vm_machdep.c revision 156199
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 * from: @(#)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/arm/arm/vm_machdep.c 156199 2006-03-02 00:37:40Z cognet $"); 45 46#include <sys/param.h> 47#include <sys/systm.h> 48#include <sys/kernel.h> 49#include <sys/malloc.h> 50#include <sys/mbuf.h> 51#include <sys/proc.h> 52#include <sys/socketvar.h> 53#include <sys/sf_buf.h> 54#include <sys/unistd.h> 55#include <machine/cpu.h> 56#include <machine/pcb.h> 57#include <machine/sysarch.h> 58#include <vm/vm.h> 59#include <vm/pmap.h> 60#include <sys/lock.h> 61#include <sys/mutex.h> 62 63#include <vm/vm.h> 64#include <vm/vm_extern.h> 65#include <vm/vm_kern.h> 66#include <vm/vm_page.h> 67#include <vm/vm_map.h> 68#include <vm/vm_param.h> 69#include <vm/uma.h> 70#include <vm/uma_int.h> 71 72#ifndef NSFBUFS 73#define NSFBUFS (512 + maxusers * 16) 74#endif 75 76static void sf_buf_init(void *arg); 77SYSINIT(sock_sf, SI_SUB_MBUF, SI_ORDER_ANY, sf_buf_init, NULL) 78 79LIST_HEAD(sf_head, sf_buf); 80 81 82/* 83 * A hash table of active sendfile(2) buffers 84 */ 85static struct sf_head *sf_buf_active; 86static u_long sf_buf_hashmask; 87 88#define SF_BUF_HASH(m) (((m) - vm_page_array) & sf_buf_hashmask) 89 90static TAILQ_HEAD(, sf_buf) sf_buf_freelist; 91static u_int sf_buf_alloc_want; 92 93/* 94 * A lock used to synchronize access to the hash table and free list 95 */ 96static struct mtx sf_buf_lock; 97 98/* 99 * Finish a fork operation, with process p2 nearly set up. 100 * Copy and update the pcb, set up the stack so that the child 101 * ready to run and return to user mode. 102 */ 103void 104cpu_fork(register struct thread *td1, register struct proc *p2, 105 struct thread *td2, int flags) 106{ 107 struct pcb *pcb1, *pcb2; 108 struct trapframe *tf; 109 struct switchframe *sf; 110 struct mdproc *mdp2; 111 112 if ((flags & RFPROC) == 0) 113 return; 114 pcb1 = td1->td_pcb; 115 pcb2 = (struct pcb *)(td2->td_kstack + td2->td_kstack_pages * PAGE_SIZE) - 1; 116#ifdef __XSCALE__ 117 pmap_use_minicache(td2->td_kstack, td2->td_kstack_pages * PAGE_SIZE); 118 if (td2->td_altkstack) 119 pmap_use_minicache(td2->td_altkstack, td2->td_altkstack_pages * 120 PAGE_SIZE); 121#endif 122 td2->td_pcb = pcb2; 123 bcopy(td1->td_pcb, pcb2, sizeof(*pcb2)); 124 mdp2 = &p2->p_md; 125 bcopy(&td1->td_proc->p_md, mdp2, sizeof(*mdp2)); 126 pcb2->un_32.pcb32_und_sp = td2->td_kstack + USPACE_UNDEF_STACK_TOP; 127 pcb2->un_32.pcb32_sp = td2->td_kstack + 128 USPACE_SVC_STACK_TOP - sizeof(*pcb2); 129 pmap_activate(td2); 130 td2->td_frame = tf = 131 (struct trapframe *)pcb2->un_32.pcb32_sp - 1; 132 *tf = *td1->td_frame; 133 sf = (struct switchframe *)tf - 1; 134 sf->sf_r4 = (u_int)fork_return; 135 sf->sf_r5 = (u_int)td2; 136 sf->sf_pc = (u_int)fork_trampoline; 137 tf->tf_spsr &= ~PSR_C_bit; 138 tf->tf_r0 = 0; 139 tf->tf_r1 = 0; 140 pcb2->un_32.pcb32_sp = (u_int)sf; 141 142 /* Setup to release sched_lock in fork_exit(). */ 143 td2->td_md.md_spinlock_count = 1; 144 td2->td_md.md_saved_cspr = 0; 145 td2->td_md.md_tp = *(uint32_t **)ARM_TP_ADDRESS; 146} 147 148void 149cpu_thread_swapin(struct thread *td) 150{ 151} 152 153void 154cpu_thread_swapout(struct thread *td) 155{ 156} 157 158/* 159 * Detatch mapped page and release resources back to the system. 160 */ 161void 162sf_buf_free(struct sf_buf *sf) 163{ 164 mtx_lock(&sf_buf_lock); 165 sf->ref_count--; 166 if (sf->ref_count == 0) { 167 TAILQ_INSERT_TAIL(&sf_buf_freelist, sf, free_entry); 168 nsfbufsused--; 169 if (sf_buf_alloc_want > 0) 170 wakeup_one(&sf_buf_freelist); 171 } 172 mtx_unlock(&sf_buf_lock); 173} 174 175/* 176 * * Allocate a pool of sf_bufs (sendfile(2) or "super-fast" if you prefer. :-)) 177 * */ 178static void 179sf_buf_init(void *arg) 180{ 181 struct sf_buf *sf_bufs; 182 vm_offset_t sf_base; 183 int i; 184 185 nsfbufs = NSFBUFS; 186 TUNABLE_INT_FETCH("kern.ipc.nsfbufs", &nsfbufs); 187 188 sf_buf_active = hashinit(nsfbufs, M_TEMP, &sf_buf_hashmask); 189 TAILQ_INIT(&sf_buf_freelist); 190 sf_base = kmem_alloc_nofault(kernel_map, nsfbufs * PAGE_SIZE); 191 sf_bufs = malloc(nsfbufs * sizeof(struct sf_buf), M_TEMP, 192 M_NOWAIT | M_ZERO); 193 for (i = 0; i < nsfbufs; i++) { 194 sf_bufs[i].kva = sf_base + i * PAGE_SIZE; 195 TAILQ_INSERT_TAIL(&sf_buf_freelist, &sf_bufs[i], free_entry); 196 } 197 sf_buf_alloc_want = 0; 198 mtx_init(&sf_buf_lock, "sf_buf", NULL, MTX_DEF); 199} 200 201/* 202 * Get an sf_buf from the freelist. Will block if none are available. 203 */ 204struct sf_buf * 205sf_buf_alloc(struct vm_page *m, int flags) 206{ 207 struct sf_head *hash_list; 208 struct sf_buf *sf; 209 int error; 210 211 hash_list = &sf_buf_active[SF_BUF_HASH(m)]; 212 mtx_lock(&sf_buf_lock); 213 LIST_FOREACH(sf, hash_list, list_entry) { 214 if (sf->m == m) { 215 sf->ref_count++; 216 if (sf->ref_count == 1) { 217 TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry); 218 nsfbufsused++; 219 nsfbufspeak = imax(nsfbufspeak, nsfbufsused); 220 } 221 goto done; 222 } 223 } 224 while ((sf = TAILQ_FIRST(&sf_buf_freelist)) == NULL) { 225 if (flags & SFB_NOWAIT) 226 goto done; 227 sf_buf_alloc_want++; 228 mbstat.sf_allocwait++; 229 error = msleep(&sf_buf_freelist, &sf_buf_lock, 230 (flags & SFB_CATCH) ? PCATCH | PVM : PVM, "sfbufa", 0); 231 sf_buf_alloc_want--; 232 233 234 /* 235 * If we got a signal, don't risk going back to sleep. 236 */ 237 if (error) 238 goto done; 239 } 240 TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry); 241 if (sf->m != NULL) 242 LIST_REMOVE(sf, list_entry); 243 LIST_INSERT_HEAD(hash_list, sf, list_entry); 244 sf->ref_count = 1; 245 sf->m = m; 246 nsfbufsused++; 247 nsfbufspeak = imax(nsfbufspeak, nsfbufsused); 248 pmap_kenter(sf->kva, VM_PAGE_TO_PHYS(sf->m)); 249done: 250 mtx_unlock(&sf_buf_lock); 251 return (sf); 252 253} 254 255/* 256 * Initialize machine state (pcb and trap frame) for a new thread about to 257 * upcall. Put enough state in the new thread's PCB to get it to go back 258 * userret(), where we can intercept it again to set the return (upcall) 259 * Address and stack, along with those from upcals that are from other sources 260 * such as those generated in thread_userret() itself. 261 */ 262void 263cpu_set_upcall(struct thread *td, struct thread *td0) 264{ 265 struct trapframe *tf; 266 struct switchframe *sf; 267 268 bcopy(td0->td_frame, td->td_frame, sizeof(struct trapframe)); 269 bcopy(td0->td_pcb, td->td_pcb, sizeof(struct pcb)); 270 tf = td->td_frame; 271 sf = (struct switchframe *)tf - 1; 272 sf->sf_r4 = (u_int)fork_return; 273 sf->sf_r5 = (u_int)td; 274 sf->sf_pc = (u_int)fork_trampoline; 275 tf->tf_spsr &= ~PSR_C_bit; 276 tf->tf_r0 = 0; 277 td->td_pcb->un_32.pcb32_sp = (u_int)sf; 278 td->td_pcb->un_32.pcb32_und_sp = td->td_kstack + USPACE_UNDEF_STACK_TOP; 279 280 /* Setup to release sched_lock in fork_exit(). */ 281 td->td_md.md_spinlock_count = 1; 282 td->td_md.md_saved_cspr = 0; 283} 284 285/* 286 * Set that machine state for performing an upcall that has to 287 * be done in thread_userret() so that those upcalls generated 288 * in thread_userret() itself can be done as well. 289 */ 290void 291cpu_set_upcall_kse(struct thread *td, void (*entry)(void *), void *arg, 292 stack_t *stack) 293{ 294 struct trapframe *tf = td->td_frame; 295 296 tf->tf_usr_sp = ((int)stack->ss_sp + stack->ss_size 297 - sizeof(struct trapframe)) & ~7; 298 tf->tf_pc = (int)entry; 299 tf->tf_r0 = (int)arg; 300 tf->tf_spsr = PSR_USR32_MODE; 301} 302 303int 304cpu_set_user_tls(struct thread *td, void *tls_base) 305{ 306 307 if (td != curthread) 308 td->td_md.md_tp = tls_base; 309 else { 310 critical_enter(); 311 *(void **)ARM_TP_ADDRESS = tls_base; 312 critical_exit(); 313 } 314 return (0); 315} 316 317void 318cpu_thread_exit(struct thread *td) 319{ 320} 321 322void 323cpu_thread_setup(struct thread *td) 324{ 325 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_pages * 326 PAGE_SIZE) - 1; 327 td->td_frame = (struct trapframe *) 328 ((u_int)td->td_kstack + USPACE_SVC_STACK_TOP - sizeof(struct pcb)) - 1; 329#ifdef __XSCALE__ 330 pmap_use_minicache(td->td_kstack, td->td_kstack_pages * PAGE_SIZE); 331#endif 332 333} 334void 335cpu_thread_clean(struct thread *td) 336{ 337} 338 339/* 340 * Intercept the return address from a freshly forked process that has NOT 341 * been scheduled yet. 342 * 343 * This is needed to make kernel threads stay in kernel mode. 344 */ 345void 346cpu_set_fork_handler(struct thread *td, void (*func)(void *), void *arg) 347{ 348 struct switchframe *sf; 349 struct trapframe *tf; 350 351 tf = td->td_frame; 352 sf = (struct switchframe *)tf - 1; 353 sf->sf_r4 = (u_int)func; 354 sf->sf_r5 = (u_int)arg; 355 td->td_pcb->un_32.pcb32_sp = (u_int)sf; 356} 357 358/* 359 * Software interrupt handler for queued VM system processing. 360 */ 361void 362swi_vm(void *dummy) 363{ 364} 365 366void 367cpu_exit(struct thread *td) 368{ 369} 370 371#define BITS_PER_INT (8 * sizeof(int)) 372vm_offset_t arm_nocache_startaddr; 373static int arm_nocache_allocated[ARM_NOCACHE_KVA_SIZE / (PAGE_SIZE * 374 BITS_PER_INT)]; 375 376/* 377 * Functions to map and unmap memory non-cached into KVA the kernel won't try 378 * to allocate. The goal is to provide uncached memory to busdma, to honor 379 * BUS_DMA_COHERENT. 380 * We can allocate at most ARM_NOCACHE_KVA_SIZE bytes. 381 * The allocator is rather dummy, each page is represented by a bit in 382 * a bitfield, 0 meaning the page is not allocated, 1 meaning it is. 383 * As soon as it finds enough contiguous pages to satisfy the request, 384 * it returns the address. 385 */ 386void * 387arm_remap_nocache(void *addr, vm_size_t size) 388{ 389 int i, j; 390 391 size = round_page(size); 392 for (i = 0; i < MIN(ARM_NOCACHE_KVA_SIZE / (PAGE_SIZE * BITS_PER_INT), 393 ARM_TP_ADDRESS); i++) { 394 if (!(arm_nocache_allocated[i / BITS_PER_INT] & (1 << (i % 395 BITS_PER_INT)))) { 396 for (j = i; j < i + (size / (PAGE_SIZE)); j++) 397 if (arm_nocache_allocated[j / BITS_PER_INT] & 398 (1 << (j % BITS_PER_INT))) 399 break; 400 if (j == i + (size / (PAGE_SIZE))) 401 break; 402 } 403 } 404 if (i < MIN(ARM_NOCACHE_KVA_SIZE / (PAGE_SIZE * BITS_PER_INT), 405 ARM_TP_ADDRESS)) { 406 vm_offset_t tomap = arm_nocache_startaddr + i * PAGE_SIZE; 407 void *ret = (void *)tomap; 408 vm_paddr_t physaddr = vtophys((vm_offset_t)addr); 409 410 for (; tomap < (vm_offset_t)ret + size; tomap += PAGE_SIZE, 411 physaddr += PAGE_SIZE, i++) { 412 pmap_kenter_nocache(tomap, physaddr); 413 arm_nocache_allocated[i / BITS_PER_INT] |= 1 << (i % 414 BITS_PER_INT); 415 } 416 return (ret); 417 } 418 return (NULL); 419} 420 421void 422arm_unmap_nocache(void *addr, vm_size_t size) 423{ 424 vm_offset_t raddr = (vm_offset_t)addr; 425 int i; 426 427 size = round_page(size); 428 i = (raddr - arm_nocache_startaddr) / (PAGE_SIZE); 429 for (; size > 0; size -= PAGE_SIZE, i++) 430 arm_nocache_allocated[i / BITS_PER_INT] &= ~(1 << (i % 431 BITS_PER_INT)); 432} 433 434#ifdef ARM_USE_SMALL_ALLOC 435 436static TAILQ_HEAD(,arm_small_page) pages_normal = 437 TAILQ_HEAD_INITIALIZER(pages_normal); 438static TAILQ_HEAD(,arm_small_page) pages_wt = 439 TAILQ_HEAD_INITIALIZER(pages_wt); 440static TAILQ_HEAD(,arm_small_page) free_pgdesc = 441 TAILQ_HEAD_INITIALIZER(free_pgdesc); 442 443extern uma_zone_t l2zone; 444 445struct mtx smallalloc_mtx; 446 447MALLOC_DEFINE(M_VMSMALLALLOC, "vm_small_alloc", "VM Small alloc data"); 448 449vm_offset_t alloc_curaddr; 450vm_offset_t alloc_firstaddr; 451 452extern int doverbose; 453 454void 455arm_add_smallalloc_pages(void *list, void *mem, int bytes, int pagetable) 456{ 457 struct arm_small_page *pg; 458 459 bytes &= ~PAGE_MASK; 460 while (bytes > 0) { 461 pg = (struct arm_small_page *)list; 462 pg->addr = mem; 463 if (pagetable) 464 TAILQ_INSERT_HEAD(&pages_wt, pg, pg_list); 465 else 466 TAILQ_INSERT_HEAD(&pages_normal, pg, pg_list); 467 list = (char *)list + sizeof(*pg); 468 mem = (char *)mem + PAGE_SIZE; 469 bytes -= PAGE_SIZE; 470 } 471} 472 473static void * 474arm_uma_do_alloc(struct arm_small_page **pglist, int bytes, int pagetable, 475 int flags) 476{ 477 void *ret; 478 vm_page_t page_array = NULL; 479 480 *pglist = (void *)kmem_malloc(kmem_map, (0x100000 / PAGE_SIZE) * 481 sizeof(struct arm_small_page), flags); 482 if (*pglist && alloc_curaddr < 0xf0000000) {/* XXX */ 483 mtx_lock(&Giant); 484 page_array = vm_page_alloc_contig(0x100000 / PAGE_SIZE, 485 0, 0xffffffff, 0x100000, 0); 486 mtx_unlock(&Giant); 487 } 488 if (page_array) { 489 vm_paddr_t pa = VM_PAGE_TO_PHYS(page_array); 490 mtx_lock(&smallalloc_mtx); 491 ret = (void*)alloc_curaddr; 492 alloc_curaddr += 0x100000; 493 /* XXX: ARM_TP_ADDRESS should probably be move elsewhere. */ 494 if (alloc_curaddr == ARM_TP_ADDRESS) 495 alloc_curaddr += 0x100000; 496 mtx_unlock(&smallalloc_mtx); 497 pmap_kenter_section((vm_offset_t)ret, pa 498 , pagetable); 499 } else { 500 if (*pglist) 501 kmem_free(kmem_map, (vm_offset_t)*pglist, 502 (0x100000 / PAGE_SIZE) * 503 sizeof(struct arm_small_page)); 504 *pglist = NULL; 505 ret = (void *)kmem_malloc(kmem_map, bytes, M_WAITOK); 506 } 507 return (ret); 508} 509 510void * 511uma_small_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 512{ 513 void *ret; 514 struct arm_small_page *sp, *tmp; 515 TAILQ_HEAD(,arm_small_page) *head; 516 static int in_alloc; 517 static int in_sleep; 518 int should_wakeup = 0; 519 520 *flags = UMA_SLAB_PRIV; 521 /* 522 * For CPUs where we setup page tables as write back, there's no 523 * need to maintain two separate pools. 524 */ 525 if (zone == l2zone && pte_l1_s_cache_mode != pte_l1_s_cache_mode_pt) 526 head = (void *)&pages_wt; 527 else 528 head = (void *)&pages_normal; 529 530 mtx_lock(&smallalloc_mtx); 531retry: 532 sp = TAILQ_FIRST(head); 533 534 if (!sp) { 535 /* No more free pages, need to alloc more. */ 536 if (in_alloc && (wait & M_WAITOK)) { 537 /* Somebody else is already doing the allocation. */ 538 in_sleep++; 539 msleep(&in_alloc, &smallalloc_mtx, PWAIT, 540 "smallalloc", 0); 541 in_sleep--; 542 goto retry; 543 } else if (in_alloc) { 544 mtx_unlock(&smallalloc_mtx); 545 return (NULL); 546 } 547 in_alloc = 1; 548 mtx_unlock(&smallalloc_mtx); 549 /* Try to alloc 1MB of contiguous memory. */ 550 ret = arm_uma_do_alloc(&sp, bytes, zone == l2zone ? 551 SECTION_PT : SECTION_CACHE, wait); 552 mtx_lock(&smallalloc_mtx); 553 in_alloc = 0; 554 if (in_sleep) 555 should_wakeup = 1; 556 if (sp) { 557 for (int i = 0; i < (0x100000 / PAGE_SIZE) - 1; 558 i++) { 559 tmp = &sp[i]; 560 tmp->addr = (char *)ret + i * PAGE_SIZE; 561 TAILQ_INSERT_HEAD(head, tmp, pg_list); 562 } 563 ret = (char *)ret + 0x100000 - PAGE_SIZE; 564 TAILQ_INSERT_HEAD(&free_pgdesc, &sp[(0x100000 / ( 565 PAGE_SIZE)) - 1], pg_list); 566 } else 567 *flags = UMA_SLAB_KMEM; 568 569 } else { 570 sp = TAILQ_FIRST(head); 571 TAILQ_REMOVE(head, sp, pg_list); 572 TAILQ_INSERT_HEAD(&free_pgdesc, sp, pg_list); 573 ret = sp->addr; 574 } 575 if (should_wakeup) 576 wakeup(&in_alloc); 577 mtx_unlock(&smallalloc_mtx); 578 if ((wait & M_ZERO)) 579 bzero(ret, bytes); 580 return (ret); 581} 582 583void 584uma_small_free(void *mem, int size, u_int8_t flags) 585{ 586 pd_entry_t *pd; 587 pt_entry_t *pt; 588 589 if (flags & UMA_SLAB_KMEM) 590 kmem_free(kmem_map, (vm_offset_t)mem, size); 591 else { 592 struct arm_small_page *sp; 593 594 mtx_lock(&smallalloc_mtx); 595 sp = TAILQ_FIRST(&free_pgdesc); 596 KASSERT(sp != NULL, ("No more free page descriptor ?")); 597 TAILQ_REMOVE(&free_pgdesc, sp, pg_list); 598 sp->addr = mem; 599 pmap_get_pde_pte(kernel_pmap, (vm_offset_t)mem, &pd, &pt); 600 if ((*pd & pte_l1_s_cache_mask) == pte_l1_s_cache_mode_pt && 601 pte_l1_s_cache_mode_pt != pte_l1_s_cache_mode) 602 TAILQ_INSERT_HEAD(&pages_wt, sp, pg_list); 603 else 604 TAILQ_INSERT_HEAD(&pages_normal, sp, pg_list); 605 mtx_unlock(&smallalloc_mtx); 606 } 607} 608 609#endif 610