| 1 | /* |
| 2 | * Copyright (c) 1991 Regents of the University of California. |
| 3 | * Copyright (c) 1994 John S. Dyson |
| 4 | * Copyright (c) 1994 David Greenman |
| 5 | * Copyright (c) 2003 Peter Wemm |
| 6 | * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu> |
| 7 | * Copyright (c) 2008, 2009 The DragonFly Project. |
| 8 | * Copyright (c) 2008, 2009 Jordan Gordeev. |
| 9 | * Copyright (c) 2011 Matthew Dillon |
| 10 | * All rights reserved. |
| 11 | * |
| 12 | * This code is derived from software contributed to Berkeley by |
| 13 | * the Systems Programming Group of the University of Utah Computer |
| 14 | * Science Department and William Jolitz of UUNET Technologies Inc. |
| 15 | * |
| 16 | * Redistribution and use in source and binary forms, with or without |
| 17 | * modification, are permitted provided that the following conditions |
| 18 | * are met: |
| 19 | * 1. Redistributions of source code must retain the above copyright |
| 20 | * notice, this list of conditions and the following disclaimer. |
| 21 | * 2. Redistributions in binary form must reproduce the above copyright |
| 22 | * notice, this list of conditions and the following disclaimer in the |
| 23 | * documentation and/or other materials provided with the distribution. |
| 24 | * 3. All advertising materials mentioning features or use of this software |
| 25 | * must display the following acknowledgement: |
| 26 | * This product includes software developed by the University of |
| 27 | * California, Berkeley and its contributors. |
| 28 | * 4. Neither the name of the University nor the names of its contributors |
| 29 | * may be used to endorse or promote products derived from this software |
| 30 | * without specific prior written permission. |
| 31 | * |
| 32 | * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND |
| 33 | * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 34 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE |
| 35 | * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE |
| 36 | * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL |
| 37 | * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS |
| 38 | * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) |
| 39 | * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT |
| 40 | * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY |
| 41 | * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF |
| 42 | * SUCH DAMAGE. |
| 43 | */ |
| 44 | /* |
| 45 | * Manage physical address maps for x86-64 systems. |
| 46 | */ |
| 47 | |
| 48 | #if JG |
| 49 | #include "opt_disable_pse.h" |
| 50 | #include "opt_pmap.h" |
| 51 | #endif |
| 52 | #include "opt_msgbuf.h" |
| 53 | |
| 54 | #include <sys/param.h> |
| 55 | #include <sys/systm.h> |
| 56 | #include <sys/kernel.h> |
| 57 | #include <sys/proc.h> |
| 58 | #include <sys/msgbuf.h> |
| 59 | #include <sys/vmmeter.h> |
| 60 | #include <sys/mman.h> |
| 61 | |
| 62 | #include <vm/vm.h> |
| 63 | #include <vm/vm_param.h> |
| 64 | #include <sys/sysctl.h> |
| 65 | #include <sys/lock.h> |
| 66 | #include <vm/vm_kern.h> |
| 67 | #include <vm/vm_page.h> |
| 68 | #include <vm/vm_map.h> |
| 69 | #include <vm/vm_object.h> |
| 70 | #include <vm/vm_extern.h> |
| 71 | #include <vm/vm_pageout.h> |
| 72 | #include <vm/vm_pager.h> |
| 73 | #include <vm/vm_zone.h> |
| 74 | |
| 75 | #include <sys/user.h> |
| 76 | #include <sys/thread2.h> |
| 77 | #include <sys/sysref2.h> |
| 78 | #include <sys/spinlock2.h> |
| 79 | #include <vm/vm_page2.h> |
| 80 | |
| 81 | #include <machine/cputypes.h> |
| 82 | #include <machine/md_var.h> |
| 83 | #include <machine/specialreg.h> |
| 84 | #include <machine/smp.h> |
| 85 | #include <machine_base/apic/apicreg.h> |
| 86 | #include <machine/globaldata.h> |
| 87 | #include <machine/pmap.h> |
| 88 | #include <machine/pmap_inval.h> |
| 89 | #include <machine/inttypes.h> |
| 90 | |
| 91 | #include <ddb/ddb.h> |
| 92 | |
| 93 | #define PMAP_KEEP_PDIRS |
| 94 | #ifndef PMAP_SHPGPERPROC |
| 95 | #define PMAP_SHPGPERPROC 2000 |
| 96 | #endif |
| 97 | |
| 98 | #if defined(DIAGNOSTIC) |
| 99 | #define PMAP_DIAGNOSTIC |
| 100 | #endif |
| 101 | |
| 102 | #define MINPV 2048 |
| 103 | |
| 104 | /* |
| 105 | * pmap debugging will report who owns a pv lock when blocking. |
| 106 | */ |
| 107 | #ifdef PMAP_DEBUG |
| 108 | |
| 109 | #define PMAP_DEBUG_DECL ,const char *func, int lineno |
| 110 | #define PMAP_DEBUG_ARGS , __func__, __LINE__ |
| 111 | #define PMAP_DEBUG_COPY , func, lineno |
| 112 | |
| 113 | #define pv_get(pmap, pindex) _pv_get(pmap, pindex \ |
| 114 | PMAP_DEBUG_ARGS) |
| 115 | #define pv_lock(pv) _pv_lock(pv \ |
| 116 | PMAP_DEBUG_ARGS) |
| 117 | #define pv_hold_try(pv) _pv_hold_try(pv \ |
| 118 | PMAP_DEBUG_ARGS) |
| 119 | #define pv_alloc(pmap, pindex, isnewp) _pv_alloc(pmap, pindex, isnewp \ |
| 120 | PMAP_DEBUG_ARGS) |
| 121 | |
| 122 | #else |
| 123 | |
| 124 | #define PMAP_DEBUG_DECL |
| 125 | #define PMAP_DEBUG_ARGS |
| 126 | #define PMAP_DEBUG_COPY |
| 127 | |
| 128 | #define pv_get(pmap, pindex) _pv_get(pmap, pindex) |
| 129 | #define pv_lock(pv) _pv_lock(pv) |
| 130 | #define pv_hold_try(pv) _pv_hold_try(pv) |
| 131 | #define pv_alloc(pmap, pindex, isnewp) _pv_alloc(pmap, pindex, isnewp) |
| 132 | |
| 133 | #endif |
| 134 | |
| 135 | /* |
| 136 | * Get PDEs and PTEs for user/kernel address space |
| 137 | */ |
| 138 | #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) |
| 139 | |
| 140 | #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & PG_V) != 0) |
| 141 | #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & PG_W) != 0) |
| 142 | #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & PG_M) != 0) |
| 143 | #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & PG_A) != 0) |
| 144 | #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & PG_V) != 0) |
| 145 | |
| 146 | /* |
| 147 | * Given a map and a machine independent protection code, |
| 148 | * convert to a vax protection code. |
| 149 | */ |
| 150 | #define pte_prot(m, p) \ |
| 151 | (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)]) |
| 152 | static int protection_codes[8]; |
| 153 | |
| 154 | struct pmap kernel_pmap; |
| 155 | static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list); |
| 156 | |
| 157 | vm_paddr_t avail_start; /* PA of first available physical page */ |
| 158 | vm_paddr_t avail_end; /* PA of last available physical page */ |
| 159 | vm_offset_t virtual2_start; /* cutout free area prior to kernel start */ |
| 160 | vm_offset_t virtual2_end; |
| 161 | vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */ |
| 162 | vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ |
| 163 | vm_offset_t KvaStart; /* VA start of KVA space */ |
| 164 | vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */ |
| 165 | vm_offset_t KvaSize; /* max size of kernel virtual address space */ |
| 166 | static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */ |
| 167 | static int pgeflag; /* PG_G or-in */ |
| 168 | static int pseflag; /* PG_PS or-in */ |
| 169 | |
| 170 | static int ndmpdp; |
| 171 | static vm_paddr_t dmaplimit; |
| 172 | static int nkpt; |
| 173 | vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS; |
| 174 | |
| 175 | static uint64_t KPTbase; |
| 176 | static uint64_t KPTphys; |
| 177 | static uint64_t KPDphys; /* phys addr of kernel level 2 */ |
| 178 | static uint64_t KPDbase; /* phys addr of kernel level 2 @ KERNBASE */ |
| 179 | uint64_t KPDPphys; /* phys addr of kernel level 3 */ |
| 180 | uint64_t KPML4phys; /* phys addr of kernel level 4 */ |
| 181 | |
| 182 | static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */ |
| 183 | static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */ |
| 184 | |
| 185 | /* |
| 186 | * Data for the pv entry allocation mechanism |
| 187 | */ |
| 188 | static vm_zone_t pvzone; |
| 189 | static struct vm_zone pvzone_store; |
| 190 | static struct vm_object pvzone_obj; |
| 191 | static int pv_entry_max=0, pv_entry_high_water=0; |
| 192 | static int pmap_pagedaemon_waken = 0; |
| 193 | static struct pv_entry *pvinit; |
| 194 | |
| 195 | /* |
| 196 | * All those kernel PT submaps that BSD is so fond of |
| 197 | */ |
| 198 | pt_entry_t *CMAP1 = 0, *ptmmap; |
| 199 | caddr_t CADDR1 = 0, ptvmmap = 0; |
| 200 | static pt_entry_t *msgbufmap; |
| 201 | struct msgbuf *msgbufp=0; |
| 202 | |
| 203 | /* |
| 204 | * Crashdump maps. |
| 205 | */ |
| 206 | static pt_entry_t *pt_crashdumpmap; |
| 207 | static caddr_t crashdumpmap; |
| 208 | |
| 209 | static int pmap_yield_count = 64; |
| 210 | SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW, |
| 211 | &pmap_yield_count, 0, "Yield during init_pt/release"); |
| 212 | |
| 213 | #define DISABLE_PSE |
| 214 | |
| 215 | static void pv_hold(pv_entry_t pv); |
| 216 | static int _pv_hold_try(pv_entry_t pv |
| 217 | PMAP_DEBUG_DECL); |
| 218 | static void pv_drop(pv_entry_t pv); |
| 219 | static void _pv_lock(pv_entry_t pv |
| 220 | PMAP_DEBUG_DECL); |
| 221 | static void pv_unlock(pv_entry_t pv); |
| 222 | static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew |
| 223 | PMAP_DEBUG_DECL); |
| 224 | static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex |
| 225 | PMAP_DEBUG_DECL); |
| 226 | static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp); |
| 227 | static pv_entry_t pv_find(pmap_t pmap, vm_pindex_t pindex); |
| 228 | static void pv_put(pv_entry_t pv); |
| 229 | static void pv_free(pv_entry_t pv); |
| 230 | static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex); |
| 231 | static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, |
| 232 | pv_entry_t *pvpp); |
| 233 | static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, |
| 234 | struct pmap_inval_info *info); |
| 235 | static vm_page_t pmap_remove_pv_page(pv_entry_t pv); |
| 236 | |
| 237 | static void pmap_remove_callback(pmap_t pmap, struct pmap_inval_info *info, |
| 238 | pv_entry_t pte_pv, pv_entry_t pt_pv, vm_offset_t va, |
| 239 | pt_entry_t *ptep, void *arg __unused); |
| 240 | static void pmap_protect_callback(pmap_t pmap, struct pmap_inval_info *info, |
| 241 | pv_entry_t pte_pv, pv_entry_t pt_pv, vm_offset_t va, |
| 242 | pt_entry_t *ptep, void *arg __unused); |
| 243 | |
| 244 | static void i386_protection_init (void); |
| 245 | static void create_pagetables(vm_paddr_t *firstaddr); |
| 246 | static void pmap_remove_all (vm_page_t m); |
| 247 | static boolean_t pmap_testbit (vm_page_t m, int bit); |
| 248 | |
| 249 | static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va); |
| 250 | static vm_offset_t pmap_kmem_choose(vm_offset_t addr); |
| 251 | |
| 252 | static unsigned pdir4mb; |
| 253 | |
| 254 | static int |
| 255 | pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2) |
| 256 | { |
| 257 | if (pv1->pv_pindex < pv2->pv_pindex) |
| 258 | return(-1); |
| 259 | if (pv1->pv_pindex > pv2->pv_pindex) |
| 260 | return(1); |
| 261 | return(0); |
| 262 | } |
| 263 | |
| 264 | RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry, |
| 265 | pv_entry_compare, vm_pindex_t, pv_pindex); |
| 266 | |
| 267 | /* |
| 268 | * Move the kernel virtual free pointer to the next |
| 269 | * 2MB. This is used to help improve performance |
| 270 | * by using a large (2MB) page for much of the kernel |
| 271 | * (.text, .data, .bss) |
| 272 | */ |
| 273 | static |
| 274 | vm_offset_t |
| 275 | pmap_kmem_choose(vm_offset_t addr) |
| 276 | { |
| 277 | vm_offset_t newaddr = addr; |
| 278 | |
| 279 | newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1); |
| 280 | return newaddr; |
| 281 | } |
| 282 | |
| 283 | /* |
| 284 | * pmap_pte_quick: |
| 285 | * |
| 286 | * Super fast pmap_pte routine best used when scanning the pv lists. |
| 287 | * This eliminates many course-grained invltlb calls. Note that many of |
| 288 | * the pv list scans are across different pmaps and it is very wasteful |
| 289 | * to do an entire invltlb when checking a single mapping. |
| 290 | */ |
| 291 | static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va); |
| 292 | |
| 293 | static |
| 294 | pt_entry_t * |
| 295 | pmap_pte_quick(pmap_t pmap, vm_offset_t va) |
| 296 | { |
| 297 | return pmap_pte(pmap, va); |
| 298 | } |
| 299 | |
| 300 | /* |
| 301 | * Returns the pindex of a page table entry (representing a terminal page). |
| 302 | * There are NUPTE_TOTAL page table entries possible (a huge number) |
| 303 | * |
| 304 | * x86-64 has a 48-bit address space, where bit 47 is sign-extended out. |
| 305 | * We want to properly translate negative KVAs. |
| 306 | */ |
| 307 | static __inline |
| 308 | vm_pindex_t |
| 309 | pmap_pte_pindex(vm_offset_t va) |
| 310 | { |
| 311 | return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1)); |
| 312 | } |
| 313 | |
| 314 | /* |
| 315 | * Returns the pindex of a page table. |
| 316 | */ |
| 317 | static __inline |
| 318 | vm_pindex_t |
| 319 | pmap_pt_pindex(vm_offset_t va) |
| 320 | { |
| 321 | return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1))); |
| 322 | } |
| 323 | |
| 324 | /* |
| 325 | * Returns the pindex of a page directory. |
| 326 | */ |
| 327 | static __inline |
| 328 | vm_pindex_t |
| 329 | pmap_pd_pindex(vm_offset_t va) |
| 330 | { |
| 331 | return (NUPTE_TOTAL + NUPT_TOTAL + |
| 332 | ((va >> PDPSHIFT) & (NUPD_TOTAL - 1))); |
| 333 | } |
| 334 | |
| 335 | static __inline |
| 336 | vm_pindex_t |
| 337 | pmap_pdp_pindex(vm_offset_t va) |
| 338 | { |
| 339 | return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + |
| 340 | ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1))); |
| 341 | } |
| 342 | |
| 343 | static __inline |
| 344 | vm_pindex_t |
| 345 | pmap_pml4_pindex(void) |
| 346 | { |
| 347 | return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL); |
| 348 | } |
| 349 | |
| 350 | /* |
| 351 | * Return various clipped indexes for a given VA |
| 352 | * |
| 353 | * Returns the index of a pte in a page table, representing a terminal |
| 354 | * page. |
| 355 | */ |
| 356 | static __inline |
| 357 | vm_pindex_t |
| 358 | pmap_pte_index(vm_offset_t va) |
| 359 | { |
| 360 | return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1)); |
| 361 | } |
| 362 | |
| 363 | /* |
| 364 | * Returns the index of a pt in a page directory, representing a page |
| 365 | * table. |
| 366 | */ |
| 367 | static __inline |
| 368 | vm_pindex_t |
| 369 | pmap_pt_index(vm_offset_t va) |
| 370 | { |
| 371 | return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1)); |
| 372 | } |
| 373 | |
| 374 | /* |
| 375 | * Returns the index of a pd in a page directory page, representing a page |
| 376 | * directory. |
| 377 | */ |
| 378 | static __inline |
| 379 | vm_pindex_t |
| 380 | pmap_pd_index(vm_offset_t va) |
| 381 | { |
| 382 | return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1)); |
| 383 | } |
| 384 | |
| 385 | /* |
| 386 | * Returns the index of a pdp in the pml4 table, representing a page |
| 387 | * directory page. |
| 388 | */ |
| 389 | static __inline |
| 390 | vm_pindex_t |
| 391 | pmap_pdp_index(vm_offset_t va) |
| 392 | { |
| 393 | return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1)); |
| 394 | } |
| 395 | |
| 396 | /* |
| 397 | * Generic procedure to index a pte from a pt, pd, or pdp. |
| 398 | */ |
| 399 | static |
| 400 | void * |
| 401 | pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex) |
| 402 | { |
| 403 | pt_entry_t *pte; |
| 404 | |
| 405 | pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m)); |
| 406 | return(&pte[pindex]); |
| 407 | } |
| 408 | |
| 409 | /* |
| 410 | * Return pointer to PDP slot in the PML4 |
| 411 | */ |
| 412 | static __inline |
| 413 | pml4_entry_t * |
| 414 | pmap_pdp(pmap_t pmap, vm_offset_t va) |
| 415 | { |
| 416 | return (&pmap->pm_pml4[pmap_pdp_index(va)]); |
| 417 | } |
| 418 | |
| 419 | /* |
| 420 | * Return pointer to PD slot in the PDP given a pointer to the PDP |
| 421 | */ |
| 422 | static __inline |
| 423 | pdp_entry_t * |
| 424 | pmap_pdp_to_pd(pml4_entry_t *pdp, vm_offset_t va) |
| 425 | { |
| 426 | pdp_entry_t *pd; |
| 427 | |
| 428 | pd = (pdp_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME); |
| 429 | return (&pd[pmap_pd_index(va)]); |
| 430 | } |
| 431 | |
| 432 | /* |
| 433 | * Return pointer to PD slot in the PDP |
| 434 | **/ |
| 435 | static __inline |
| 436 | pdp_entry_t * |
| 437 | pmap_pd(pmap_t pmap, vm_offset_t va) |
| 438 | { |
| 439 | pml4_entry_t *pdp; |
| 440 | |
| 441 | pdp = pmap_pdp(pmap, va); |
| 442 | if ((*pdp & PG_V) == 0) |
| 443 | return NULL; |
| 444 | return (pmap_pdp_to_pd(pdp, va)); |
| 445 | } |
| 446 | |
| 447 | /* |
| 448 | * Return pointer to PT slot in the PD given a pointer to the PD |
| 449 | */ |
| 450 | static __inline |
| 451 | pd_entry_t * |
| 452 | pmap_pd_to_pt(pdp_entry_t *pd, vm_offset_t va) |
| 453 | { |
| 454 | pd_entry_t *pt; |
| 455 | |
| 456 | pt = (pd_entry_t *)PHYS_TO_DMAP(*pd & PG_FRAME); |
| 457 | return (&pt[pmap_pt_index(va)]); |
| 458 | } |
| 459 | |
| 460 | /* |
| 461 | * Return pointer to PT slot in the PD |
| 462 | */ |
| 463 | static __inline |
| 464 | pd_entry_t * |
| 465 | pmap_pt(pmap_t pmap, vm_offset_t va) |
| 466 | { |
| 467 | pdp_entry_t *pd; |
| 468 | |
| 469 | pd = pmap_pd(pmap, va); |
| 470 | if (pd == NULL || (*pd & PG_V) == 0) |
| 471 | return NULL; |
| 472 | return (pmap_pd_to_pt(pd, va)); |
| 473 | } |
| 474 | |
| 475 | /* |
| 476 | * Return pointer to PTE slot in the PT given a pointer to the PT |
| 477 | */ |
| 478 | static __inline |
| 479 | pt_entry_t * |
| 480 | pmap_pt_to_pte(pd_entry_t *pt, vm_offset_t va) |
| 481 | { |
| 482 | pt_entry_t *pte; |
| 483 | |
| 484 | pte = (pt_entry_t *)PHYS_TO_DMAP(*pt & PG_FRAME); |
| 485 | return (&pte[pmap_pte_index(va)]); |
| 486 | } |
| 487 | |
| 488 | /* |
| 489 | * Return pointer to PTE slot in the PT |
| 490 | */ |
| 491 | static __inline |
| 492 | pt_entry_t * |
| 493 | pmap_pte(pmap_t pmap, vm_offset_t va) |
| 494 | { |
| 495 | pd_entry_t *pt; |
| 496 | |
| 497 | pt = pmap_pt(pmap, va); |
| 498 | if (pt == NULL || (*pt & PG_V) == 0) |
| 499 | return NULL; |
| 500 | if ((*pt & PG_PS) != 0) |
| 501 | return ((pt_entry_t *)pt); |
| 502 | return (pmap_pt_to_pte(pt, va)); |
| 503 | } |
| 504 | |
| 505 | /* |
| 506 | * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is |
| 507 | * the PT layer. This will speed up core pmap operations considerably. |
| 508 | */ |
| 509 | static __inline |
| 510 | void |
| 511 | pv_cache(pv_entry_t pv, vm_pindex_t pindex) |
| 512 | { |
| 513 | if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0)) |
| 514 | pv->pv_pmap->pm_pvhint = pv; |
| 515 | } |
| 516 | |
| 517 | |
| 518 | /* |
| 519 | * KVM - return address of PT slot in PD |
| 520 | */ |
| 521 | static __inline |
| 522 | pd_entry_t * |
| 523 | vtopt(vm_offset_t va) |
| 524 | { |
| 525 | uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + |
| 526 | NPML4EPGSHIFT)) - 1); |
| 527 | |
| 528 | return (PDmap + ((va >> PDRSHIFT) & mask)); |
| 529 | } |
| 530 | |
| 531 | /* |
| 532 | * KVM - return address of PTE slot in PT |
| 533 | */ |
| 534 | static __inline |
| 535 | pt_entry_t * |
| 536 | vtopte(vm_offset_t va) |
| 537 | { |
| 538 | uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + |
| 539 | NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1); |
| 540 | |
| 541 | return (PTmap + ((va >> PAGE_SHIFT) & mask)); |
| 542 | } |
| 543 | |
| 544 | static uint64_t |
| 545 | allocpages(vm_paddr_t *firstaddr, long n) |
| 546 | { |
| 547 | uint64_t ret; |
| 548 | |
| 549 | ret = *firstaddr; |
| 550 | bzero((void *)ret, n * PAGE_SIZE); |
| 551 | *firstaddr += n * PAGE_SIZE; |
| 552 | return (ret); |
| 553 | } |
| 554 | |
| 555 | static |
| 556 | void |
| 557 | create_pagetables(vm_paddr_t *firstaddr) |
| 558 | { |
| 559 | long i; /* must be 64 bits */ |
| 560 | long nkpt_base; |
| 561 | long nkpt_phys; |
| 562 | int j; |
| 563 | |
| 564 | /* |
| 565 | * We are running (mostly) V=P at this point |
| 566 | * |
| 567 | * Calculate NKPT - number of kernel page tables. We have to |
| 568 | * accomodoate prealloction of the vm_page_array, dump bitmap, |
| 569 | * MSGBUF_SIZE, and other stuff. Be generous. |
| 570 | * |
| 571 | * Maxmem is in pages. |
| 572 | * |
| 573 | * ndmpdp is the number of 1GB pages we wish to map. |
| 574 | */ |
| 575 | ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT; |
| 576 | if (ndmpdp < 4) /* Minimum 4GB of dirmap */ |
| 577 | ndmpdp = 4; |
| 578 | KKASSERT(ndmpdp <= NKPDPE * NPDEPG); |
| 579 | |
| 580 | /* |
| 581 | * Starting at the beginning of kvm (not KERNBASE). |
| 582 | */ |
| 583 | nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR; |
| 584 | nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR; |
| 585 | nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E + |
| 586 | ndmpdp) + 511) / 512; |
| 587 | nkpt_phys += 128; |
| 588 | |
| 589 | /* |
| 590 | * Starting at KERNBASE - map 2G worth of page table pages. |
| 591 | * KERNBASE is offset -2G from the end of kvm. |
| 592 | */ |
| 593 | nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */ |
| 594 | |
| 595 | /* |
| 596 | * Allocate pages |
| 597 | */ |
| 598 | KPTbase = allocpages(firstaddr, nkpt_base); |
| 599 | KPTphys = allocpages(firstaddr, nkpt_phys); |
| 600 | KPML4phys = allocpages(firstaddr, 1); |
| 601 | KPDPphys = allocpages(firstaddr, NKPML4E); |
| 602 | KPDphys = allocpages(firstaddr, NKPDPE); |
| 603 | |
| 604 | /* |
| 605 | * Calculate the page directory base for KERNBASE, |
| 606 | * that is where we start populating the page table pages. |
| 607 | * Basically this is the end - 2. |
| 608 | */ |
| 609 | KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT); |
| 610 | |
| 611 | DMPDPphys = allocpages(firstaddr, NDMPML4E); |
| 612 | if ((amd_feature & AMDID_PAGE1GB) == 0) |
| 613 | DMPDphys = allocpages(firstaddr, ndmpdp); |
| 614 | dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT; |
| 615 | |
| 616 | /* |
| 617 | * Fill in the underlying page table pages for the area around |
| 618 | * KERNBASE. This remaps low physical memory to KERNBASE. |
| 619 | * |
| 620 | * Read-only from zero to physfree |
| 621 | * XXX not fully used, underneath 2M pages |
| 622 | */ |
| 623 | for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) { |
| 624 | ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT; |
| 625 | ((pt_entry_t *)KPTbase)[i] |= PG_RW | PG_V | PG_G; |
| 626 | } |
| 627 | |
| 628 | /* |
| 629 | * Now map the initial kernel page tables. One block of page |
| 630 | * tables is placed at the beginning of kernel virtual memory, |
| 631 | * and another block is placed at KERNBASE to map the kernel binary, |
| 632 | * data, bss, and initial pre-allocations. |
| 633 | */ |
| 634 | for (i = 0; i < nkpt_base; i++) { |
| 635 | ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT); |
| 636 | ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V; |
| 637 | } |
| 638 | for (i = 0; i < nkpt_phys; i++) { |
| 639 | ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT); |
| 640 | ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V; |
| 641 | } |
| 642 | |
| 643 | /* |
| 644 | * Map from zero to end of allocations using 2M pages as an |
| 645 | * optimization. This will bypass some of the KPTBase pages |
| 646 | * above in the KERNBASE area. |
| 647 | */ |
| 648 | for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) { |
| 649 | ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT; |
| 650 | ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V | PG_PS | PG_G; |
| 651 | } |
| 652 | |
| 653 | /* |
| 654 | * And connect up the PD to the PDP. The kernel pmap is expected |
| 655 | * to pre-populate all of its PDs. See NKPDPE in vmparam.h. |
| 656 | */ |
| 657 | for (i = 0; i < NKPDPE; i++) { |
| 658 | ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] = |
| 659 | KPDphys + (i << PAGE_SHIFT); |
| 660 | ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |= |
| 661 | PG_RW | PG_V | PG_U; |
| 662 | } |
| 663 | |
| 664 | /* |
| 665 | * Now set up the direct map space using either 2MB or 1GB pages |
| 666 | * Preset PG_M and PG_A because demotion expects it. |
| 667 | * |
| 668 | * When filling in entries in the PD pages make sure any excess |
| 669 | * entries are set to zero as we allocated enough PD pages |
| 670 | */ |
| 671 | if ((amd_feature & AMDID_PAGE1GB) == 0) { |
| 672 | for (i = 0; i < NPDEPG * ndmpdp; i++) { |
| 673 | ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT; |
| 674 | ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS | |
| 675 | PG_G | PG_M | PG_A; |
| 676 | } |
| 677 | |
| 678 | /* |
| 679 | * And the direct map space's PDP |
| 680 | */ |
| 681 | for (i = 0; i < ndmpdp; i++) { |
| 682 | ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys + |
| 683 | (i << PAGE_SHIFT); |
| 684 | ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U; |
| 685 | } |
| 686 | } else { |
| 687 | for (i = 0; i < ndmpdp; i++) { |
| 688 | ((pdp_entry_t *)DMPDPphys)[i] = |
| 689 | (vm_paddr_t)i << PDPSHIFT; |
| 690 | ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS | |
| 691 | PG_G | PG_M | PG_A; |
| 692 | } |
| 693 | } |
| 694 | |
| 695 | /* And recursively map PML4 to itself in order to get PTmap */ |
| 696 | ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys; |
| 697 | ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U; |
| 698 | |
| 699 | /* |
| 700 | * Connect the Direct Map slots up to the PML4 |
| 701 | */ |
| 702 | for (j = 0; j < NDMPML4E; ++j) { |
| 703 | ((pdp_entry_t *)KPML4phys)[DMPML4I + j] = |
| 704 | (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) | |
| 705 | PG_RW | PG_V | PG_U; |
| 706 | } |
| 707 | |
| 708 | /* |
| 709 | * Connect the KVA slot up to the PML4 |
| 710 | */ |
| 711 | ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys; |
| 712 | ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U; |
| 713 | } |
| 714 | |
| 715 | /* |
| 716 | * Bootstrap the system enough to run with virtual memory. |
| 717 | * |
| 718 | * On the i386 this is called after mapping has already been enabled |
| 719 | * and just syncs the pmap module with what has already been done. |
| 720 | * [We can't call it easily with mapping off since the kernel is not |
| 721 | * mapped with PA == VA, hence we would have to relocate every address |
| 722 | * from the linked base (virtual) address "KERNBASE" to the actual |
| 723 | * (physical) address starting relative to 0] |
| 724 | */ |
| 725 | void |
| 726 | pmap_bootstrap(vm_paddr_t *firstaddr) |
| 727 | { |
| 728 | vm_offset_t va; |
| 729 | pt_entry_t *pte; |
| 730 | struct mdglobaldata *gd; |
| 731 | int pg; |
| 732 | |
| 733 | KvaStart = VM_MIN_KERNEL_ADDRESS; |
| 734 | KvaEnd = VM_MAX_KERNEL_ADDRESS; |
| 735 | KvaSize = KvaEnd - KvaStart; |
| 736 | |
| 737 | avail_start = *firstaddr; |
| 738 | |
| 739 | /* |
| 740 | * Create an initial set of page tables to run the kernel in. |
| 741 | */ |
| 742 | create_pagetables(firstaddr); |
| 743 | |
| 744 | virtual2_start = KvaStart; |
| 745 | virtual2_end = PTOV_OFFSET; |
| 746 | |
| 747 | virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr; |
| 748 | virtual_start = pmap_kmem_choose(virtual_start); |
| 749 | |
| 750 | virtual_end = VM_MAX_KERNEL_ADDRESS; |
| 751 | |
| 752 | /* XXX do %cr0 as well */ |
| 753 | load_cr4(rcr4() | CR4_PGE | CR4_PSE); |
| 754 | load_cr3(KPML4phys); |
| 755 | |
| 756 | /* |
| 757 | * Initialize protection array. |
| 758 | */ |
| 759 | i386_protection_init(); |
| 760 | |
| 761 | /* |
| 762 | * The kernel's pmap is statically allocated so we don't have to use |
| 763 | * pmap_create, which is unlikely to work correctly at this part of |
| 764 | * the boot sequence (XXX and which no longer exists). |
| 765 | */ |
| 766 | kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys); |
| 767 | kernel_pmap.pm_count = 1; |
| 768 | kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK; |
| 769 | RB_INIT(&kernel_pmap.pm_pvroot); |
| 770 | spin_init(&kernel_pmap.pm_spin); |
| 771 | lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok"); |
| 772 | |
| 773 | /* |
| 774 | * Reserve some special page table entries/VA space for temporary |
| 775 | * mapping of pages. |
| 776 | */ |
| 777 | #define SYSMAP(c, p, v, n) \ |
| 778 | v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); |
| 779 | |
| 780 | va = virtual_start; |
| 781 | pte = vtopte(va); |
| 782 | |
| 783 | /* |
| 784 | * CMAP1/CMAP2 are used for zeroing and copying pages. |
| 785 | */ |
| 786 | SYSMAP(caddr_t, CMAP1, CADDR1, 1) |
| 787 | |
| 788 | /* |
| 789 | * Crashdump maps. |
| 790 | */ |
| 791 | SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS); |
| 792 | |
| 793 | /* |
| 794 | * ptvmmap is used for reading arbitrary physical pages via |
| 795 | * /dev/mem. |
| 796 | */ |
| 797 | SYSMAP(caddr_t, ptmmap, ptvmmap, 1) |
| 798 | |
| 799 | /* |
| 800 | * msgbufp is used to map the system message buffer. |
| 801 | * XXX msgbufmap is not used. |
| 802 | */ |
| 803 | SYSMAP(struct msgbuf *, msgbufmap, msgbufp, |
| 804 | atop(round_page(MSGBUF_SIZE))) |
| 805 | |
| 806 | virtual_start = va; |
| 807 | |
| 808 | *CMAP1 = 0; |
| 809 | |
| 810 | /* |
| 811 | * PG_G is terribly broken on SMP because we IPI invltlb's in some |
| 812 | * cases rather then invl1pg. Actually, I don't even know why it |
| 813 | * works under UP because self-referential page table mappings |
| 814 | */ |
| 815 | #ifdef SMP |
| 816 | pgeflag = 0; |
| 817 | #else |
| 818 | if (cpu_feature & CPUID_PGE) |
| 819 | pgeflag = PG_G; |
| 820 | #endif |
| 821 | |
| 822 | /* |
| 823 | * Initialize the 4MB page size flag |
| 824 | */ |
| 825 | pseflag = 0; |
| 826 | /* |
| 827 | * The 4MB page version of the initial |
| 828 | * kernel page mapping. |
| 829 | */ |
| 830 | pdir4mb = 0; |
| 831 | |
| 832 | #if !defined(DISABLE_PSE) |
| 833 | if (cpu_feature & CPUID_PSE) { |
| 834 | pt_entry_t ptditmp; |
| 835 | /* |
| 836 | * Note that we have enabled PSE mode |
| 837 | */ |
| 838 | pseflag = PG_PS; |
| 839 | ptditmp = *(PTmap + x86_64_btop(KERNBASE)); |
| 840 | ptditmp &= ~(NBPDR - 1); |
| 841 | ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag; |
| 842 | pdir4mb = ptditmp; |
| 843 | |
| 844 | #ifndef SMP |
| 845 | /* |
| 846 | * Enable the PSE mode. If we are SMP we can't do this |
| 847 | * now because the APs will not be able to use it when |
| 848 | * they boot up. |
| 849 | */ |
| 850 | load_cr4(rcr4() | CR4_PSE); |
| 851 | |
| 852 | /* |
| 853 | * We can do the mapping here for the single processor |
| 854 | * case. We simply ignore the old page table page from |
| 855 | * now on. |
| 856 | */ |
| 857 | /* |
| 858 | * For SMP, we still need 4K pages to bootstrap APs, |
| 859 | * PSE will be enabled as soon as all APs are up. |
| 860 | */ |
| 861 | PTD[KPTDI] = (pd_entry_t)ptditmp; |
| 862 | cpu_invltlb(); |
| 863 | #endif |
| 864 | } |
| 865 | #endif |
| 866 | |
| 867 | /* |
| 868 | * We need to finish setting up the globaldata page for the BSP. |
| 869 | * locore has already populated the page table for the mdglobaldata |
| 870 | * portion. |
| 871 | */ |
| 872 | pg = MDGLOBALDATA_BASEALLOC_PAGES; |
| 873 | gd = &CPU_prvspace[0].mdglobaldata; |
| 874 | |
| 875 | cpu_invltlb(); |
| 876 | } |
| 877 | |
| 878 | #ifdef SMP |
| 879 | /* |
| 880 | * Set 4mb pdir for mp startup |
| 881 | */ |
| 882 | void |
| 883 | pmap_set_opt(void) |
| 884 | { |
| 885 | if (pseflag && (cpu_feature & CPUID_PSE)) { |
| 886 | load_cr4(rcr4() | CR4_PSE); |
| 887 | if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */ |
| 888 | cpu_invltlb(); |
| 889 | } |
| 890 | } |
| 891 | } |
| 892 | #endif |
| 893 | |
| 894 | /* |
| 895 | * Initialize the pmap module. |
| 896 | * Called by vm_init, to initialize any structures that the pmap |
| 897 | * system needs to map virtual memory. |
| 898 | * pmap_init has been enhanced to support in a fairly consistant |
| 899 | * way, discontiguous physical memory. |
| 900 | */ |
| 901 | void |
| 902 | pmap_init(void) |
| 903 | { |
| 904 | int i; |
| 905 | int initial_pvs; |
| 906 | |
| 907 | /* |
| 908 | * Allocate memory for random pmap data structures. Includes the |
| 909 | * pv_head_table. |
| 910 | */ |
| 911 | |
| 912 | for (i = 0; i < vm_page_array_size; i++) { |
| 913 | vm_page_t m; |
| 914 | |
| 915 | m = &vm_page_array[i]; |
| 916 | TAILQ_INIT(&m->md.pv_list); |
| 917 | } |
| 918 | |
| 919 | /* |
| 920 | * init the pv free list |
| 921 | */ |
| 922 | initial_pvs = vm_page_array_size; |
| 923 | if (initial_pvs < MINPV) |
| 924 | initial_pvs = MINPV; |
| 925 | pvzone = &pvzone_store; |
| 926 | pvinit = (void *)kmem_alloc(&kernel_map, |
| 927 | initial_pvs * sizeof (struct pv_entry)); |
| 928 | zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), |
| 929 | pvinit, initial_pvs); |
| 930 | |
| 931 | /* |
| 932 | * Now it is safe to enable pv_table recording. |
| 933 | */ |
| 934 | pmap_initialized = TRUE; |
| 935 | } |
| 936 | |
| 937 | /* |
| 938 | * Initialize the address space (zone) for the pv_entries. Set a |
| 939 | * high water mark so that the system can recover from excessive |
| 940 | * numbers of pv entries. |
| 941 | */ |
| 942 | void |
| 943 | pmap_init2(void) |
| 944 | { |
| 945 | int shpgperproc = PMAP_SHPGPERPROC; |
| 946 | int entry_max; |
| 947 | |
| 948 | TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); |
| 949 | pv_entry_max = shpgperproc * maxproc + vm_page_array_size; |
| 950 | TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); |
| 951 | pv_entry_high_water = 9 * (pv_entry_max / 10); |
| 952 | |
| 953 | /* |
| 954 | * Subtract out pages already installed in the zone (hack) |
| 955 | */ |
| 956 | entry_max = pv_entry_max - vm_page_array_size; |
| 957 | if (entry_max <= 0) |
| 958 | entry_max = 1; |
| 959 | |
| 960 | zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1); |
| 961 | } |
| 962 | |
| 963 | |
| 964 | /*************************************************** |
| 965 | * Low level helper routines..... |
| 966 | ***************************************************/ |
| 967 | |
| 968 | /* |
| 969 | * this routine defines the region(s) of memory that should |
| 970 | * not be tested for the modified bit. |
| 971 | */ |
| 972 | static __inline |
| 973 | int |
| 974 | pmap_track_modified(vm_pindex_t pindex) |
| 975 | { |
| 976 | vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT; |
| 977 | if ((va < clean_sva) || (va >= clean_eva)) |
| 978 | return 1; |
| 979 | else |
| 980 | return 0; |
| 981 | } |
| 982 | |
| 983 | /* |
| 984 | * Extract the physical page address associated with the map/VA pair. |
| 985 | * The page must be wired for this to work reliably. |
| 986 | * |
| 987 | * XXX for the moment we're using pv_find() instead of pv_get(), as |
| 988 | * callers might be expecting non-blocking operation. |
| 989 | */ |
| 990 | vm_paddr_t |
| 991 | pmap_extract(pmap_t pmap, vm_offset_t va) |
| 992 | { |
| 993 | vm_paddr_t rtval; |
| 994 | pv_entry_t pt_pv; |
| 995 | pt_entry_t *ptep; |
| 996 | |
| 997 | rtval = 0; |
| 998 | if (va >= VM_MAX_USER_ADDRESS) { |
| 999 | /* |
| 1000 | * Kernel page directories might be direct-mapped and |
| 1001 | * there is typically no PV tracking of pte's |
| 1002 | */ |
| 1003 | pd_entry_t *pt; |
| 1004 | |
| 1005 | pt = pmap_pt(pmap, va); |
| 1006 | if (pt && (*pt & PG_V)) { |
| 1007 | if (*pt & PG_PS) { |
| 1008 | rtval = *pt & PG_PS_FRAME; |
| 1009 | rtval |= va & PDRMASK; |
| 1010 | } else { |
| 1011 | ptep = pmap_pt_to_pte(pt, va); |
| 1012 | if (*pt & PG_V) { |
| 1013 | rtval = *ptep & PG_FRAME; |
| 1014 | rtval |= va & PAGE_MASK; |
| 1015 | } |
| 1016 | } |
| 1017 | } |
| 1018 | } else { |
| 1019 | /* |
| 1020 | * User pages currently do not direct-map the page directory |
| 1021 | * and some pages might not used managed PVs. But all PT's |
| 1022 | * will have a PV. |
| 1023 | */ |
| 1024 | pt_pv = pv_find(pmap, pmap_pt_pindex(va)); |
| 1025 | if (pt_pv) { |
| 1026 | ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va)); |
| 1027 | if (*ptep & PG_V) { |
| 1028 | rtval = *ptep & PG_FRAME; |
| 1029 | rtval |= va & PAGE_MASK; |
| 1030 | } |
| 1031 | pv_drop(pt_pv); |
| 1032 | } |
| 1033 | } |
| 1034 | return rtval; |
| 1035 | } |
| 1036 | |
| 1037 | /* |
| 1038 | * Extract the physical page address associated kernel virtual address. |
| 1039 | */ |
| 1040 | vm_paddr_t |
| 1041 | pmap_kextract(vm_offset_t va) |
| 1042 | { |
| 1043 | pd_entry_t pt; /* pt entry in pd */ |
| 1044 | vm_paddr_t pa; |
| 1045 | |
| 1046 | if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) { |
| 1047 | pa = DMAP_TO_PHYS(va); |
| 1048 | } else { |
| 1049 | pt = *vtopt(va); |
| 1050 | if (pt & PG_PS) { |
| 1051 | pa = (pt & PG_PS_FRAME) | (va & PDRMASK); |
| 1052 | } else { |
| 1053 | /* |
| 1054 | * Beware of a concurrent promotion that changes the |
| 1055 | * PDE at this point! For example, vtopte() must not |
| 1056 | * be used to access the PTE because it would use the |
| 1057 | * new PDE. It is, however, safe to use the old PDE |
| 1058 | * because the page table page is preserved by the |
| 1059 | * promotion. |
| 1060 | */ |
| 1061 | pa = *pmap_pt_to_pte(&pt, va); |
| 1062 | pa = (pa & PG_FRAME) | (va & PAGE_MASK); |
| 1063 | } |
| 1064 | } |
| 1065 | return pa; |
| 1066 | } |
| 1067 | |
| 1068 | /*************************************************** |
| 1069 | * Low level mapping routines..... |
| 1070 | ***************************************************/ |
| 1071 | |
| 1072 | /* |
| 1073 | * Routine: pmap_kenter |
| 1074 | * Function: |
| 1075 | * Add a wired page to the KVA |
| 1076 | * NOTE! note that in order for the mapping to take effect -- you |
| 1077 | * should do an invltlb after doing the pmap_kenter(). |
| 1078 | */ |
| 1079 | void |
| 1080 | pmap_kenter(vm_offset_t va, vm_paddr_t pa) |
| 1081 | { |
| 1082 | pt_entry_t *pte; |
| 1083 | pt_entry_t npte; |
| 1084 | pmap_inval_info info; |
| 1085 | |
| 1086 | pmap_inval_init(&info); /* XXX remove */ |
| 1087 | npte = pa | PG_RW | PG_V | pgeflag; |
| 1088 | pte = vtopte(va); |
| 1089 | pmap_inval_interlock(&info, &kernel_pmap, va); /* XXX remove */ |
| 1090 | *pte = npte; |
| 1091 | pmap_inval_deinterlock(&info, &kernel_pmap); /* XXX remove */ |
| 1092 | pmap_inval_done(&info); /* XXX remove */ |
| 1093 | } |
| 1094 | |
| 1095 | /* |
| 1096 | * Routine: pmap_kenter_quick |
| 1097 | * Function: |
| 1098 | * Similar to pmap_kenter(), except we only invalidate the |
| 1099 | * mapping on the current CPU. |
| 1100 | */ |
| 1101 | void |
| 1102 | pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa) |
| 1103 | { |
| 1104 | pt_entry_t *pte; |
| 1105 | pt_entry_t npte; |
| 1106 | |
| 1107 | npte = pa | PG_RW | PG_V | pgeflag; |
| 1108 | pte = vtopte(va); |
| 1109 | *pte = npte; |
| 1110 | cpu_invlpg((void *)va); |
| 1111 | } |
| 1112 | |
| 1113 | void |
| 1114 | pmap_kenter_sync(vm_offset_t va) |
| 1115 | { |
| 1116 | pmap_inval_info info; |
| 1117 | |
| 1118 | pmap_inval_init(&info); |
| 1119 | pmap_inval_interlock(&info, &kernel_pmap, va); |
| 1120 | pmap_inval_deinterlock(&info, &kernel_pmap); |
| 1121 | pmap_inval_done(&info); |
| 1122 | } |
| 1123 | |
| 1124 | void |
| 1125 | pmap_kenter_sync_quick(vm_offset_t va) |
| 1126 | { |
| 1127 | cpu_invlpg((void *)va); |
| 1128 | } |
| 1129 | |
| 1130 | /* |
| 1131 | * remove a page from the kernel pagetables |
| 1132 | */ |
| 1133 | void |
| 1134 | pmap_kremove(vm_offset_t va) |
| 1135 | { |
| 1136 | pt_entry_t *pte; |
| 1137 | pmap_inval_info info; |
| 1138 | |
| 1139 | pmap_inval_init(&info); |
| 1140 | pte = vtopte(va); |
| 1141 | pmap_inval_interlock(&info, &kernel_pmap, va); |
| 1142 | (void)pte_load_clear(pte); |
| 1143 | pmap_inval_deinterlock(&info, &kernel_pmap); |
| 1144 | pmap_inval_done(&info); |
| 1145 | } |
| 1146 | |
| 1147 | void |
| 1148 | pmap_kremove_quick(vm_offset_t va) |
| 1149 | { |
| 1150 | pt_entry_t *pte; |
| 1151 | pte = vtopte(va); |
| 1152 | (void)pte_load_clear(pte); |
| 1153 | cpu_invlpg((void *)va); |
| 1154 | } |
| 1155 | |
| 1156 | /* |
| 1157 | * XXX these need to be recoded. They are not used in any critical path. |
| 1158 | */ |
| 1159 | void |
| 1160 | pmap_kmodify_rw(vm_offset_t va) |
| 1161 | { |
| 1162 | atomic_set_long(vtopte(va), PG_RW); |
| 1163 | cpu_invlpg((void *)va); |
| 1164 | } |
| 1165 | |
| 1166 | void |
| 1167 | pmap_kmodify_nc(vm_offset_t va) |
| 1168 | { |
| 1169 | atomic_set_long(vtopte(va), PG_N); |
| 1170 | cpu_invlpg((void *)va); |
| 1171 | } |
| 1172 | |
| 1173 | /* |
| 1174 | * Used to map a range of physical addresses into kernel virtual |
| 1175 | * address space during the low level boot, typically to map the |
| 1176 | * dump bitmap, message buffer, and vm_page_array. |
| 1177 | * |
| 1178 | * These mappings are typically made at some pointer after the end of the |
| 1179 | * kernel text+data. |
| 1180 | * |
| 1181 | * We could return PHYS_TO_DMAP(start) here and not allocate any |
| 1182 | * via (*virtp), but then kmem from userland and kernel dumps won't |
| 1183 | * have access to the related pointers. |
| 1184 | */ |
| 1185 | vm_offset_t |
| 1186 | pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot) |
| 1187 | { |
| 1188 | vm_offset_t va; |
| 1189 | vm_offset_t va_start; |
| 1190 | |
| 1191 | /*return PHYS_TO_DMAP(start);*/ |
| 1192 | |
| 1193 | va_start = *virtp; |
| 1194 | va = va_start; |
| 1195 | |
| 1196 | while (start < end) { |
| 1197 | pmap_kenter_quick(va, start); |
| 1198 | va += PAGE_SIZE; |
| 1199 | start += PAGE_SIZE; |
| 1200 | } |
| 1201 | *virtp = va; |
| 1202 | return va_start; |
| 1203 | } |
| 1204 | |
| 1205 | |
| 1206 | /* |
| 1207 | * Add a list of wired pages to the kva |
| 1208 | * this routine is only used for temporary |
| 1209 | * kernel mappings that do not need to have |
| 1210 | * page modification or references recorded. |
| 1211 | * Note that old mappings are simply written |
| 1212 | * over. The page *must* be wired. |
| 1213 | */ |
| 1214 | void |
| 1215 | pmap_qenter(vm_offset_t va, vm_page_t *m, int count) |
| 1216 | { |
| 1217 | vm_offset_t end_va; |
| 1218 | |
| 1219 | end_va = va + count * PAGE_SIZE; |
| 1220 | |
| 1221 | while (va < end_va) { |
| 1222 | pt_entry_t *pte; |
| 1223 | |
| 1224 | pte = vtopte(va); |
| 1225 | *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag; |
| 1226 | cpu_invlpg((void *)va); |
| 1227 | va += PAGE_SIZE; |
| 1228 | m++; |
| 1229 | } |
| 1230 | smp_invltlb(); |
| 1231 | } |
| 1232 | |
| 1233 | /* |
| 1234 | * This routine jerks page mappings from the |
| 1235 | * kernel -- it is meant only for temporary mappings. |
| 1236 | * |
| 1237 | * MPSAFE, INTERRUPT SAFE (cluster callback) |
| 1238 | */ |
| 1239 | void |
| 1240 | pmap_qremove(vm_offset_t va, int count) |
| 1241 | { |
| 1242 | vm_offset_t end_va; |
| 1243 | |
| 1244 | end_va = va + count * PAGE_SIZE; |
| 1245 | |
| 1246 | while (va < end_va) { |
| 1247 | pt_entry_t *pte; |
| 1248 | |
| 1249 | pte = vtopte(va); |
| 1250 | (void)pte_load_clear(pte); |
| 1251 | cpu_invlpg((void *)va); |
| 1252 | va += PAGE_SIZE; |
| 1253 | } |
| 1254 | smp_invltlb(); |
| 1255 | } |
| 1256 | |
| 1257 | /* |
| 1258 | * Create a new thread and optionally associate it with a (new) process. |
| 1259 | * NOTE! the new thread's cpu may not equal the current cpu. |
| 1260 | */ |
| 1261 | void |
| 1262 | pmap_init_thread(thread_t td) |
| 1263 | { |
| 1264 | /* enforce pcb placement & alignment */ |
| 1265 | td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1; |
| 1266 | td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF); |
| 1267 | td->td_savefpu = &td->td_pcb->pcb_save; |
| 1268 | td->td_sp = (char *)td->td_pcb; /* no -16 */ |
| 1269 | } |
| 1270 | |
| 1271 | /* |
| 1272 | * This routine directly affects the fork perf for a process. |
| 1273 | */ |
| 1274 | void |
| 1275 | pmap_init_proc(struct proc *p) |
| 1276 | { |
| 1277 | } |
| 1278 | |
| 1279 | /* |
| 1280 | * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because |
| 1281 | * it, and IdlePTD, represents the template used to update all other pmaps. |
| 1282 | * |
| 1283 | * On architectures where the kernel pmap is not integrated into the user |
| 1284 | * process pmap, this pmap represents the process pmap, not the kernel pmap. |
| 1285 | * kernel_pmap should be used to directly access the kernel_pmap. |
| 1286 | */ |
| 1287 | void |
| 1288 | pmap_pinit0(struct pmap *pmap) |
| 1289 | { |
| 1290 | pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys); |
| 1291 | pmap->pm_count = 1; |
| 1292 | pmap->pm_active = 0; |
| 1293 | pmap->pm_pvhint = NULL; |
| 1294 | RB_INIT(&pmap->pm_pvroot); |
| 1295 | spin_init(&pmap->pm_spin); |
| 1296 | lwkt_token_init(&pmap->pm_token, "pmap_tok"); |
| 1297 | bzero(&pmap->pm_stats, sizeof pmap->pm_stats); |
| 1298 | } |
| 1299 | |
| 1300 | /* |
| 1301 | * Initialize a preallocated and zeroed pmap structure, |
| 1302 | * such as one in a vmspace structure. |
| 1303 | */ |
| 1304 | void |
| 1305 | pmap_pinit(struct pmap *pmap) |
| 1306 | { |
| 1307 | pv_entry_t pv; |
| 1308 | int j; |
| 1309 | |
| 1310 | /* |
| 1311 | * Misc initialization |
| 1312 | */ |
| 1313 | pmap->pm_count = 1; |
| 1314 | pmap->pm_active = 0; |
| 1315 | pmap->pm_pvhint = NULL; |
| 1316 | if (pmap->pm_pmlpv == NULL) { |
| 1317 | RB_INIT(&pmap->pm_pvroot); |
| 1318 | bzero(&pmap->pm_stats, sizeof pmap->pm_stats); |
| 1319 | spin_init(&pmap->pm_spin); |
| 1320 | lwkt_token_init(&pmap->pm_token, "pmap_tok"); |
| 1321 | } |
| 1322 | |
| 1323 | /* |
| 1324 | * No need to allocate page table space yet but we do need a valid |
| 1325 | * page directory table. |
| 1326 | */ |
| 1327 | if (pmap->pm_pml4 == NULL) { |
| 1328 | pmap->pm_pml4 = |
| 1329 | (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE); |
| 1330 | } |
| 1331 | |
| 1332 | /* |
| 1333 | * Allocate the page directory page, which wires it even though |
| 1334 | * it isn't being entered into some higher level page table (it |
| 1335 | * being the highest level). If one is already cached we don't |
| 1336 | * have to do anything. |
| 1337 | */ |
| 1338 | if ((pv = pmap->pm_pmlpv) == NULL) { |
| 1339 | pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL); |
| 1340 | pmap->pm_pmlpv = pv; |
| 1341 | pmap_kenter((vm_offset_t)pmap->pm_pml4, |
| 1342 | VM_PAGE_TO_PHYS(pv->pv_m)); |
| 1343 | pv_put(pv); |
| 1344 | |
| 1345 | /* |
| 1346 | * Install DMAP and KMAP. |
| 1347 | */ |
| 1348 | for (j = 0; j < NDMPML4E; ++j) { |
| 1349 | pmap->pm_pml4[DMPML4I + j] = |
| 1350 | (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) | |
| 1351 | PG_RW | PG_V | PG_U; |
| 1352 | } |
| 1353 | pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U; |
| 1354 | |
| 1355 | /* |
| 1356 | * install self-referential address mapping entry |
| 1357 | */ |
| 1358 | pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) | |
| 1359 | PG_V | PG_RW | PG_A | PG_M; |
| 1360 | } else { |
| 1361 | KKASSERT(pv->pv_m->flags & PG_MAPPED); |
| 1362 | KKASSERT(pv->pv_m->flags & PG_WRITEABLE); |
| 1363 | } |
| 1364 | KKASSERT(pmap->pm_pml4[255] == 0); |
| 1365 | KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv); |
| 1366 | KKASSERT(pv->pv_entry.rbe_left == NULL); |
| 1367 | KKASSERT(pv->pv_entry.rbe_right == NULL); |
| 1368 | } |
| 1369 | |
| 1370 | /* |
| 1371 | * Clean up a pmap structure so it can be physically freed. This routine |
| 1372 | * is called by the vmspace dtor function. A great deal of pmap data is |
| 1373 | * left passively mapped to improve vmspace management so we have a bit |
| 1374 | * of cleanup work to do here. |
| 1375 | */ |
| 1376 | void |
| 1377 | pmap_puninit(pmap_t pmap) |
| 1378 | { |
| 1379 | pv_entry_t pv; |
| 1380 | vm_page_t p; |
| 1381 | |
| 1382 | KKASSERT(pmap->pm_active == 0); |
| 1383 | if ((pv = pmap->pm_pmlpv) != NULL) { |
| 1384 | if (pv_hold_try(pv) == 0) |
| 1385 | pv_lock(pv); |
| 1386 | p = pmap_remove_pv_page(pv); |
| 1387 | pv_free(pv); |
| 1388 | pmap_kremove((vm_offset_t)pmap->pm_pml4); |
| 1389 | vm_page_busy_wait(p, FALSE, "pgpun"); |
| 1390 | KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED)); |
| 1391 | vm_page_unwire(p, 0); |
| 1392 | vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE); |
| 1393 | |
| 1394 | /* |
| 1395 | * XXX eventually clean out PML4 static entries and |
| 1396 | * use vm_page_free_zero() |
| 1397 | */ |
| 1398 | vm_page_free(p); |
| 1399 | pmap->pm_pmlpv = NULL; |
| 1400 | } |
| 1401 | if (pmap->pm_pml4) { |
| 1402 | KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys)); |
| 1403 | kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE); |
| 1404 | pmap->pm_pml4 = NULL; |
| 1405 | } |
| 1406 | KKASSERT(pmap->pm_stats.resident_count == 0); |
| 1407 | KKASSERT(pmap->pm_stats.wired_count == 0); |
| 1408 | } |
| 1409 | |
| 1410 | /* |
| 1411 | * Wire in kernel global address entries. To avoid a race condition |
| 1412 | * between pmap initialization and pmap_growkernel, this procedure |
| 1413 | * adds the pmap to the master list (which growkernel scans to update), |
| 1414 | * then copies the template. |
| 1415 | */ |
| 1416 | void |
| 1417 | pmap_pinit2(struct pmap *pmap) |
| 1418 | { |
| 1419 | /* |
| 1420 | * XXX copies current process, does not fill in MPPTDI |
| 1421 | */ |
| 1422 | spin_lock(&pmap_spin); |
| 1423 | TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode); |
| 1424 | spin_unlock(&pmap_spin); |
| 1425 | } |
| 1426 | |
| 1427 | /* |
| 1428 | * This routine is called when various levels in the page table need to |
| 1429 | * be populated. This routine cannot fail. |
| 1430 | * |
| 1431 | * This function returns two locked pv_entry's, one representing the |
| 1432 | * requested pv and one representing the requested pv's parent pv. If |
| 1433 | * the pv did not previously exist it will be mapped into its parent |
| 1434 | * and wired, otherwise no additional wire count will be added. |
| 1435 | */ |
| 1436 | static |
| 1437 | pv_entry_t |
| 1438 | pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp) |
| 1439 | { |
| 1440 | pt_entry_t *ptep; |
| 1441 | pv_entry_t pv; |
| 1442 | pv_entry_t pvp; |
| 1443 | vm_pindex_t pt_pindex; |
| 1444 | vm_page_t m; |
| 1445 | int isnew; |
| 1446 | |
| 1447 | /* |
| 1448 | * If the pv already exists and we aren't being asked for the |
| 1449 | * parent page table page we can just return it. A locked+held pv |
| 1450 | * is returned. |
| 1451 | */ |
| 1452 | pv = pv_alloc(pmap, ptepindex, &isnew); |
| 1453 | if (isnew == 0 && pvpp == NULL) |
| 1454 | return(pv); |
| 1455 | |
| 1456 | /* |
| 1457 | * This is a new PV, we have to resolve its parent page table and |
| 1458 | * add an additional wiring to the page if necessary. |
| 1459 | */ |
| 1460 | |
| 1461 | /* |
| 1462 | * Special case terminal PVs. These are not page table pages so |
| 1463 | * no vm_page is allocated (the caller supplied the vm_page). If |
| 1464 | * pvpp is non-NULL we are being asked to also removed the pt_pv |
| 1465 | * for this pv. |
| 1466 | * |
| 1467 | * Note that pt_pv's are only returned for user VAs. We assert that |
| 1468 | * a pt_pv is not being requested for kernel VAs. |
| 1469 | */ |
| 1470 | if (ptepindex < pmap_pt_pindex(0)) { |
| 1471 | if (ptepindex >= NUPTE_USER) |
| 1472 | KKASSERT(pvpp == NULL); |
| 1473 | else |
| 1474 | KKASSERT(pvpp != NULL); |
| 1475 | if (pvpp) { |
| 1476 | pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT); |
| 1477 | pvp = pmap_allocpte(pmap, pt_pindex, NULL); |
| 1478 | if (isnew) |
| 1479 | vm_page_wire_quick(pvp->pv_m); |
| 1480 | *pvpp = pvp; |
| 1481 | } else { |
| 1482 | pvp = NULL; |
| 1483 | } |
| 1484 | return(pv); |
| 1485 | } |
| 1486 | |
| 1487 | /* |
| 1488 | * Non-terminal PVs allocate a VM page to represent the page table, |
| 1489 | * so we have to resolve pvp and calculate ptepindex for the pvp |
| 1490 | * and then for the page table entry index in the pvp for |
| 1491 | * fall-through. |
| 1492 | */ |
| 1493 | if (ptepindex < pmap_pd_pindex(0)) { |
| 1494 | /* |
| 1495 | * pv is PT, pvp is PD |
| 1496 | */ |
| 1497 | ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT; |
| 1498 | ptepindex += NUPTE_TOTAL + NUPT_TOTAL; |
| 1499 | pvp = pmap_allocpte(pmap, ptepindex, NULL); |
| 1500 | if (!isnew) |
| 1501 | goto notnew; |
| 1502 | |
| 1503 | /* |
| 1504 | * PT index in PD |
| 1505 | */ |
| 1506 | ptepindex = pv->pv_pindex - pmap_pt_pindex(0); |
| 1507 | ptepindex &= ((1ul << NPDEPGSHIFT) - 1); |
| 1508 | } else if (ptepindex < pmap_pdp_pindex(0)) { |
| 1509 | /* |
| 1510 | * pv is PD, pvp is PDP |
| 1511 | */ |
| 1512 | ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT; |
| 1513 | ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL; |
| 1514 | pvp = pmap_allocpte(pmap, ptepindex, NULL); |
| 1515 | if (!isnew) |
| 1516 | goto notnew; |
| 1517 | |
| 1518 | /* |
| 1519 | * PD index in PDP |
| 1520 | */ |
| 1521 | ptepindex = pv->pv_pindex - pmap_pd_pindex(0); |
| 1522 | ptepindex &= ((1ul << NPDPEPGSHIFT) - 1); |
| 1523 | } else if (ptepindex < pmap_pml4_pindex()) { |
| 1524 | /* |
| 1525 | * pv is PDP, pvp is the root pml4 table |
| 1526 | */ |
| 1527 | pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL); |
| 1528 | if (!isnew) |
| 1529 | goto notnew; |
| 1530 | |
| 1531 | /* |
| 1532 | * PDP index in PML4 |
| 1533 | */ |
| 1534 | ptepindex = pv->pv_pindex - pmap_pdp_pindex(0); |
| 1535 | ptepindex &= ((1ul << NPML4EPGSHIFT) - 1); |
| 1536 | } else { |
| 1537 | /* |
| 1538 | * pv represents the top-level PML4, there is no parent. |
| 1539 | */ |
| 1540 | pvp = NULL; |
| 1541 | if (!isnew) |
| 1542 | goto notnew; |
| 1543 | } |
| 1544 | |
| 1545 | /* |
| 1546 | * This code is only reached if isnew is TRUE and this is not a |
| 1547 | * terminal PV. We need to allocate a vm_page for the page table |
| 1548 | * at this level and enter it into the parent page table. |
| 1549 | * |
| 1550 | * page table pages are marked PG_WRITEABLE and PG_MAPPED. |
| 1551 | */ |
| 1552 | for (;;) { |
| 1553 | m = vm_page_alloc(NULL, pv->pv_pindex, |
| 1554 | VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | |
| 1555 | VM_ALLOC_INTERRUPT); |
| 1556 | if (m) |
| 1557 | break; |
| 1558 | vm_wait(0); |
| 1559 | } |
| 1560 | vm_page_spin_lock(m); |
| 1561 | TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); |
| 1562 | pv->pv_m = m; |
| 1563 | vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE); |
| 1564 | vm_page_spin_unlock(m); |
| 1565 | vm_page_unmanage(m); /* m must be spinunlocked */ |
| 1566 | |
| 1567 | if ((m->flags & PG_ZERO) == 0) { |
| 1568 | pmap_zero_page(VM_PAGE_TO_PHYS(m)); |
| 1569 | } |
| 1570 | #ifdef PMAP_DEBUG |
| 1571 | else { |
| 1572 | pmap_page_assertzero(VM_PAGE_TO_PHYS(m)); |
| 1573 | } |
| 1574 | #endif |
| 1575 | m->valid = VM_PAGE_BITS_ALL; |
| 1576 | vm_page_flag_clear(m, PG_ZERO); |
| 1577 | vm_page_wire(m); /* wire for mapping in parent */ |
| 1578 | |
| 1579 | /* |
| 1580 | * Wire the page into pvp, bump the wire-count for pvp's page table |
| 1581 | * page. Bump the resident_count for the pmap. There is no pvp |
| 1582 | * for the top level, address the pm_pml4[] array directly. |
| 1583 | * |
| 1584 | * If the caller wants the parent we return it, otherwise |
| 1585 | * we just put it away. |
| 1586 | * |
| 1587 | * No interlock is needed for pte 0 -> non-zero. |
| 1588 | */ |
| 1589 | if (pvp) { |
| 1590 | vm_page_wire_quick(pvp->pv_m); |
| 1591 | ptep = pv_pte_lookup(pvp, ptepindex); |
| 1592 | KKASSERT((*ptep & PG_V) == 0); |
| 1593 | *ptep = VM_PAGE_TO_PHYS(m) | (PG_U | PG_RW | PG_V | |
| 1594 | PG_A | PG_M); |
| 1595 | } |
| 1596 | vm_page_wakeup(m); |
| 1597 | notnew: |
| 1598 | if (pvpp) |
| 1599 | *pvpp = pvp; |
| 1600 | else if (pvp) |
| 1601 | pv_put(pvp); |
| 1602 | return (pv); |
| 1603 | } |
| 1604 | |
| 1605 | /* |
| 1606 | * Release any resources held by the given physical map. |
| 1607 | * |
| 1608 | * Called when a pmap initialized by pmap_pinit is being released. Should |
| 1609 | * only be called if the map contains no valid mappings. |
| 1610 | * |
| 1611 | * Caller must hold pmap->pm_token |
| 1612 | */ |
| 1613 | struct pmap_release_info { |
| 1614 | pmap_t pmap; |
| 1615 | int retry; |
| 1616 | }; |
| 1617 | |
| 1618 | static int pmap_release_callback(pv_entry_t pv, void *data); |
| 1619 | |
| 1620 | void |
| 1621 | pmap_release(struct pmap *pmap) |
| 1622 | { |
| 1623 | struct pmap_release_info info; |
| 1624 | |
| 1625 | KASSERT(pmap->pm_active == 0, |
| 1626 | ("pmap still active! %016jx", (uintmax_t)pmap->pm_active)); |
| 1627 | |
| 1628 | spin_lock(&pmap_spin); |
| 1629 | TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode); |
| 1630 | spin_unlock(&pmap_spin); |
| 1631 | |
| 1632 | /* |
| 1633 | * Pull pv's off the RB tree in order from low to high and release |
| 1634 | * each page. |
| 1635 | */ |
| 1636 | info.pmap = pmap; |
| 1637 | do { |
| 1638 | info.retry = 0; |
| 1639 | spin_lock(&pmap->pm_spin); |
| 1640 | RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL, |
| 1641 | pmap_release_callback, &info); |
| 1642 | spin_unlock(&pmap->pm_spin); |
| 1643 | } while (info.retry); |
| 1644 | |
| 1645 | |
| 1646 | /* |
| 1647 | * One resident page (the pml4 page) should remain. |
| 1648 | * No wired pages should remain. |
| 1649 | */ |
| 1650 | KKASSERT(pmap->pm_stats.resident_count == 1); |
| 1651 | KKASSERT(pmap->pm_stats.wired_count == 0); |
| 1652 | } |
| 1653 | |
| 1654 | static int |
| 1655 | pmap_release_callback(pv_entry_t pv, void *data) |
| 1656 | { |
| 1657 | struct pmap_release_info *info = data; |
| 1658 | pmap_t pmap = info->pmap; |
| 1659 | vm_page_t p; |
| 1660 | |
| 1661 | if (pv_hold_try(pv)) { |
| 1662 | spin_unlock(&pmap->pm_spin); |
| 1663 | } else { |
| 1664 | spin_unlock(&pmap->pm_spin); |
| 1665 | pv_lock(pv); |
| 1666 | if (pv->pv_pmap != pmap) { |
| 1667 | pv_put(pv); |
| 1668 | spin_lock(&pmap->pm_spin); |
| 1669 | info->retry = 1; |
| 1670 | return(-1); |
| 1671 | } |
| 1672 | } |
| 1673 | |
| 1674 | /* |
| 1675 | * The pmap is currently not spinlocked, pv is held+locked. |
| 1676 | * Remove the pv's page from its parent's page table. The |
| 1677 | * parent's page table page's wire_count will be decremented. |
| 1678 | */ |
| 1679 | pmap_remove_pv_pte(pv, NULL, NULL); |
| 1680 | |
| 1681 | /* |
| 1682 | * Terminal pvs are unhooked from their vm_pages. Because |
| 1683 | * terminal pages aren't page table pages they aren't wired |
| 1684 | * by us, so we have to be sure not to unwire them either. |
| 1685 | */ |
| 1686 | if (pv->pv_pindex < pmap_pt_pindex(0)) { |
| 1687 | pmap_remove_pv_page(pv); |
| 1688 | goto skip; |
| 1689 | } |
| 1690 | |
| 1691 | /* |
| 1692 | * We leave the top-level page table page cached, wired, and |
| 1693 | * mapped in the pmap until the dtor function (pmap_puninit()) |
| 1694 | * gets called. |
| 1695 | * |
| 1696 | * Since we are leaving the top-level pv intact we need |
| 1697 | * to break out of what would otherwise be an infinite loop. |
| 1698 | */ |
| 1699 | if (pv->pv_pindex == pmap_pml4_pindex()) { |
| 1700 | pv_put(pv); |
| 1701 | spin_lock(&pmap->pm_spin); |
| 1702 | return(-1); |
| 1703 | } |
| 1704 | |
| 1705 | /* |
| 1706 | * For page table pages (other than the top-level page), |
| 1707 | * remove and free the vm_page. The representitive mapping |
| 1708 | * removed above by pmap_remove_pv_pte() did not undo the |
| 1709 | * last wire_count so we have to do that as well. |
| 1710 | */ |
| 1711 | p = pmap_remove_pv_page(pv); |
| 1712 | vm_page_busy_wait(p, FALSE, "pmaprl"); |
| 1713 | if (p->wire_count != 1) { |
| 1714 | kprintf("p->wire_count was %016lx %d\n", |
| 1715 | pv->pv_pindex, p->wire_count); |
| 1716 | } |
| 1717 | KKASSERT(p->wire_count == 1); |
| 1718 | KKASSERT(p->flags & PG_UNMANAGED); |
| 1719 | |
| 1720 | vm_page_unwire(p, 0); |
| 1721 | KKASSERT(p->wire_count == 0); |
| 1722 | /* JG eventually revert to using vm_page_free_zero() */ |
| 1723 | vm_page_free(p); |
| 1724 | skip: |
| 1725 | pv_free(pv); |
| 1726 | spin_lock(&pmap->pm_spin); |
| 1727 | return(0); |
| 1728 | } |
| 1729 | |
| 1730 | /* |
| 1731 | * This function will remove the pte associated with a pv from its parent. |
| 1732 | * Terminal pv's are supported. The removal will be interlocked if info |
| 1733 | * is non-NULL. The caller must dispose of pv instead of just unlocking |
| 1734 | * it. |
| 1735 | * |
| 1736 | * The wire count will be dropped on the parent page table. The wire |
| 1737 | * count on the page being removed (pv->pv_m) from the parent page table |
| 1738 | * is NOT touched. Note that terminal pages will not have any additional |
| 1739 | * wire counts while page table pages will have at least one representing |
| 1740 | * the mapping, plus others representing sub-mappings. |
| 1741 | * |
| 1742 | * NOTE: Cannot be called on kernel page table pages, only KVM terminal |
| 1743 | * pages and user page table and terminal pages. |
| 1744 | * |
| 1745 | * The pv must be locked. |
| 1746 | * |
| 1747 | * XXX must lock parent pv's if they exist to remove pte XXX |
| 1748 | */ |
| 1749 | static |
| 1750 | void |
| 1751 | pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, struct pmap_inval_info *info) |
| 1752 | { |
| 1753 | vm_pindex_t ptepindex = pv->pv_pindex; |
| 1754 | pmap_t pmap = pv->pv_pmap; |
| 1755 | vm_page_t p; |
| 1756 | int gotpvp = 0; |
| 1757 | |
| 1758 | KKASSERT(pmap); |
| 1759 | |
| 1760 | if (ptepindex == pmap_pml4_pindex()) { |
| 1761 | /* |
| 1762 | * We are the top level pml4 table, there is no parent. |
| 1763 | */ |
| 1764 | p = pmap->pm_pmlpv->pv_m; |
| 1765 | } else if (ptepindex >= pmap_pdp_pindex(0)) { |
| 1766 | /* |
| 1767 | * Remove a PDP page from the pml4e. This can only occur |
| 1768 | * with user page tables. We do not have to lock the |
| 1769 | * pml4 PV so just ignore pvp. |
| 1770 | */ |
| 1771 | vm_pindex_t pml4_pindex; |
| 1772 | vm_pindex_t pdp_index; |
| 1773 | pml4_entry_t *pdp; |
| 1774 | |
| 1775 | pdp_index = ptepindex - pmap_pdp_pindex(0); |
| 1776 | if (pvp == NULL) { |
| 1777 | pml4_pindex = pmap_pml4_pindex(); |
| 1778 | pvp = pv_get(pv->pv_pmap, pml4_pindex); |
| 1779 | gotpvp = 1; |
| 1780 | } |
| 1781 | pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)]; |
| 1782 | KKASSERT((*pdp & PG_V) != 0); |
| 1783 | p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME); |
| 1784 | *pdp = 0; |
| 1785 | KKASSERT(info == NULL); |
| 1786 | } else if (ptepindex >= pmap_pd_pindex(0)) { |
| 1787 | /* |
| 1788 | * Remove a PD page from the pdp |
| 1789 | */ |
| 1790 | vm_pindex_t pdp_pindex; |
| 1791 | vm_pindex_t pd_index; |
| 1792 | pdp_entry_t *pd; |
| 1793 | |
| 1794 | pd_index = ptepindex - pmap_pd_pindex(0); |
| 1795 | |
| 1796 | if (pvp == NULL) { |
| 1797 | pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + |
| 1798 | (pd_index >> NPML4EPGSHIFT); |
| 1799 | pvp = pv_get(pv->pv_pmap, pdp_pindex); |
| 1800 | gotpvp = 1; |
| 1801 | } |
| 1802 | pd = pv_pte_lookup(pvp, pd_index & ((1ul << NPDPEPGSHIFT) - 1)); |
| 1803 | KKASSERT((*pd & PG_V) != 0); |
| 1804 | p = PHYS_TO_VM_PAGE(*pd & PG_FRAME); |
| 1805 | *pd = 0; |
| 1806 | KKASSERT(info == NULL); |
| 1807 | } else if (ptepindex >= pmap_pt_pindex(0)) { |
| 1808 | /* |
| 1809 | * Remove a PT page from the pd |
| 1810 | */ |
| 1811 | vm_pindex_t pd_pindex; |
| 1812 | vm_pindex_t pt_index; |
| 1813 | pd_entry_t *pt; |
| 1814 | |
| 1815 | pt_index = ptepindex - pmap_pt_pindex(0); |
| 1816 | |
| 1817 | if (pvp == NULL) { |
| 1818 | pd_pindex = NUPTE_TOTAL + NUPT_TOTAL + |
| 1819 | (pt_index >> NPDPEPGSHIFT); |
| 1820 | pvp = pv_get(pv->pv_pmap, pd_pindex); |
| 1821 | gotpvp = 1; |
| 1822 | } |
| 1823 | pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1)); |
| 1824 | KKASSERT((*pt & PG_V) != 0); |
| 1825 | p = PHYS_TO_VM_PAGE(*pt & PG_FRAME); |
| 1826 | *pt = 0; |
| 1827 | KKASSERT(info == NULL); |
| 1828 | } else { |
| 1829 | /* |
| 1830 | * Remove a PTE from the PT page |
| 1831 | * |
| 1832 | * NOTE: pv's must be locked bottom-up to avoid deadlocking. |
| 1833 | * pv is a pte_pv so we can safely lock pt_pv. |
| 1834 | */ |
| 1835 | vm_pindex_t pt_pindex; |
| 1836 | pt_entry_t *ptep; |
| 1837 | pt_entry_t pte; |
| 1838 | vm_offset_t va; |
| 1839 | |
| 1840 | pt_pindex = ptepindex >> NPTEPGSHIFT; |
| 1841 | va = (vm_offset_t)ptepindex << PAGE_SHIFT; |
| 1842 | |
| 1843 | if (ptepindex >= NUPTE_USER) { |
| 1844 | ptep = vtopte(ptepindex << PAGE_SHIFT); |
| 1845 | KKASSERT(pvp == NULL); |
| 1846 | } else { |
| 1847 | if (pvp == NULL) { |
| 1848 | pt_pindex = NUPTE_TOTAL + |
| 1849 | (ptepindex >> NPDPEPGSHIFT); |
| 1850 | pvp = pv_get(pv->pv_pmap, pt_pindex); |
| 1851 | gotpvp = 1; |
| 1852 | } |
| 1853 | ptep = pv_pte_lookup(pvp, ptepindex & |
| 1854 | ((1ul << NPDPEPGSHIFT) - 1)); |
| 1855 | } |
| 1856 | |
| 1857 | if (info) |
| 1858 | pmap_inval_interlock(info, pmap, va); |
| 1859 | pte = pte_load_clear(ptep); |
| 1860 | if (info) |
| 1861 | pmap_inval_deinterlock(info, pmap); |
| 1862 | else |
| 1863 | cpu_invlpg((void *)va); |
| 1864 | |
| 1865 | /* |
| 1866 | * Now update the vm_page_t |
| 1867 | */ |
| 1868 | if ((pte & (PG_MANAGED|PG_V)) != (PG_MANAGED|PG_V)) { |
| 1869 | kprintf("remove_pte badpte %016lx %016lx %d\n", |
| 1870 | pte, pv->pv_pindex, |
| 1871 | pv->pv_pindex < pmap_pt_pindex(0)); |
| 1872 | } |
| 1873 | /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/ |
| 1874 | p = PHYS_TO_VM_PAGE(pte & PG_FRAME); |
| 1875 | |
| 1876 | if (pte & PG_M) { |
| 1877 | if (pmap_track_modified(ptepindex)) |
| 1878 | vm_page_dirty(p); |
| 1879 | } |
| 1880 | if (pte & PG_A) { |
| 1881 | vm_page_flag_set(p, PG_REFERENCED); |
| 1882 | } |
| 1883 | if (pte & PG_W) |
| 1884 | atomic_add_long(&pmap->pm_stats.wired_count, -1); |
| 1885 | if (pte & PG_G) |
| 1886 | cpu_invlpg((void *)va); |
| 1887 | } |
| 1888 | |
| 1889 | /* |
| 1890 | * Unwire the parent page table page. The wire_count cannot go below |
| 1891 | * 1 here because the parent page table page is itself still mapped. |
| 1892 | * |
| 1893 | * XXX remove the assertions later. |
| 1894 | */ |
| 1895 | KKASSERT(pv->pv_m == p); |
| 1896 | if (pvp && vm_page_unwire_quick(pvp->pv_m)) |
| 1897 | panic("pmap_remove_pv_pte: Insufficient wire_count"); |
| 1898 | |
| 1899 | if (gotpvp) |
| 1900 | pv_put(pvp); |
| 1901 | } |
| 1902 | |
| 1903 | static |
| 1904 | vm_page_t |
| 1905 | pmap_remove_pv_page(pv_entry_t pv) |
| 1906 | { |
| 1907 | vm_page_t m; |
| 1908 | |
| 1909 | m = pv->pv_m; |
| 1910 | KKASSERT(m); |
| 1911 | vm_page_spin_lock(m); |
| 1912 | pv->pv_m = NULL; |
| 1913 | TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); |
| 1914 | /* |
| 1915 | if (m->object) |
| 1916 | atomic_add_int(&m->object->agg_pv_list_count, -1); |
| 1917 | */ |
| 1918 | if (TAILQ_EMPTY(&m->md.pv_list)) |
| 1919 | vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE); |
| 1920 | vm_page_spin_unlock(m); |
| 1921 | return(m); |
| 1922 | } |
| 1923 | |
| 1924 | /* |
| 1925 | * Grow the number of kernel page table entries, if needed. |
| 1926 | * |
| 1927 | * This routine is always called to validate any address space |
| 1928 | * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address |
| 1929 | * space below KERNBASE. |
| 1930 | */ |
| 1931 | void |
| 1932 | pmap_growkernel(vm_offset_t kstart, vm_offset_t kend) |
| 1933 | { |
| 1934 | vm_paddr_t paddr; |
| 1935 | vm_offset_t ptppaddr; |
| 1936 | vm_page_t nkpg; |
| 1937 | pd_entry_t *pt, newpt; |
| 1938 | pdp_entry_t newpd; |
| 1939 | int update_kernel_vm_end; |
| 1940 | |
| 1941 | /* |
| 1942 | * bootstrap kernel_vm_end on first real VM use |
| 1943 | */ |
| 1944 | if (kernel_vm_end == 0) { |
| 1945 | kernel_vm_end = VM_MIN_KERNEL_ADDRESS; |
| 1946 | nkpt = 0; |
| 1947 | while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & PG_V) != 0) { |
| 1948 | kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & |
| 1949 | ~(PAGE_SIZE * NPTEPG - 1); |
| 1950 | nkpt++; |
| 1951 | if (kernel_vm_end - 1 >= kernel_map.max_offset) { |
| 1952 | kernel_vm_end = kernel_map.max_offset; |
| 1953 | break; |
| 1954 | } |
| 1955 | } |
| 1956 | } |
| 1957 | |
| 1958 | /* |
| 1959 | * Fill in the gaps. kernel_vm_end is only adjusted for ranges |
| 1960 | * below KERNBASE. Ranges above KERNBASE are kldloaded and we |
| 1961 | * do not want to force-fill 128G worth of page tables. |
| 1962 | */ |
| 1963 | if (kstart < KERNBASE) { |
| 1964 | if (kstart > kernel_vm_end) |
| 1965 | kstart = kernel_vm_end; |
| 1966 | KKASSERT(kend <= KERNBASE); |
| 1967 | update_kernel_vm_end = 1; |
| 1968 | } else { |
| 1969 | update_kernel_vm_end = 0; |
| 1970 | } |
| 1971 | |
| 1972 | kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG); |
| 1973 | kend = roundup2(kend, PAGE_SIZE * NPTEPG); |
| 1974 | |
| 1975 | if (kend - 1 >= kernel_map.max_offset) |
| 1976 | kend = kernel_map.max_offset; |
| 1977 | |
| 1978 | while (kstart < kend) { |
| 1979 | pt = pmap_pt(&kernel_pmap, kstart); |
| 1980 | if (pt == NULL) { |
| 1981 | /* We need a new PDP entry */ |
| 1982 | nkpg = vm_page_alloc(NULL, nkpt, |
| 1983 | VM_ALLOC_NORMAL | |
| 1984 | VM_ALLOC_SYSTEM | |
| 1985 | VM_ALLOC_INTERRUPT); |
| 1986 | if (nkpg == NULL) { |
| 1987 | panic("pmap_growkernel: no memory to grow " |
| 1988 | "kernel"); |
| 1989 | } |
| 1990 | paddr = VM_PAGE_TO_PHYS(nkpg); |
| 1991 | if ((nkpg->flags & PG_ZERO) == 0) |
| 1992 | pmap_zero_page(paddr); |
| 1993 | vm_page_flag_clear(nkpg, PG_ZERO); |
| 1994 | newpd = (pdp_entry_t) |
| 1995 | (paddr | PG_V | PG_RW | PG_A | PG_M); |
| 1996 | *pmap_pd(&kernel_pmap, kstart) = newpd; |
| 1997 | nkpt++; |
| 1998 | continue; /* try again */ |
| 1999 | } |
| 2000 | if ((*pt & PG_V) != 0) { |
| 2001 | kstart = (kstart + PAGE_SIZE * NPTEPG) & |
| 2002 | ~(PAGE_SIZE * NPTEPG - 1); |
| 2003 | if (kstart - 1 >= kernel_map.max_offset) { |
| 2004 | kstart = kernel_map.max_offset; |
| 2005 | break; |
| 2006 | } |
| 2007 | continue; |
| 2008 | } |
| 2009 | |
| 2010 | /* |
| 2011 | * This index is bogus, but out of the way |
| 2012 | */ |
| 2013 | nkpg = vm_page_alloc(NULL, nkpt, |
| 2014 | VM_ALLOC_NORMAL | |
| 2015 | VM_ALLOC_SYSTEM | |
| 2016 | VM_ALLOC_INTERRUPT); |
| 2017 | if (nkpg == NULL) |
| 2018 | panic("pmap_growkernel: no memory to grow kernel"); |
| 2019 | |
| 2020 | vm_page_wire(nkpg); |
| 2021 | ptppaddr = VM_PAGE_TO_PHYS(nkpg); |
| 2022 | pmap_zero_page(ptppaddr); |
| 2023 | vm_page_flag_clear(nkpg, PG_ZERO); |
| 2024 | newpt = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); |
| 2025 | *pmap_pt(&kernel_pmap, kstart) = newpt; |
| 2026 | nkpt++; |
| 2027 | |
| 2028 | kstart = (kstart + PAGE_SIZE * NPTEPG) & |
| 2029 | ~(PAGE_SIZE * NPTEPG - 1); |
| 2030 | |
| 2031 | if (kstart - 1 >= kernel_map.max_offset) { |
| 2032 | kstart = kernel_map.max_offset; |
| 2033 | break; |
| 2034 | } |
| 2035 | } |
| 2036 | |
| 2037 | /* |
| 2038 | * Only update kernel_vm_end for areas below KERNBASE. |
| 2039 | */ |
| 2040 | if (update_kernel_vm_end && kernel_vm_end < kstart) |
| 2041 | kernel_vm_end = kstart; |
| 2042 | } |
| 2043 | |
| 2044 | /* |
| 2045 | * Retire the given physical map from service. |
| 2046 | * Should only be called if the map contains |
| 2047 | * no valid mappings. |
| 2048 | */ |
| 2049 | void |
| 2050 | pmap_destroy(pmap_t pmap) |
| 2051 | { |
| 2052 | int count; |
| 2053 | |
| 2054 | if (pmap == NULL) |
| 2055 | return; |
| 2056 | |
| 2057 | lwkt_gettoken(&pmap->pm_token); |
| 2058 | count = --pmap->pm_count; |
| 2059 | if (count == 0) { |
| 2060 | pmap_release(pmap); /* eats pm_token */ |
| 2061 | panic("destroying a pmap is not yet implemented"); |
| 2062 | } |
| 2063 | lwkt_reltoken(&pmap->pm_token); |
| 2064 | } |
| 2065 | |
| 2066 | /* |
| 2067 | * Add a reference to the specified pmap. |
| 2068 | */ |
| 2069 | void |
| 2070 | pmap_reference(pmap_t pmap) |
| 2071 | { |
| 2072 | if (pmap != NULL) { |
| 2073 | lwkt_gettoken(&pmap->pm_token); |
| 2074 | pmap->pm_count++; |
| 2075 | lwkt_reltoken(&pmap->pm_token); |
| 2076 | } |
| 2077 | } |
| 2078 | |
| 2079 | /*************************************************** |
| 2080 | * page management routines. |
| 2081 | ***************************************************/ |
| 2082 | |
| 2083 | /* |
| 2084 | * Hold a pv without locking it |
| 2085 | */ |
| 2086 | static void |
| 2087 | pv_hold(pv_entry_t pv) |
| 2088 | { |
| 2089 | u_int count; |
| 2090 | |
| 2091 | if (atomic_cmpset_int(&pv->pv_hold, 0, 1)) |
| 2092 | return; |
| 2093 | |
| 2094 | for (;;) { |
| 2095 | count = pv->pv_hold; |
| 2096 | cpu_ccfence(); |
| 2097 | if (atomic_cmpset_int(&pv->pv_hold, count, count + 1)) |
| 2098 | return; |
| 2099 | /* retry */ |
| 2100 | } |
| 2101 | } |
| 2102 | |
| 2103 | /* |
| 2104 | * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv |
| 2105 | * was successfully locked, FALSE if it wasn't. The caller must dispose of |
| 2106 | * the pv properly. |
| 2107 | * |
| 2108 | * Either the pmap->pm_spin or the related vm_page_spin (if traversing a |
| 2109 | * pv list via its page) must be held by the caller. |
| 2110 | */ |
| 2111 | static int |
| 2112 | _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL) |
| 2113 | { |
| 2114 | u_int count; |
| 2115 | |
| 2116 | if (atomic_cmpset_int(&pv->pv_hold, 0, PV_HOLD_LOCKED | 1)) { |
| 2117 | #ifdef PMAP_DEBUG |
| 2118 | pv->pv_func = func; |
| 2119 | pv->pv_line = lineno; |
| 2120 | #endif |
| 2121 | return TRUE; |
| 2122 | } |
| 2123 | |
| 2124 | for (;;) { |
| 2125 | count = pv->pv_hold; |
| 2126 | cpu_ccfence(); |
| 2127 | if ((count & PV_HOLD_LOCKED) == 0) { |
| 2128 | if (atomic_cmpset_int(&pv->pv_hold, count, |
| 2129 | (count + 1) | PV_HOLD_LOCKED)) { |
| 2130 | #ifdef PMAP_DEBUG |
| 2131 | pv->pv_func = func; |
| 2132 | pv->pv_line = lineno; |
| 2133 | #endif |
| 2134 | return TRUE; |
| 2135 | } |
| 2136 | } else { |
| 2137 | if (atomic_cmpset_int(&pv->pv_hold, count, count + 1)) |
| 2138 | return FALSE; |
| 2139 | } |
| 2140 | /* retry */ |
| 2141 | } |
| 2142 | } |
| 2143 | |
| 2144 | /* |
| 2145 | * Drop a previously held pv_entry which could not be locked, allowing its |
| 2146 | * destruction. |
| 2147 | * |
| 2148 | * Must not be called with a spinlock held as we might zfree() the pv if it |
| 2149 | * is no longer associated with a pmap and this was the last hold count. |
| 2150 | */ |
| 2151 | static void |
| 2152 | pv_drop(pv_entry_t pv) |
| 2153 | { |
| 2154 | u_int count; |
| 2155 | |
| 2156 | if (atomic_cmpset_int(&pv->pv_hold, 1, 0)) { |
| 2157 | if (pv->pv_pmap == NULL) |
| 2158 | zfree(pvzone, pv); |
| 2159 | return; |
| 2160 | } |
| 2161 | |
| 2162 | for (;;) { |
| 2163 | count = pv->pv_hold; |
| 2164 | cpu_ccfence(); |
| 2165 | KKASSERT((count & PV_HOLD_MASK) > 0); |
| 2166 | KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) != |
| 2167 | (PV_HOLD_LOCKED | 1)); |
| 2168 | if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) { |
| 2169 | if (count == 1 && pv->pv_pmap == NULL) |
| 2170 | zfree(pvzone, pv); |
| 2171 | return; |
| 2172 | } |
| 2173 | /* retry */ |
| 2174 | } |
| 2175 | } |
| 2176 | |
| 2177 | /* |
| 2178 | * Find or allocate the requested PV entry, returning a locked pv |
| 2179 | */ |
| 2180 | static |
| 2181 | pv_entry_t |
| 2182 | _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL) |
| 2183 | { |
| 2184 | pv_entry_t pv; |
| 2185 | pv_entry_t pnew = NULL; |
| 2186 | |
| 2187 | spin_lock(&pmap->pm_spin); |
| 2188 | for (;;) { |
| 2189 | if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) { |
| 2190 | pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, |
| 2191 | pindex); |
| 2192 | } |
| 2193 | if (pv == NULL) { |
| 2194 | if (pnew == NULL) { |
| 2195 | spin_unlock(&pmap->pm_spin); |
| 2196 | pnew = zalloc(pvzone); |
| 2197 | spin_lock(&pmap->pm_spin); |
| 2198 | continue; |
| 2199 | } |
| 2200 | pnew->pv_pmap = pmap; |
| 2201 | pnew->pv_pindex = pindex; |
| 2202 | pnew->pv_hold = PV_HOLD_LOCKED | 1; |
| 2203 | #ifdef PMAP_DEBUG |
| 2204 | pnew->pv_func = func; |
| 2205 | pnew->pv_line = lineno; |
| 2206 | #endif |
| 2207 | pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew); |
| 2208 | atomic_add_long(&pmap->pm_stats.resident_count, 1); |
| 2209 | spin_unlock(&pmap->pm_spin); |
| 2210 | *isnew = 1; |
| 2211 | return(pnew); |
| 2212 | } |
| 2213 | if (pnew) { |
| 2214 | spin_unlock(&pmap->pm_spin); |
| 2215 | zfree(pvzone, pnew); |
| 2216 | pnew = NULL; |
| 2217 | spin_lock(&pmap->pm_spin); |
| 2218 | continue; |
| 2219 | } |
| 2220 | if (_pv_hold_try(pv PMAP_DEBUG_COPY)) { |
| 2221 | spin_unlock(&pmap->pm_spin); |
| 2222 | *isnew = 0; |
| 2223 | return(pv); |
| 2224 | } |
| 2225 | spin_unlock(&pmap->pm_spin); |
| 2226 | _pv_lock(pv PMAP_DEBUG_COPY); |
| 2227 | if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) { |
| 2228 | *isnew = 0; |
| 2229 | return(pv); |
| 2230 | } |
| 2231 | pv_put(pv); |
| 2232 | spin_lock(&pmap->pm_spin); |
| 2233 | } |
| 2234 | |
| 2235 | |
| 2236 | } |
| 2237 | |
| 2238 | /* |
| 2239 | * Find the requested PV entry, returning a locked+held pv or NULL |
| 2240 | */ |
| 2241 | static |
| 2242 | pv_entry_t |
| 2243 | _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL) |
| 2244 | { |
| 2245 | pv_entry_t pv; |
| 2246 | |
| 2247 | spin_lock(&pmap->pm_spin); |
| 2248 | for (;;) { |
| 2249 | /* |
| 2250 | * Shortcut cache |
| 2251 | */ |
| 2252 | if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) { |
| 2253 | pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, |
| 2254 | pindex); |
| 2255 | } |
| 2256 | if (pv == NULL) { |
| 2257 | spin_unlock(&pmap->pm_spin); |
| 2258 | return NULL; |
| 2259 | } |
| 2260 | if (_pv_hold_try(pv PMAP_DEBUG_COPY)) { |
| 2261 | pv_cache(pv, pindex); |
| 2262 | spin_unlock(&pmap->pm_spin); |
| 2263 | return(pv); |
| 2264 | } |
| 2265 | spin_unlock(&pmap->pm_spin); |
| 2266 | _pv_lock(pv PMAP_DEBUG_COPY); |
| 2267 | if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) |
| 2268 | return(pv); |
| 2269 | pv_put(pv); |
| 2270 | spin_lock(&pmap->pm_spin); |
| 2271 | } |
| 2272 | } |
| 2273 | |
| 2274 | /* |
| 2275 | * Lookup, hold, and attempt to lock (pmap,pindex). |
| 2276 | * |
| 2277 | * If the entry does not exist NULL is returned and *errorp is set to 0 |
| 2278 | * |
| 2279 | * If the entry exists and could be successfully locked it is returned and |
| 2280 | * errorp is set to 0. |
| 2281 | * |
| 2282 | * If the entry exists but could NOT be successfully locked it is returned |
| 2283 | * held and *errorp is set to 1. |
| 2284 | */ |
| 2285 | static |
| 2286 | pv_entry_t |
| 2287 | pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp) |
| 2288 | { |
| 2289 | pv_entry_t pv; |
| 2290 | |
| 2291 | spin_lock(&pmap->pm_spin); |
| 2292 | if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) |
| 2293 | pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex); |
| 2294 | if (pv == NULL) { |
| 2295 | spin_unlock(&pmap->pm_spin); |
| 2296 | *errorp = 0; |
| 2297 | return NULL; |
| 2298 | } |
| 2299 | if (pv_hold_try(pv)) { |
| 2300 | pv_cache(pv, pindex); |
| 2301 | spin_unlock(&pmap->pm_spin); |
| 2302 | *errorp = 0; |
| 2303 | return(pv); /* lock succeeded */ |
| 2304 | } |
| 2305 | spin_unlock(&pmap->pm_spin); |
| 2306 | *errorp = 1; |
| 2307 | return (pv); /* lock failed */ |
| 2308 | } |
| 2309 | |
| 2310 | /* |
| 2311 | * Find the requested PV entry, returning a held pv or NULL |
| 2312 | */ |
| 2313 | static |
| 2314 | pv_entry_t |
| 2315 | pv_find(pmap_t pmap, vm_pindex_t pindex) |
| 2316 | { |
| 2317 | pv_entry_t pv; |
| 2318 | |
| 2319 | spin_lock(&pmap->pm_spin); |
| 2320 | |
| 2321 | if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) |
| 2322 | pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex); |
| 2323 | if (pv == NULL) { |
| 2324 | spin_unlock(&pmap->pm_spin); |
| 2325 | return NULL; |
| 2326 | } |
| 2327 | pv_hold(pv); |
| 2328 | pv_cache(pv, pindex); |
| 2329 | spin_unlock(&pmap->pm_spin); |
| 2330 | return(pv); |
| 2331 | } |
| 2332 | |
| 2333 | /* |
| 2334 | * Lock a held pv, keeping the hold count |
| 2335 | */ |
| 2336 | static |
| 2337 | void |
| 2338 | _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL) |
| 2339 | { |
| 2340 | u_int count; |
| 2341 | |
| 2342 | for (;;) { |
| 2343 | count = pv->pv_hold; |
| 2344 | cpu_ccfence(); |
| 2345 | if ((count & PV_HOLD_LOCKED) == 0) { |
| 2346 | if (atomic_cmpset_int(&pv->pv_hold, count, |
| 2347 | count | PV_HOLD_LOCKED)) { |
| 2348 | #ifdef PMAP_DEBUG |
| 2349 | pv->pv_func = func; |
| 2350 | pv->pv_line = lineno; |
| 2351 | #endif |
| 2352 | return; |
| 2353 | } |
| 2354 | continue; |
| 2355 | } |
| 2356 | tsleep_interlock(pv, 0); |
| 2357 | if (atomic_cmpset_int(&pv->pv_hold, count, |
| 2358 | count | PV_HOLD_WAITING)) { |
| 2359 | #ifdef PMAP_DEBUG |
| 2360 | kprintf("pv waiting on %s:%d\n", |
| 2361 | pv->pv_func, pv->pv_line); |
| 2362 | #endif |
| 2363 | tsleep(pv, PINTERLOCKED, "pvwait", hz); |
| 2364 | } |
| 2365 | /* retry */ |
| 2366 | } |
| 2367 | } |
| 2368 | |
| 2369 | /* |
| 2370 | * Unlock a held and locked pv, keeping the hold count. |
| 2371 | */ |
| 2372 | static |
| 2373 | void |
| 2374 | pv_unlock(pv_entry_t pv) |
| 2375 | { |
| 2376 | u_int count; |
| 2377 | |
| 2378 | if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 1)) |
| 2379 | return; |
| 2380 | |
| 2381 | for (;;) { |
| 2382 | count = pv->pv_hold; |
| 2383 | cpu_ccfence(); |
| 2384 | KKASSERT((count & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >= |
| 2385 | (PV_HOLD_LOCKED | 1)); |
| 2386 | if (atomic_cmpset_int(&pv->pv_hold, count, |
| 2387 | count & |
| 2388 | ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) { |
| 2389 | if (count & PV_HOLD_WAITING) |
| 2390 | wakeup(pv); |
| 2391 | break; |
| 2392 | } |
| 2393 | } |
| 2394 | } |
| 2395 | |
| 2396 | /* |
| 2397 | * Unlock and drop a pv. If the pv is no longer associated with a pmap |
| 2398 | * and the hold count drops to zero we will free it. |
| 2399 | * |
| 2400 | * Caller should not hold any spin locks. We are protected from hold races |
| 2401 | * by virtue of holds only occuring only with a pmap_spin or vm_page_spin |
| 2402 | * lock held. A pv cannot be located otherwise. |
| 2403 | */ |
| 2404 | static |
| 2405 | void |
| 2406 | pv_put(pv_entry_t pv) |
| 2407 | { |
| 2408 | if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 0)) { |
| 2409 | if (pv->pv_pmap == NULL) |
| 2410 | zfree(pvzone, pv); |
| 2411 | return; |
| 2412 | } |
| 2413 | pv_unlock(pv); |
| 2414 | pv_drop(pv); |
| 2415 | } |
| 2416 | |
| 2417 | /* |
| 2418 | * Unlock, drop, and free a pv, destroying it. The pv is removed from its |
| 2419 | * pmap. Any pte operations must have already been completed. |
| 2420 | */ |
| 2421 | static |
| 2422 | void |
| 2423 | pv_free(pv_entry_t pv) |
| 2424 | { |
| 2425 | pmap_t pmap; |
| 2426 | |
| 2427 | KKASSERT(pv->pv_m == NULL); |
| 2428 | if ((pmap = pv->pv_pmap) != NULL) { |
| 2429 | spin_lock(&pmap->pm_spin); |
| 2430 | pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv); |
| 2431 | if (pmap->pm_pvhint == pv) |
| 2432 | pmap->pm_pvhint = NULL; |
| 2433 | atomic_add_long(&pmap->pm_stats.resident_count, -1); |
| 2434 | pv->pv_pmap = NULL; |
| 2435 | pv->pv_pindex = 0; |
| 2436 | spin_unlock(&pmap->pm_spin); |
| 2437 | } |
| 2438 | pv_put(pv); |
| 2439 | } |
| 2440 | |
| 2441 | /* |
| 2442 | * This routine is very drastic, but can save the system |
| 2443 | * in a pinch. |
| 2444 | */ |
| 2445 | void |
| 2446 | pmap_collect(void) |
| 2447 | { |
| 2448 | int i; |
| 2449 | vm_page_t m; |
| 2450 | static int warningdone=0; |
| 2451 | |
| 2452 | if (pmap_pagedaemon_waken == 0) |
| 2453 | return; |
| 2454 | pmap_pagedaemon_waken = 0; |
| 2455 | if (warningdone < 5) { |
| 2456 | kprintf("pmap_collect: collecting pv entries -- " |
| 2457 | "suggest increasing PMAP_SHPGPERPROC\n"); |
| 2458 | warningdone++; |
| 2459 | } |
| 2460 | |
| 2461 | for (i = 0; i < vm_page_array_size; i++) { |
| 2462 | m = &vm_page_array[i]; |
| 2463 | if (m->wire_count || m->hold_count) |
| 2464 | continue; |
| 2465 | if (vm_page_busy_try(m, TRUE) == 0) { |
| 2466 | if (m->wire_count == 0 && m->hold_count == 0) { |
| 2467 | pmap_remove_all(m); |
| 2468 | } |
| 2469 | vm_page_wakeup(m); |
| 2470 | } |
| 2471 | } |
| 2472 | } |
| 2473 | |
| 2474 | /* |
| 2475 | * Scan the pmap for active page table entries and issue a callback. |
| 2476 | * The callback must dispose of pte_pv. |
| 2477 | * |
| 2478 | * NOTE: Unmanaged page table entries will not have a pte_pv |
| 2479 | * |
| 2480 | * NOTE: Kernel page table entries will not have a pt_pv. That is, wiring |
| 2481 | * counts are not tracked in kernel page table pages. |
| 2482 | * |
| 2483 | * It is assumed that the start and end are properly rounded to the page size. |
| 2484 | */ |
| 2485 | static void |
| 2486 | pmap_scan(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva, |
| 2487 | void (*func)(pmap_t, struct pmap_inval_info *, |
| 2488 | pv_entry_t, pv_entry_t, vm_offset_t, |
| 2489 | pt_entry_t *, void *), |
| 2490 | void *arg) |
| 2491 | { |
| 2492 | pv_entry_t pdp_pv; /* A page directory page PV */ |
| 2493 | pv_entry_t pd_pv; /* A page directory PV */ |
| 2494 | pv_entry_t pt_pv; /* A page table PV */ |
| 2495 | pv_entry_t pte_pv; /* A page table entry PV */ |
| 2496 | pt_entry_t *ptep; |
| 2497 | vm_offset_t va_next; |
| 2498 | struct pmap_inval_info info; |
| 2499 | int error; |
| 2500 | |
| 2501 | if (pmap == NULL) |
| 2502 | return; |
| 2503 | |
| 2504 | /* |
| 2505 | * Hold the token for stability; if the pmap is empty we have nothing |
| 2506 | * to do. |
| 2507 | */ |
| 2508 | lwkt_gettoken(&pmap->pm_token); |
| 2509 | #if 0 |
| 2510 | if (pmap->pm_stats.resident_count == 0) { |
| 2511 | lwkt_reltoken(&pmap->pm_token); |
| 2512 | return; |
| 2513 | } |
| 2514 | #endif |
| 2515 | |
| 2516 | pmap_inval_init(&info); |
| 2517 | |
| 2518 | /* |
| 2519 | * Special handling for removing one page, which is a very common |
| 2520 | * operation (it is?). |
| 2521 | * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4 |
| 2522 | */ |
| 2523 | if (sva + PAGE_SIZE == eva) { |
| 2524 | if (sva >= VM_MAX_USER_ADDRESS) { |
| 2525 | /* |
| 2526 | * Kernel mappings do not track wire counts on |
| 2527 | * page table pages. |
| 2528 | */ |
| 2529 | pt_pv = NULL; |
| 2530 | pte_pv = pv_get(pmap, pmap_pte_pindex(sva)); |
| 2531 | ptep = vtopte(sva); |
| 2532 | } else { |
| 2533 | /* |
| 2534 | * User mappings may or may not have a pte_pv but |
| 2535 | * will always have a pt_pv if the page is present. |
| 2536 | */ |
| 2537 | pte_pv = pv_get(pmap, pmap_pte_pindex(sva)); |
| 2538 | pt_pv = pv_get(pmap, pmap_pt_pindex(sva)); |
| 2539 | if (pt_pv == NULL) { |
| 2540 | KKASSERT(pte_pv == NULL); |
| 2541 | goto fast_skip; |
| 2542 | } |
| 2543 | ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva)); |
| 2544 | } |
| 2545 | if (*ptep == 0) { |
| 2546 | /* |
| 2547 | * Unlike the pv_find() case below we actually |
| 2548 | * acquired a locked pv in this case so any |
| 2549 | * race should have been resolved. It is expected |
| 2550 | * to not exist. |
| 2551 | */ |
| 2552 | KKASSERT(pte_pv == NULL); |
| 2553 | } else if (pte_pv) { |
| 2554 | KASSERT((*ptep & (PG_MANAGED|PG_V)) == (PG_MANAGED| |
| 2555 | PG_V), |
| 2556 | ("bad *ptep %016lx sva %016lx pte_pv %p", |
| 2557 | *ptep, sva, pte_pv)); |
| 2558 | func(pmap, &info, pte_pv, pt_pv, sva, ptep, arg); |
| 2559 | } else { |
| 2560 | KASSERT((*ptep & (PG_MANAGED|PG_V)) == PG_V, |
| 2561 | ("bad *ptep %016lx sva %016lx pte_pv NULL", |
| 2562 | *ptep, sva)); |
| 2563 | func(pmap, &info, pte_pv, pt_pv, sva, ptep, arg); |
| 2564 | } |
| 2565 | if (pt_pv) |
| 2566 | pv_put(pt_pv); |
| 2567 | fast_skip: |
| 2568 | pmap_inval_done(&info); |
| 2569 | lwkt_reltoken(&pmap->pm_token); |
| 2570 | return; |
| 2571 | } |
| 2572 | |
| 2573 | /* |
| 2574 | * NOTE: kernel mappings do not track page table pages, only |
| 2575 | * terminal pages. |
| 2576 | * |
| 2577 | * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4. |
| 2578 | * However, for the scan to be efficient we try to |
| 2579 | * cache items top-down. |
| 2580 | */ |
| 2581 | pdp_pv = NULL; |
| 2582 | pd_pv = NULL; |
| 2583 | pt_pv = NULL; |
| 2584 | |
| 2585 | for (; sva < eva; sva = va_next) { |
| 2586 | lwkt_yield(); |
| 2587 | if (sva >= VM_MAX_USER_ADDRESS) { |
| 2588 | if (pt_pv) { |
| 2589 | pv_put(pt_pv); |
| 2590 | pt_pv = NULL; |
| 2591 | } |
| 2592 | goto kernel_skip; |
| 2593 | } |
| 2594 | |
| 2595 | /* |
| 2596 | * PDP cache |
| 2597 | */ |
| 2598 | if (pdp_pv == NULL) { |
| 2599 | pdp_pv = pv_get(pmap, pmap_pdp_pindex(sva)); |
| 2600 | } else if (pdp_pv->pv_pindex != pmap_pdp_pindex(sva)) { |
| 2601 | pv_put(pdp_pv); |
| 2602 | pdp_pv = pv_get(pmap, pmap_pdp_pindex(sva)); |
| 2603 | } |
| 2604 | if (pdp_pv == NULL) { |
| 2605 | va_next = (sva + NBPML4) & ~PML4MASK; |
| 2606 | if (va_next < sva) |
| 2607 | va_next = eva; |
| 2608 | continue; |
| 2609 | } |
| 2610 | |
| 2611 | /* |
| 2612 | * PD cache |
| 2613 | */ |
| 2614 | if (pd_pv == NULL) { |
| 2615 | if (pdp_pv) { |
| 2616 | pv_put(pdp_pv); |
| 2617 | pdp_pv = NULL; |
| 2618 | } |
| 2619 | pd_pv = pv_get(pmap, pmap_pd_pindex(sva)); |
| 2620 | } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) { |
| 2621 | pv_put(pd_pv); |
| 2622 | if (pdp_pv) { |
| 2623 | pv_put(pdp_pv); |
| 2624 | pdp_pv = NULL; |
| 2625 | } |
| 2626 | pd_pv = pv_get(pmap, pmap_pd_pindex(sva)); |
| 2627 | } |
| 2628 | if (pd_pv == NULL) { |
| 2629 | va_next = (sva + NBPDP) & ~PDPMASK; |
| 2630 | if (va_next < sva) |
| 2631 | va_next = eva; |
| 2632 | continue; |
| 2633 | } |
| 2634 | |
| 2635 | /* |
| 2636 | * PT cache |
| 2637 | */ |
| 2638 | if (pt_pv == NULL) { |
| 2639 | if (pdp_pv) { |
| 2640 | pv_put(pdp_pv); |
| 2641 | pdp_pv = NULL; |
| 2642 | } |
| 2643 | if (pd_pv) { |
| 2644 | pv_put(pd_pv); |
| 2645 | pd_pv = NULL; |
| 2646 | } |
| 2647 | pt_pv = pv_get(pmap, pmap_pt_pindex(sva)); |
| 2648 | } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) { |
| 2649 | if (pdp_pv) { |
| 2650 | pv_put(pdp_pv); |
| 2651 | pdp_pv = NULL; |
| 2652 | } |
| 2653 | if (pd_pv) { |
| 2654 | pv_put(pd_pv); |
| 2655 | pd_pv = NULL; |
| 2656 | } |
| 2657 | pv_put(pt_pv); |
| 2658 | pt_pv = pv_get(pmap, pmap_pt_pindex(sva)); |
| 2659 | } |
| 2660 | |
| 2661 | /* |
| 2662 | * We will scan or skip a page table page so adjust va_next |
| 2663 | * either way. |
| 2664 | */ |
| 2665 | if (pt_pv == NULL) { |
| 2666 | va_next = (sva + NBPDR) & ~PDRMASK; |
| 2667 | if (va_next < sva) |
| 2668 | va_next = eva; |
| 2669 | continue; |
| 2670 | } |
| 2671 | |
| 2672 | /* |
| 2673 | * From this point in the loop testing pt_pv for non-NULL |
| 2674 | * means we are in UVM, else if it is NULL we are in KVM. |
| 2675 | */ |
| 2676 | kernel_skip: |
| 2677 | va_next = (sva + NBPDR) & ~PDRMASK; |
| 2678 | if (va_next < sva) |
| 2679 | va_next = eva; |
| 2680 | |
| 2681 | /* |
| 2682 | * Limit our scan to either the end of the va represented |
| 2683 | * by the current page table page, or to the end of the |
| 2684 | * range being removed. |
| 2685 | * |
| 2686 | * Scan the page table for pages. Some pages may not be |
| 2687 | * managed (might not have a pv_entry). |
| 2688 | * |
| 2689 | * There is no page table management for kernel pages so |
| 2690 | * pt_pv will be NULL in that case, but otherwise pt_pv |
| 2691 | * is non-NULL, locked, and referenced. |
| 2692 | */ |
| 2693 | if (va_next > eva) |
| 2694 | va_next = eva; |
| 2695 | |
| 2696 | /* |
| 2697 | * At this point a non-NULL pt_pv means a UVA, and a NULL |
| 2698 | * pt_pv means a KVA. |
| 2699 | */ |
| 2700 | if (pt_pv) |
| 2701 | ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva)); |
| 2702 | else |
| 2703 | ptep = vtopte(sva); |
| 2704 | |
| 2705 | while (sva < va_next) { |
| 2706 | /* |
| 2707 | * Acquire the related pte_pv, if any. If *ptep == 0 |
| 2708 | * the related pte_pv should not exist, but if *ptep |
| 2709 | * is not zero the pte_pv may or may not exist (e.g. |
| 2710 | * will not exist for an unmanaged page). |
| 2711 | * |
| 2712 | * However a multitude of races are possible here. |
| 2713 | * |
| 2714 | * In addition, the (pt_pv, pte_pv) lock order is |
| 2715 | * backwards, so we have to be careful in aquiring |
| 2716 | * a properly locked pte_pv. |
| 2717 | */ |
| 2718 | lwkt_yield(); |
| 2719 | if (pt_pv) { |
| 2720 | pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva), |
| 2721 | &error); |
| 2722 | if (error) { |
| 2723 | if (pdp_pv) { |
| 2724 | pv_put(pdp_pv); |
| 2725 | pdp_pv = NULL; |
| 2726 | } |
| 2727 | if (pd_pv) { |
| 2728 | pv_put(pd_pv); |
| 2729 | pd_pv = NULL; |
| 2730 | } |
| 2731 | pv_put(pt_pv); /* must be non-NULL */ |
| 2732 | pt_pv = NULL; |
| 2733 | pv_lock(pte_pv); /* safe to block now */ |
| 2734 | pv_put(pte_pv); |
| 2735 | pte_pv = NULL; |
| 2736 | pt_pv = pv_get(pmap, |
| 2737 | pmap_pt_pindex(sva)); |
| 2738 | continue; |
| 2739 | } |
| 2740 | } else { |
| 2741 | pte_pv = pv_get(pmap, pmap_pte_pindex(sva)); |
| 2742 | } |
| 2743 | |
| 2744 | /* |
| 2745 | * Ok, if *ptep == 0 we had better NOT have a pte_pv. |
| 2746 | */ |
| 2747 | if (*ptep == 0) { |
| 2748 | if (pte_pv) { |
| 2749 | kprintf("Unexpected non-NULL pte_pv " |
| 2750 | "%p pt_pv %p *ptep = %016lx\n", |
| 2751 | pte_pv, pt_pv, *ptep); |
| 2752 | panic("Unexpected non-NULL pte_pv"); |
| 2753 | } |
| 2754 | sva += PAGE_SIZE; |
| 2755 | ++ptep; |
| 2756 | continue; |
| 2757 | } |
| 2758 | |
| 2759 | /* |
| 2760 | * Ready for the callback. The locked pte_pv (if any) |
| 2761 | * is consumed by the callback. pte_pv will exist if |
| 2762 | * the page is managed, and will not exist if it |
| 2763 | * isn't. |
| 2764 | */ |
| 2765 | if (pte_pv) { |
| 2766 | KASSERT((*ptep & (PG_MANAGED|PG_V)) == |
| 2767 | (PG_MANAGED|PG_V), |
| 2768 | ("bad *ptep %016lx sva %016lx " |
| 2769 | "pte_pv %p", |
| 2770 | *ptep, sva, pte_pv)); |
| 2771 | func(pmap, &info, pte_pv, pt_pv, sva, |
| 2772 | ptep, arg); |
| 2773 | } else { |
| 2774 | KASSERT((*ptep & (PG_MANAGED|PG_V)) == |
| 2775 | PG_V, |
| 2776 | ("bad *ptep %016lx sva %016lx " |
| 2777 | "pte_pv NULL", |
| 2778 | *ptep, sva)); |
| 2779 | func(pmap, &info, pte_pv, pt_pv, sva, |
| 2780 | ptep, arg); |
| 2781 | } |
| 2782 | pte_pv = NULL; |
| 2783 | sva += PAGE_SIZE; |
| 2784 | ++ptep; |
| 2785 | } |
| 2786 | } |
| 2787 | if (pdp_pv) { |
| 2788 | pv_put(pdp_pv); |
| 2789 | pdp_pv = NULL; |
| 2790 | } |
| 2791 | if (pd_pv) { |
| 2792 | pv_put(pd_pv); |
| 2793 | pd_pv = NULL; |
| 2794 | } |
| 2795 | if (pt_pv) { |
| 2796 | pv_put(pt_pv); |
| 2797 | pt_pv = NULL; |
| 2798 | } |
| 2799 | pmap_inval_done(&info); |
| 2800 | lwkt_reltoken(&pmap->pm_token); |
| 2801 | } |
| 2802 | |
| 2803 | void |
| 2804 | pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva) |
| 2805 | { |
| 2806 | pmap_scan(pmap, sva, eva, pmap_remove_callback, NULL); |
| 2807 | } |
| 2808 | |
| 2809 | static void |
| 2810 | pmap_remove_callback(pmap_t pmap, struct pmap_inval_info *info, |
| 2811 | pv_entry_t pte_pv, pv_entry_t pt_pv, vm_offset_t va, |
| 2812 | pt_entry_t *ptep, void *arg __unused) |
| 2813 | { |
| 2814 | pt_entry_t pte; |
| 2815 | |
| 2816 | if (pte_pv) { |
| 2817 | /* |
| 2818 | * This will also drop pt_pv's wire_count. Note that |
| 2819 | * terminal pages are not wired based on mmu presence. |
| 2820 | */ |
| 2821 | pmap_remove_pv_pte(pte_pv, pt_pv, info); |
| 2822 | pmap_remove_pv_page(pte_pv); |
| 2823 | pv_free(pte_pv); |
| 2824 | } else { |
| 2825 | /* |
| 2826 | * pt_pv's wire_count is still bumped by unmanaged pages |
| 2827 | * so we must decrement it manually. |
| 2828 | */ |
| 2829 | pmap_inval_interlock(info, pmap, va); |
| 2830 | pte = pte_load_clear(ptep); |
| 2831 | pmap_inval_deinterlock(info, pmap); |
| 2832 | if (pte & PG_W) |
| 2833 | atomic_add_long(&pmap->pm_stats.wired_count, -1); |
| 2834 | atomic_add_long(&pmap->pm_stats.resident_count, -1); |
| 2835 | if (pt_pv && vm_page_unwire_quick(pt_pv->pv_m)) |
| 2836 | panic("pmap_remove: insufficient wirecount"); |
| 2837 | } |
| 2838 | } |
| 2839 | |
| 2840 | /* |
| 2841 | * Removes this physical page from all physical maps in which it resides. |
| 2842 | * Reflects back modify bits to the pager. |
| 2843 | * |
| 2844 | * This routine may not be called from an interrupt. |
| 2845 | */ |
| 2846 | static |
| 2847 | void |
| 2848 | pmap_remove_all(vm_page_t m) |
| 2849 | { |
| 2850 | struct pmap_inval_info info; |
| 2851 | pv_entry_t pv; |
| 2852 | |
| 2853 | if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) |
| 2854 | return; |
| 2855 | |
| 2856 | pmap_inval_init(&info); |
| 2857 | vm_page_spin_lock(m); |
| 2858 | while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { |
| 2859 | KKASSERT(pv->pv_m == m); |
| 2860 | if (pv_hold_try(pv)) { |
| 2861 | vm_page_spin_unlock(m); |
| 2862 | } else { |
| 2863 | vm_page_spin_unlock(m); |
| 2864 | pv_lock(pv); |
| 2865 | if (pv->pv_m != m) { |
| 2866 | pv_put(pv); |
| 2867 | vm_page_spin_lock(m); |
| 2868 | continue; |
| 2869 | } |
| 2870 | } |
| 2871 | /* |
| 2872 | * Holding no spinlocks, pv is locked. |
| 2873 | */ |
| 2874 | pmap_remove_pv_pte(pv, NULL, &info); |
| 2875 | pmap_remove_pv_page(pv); |
| 2876 | pv_free(pv); |
| 2877 | vm_page_spin_lock(m); |
| 2878 | } |
| 2879 | KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0); |
| 2880 | vm_page_spin_unlock(m); |
| 2881 | pmap_inval_done(&info); |
| 2882 | } |
| 2883 | |
| 2884 | /* |
| 2885 | * pmap_protect: |
| 2886 | * |
| 2887 | * Set the physical protection on the specified range of this map |
| 2888 | * as requested. |
| 2889 | * |
| 2890 | * This function may not be called from an interrupt if the map is |
| 2891 | * not the kernel_pmap. |
| 2892 | */ |
| 2893 | void |
| 2894 | pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) |
| 2895 | { |
| 2896 | /* JG review for NX */ |
| 2897 | |
| 2898 | if (pmap == NULL) |
| 2899 | return; |
| 2900 | if ((prot & VM_PROT_READ) == VM_PROT_NONE) { |
| 2901 | pmap_remove(pmap, sva, eva); |
| 2902 | return; |
| 2903 | } |
| 2904 | if (prot & VM_PROT_WRITE) |
| 2905 | return; |
| 2906 | pmap_scan(pmap, sva, eva, pmap_protect_callback, &prot); |
| 2907 | } |
| 2908 | |
| 2909 | static |
| 2910 | void |
| 2911 | pmap_protect_callback(pmap_t pmap, struct pmap_inval_info *info, |
| 2912 | pv_entry_t pte_pv, pv_entry_t pt_pv, vm_offset_t va, |
| 2913 | pt_entry_t *ptep, void *arg __unused) |
| 2914 | { |
| 2915 | pt_entry_t pbits; |
| 2916 | pt_entry_t cbits; |
| 2917 | vm_page_t m; |
| 2918 | |
| 2919 | /* |
| 2920 | * XXX non-optimal. |
| 2921 | */ |
| 2922 | pmap_inval_interlock(info, pmap, va); |
| 2923 | again: |
| 2924 | pbits = *ptep; |
| 2925 | cbits = pbits; |
| 2926 | if (pte_pv) { |
| 2927 | m = NULL; |
| 2928 | if (pbits & PG_A) { |
| 2929 | m = PHYS_TO_VM_PAGE(pbits & PG_FRAME); |
| 2930 | KKASSERT(m == pte_pv->pv_m); |
| 2931 | vm_page_flag_set(m, PG_REFERENCED); |
| 2932 | cbits &= ~PG_A; |
| 2933 | } |
| 2934 | if (pbits & PG_M) { |
| 2935 | if (pmap_track_modified(pte_pv->pv_pindex)) { |
| 2936 | if (m == NULL) |
| 2937 | m = PHYS_TO_VM_PAGE(pbits & PG_FRAME); |
| 2938 | vm_page_dirty(m); |
| 2939 | cbits &= ~PG_M; |
| 2940 | } |
| 2941 | } |
| 2942 | } |
| 2943 | cbits &= ~PG_RW; |
| 2944 | if (pbits != cbits && !atomic_cmpset_long(ptep, pbits, cbits)) { |
| 2945 | goto again; |
| 2946 | } |
| 2947 | pmap_inval_deinterlock(info, pmap); |
| 2948 | if (pte_pv) |
| 2949 | pv_put(pte_pv); |
| 2950 | } |
| 2951 | |
| 2952 | /* |
| 2953 | * Insert the vm_page (m) at the virtual address (va), replacing any prior |
| 2954 | * mapping at that address. Set protection and wiring as requested. |
| 2955 | * |
| 2956 | * NOTE: This routine MUST insert the page into the pmap now, it cannot |
| 2957 | * lazy-evaluate. |
| 2958 | */ |
| 2959 | void |
| 2960 | pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, |
| 2961 | boolean_t wired) |
| 2962 | { |
| 2963 | pmap_inval_info info; |
| 2964 | pv_entry_t pt_pv; /* page table */ |
| 2965 | pv_entry_t pte_pv; /* page table entry */ |
| 2966 | pt_entry_t *ptep; |
| 2967 | vm_paddr_t opa; |
| 2968 | pt_entry_t origpte, newpte; |
| 2969 | vm_paddr_t pa; |
| 2970 | |
| 2971 | if (pmap == NULL) |
| 2972 | return; |
| 2973 | va = trunc_page(va); |
| 2974 | #ifdef PMAP_DIAGNOSTIC |
| 2975 | if (va >= KvaEnd) |
| 2976 | panic("pmap_enter: toobig"); |
| 2977 | if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) |
| 2978 | panic("pmap_enter: invalid to pmap_enter page table " |
| 2979 | "pages (va: 0x%lx)", va); |
| 2980 | #endif |
| 2981 | if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) { |
| 2982 | kprintf("Warning: pmap_enter called on UVA with " |
| 2983 | "kernel_pmap\n"); |
| 2984 | #ifdef DDB |
| 2985 | db_print_backtrace(); |
| 2986 | #endif |
| 2987 | } |
| 2988 | if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) { |
| 2989 | kprintf("Warning: pmap_enter called on KVA without" |
| 2990 | "kernel_pmap\n"); |
| 2991 | #ifdef DDB |
| 2992 | db_print_backtrace(); |
| 2993 | #endif |
| 2994 | } |
| 2995 | |
| 2996 | /* |
| 2997 | * Get locked PV entries for our new page table entry (pte_pv) |
| 2998 | * and for its parent page table (pt_pv). We need the parent |
| 2999 | * so we can resolve the location of the ptep. |
| 3000 | * |
| 3001 | * Only hardware MMU actions can modify the ptep out from |
| 3002 | * under us. |
| 3003 | * |
| 3004 | * if (m) is fictitious or unmanaged we do not create a managing |
| 3005 | * pte_pv for it. Any pre-existing page's management state must |
| 3006 | * match (avoiding code complexity). |
| 3007 | * |
| 3008 | * If the pmap is still being initialized we assume existing |
| 3009 | * page tables. |
| 3010 | * |
| 3011 | * Kernel mapppings do not track page table pages (i.e. pt_pv). |
| 3012 | * pmap_allocpte() checks the |
| 3013 | */ |
| 3014 | if (pmap_initialized == FALSE) { |
| 3015 | pte_pv = NULL; |
| 3016 | pt_pv = NULL; |
| 3017 | ptep = vtopte(va); |
| 3018 | } else if (m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) { |
| 3019 | pte_pv = NULL; |
| 3020 | if (va >= VM_MAX_USER_ADDRESS) { |
| 3021 | pt_pv = NULL; |
| 3022 | ptep = vtopte(va); |
| 3023 | } else { |
| 3024 | pt_pv = pmap_allocpte(pmap, pmap_pt_pindex(va), NULL); |
| 3025 | ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va)); |
| 3026 | } |
| 3027 | KKASSERT(*ptep == 0 || (*ptep & PG_MANAGED) == 0); |
| 3028 | } else { |
| 3029 | if (va >= VM_MAX_USER_ADDRESS) { |
| 3030 | pt_pv = NULL; |
| 3031 | pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL); |
| 3032 | ptep = vtopte(va); |
| 3033 | } else { |
| 3034 | pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), |
| 3035 | &pt_pv); |
| 3036 | ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va)); |
| 3037 | } |
| 3038 | KKASSERT(*ptep == 0 || (*ptep & PG_MANAGED)); |
| 3039 | } |
| 3040 | |
| 3041 | pa = VM_PAGE_TO_PHYS(m); |
| 3042 | origpte = *ptep; |
| 3043 | opa = origpte & PG_FRAME; |
| 3044 | |
| 3045 | newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) | PG_V | PG_A); |
| 3046 | if (wired) |
| 3047 | newpte |= PG_W; |
| 3048 | if (va < VM_MAX_USER_ADDRESS) |
| 3049 | newpte |= PG_U; |
| 3050 | if (pte_pv) |
| 3051 | newpte |= PG_MANAGED; |
| 3052 | if (pmap == &kernel_pmap) |
| 3053 | newpte |= pgeflag; |
| 3054 | |
| 3055 | /* |
| 3056 | * It is possible for multiple faults to occur in threaded |
| 3057 | * environments, the existing pte might be correct. |
| 3058 | */ |
| 3059 | if (((origpte ^ newpte) & ~(pt_entry_t)(PG_M|PG_A)) == 0) |
| 3060 | goto done; |
| 3061 | |
| 3062 | if ((prot & VM_PROT_NOSYNC) == 0) |
| 3063 | pmap_inval_init(&info); |
| 3064 | |
| 3065 | /* |
| 3066 | * Ok, either the address changed or the protection or wiring |
| 3067 | * changed. |
| 3068 | * |
| 3069 | * Clear the current entry, interlocking the removal. For managed |
| 3070 | * pte's this will also flush the modified state to the vm_page. |
| 3071 | * Atomic ops are mandatory in order to ensure that PG_M events are |
| 3072 | * not lost during any transition. |
| 3073 | */ |
| 3074 | if (opa) { |
| 3075 | if (pte_pv) { |
| 3076 | /* |
| 3077 | * pmap_remove_pv_pte() unwires pt_pv and assumes |
| 3078 | * we will free pte_pv, but since we are reusing |
| 3079 | * pte_pv we want to retain the wire count. |
| 3080 | * |
| 3081 | * pt_pv won't exist for a kernel page (managed or |
| 3082 | * otherwise). |
| 3083 | */ |
| 3084 | if (pt_pv) |
| 3085 | vm_page_wire_quick(pt_pv->pv_m); |
| 3086 | if (prot & VM_PROT_NOSYNC) |
| 3087 | pmap_remove_pv_pte(pte_pv, pt_pv, NULL); |
| 3088 | else |
| 3089 | pmap_remove_pv_pte(pte_pv, pt_pv, &info); |
| 3090 | if (pte_pv->pv_m) |
| 3091 | pmap_remove_pv_page(pte_pv); |
| 3092 | } else if (prot & VM_PROT_NOSYNC) { |
| 3093 | /* leave wire count on PT page intact */ |
| 3094 | (void)pte_load_clear(ptep); |
| 3095 | cpu_invlpg((void *)va); |
| 3096 | atomic_add_long(&pmap->pm_stats.resident_count, -1); |
| 3097 | } else { |
| 3098 | /* leave wire count on PT page intact */ |
| 3099 | pmap_inval_interlock(&info, pmap, va); |
| 3100 | (void)pte_load_clear(ptep); |
| 3101 | pmap_inval_deinterlock(&info, pmap); |
| 3102 | atomic_add_long(&pmap->pm_stats.resident_count, -1); |
| 3103 | } |
| 3104 | KKASSERT(*ptep == 0); |
| 3105 | } |
| 3106 | |
| 3107 | if (pte_pv) { |
| 3108 | /* |
| 3109 | * Enter on the PV list if part of our managed memory. |
| 3110 | * Wiring of the PT page is already handled. |
| 3111 | */ |
| 3112 | KKASSERT(pte_pv->pv_m == NULL); |
| 3113 | vm_page_spin_lock(m); |
| 3114 | pte_pv->pv_m = m; |
| 3115 | TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list); |
| 3116 | /* |
| 3117 | if (m->object) |
| 3118 | atomic_add_int(&m->object->agg_pv_list_count, 1); |
| 3119 | */ |
| 3120 | vm_page_flag_set(m, PG_MAPPED); |
| 3121 | vm_page_spin_unlock(m); |
| 3122 | } else if (pt_pv && opa == 0) { |
| 3123 | /* |
| 3124 | * We have to adjust the wire count on the PT page ourselves |
| 3125 | * for unmanaged entries. If opa was non-zero we retained |
| 3126 | * the existing wire count from the removal. |
| 3127 | */ |
| 3128 | vm_page_wire_quick(pt_pv->pv_m); |
| 3129 | } |
| 3130 | |
| 3131 | /* |
| 3132 | * Ok, for UVM (pt_pv != NULL) we don't need to interlock or |
| 3133 | * invalidate anything, the TLB won't have any stale entries to |
| 3134 | * remove. |
| 3135 | * |
| 3136 | * For KVM there appear to still be issues. Theoretically we |
| 3137 | * should be able to scrap the interlocks entirely but we |
| 3138 | * get crashes. |
| 3139 | */ |
| 3140 | if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL) |
| 3141 | pmap_inval_interlock(&info, pmap, va); |
| 3142 | *(volatile pt_entry_t *)ptep = newpte; |
| 3143 | |
| 3144 | if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL) |
| 3145 | pmap_inval_deinterlock(&info, pmap); |
| 3146 | else if (pt_pv == NULL) |
| 3147 | cpu_invlpg((void *)va); |
| 3148 | |
| 3149 | if (wired) |
| 3150 | atomic_add_long(&pmap->pm_stats.wired_count, 1); |
| 3151 | if (newpte & PG_RW) |
| 3152 | vm_page_flag_set(m, PG_WRITEABLE); |
| 3153 | if (pte_pv == NULL) |
| 3154 | atomic_add_long(&pmap->pm_stats.resident_count, 1); |
| 3155 | |
| 3156 | /* |
| 3157 | * Cleanup |
| 3158 | */ |
| 3159 | if ((prot & VM_PROT_NOSYNC) == 0 || pte_pv == NULL) |
| 3160 | pmap_inval_done(&info); |
| 3161 | done: |
| 3162 | KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED)); |
| 3163 | |
| 3164 | /* |
| 3165 | * Cleanup the pv entry, allowing other accessors. |
| 3166 | */ |
| 3167 | if (pte_pv) |
| 3168 | pv_put(pte_pv); |
| 3169 | if (pt_pv) |
| 3170 | pv_put(pt_pv); |
| 3171 | } |
| 3172 | |
| 3173 | /* |
| 3174 | * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired. |
| 3175 | * This code also assumes that the pmap has no pre-existing entry for this |
| 3176 | * VA. |
| 3177 | * |
| 3178 | * This code currently may only be used on user pmaps, not kernel_pmap. |
| 3179 | */ |
| 3180 | void |
| 3181 | pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m) |
| 3182 | { |
| 3183 | pmap_enter(pmap, va, m, VM_PROT_READ, FALSE); |
| 3184 | } |
| 3185 | |
| 3186 | /* |
| 3187 | * Make a temporary mapping for a physical address. This is only intended |
| 3188 | * to be used for panic dumps. |
| 3189 | * |
| 3190 | * The caller is responsible for calling smp_invltlb(). |
| 3191 | */ |
| 3192 | void * |
| 3193 | pmap_kenter_temporary(vm_paddr_t pa, long i) |
| 3194 | { |
| 3195 | pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa); |
| 3196 | return ((void *)crashdumpmap); |
| 3197 | } |
| 3198 | |
| 3199 | #define MAX_INIT_PT (96) |
| 3200 | |
| 3201 | /* |
| 3202 | * This routine preloads the ptes for a given object into the specified pmap. |
| 3203 | * This eliminates the blast of soft faults on process startup and |
| 3204 | * immediately after an mmap. |
| 3205 | */ |
| 3206 | static int pmap_object_init_pt_callback(vm_page_t p, void *data); |
| 3207 | |
| 3208 | void |
| 3209 | pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot, |
| 3210 | vm_object_t object, vm_pindex_t pindex, |
| 3211 | vm_size_t size, int limit) |
| 3212 | { |
| 3213 | struct rb_vm_page_scan_info info; |
| 3214 | struct lwp *lp; |
| 3215 | vm_size_t psize; |
| 3216 | |
| 3217 | /* |
| 3218 | * We can't preinit if read access isn't set or there is no pmap |
| 3219 | * or object. |
| 3220 | */ |
| 3221 | if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL) |
| 3222 | return; |
| 3223 | |
| 3224 | /* |
| 3225 | * We can't preinit if the pmap is not the current pmap |
| 3226 | */ |
| 3227 | lp = curthread->td_lwp; |
| 3228 | if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace)) |
| 3229 | return; |
| 3230 | |
| 3231 | psize = x86_64_btop(size); |
| 3232 | |
| 3233 | if ((object->type != OBJT_VNODE) || |
| 3234 | ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) && |
| 3235 | (object->resident_page_count > MAX_INIT_PT))) { |
| 3236 | return; |
| 3237 | } |
| 3238 | |
| 3239 | if (pindex + psize > object->size) { |
| 3240 | if (object->size < pindex) |
| 3241 | return; |
| 3242 | psize = object->size - pindex; |
| 3243 | } |
| 3244 | |
| 3245 | if (psize == 0) |
| 3246 | return; |
| 3247 | |
| 3248 | /* |
| 3249 | * Use a red-black scan to traverse the requested range and load |
| 3250 | * any valid pages found into the pmap. |
| 3251 | * |
| 3252 | * We cannot safely scan the object's memq without holding the |
| 3253 | * object token. |
| 3254 | */ |
| 3255 | info.start_pindex = pindex; |
| 3256 | info.end_pindex = pindex + psize - 1; |
| 3257 | info.limit = limit; |
| 3258 | info.mpte = NULL; |
| 3259 | info.addr = addr; |
| 3260 | info.pmap = pmap; |
| 3261 | |
| 3262 | vm_object_hold_shared(object); |
| 3263 | vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, |
| 3264 | pmap_object_init_pt_callback, &info); |
| 3265 | vm_object_drop(object); |
| 3266 | } |
| 3267 | |
| 3268 | static |
| 3269 | int |
| 3270 | pmap_object_init_pt_callback(vm_page_t p, void *data) |
| 3271 | { |
| 3272 | struct rb_vm_page_scan_info *info = data; |
| 3273 | vm_pindex_t rel_index; |
| 3274 | |
| 3275 | /* |
| 3276 | * don't allow an madvise to blow away our really |
| 3277 | * free pages allocating pv entries. |
| 3278 | */ |
| 3279 | if ((info->limit & MAP_PREFAULT_MADVISE) && |
| 3280 | vmstats.v_free_count < vmstats.v_free_reserved) { |
| 3281 | return(-1); |
| 3282 | } |
| 3283 | |
| 3284 | /* |
| 3285 | * Ignore list markers and ignore pages we cannot instantly |
| 3286 | * busy (while holding the object token). |
| 3287 | */ |
| 3288 | if (p->flags & PG_MARKER) |
| 3289 | return 0; |
| 3290 | if (vm_page_busy_try(p, TRUE)) |
| 3291 | return 0; |
| 3292 | if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && |
| 3293 | (p->flags & PG_FICTITIOUS) == 0) { |
| 3294 | if ((p->queue - p->pc) == PQ_CACHE) |
| 3295 | vm_page_deactivate(p); |
| 3296 | rel_index = p->pindex - info->start_pindex; |
| 3297 | pmap_enter_quick(info->pmap, |
| 3298 | info->addr + x86_64_ptob(rel_index), p); |
| 3299 | } |
| 3300 | vm_page_wakeup(p); |
| 3301 | lwkt_yield(); |
| 3302 | return(0); |
| 3303 | } |
| 3304 | |
| 3305 | /* |
| 3306 | * Return TRUE if the pmap is in shape to trivially pre-fault the specified |
| 3307 | * address. |
| 3308 | * |
| 3309 | * Returns FALSE if it would be non-trivial or if a pte is already loaded |
| 3310 | * into the slot. |
| 3311 | * |
| 3312 | * XXX This is safe only because page table pages are not freed. |
| 3313 | */ |
| 3314 | int |
| 3315 | pmap_prefault_ok(pmap_t pmap, vm_offset_t addr) |
| 3316 | { |
| 3317 | pt_entry_t *pte; |
| 3318 | |
| 3319 | /*spin_lock(&pmap->pm_spin);*/ |
| 3320 | if ((pte = pmap_pte(pmap, addr)) != NULL) { |
| 3321 | if (*pte & PG_V) { |
| 3322 | /*spin_unlock(&pmap->pm_spin);*/ |
| 3323 | return FALSE; |
| 3324 | } |
| 3325 | } |
| 3326 | /*spin_unlock(&pmap->pm_spin);*/ |
| 3327 | return TRUE; |
| 3328 | } |
| 3329 | |
| 3330 | /* |
| 3331 | * Change the wiring attribute for a pmap/va pair. The mapping must already |
| 3332 | * exist in the pmap. The mapping may or may not be managed. |
| 3333 | */ |
| 3334 | void |
| 3335 | pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired) |
| 3336 | { |
| 3337 | pt_entry_t *ptep; |
| 3338 | pv_entry_t pv; |
| 3339 | |
| 3340 | if (pmap == NULL) |
| 3341 | return; |
| 3342 | lwkt_gettoken(&pmap->pm_token); |
| 3343 | pv = pmap_allocpte(pmap, pmap_pt_pindex(va), NULL); |
| 3344 | ptep = pv_pte_lookup(pv, pmap_pte_index(va)); |
| 3345 | |
| 3346 | if (wired && !pmap_pte_w(ptep)) |
| 3347 | atomic_add_long(&pmap->pm_stats.wired_count, 1); |
| 3348 | else if (!wired && pmap_pte_w(ptep)) |
| 3349 | atomic_add_long(&pmap->pm_stats.wired_count, -1); |
| 3350 | |
| 3351 | /* |
| 3352 | * Wiring is not a hardware characteristic so there is no need to |
| 3353 | * invalidate TLB. However, in an SMP environment we must use |
| 3354 | * a locked bus cycle to update the pte (if we are not using |
| 3355 | * the pmap_inval_*() API that is)... it's ok to do this for simple |
| 3356 | * wiring changes. |
| 3357 | */ |
| 3358 | #ifdef SMP |
| 3359 | if (wired) |
| 3360 | atomic_set_long(ptep, PG_W); |
| 3361 | else |
| 3362 | atomic_clear_long(ptep, PG_W); |
| 3363 | #else |
| 3364 | if (wired) |
| 3365 | atomic_set_long_nonlocked(ptep, PG_W); |
| 3366 | else |
| 3367 | atomic_clear_long_nonlocked(ptep, PG_W); |
| 3368 | #endif |
| 3369 | pv_put(pv); |
| 3370 | lwkt_reltoken(&pmap->pm_token); |
| 3371 | } |
| 3372 | |
| 3373 | |
| 3374 | |
| 3375 | /* |
| 3376 | * Copy the range specified by src_addr/len from the source map to |
| 3377 | * the range dst_addr/len in the destination map. |
| 3378 | * |
| 3379 | * This routine is only advisory and need not do anything. |
| 3380 | */ |
| 3381 | void |
| 3382 | pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, |
| 3383 | vm_size_t len, vm_offset_t src_addr) |
| 3384 | { |
| 3385 | } |
| 3386 | |
| 3387 | /* |
| 3388 | * pmap_zero_page: |
| 3389 | * |
| 3390 | * Zero the specified physical page. |
| 3391 | * |
| 3392 | * This function may be called from an interrupt and no locking is |
| 3393 | * required. |
| 3394 | */ |
| 3395 | void |
| 3396 | pmap_zero_page(vm_paddr_t phys) |
| 3397 | { |
| 3398 | vm_offset_t va = PHYS_TO_DMAP(phys); |
| 3399 | |
| 3400 | pagezero((void *)va); |
| 3401 | } |
| 3402 | |
| 3403 | /* |
| 3404 | * pmap_page_assertzero: |
| 3405 | * |
| 3406 | * Assert that a page is empty, panic if it isn't. |
| 3407 | */ |
| 3408 | void |
| 3409 | pmap_page_assertzero(vm_paddr_t phys) |
| 3410 | { |
| 3411 | vm_offset_t va = PHYS_TO_DMAP(phys); |
| 3412 | size_t i; |
| 3413 | |
| 3414 | for (i = 0; i < PAGE_SIZE; i += sizeof(long)) { |
| 3415 | if (*(long *)((char *)va + i) != 0) { |
| 3416 | panic("pmap_page_assertzero() @ %p not zero!\n", |
| 3417 | (void *)(intptr_t)va); |
| 3418 | } |
| 3419 | } |
| 3420 | } |
| 3421 | |
| 3422 | /* |
| 3423 | * pmap_zero_page: |
| 3424 | * |
| 3425 | * Zero part of a physical page by mapping it into memory and clearing |
| 3426 | * its contents with bzero. |
| 3427 | * |
| 3428 | * off and size may not cover an area beyond a single hardware page. |
| 3429 | */ |
| 3430 | void |
| 3431 | pmap_zero_page_area(vm_paddr_t phys, int off, int size) |
| 3432 | { |
| 3433 | vm_offset_t virt = PHYS_TO_DMAP(phys); |
| 3434 | |
| 3435 | bzero((char *)virt + off, size); |
| 3436 | } |
| 3437 | |
| 3438 | /* |
| 3439 | * pmap_copy_page: |
| 3440 | * |
| 3441 | * Copy the physical page from the source PA to the target PA. |
| 3442 | * This function may be called from an interrupt. No locking |
| 3443 | * is required. |
| 3444 | */ |
| 3445 | void |
| 3446 | pmap_copy_page(vm_paddr_t src, vm_paddr_t dst) |
| 3447 | { |
| 3448 | vm_offset_t src_virt, dst_virt; |
| 3449 | |
| 3450 | src_virt = PHYS_TO_DMAP(src); |
| 3451 | dst_virt = PHYS_TO_DMAP(dst); |
| 3452 | bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE); |
| 3453 | } |
| 3454 | |
| 3455 | /* |
| 3456 | * pmap_copy_page_frag: |
| 3457 | * |
| 3458 | * Copy the physical page from the source PA to the target PA. |
| 3459 | * This function may be called from an interrupt. No locking |
| 3460 | * is required. |
| 3461 | */ |
| 3462 | void |
| 3463 | pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes) |
| 3464 | { |
| 3465 | vm_offset_t src_virt, dst_virt; |
| 3466 | |
| 3467 | src_virt = PHYS_TO_DMAP(src); |
| 3468 | dst_virt = PHYS_TO_DMAP(dst); |
| 3469 | |
| 3470 | bcopy((char *)src_virt + (src & PAGE_MASK), |
| 3471 | (char *)dst_virt + (dst & PAGE_MASK), |
| 3472 | bytes); |
| 3473 | } |
| 3474 | |
| 3475 | /* |
| 3476 | * Returns true if the pmap's pv is one of the first 16 pvs linked to from |
| 3477 | * this page. This count may be changed upwards or downwards in the future; |
| 3478 | * it is only necessary that true be returned for a small subset of pmaps |
| 3479 | * for proper page aging. |
| 3480 | */ |
| 3481 | boolean_t |
| 3482 | pmap_page_exists_quick(pmap_t pmap, vm_page_t m) |
| 3483 | { |
| 3484 | pv_entry_t pv; |
| 3485 | int loops = 0; |
| 3486 | |
| 3487 | if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) |
| 3488 | return FALSE; |
| 3489 | |
| 3490 | vm_page_spin_lock(m); |
| 3491 | TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { |
| 3492 | if (pv->pv_pmap == pmap) { |
| 3493 | vm_page_spin_unlock(m); |
| 3494 | return TRUE; |
| 3495 | } |
| 3496 | loops++; |
| 3497 | if (loops >= 16) |
| 3498 | break; |
| 3499 | } |
| 3500 | vm_page_spin_unlock(m); |
| 3501 | return (FALSE); |
| 3502 | } |
| 3503 | |
| 3504 | /* |
| 3505 | * Remove all pages from specified address space this aids process exit |
| 3506 | * speeds. Also, this code may be special cased for the current process |
| 3507 | * only. |
| 3508 | */ |
| 3509 | void |
| 3510 | pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) |
| 3511 | { |
| 3512 | pmap_remove(pmap, sva, eva); |
| 3513 | } |
| 3514 | |
| 3515 | /* |
| 3516 | * pmap_testbit tests bits in pte's note that the testbit/clearbit |
| 3517 | * routines are inline, and a lot of things compile-time evaluate. |
| 3518 | */ |
| 3519 | static |
| 3520 | boolean_t |
| 3521 | pmap_testbit(vm_page_t m, int bit) |
| 3522 | { |
| 3523 | pv_entry_t pv; |
| 3524 | pt_entry_t *pte; |
| 3525 | |
| 3526 | if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) |
| 3527 | return FALSE; |
| 3528 | |
| 3529 | if (TAILQ_FIRST(&m->md.pv_list) == NULL) |
| 3530 | return FALSE; |
| 3531 | vm_page_spin_lock(m); |
| 3532 | if (TAILQ_FIRST(&m->md.pv_list) == NULL) { |
| 3533 | vm_page_spin_unlock(m); |
| 3534 | return FALSE; |
| 3535 | } |
| 3536 | |
| 3537 | TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { |
| 3538 | /* |
| 3539 | * if the bit being tested is the modified bit, then |
| 3540 | * mark clean_map and ptes as never |
| 3541 | * modified. |
| 3542 | */ |
| 3543 | if (bit & (PG_A|PG_M)) { |
| 3544 | if (!pmap_track_modified(pv->pv_pindex)) |
| 3545 | continue; |
| 3546 | } |
| 3547 | |
| 3548 | #if defined(PMAP_DIAGNOSTIC) |
| 3549 | if (pv->pv_pmap == NULL) { |
| 3550 | kprintf("Null pmap (tb) at pindex: %"PRIu64"\n", |
| 3551 | pv->pv_pindex); |
| 3552 | continue; |
| 3553 | } |
| 3554 | #endif |
| 3555 | pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT); |
| 3556 | if (*pte & bit) { |
| 3557 | vm_page_spin_unlock(m); |
| 3558 | return TRUE; |
| 3559 | } |
| 3560 | } |
| 3561 | vm_page_spin_unlock(m); |
| 3562 | return (FALSE); |
| 3563 | } |
| 3564 | |
| 3565 | /* |
| 3566 | * This routine is used to modify bits in ptes. Only one bit should be |
| 3567 | * specified. PG_RW requires special handling. |
| 3568 | * |
| 3569 | * Caller must NOT hold any spin locks |
| 3570 | */ |
| 3571 | static __inline |
| 3572 | void |
| 3573 | pmap_clearbit(vm_page_t m, int bit) |
| 3574 | { |
| 3575 | struct pmap_inval_info info; |
| 3576 | pv_entry_t pv; |
| 3577 | pt_entry_t *pte; |
| 3578 | pt_entry_t pbits; |
| 3579 | pmap_t save_pmap; |
| 3580 | |
| 3581 | if (bit == PG_RW) |
| 3582 | vm_page_flag_clear(m, PG_WRITEABLE); |
| 3583 | if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) { |
| 3584 | return; |
| 3585 | } |
| 3586 | |
| 3587 | /* |
| 3588 | * PG_M or PG_A case |
| 3589 | * |
| 3590 | * Loop over all current mappings setting/clearing as appropos If |
| 3591 | * setting RO do we need to clear the VAC? |
| 3592 | * |
| 3593 | * NOTE: When clearing PG_M we could also (not implemented) drop |
| 3594 | * through to the PG_RW code and clear PG_RW too, forcing |
| 3595 | * a fault on write to redetect PG_M for virtual kernels, but |
| 3596 | * it isn't necessary since virtual kernels invalidate the |
| 3597 | * pte when they clear the VPTE_M bit in their virtual page |
| 3598 | * tables. |
| 3599 | * |
| 3600 | * NOTE: Does not re-dirty the page when clearing only PG_M. |
| 3601 | */ |
| 3602 | if ((bit & PG_RW) == 0) { |
| 3603 | vm_page_spin_lock(m); |
| 3604 | TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { |
| 3605 | #if defined(PMAP_DIAGNOSTIC) |
| 3606 | if (pv->pv_pmap == NULL) { |
| 3607 | kprintf("Null pmap (cb) at pindex: %"PRIu64"\n", |
| 3608 | pv->pv_pindex); |
| 3609 | continue; |
| 3610 | } |
| 3611 | #endif |
| 3612 | pte = pmap_pte_quick(pv->pv_pmap, |
| 3613 | pv->pv_pindex << PAGE_SHIFT); |
| 3614 | pbits = *pte; |
| 3615 | if (pbits & bit) |
| 3616 | atomic_clear_long(pte, bit); |
| 3617 | } |
| 3618 | vm_page_spin_unlock(m); |
| 3619 | return; |
| 3620 | } |
| 3621 | |
| 3622 | /* |
| 3623 | * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M |
| 3624 | * was set. |
| 3625 | */ |
| 3626 | pmap_inval_init(&info); |
| 3627 | |
| 3628 | restart: |
| 3629 | vm_page_spin_lock(m); |
| 3630 | TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { |
| 3631 | /* |
| 3632 | * don't write protect pager mappings |
| 3633 | */ |
| 3634 | if (!pmap_track_modified(pv->pv_pindex)) |
| 3635 | continue; |
| 3636 | |
| 3637 | #if defined(PMAP_DIAGNOSTIC) |
| 3638 | if (pv->pv_pmap == NULL) { |
| 3639 | kprintf("Null pmap (cb) at pindex: %"PRIu64"\n", |
| 3640 | pv->pv_pindex); |
| 3641 | continue; |
| 3642 | } |
| 3643 | #endif |
| 3644 | /* |
| 3645 | * Skip pages which do not have PG_RW set. |
| 3646 | */ |
| 3647 | pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT); |
| 3648 | if ((*pte & PG_RW) == 0) |
| 3649 | continue; |
| 3650 | |
| 3651 | /* |
| 3652 | * Lock the PV |
| 3653 | */ |
| 3654 | if (pv_hold_try(pv) == 0) { |
| 3655 | vm_page_spin_unlock(m); |
| 3656 | pv_lock(pv); /* held, now do a blocking lock */ |
| 3657 | pv_put(pv); /* and release */ |
| 3658 | goto restart; /* anything could have happened */ |
| 3659 | } |
| 3660 | |
| 3661 | save_pmap = pv->pv_pmap; |
| 3662 | vm_page_spin_unlock(m); |
| 3663 | pmap_inval_interlock(&info, save_pmap, |
| 3664 | (vm_offset_t)pv->pv_pindex << PAGE_SHIFT); |
| 3665 | KKASSERT(pv->pv_pmap == save_pmap); |
| 3666 | for (;;) { |
| 3667 | pbits = *pte; |
| 3668 | cpu_ccfence(); |
| 3669 | if (atomic_cmpset_long(pte, pbits, |
| 3670 | pbits & ~(PG_RW|PG_M))) { |
| 3671 | break; |
| 3672 | } |
| 3673 | } |
| 3674 | pmap_inval_deinterlock(&info, save_pmap); |
| 3675 | vm_page_spin_lock(m); |
| 3676 | |
| 3677 | /* |
| 3678 | * If PG_M was found to be set while we were clearing PG_RW |
| 3679 | * we also clear PG_M (done above) and mark the page dirty. |
| 3680 | * Callers expect this behavior. |
| 3681 | */ |
| 3682 | if (pbits & PG_M) |
| 3683 | vm_page_dirty(m); |
| 3684 | pv_put(pv); |
| 3685 | } |
| 3686 | vm_page_spin_unlock(m); |
| 3687 | pmap_inval_done(&info); |
| 3688 | } |
| 3689 | |
| 3690 | /* |
| 3691 | * Lower the permission for all mappings to a given page. |
| 3692 | * |
| 3693 | * Page must be busied by caller. |
| 3694 | */ |
| 3695 | void |
| 3696 | pmap_page_protect(vm_page_t m, vm_prot_t prot) |
| 3697 | { |
| 3698 | /* JG NX support? */ |
| 3699 | if ((prot & VM_PROT_WRITE) == 0) { |
| 3700 | if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { |
| 3701 | /* |
| 3702 | * NOTE: pmap_clearbit(.. PG_RW) also clears |
| 3703 | * the PG_WRITEABLE flag in (m). |
| 3704 | */ |
| 3705 | pmap_clearbit(m, PG_RW); |
| 3706 | } else { |
| 3707 | pmap_remove_all(m); |
| 3708 | } |
| 3709 | } |
| 3710 | } |
| 3711 | |
| 3712 | vm_paddr_t |
| 3713 | pmap_phys_address(vm_pindex_t ppn) |
| 3714 | { |
| 3715 | return (x86_64_ptob(ppn)); |
| 3716 | } |
| 3717 | |
| 3718 | /* |
| 3719 | * Return a count of reference bits for a page, clearing those bits. |
| 3720 | * It is not necessary for every reference bit to be cleared, but it |
| 3721 | * is necessary that 0 only be returned when there are truly no |
| 3722 | * reference bits set. |
| 3723 | * |
| 3724 | * XXX: The exact number of bits to check and clear is a matter that |
| 3725 | * should be tested and standardized at some point in the future for |
| 3726 | * optimal aging of shared pages. |
| 3727 | * |
| 3728 | * This routine may not block. |
| 3729 | */ |
| 3730 | int |
| 3731 | pmap_ts_referenced(vm_page_t m) |
| 3732 | { |
| 3733 | pv_entry_t pv; |
| 3734 | pt_entry_t *pte; |
| 3735 | int rtval = 0; |
| 3736 | |
| 3737 | if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) |
| 3738 | return (rtval); |
| 3739 | |
| 3740 | vm_page_spin_lock(m); |
| 3741 | TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { |
| 3742 | if (!pmap_track_modified(pv->pv_pindex)) |
| 3743 | continue; |
| 3744 | pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT); |
| 3745 | if (pte && (*pte & PG_A)) { |
| 3746 | #ifdef SMP |
| 3747 | atomic_clear_long(pte, PG_A); |
| 3748 | #else |
| 3749 | atomic_clear_long_nonlocked(pte, PG_A); |
| 3750 | #endif |
| 3751 | rtval++; |
| 3752 | if (rtval > 4) |
| 3753 | break; |
| 3754 | } |
| 3755 | } |
| 3756 | vm_page_spin_unlock(m); |
| 3757 | return (rtval); |
| 3758 | } |
| 3759 | |
| 3760 | /* |
| 3761 | * pmap_is_modified: |
| 3762 | * |
| 3763 | * Return whether or not the specified physical page was modified |
| 3764 | * in any physical maps. |
| 3765 | */ |
| 3766 | boolean_t |
| 3767 | pmap_is_modified(vm_page_t m) |
| 3768 | { |
| 3769 | boolean_t res; |
| 3770 | |
| 3771 | res = pmap_testbit(m, PG_M); |
| 3772 | return (res); |
| 3773 | } |
| 3774 | |
| 3775 | /* |
| 3776 | * Clear the modify bits on the specified physical page. |
| 3777 | */ |
| 3778 | void |
| 3779 | pmap_clear_modify(vm_page_t m) |
| 3780 | { |
| 3781 | pmap_clearbit(m, PG_M); |
| 3782 | } |
| 3783 | |
| 3784 | /* |
| 3785 | * pmap_clear_reference: |
| 3786 | * |
| 3787 | * Clear the reference bit on the specified physical page. |
| 3788 | */ |
| 3789 | void |
| 3790 | pmap_clear_reference(vm_page_t m) |
| 3791 | { |
| 3792 | pmap_clearbit(m, PG_A); |
| 3793 | } |
| 3794 | |
| 3795 | /* |
| 3796 | * Miscellaneous support routines follow |
| 3797 | */ |
| 3798 | |
| 3799 | static |
| 3800 | void |
| 3801 | i386_protection_init(void) |
| 3802 | { |
| 3803 | int *kp, prot; |
| 3804 | |
| 3805 | /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */ |
| 3806 | kp = protection_codes; |
| 3807 | for (prot = 0; prot < 8; prot++) { |
| 3808 | switch (prot) { |
| 3809 | case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE: |
| 3810 | /* |
| 3811 | * Read access is also 0. There isn't any execute bit, |
| 3812 | * so just make it readable. |
| 3813 | */ |
| 3814 | case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE: |
| 3815 | case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE: |
| 3816 | case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE: |
| 3817 | *kp++ = 0; |
| 3818 | break; |
| 3819 | case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE: |
| 3820 | case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE: |
| 3821 | case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE: |
| 3822 | case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE: |
| 3823 | *kp++ = PG_RW; |
| 3824 | break; |
| 3825 | } |
| 3826 | } |
| 3827 | } |
| 3828 | |
| 3829 | /* |
| 3830 | * Map a set of physical memory pages into the kernel virtual |
| 3831 | * address space. Return a pointer to where it is mapped. This |
| 3832 | * routine is intended to be used for mapping device memory, |
| 3833 | * NOT real memory. |
| 3834 | * |
| 3835 | * NOTE: we can't use pgeflag unless we invalidate the pages one at |
| 3836 | * a time. |
| 3837 | */ |
| 3838 | void * |
| 3839 | pmap_mapdev(vm_paddr_t pa, vm_size_t size) |
| 3840 | { |
| 3841 | vm_offset_t va, tmpva, offset; |
| 3842 | pt_entry_t *pte; |
| 3843 | |
| 3844 | offset = pa & PAGE_MASK; |
| 3845 | size = roundup(offset + size, PAGE_SIZE); |
| 3846 | |
| 3847 | va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE); |
| 3848 | if (va == 0) |
| 3849 | panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); |
| 3850 | |
| 3851 | pa = pa & ~PAGE_MASK; |
| 3852 | for (tmpva = va; size > 0;) { |
| 3853 | pte = vtopte(tmpva); |
| 3854 | *pte = pa | PG_RW | PG_V; /* | pgeflag; */ |
| 3855 | size -= PAGE_SIZE; |
| 3856 | tmpva += PAGE_SIZE; |
| 3857 | pa += PAGE_SIZE; |
| 3858 | } |
| 3859 | cpu_invltlb(); |
| 3860 | smp_invltlb(); |
| 3861 | |
| 3862 | return ((void *)(va + offset)); |
| 3863 | } |
| 3864 | |
| 3865 | void * |
| 3866 | pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size) |
| 3867 | { |
| 3868 | vm_offset_t va, tmpva, offset; |
| 3869 | pt_entry_t *pte; |
| 3870 | |
| 3871 | offset = pa & PAGE_MASK; |
| 3872 | size = roundup(offset + size, PAGE_SIZE); |
| 3873 | |
| 3874 | va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE); |
| 3875 | if (va == 0) |
| 3876 | panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); |
| 3877 | |
| 3878 | pa = pa & ~PAGE_MASK; |
| 3879 | for (tmpva = va; size > 0;) { |
| 3880 | pte = vtopte(tmpva); |
| 3881 | *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */ |
| 3882 | size -= PAGE_SIZE; |
| 3883 | tmpva += PAGE_SIZE; |
| 3884 | pa += PAGE_SIZE; |
| 3885 | } |
| 3886 | cpu_invltlb(); |
| 3887 | smp_invltlb(); |
| 3888 | |
| 3889 | return ((void *)(va + offset)); |
| 3890 | } |
| 3891 | |
| 3892 | void |
| 3893 | pmap_unmapdev(vm_offset_t va, vm_size_t size) |
| 3894 | { |
| 3895 | vm_offset_t base, offset; |
| 3896 | |
| 3897 | base = va & ~PAGE_MASK; |
| 3898 | offset = va & PAGE_MASK; |
| 3899 | size = roundup(offset + size, PAGE_SIZE); |
| 3900 | pmap_qremove(va, size >> PAGE_SHIFT); |
| 3901 | kmem_free(&kernel_map, base, size); |
| 3902 | } |
| 3903 | |
| 3904 | /* |
| 3905 | * perform the pmap work for mincore |
| 3906 | */ |
| 3907 | int |
| 3908 | pmap_mincore(pmap_t pmap, vm_offset_t addr) |
| 3909 | { |
| 3910 | pt_entry_t *ptep, pte; |
| 3911 | vm_page_t m; |
| 3912 | int val = 0; |
| 3913 | |
| 3914 | lwkt_gettoken(&pmap->pm_token); |
| 3915 | ptep = pmap_pte(pmap, addr); |
| 3916 | |
| 3917 | if (ptep && (pte = *ptep) != 0) { |
| 3918 | vm_offset_t pa; |
| 3919 | |
| 3920 | val = MINCORE_INCORE; |
| 3921 | if ((pte & PG_MANAGED) == 0) |
| 3922 | goto done; |
| 3923 | |
| 3924 | pa = pte & PG_FRAME; |
| 3925 | |
| 3926 | m = PHYS_TO_VM_PAGE(pa); |
| 3927 | |
| 3928 | /* |
| 3929 | * Modified by us |
| 3930 | */ |
| 3931 | if (pte & PG_M) |
| 3932 | val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; |
| 3933 | /* |
| 3934 | * Modified by someone |
| 3935 | */ |
| 3936 | else if (m->dirty || pmap_is_modified(m)) |
| 3937 | val |= MINCORE_MODIFIED_OTHER; |
| 3938 | /* |
| 3939 | * Referenced by us |
| 3940 | */ |
| 3941 | if (pte & PG_A) |
| 3942 | val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; |
| 3943 | |
| 3944 | /* |
| 3945 | * Referenced by someone |
| 3946 | */ |
| 3947 | else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) { |
| 3948 | val |= MINCORE_REFERENCED_OTHER; |
| 3949 | vm_page_flag_set(m, PG_REFERENCED); |
| 3950 | } |
| 3951 | } |
| 3952 | done: |
| 3953 | lwkt_reltoken(&pmap->pm_token); |
| 3954 | |
| 3955 | return val; |
| 3956 | } |
| 3957 | |
| 3958 | /* |
| 3959 | * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new |
| 3960 | * vmspace will be ref'd and the old one will be deref'd. |
| 3961 | * |
| 3962 | * The vmspace for all lwps associated with the process will be adjusted |
| 3963 | * and cr3 will be reloaded if any lwp is the current lwp. |
| 3964 | * |
| 3965 | * The process must hold the vmspace->vm_map.token for oldvm and newvm |
| 3966 | */ |
| 3967 | void |
| 3968 | pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs) |
| 3969 | { |
| 3970 | struct vmspace *oldvm; |
| 3971 | struct lwp *lp; |
| 3972 | |
| 3973 | oldvm = p->p_vmspace; |
| 3974 | if (oldvm != newvm) { |
| 3975 | if (adjrefs) |
| 3976 | sysref_get(&newvm->vm_sysref); |
| 3977 | p->p_vmspace = newvm; |
| 3978 | KKASSERT(p->p_nthreads == 1); |
| 3979 | lp = RB_ROOT(&p->p_lwp_tree); |
| 3980 | pmap_setlwpvm(lp, newvm); |
| 3981 | if (adjrefs) |
| 3982 | sysref_put(&oldvm->vm_sysref); |
| 3983 | } |
| 3984 | } |
| 3985 | |
| 3986 | /* |
| 3987 | * Set the vmspace for a LWP. The vmspace is almost universally set the |
| 3988 | * same as the process vmspace, but virtual kernels need to swap out contexts |
| 3989 | * on a per-lwp basis. |
| 3990 | * |
| 3991 | * Caller does not necessarily hold any vmspace tokens. Caller must control |
| 3992 | * the lwp (typically be in the context of the lwp). We use a critical |
| 3993 | * section to protect against statclock and hardclock (statistics collection). |
| 3994 | */ |
| 3995 | void |
| 3996 | pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm) |
| 3997 | { |
| 3998 | struct vmspace *oldvm; |
| 3999 | struct pmap *pmap; |
| 4000 | |
| 4001 | oldvm = lp->lwp_vmspace; |
| 4002 | |
| 4003 | if (oldvm != newvm) { |
| 4004 | crit_enter(); |
| 4005 | lp->lwp_vmspace = newvm; |
| 4006 | if (curthread->td_lwp == lp) { |
| 4007 | pmap = vmspace_pmap(newvm); |
| 4008 | #if defined(SMP) |
| 4009 | atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask); |
| 4010 | if (pmap->pm_active & CPUMASK_LOCK) |
| 4011 | pmap_interlock_wait(newvm); |
| 4012 | #else |
| 4013 | pmap->pm_active |= 1; |
| 4014 | #endif |
| 4015 | #if defined(SWTCH_OPTIM_STATS) |
| 4016 | tlb_flush_count++; |
| 4017 | #endif |
| 4018 | curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4); |
| 4019 | curthread->td_pcb->pcb_cr3 |= PG_RW | PG_U | PG_V; |
| 4020 | load_cr3(curthread->td_pcb->pcb_cr3); |
| 4021 | pmap = vmspace_pmap(oldvm); |
| 4022 | #if defined(SMP) |
| 4023 | atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask); |
| 4024 | #else |
| 4025 | pmap->pm_active &= ~(cpumask_t)1; |
| 4026 | #endif |
| 4027 | } |
| 4028 | crit_exit(); |
| 4029 | } |
| 4030 | } |
| 4031 | |
| 4032 | #ifdef SMP |
| 4033 | |
| 4034 | /* |
| 4035 | * Called when switching to a locked pmap, used to interlock against pmaps |
| 4036 | * undergoing modifications to prevent us from activating the MMU for the |
| 4037 | * target pmap until all such modifications have completed. We have to do |
| 4038 | * this because the thread making the modifications has already set up its |
| 4039 | * SMP synchronization mask. |
| 4040 | * |
| 4041 | * This function cannot sleep! |
| 4042 | * |
| 4043 | * No requirements. |
| 4044 | */ |
| 4045 | void |
| 4046 | pmap_interlock_wait(struct vmspace *vm) |
| 4047 | { |
| 4048 | struct pmap *pmap = &vm->vm_pmap; |
| 4049 | |
| 4050 | if (pmap->pm_active & CPUMASK_LOCK) { |
| 4051 | crit_enter(); |
| 4052 | KKASSERT(curthread->td_critcount >= 2); |
| 4053 | DEBUG_PUSH_INFO("pmap_interlock_wait"); |
| 4054 | while (pmap->pm_active & CPUMASK_LOCK) { |
| 4055 | cpu_ccfence(); |
| 4056 | lwkt_process_ipiq(); |
| 4057 | } |
| 4058 | DEBUG_POP_INFO(); |
| 4059 | crit_exit(); |
| 4060 | } |
| 4061 | } |
| 4062 | |
| 4063 | #endif |
| 4064 | |
| 4065 | vm_offset_t |
| 4066 | pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) |
| 4067 | { |
| 4068 | |
| 4069 | if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { |
| 4070 | return addr; |
| 4071 | } |
| 4072 | |
| 4073 | addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1); |
| 4074 | return addr; |
| 4075 | } |
| 4076 | |
| 4077 | /* |
| 4078 | * Used by kmalloc/kfree, page already exists at va |
| 4079 | */ |
| 4080 | vm_page_t |
| 4081 | pmap_kvtom(vm_offset_t va) |
| 4082 | { |
| 4083 | return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME)); |
| 4084 | } |