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 <xen/hypervisor.h>
96 #include <pc98/cbus/cbus.h>
98 #include <isa/isareg.h>
102 #include <machine/xbox.h>
106 #define NSFBUFS (512 + maxusers * 16)
109 _Static_assert(OFFSETOF_CURTHREAD == offsetof(struct pcpu, pc_curthread),
110 "OFFSETOF_CURTHREAD does not correspond with offset of pc_curthread.");
111 _Static_assert(OFFSETOF_CURPCB == offsetof(struct pcpu, pc_curpcb),
112 "OFFSETOF_CURPCB does not correspond with offset of pc_curpcb.");
114 static void cpu_reset_real(void);
116 static void cpu_reset_proxy(void);
117 static u_int cpu_reset_proxyid;
118 static volatile u_int cpu_reset_proxy_active;
121 extern int _ucodesel, _udatasel;
124 * Finish a fork operation, with process p2 nearly set up.
125 * Copy and update the pcb, set up the stack so that the child
126 * ready to run and return to user mode.
129 cpu_fork(td1, p2, td2, flags)
130 register struct thread *td1;
131 register struct proc *p2;
135 register struct proc *p1;
140 if ((flags & RFPROC) == 0) {
141 if ((flags & RFMEM) == 0) {
142 /* unshare user LDT */
143 struct mdproc *mdp1 = &p1->p_md;
144 struct proc_ldt *pldt, *pldt1;
146 mtx_lock_spin(&dt_lock);
147 if ((pldt1 = mdp1->md_ldt) != NULL &&
148 pldt1->ldt_refcnt > 1) {
149 pldt = user_ldt_alloc(mdp1, pldt1->ldt_len);
151 panic("could not copy LDT");
154 user_ldt_deref(pldt1);
156 mtx_unlock_spin(&dt_lock);
161 /* Ensure that td1's pcb is up to date. */
162 if (td1 == curthread)
163 td1->td_pcb->pcb_gs = rgs();
166 if (PCPU_GET(fpcurthread) == td1)
167 npxsave(td1->td_pcb->pcb_save);
171 /* Point the pcb to the top of the stack */
172 pcb2 = (struct pcb *)(td2->td_kstack +
173 td2->td_kstack_pages * PAGE_SIZE) - 1;
177 bcopy(td1->td_pcb, pcb2, sizeof(*pcb2));
179 /* Properly initialize pcb_save */
180 pcb2->pcb_save = &pcb2->pcb_user_save;
182 /* Point mdproc and then copy over td1's contents */
184 bcopy(&p1->p_md, mdp2, sizeof(*mdp2));
187 * Create a new fresh stack for the new process.
188 * Copy the trap frame for the return to user mode as if from a
189 * syscall. This copies most of the user mode register values.
190 * The -16 is so we can expand the trapframe if we go to vm86.
192 td2->td_frame = (struct trapframe *)((caddr_t)td2->td_pcb - 16) - 1;
193 bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe));
195 td2->td_frame->tf_eax = 0; /* Child returns zero */
196 td2->td_frame->tf_eflags &= ~PSL_C; /* success */
197 td2->td_frame->tf_edx = 1;
200 * If the parent process has the trap bit set (i.e. a debugger had
201 * single stepped the process to the system call), we need to clear
202 * the trap flag from the new frame unless the debugger had set PF_FORK
203 * on the parent. Otherwise, the child will receive a (likely
204 * unexpected) SIGTRAP when it executes the first instruction after
205 * returning to userland.
207 if ((p1->p_pfsflags & PF_FORK) == 0)
208 td2->td_frame->tf_eflags &= ~PSL_T;
211 * Set registers for trampoline to user mode. Leave space for the
212 * return address on stack. These are the kernel mode register values.
215 pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdpt);
217 pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdir);
220 pcb2->pcb_esi = (int)fork_return; /* fork_trampoline argument */
222 pcb2->pcb_esp = (int)td2->td_frame - sizeof(void *);
223 pcb2->pcb_ebx = (int)td2; /* fork_trampoline argument */
224 pcb2->pcb_eip = (int)fork_trampoline;
225 pcb2->pcb_psl = PSL_KERNEL; /* ints disabled */
227 * pcb2->pcb_dr*: cloned above.
228 * pcb2->pcb_savefpu: cloned above.
229 * pcb2->pcb_flags: cloned above.
230 * pcb2->pcb_onfault: cloned above (always NULL here?).
231 * pcb2->pcb_gs: cloned above.
232 * pcb2->pcb_ext: cleared below.
236 * XXX don't copy the i/o pages. this should probably be fixed.
240 /* Copy the LDT, if necessary. */
241 mtx_lock_spin(&dt_lock);
242 if (mdp2->md_ldt != NULL) {
244 mdp2->md_ldt->ldt_refcnt++;
246 mdp2->md_ldt = user_ldt_alloc(mdp2,
247 mdp2->md_ldt->ldt_len);
248 if (mdp2->md_ldt == NULL)
249 panic("could not copy LDT");
252 mtx_unlock_spin(&dt_lock);
254 /* Setup to release spin count in fork_exit(). */
255 td2->td_md.md_spinlock_count = 1;
257 * XXX XEN need to check on PSL_USER is handled
259 td2->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
261 * Now, cpu_switch() can schedule the new process.
262 * pcb_esp is loaded pointing to the cpu_switch() stack frame
263 * containing the return address when exiting cpu_switch.
264 * This will normally be to fork_trampoline(), which will have
265 * %ebx loaded with the new proc's pointer. fork_trampoline()
266 * will set up a stack to call fork_return(p, frame); to complete
267 * the return to user-mode.
272 * Intercept the return address from a freshly forked process that has NOT
273 * been scheduled yet.
275 * This is needed to make kernel threads stay in kernel mode.
278 cpu_set_fork_handler(td, func, arg)
280 void (*func)(void *);
284 * Note that the trap frame follows the args, so the function
285 * is really called like this: func(arg, frame);
287 td->td_pcb->pcb_esi = (int) func; /* function */
288 td->td_pcb->pcb_ebx = (int) arg; /* first arg */
292 cpu_exit(struct thread *td)
296 * If this process has a custom LDT, release it. Reset pc->pcb_gs
297 * and %gs before we free it in case they refer to an LDT entry.
299 mtx_lock_spin(&dt_lock);
300 if (td->td_proc->p_md.md_ldt) {
301 td->td_pcb->pcb_gs = _udatasel;
305 mtx_unlock_spin(&dt_lock);
309 cpu_thread_exit(struct thread *td)
314 if (td == PCPU_GET(fpcurthread))
319 /* Disable any hardware breakpoints. */
320 if (td->td_pcb->pcb_flags & PCB_DBREGS) {
322 td->td_pcb->pcb_flags &= ~PCB_DBREGS;
327 cpu_thread_clean(struct thread *td)
332 if (pcb->pcb_ext != NULL) {
333 /* if (pcb->pcb_ext->ext_refcount-- == 1) ?? */
335 * XXX do we need to move the TSS off the allocated pages
336 * before freeing them? (not done here)
338 kmem_free(kernel_arena, (vm_offset_t)pcb->pcb_ext,
345 cpu_thread_swapin(struct thread *td)
350 cpu_thread_swapout(struct thread *td)
355 cpu_thread_alloc(struct thread *td)
358 td->td_pcb = (struct pcb *)(td->td_kstack +
359 td->td_kstack_pages * PAGE_SIZE) - 1;
360 td->td_frame = (struct trapframe *)((caddr_t)td->td_pcb - 16) - 1;
361 td->td_pcb->pcb_ext = NULL;
362 td->td_pcb->pcb_save = &td->td_pcb->pcb_user_save;
366 cpu_thread_free(struct thread *td)
369 cpu_thread_clean(td);
373 cpu_set_syscall_retval(struct thread *td, int error)
378 td->td_frame->tf_eax = td->td_retval[0];
379 td->td_frame->tf_edx = td->td_retval[1];
380 td->td_frame->tf_eflags &= ~PSL_C;
385 * Reconstruct pc, assuming lcall $X,y is 7 bytes, int
386 * 0x80 is 2 bytes. We saved this in tf_err.
388 td->td_frame->tf_eip -= td->td_frame->tf_err;
395 if (td->td_proc->p_sysent->sv_errsize) {
396 if (error >= td->td_proc->p_sysent->sv_errsize)
397 error = -1; /* XXX */
399 error = td->td_proc->p_sysent->sv_errtbl[error];
401 td->td_frame->tf_eax = error;
402 td->td_frame->tf_eflags |= PSL_C;
408 * Initialize machine state (pcb and trap frame) for a new thread about to
409 * upcall. Put enough state in the new thread's PCB to get it to go back
410 * userret(), where we can intercept it again to set the return (upcall)
411 * Address and stack, along with those from upcals that are from other sources
412 * such as those generated in thread_userret() itself.
415 cpu_set_upcall(struct thread *td, struct thread *td0)
419 /* Point the pcb to the top of the stack. */
423 * Copy the upcall pcb. This loads kernel regs.
424 * Those not loaded individually below get their default
427 bcopy(td0->td_pcb, pcb2, sizeof(*pcb2));
428 pcb2->pcb_flags &= ~(PCB_NPXINITDONE | PCB_NPXUSERINITDONE |
430 pcb2->pcb_save = &pcb2->pcb_user_save;
433 * Create a new fresh stack for the new thread.
435 bcopy(td0->td_frame, td->td_frame, sizeof(struct trapframe));
437 /* If the current thread has the trap bit set (i.e. a debugger had
438 * single stepped the process to the system call), we need to clear
439 * the trap flag from the new frame. Otherwise, the new thread will
440 * receive a (likely unexpected) SIGTRAP when it executes the first
441 * instruction after returning to userland.
443 td->td_frame->tf_eflags &= ~PSL_T;
446 * Set registers for trampoline to user mode. Leave space for the
447 * return address on stack. These are the kernel mode register values.
450 pcb2->pcb_esi = (int)fork_return; /* trampoline arg */
452 pcb2->pcb_esp = (int)td->td_frame - sizeof(void *); /* trampoline arg */
453 pcb2->pcb_ebx = (int)td; /* trampoline arg */
454 pcb2->pcb_eip = (int)fork_trampoline;
455 pcb2->pcb_psl &= ~(PSL_I); /* interrupts must be disabled */
456 pcb2->pcb_gs = rgs();
458 * If we didn't copy the pcb, we'd need to do the following registers:
459 * pcb2->pcb_cr3: cloned above.
460 * pcb2->pcb_dr*: cloned above.
461 * pcb2->pcb_savefpu: cloned above.
462 * pcb2->pcb_flags: cloned above.
463 * pcb2->pcb_onfault: cloned above (always NULL here?).
464 * pcb2->pcb_gs: cloned above.
465 * pcb2->pcb_ext: cleared below.
467 pcb2->pcb_ext = NULL;
469 /* Setup to release spin count in fork_exit(). */
470 td->td_md.md_spinlock_count = 1;
471 td->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
475 * Set that machine state for performing an upcall that has to
476 * be done in thread_userret() so that those upcalls generated
477 * in thread_userret() itself can be done as well.
480 cpu_set_upcall_kse(struct thread *td, void (*entry)(void *), void *arg,
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.
491 cpu_thread_clean(td);
494 * Set the trap frame to point at the beginning of the uts
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;
503 * Pass the address of the mailbox for this kse to the uts
504 * function as a parameter on the stack.
506 suword((void *)(td->td_frame->tf_esp + sizeof(void *)),
511 cpu_set_user_tls(struct thread *td, void *tls_base)
513 struct segment_descriptor sd;
517 * Construct a descriptor and store it in the pcb for
518 * the next context switch. Also store it in the gdt
519 * so that the load of tf_fs into %fs will activate it
520 * at return to userland.
522 base = (uint32_t)tls_base;
523 sd.sd_lobase = base & 0xffffff;
524 sd.sd_hibase = (base >> 24) & 0xff;
525 sd.sd_lolimit = 0xffff; /* 4GB limit, wraps around */
527 sd.sd_type = SDT_MEMRWA;
535 td->td_pcb->pcb_gsd = sd;
536 if (td == curthread) {
537 PCPU_GET(fsgs_gdt)[1] = sd;
538 load_gs(GSEL(GUGS_SEL, SEL_UPL));
545 * Convert kernel VA to physical address
552 pa = pmap_kextract((vm_offset_t)addr);
554 panic("kvtop: zero page frame");
564 cpu_reset_proxy_active = 1;
565 while (cpu_reset_proxy_active == 1)
566 ; /* Wait for other cpu to see that we've started */
567 CPU_SETOF(cpu_reset_proxyid, &tcrp);
569 printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
579 if (arch_i386_is_xbox) {
580 /* Kick the PIC16L, it can reboot the box */
592 CPU_CLR(PCPU_GET(cpuid), &map);
593 CPU_NAND(&map, &stopped_cpus);
594 if (!CPU_EMPTY(&map)) {
595 printf("cpu_reset: Stopping other CPUs\n");
599 if (PCPU_GET(cpuid) != 0) {
600 cpu_reset_proxyid = PCPU_GET(cpuid);
601 cpustop_restartfunc = cpu_reset_proxy;
602 cpu_reset_proxy_active = 0;
603 printf("cpu_reset: Restarting BSP\n");
605 /* Restart CPU #0. */
606 /* XXX: restart_cpus(1 << 0); */
607 CPU_SETOF(0, &started_cpus);
611 while (cpu_reset_proxy_active == 0 && cnt < 10000000)
612 cnt++; /* Wait for BSP to announce restart */
613 if (cpu_reset_proxy_active == 0)
614 printf("cpu_reset: Failed to restart BSP\n");
616 cpu_reset_proxy_active = 2;
632 struct region_descriptor null_idt;
639 if (smp_processor_id() == 0)
640 HYPERVISOR_shutdown(SHUTDOWN_reboot);
642 HYPERVISOR_shutdown(SHUTDOWN_poweroff);
645 if (elan_mmcr != NULL)
646 elan_mmcr->RESCFG = 1;
649 if (cpu == CPU_GEODE1100) {
650 /* Attempt Geode's own reset */
651 outl(0xcf8, 0x80009044ul);
657 * Attempt to do a CPU reset via CPU reset port.
659 if ((inb(0x35) & 0xa0) != 0xa0) {
660 outb(0x37, 0x0f); /* SHUT0 = 0. */
661 outb(0x37, 0x0b); /* SHUT1 = 0. */
663 outb(0xf0, 0x00); /* Reset. */
665 #if !defined(BROKEN_KEYBOARD_RESET)
667 * Attempt to do a CPU reset via the keyboard controller,
668 * do not turn off GateA20, as any machine that fails
669 * to do the reset here would then end up in no man's land.
671 outb(IO_KBD + 4, 0xFE);
672 DELAY(500000); /* wait 0.5 sec to see if that did it */
676 * Attempt to force a reset via the Reset Control register at
677 * I/O port 0xcf9. Bit 2 forces a system reset when it
678 * transitions from 0 to 1. Bit 1 selects the type of reset
679 * to attempt: 0 selects a "soft" reset, and 1 selects a
680 * "hard" reset. We try a "hard" reset. The first write sets
681 * bit 1 to select a "hard" reset and clears bit 2. The
682 * second write forces a 0 -> 1 transition in bit 2 to trigger
687 DELAY(500000); /* wait 0.5 sec to see if that did it */
690 * Attempt to force a reset via the Fast A20 and Init register
691 * at I/O port 0x92. Bit 1 serves as an alternate A20 gate.
692 * Bit 0 asserts INIT# when set to 1. We are careful to only
693 * preserve bit 1 while setting bit 0. We also must clear bit
694 * 0 before setting it if it isn't already clear.
699 outb(0x92, b & 0xfe);
701 DELAY(500000); /* wait 0.5 sec to see if that did it */
705 printf("No known reset method worked, attempting CPU shutdown\n");
706 DELAY(1000000); /* wait 1 sec for printf to complete */
709 null_idt.rd_limit = 0;
710 null_idt.rd_base = 0;
713 /* "good night, sweet prince .... <THUNK!>" */
721 * Get an sf_buf from the freelist. May block if none are available.
724 sf_buf_map(struct sf_buf *sf, int flags)
726 pt_entry_t opte, *ptep;
729 * Update the sf_buf's virtual-to-physical mapping, flushing the
730 * virtual address from the TLB. Since the reference count for
731 * the sf_buf's old mapping was zero, that mapping is not
732 * currently in use. Consequently, there is no need to exchange
733 * the old and new PTEs atomically, even under PAE.
735 ptep = vtopte(sf->kva);
738 PT_SET_MA(sf->kva, xpmap_ptom(VM_PAGE_TO_PHYS(sf->m)) | pgeflag
739 | PG_RW | PG_V | pmap_cache_bits(sf->m->md.pat_mode, 0));
741 *ptep = VM_PAGE_TO_PHYS(sf->m) | pgeflag | PG_RW | PG_V |
742 pmap_cache_bits(sf->m->md.pat_mode, 0);
746 * Avoid unnecessary TLB invalidations: If the sf_buf's old
747 * virtual-to-physical mapping was not used, then any processor
748 * that has invalidated the sf_buf's virtual address from its TLB
749 * since the last used mapping need not invalidate again.
752 if ((opte & (PG_V | PG_A)) == (PG_V | PG_A))
753 CPU_ZERO(&sf->cpumask);
755 sf_buf_shootdown(sf, flags);
757 if ((opte & (PG_V | PG_A)) == (PG_V | PG_A))
758 pmap_invalidate_page(kernel_pmap, sf->kva);
764 sf_buf_shootdown(struct sf_buf *sf, int flags)
770 cpuid = PCPU_GET(cpuid);
771 if (!CPU_ISSET(cpuid, &sf->cpumask)) {
772 CPU_SET(cpuid, &sf->cpumask);
775 if ((flags & SFB_CPUPRIVATE) == 0) {
776 other_cpus = all_cpus;
777 CPU_CLR(cpuid, &other_cpus);
778 CPU_NAND(&other_cpus, &sf->cpumask);
779 if (!CPU_EMPTY(&other_cpus)) {
780 CPU_OR(&sf->cpumask, &other_cpus);
781 smp_masked_invlpg(other_cpus, sf->kva);
789 * MD part of sf_buf_free().
792 sf_buf_unmap(struct sf_buf *sf)
796 * Xen doesn't like having dangling R/W mappings
798 pmap_qremove(sf->kva, 1);
806 sf_buf_invalidate(struct sf_buf *sf)
811 * Use pmap_qenter to update the pte for
812 * existing mapping, in particular, the PAT
813 * settings are recalculated.
815 pmap_qenter(sf->kva, &m, 1);
816 pmap_invalidate_cache_range(sf->kva, sf->kva + PAGE_SIZE);
820 * Invalidate the cache lines that may belong to the page, if
821 * (possibly old) mapping of the page by sf buffer exists. Returns
822 * TRUE when mapping was found and cache invalidated.
825 sf_buf_invalidate_cache(vm_page_t m)
828 return (sf_buf_process_page(m, sf_buf_invalidate));
832 * Software interrupt handler for queued VM system processing.
837 if (busdma_swi_pending != 0)
842 * Tell whether this address is in some physical memory region.
843 * Currently used by the kernel coredump code in order to avoid
844 * dumping the ``ISA memory hole'' which could cause indefinite hangs,
845 * or other unpredictable behaviour.
849 is_physical_memory(vm_paddr_t addr)
853 /* The ISA ``memory hole''. */
854 if (addr >= 0xa0000 && addr < 0x100000)
859 * stuff other tests for known memory-mapped devices (PCI?)