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/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 <i386/isa/isa.h>
99 #include <machine/xbox.h>
103 #define NSFBUFS (512 + maxusers * 16)
106 static void cpu_reset_real(void);
108 static void cpu_reset_proxy(void);
109 static u_int cpu_reset_proxyid;
110 static volatile u_int cpu_reset_proxy_active;
112 static void sf_buf_init(void *arg);
113 SYSINIT(sock_sf, SI_SUB_MBUF, SI_ORDER_ANY, sf_buf_init, NULL)
115 LIST_HEAD(sf_head, sf_buf);
118 * A hash table of active sendfile(2) buffers
120 static struct sf_head *sf_buf_active;
121 static u_long sf_buf_hashmask;
123 #define SF_BUF_HASH(m) (((m) - vm_page_array) & sf_buf_hashmask)
125 static TAILQ_HEAD(, sf_buf) sf_buf_freelist;
126 static u_int sf_buf_alloc_want;
129 * A lock used to synchronize access to the hash table and free list
131 static struct mtx sf_buf_lock;
133 extern int _ucodesel, _udatasel;
136 * Finish a fork operation, with process p2 nearly set up.
137 * Copy and update the pcb, set up the stack so that the child
138 * ready to run and return to user mode.
141 cpu_fork(td1, p2, td2, flags)
142 register struct thread *td1;
143 register struct proc *p2;
147 register struct proc *p1;
155 if ((flags & RFPROC) == 0) {
156 if ((flags & RFMEM) == 0) {
157 /* unshare user LDT */
158 struct mdproc *mdp1 = &p1->p_md;
159 struct proc_ldt *pldt;
161 mtx_lock_spin(&dt_lock);
162 if ((pldt = mdp1->md_ldt) != NULL &&
163 pldt->ldt_refcnt > 1) {
164 pldt = user_ldt_alloc(mdp1, pldt->ldt_len);
166 panic("could not copy LDT");
171 mtx_unlock_spin(&dt_lock);
176 /* Ensure that p1's pcb is up to date. */
177 if (td1 == curthread)
178 td1->td_pcb->pcb_gs = rgs();
180 savecrit = intr_disable();
181 if (PCPU_GET(fpcurthread) == td1)
182 npxsave(&td1->td_pcb->pcb_save);
183 intr_restore(savecrit);
186 /* Point the pcb to the top of the stack */
187 pcb2 = (struct pcb *)(td2->td_kstack +
188 td2->td_kstack_pages * PAGE_SIZE) - 1;
192 bcopy(td1->td_pcb, pcb2, sizeof(*pcb2));
194 /* Point mdproc and then copy over td1's contents */
196 bcopy(&p1->p_md, mdp2, sizeof(*mdp2));
199 * Create a new fresh stack for the new process.
200 * Copy the trap frame for the return to user mode as if from a
201 * syscall. This copies most of the user mode register values.
202 * The -16 is so we can expand the trapframe if we go to vm86.
204 td2->td_frame = (struct trapframe *)((caddr_t)td2->td_pcb - 16) - 1;
205 bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe));
207 td2->td_frame->tf_eax = 0; /* Child returns zero */
208 td2->td_frame->tf_eflags &= ~PSL_C; /* success */
209 td2->td_frame->tf_edx = 1;
212 * If the parent process has the trap bit set (i.e. a debugger had
213 * single stepped the process to the system call), we need to clear
214 * the trap flag from the new frame unless the debugger had set PF_FORK
215 * on the parent. Otherwise, the child will receive a (likely
216 * unexpected) SIGTRAP when it executes the first instruction after
217 * returning to userland.
219 if ((p1->p_pfsflags & PF_FORK) == 0)
220 td2->td_frame->tf_eflags &= ~PSL_T;
223 * Set registers for trampoline to user mode. Leave space for the
224 * return address on stack. These are the kernel mode register values.
227 pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdpt);
229 pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdir);
232 pcb2->pcb_esi = (int)fork_return; /* fork_trampoline argument */
234 pcb2->pcb_esp = (int)td2->td_frame - sizeof(void *);
235 pcb2->pcb_ebx = (int)td2; /* fork_trampoline argument */
236 pcb2->pcb_eip = (int)fork_trampoline;
237 pcb2->pcb_psl = PSL_KERNEL; /* ints disabled */
239 * pcb2->pcb_dr*: cloned above.
240 * pcb2->pcb_savefpu: cloned above.
241 * pcb2->pcb_flags: cloned above.
242 * pcb2->pcb_onfault: cloned above (always NULL here?).
243 * pcb2->pcb_gs: cloned above.
244 * pcb2->pcb_ext: cleared below.
248 * XXX don't copy the i/o pages. this should probably be fixed.
252 /* Copy the LDT, if necessary. */
253 mtx_lock_spin(&dt_lock);
254 if (mdp2->md_ldt != NULL) {
256 mdp2->md_ldt->ldt_refcnt++;
258 mdp2->md_ldt = user_ldt_alloc(mdp2,
259 mdp2->md_ldt->ldt_len);
260 if (mdp2->md_ldt == NULL)
261 panic("could not copy LDT");
264 mtx_unlock_spin(&dt_lock);
266 /* Setup to release spin count in fork_exit(). */
267 td2->td_md.md_spinlock_count = 1;
268 td2->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
271 * Now, cpu_switch() can schedule the new process.
272 * pcb_esp is loaded pointing to the cpu_switch() stack frame
273 * containing the return address when exiting cpu_switch.
274 * This will normally be to fork_trampoline(), which will have
275 * %ebx loaded with the new proc's pointer. fork_trampoline()
276 * will set up a stack to call fork_return(p, frame); to complete
277 * the return to user-mode.
282 * Intercept the return address from a freshly forked process that has NOT
283 * been scheduled yet.
285 * This is needed to make kernel threads stay in kernel mode.
288 cpu_set_fork_handler(td, func, arg)
290 void (*func)(void *);
294 * Note that the trap frame follows the args, so the function
295 * is really called like this: func(arg, frame);
297 td->td_pcb->pcb_esi = (int) func; /* function */
298 td->td_pcb->pcb_ebx = (int) arg; /* first arg */
302 cpu_exit(struct thread *td)
306 * If this process has a custom LDT, release it. Reset pc->pcb_gs
307 * and %gs before we free it in case they refer to an LDT entry.
309 mtx_lock_spin(&dt_lock);
310 if (td->td_proc->p_md.md_ldt) {
311 td->td_pcb->pcb_gs = _udatasel;
315 mtx_unlock_spin(&dt_lock);
319 cpu_thread_exit(struct thread *td)
323 if (td == PCPU_GET(fpcurthread))
327 /* Disable any hardware breakpoints. */
328 if (td->td_pcb->pcb_flags & PCB_DBREGS) {
330 td->td_pcb->pcb_flags &= ~PCB_DBREGS;
335 cpu_thread_clean(struct thread *td)
340 if (pcb->pcb_ext != NULL) {
341 /* XXXKSE XXXSMP not SMP SAFE.. what locks do we have? */
342 /* if (pcb->pcb_ext->ext_refcount-- == 1) ?? */
344 * XXX do we need to move the TSS off the allocated pages
345 * before freeing them? (not done here)
347 kmem_free(kernel_map, (vm_offset_t)pcb->pcb_ext,
354 cpu_thread_swapin(struct thread *td)
359 cpu_thread_swapout(struct thread *td)
364 cpu_thread_setup(struct thread *td)
367 td->td_pcb = (struct pcb *)(td->td_kstack +
368 td->td_kstack_pages * PAGE_SIZE) - 1;
369 td->td_frame = (struct trapframe *)((caddr_t)td->td_pcb - 16) - 1;
370 td->td_pcb->pcb_ext = NULL;
374 * Initialize machine state (pcb and trap frame) for a new thread about to
375 * upcall. Put enough state in the new thread's PCB to get it to go back
376 * userret(), where we can intercept it again to set the return (upcall)
377 * Address and stack, along with those from upcals that are from other sources
378 * such as those generated in thread_userret() itself.
381 cpu_set_upcall(struct thread *td, struct thread *td0)
385 /* Point the pcb to the top of the stack. */
389 * Copy the upcall pcb. This loads kernel regs.
390 * Those not loaded individually below get their default
393 * XXXKSE It might be a good idea to simply skip this as
394 * the values of the other registers may be unimportant.
395 * This would remove any requirement for knowing the KSE
396 * at this time (see the matching comment below for
397 * more analysis) (need a good safe default).
399 bcopy(td0->td_pcb, pcb2, sizeof(*pcb2));
400 pcb2->pcb_flags &= ~(PCB_NPXTRAP|PCB_NPXINITDONE);
403 * Create a new fresh stack for the new thread.
404 * The -16 is so we can expand the trapframe if we go to vm86.
405 * Don't forget to set this stack value into whatever supplies
406 * the address for the fault handlers.
407 * The contexts are filled in at the time we actually DO the
408 * upcall as only then do we know which KSE we got.
410 bcopy(td0->td_frame, td->td_frame, sizeof(struct trapframe));
413 * Set registers for trampoline to user mode. Leave space for the
414 * return address on stack. These are the kernel mode register values.
417 pcb2->pcb_cr3 = vtophys(vmspace_pmap(td->td_proc->p_vmspace)->pm_pdpt);
419 pcb2->pcb_cr3 = vtophys(vmspace_pmap(td->td_proc->p_vmspace)->pm_pdir);
422 pcb2->pcb_esi = (int)fork_return; /* trampoline arg */
424 pcb2->pcb_esp = (int)td->td_frame - sizeof(void *); /* trampoline arg */
425 pcb2->pcb_ebx = (int)td; /* trampoline arg */
426 pcb2->pcb_eip = (int)fork_trampoline;
427 pcb2->pcb_psl &= ~(PSL_I); /* interrupts must be disabled */
428 pcb2->pcb_gs = rgs();
430 * If we didn't copy the pcb, we'd need to do the following registers:
431 * pcb2->pcb_dr*: cloned above.
432 * pcb2->pcb_savefpu: cloned above.
433 * pcb2->pcb_flags: cloned above.
434 * pcb2->pcb_onfault: cloned above (always NULL here?).
435 * pcb2->pcb_gs: cloned above. XXXKSE ???
436 * pcb2->pcb_ext: cleared below.
438 pcb2->pcb_ext = NULL;
440 /* Setup to release spin count in fork_exit(). */
441 td->td_md.md_spinlock_count = 1;
442 td->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
446 * Set that machine state for performing an upcall that has to
447 * be done in thread_userret() so that those upcalls generated
448 * in thread_userret() itself can be done as well.
451 cpu_set_upcall_kse(struct thread *td, void (*entry)(void *), void *arg,
456 * Do any extra cleaning that needs to be done.
457 * The thread may have optional components
458 * that are not present in a fresh thread.
459 * This may be a recycled thread so make it look
460 * as though it's newly allocated.
462 cpu_thread_clean(td);
465 * Set the trap frame to point at the beginning of the uts
468 td->td_frame->tf_ebp = 0;
469 td->td_frame->tf_esp =
470 (((int)stack->ss_sp + stack->ss_size - 4) & ~0x0f) - 4;
471 td->td_frame->tf_eip = (int)entry;
474 * Pass the address of the mailbox for this kse to the uts
475 * function as a parameter on the stack.
477 suword((void *)(td->td_frame->tf_esp + sizeof(void *)),
482 cpu_set_user_tls(struct thread *td, void *tls_base)
484 struct segment_descriptor sd;
488 * Construct a descriptor and store it in the pcb for
489 * the next context switch. Also store it in the gdt
490 * so that the load of tf_fs into %fs will activate it
491 * at return to userland.
493 base = (uint32_t)tls_base;
494 sd.sd_lobase = base & 0xffffff;
495 sd.sd_hibase = (base >> 24) & 0xff;
496 sd.sd_lolimit = 0xffff; /* 4GB limit, wraps around */
498 sd.sd_type = SDT_MEMRWA;
506 td->td_pcb->pcb_gsd = sd;
507 if (td == curthread) {
508 PCPU_GET(fsgs_gdt)[1] = sd;
509 load_gs(GSEL(GUGS_SEL, SEL_UPL));
516 * Convert kernel VA to physical address
523 pa = pmap_kextract((vm_offset_t)addr);
525 panic("kvtop: zero page frame");
534 cpu_reset_proxy_active = 1;
535 while (cpu_reset_proxy_active == 1)
536 ; /* Wait for other cpu to see that we've started */
537 stop_cpus((1<<cpu_reset_proxyid));
538 printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
548 if (arch_i386_is_xbox) {
549 /* Kick the PIC16L, it can reboot the box */
559 map = PCPU_GET(other_cpus) & ~stopped_cpus;
561 printf("cpu_reset: Stopping other CPUs\n");
565 if (PCPU_GET(cpuid) != 0) {
566 cpu_reset_proxyid = PCPU_GET(cpuid);
567 cpustop_restartfunc = cpu_reset_proxy;
568 cpu_reset_proxy_active = 0;
569 printf("cpu_reset: Restarting BSP\n");
571 /* Restart CPU #0. */
572 /* XXX: restart_cpus(1 << 0); */
573 atomic_store_rel_int(&started_cpus, (1 << 0));
576 while (cpu_reset_proxy_active == 0 && cnt < 10000000)
577 cnt++; /* Wait for BSP to announce restart */
578 if (cpu_reset_proxy_active == 0)
579 printf("cpu_reset: Failed to restart BSP\n");
581 cpu_reset_proxy_active = 2;
597 struct region_descriptor null_idt;
604 if (elan_mmcr != NULL)
605 elan_mmcr->RESCFG = 1;
608 if (cpu == CPU_GEODE1100) {
609 /* Attempt Geode's own reset */
610 outl(0xcf8, 0x80009044ul);
616 * Attempt to do a CPU reset via CPU reset port.
618 if ((inb(0x35) & 0xa0) != 0xa0) {
619 outb(0x37, 0x0f); /* SHUT0 = 0. */
620 outb(0x37, 0x0b); /* SHUT1 = 0. */
622 outb(0xf0, 0x00); /* Reset. */
624 #if !defined(BROKEN_KEYBOARD_RESET)
626 * Attempt to do a CPU reset via the keyboard controller,
627 * do not turn off GateA20, as any machine that fails
628 * to do the reset here would then end up in no man's land.
630 outb(IO_KBD + 4, 0xFE);
631 DELAY(500000); /* wait 0.5 sec to see if that did it */
635 * Attempt to force a reset via the Reset Control register at
636 * I/O port 0xcf9. Bit 2 forces a system reset when it is
637 * written as 1. Bit 1 selects the type of reset to attempt:
638 * 0 selects a "soft" reset, and 1 selects a "hard" reset. We
639 * try to do a "soft" reset first, and then a "hard" reset.
643 DELAY(500000); /* wait 0.5 sec to see if that did it */
646 * Attempt to force a reset via the Fast A20 and Init register
647 * at I/O port 0x92. Bit 1 serves as an alternate A20 gate.
648 * Bit 0 asserts INIT# when set to 1. We are careful to only
649 * preserve bit 1 while setting bit 0. We also must clear bit
650 * 0 before setting it if it isn't already clear.
655 outb(0x92, b & 0xfe);
657 DELAY(500000); /* wait 0.5 sec to see if that did it */
661 printf("No known reset method worked, attempting CPU shutdown\n");
662 DELAY(1000000); /* wait 1 sec for printf to complete */
665 null_idt.rd_limit = 0;
666 null_idt.rd_base = 0;
669 /* "good night, sweet prince .... <THUNK!>" */
677 * Allocate a pool of sf_bufs (sendfile(2) or "super-fast" if you prefer. :-))
680 sf_buf_init(void *arg)
682 struct sf_buf *sf_bufs;
687 TUNABLE_INT_FETCH("kern.ipc.nsfbufs", &nsfbufs);
689 sf_buf_active = hashinit(nsfbufs, M_TEMP, &sf_buf_hashmask);
690 TAILQ_INIT(&sf_buf_freelist);
691 sf_base = kmem_alloc_nofault(kernel_map, nsfbufs * PAGE_SIZE);
692 sf_bufs = malloc(nsfbufs * sizeof(struct sf_buf), M_TEMP,
694 for (i = 0; i < nsfbufs; i++) {
695 sf_bufs[i].kva = sf_base + i * PAGE_SIZE;
696 TAILQ_INSERT_TAIL(&sf_buf_freelist, &sf_bufs[i], free_entry);
698 sf_buf_alloc_want = 0;
699 mtx_init(&sf_buf_lock, "sf_buf", NULL, MTX_DEF);
703 * Get an sf_buf from the freelist. May block if none are available.
706 sf_buf_alloc(struct vm_page *m, int flags)
708 pt_entry_t opte, *ptep;
709 struct sf_head *hash_list;
712 cpumask_t cpumask, other_cpus;
716 KASSERT(curthread->td_pinned > 0 || (flags & SFB_CPUPRIVATE) == 0,
717 ("sf_buf_alloc(SFB_CPUPRIVATE): curthread not pinned"));
718 hash_list = &sf_buf_active[SF_BUF_HASH(m)];
719 mtx_lock(&sf_buf_lock);
720 LIST_FOREACH(sf, hash_list, list_entry) {
723 if (sf->ref_count == 1) {
724 TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
726 nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
735 while ((sf = TAILQ_FIRST(&sf_buf_freelist)) == NULL) {
736 if (flags & SFB_NOWAIT)
739 mbstat.sf_allocwait++;
740 error = msleep(&sf_buf_freelist, &sf_buf_lock,
741 (flags & SFB_CATCH) ? PCATCH | PVM : PVM, "sfbufa", 0);
745 * If we got a signal, don't risk going back to sleep.
750 TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
752 LIST_REMOVE(sf, list_entry);
753 LIST_INSERT_HEAD(hash_list, sf, list_entry);
757 nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
760 * Update the sf_buf's virtual-to-physical mapping, flushing the
761 * virtual address from the TLB. Since the reference count for
762 * the sf_buf's old mapping was zero, that mapping is not
763 * currently in use. Consequently, there is no need to exchange
764 * the old and new PTEs atomically, even under PAE.
766 ptep = vtopte(sf->kva);
768 *ptep = VM_PAGE_TO_PHYS(m) | pgeflag | PG_RW | PG_V;
771 * Avoid unnecessary TLB invalidations: If the sf_buf's old
772 * virtual-to-physical mapping was not used, then any processor
773 * that has invalidated the sf_buf's virtual address from its TLB
774 * since the last used mapping need not invalidate again.
777 if ((opte & (PG_V | PG_A)) == (PG_V | PG_A))
781 cpumask = PCPU_GET(cpumask);
782 if ((sf->cpumask & cpumask) == 0) {
783 sf->cpumask |= cpumask;
786 if ((flags & SFB_CPUPRIVATE) == 0) {
787 other_cpus = PCPU_GET(other_cpus) & ~sf->cpumask;
788 if (other_cpus != 0) {
789 sf->cpumask |= other_cpus;
790 smp_masked_invlpg(other_cpus, sf->kva);
795 if ((opte & (PG_V | PG_A)) == (PG_V | PG_A))
796 pmap_invalidate_page(kernel_pmap, sf->kva);
799 mtx_unlock(&sf_buf_lock);
804 * Remove a reference from the given sf_buf, adding it to the free
805 * list when its reference count reaches zero. A freed sf_buf still,
806 * however, retains its virtual-to-physical mapping until it is
807 * recycled or reactivated by sf_buf_alloc(9).
810 sf_buf_free(struct sf_buf *sf)
813 mtx_lock(&sf_buf_lock);
815 if (sf->ref_count == 0) {
816 TAILQ_INSERT_TAIL(&sf_buf_freelist, sf, free_entry);
818 if (sf_buf_alloc_want > 0)
819 wakeup_one(&sf_buf_freelist);
821 mtx_unlock(&sf_buf_lock);
825 * Software interrupt handler for queued VM system processing.
830 if (busdma_swi_pending != 0)
835 * Tell whether this address is in some physical memory region.
836 * Currently used by the kernel coredump code in order to avoid
837 * dumping the ``ISA memory hole'' which could cause indefinite hangs,
838 * or other unpredictable behaviour.
842 is_physical_memory(vm_paddr_t addr)
846 /* The ISA ``memory hole''. */
847 if (addr >= 0xa0000 && addr < 0x100000)
852 * stuff other tests for known memory-mapped devices (PCI?)