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/sf_buf.h>
66 #include <sys/sched.h>
67 #include <sys/sysctl.h>
68 #include <sys/unistd.h>
69 #include <sys/vnode.h>
70 #include <sys/vmmeter.h>
72 #include <machine/cpu.h>
73 #include <machine/cputypes.h>
74 #include <machine/md_var.h>
75 #include <machine/pcb.h>
76 #include <machine/pcb_ext.h>
77 #include <machine/smp.h>
78 #include <machine/vm86.h>
81 #include <machine/elan_mmcr.h>
85 #include <vm/vm_extern.h>
86 #include <vm/vm_kern.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_param.h>
92 #include <pc98/cbus/cbus.h>
94 #include <i386/isa/isa.h>
98 #include <machine/xbox.h>
102 #define NSFBUFS (512 + maxusers * 16)
105 static void cpu_reset_real(void);
107 static void cpu_reset_proxy(void);
108 static u_int cpu_reset_proxyid;
109 static volatile u_int cpu_reset_proxy_active;
111 static void sf_buf_init(void *arg);
112 SYSINIT(sock_sf, SI_SUB_MBUF, SI_ORDER_ANY, sf_buf_init, NULL);
114 LIST_HEAD(sf_head, sf_buf);
117 * A hash table of active sendfile(2) buffers
119 static struct sf_head *sf_buf_active;
120 static u_long sf_buf_hashmask;
122 #define SF_BUF_HASH(m) (((m) - vm_page_array) & sf_buf_hashmask)
124 static TAILQ_HEAD(, sf_buf) sf_buf_freelist;
125 static u_int sf_buf_alloc_want;
128 * A lock used to synchronize access to the hash table and free list
130 static struct mtx sf_buf_lock;
132 extern int _ucodesel, _udatasel;
135 * Finish a fork operation, with process p2 nearly set up.
136 * Copy and update the pcb, set up the stack so that the child
137 * ready to run and return to user mode.
140 cpu_fork(td1, p2, td2, flags)
141 register struct thread *td1;
142 register struct proc *p2;
146 register struct proc *p1;
154 if ((flags & RFPROC) == 0) {
155 if ((flags & RFMEM) == 0) {
156 /* unshare user LDT */
157 struct mdproc *mdp1 = &p1->p_md;
158 struct proc_ldt *pldt;
160 mtx_lock_spin(&dt_lock);
161 if ((pldt = mdp1->md_ldt) != NULL &&
162 pldt->ldt_refcnt > 1) {
163 pldt = user_ldt_alloc(mdp1, pldt->ldt_len);
165 panic("could not copy LDT");
170 mtx_unlock_spin(&dt_lock);
175 /* Ensure that p1's pcb is up to date. */
176 if (td1 == curthread)
177 td1->td_pcb->pcb_gs = rgs();
179 savecrit = intr_disable();
180 if (PCPU_GET(fpcurthread) == td1)
181 npxsave(&td1->td_pcb->pcb_save);
182 intr_restore(savecrit);
185 /* Point the pcb to the top of the stack */
186 pcb2 = (struct pcb *)(td2->td_kstack +
187 td2->td_kstack_pages * PAGE_SIZE) - 1;
191 bcopy(td1->td_pcb, pcb2, sizeof(*pcb2));
193 /* Point mdproc and then copy over td1's contents */
195 bcopy(&p1->p_md, mdp2, sizeof(*mdp2));
198 * Create a new fresh stack for the new process.
199 * Copy the trap frame for the return to user mode as if from a
200 * syscall. This copies most of the user mode register values.
201 * The -16 is so we can expand the trapframe if we go to vm86.
203 td2->td_frame = (struct trapframe *)((caddr_t)td2->td_pcb - 16) - 1;
204 bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe));
206 td2->td_frame->tf_eax = 0; /* Child returns zero */
207 td2->td_frame->tf_eflags &= ~PSL_C; /* success */
208 td2->td_frame->tf_edx = 1;
211 * If the parent process has the trap bit set (i.e. a debugger had
212 * single stepped the process to the system call), we need to clear
213 * the trap flag from the new frame unless the debugger had set PF_FORK
214 * on the parent. Otherwise, the child will receive a (likely
215 * unexpected) SIGTRAP when it executes the first instruction after
216 * returning to userland.
218 if ((p1->p_pfsflags & PF_FORK) == 0)
219 td2->td_frame->tf_eflags &= ~PSL_T;
222 * Set registers for trampoline to user mode. Leave space for the
223 * return address on stack. These are the kernel mode register values.
226 pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdpt);
228 pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdir);
231 pcb2->pcb_esi = (int)fork_return; /* fork_trampoline argument */
233 pcb2->pcb_esp = (int)td2->td_frame - sizeof(void *);
234 pcb2->pcb_ebx = (int)td2; /* fork_trampoline argument */
235 pcb2->pcb_eip = (int)fork_trampoline;
236 pcb2->pcb_psl = PSL_KERNEL; /* ints disabled */
238 * pcb2->pcb_dr*: cloned above.
239 * pcb2->pcb_savefpu: cloned above.
240 * pcb2->pcb_flags: cloned above.
241 * pcb2->pcb_onfault: cloned above (always NULL here?).
242 * pcb2->pcb_gs: cloned above.
243 * pcb2->pcb_ext: cleared below.
247 * XXX don't copy the i/o pages. this should probably be fixed.
251 /* Copy the LDT, if necessary. */
252 mtx_lock_spin(&dt_lock);
253 if (mdp2->md_ldt != NULL) {
255 mdp2->md_ldt->ldt_refcnt++;
257 mdp2->md_ldt = user_ldt_alloc(mdp2,
258 mdp2->md_ldt->ldt_len);
259 if (mdp2->md_ldt == NULL)
260 panic("could not copy LDT");
263 mtx_unlock_spin(&dt_lock);
265 /* Setup to release spin count in fork_exit(). */
266 td2->td_md.md_spinlock_count = 1;
267 td2->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
270 * Now, cpu_switch() can schedule the new process.
271 * pcb_esp is loaded pointing to the cpu_switch() stack frame
272 * containing the return address when exiting cpu_switch.
273 * This will normally be to fork_trampoline(), which will have
274 * %ebx loaded with the new proc's pointer. fork_trampoline()
275 * will set up a stack to call fork_return(p, frame); to complete
276 * the return to user-mode.
281 * Intercept the return address from a freshly forked process that has NOT
282 * been scheduled yet.
284 * This is needed to make kernel threads stay in kernel mode.
287 cpu_set_fork_handler(td, func, arg)
289 void (*func)(void *);
293 * Note that the trap frame follows the args, so the function
294 * is really called like this: func(arg, frame);
296 td->td_pcb->pcb_esi = (int) func; /* function */
297 td->td_pcb->pcb_ebx = (int) arg; /* first arg */
301 cpu_exit(struct thread *td)
305 * If this process has a custom LDT, release it. Reset pc->pcb_gs
306 * and %gs before we free it in case they refer to an LDT entry.
308 mtx_lock_spin(&dt_lock);
309 if (td->td_proc->p_md.md_ldt) {
310 td->td_pcb->pcb_gs = _udatasel;
314 mtx_unlock_spin(&dt_lock);
318 cpu_thread_exit(struct thread *td)
322 if (td == PCPU_GET(fpcurthread))
326 /* Disable any hardware breakpoints. */
327 if (td->td_pcb->pcb_flags & PCB_DBREGS) {
329 td->td_pcb->pcb_flags &= ~PCB_DBREGS;
334 cpu_thread_clean(struct thread *td)
339 if (pcb->pcb_ext != NULL) {
340 /* if (pcb->pcb_ext->ext_refcount-- == 1) ?? */
342 * XXX do we need to move the TSS off the allocated pages
343 * before freeing them? (not done here)
345 kmem_free(kernel_map, (vm_offset_t)pcb->pcb_ext,
352 cpu_thread_swapin(struct thread *td)
357 cpu_thread_swapout(struct thread *td)
362 cpu_thread_alloc(struct thread *td)
365 td->td_pcb = (struct pcb *)(td->td_kstack +
366 td->td_kstack_pages * PAGE_SIZE) - 1;
367 td->td_frame = (struct trapframe *)((caddr_t)td->td_pcb - 16) - 1;
368 td->td_pcb->pcb_ext = NULL;
372 cpu_thread_free(struct thread *td)
375 cpu_thread_clean(td);
379 * Initialize machine state (pcb and trap frame) for a new thread about to
380 * upcall. Put enough state in the new thread's PCB to get it to go back
381 * userret(), where we can intercept it again to set the return (upcall)
382 * Address and stack, along with those from upcals that are from other sources
383 * such as those generated in thread_userret() itself.
386 cpu_set_upcall(struct thread *td, struct thread *td0)
390 /* Point the pcb to the top of the stack. */
394 * Copy the upcall pcb. This loads kernel regs.
395 * Those not loaded individually below get their default
398 bcopy(td0->td_pcb, pcb2, sizeof(*pcb2));
399 pcb2->pcb_flags &= ~(PCB_NPXTRAP|PCB_NPXINITDONE);
402 * Create a new fresh stack for the new thread.
404 bcopy(td0->td_frame, td->td_frame, sizeof(struct trapframe));
407 * Set registers for trampoline to user mode. Leave space for the
408 * return address on stack. These are the kernel mode register values.
411 pcb2->pcb_cr3 = vtophys(vmspace_pmap(td->td_proc->p_vmspace)->pm_pdpt);
413 pcb2->pcb_cr3 = vtophys(vmspace_pmap(td->td_proc->p_vmspace)->pm_pdir);
416 pcb2->pcb_esi = (int)fork_return; /* trampoline arg */
418 pcb2->pcb_esp = (int)td->td_frame - sizeof(void *); /* trampoline arg */
419 pcb2->pcb_ebx = (int)td; /* trampoline arg */
420 pcb2->pcb_eip = (int)fork_trampoline;
421 pcb2->pcb_psl &= ~(PSL_I); /* interrupts must be disabled */
422 pcb2->pcb_gs = rgs();
424 * If we didn't copy the pcb, we'd need to do the following registers:
425 * pcb2->pcb_dr*: cloned above.
426 * pcb2->pcb_savefpu: cloned above.
427 * pcb2->pcb_flags: cloned above.
428 * pcb2->pcb_onfault: cloned above (always NULL here?).
429 * pcb2->pcb_gs: cloned above.
430 * pcb2->pcb_ext: cleared below.
432 pcb2->pcb_ext = NULL;
434 /* Setup to release spin count in fork_exit(). */
435 td->td_md.md_spinlock_count = 1;
436 td->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
440 * Set that machine state for performing an upcall that has to
441 * be done in thread_userret() so that those upcalls generated
442 * in thread_userret() itself can be done as well.
445 cpu_set_upcall_kse(struct thread *td, void (*entry)(void *), void *arg,
450 * Do any extra cleaning that needs to be done.
451 * The thread may have optional components
452 * that are not present in a fresh thread.
453 * This may be a recycled thread so make it look
454 * as though it's newly allocated.
456 cpu_thread_clean(td);
459 * Set the trap frame to point at the beginning of the uts
462 td->td_frame->tf_ebp = 0;
463 td->td_frame->tf_esp =
464 (((int)stack->ss_sp + stack->ss_size - 4) & ~0x0f) - 4;
465 td->td_frame->tf_eip = (int)entry;
468 * Pass the address of the mailbox for this kse to the uts
469 * function as a parameter on the stack.
471 suword((void *)(td->td_frame->tf_esp + sizeof(void *)),
476 cpu_set_user_tls(struct thread *td, void *tls_base)
478 struct segment_descriptor sd;
482 * Construct a descriptor and store it in the pcb for
483 * the next context switch. Also store it in the gdt
484 * so that the load of tf_fs into %fs will activate it
485 * at return to userland.
487 base = (uint32_t)tls_base;
488 sd.sd_lobase = base & 0xffffff;
489 sd.sd_hibase = (base >> 24) & 0xff;
490 sd.sd_lolimit = 0xffff; /* 4GB limit, wraps around */
492 sd.sd_type = SDT_MEMRWA;
500 td->td_pcb->pcb_gsd = sd;
501 if (td == curthread) {
502 PCPU_GET(fsgs_gdt)[1] = sd;
503 load_gs(GSEL(GUGS_SEL, SEL_UPL));
510 * Convert kernel VA to physical address
517 pa = pmap_kextract((vm_offset_t)addr);
519 panic("kvtop: zero page frame");
528 cpu_reset_proxy_active = 1;
529 while (cpu_reset_proxy_active == 1)
530 ; /* Wait for other cpu to see that we've started */
531 stop_cpus((1<<cpu_reset_proxyid));
532 printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
542 if (arch_i386_is_xbox) {
543 /* Kick the PIC16L, it can reboot the box */
553 map = PCPU_GET(other_cpus) & ~stopped_cpus;
555 printf("cpu_reset: Stopping other CPUs\n");
559 if (PCPU_GET(cpuid) != 0) {
560 cpu_reset_proxyid = PCPU_GET(cpuid);
561 cpustop_restartfunc = cpu_reset_proxy;
562 cpu_reset_proxy_active = 0;
563 printf("cpu_reset: Restarting BSP\n");
565 /* Restart CPU #0. */
566 /* XXX: restart_cpus(1 << 0); */
567 atomic_store_rel_int(&started_cpus, (1 << 0));
570 while (cpu_reset_proxy_active == 0 && cnt < 10000000)
571 cnt++; /* Wait for BSP to announce restart */
572 if (cpu_reset_proxy_active == 0)
573 printf("cpu_reset: Failed to restart BSP\n");
575 cpu_reset_proxy_active = 2;
591 struct region_descriptor null_idt;
598 if (elan_mmcr != NULL)
599 elan_mmcr->RESCFG = 1;
602 if (cpu == CPU_GEODE1100) {
603 /* Attempt Geode's own reset */
604 outl(0xcf8, 0x80009044ul);
610 * Attempt to do a CPU reset via CPU reset port.
612 if ((inb(0x35) & 0xa0) != 0xa0) {
613 outb(0x37, 0x0f); /* SHUT0 = 0. */
614 outb(0x37, 0x0b); /* SHUT1 = 0. */
616 outb(0xf0, 0x00); /* Reset. */
618 #if !defined(BROKEN_KEYBOARD_RESET)
620 * Attempt to do a CPU reset via the keyboard controller,
621 * do not turn off GateA20, as any machine that fails
622 * to do the reset here would then end up in no man's land.
624 outb(IO_KBD + 4, 0xFE);
625 DELAY(500000); /* wait 0.5 sec to see if that did it */
629 * Attempt to force a reset via the Reset Control register at
630 * I/O port 0xcf9. Bit 2 forces a system reset when it is
631 * written as 1. Bit 1 selects the type of reset to attempt:
632 * 0 selects a "soft" reset, and 1 selects a "hard" reset. We
633 * try to do a "soft" reset first, and then a "hard" reset.
637 DELAY(500000); /* wait 0.5 sec to see if that did it */
640 * Attempt to force a reset via the Fast A20 and Init register
641 * at I/O port 0x92. Bit 1 serves as an alternate A20 gate.
642 * Bit 0 asserts INIT# when set to 1. We are careful to only
643 * preserve bit 1 while setting bit 0. We also must clear bit
644 * 0 before setting it if it isn't already clear.
649 outb(0x92, b & 0xfe);
651 DELAY(500000); /* wait 0.5 sec to see if that did it */
655 printf("No known reset method worked, attempting CPU shutdown\n");
656 DELAY(1000000); /* wait 1 sec for printf to complete */
659 null_idt.rd_limit = 0;
660 null_idt.rd_base = 0;
663 /* "good night, sweet prince .... <THUNK!>" */
671 * Allocate a pool of sf_bufs (sendfile(2) or "super-fast" if you prefer. :-))
674 sf_buf_init(void *arg)
676 struct sf_buf *sf_bufs;
681 TUNABLE_INT_FETCH("kern.ipc.nsfbufs", &nsfbufs);
683 sf_buf_active = hashinit(nsfbufs, M_TEMP, &sf_buf_hashmask);
684 TAILQ_INIT(&sf_buf_freelist);
685 sf_base = kmem_alloc_nofault(kernel_map, nsfbufs * PAGE_SIZE);
686 sf_bufs = malloc(nsfbufs * sizeof(struct sf_buf), M_TEMP,
688 for (i = 0; i < nsfbufs; i++) {
689 sf_bufs[i].kva = sf_base + i * PAGE_SIZE;
690 TAILQ_INSERT_TAIL(&sf_buf_freelist, &sf_bufs[i], free_entry);
692 sf_buf_alloc_want = 0;
693 mtx_init(&sf_buf_lock, "sf_buf", NULL, MTX_DEF);
697 * Get an sf_buf from the freelist. May block if none are available.
700 sf_buf_alloc(struct vm_page *m, int flags)
702 pt_entry_t opte, *ptep;
703 struct sf_head *hash_list;
706 cpumask_t cpumask, other_cpus;
710 KASSERT(curthread->td_pinned > 0 || (flags & SFB_CPUPRIVATE) == 0,
711 ("sf_buf_alloc(SFB_CPUPRIVATE): curthread not pinned"));
712 hash_list = &sf_buf_active[SF_BUF_HASH(m)];
713 mtx_lock(&sf_buf_lock);
714 LIST_FOREACH(sf, hash_list, list_entry) {
717 if (sf->ref_count == 1) {
718 TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
720 nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
729 while ((sf = TAILQ_FIRST(&sf_buf_freelist)) == NULL) {
730 if (flags & SFB_NOWAIT)
733 mbstat.sf_allocwait++;
734 error = msleep(&sf_buf_freelist, &sf_buf_lock,
735 (flags & SFB_CATCH) ? PCATCH | PVM : PVM, "sfbufa", 0);
739 * If we got a signal, don't risk going back to sleep.
744 TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
746 LIST_REMOVE(sf, list_entry);
747 LIST_INSERT_HEAD(hash_list, sf, list_entry);
751 nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
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(m) | pgeflag | PG_RW | PG_V;
765 * Avoid unnecessary TLB invalidations: If the sf_buf's old
766 * virtual-to-physical mapping was not used, then any processor
767 * that has invalidated the sf_buf's virtual address from its TLB
768 * since the last used mapping need not invalidate again.
771 if ((opte & (PG_V | PG_A)) == (PG_V | PG_A))
775 cpumask = PCPU_GET(cpumask);
776 if ((sf->cpumask & cpumask) == 0) {
777 sf->cpumask |= cpumask;
780 if ((flags & SFB_CPUPRIVATE) == 0) {
781 other_cpus = PCPU_GET(other_cpus) & ~sf->cpumask;
782 if (other_cpus != 0) {
783 sf->cpumask |= other_cpus;
784 smp_masked_invlpg(other_cpus, sf->kva);
789 if ((opte & (PG_V | PG_A)) == (PG_V | PG_A))
790 pmap_invalidate_page(kernel_pmap, sf->kva);
793 mtx_unlock(&sf_buf_lock);
798 * Remove a reference from the given sf_buf, adding it to the free
799 * list when its reference count reaches zero. A freed sf_buf still,
800 * however, retains its virtual-to-physical mapping until it is
801 * recycled or reactivated by sf_buf_alloc(9).
804 sf_buf_free(struct sf_buf *sf)
807 mtx_lock(&sf_buf_lock);
809 if (sf->ref_count == 0) {
810 TAILQ_INSERT_TAIL(&sf_buf_freelist, sf, free_entry);
812 if (sf_buf_alloc_want > 0)
813 wakeup_one(&sf_buf_freelist);
815 mtx_unlock(&sf_buf_lock);
819 * Software interrupt handler for queued VM system processing.
824 if (busdma_swi_pending != 0)
829 * Tell whether this address is in some physical memory region.
830 * Currently used by the kernel coredump code in order to avoid
831 * dumping the ``ISA memory hole'' which could cause indefinite hangs,
832 * or other unpredictable behaviour.
836 is_physical_memory(vm_paddr_t addr)
840 /* The ISA ``memory hole''. */
841 if (addr >= 0xa0000 && addr < 0x100000)
846 * stuff other tests for known memory-mapped devices (PCI?)