vm_machdep.c revision 121751
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 121751 2003-10-30 19:04:58Z 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/socketvar.h> 63#include <sys/sysctl.h> 64#include <sys/unistd.h> 65 66#include <machine/cpu.h> 67#include <machine/md_var.h> 68#include <machine/pcb.h> 69 70#include <vm/vm.h> 71#include <vm/vm_param.h> 72#include <sys/lock.h> 73#include <vm/vm_kern.h> 74#include <vm/vm_page.h> 75#include <vm/vm_map.h> 76#include <vm/vm_extern.h> 77 78#include <sys/user.h> 79 80#include <amd64/isa/isa.h> 81 82static void cpu_reset_real(void); 83static void sf_buf_init(void *arg); 84SYSINIT(sock_sf, SI_SUB_MBUF, SI_ORDER_ANY, sf_buf_init, NULL) 85 86/* 87 * Expanded sf_freelist head. Really an SLIST_HEAD() in disguise, with the 88 * sf_freelist head with the sf_lock mutex. 89 */ 90static struct { 91 SLIST_HEAD(, sf_buf) sf_head; 92 struct mtx sf_lock; 93} sf_freelist; 94 95static u_int sf_buf_alloc_want; 96 97/* 98 * Finish a fork operation, with process p2 nearly set up. 99 * Copy and update the pcb, set up the stack so that the child 100 * ready to run and return to user mode. 101 */ 102void 103cpu_fork(td1, p2, td2, flags) 104 register struct thread *td1; 105 register struct proc *p2; 106 struct thread *td2; 107 int flags; 108{ 109 register struct proc *p1; 110 struct pcb *pcb2; 111 struct mdproc *mdp2; 112 register_t savecrit; 113 114 p1 = td1->td_proc; 115 if ((flags & RFPROC) == 0) 116 return; 117 118 /* Ensure that p1's pcb is up to date. */ 119 savecrit = intr_disable(); 120 if (PCPU_GET(fpcurthread) == td1) 121 npxsave(&td1->td_pcb->pcb_save); 122 intr_restore(savecrit); 123 124 /* Point the pcb to the top of the stack */ 125 pcb2 = (struct pcb *)(td2->td_kstack + KSTACK_PAGES * PAGE_SIZE) - 1; 126 td2->td_pcb = pcb2; 127 128 /* Copy p1's pcb */ 129 bcopy(td1->td_pcb, pcb2, sizeof(*pcb2)); 130 131 /* Point mdproc and then copy over td1's contents */ 132 mdp2 = &p2->p_md; 133 bcopy(&p1->p_md, mdp2, sizeof(*mdp2)); 134 135 /* 136 * Create a new fresh stack for the new process. 137 * Copy the trap frame for the return to user mode as if from a 138 * syscall. This copies most of the user mode register values. 139 */ 140 td2->td_frame = (struct trapframe *)td2->td_pcb - 1; 141 bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe)); 142 143 td2->td_frame->tf_rax = 0; /* Child returns zero */ 144 td2->td_frame->tf_rflags &= ~PSL_C; /* success */ 145 td2->td_frame->tf_rdx = 1; 146 147 /* 148 * Set registers for trampoline to user mode. Leave space for the 149 * return address on stack. These are the kernel mode register values. 150 */ 151 pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pml4); 152 pcb2->pcb_r12 = (register_t)fork_return; /* fork_trampoline argument */ 153 pcb2->pcb_rbp = 0; 154 pcb2->pcb_rsp = (register_t)td2->td_frame - sizeof(void *); 155 pcb2->pcb_rbx = (register_t)td2; /* fork_trampoline argument */ 156 pcb2->pcb_rip = (register_t)fork_trampoline; 157 pcb2->pcb_rflags = td2->td_frame->tf_rflags & ~PSL_I; /* ints disabled */ 158 /*- 159 * pcb2->pcb_savefpu: cloned above. 160 * pcb2->pcb_flags: cloned above. 161 * pcb2->pcb_onfault: cloned above (always NULL here?). 162 * pcb2->pcb_[fg]sbase: cloned above 163 */ 164 165 /* 166 * Now, cpu_switch() can schedule the new process. 167 * pcb_rsp is loaded pointing to the cpu_switch() stack frame 168 * containing the return address when exiting cpu_switch. 169 * This will normally be to fork_trampoline(), which will have 170 * %ebx loaded with the new proc's pointer. fork_trampoline() 171 * will set up a stack to call fork_return(p, frame); to complete 172 * the return to user-mode. 173 */ 174} 175 176/* 177 * Intercept the return address from a freshly forked process that has NOT 178 * been scheduled yet. 179 * 180 * This is needed to make kernel threads stay in kernel mode. 181 */ 182void 183cpu_set_fork_handler(td, func, arg) 184 struct thread *td; 185 void (*func)(void *); 186 void *arg; 187{ 188 /* 189 * Note that the trap frame follows the args, so the function 190 * is really called like this: func(arg, frame); 191 */ 192 td->td_pcb->pcb_r12 = (long) func; /* function */ 193 td->td_pcb->pcb_rbx = (long) arg; /* first arg */ 194} 195 196void 197cpu_exit(struct thread *td) 198{ 199 struct mdproc *mdp; 200 201 mdp = &td->td_proc->p_md; 202} 203 204void 205cpu_thread_exit(struct thread *td) 206{ 207 208 if (td == PCPU_GET(fpcurthread)) 209 npxdrop(); 210} 211 212void 213cpu_thread_clean(struct thread *td) 214{ 215} 216 217void 218cpu_thread_swapin(struct thread *td) 219{ 220} 221 222void 223cpu_thread_swapout(struct thread *td) 224{ 225} 226 227void 228cpu_sched_exit(td) 229 register struct thread *td; 230{ 231} 232 233void 234cpu_thread_setup(struct thread *td) 235{ 236 237 td->td_pcb = 238 (struct pcb *)(td->td_kstack + KSTACK_PAGES * PAGE_SIZE) - 1; 239 td->td_frame = (struct trapframe *)td->td_pcb - 1; 240} 241 242/* 243 * Initialize machine state (pcb and trap frame) for a new thread about to 244 * upcall. Pu t enough state in the new thread's PCB to get it to go back 245 * userret(), where we can intercept it again to set the return (upcall) 246 * Address and stack, along with those from upcals that are from other sources 247 * such as those generated in thread_userret() itself. 248 */ 249void 250cpu_set_upcall(struct thread *td, struct thread *td0) 251{ 252 struct pcb *pcb2; 253 254 /* Point the pcb to the top of the stack. */ 255 pcb2 = td->td_pcb; 256 257 /* 258 * Copy the upcall pcb. This loads kernel regs. 259 * Those not loaded individually below get their default 260 * values here. 261 * 262 * XXXKSE It might be a good idea to simply skip this as 263 * the values of the other registers may be unimportant. 264 * This would remove any requirement for knowing the KSE 265 * at this time (see the matching comment below for 266 * more analysis) (need a good safe default). 267 */ 268 bcopy(td0->td_pcb, pcb2, sizeof(*pcb2)); 269 pcb2->pcb_flags &= ~(PCB_NPXTRAP|PCB_NPXINITDONE); 270 271 /* 272 * Create a new fresh stack for the new thread. 273 * Don't forget to set this stack value into whatever supplies 274 * the address for the fault handlers. 275 * The contexts are filled in at the time we actually DO the 276 * upcall as only then do we know which KSE we got. 277 */ 278 bcopy(td0->td_frame, td->td_frame, sizeof(struct trapframe)); 279 280 /* 281 * Set registers for trampoline to user mode. Leave space for the 282 * return address on stack. These are the kernel mode register values. 283 */ 284 pcb2->pcb_cr3 = vtophys(vmspace_pmap(td->td_proc->p_vmspace)->pm_pml4); 285 pcb2->pcb_r12 = (register_t)fork_return; /* trampoline arg */ 286 pcb2->pcb_rbp = 0; 287 pcb2->pcb_rsp = (register_t)td->td_frame - sizeof(void *); /* trampoline arg */ 288 pcb2->pcb_rbx = (register_t)td; /* trampoline arg */ 289 pcb2->pcb_rip = (register_t)fork_trampoline; 290 pcb2->pcb_rflags = PSL_KERNEL; /* ints disabled */ 291 /* 292 * If we didn't copy the pcb, we'd need to do the following registers: 293 * pcb2->pcb_savefpu: cloned above. 294 * pcb2->pcb_rflags: cloned above. 295 * pcb2->pcb_onfault: cloned above (always NULL here?). 296 * pcb2->pcb_[fg]sbase: cloned above 297 */ 298} 299 300/* 301 * Set that machine state for performing an upcall that has to 302 * be done in thread_userret() so that those upcalls generated 303 * in thread_userret() itself can be done as well. 304 */ 305void 306cpu_set_upcall_kse(struct thread *td, struct kse_upcall *ku) 307{ 308 309 /* 310 * Do any extra cleaning that needs to be done. 311 * The thread may have optional components 312 * that are not present in a fresh thread. 313 * This may be a recycled thread so make it look 314 * as though it's newly allocated. 315 */ 316 cpu_thread_clean(td); 317 318 /* 319 * Set the trap frame to point at the beginning of the uts 320 * function. 321 */ 322 td->td_frame->tf_rsp = 323 ((register_t)ku->ku_stack.ss_sp + ku->ku_stack.ss_size) & ~0x0f; 324 td->td_frame->tf_rsp -= 8; 325 td->td_frame->tf_rip = (register_t)ku->ku_func; 326 327 /* 328 * Pass the address of the mailbox for this kse to the uts 329 * function as a parameter on the stack. 330 */ 331 td->td_frame->tf_rdi = (register_t)ku->ku_mailbox; 332} 333 334 335/* 336 * Force reset the processor by invalidating the entire address space! 337 */ 338 339void 340cpu_reset() 341{ 342 cpu_reset_real(); 343} 344 345static void 346cpu_reset_real() 347{ 348 349 /* 350 * Attempt to do a CPU reset via the keyboard controller, 351 * do not turn of the GateA20, as any machine that fails 352 * to do the reset here would then end up in no man's land. 353 */ 354 355 outb(IO_KBD + 4, 0xFE); 356 DELAY(500000); /* wait 0.5 sec to see if that did it */ 357 printf("Keyboard reset did not work, attempting CPU shutdown\n"); 358 DELAY(1000000); /* wait 1 sec for printf to complete */ 359 /* force a shutdown by unmapping entire address space ! */ 360 bzero((caddr_t)PML4map, PAGE_SIZE); 361 362 /* "good night, sweet prince .... <THUNK!>" */ 363 invltlb(); 364 /* NOTREACHED */ 365 while(1); 366} 367 368/* 369 * Allocate a pool of sf_bufs (sendfile(2) or "super-fast" if you prefer. :-)) 370 */ 371static void 372sf_buf_init(void *arg) 373{ 374 struct sf_buf *sf_bufs; 375 int i; 376 377 mtx_init(&sf_freelist.sf_lock, "sf_bufs list lock", NULL, MTX_DEF); 378 mtx_lock(&sf_freelist.sf_lock); 379 SLIST_INIT(&sf_freelist.sf_head); 380 sf_bufs = malloc(nsfbufs * sizeof(struct sf_buf), M_TEMP, 381 M_NOWAIT | M_ZERO); 382 for (i = 0; i < nsfbufs; i++) 383 SLIST_INSERT_HEAD(&sf_freelist.sf_head, &sf_bufs[i], free_list); 384 sf_buf_alloc_want = 0; 385 mtx_unlock(&sf_freelist.sf_lock); 386} 387 388/* 389 * Get an sf_buf from the freelist. Will block if none are available. 390 */ 391struct sf_buf * 392sf_buf_alloc(struct vm_page *m) 393{ 394 struct sf_buf *sf; 395 int error; 396 397 mtx_lock(&sf_freelist.sf_lock); 398 while ((sf = SLIST_FIRST(&sf_freelist.sf_head)) == NULL) { 399 sf_buf_alloc_want++; 400 error = msleep(&sf_freelist, &sf_freelist.sf_lock, PVM|PCATCH, 401 "sfbufa", 0); 402 sf_buf_alloc_want--; 403 404 /* 405 * If we got a signal, don't risk going back to sleep. 406 */ 407 if (error) 408 break; 409 } 410 if (sf != NULL) { 411 SLIST_REMOVE_HEAD(&sf_freelist.sf_head, free_list); 412 sf->m = m; 413 sf->kva = PHYS_TO_DMAP(m->phys_addr); 414 } 415 mtx_unlock(&sf_freelist.sf_lock); 416 return (sf); 417} 418 419/* 420 * Detatch mapped page and release resources back to the system. 421 */ 422void 423sf_buf_free(void *addr, void *args) 424{ 425 struct sf_buf *sf; 426 struct vm_page *m; 427 428 sf = args; 429 m = sf->m; 430 vm_page_lock_queues(); 431 vm_page_unwire(m, 0); 432 /* 433 * Check for the object going away on us. This can 434 * happen since we don't hold a reference to it. 435 * If so, we're responsible for freeing the page. 436 */ 437 if (m->wire_count == 0 && m->object == NULL) 438 vm_page_free(m); 439 vm_page_unlock_queues(); 440 sf->m = NULL; 441 mtx_lock(&sf_freelist.sf_lock); 442 SLIST_INSERT_HEAD(&sf_freelist.sf_head, sf, free_list); 443 if (sf_buf_alloc_want > 0) 444 wakeup_one(&sf_freelist); 445 mtx_unlock(&sf_freelist.sf_lock); 446} 447 448/* 449 * Software interrupt handler for queued VM system processing. 450 */ 451void 452swi_vm(void *dummy) 453{ 454 if (busdma_swi_pending != 0) 455 busdma_swi(); 456} 457 458/* 459 * Tell whether this address is in some physical memory region. 460 * Currently used by the kernel coredump code in order to avoid 461 * dumping the ``ISA memory hole'' which could cause indefinite hangs, 462 * or other unpredictable behaviour. 463 */ 464 465int 466is_physical_memory(addr) 467 vm_offset_t addr; 468{ 469 470#ifdef DEV_ISA 471 /* The ISA ``memory hole''. */ 472 if (addr >= 0xa0000 && addr < 0x100000) 473 return 0; 474#endif 475 476 /* 477 * stuff other tests for known memory-mapped devices (PCI?) 478 * here 479 */ 480 481 return 1; 482} 483