269 * pcb2->pcb_ext: cleared below.
273 * XXX don't copy the i/o pages. this should probably be fixed.
277 /* Copy the LDT, if necessary. */
278 mtx_lock_spin(&dt_lock);
279 if (mdp2->md_ldt != NULL) {
281 mdp2->md_ldt->ldt_refcnt++;
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");
289 mtx_unlock_spin(&dt_lock);
291 /* Setup to release spin count in fork_exit(). */
292 td2->td_md.md_spinlock_count = 1;
293 td2->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
296 * Now, cpu_switch() can schedule the new process.
297 * pcb_esp is loaded pointing to the cpu_switch() stack frame
298 * containing the return address when exiting cpu_switch.
299 * This will normally be to fork_trampoline(), which will have
300 * %ebx loaded with the new proc's pointer. fork_trampoline()
301 * will set up a stack to call fork_return(p, frame); to complete
302 * the return to user-mode.
307 * Intercept the return address from a freshly forked process that has NOT
308 * been scheduled yet.
310 * This is needed to make kernel threads stay in kernel mode.
313 cpu_fork_kthread_handler(struct thread *td, void (*func)(void *), void *arg)
316 * Note that the trap frame follows the args, so the function
317 * is really called like this: func(arg, frame);
319 td->td_pcb->pcb_esi = (int) func; /* function */
320 td->td_pcb->pcb_ebx = (int) arg; /* first arg */
324 cpu_exit(struct thread *td)
328 * If this process has a custom LDT, release it. Reset pc->pcb_gs
329 * and %gs before we free it in case they refer to an LDT entry.
331 mtx_lock_spin(&dt_lock);
332 if (td->td_proc->p_md.md_ldt) {
333 td->td_pcb->pcb_gs = _udatasel;
337 mtx_unlock_spin(&dt_lock);
341 cpu_thread_exit(struct thread *td)
345 if (td == PCPU_GET(fpcurthread))
349 /* Disable any hardware breakpoints. */
350 if (td->td_pcb->pcb_flags & PCB_DBREGS) {
352 td->td_pcb->pcb_flags &= ~PCB_DBREGS;
357 cpu_thread_clean(struct thread *td)
362 if (pcb->pcb_ext != NULL) {
363 /* if (pcb->pcb_ext->ext_refcount-- == 1) ?? */
365 * XXX do we need to move the TSS off the allocated pages
366 * before freeing them? (not done here)
368 kmem_free(kernel_arena, (vm_offset_t)pcb->pcb_ext,
375 cpu_thread_swapin(struct thread *td)
380 cpu_thread_swapout(struct thread *td)
385 cpu_thread_alloc(struct thread *td)
388 struct xstate_hdr *xhdr;
390 td->td_pcb = pcb = get_pcb_td(td);
391 td->td_frame = (struct trapframe *)((caddr_t)pcb - 16) - 1;
393 pcb->pcb_save = get_pcb_user_save_pcb(pcb);
395 xhdr = (struct xstate_hdr *)(pcb->pcb_save + 1);
396 bzero(xhdr, sizeof(*xhdr));
397 xhdr->xstate_bv = xsave_mask;
402 cpu_thread_free(struct thread *td)
405 cpu_thread_clean(td);
409 cpu_set_syscall_retval(struct thread *td, int error)
414 td->td_frame->tf_eax = td->td_retval[0];
415 td->td_frame->tf_edx = td->td_retval[1];
416 td->td_frame->tf_eflags &= ~PSL_C;
421 * Reconstruct pc, assuming lcall $X,y is 7 bytes, int
422 * 0x80 is 2 bytes. We saved this in tf_err.
424 td->td_frame->tf_eip -= td->td_frame->tf_err;
431 td->td_frame->tf_eax = SV_ABI_ERRNO(td->td_proc, error);
432 td->td_frame->tf_eflags |= PSL_C;
438 * Initialize machine state, mostly pcb and trap frame for a new
439 * thread, about to return to userspace. Put enough state in the new
440 * thread's PCB to get it to go back to the fork_return(), which
441 * finalizes the thread state and handles peculiarities of the first
442 * return to userspace for the new thread.
445 cpu_copy_thread(struct thread *td, struct thread *td0)
449 /* Point the pcb to the top of the stack. */
453 * Copy the upcall pcb. This loads kernel regs.
454 * Those not loaded individually below get their default
457 bcopy(td0->td_pcb, pcb2, sizeof(*pcb2));
458 pcb2->pcb_flags &= ~(PCB_NPXINITDONE | PCB_NPXUSERINITDONE |
460 pcb2->pcb_save = get_pcb_user_save_pcb(pcb2);
461 bcopy(get_pcb_user_save_td(td0), pcb2->pcb_save,
462 cpu_max_ext_state_size);
465 * Create a new fresh stack for the new thread.
467 bcopy(td0->td_frame, td->td_frame, sizeof(struct trapframe));
469 /* If the current thread has the trap bit set (i.e. a debugger had
470 * single stepped the process to the system call), we need to clear
471 * the trap flag from the new frame. Otherwise, the new thread will
472 * receive a (likely unexpected) SIGTRAP when it executes the first
473 * instruction after returning to userland.
475 td->td_frame->tf_eflags &= ~PSL_T;
478 * Set registers for trampoline to user mode. Leave space for the
479 * return address on stack. These are the kernel mode register values.
482 pcb2->pcb_esi = (int)fork_return; /* trampoline arg */
484 pcb2->pcb_esp = (int)td->td_frame - sizeof(void *); /* trampoline arg */
485 pcb2->pcb_ebx = (int)td; /* trampoline arg */
486 pcb2->pcb_eip = (int)fork_trampoline;
487 pcb2->pcb_psl &= ~(PSL_I); /* interrupts must be disabled */
488 pcb2->pcb_gs = rgs();
490 * If we didn't copy the pcb, we'd need to do the following registers:
491 * pcb2->pcb_cr3: cloned above.
492 * pcb2->pcb_dr*: cloned above.
493 * pcb2->pcb_savefpu: cloned above.
494 * pcb2->pcb_flags: cloned above.
495 * pcb2->pcb_onfault: cloned above (always NULL here?).
496 * pcb2->pcb_gs: cloned above.
497 * pcb2->pcb_ext: cleared below.
499 pcb2->pcb_ext = NULL;
501 /* Setup to release spin count in fork_exit(). */
502 td->td_md.md_spinlock_count = 1;
503 td->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
507 * Set that machine state for performing an upcall that starts
508 * the entry function with the given argument.
511 cpu_set_upcall(struct thread *td, void (*entry)(void *), void *arg,
516 * Do any extra cleaning that needs to be done.
517 * The thread may have optional components
518 * that are not present in a fresh thread.
519 * This may be a recycled thread so make it look
520 * as though it's newly allocated.
522 cpu_thread_clean(td);
525 * Set the trap frame to point at the beginning of the entry
528 td->td_frame->tf_ebp = 0;
529 td->td_frame->tf_esp =
530 (((int)stack->ss_sp + stack->ss_size - 4) & ~0x0f) - 4;
531 td->td_frame->tf_eip = (int)entry;
533 /* Return address sentinel value to stop stack unwinding. */
534 suword((void *)td->td_frame->tf_esp, 0);
536 /* Pass the argument to the entry point. */
537 suword((void *)(td->td_frame->tf_esp + sizeof(void *)),
542 cpu_set_user_tls(struct thread *td, void *tls_base)
544 struct segment_descriptor sd;
548 * Construct a descriptor and store it in the pcb for
549 * the next context switch. Also store it in the gdt
550 * so that the load of tf_fs into %fs will activate it
551 * at return to userland.
553 base = (uint32_t)tls_base;
554 sd.sd_lobase = base & 0xffffff;
555 sd.sd_hibase = (base >> 24) & 0xff;
556 sd.sd_lolimit = 0xffff; /* 4GB limit, wraps around */
558 sd.sd_type = SDT_MEMRWA;
566 td->td_pcb->pcb_gsd = sd;
567 if (td == curthread) {
568 PCPU_GET(fsgs_gdt)[1] = sd;
569 load_gs(GSEL(GUGS_SEL, SEL_UPL));
576 * Convert kernel VA to physical address
583 pa = pmap_kextract((vm_offset_t)addr);
585 panic("kvtop: zero page frame");
595 cpu_reset_proxy_active = 1;
596 while (cpu_reset_proxy_active == 1)
597 ; /* Wait for other cpu to see that we've started */
598 CPU_SETOF(cpu_reset_proxyid, &tcrp);
600 printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
610 if (arch_i386_is_xbox) {
611 /* Kick the PIC16L, it can reboot the box */
623 CPU_CLR(PCPU_GET(cpuid), &map);
624 CPU_NAND(&map, &stopped_cpus);
625 if (!CPU_EMPTY(&map)) {
626 printf("cpu_reset: Stopping other CPUs\n");
630 if (PCPU_GET(cpuid) != 0) {
631 cpu_reset_proxyid = PCPU_GET(cpuid);
632 cpustop_restartfunc = cpu_reset_proxy;
633 cpu_reset_proxy_active = 0;
634 printf("cpu_reset: Restarting BSP\n");
636 /* Restart CPU #0. */
637 /* XXX: restart_cpus(1 << 0); */
638 CPU_SETOF(0, &started_cpus);
642 while (cpu_reset_proxy_active == 0 && cnt < 10000000)
643 cnt++; /* Wait for BSP to announce restart */
644 if (cpu_reset_proxy_active == 0)
645 printf("cpu_reset: Failed to restart BSP\n");
647 cpu_reset_proxy_active = 2;
663 struct region_descriptor null_idt;
670 if (elan_mmcr != NULL)
671 elan_mmcr->RESCFG = 1;
674 if (cpu == CPU_GEODE1100) {
675 /* Attempt Geode's own reset */
676 outl(0xcf8, 0x80009044ul);
682 * Attempt to do a CPU reset via CPU reset port.
684 if ((inb(0x35) & 0xa0) != 0xa0) {
685 outb(0x37, 0x0f); /* SHUT0 = 0. */
686 outb(0x37, 0x0b); /* SHUT1 = 0. */
688 outb(0xf0, 0x00); /* Reset. */
690 #if !defined(BROKEN_KEYBOARD_RESET)
692 * Attempt to do a CPU reset via the keyboard controller,
693 * do not turn off GateA20, as any machine that fails
694 * to do the reset here would then end up in no man's land.
696 outb(IO_KBD + 4, 0xFE);
697 DELAY(500000); /* wait 0.5 sec to see if that did it */
701 * Attempt to force a reset via the Reset Control register at
702 * I/O port 0xcf9. Bit 2 forces a system reset when it
703 * transitions from 0 to 1. Bit 1 selects the type of reset
704 * to attempt: 0 selects a "soft" reset, and 1 selects a
705 * "hard" reset. We try a "hard" reset. The first write sets
706 * bit 1 to select a "hard" reset and clears bit 2. The
707 * second write forces a 0 -> 1 transition in bit 2 to trigger
712 DELAY(500000); /* wait 0.5 sec to see if that did it */
715 * Attempt to force a reset via the Fast A20 and Init register
716 * at I/O port 0x92. Bit 1 serves as an alternate A20 gate.
717 * Bit 0 asserts INIT# when set to 1. We are careful to only
718 * preserve bit 1 while setting bit 0. We also must clear bit
719 * 0 before setting it if it isn't already clear.
724 outb(0x92, b & 0xfe);
726 DELAY(500000); /* wait 0.5 sec to see if that did it */
730 printf("No known reset method worked, attempting CPU shutdown\n");
731 DELAY(1000000); /* wait 1 sec for printf to complete */
734 null_idt.rd_limit = 0;
735 null_idt.rd_base = 0;
738 /* "good night, sweet prince .... <THUNK!>" */
746 * Get an sf_buf from the freelist. May block if none are available.
749 sf_buf_map(struct sf_buf *sf, int flags)
751 pt_entry_t opte, *ptep;
754 * Update the sf_buf's virtual-to-physical mapping, flushing the
755 * virtual address from the TLB. Since the reference count for
756 * the sf_buf's old mapping was zero, that mapping is not
757 * currently in use. Consequently, there is no need to exchange
758 * the old and new PTEs atomically, even under PAE.
760 ptep = vtopte(sf->kva);
762 *ptep = VM_PAGE_TO_PHYS(sf->m) | pgeflag | PG_RW | PG_V |
763 pmap_cache_bits(sf->m->md.pat_mode, 0);
766 * Avoid unnecessary TLB invalidations: If the sf_buf's old
767 * virtual-to-physical mapping was not used, then any processor
768 * that has invalidated the sf_buf's virtual address from its TLB
769 * since the last used mapping need not invalidate again.
772 if ((opte & (PG_V | PG_A)) == (PG_V | PG_A))
773 CPU_ZERO(&sf->cpumask);
775 sf_buf_shootdown(sf, flags);
777 if ((opte & (PG_V | PG_A)) == (PG_V | PG_A))
778 pmap_invalidate_page(kernel_pmap, sf->kva);
784 sf_buf_shootdown(struct sf_buf *sf, int flags)
790 cpuid = PCPU_GET(cpuid);
791 if (!CPU_ISSET(cpuid, &sf->cpumask)) {
792 CPU_SET(cpuid, &sf->cpumask);
795 if ((flags & SFB_CPUPRIVATE) == 0) {
796 other_cpus = all_cpus;
797 CPU_CLR(cpuid, &other_cpus);
798 CPU_NAND(&other_cpus, &sf->cpumask);
799 if (!CPU_EMPTY(&other_cpus)) {
800 CPU_OR(&sf->cpumask, &other_cpus);
801 smp_masked_invlpg(other_cpus, sf->kva, kernel_pmap);
809 * MD part of sf_buf_free().
812 sf_buf_unmap(struct sf_buf *sf)
819 sf_buf_invalidate(struct sf_buf *sf)
824 * Use pmap_qenter to update the pte for
825 * existing mapping, in particular, the PAT
826 * settings are recalculated.
828 pmap_qenter(sf->kva, &m, 1);
829 pmap_invalidate_cache_range(sf->kva, sf->kva + PAGE_SIZE, FALSE);
833 * Invalidate the cache lines that may belong to the page, if
834 * (possibly old) mapping of the page by sf buffer exists. Returns
835 * TRUE when mapping was found and cache invalidated.
838 sf_buf_invalidate_cache(vm_page_t m)
841 return (sf_buf_process_page(m, sf_buf_invalidate));
845 * Software interrupt handler for queued VM system processing.
850 if (busdma_swi_pending != 0)
855 * Tell whether this address is in some physical memory region.
856 * Currently used by the kernel coredump code in order to avoid
857 * dumping the ``ISA memory hole'' which could cause indefinite hangs,
858 * or other unpredictable behaviour.
862 is_physical_memory(vm_paddr_t addr)
866 /* The ISA ``memory hole''. */
867 if (addr >= 0xa0000 && addr < 0x100000)
872 * stuff other tests for known memory-mapped devices (PCI?)