vm_machdep.c revision 122849
1/*- 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 5 * All rights reserved. 6 * 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. 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 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. 26 * 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 37 * SUCH DAMAGE. 38 * 39 * @(#)vm_machdep.c 7.3 (Berkeley) 5/13/91 40 * Utah $Hdr: vm_machdep.c 1.16.1.1 89/06/23$ 41 */ 42 43#include <sys/cdefs.h> 44__FBSDID("$FreeBSD: head/sys/amd64/amd64/vm_machdep.c 122849 2003-11-17 08:58:16Z peter $"); 45 46#include "opt_isa.h" 47#include "opt_kstack_pages.h" 48 49#include <sys/param.h> 50#include <sys/systm.h> 51#include <sys/malloc.h> 52#include <sys/proc.h> 53#include <sys/kse.h> 54#include <sys/bio.h> 55#include <sys/buf.h> 56#include <sys/vnode.h> 57#include <sys/vmmeter.h> 58#include <sys/kernel.h> 59#include <sys/ktr.h> 60#include <sys/mbuf.h> 61#include <sys/mutex.h> 62#include <sys/sf_buf.h> 63#include <sys/smp.h> 64#include <sys/sysctl.h> 65#include <sys/unistd.h> 66 67#include <machine/cpu.h> 68#include <machine/md_var.h> 69#include <machine/pcb.h> 70 71#include <vm/vm.h> 72#include <vm/vm_param.h> 73#include <sys/lock.h> 74#include <vm/vm_kern.h> 75#include <vm/vm_page.h> 76#include <vm/vm_map.h> 77#include <vm/vm_extern.h> 78 79#include <sys/user.h> 80 81#include <amd64/isa/isa.h> 82 83static void cpu_reset_real(void); 84#ifdef SMP 85static void cpu_reset_proxy(void); 86static u_int cpu_reset_proxyid; 87static volatile u_int cpu_reset_proxy_active; 88#endif 89static void sf_buf_init(void *arg); 90SYSINIT(sock_sf, SI_SUB_MBUF, SI_ORDER_ANY, sf_buf_init, NULL) 91 92/* 93 * Expanded sf_freelist head. Really an SLIST_HEAD() in disguise, with the 94 * sf_freelist head with the sf_lock mutex. 95 */ 96static struct { 97 SLIST_HEAD(, sf_buf) sf_head; 98 struct mtx sf_lock; 99} sf_freelist; 100 101static u_int sf_buf_alloc_want; 102 103/* 104 * Finish a fork operation, with process p2 nearly set up. 105 * Copy and update the pcb, set up the stack so that the child 106 * ready to run and return to user mode. 107 */ 108void 109cpu_fork(td1, p2, td2, flags) 110 register struct thread *td1; 111 register struct proc *p2; 112 struct thread *td2; 113 int flags; 114{ 115 register struct proc *p1; 116 struct pcb *pcb2; 117 struct mdproc *mdp2; 118 register_t savecrit; 119 120 p1 = td1->td_proc; 121 if ((flags & RFPROC) == 0) 122 return; 123 124 /* Ensure that p1's pcb is up to date. */ 125 savecrit = intr_disable(); 126 if (PCPU_GET(fpcurthread) == td1) 127 fpusave(&td1->td_pcb->pcb_save); 128 intr_restore(savecrit); 129 130 /* Point the pcb to the top of the stack */ 131 pcb2 = (struct pcb *)(td2->td_kstack + KSTACK_PAGES * PAGE_SIZE) - 1; 132 td2->td_pcb = pcb2; 133 134 /* Copy p1's pcb */ 135 bcopy(td1->td_pcb, pcb2, sizeof(*pcb2)); 136 137 /* Point mdproc and then copy over td1's contents */ 138 mdp2 = &p2->p_md; 139 bcopy(&p1->p_md, mdp2, sizeof(*mdp2)); 140 141 /* 142 * Create a new fresh stack for the new process. 143 * Copy the trap frame for the return to user mode as if from a 144 * syscall. This copies most of the user mode register values. 145 */ 146 td2->td_frame = (struct trapframe *)td2->td_pcb - 1; 147 bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe)); 148 149 td2->td_frame->tf_rax = 0; /* Child returns zero */ 150 td2->td_frame->tf_rflags &= ~PSL_C; /* success */ 151 td2->td_frame->tf_rdx = 1; 152 153 /* 154 * Set registers for trampoline to user mode. Leave space for the 155 * return address on stack. These are the kernel mode register values. 156 */ 157 pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pml4); 158 pcb2->pcb_r12 = (register_t)fork_return; /* fork_trampoline argument */ 159 pcb2->pcb_rbp = 0; 160 pcb2->pcb_rsp = (register_t)td2->td_frame - sizeof(void *); 161 pcb2->pcb_rbx = (register_t)td2; /* fork_trampoline argument */ 162 pcb2->pcb_rip = (register_t)fork_trampoline; 163 pcb2->pcb_rflags = td2->td_frame->tf_rflags & ~PSL_I; /* ints disabled */ 164 /*- 165 * pcb2->pcb_savefpu: cloned above. 166 * pcb2->pcb_flags: cloned above. 167 * pcb2->pcb_onfault: cloned above (always NULL here?). 168 * pcb2->pcb_[fg]sbase: cloned above 169 */ 170 171 /* 172 * Now, cpu_switch() can schedule the new process. 173 * pcb_rsp is loaded pointing to the cpu_switch() stack frame 174 * containing the return address when exiting cpu_switch. 175 * This will normally be to fork_trampoline(), which will have 176 * %ebx loaded with the new proc's pointer. fork_trampoline() 177 * will set up a stack to call fork_return(p, frame); to complete 178 * the return to user-mode. 179 */ 180} 181 182/* 183 * Intercept the return address from a freshly forked process that has NOT 184 * been scheduled yet. 185 * 186 * This is needed to make kernel threads stay in kernel mode. 187 */ 188void 189cpu_set_fork_handler(td, func, arg) 190 struct thread *td; 191 void (*func)(void *); 192 void *arg; 193{ 194 /* 195 * Note that the trap frame follows the args, so the function 196 * is really called like this: func(arg, frame); 197 */ 198 td->td_pcb->pcb_r12 = (long) func; /* function */ 199 td->td_pcb->pcb_rbx = (long) arg; /* first arg */ 200} 201 202void 203cpu_exit(struct thread *td) 204{ 205 struct mdproc *mdp; 206 207 mdp = &td->td_proc->p_md; 208} 209 210void 211cpu_thread_exit(struct thread *td) 212{ 213 214 if (td == PCPU_GET(fpcurthread)) 215 fpudrop(); 216} 217 218void 219cpu_thread_clean(struct thread *td) 220{ 221} 222 223void 224cpu_thread_swapin(struct thread *td) 225{ 226} 227 228void 229cpu_thread_swapout(struct thread *td) 230{ 231} 232 233void 234cpu_sched_exit(td) 235 register struct thread *td; 236{ 237} 238 239void 240cpu_thread_setup(struct thread *td) 241{ 242 243 td->td_pcb = 244 (struct pcb *)(td->td_kstack + KSTACK_PAGES * PAGE_SIZE) - 1; 245 td->td_frame = (struct trapframe *)td->td_pcb - 1; 246} 247 248/* 249 * Initialize machine state (pcb and trap frame) for a new thread about to 250 * upcall. Pu t enough state in the new thread's PCB to get it to go back 251 * userret(), where we can intercept it again to set the return (upcall) 252 * Address and stack, along with those from upcals that are from other sources 253 * such as those generated in thread_userret() itself. 254 */ 255void 256cpu_set_upcall(struct thread *td, struct thread *td0) 257{ 258 struct pcb *pcb2; 259 260 /* Point the pcb to the top of the stack. */ 261 pcb2 = td->td_pcb; 262 263 /* 264 * Copy the upcall pcb. This loads kernel regs. 265 * Those not loaded individually below get their default 266 * values here. 267 * 268 * XXXKSE It might be a good idea to simply skip this as 269 * the values of the other registers may be unimportant. 270 * This would remove any requirement for knowing the KSE 271 * at this time (see the matching comment below for 272 * more analysis) (need a good safe default). 273 */ 274 bcopy(td0->td_pcb, pcb2, sizeof(*pcb2)); 275 pcb2->pcb_flags &= ~PCB_FPUINITDONE; 276 277 /* 278 * Create a new fresh stack for the new thread. 279 * Don't forget to set this stack value into whatever supplies 280 * the address for the fault handlers. 281 * The contexts are filled in at the time we actually DO the 282 * upcall as only then do we know which KSE we got. 283 */ 284 bcopy(td0->td_frame, td->td_frame, sizeof(struct trapframe)); 285 286 /* 287 * Set registers for trampoline to user mode. Leave space for the 288 * return address on stack. These are the kernel mode register values. 289 */ 290 pcb2->pcb_cr3 = vtophys(vmspace_pmap(td->td_proc->p_vmspace)->pm_pml4); 291 pcb2->pcb_r12 = (register_t)fork_return; /* trampoline arg */ 292 pcb2->pcb_rbp = 0; 293 pcb2->pcb_rsp = (register_t)td->td_frame - sizeof(void *); /* trampoline arg */ 294 pcb2->pcb_rbx = (register_t)td; /* trampoline arg */ 295 pcb2->pcb_rip = (register_t)fork_trampoline; 296 pcb2->pcb_rflags = PSL_KERNEL; /* ints disabled */ 297 /* 298 * If we didn't copy the pcb, we'd need to do the following registers: 299 * pcb2->pcb_savefpu: cloned above. 300 * pcb2->pcb_rflags: cloned above. 301 * pcb2->pcb_onfault: cloned above (always NULL here?). 302 * pcb2->pcb_[fg]sbase: cloned above 303 */ 304} 305 306/* 307 * Set that machine state for performing an upcall that has to 308 * be done in thread_userret() so that those upcalls generated 309 * in thread_userret() itself can be done as well. 310 */ 311void 312cpu_set_upcall_kse(struct thread *td, struct kse_upcall *ku) 313{ 314 315 /* 316 * Do any extra cleaning that needs to be done. 317 * The thread may have optional components 318 * that are not present in a fresh thread. 319 * This may be a recycled thread so make it look 320 * as though it's newly allocated. 321 */ 322 cpu_thread_clean(td); 323 324 /* 325 * Set the trap frame to point at the beginning of the uts 326 * function. 327 */ 328 td->td_frame->tf_rsp = 329 ((register_t)ku->ku_stack.ss_sp + ku->ku_stack.ss_size) & ~0x0f; 330 td->td_frame->tf_rsp -= 8; 331 td->td_frame->tf_rip = (register_t)ku->ku_func; 332 333 /* 334 * Pass the address of the mailbox for this kse to the uts 335 * function as a parameter on the stack. 336 */ 337 td->td_frame->tf_rdi = (register_t)ku->ku_mailbox; 338} 339 340 341/* 342 * Force reset the processor by invalidating the entire address space! 343 */ 344 345#ifdef SMP 346static void 347cpu_reset_proxy() 348{ 349 350 cpu_reset_proxy_active = 1; 351 while (cpu_reset_proxy_active == 1) 352 ; /* Wait for other cpu to see that we've started */ 353 stop_cpus((1<<cpu_reset_proxyid)); 354 printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid); 355 DELAY(1000000); 356 cpu_reset_real(); 357} 358#endif 359 360void 361cpu_reset() 362{ 363#ifdef SMP 364 if (smp_active == 0) { 365 cpu_reset_real(); 366 /* NOTREACHED */ 367 } else { 368 369 u_int map; 370 int cnt; 371 printf("cpu_reset called on cpu#%d\n", PCPU_GET(cpuid)); 372 373 map = PCPU_GET(other_cpus) & ~ stopped_cpus; 374 375 if (map != 0) { 376 printf("cpu_reset: Stopping other CPUs\n"); 377 stop_cpus(map); /* Stop all other CPUs */ 378 } 379 380 if (PCPU_GET(cpuid) == 0) { 381 DELAY(1000000); 382 cpu_reset_real(); 383 /* NOTREACHED */ 384 } else { 385 /* We are not BSP (CPU #0) */ 386 387 cpu_reset_proxyid = PCPU_GET(cpuid); 388 cpustop_restartfunc = cpu_reset_proxy; 389 cpu_reset_proxy_active = 0; 390 printf("cpu_reset: Restarting BSP\n"); 391 started_cpus = (1<<0); /* Restart CPU #0 */ 392 393 cnt = 0; 394 while (cpu_reset_proxy_active == 0 && cnt < 10000000) 395 cnt++; /* Wait for BSP to announce restart */ 396 if (cpu_reset_proxy_active == 0) 397 printf("cpu_reset: Failed to restart BSP\n"); 398 enable_intr(); 399 cpu_reset_proxy_active = 2; 400 401 while (1); 402 /* NOTREACHED */ 403 } 404 } 405#else 406 cpu_reset_real(); 407#endif 408} 409 410static void 411cpu_reset_real() 412{ 413 414 /* 415 * Attempt to do a CPU reset via the keyboard controller, 416 * do not turn of the GateA20, as any machine that fails 417 * to do the reset here would then end up in no man's land. 418 */ 419 420 outb(IO_KBD + 4, 0xFE); 421 DELAY(500000); /* wait 0.5 sec to see if that did it */ 422 printf("Keyboard reset did not work, attempting CPU shutdown\n"); 423 DELAY(1000000); /* wait 1 sec for printf to complete */ 424 /* force a shutdown by unmapping entire address space ! */ 425 bzero((caddr_t)PML4map, PAGE_SIZE); 426 427 /* "good night, sweet prince .... <THUNK!>" */ 428 invltlb(); 429 /* NOTREACHED */ 430 while(1); 431} 432 433/* 434 * Allocate a pool of sf_bufs (sendfile(2) or "super-fast" if you prefer. :-)) 435 */ 436static void 437sf_buf_init(void *arg) 438{ 439 struct sf_buf *sf_bufs; 440 int i; 441 442 mtx_init(&sf_freelist.sf_lock, "sf_bufs list lock", NULL, MTX_DEF); 443 SLIST_INIT(&sf_freelist.sf_head); 444 sf_bufs = malloc(nsfbufs * sizeof(struct sf_buf), M_TEMP, 445 M_NOWAIT | M_ZERO); 446 for (i = 0; i < nsfbufs; i++) 447 SLIST_INSERT_HEAD(&sf_freelist.sf_head, &sf_bufs[i], free_list); 448 sf_buf_alloc_want = 0; 449} 450 451/* 452 * Get an sf_buf from the freelist. Will block if none are available. 453 */ 454struct sf_buf * 455sf_buf_alloc(struct vm_page *m) 456{ 457 struct sf_buf *sf; 458 int error; 459 460 mtx_lock(&sf_freelist.sf_lock); 461 while ((sf = SLIST_FIRST(&sf_freelist.sf_head)) == NULL) { 462 sf_buf_alloc_want++; 463 error = msleep(&sf_freelist, &sf_freelist.sf_lock, PVM|PCATCH, 464 "sfbufa", 0); 465 sf_buf_alloc_want--; 466 467 /* 468 * If we got a signal, don't risk going back to sleep. 469 */ 470 if (error) 471 break; 472 } 473 if (sf != NULL) { 474 SLIST_REMOVE_HEAD(&sf_freelist.sf_head, free_list); 475 sf->m = m; 476 } 477 mtx_unlock(&sf_freelist.sf_lock); 478 return (sf); 479} 480 481/* 482 * Detatch mapped page and release resources back to the system. 483 */ 484void 485sf_buf_free(void *addr, void *args) 486{ 487 struct sf_buf *sf; 488 struct vm_page *m; 489 490 sf = args; 491 m = sf->m; 492 vm_page_lock_queues(); 493 vm_page_unwire(m, 0); 494 /* 495 * Check for the object going away on us. This can 496 * happen since we don't hold a reference to it. 497 * If so, we're responsible for freeing the page. 498 */ 499 if (m->wire_count == 0 && m->object == NULL) 500 vm_page_free(m); 501 vm_page_unlock_queues(); 502 sf->m = NULL; 503 mtx_lock(&sf_freelist.sf_lock); 504 SLIST_INSERT_HEAD(&sf_freelist.sf_head, sf, free_list); 505 if (sf_buf_alloc_want > 0) 506 wakeup_one(&sf_freelist); 507 mtx_unlock(&sf_freelist.sf_lock); 508} 509 510/* 511 * Software interrupt handler for queued VM system processing. 512 */ 513void 514swi_vm(void *dummy) 515{ 516 if (busdma_swi_pending != 0) 517 busdma_swi(); 518} 519 520/* 521 * Tell whether this address is in some physical memory region. 522 * Currently used by the kernel coredump code in order to avoid 523 * dumping the ``ISA memory hole'' which could cause indefinite hangs, 524 * or other unpredictable behaviour. 525 */ 526 527int 528is_physical_memory(addr) 529 vm_offset_t addr; 530{ 531 532#ifdef DEV_ISA 533 /* The ISA ``memory hole''. */ 534 if (addr >= 0xa0000 && addr < 0x100000) 535 return 0; 536#endif 537 538 /* 539 * stuff other tests for known memory-mapped devices (PCI?) 540 * here 541 */ 542 543 return 1; 544} 545