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>
57 #include <sys/kernel.h>
60 #include <sys/malloc.h>
62 #include <sys/mutex.h>
63 #include <sys/pioctl.h>
65 #include <sys/refcount.h>
66 #include <sys/sf_buf.h>
68 #include <sys/sched.h>
69 #include <sys/sysctl.h>
70 #include <sys/unistd.h>
71 #include <sys/vnode.h>
72 #include <sys/vmmeter.h>
74 #include <machine/cpu.h>
75 #include <machine/cputypes.h>
76 #include <machine/md_var.h>
77 #include <machine/pcb.h>
78 #include <machine/pcb_ext.h>
79 #include <machine/smp.h>
80 #include <machine/vm86.h>
83 #include <machine/elan_mmcr.h>
87 #include <vm/vm_extern.h>
88 #include <vm/vm_kern.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_param.h>
94 #include <pc98/cbus/cbus.h>
96 #include <i386/isa/isa.h>
100 #include <machine/xbox.h>
104 #define NSFBUFS (512 + maxusers * 16)
107 static void cpu_reset_real(void);
109 static void cpu_reset_proxy(void);
110 static u_int cpu_reset_proxyid;
111 static volatile u_int cpu_reset_proxy_active;
113 static void sf_buf_init(void *arg);
114 SYSINIT(sock_sf, SI_SUB_MBUF, SI_ORDER_ANY, sf_buf_init, NULL)
116 LIST_HEAD(sf_head, sf_buf);
119 * A hash table of active sendfile(2) buffers
121 static struct sf_head *sf_buf_active;
122 static u_long sf_buf_hashmask;
124 #define SF_BUF_HASH(m) (((m) - vm_page_array) & sf_buf_hashmask)
126 static TAILQ_HEAD(, sf_buf) sf_buf_freelist;
127 static u_int sf_buf_alloc_want;
130 * A lock used to synchronize access to the hash table and free list
132 static struct mtx sf_buf_lock;
134 extern int _ucodesel, _udatasel;
137 * Finish a fork operation, with process p2 nearly set up.
138 * Copy and update the pcb, set up the stack so that the child
139 * ready to run and return to user mode.
142 cpu_fork(td1, p2, td2, flags)
143 register struct thread *td1;
144 register struct proc *p2;
148 register struct proc *p1;
156 if ((flags & RFPROC) == 0) {
157 if ((flags & RFMEM) == 0) {
158 /* unshare user LDT */
159 struct mdproc *mdp1 = &p1->p_md;
160 struct proc_ldt *pldt;
162 mtx_lock_spin(&dt_lock);
163 if ((pldt = mdp1->md_ldt) != NULL &&
164 pldt->ldt_refcnt > 1) {
165 pldt = user_ldt_alloc(mdp1, pldt->ldt_len);
167 panic("could not copy LDT");
172 mtx_unlock_spin(&dt_lock);
177 /* Ensure that p1's pcb is up to date. */
178 if (td1 == curthread)
179 td1->td_pcb->pcb_gs = rgs();
181 savecrit = intr_disable();
182 if (PCPU_GET(fpcurthread) == td1)
183 npxsave(&td1->td_pcb->pcb_save);
184 intr_restore(savecrit);
187 /* Point the pcb to the top of the stack */
188 pcb2 = (struct pcb *)(td2->td_kstack +
189 td2->td_kstack_pages * PAGE_SIZE) - 1;
193 bcopy(td1->td_pcb, pcb2, sizeof(*pcb2));
195 /* Point mdproc and then copy over td1's contents */
197 bcopy(&p1->p_md, mdp2, sizeof(*mdp2));
200 * Create a new fresh stack for the new process.
201 * Copy the trap frame for the return to user mode as if from a
202 * syscall. This copies most of the user mode register values.
203 * The -16 is so we can expand the trapframe if we go to vm86.
205 td2->td_frame = (struct trapframe *)((caddr_t)td2->td_pcb - 16) - 1;
206 bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe));
208 td2->td_frame->tf_eax = 0; /* Child returns zero */
209 td2->td_frame->tf_eflags &= ~PSL_C; /* success */
210 td2->td_frame->tf_edx = 1;
213 * If the parent process has the trap bit set (i.e. a debugger had
214 * single stepped the process to the system call), we need to clear
215 * the trap flag from the new frame unless the debugger had set PF_FORK
216 * on the parent. Otherwise, the child will receive a (likely
217 * unexpected) SIGTRAP when it executes the first instruction after
218 * returning to userland.
220 if ((p1->p_pfsflags & PF_FORK) == 0)
221 td2->td_frame->tf_eflags &= ~PSL_T;
224 * Set registers for trampoline to user mode. Leave space for the
225 * return address on stack. These are the kernel mode register values.
228 pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdpt);
230 pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdir);
233 pcb2->pcb_esi = (int)fork_return; /* fork_trampoline argument */
235 pcb2->pcb_esp = (int)td2->td_frame - sizeof(void *);
236 pcb2->pcb_ebx = (int)td2; /* fork_trampoline argument */
237 pcb2->pcb_eip = (int)fork_trampoline;
238 pcb2->pcb_psl = PSL_KERNEL; /* ints disabled */
240 * pcb2->pcb_dr*: cloned above.
241 * pcb2->pcb_savefpu: cloned above.
242 * pcb2->pcb_flags: cloned above.
243 * pcb2->pcb_onfault: cloned above (always NULL here?).
244 * pcb2->pcb_gs: cloned above.
245 * pcb2->pcb_ext: cleared below.
249 * XXX don't copy the i/o pages. this should probably be fixed.
253 /* Copy the LDT, if necessary. */
254 mtx_lock_spin(&dt_lock);
255 if (mdp2->md_ldt != NULL) {
257 refcount_acquire(&mdp2->md_ldt->ldt_refcnt);
259 mdp2->md_ldt = user_ldt_alloc(mdp2,
260 mdp2->md_ldt->ldt_len);
261 if (mdp2->md_ldt == NULL)
262 panic("could not copy LDT");
265 mtx_unlock_spin(&dt_lock);
267 /* Setup to release spin count in fork_exit(). */
268 td2->td_md.md_spinlock_count = 1;
269 td2->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
272 * Now, cpu_switch() can schedule the new process.
273 * pcb_esp is loaded pointing to the cpu_switch() stack frame
274 * containing the return address when exiting cpu_switch.
275 * This will normally be to fork_trampoline(), which will have
276 * %ebx loaded with the new proc's pointer. fork_trampoline()
277 * will set up a stack to call fork_return(p, frame); to complete
278 * the return to user-mode.
283 * Intercept the return address from a freshly forked process that has NOT
284 * been scheduled yet.
286 * This is needed to make kernel threads stay in kernel mode.
289 cpu_set_fork_handler(td, func, arg)
291 void (*func)(void *);
295 * Note that the trap frame follows the args, so the function
296 * is really called like this: func(arg, frame);
298 td->td_pcb->pcb_esi = (int) func; /* function */
299 td->td_pcb->pcb_ebx = (int) arg; /* first arg */
303 cpu_exit(struct thread *td)
307 * If this process has a custom LDT, release it. Reset pc->pcb_gs
308 * and %gs before we free it in case they refer to an LDT entry.
310 mtx_lock_spin(&dt_lock);
311 if (td->td_proc->p_md.md_ldt) {
312 td->td_pcb->pcb_gs = _udatasel;
316 mtx_unlock_spin(&dt_lock);
320 cpu_thread_exit(struct thread *td)
324 if (td == PCPU_GET(fpcurthread))
328 /* Disable any hardware breakpoints. */
329 if (td->td_pcb->pcb_flags & PCB_DBREGS) {
331 td->td_pcb->pcb_flags &= ~PCB_DBREGS;
336 cpu_thread_clean(struct thread *td)
341 if (pcb->pcb_ext != NULL) {
342 /* XXXKSE XXXSMP not SMP SAFE.. what locks do we have? */
343 /* if (pcb->pcb_ext->ext_refcount-- == 1) ?? */
345 * XXX do we need to move the TSS off the allocated pages
346 * before freeing them? (not done here)
348 kmem_free(kernel_map, (vm_offset_t)pcb->pcb_ext,
355 cpu_thread_swapin(struct thread *td)
360 cpu_thread_swapout(struct thread *td)
365 cpu_thread_setup(struct thread *td)
368 td->td_pcb = (struct pcb *)(td->td_kstack +
369 td->td_kstack_pages * PAGE_SIZE) - 1;
370 td->td_frame = (struct trapframe *)((caddr_t)td->td_pcb - 16) - 1;
371 td->td_pcb->pcb_ext = NULL;
375 * Initialize machine state (pcb and trap frame) for a new thread about to
376 * upcall. Put enough state in the new thread's PCB to get it to go back
377 * userret(), where we can intercept it again to set the return (upcall)
378 * Address and stack, along with those from upcals that are from other sources
379 * such as those generated in thread_userret() itself.
382 cpu_set_upcall(struct thread *td, struct thread *td0)
386 /* Point the pcb to the top of the stack. */
390 * Copy the upcall pcb. This loads kernel regs.
391 * Those not loaded individually below get their default
394 * XXXKSE It might be a good idea to simply skip this as
395 * the values of the other registers may be unimportant.
396 * This would remove any requirement for knowing the KSE
397 * at this time (see the matching comment below for
398 * more analysis) (need a good safe default).
400 bcopy(td0->td_pcb, pcb2, sizeof(*pcb2));
401 pcb2->pcb_flags &= ~(PCB_NPXTRAP|PCB_NPXINITDONE);
404 * Create a new fresh stack for the new thread.
405 * The -16 is so we can expand the trapframe if we go to vm86.
406 * Don't forget to set this stack value into whatever supplies
407 * the address for the fault handlers.
408 * The contexts are filled in at the time we actually DO the
409 * upcall as only then do we know which KSE we got.
411 bcopy(td0->td_frame, td->td_frame, sizeof(struct trapframe));
414 * Set registers for trampoline to user mode. Leave space for the
415 * return address on stack. These are the kernel mode register values.
418 pcb2->pcb_cr3 = vtophys(vmspace_pmap(td->td_proc->p_vmspace)->pm_pdpt);
420 pcb2->pcb_cr3 = vtophys(vmspace_pmap(td->td_proc->p_vmspace)->pm_pdir);
423 pcb2->pcb_esi = (int)fork_return; /* trampoline arg */
425 pcb2->pcb_esp = (int)td->td_frame - sizeof(void *); /* trampoline arg */
426 pcb2->pcb_ebx = (int)td; /* trampoline arg */
427 pcb2->pcb_eip = (int)fork_trampoline;
428 pcb2->pcb_psl &= ~(PSL_I); /* interrupts must be disabled */
429 pcb2->pcb_gs = rgs();
431 * If we didn't copy the pcb, we'd need to do the following registers:
432 * pcb2->pcb_dr*: cloned above.
433 * pcb2->pcb_savefpu: cloned above.
434 * pcb2->pcb_flags: cloned above.
435 * pcb2->pcb_onfault: cloned above (always NULL here?).
436 * pcb2->pcb_gs: cloned above. XXXKSE ???
437 * pcb2->pcb_ext: cleared below.
439 pcb2->pcb_ext = NULL;
441 /* Setup to release spin count in fork_exit(). */
442 td->td_md.md_spinlock_count = 1;
443 td->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
447 * Set that machine state for performing an upcall that has to
448 * be done in thread_userret() so that those upcalls generated
449 * in thread_userret() itself can be done as well.
452 cpu_set_upcall_kse(struct thread *td, void (*entry)(void *), void *arg,
457 * Do any extra cleaning that needs to be done.
458 * The thread may have optional components
459 * that are not present in a fresh thread.
460 * This may be a recycled thread so make it look
461 * as though it's newly allocated.
463 cpu_thread_clean(td);
466 * Set the trap frame to point at the beginning of the uts
469 td->td_frame->tf_ebp = 0;
470 td->td_frame->tf_esp =
471 (((int)stack->ss_sp + stack->ss_size - 4) & ~0x0f) - 4;
472 td->td_frame->tf_eip = (int)entry;
475 * Pass the address of the mailbox for this kse to the uts
476 * function as a parameter on the stack.
478 suword((void *)(td->td_frame->tf_esp + sizeof(void *)),
483 cpu_set_user_tls(struct thread *td, void *tls_base)
485 struct segment_descriptor sd;
489 * Construct a descriptor and store it in the pcb for
490 * the next context switch. Also store it in the gdt
491 * so that the load of tf_fs into %fs will activate it
492 * at return to userland.
494 base = (uint32_t)tls_base;
495 sd.sd_lobase = base & 0xffffff;
496 sd.sd_hibase = (base >> 24) & 0xff;
497 sd.sd_lolimit = 0xffff; /* 4GB limit, wraps around */
499 sd.sd_type = SDT_MEMRWA;
507 td->td_pcb->pcb_gsd = sd;
508 if (td == curthread) {
509 PCPU_GET(fsgs_gdt)[1] = sd;
510 load_gs(GSEL(GUGS_SEL, SEL_UPL));
517 * Convert kernel VA to physical address
524 pa = pmap_kextract((vm_offset_t)addr);
526 panic("kvtop: zero page frame");
535 cpu_reset_proxy_active = 1;
536 while (cpu_reset_proxy_active == 1)
537 ; /* Wait for other cpu to see that we've started */
538 stop_cpus((1<<cpu_reset_proxyid));
539 printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
549 if (arch_i386_is_xbox) {
550 /* Kick the PIC16L, it can reboot the box */
560 map = PCPU_GET(other_cpus) & ~stopped_cpus;
562 printf("cpu_reset: Stopping other CPUs\n");
566 if (PCPU_GET(cpuid) != 0) {
567 cpu_reset_proxyid = PCPU_GET(cpuid);
568 cpustop_restartfunc = cpu_reset_proxy;
569 cpu_reset_proxy_active = 0;
570 printf("cpu_reset: Restarting BSP\n");
572 /* Restart CPU #0. */
573 /* XXX: restart_cpus(1 << 0); */
574 atomic_store_rel_int(&started_cpus, (1 << 0));
577 while (cpu_reset_proxy_active == 0 && cnt < 10000000)
578 cnt++; /* Wait for BSP to announce restart */
579 if (cpu_reset_proxy_active == 0)
580 printf("cpu_reset: Failed to restart BSP\n");
582 cpu_reset_proxy_active = 2;
598 struct region_descriptor null_idt;
605 if (elan_mmcr != NULL)
606 elan_mmcr->RESCFG = 1;
609 if (cpu == CPU_GEODE1100) {
610 /* Attempt Geode's own reset */
611 outl(0xcf8, 0x80009044ul);
617 * Attempt to do a CPU reset via CPU reset port.
619 if ((inb(0x35) & 0xa0) != 0xa0) {
620 outb(0x37, 0x0f); /* SHUT0 = 0. */
621 outb(0x37, 0x0b); /* SHUT1 = 0. */
623 outb(0xf0, 0x00); /* Reset. */
625 #if !defined(BROKEN_KEYBOARD_RESET)
627 * Attempt to do a CPU reset via the keyboard controller,
628 * do not turn off GateA20, as any machine that fails
629 * to do the reset here would then end up in no man's land.
631 outb(IO_KBD + 4, 0xFE);
632 DELAY(500000); /* wait 0.5 sec to see if that did it */
636 * Attempt to force a reset via the Reset Control register at
637 * I/O port 0xcf9. Bit 2 forces a system reset when it is
638 * written as 1. Bit 1 selects the type of reset to attempt:
639 * 0 selects a "soft" reset, and 1 selects a "hard" reset. We
640 * try to do a "soft" reset first, and then a "hard" reset.
644 DELAY(500000); /* wait 0.5 sec to see if that did it */
647 * Attempt to force a reset via the Fast A20 and Init register
648 * at I/O port 0x92. Bit 1 serves as an alternate A20 gate.
649 * Bit 0 asserts INIT# when set to 1. We are careful to only
650 * preserve bit 1 while setting bit 0. We also must clear bit
651 * 0 before setting it if it isn't already clear.
656 outb(0x92, b & 0xfe);
658 DELAY(500000); /* wait 0.5 sec to see if that did it */
662 printf("No known reset method worked, attempting CPU shutdown\n");
663 DELAY(1000000); /* wait 1 sec for printf to complete */
666 null_idt.rd_limit = 0;
667 null_idt.rd_base = 0;
670 /* "good night, sweet prince .... <THUNK!>" */
678 * Allocate a pool of sf_bufs (sendfile(2) or "super-fast" if you prefer. :-))
681 sf_buf_init(void *arg)
683 struct sf_buf *sf_bufs;
688 TUNABLE_INT_FETCH("kern.ipc.nsfbufs", &nsfbufs);
690 sf_buf_active = hashinit(nsfbufs, M_TEMP, &sf_buf_hashmask);
691 TAILQ_INIT(&sf_buf_freelist);
692 sf_base = kmem_alloc_nofault(kernel_map, nsfbufs * PAGE_SIZE);
693 sf_bufs = malloc(nsfbufs * sizeof(struct sf_buf), M_TEMP,
695 for (i = 0; i < nsfbufs; i++) {
696 sf_bufs[i].kva = sf_base + i * PAGE_SIZE;
697 TAILQ_INSERT_TAIL(&sf_buf_freelist, &sf_bufs[i], free_entry);
699 sf_buf_alloc_want = 0;
700 mtx_init(&sf_buf_lock, "sf_buf", NULL, MTX_DEF);
704 * Get an sf_buf from the freelist. May block if none are available.
707 sf_buf_alloc(struct vm_page *m, int flags)
709 pt_entry_t opte, *ptep;
710 struct sf_head *hash_list;
713 cpumask_t cpumask, other_cpus;
717 KASSERT(curthread->td_pinned > 0 || (flags & SFB_CPUPRIVATE) == 0,
718 ("sf_buf_alloc(SFB_CPUPRIVATE): curthread not pinned"));
719 hash_list = &sf_buf_active[SF_BUF_HASH(m)];
720 mtx_lock(&sf_buf_lock);
721 LIST_FOREACH(sf, hash_list, list_entry) {
724 if (sf->ref_count == 1) {
725 TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
727 nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
736 while ((sf = TAILQ_FIRST(&sf_buf_freelist)) == NULL) {
737 if (flags & SFB_NOWAIT)
740 mbstat.sf_allocwait++;
741 error = msleep(&sf_buf_freelist, &sf_buf_lock,
742 (flags & SFB_CATCH) ? PCATCH | PVM : PVM, "sfbufa", 0);
746 * If we got a signal, don't risk going back to sleep.
751 TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
753 LIST_REMOVE(sf, list_entry);
754 LIST_INSERT_HEAD(hash_list, sf, list_entry);
758 nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
761 * Update the sf_buf's virtual-to-physical mapping, flushing the
762 * virtual address from the TLB. Since the reference count for
763 * the sf_buf's old mapping was zero, that mapping is not
764 * currently in use. Consequently, there is no need to exchange
765 * the old and new PTEs atomically, even under PAE.
767 ptep = vtopte(sf->kva);
769 *ptep = VM_PAGE_TO_PHYS(m) | pgeflag | PG_RW | PG_V;
772 * Avoid unnecessary TLB invalidations: If the sf_buf's old
773 * virtual-to-physical mapping was not used, then any processor
774 * that has invalidated the sf_buf's virtual address from its TLB
775 * since the last used mapping need not invalidate again.
778 if ((opte & (PG_V | PG_A)) == (PG_V | PG_A))
782 cpumask = PCPU_GET(cpumask);
783 if ((sf->cpumask & cpumask) == 0) {
784 sf->cpumask |= cpumask;
787 if ((flags & SFB_CPUPRIVATE) == 0) {
788 other_cpus = PCPU_GET(other_cpus) & ~sf->cpumask;
789 if (other_cpus != 0) {
790 sf->cpumask |= other_cpus;
791 smp_masked_invlpg(other_cpus, sf->kva);
796 if ((opte & (PG_V | PG_A)) == (PG_V | PG_A))
797 pmap_invalidate_page(kernel_pmap, sf->kva);
800 mtx_unlock(&sf_buf_lock);
805 * Remove a reference from the given sf_buf, adding it to the free
806 * list when its reference count reaches zero. A freed sf_buf still,
807 * however, retains its virtual-to-physical mapping until it is
808 * recycled or reactivated by sf_buf_alloc(9).
811 sf_buf_free(struct sf_buf *sf)
814 mtx_lock(&sf_buf_lock);
816 if (sf->ref_count == 0) {
817 TAILQ_INSERT_TAIL(&sf_buf_freelist, sf, free_entry);
819 if (sf_buf_alloc_want > 0)
820 wakeup_one(&sf_buf_freelist);
822 mtx_unlock(&sf_buf_lock);
826 * Software interrupt handler for queued VM system processing.
831 if (busdma_swi_pending != 0)
836 * Tell whether this address is in some physical memory region.
837 * Currently used by the kernel coredump code in order to avoid
838 * dumping the ``ISA memory hole'' which could cause indefinite hangs,
839 * or other unpredictable behaviour.
843 is_physical_memory(vm_paddr_t addr)
847 /* The ISA ``memory hole''. */
848 if (addr >= 0xa0000 && addr < 0x100000)
853 * stuff other tests for known memory-mapped devices (PCI?)