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>
92 #include <isa/isareg.h>
95 #include <machine/xbox.h>
99 #define NSFBUFS (512 + maxusers * 16)
102 _Static_assert(OFFSETOF_CURTHREAD == offsetof(struct pcpu, pc_curthread),
103 "OFFSETOF_CURTHREAD does not correspond with offset of pc_curthread.");
104 _Static_assert(OFFSETOF_CURPCB == offsetof(struct pcpu, pc_curpcb),
105 "OFFSETOF_CURPCB does not correspond with offset of pc_curpcb.");
106 _Static_assert(__OFFSETOF_MONITORBUF == offsetof(struct pcpu, pc_monitorbuf),
107 "__OFFSETOF_MONINORBUF does not correspond with offset of pc_monitorbuf.");
109 static void cpu_reset_real(void);
111 static void cpu_reset_proxy(void);
112 static u_int cpu_reset_proxyid;
113 static volatile u_int cpu_reset_proxy_active;
117 get_pcb_user_save_td(struct thread *td)
121 p = td->td_kstack + td->td_kstack_pages * PAGE_SIZE -
122 roundup2(cpu_max_ext_state_size, XSAVE_AREA_ALIGN);
123 KASSERT((p % XSAVE_AREA_ALIGN) == 0, ("Unaligned pcb_user_save area"));
124 return ((union savefpu *)p);
128 get_pcb_user_save_pcb(struct pcb *pcb)
132 p = (vm_offset_t)(pcb + 1);
133 return ((union savefpu *)p);
137 get_pcb_td(struct thread *td)
141 p = td->td_kstack + td->td_kstack_pages * PAGE_SIZE -
142 roundup2(cpu_max_ext_state_size, XSAVE_AREA_ALIGN) -
144 return ((struct pcb *)p);
148 alloc_fpusave(int flags)
151 struct savefpu_ymm *sf;
153 res = malloc(cpu_max_ext_state_size, M_DEVBUF, flags);
155 sf = (struct savefpu_ymm *)res;
156 bzero(&sf->sv_xstate.sx_hd, sizeof(sf->sv_xstate.sx_hd));
157 sf->sv_xstate.sx_hd.xstate_bv = xsave_mask;
162 * Finish a fork operation, with process p2 nearly set up.
163 * Copy and update the pcb, set up the stack so that the child
164 * ready to run and return to user mode.
167 cpu_fork(struct thread *td1, struct proc *p2, struct thread *td2, int flags)
174 if ((flags & RFPROC) == 0) {
175 if ((flags & RFMEM) == 0) {
176 /* unshare user LDT */
177 struct mdproc *mdp1 = &p1->p_md;
178 struct proc_ldt *pldt, *pldt1;
180 mtx_lock_spin(&dt_lock);
181 if ((pldt1 = mdp1->md_ldt) != NULL &&
182 pldt1->ldt_refcnt > 1) {
183 pldt = user_ldt_alloc(mdp1, pldt1->ldt_len);
185 panic("could not copy LDT");
188 user_ldt_deref(pldt1);
190 mtx_unlock_spin(&dt_lock);
195 /* Ensure that td1's pcb is up to date. */
196 if (td1 == curthread)
197 td1->td_pcb->pcb_gs = rgs();
199 if (PCPU_GET(fpcurthread) == td1)
200 npxsave(td1->td_pcb->pcb_save);
203 /* Point the pcb to the top of the stack */
204 pcb2 = get_pcb_td(td2);
208 bcopy(td1->td_pcb, pcb2, sizeof(*pcb2));
210 /* Properly initialize pcb_save */
211 pcb2->pcb_save = get_pcb_user_save_pcb(pcb2);
212 bcopy(get_pcb_user_save_td(td1), get_pcb_user_save_pcb(pcb2),
213 cpu_max_ext_state_size);
215 /* Point mdproc and then copy over td1's contents */
217 bcopy(&p1->p_md, mdp2, sizeof(*mdp2));
220 * Create a new fresh stack for the new process.
221 * Copy the trap frame for the return to user mode as if from a
222 * syscall. This copies most of the user mode register values.
223 * The -16 is so we can expand the trapframe if we go to vm86.
225 td2->td_frame = (struct trapframe *)((caddr_t)td2->td_pcb - 16) - 1;
226 bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe));
228 td2->td_frame->tf_eax = 0; /* Child returns zero */
229 td2->td_frame->tf_eflags &= ~PSL_C; /* success */
230 td2->td_frame->tf_edx = 1;
233 * If the parent process has the trap bit set (i.e. a debugger had
234 * single stepped the process to the system call), we need to clear
235 * the trap flag from the new frame unless the debugger had set PF_FORK
236 * on the parent. Otherwise, the child will receive a (likely
237 * unexpected) SIGTRAP when it executes the first instruction after
238 * returning to userland.
240 if ((p1->p_pfsflags & PF_FORK) == 0)
241 td2->td_frame->tf_eflags &= ~PSL_T;
244 * Set registers for trampoline to user mode. Leave space for the
245 * return address on stack. These are the kernel mode register values.
247 #if defined(PAE) || defined(PAE_TABLES)
248 pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdpt);
250 pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdir);
253 pcb2->pcb_esi = (int)fork_return; /* fork_trampoline argument */
255 pcb2->pcb_esp = (int)td2->td_frame - sizeof(void *);
256 pcb2->pcb_ebx = (int)td2; /* fork_trampoline argument */
257 pcb2->pcb_eip = (int)fork_trampoline;
259 * pcb2->pcb_dr*: cloned above.
260 * pcb2->pcb_savefpu: cloned above.
261 * pcb2->pcb_flags: cloned above.
262 * pcb2->pcb_onfault: cloned above (always NULL here?).
263 * pcb2->pcb_gs: cloned above.
264 * pcb2->pcb_ext: cleared below.
268 * XXX don't copy the i/o pages. this should probably be fixed.
272 /* Copy the LDT, if necessary. */
273 mtx_lock_spin(&dt_lock);
274 if (mdp2->md_ldt != NULL) {
276 mdp2->md_ldt->ldt_refcnt++;
278 mdp2->md_ldt = user_ldt_alloc(mdp2,
279 mdp2->md_ldt->ldt_len);
280 if (mdp2->md_ldt == NULL)
281 panic("could not copy LDT");
284 mtx_unlock_spin(&dt_lock);
286 /* Setup to release spin count in fork_exit(). */
287 td2->td_md.md_spinlock_count = 1;
288 td2->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
291 * Now, cpu_switch() can schedule the new process.
292 * pcb_esp is loaded pointing to the cpu_switch() stack frame
293 * containing the return address when exiting cpu_switch.
294 * This will normally be to fork_trampoline(), which will have
295 * %ebx loaded with the new proc's pointer. fork_trampoline()
296 * will set up a stack to call fork_return(p, frame); to complete
297 * the return to user-mode.
302 * Intercept the return address from a freshly forked process that has NOT
303 * been scheduled yet.
305 * This is needed to make kernel threads stay in kernel mode.
308 cpu_fork_kthread_handler(struct thread *td, void (*func)(void *), void *arg)
311 * Note that the trap frame follows the args, so the function
312 * is really called like this: func(arg, frame);
314 td->td_pcb->pcb_esi = (int) func; /* function */
315 td->td_pcb->pcb_ebx = (int) arg; /* first arg */
319 cpu_exit(struct thread *td)
323 * If this process has a custom LDT, release it. Reset pc->pcb_gs
324 * and %gs before we free it in case they refer to an LDT entry.
326 mtx_lock_spin(&dt_lock);
327 if (td->td_proc->p_md.md_ldt) {
328 td->td_pcb->pcb_gs = _udatasel;
332 mtx_unlock_spin(&dt_lock);
336 cpu_thread_exit(struct thread *td)
340 if (td == PCPU_GET(fpcurthread))
344 /* Disable any hardware breakpoints. */
345 if (td->td_pcb->pcb_flags & PCB_DBREGS) {
347 td->td_pcb->pcb_flags &= ~PCB_DBREGS;
352 cpu_thread_clean(struct thread *td)
357 if (pcb->pcb_ext != NULL) {
358 /* if (pcb->pcb_ext->ext_refcount-- == 1) ?? */
360 * XXX do we need to move the TSS off the allocated pages
361 * before freeing them? (not done here)
363 kmem_free(kernel_arena, (vm_offset_t)pcb->pcb_ext,
370 cpu_thread_swapin(struct thread *td)
375 cpu_thread_swapout(struct thread *td)
380 cpu_thread_alloc(struct thread *td)
383 struct xstate_hdr *xhdr;
385 td->td_pcb = pcb = get_pcb_td(td);
386 td->td_frame = (struct trapframe *)((caddr_t)pcb - 16) - 1;
388 pcb->pcb_save = get_pcb_user_save_pcb(pcb);
390 xhdr = (struct xstate_hdr *)(pcb->pcb_save + 1);
391 bzero(xhdr, sizeof(*xhdr));
392 xhdr->xstate_bv = xsave_mask;
397 cpu_thread_free(struct thread *td)
400 cpu_thread_clean(td);
404 cpu_set_syscall_retval(struct thread *td, int error)
409 td->td_frame->tf_eax = td->td_retval[0];
410 td->td_frame->tf_edx = td->td_retval[1];
411 td->td_frame->tf_eflags &= ~PSL_C;
416 * Reconstruct pc, assuming lcall $X,y is 7 bytes, int
417 * 0x80 is 2 bytes. We saved this in tf_err.
419 td->td_frame->tf_eip -= td->td_frame->tf_err;
426 td->td_frame->tf_eax = SV_ABI_ERRNO(td->td_proc, error);
427 td->td_frame->tf_eflags |= PSL_C;
433 * Initialize machine state, mostly pcb and trap frame for a new
434 * thread, about to return to userspace. Put enough state in the new
435 * thread's PCB to get it to go back to the fork_return(), which
436 * finalizes the thread state and handles peculiarities of the first
437 * return to userspace for the new thread.
440 cpu_copy_thread(struct thread *td, struct thread *td0)
444 /* Point the pcb to the top of the stack. */
448 * Copy the upcall pcb. This loads kernel regs.
449 * Those not loaded individually below get their default
452 bcopy(td0->td_pcb, pcb2, sizeof(*pcb2));
453 pcb2->pcb_flags &= ~(PCB_NPXINITDONE | PCB_NPXUSERINITDONE |
455 pcb2->pcb_save = get_pcb_user_save_pcb(pcb2);
456 bcopy(get_pcb_user_save_td(td0), pcb2->pcb_save,
457 cpu_max_ext_state_size);
460 * Create a new fresh stack for the new thread.
462 bcopy(td0->td_frame, td->td_frame, sizeof(struct trapframe));
464 /* If the current thread has the trap bit set (i.e. a debugger had
465 * single stepped the process to the system call), we need to clear
466 * the trap flag from the new frame. Otherwise, the new thread will
467 * receive a (likely unexpected) SIGTRAP when it executes the first
468 * instruction after returning to userland.
470 td->td_frame->tf_eflags &= ~PSL_T;
473 * Set registers for trampoline to user mode. Leave space for the
474 * return address on stack. These are the kernel mode register values.
477 pcb2->pcb_esi = (int)fork_return; /* trampoline arg */
479 pcb2->pcb_esp = (int)td->td_frame - sizeof(void *); /* trampoline arg */
480 pcb2->pcb_ebx = (int)td; /* trampoline arg */
481 pcb2->pcb_eip = (int)fork_trampoline;
482 pcb2->pcb_gs = rgs();
484 * If we didn't copy the pcb, we'd need to do the following registers:
485 * pcb2->pcb_cr3: cloned above.
486 * pcb2->pcb_dr*: cloned above.
487 * pcb2->pcb_savefpu: cloned above.
488 * pcb2->pcb_flags: cloned above.
489 * pcb2->pcb_onfault: cloned above (always NULL here?).
490 * pcb2->pcb_gs: cloned above.
491 * pcb2->pcb_ext: cleared below.
493 pcb2->pcb_ext = NULL;
495 /* Setup to release spin count in fork_exit(). */
496 td->td_md.md_spinlock_count = 1;
497 td->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
501 * Set that machine state for performing an upcall that starts
502 * the entry function with the given argument.
505 cpu_set_upcall(struct thread *td, void (*entry)(void *), void *arg,
510 * Do any extra cleaning that needs to be done.
511 * The thread may have optional components
512 * that are not present in a fresh thread.
513 * This may be a recycled thread so make it look
514 * as though it's newly allocated.
516 cpu_thread_clean(td);
519 * Set the trap frame to point at the beginning of the entry
522 td->td_frame->tf_ebp = 0;
523 td->td_frame->tf_esp =
524 (((int)stack->ss_sp + stack->ss_size - 4) & ~0x0f) - 4;
525 td->td_frame->tf_eip = (int)entry;
527 /* Pass the argument to the entry point. */
528 suword((void *)(td->td_frame->tf_esp + sizeof(void *)),
533 cpu_set_user_tls(struct thread *td, void *tls_base)
535 struct segment_descriptor sd;
539 * Construct a descriptor and store it in the pcb for
540 * the next context switch. Also store it in the gdt
541 * so that the load of tf_fs into %fs will activate it
542 * at return to userland.
544 base = (uint32_t)tls_base;
545 sd.sd_lobase = base & 0xffffff;
546 sd.sd_hibase = (base >> 24) & 0xff;
547 sd.sd_lolimit = 0xffff; /* 4GB limit, wraps around */
549 sd.sd_type = SDT_MEMRWA;
557 td->td_pcb->pcb_gsd = sd;
558 if (td == curthread) {
559 PCPU_GET(fsgs_gdt)[1] = sd;
560 load_gs(GSEL(GUGS_SEL, SEL_UPL));
567 * Convert kernel VA to physical address
574 pa = pmap_kextract((vm_offset_t)addr);
576 panic("kvtop: zero page frame");
586 cpu_reset_proxy_active = 1;
587 while (cpu_reset_proxy_active == 1)
588 ; /* Wait for other cpu to see that we've started */
589 CPU_SETOF(cpu_reset_proxyid, &tcrp);
591 printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
601 if (arch_i386_is_xbox) {
602 /* Kick the PIC16L, it can reboot the box */
614 CPU_CLR(PCPU_GET(cpuid), &map);
615 CPU_NAND(&map, &stopped_cpus);
616 if (!CPU_EMPTY(&map)) {
617 printf("cpu_reset: Stopping other CPUs\n");
621 if (PCPU_GET(cpuid) != 0) {
622 cpu_reset_proxyid = PCPU_GET(cpuid);
623 cpustop_restartfunc = cpu_reset_proxy;
624 cpu_reset_proxy_active = 0;
625 printf("cpu_reset: Restarting BSP\n");
627 /* Restart CPU #0. */
628 /* XXX: restart_cpus(1 << 0); */
629 CPU_SETOF(0, &started_cpus);
633 while (cpu_reset_proxy_active == 0 && cnt < 10000000)
634 cnt++; /* Wait for BSP to announce restart */
635 if (cpu_reset_proxy_active == 0)
636 printf("cpu_reset: Failed to restart BSP\n");
638 cpu_reset_proxy_active = 2;
654 struct region_descriptor null_idt;
659 if (elan_mmcr != NULL)
660 elan_mmcr->RESCFG = 1;
663 if (cpu == CPU_GEODE1100) {
664 /* Attempt Geode's own reset */
665 outl(0xcf8, 0x80009044ul);
669 #if !defined(BROKEN_KEYBOARD_RESET)
671 * Attempt to do a CPU reset via the keyboard controller,
672 * do not turn off GateA20, as any machine that fails
673 * to do the reset here would then end up in no man's land.
675 outb(IO_KBD + 4, 0xFE);
676 DELAY(500000); /* wait 0.5 sec to see if that did it */
680 * Attempt to force a reset via the Reset Control register at
681 * I/O port 0xcf9. Bit 2 forces a system reset when it
682 * transitions from 0 to 1. Bit 1 selects the type of reset
683 * to attempt: 0 selects a "soft" reset, and 1 selects a
684 * "hard" reset. We try a "hard" reset. The first write sets
685 * bit 1 to select a "hard" reset and clears bit 2. The
686 * second write forces a 0 -> 1 transition in bit 2 to trigger
691 DELAY(500000); /* wait 0.5 sec to see if that did it */
694 * Attempt to force a reset via the Fast A20 and Init register
695 * at I/O port 0x92. Bit 1 serves as an alternate A20 gate.
696 * Bit 0 asserts INIT# when set to 1. We are careful to only
697 * preserve bit 1 while setting bit 0. We also must clear bit
698 * 0 before setting it if it isn't already clear.
703 outb(0x92, b & 0xfe);
705 DELAY(500000); /* wait 0.5 sec to see if that did it */
708 printf("No known reset method worked, attempting CPU shutdown\n");
709 DELAY(1000000); /* wait 1 sec for printf to complete */
712 null_idt.rd_limit = 0;
713 null_idt.rd_base = 0;
716 /* "good night, sweet prince .... <THUNK!>" */
724 * Get an sf_buf from the freelist. May block if none are available.
727 sf_buf_map(struct sf_buf *sf, int flags)
729 pt_entry_t opte, *ptep;
732 * Update the sf_buf's virtual-to-physical mapping, flushing the
733 * virtual address from the TLB. Since the reference count for
734 * the sf_buf's old mapping was zero, that mapping is not
735 * currently in use. Consequently, there is no need to exchange
736 * the old and new PTEs atomically, even under PAE.
738 ptep = vtopte(sf->kva);
740 *ptep = VM_PAGE_TO_PHYS(sf->m) | pgeflag | PG_RW | PG_V |
741 pmap_cache_bits(sf->m->md.pat_mode, 0);
744 * Avoid unnecessary TLB invalidations: If the sf_buf's old
745 * virtual-to-physical mapping was not used, then any processor
746 * that has invalidated the sf_buf's virtual address from its TLB
747 * since the last used mapping need not invalidate again.
750 if ((opte & (PG_V | PG_A)) == (PG_V | PG_A))
751 CPU_ZERO(&sf->cpumask);
753 sf_buf_shootdown(sf, flags);
755 if ((opte & (PG_V | PG_A)) == (PG_V | PG_A))
756 pmap_invalidate_page(kernel_pmap, sf->kva);
762 sf_buf_shootdown(struct sf_buf *sf, int flags)
768 cpuid = PCPU_GET(cpuid);
769 if (!CPU_ISSET(cpuid, &sf->cpumask)) {
770 CPU_SET(cpuid, &sf->cpumask);
773 if ((flags & SFB_CPUPRIVATE) == 0) {
774 other_cpus = all_cpus;
775 CPU_CLR(cpuid, &other_cpus);
776 CPU_NAND(&other_cpus, &sf->cpumask);
777 if (!CPU_EMPTY(&other_cpus)) {
778 CPU_OR(&sf->cpumask, &other_cpus);
779 smp_masked_invlpg(other_cpus, sf->kva);
787 * MD part of sf_buf_free().
790 sf_buf_unmap(struct sf_buf *sf)
797 sf_buf_invalidate(struct sf_buf *sf)
802 * Use pmap_qenter to update the pte for
803 * existing mapping, in particular, the PAT
804 * settings are recalculated.
806 pmap_qenter(sf->kva, &m, 1);
807 pmap_invalidate_cache_range(sf->kva, sf->kva + PAGE_SIZE, FALSE);
811 * Invalidate the cache lines that may belong to the page, if
812 * (possibly old) mapping of the page by sf buffer exists. Returns
813 * TRUE when mapping was found and cache invalidated.
816 sf_buf_invalidate_cache(vm_page_t m)
819 return (sf_buf_process_page(m, sf_buf_invalidate));
823 * Software interrupt handler for queued VM system processing.
828 if (busdma_swi_pending != 0)
833 * Tell whether this address is in some physical memory region.
834 * Currently used by the kernel coredump code in order to avoid
835 * dumping the ``ISA memory hole'' which could cause indefinite hangs,
836 * or other unpredictable behaviour.
840 is_physical_memory(vm_paddr_t addr)
844 /* The ISA ``memory hole''. */
845 if (addr >= 0xa0000 && addr < 0x100000)
850 * stuff other tests for known memory-mapped devices (PCI?)