pmap.c revision 121823
162587Sitojun/*- 278064Sume * Copyright (c) 1991 Regents of the University of California. 362587Sitojun * All rights reserved. 4139826Simp * Copyright (c) 1994 John S. Dyson 552904Sshin * All rights reserved. 652904Sshin * Copyright (c) 1994 David Greenman 753541Sshin * All rights reserved. 852904Sshin * 952904Sshin * This code is derived from software contributed to Berkeley by 1052904Sshin * the Systems Programming Group of the University of Utah Computer 1152904Sshin * Science Department and William Jolitz of UUNET Technologies Inc. 1252904Sshin * 1352904Sshin * Redistribution and use in source and binary forms, with or without 1452904Sshin * modification, are permitted provided that the following conditions 1552904Sshin * are met: 1652904Sshin * 1. Redistributions of source code must retain the above copyright 1752904Sshin * notice, this list of conditions and the following disclaimer. 1852904Sshin * 2. Redistributions in binary form must reproduce the above copyright 1953541Sshin * notice, this list of conditions and the following disclaimer in the 2052904Sshin * documentation and/or other materials provided with the distribution. 2152904Sshin * 3. All advertising materials mentioning features or use of this software 2252904Sshin * must display the following acknowledgement: 2352904Sshin * This product includes software developed by the University of 2452904Sshin * California, Berkeley and its contributors. 2552904Sshin * 4. Neither the name of the University nor the names of its contributors 2652904Sshin * may be used to endorse or promote products derived from this software 2752904Sshin * without specific prior written permission. 2852904Sshin * 2952904Sshin * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 3052904Sshin * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 3152904Sshin * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 3252904Sshin * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 33139826Simp * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 3452904Sshin * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 3552904Sshin * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 3652904Sshin * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 3752904Sshin * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 3852904Sshin * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 3952904Sshin * SUCH DAMAGE. 4052904Sshin * 4152904Sshin * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 4252904Sshin */ 4352904Sshin/*- 4452904Sshin * Copyright (c) 2003 Networks Associates Technology, Inc. 4552904Sshin * All rights reserved. 4652904Sshin * 4752904Sshin * This software was developed for the FreeBSD Project by Jake Burkholder, 4852904Sshin * Safeport Network Services, and Network Associates Laboratories, the 4952904Sshin * Security Research Division of Network Associates, Inc. under 5052904Sshin * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA 5152904Sshin * CHATS research program. 5252904Sshin * 5352904Sshin * Redistribution and use in source and binary forms, with or without 5452904Sshin * modification, are permitted provided that the following conditions 5552904Sshin * are met: 5652904Sshin * 1. Redistributions of source code must retain the above copyright 5752904Sshin * notice, this list of conditions and the following disclaimer. 5852904Sshin * 2. Redistributions in binary form must reproduce the above copyright 5952904Sshin * notice, this list of conditions and the following disclaimer in the 6052904Sshin * documentation and/or other materials provided with the distribution. 6152904Sshin * 6252904Sshin * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 6352904Sshin * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 6462587Sitojun * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 6578064Sume * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 6657120Sshin * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 6757120Sshin * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 6862587Sitojun * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 6962587Sitojun * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 7062587Sitojun * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 7152904Sshin * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 7252904Sshin * SUCH DAMAGE. 7378064Sume */ 7478064Sume 7552904Sshin#include <sys/cdefs.h> 7662587Sitojun__FBSDID("$FreeBSD: head/sys/i386/i386/pmap.c 121823 2003-10-31 21:02:04Z jhb $"); 7778064Sume 7852904Sshin/* 7952904Sshin * Manages physical address maps. 8052904Sshin * 8152904Sshin * In addition to hardware address maps, this 8252904Sshin * module is called upon to provide software-use-only 8352904Sshin * maps which may or may not be stored in the same 8452904Sshin * form as hardware maps. These pseudo-maps are 8552904Sshin * used to store intermediate results from copy 8652904Sshin * operations to and from address spaces. 8752904Sshin * 8862587Sitojun * Since the information managed by this module is 8952904Sshin * also stored by the logical address mapping module, 9052904Sshin * this module may throw away valid virtual-to-physical 9152904Sshin * mappings at almost any time. However, invalidations 9252904Sshin * of virtual-to-physical mappings must be done as 9352904Sshin * requested. 9452904Sshin * 9552904Sshin * In order to cope with hardware architectures which 9652904Sshin * make virtual-to-physical map invalidates expensive, 9752904Sshin * this module may delay invalidate or reduced protection 9852904Sshin * operations until such time as they are actually 9952904Sshin * necessary. This module is given full information as 10052904Sshin * to which processors are currently using which maps, 10152904Sshin * and to when physical maps must be made correct. 10252904Sshin */ 10352904Sshin 10452904Sshin#include "opt_pmap.h" 10552904Sshin#include "opt_msgbuf.h" 10697181Smike#include "opt_kstack_pages.h" 10752904Sshin 10852904Sshin#include <sys/param.h> 10952904Sshin#include <sys/systm.h> 11052904Sshin#include <sys/kernel.h> 11152904Sshin#include <sys/lock.h> 11297181Smike#include <sys/mman.h> 11352904Sshin#include <sys/msgbuf.h> 11452904Sshin#include <sys/mutex.h> 11552904Sshin#include <sys/proc.h> 11652904Sshin#include <sys/sx.h> 11752904Sshin#include <sys/user.h> 11852904Sshin#include <sys/vmmeter.h> 11997181Smike#include <sys/sysctl.h> 12097181Smike#ifdef SMP 12197181Smike#include <sys/smp.h> 12252904Sshin#endif 12352904Sshin 12452904Sshin#include <vm/vm.h> 12562587Sitojun#include <vm/vm_param.h> 12695023Ssuz#include <vm/vm_kern.h> 12762587Sitojun#include <vm/vm_page.h> 12862587Sitojun#include <vm/vm_map.h> 12962587Sitojun#include <vm/vm_object.h> 13052904Sshin#include <vm/vm_extern.h> 13152904Sshin#include <vm/vm_pageout.h> 13262587Sitojun#include <vm/vm_pager.h> 13352904Sshin#include <vm/uma.h> 13452904Sshin 13597181Smike#include <machine/cpu.h> 13697181Smike#include <machine/cputypes.h> 13797181Smike#include <machine/md_var.h> 13897181Smike#include <machine/specialreg.h> 13952904Sshin#if defined(SMP) || defined(APIC_IO) 14052904Sshin#include <machine/smp.h> 14197181Smike#include <machine/apic.h> 14262587Sitojun#include <machine/segments.h> 14352904Sshin#include <machine/tss.h> 14497181Smike#endif /* SMP || APIC_IO */ 14552904Sshin 146100503Sume#define PMAP_KEEP_PDIRS 14797181Smike#ifndef PMAP_SHPGPERPROC 14897181Smike#define PMAP_SHPGPERPROC 200 14997181Smike#endif 15062587Sitojun 15197181Smike#if defined(DIAGNOSTIC) 15252904Sshin#define PMAP_DIAGNOSTIC 15352904Sshin#endif 15452904Sshin 15552904Sshin#define MINPV 2048 15652904Sshin 15795023Ssuz#if !defined(PMAP_DIAGNOSTIC) 15862587Sitojun#define PMAP_INLINE __inline 15962587Sitojun#else 16052904Sshin#define PMAP_INLINE 16162587Sitojun#endif 16252904Sshin 16362587Sitojun/* 16452904Sshin * Get PDEs and PTEs for user/kernel address space 16562587Sitojun */ 16652904Sshin#define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT])) 16752904Sshin#define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) 16852904Sshin 16952904Sshin#define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0) 17078064Sume#define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0) 17178064Sume#define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0) 17252904Sshin#define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0) 17352904Sshin#define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0) 17452904Sshin 17552904Sshin#define pmap_pte_set_w(pte, v) ((v)?(*(int *)pte |= PG_W):(*(int *)pte &= ~PG_W)) 17652904Sshin#define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v))) 17752904Sshin 17852904Sshin/* 17952904Sshin * Given a map and a machine independent protection code, 18052904Sshin * convert to a vax protection code. 18152904Sshin */ 18295023Ssuz#define pte_prot(m, p) (protection_codes[p]) 18397181Smikestatic int protection_codes[8]; 18462587Sitojun 18562587Sitojunstruct pmap kernel_pmap_store; 18662587SitojunLIST_HEAD(pmaplist, pmap); 18762587Sitojunstatic struct pmaplist allpmaps; 18862587Sitojunstatic struct mtx allpmaps_lock; 18962587Sitojun#ifdef SMP 19062587Sitojunstatic struct mtx lazypmap_lock; 19162587Sitojun#endif 19297181Smike 19362587Sitojunvm_paddr_t avail_start; /* PA of first available physical page */ 19462587Sitojunvm_paddr_t avail_end; /* PA of last available physical page */ 19562587Sitojunvm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ 19662587Sitojunvm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ 19762587Sitojunstatic boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */ 19862587Sitojunint pgeflag = 0; /* PG_G or-in */ 19962587Sitojunint pseflag = 0; /* PG_PS or-in */ 20062587Sitojun 20152904Sshinstatic int nkpt; 20252904Sshinvm_offset_t kernel_vm_end; 20352904Sshinextern u_int32_t KERNend; 20452904Sshin 20552904Sshin#ifdef PAE 20652904Sshinstatic uma_zone_t pdptzone; 20797181Smike#endif 20862587Sitojun 20952904Sshin/* 21052904Sshin * Data for the pv entry allocation mechanism 21162587Sitojun */ 21252904Sshinstatic uma_zone_t pvzone; 21352904Sshinstatic struct vm_object pvzone_obj; 21462587Sitojunstatic int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; 21552904Sshinint pmap_pagedaemon_waken; 21652904Sshin 217121315Sume/* 218121315Sume * All those kernel PT submaps that BSD is so fond of 219121315Sume */ 22062587Sitojunpt_entry_t *CMAP1 = 0; 22152904Sshinstatic pt_entry_t *CMAP2, *CMAP3, *ptmmap; 22252904Sshincaddr_t CADDR1 = 0, ptvmmap = 0; 22362587Sitojunstatic caddr_t CADDR2, CADDR3; 22452904Sshinstatic struct mtx CMAPCADDR12_lock; 22552904Sshinstatic pt_entry_t *msgbufmap; 22697181Smikestruct msgbuf *msgbufp = 0; 22752904Sshin 22852904Sshin/* 22952904Sshin * Crashdump maps. 23097181Smike */ 23152904Sshinstatic pt_entry_t *pt_crashdumpmap; 23252904Sshinstatic caddr_t crashdumpmap; 23352904Sshin 23497181Smike#ifdef SMP 23552904Sshinextern pt_entry_t *SMPpt; 23652904Sshin#endif 23752904Sshinstatic pt_entry_t *PMAP1 = 0; 23853877Sitojunstatic pt_entry_t *PADDR1 = 0; 23953877Sitojun 24053877Sitojunstatic PMAP_INLINE void free_pv_entry(pv_entry_t pv); 24152904Sshinstatic pv_entry_t get_pv_entry(void); 24253877Sitojunstatic void i386_protection_init(void); 24362587Sitojunstatic void pmap_clear_ptes(vm_page_t m, int bit) 24462587Sitojun __always_inline; 24553877Sitojun 24697181Smikestatic int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva); 24762587Sitojunstatic void pmap_remove_page(struct pmap *pmap, vm_offset_t va); 24862587Sitojunstatic int pmap_remove_entry(struct pmap *pmap, vm_page_t m, 24953877Sitojun vm_offset_t va); 25097181Smikestatic void pmap_insert_entry(pmap_t pmap, vm_offset_t va, 25152904Sshin vm_page_t mpte, vm_page_t m); 25278064Sume 25378064Sumestatic vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va); 25478064Sume 25578064Sumestatic vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex); 25678064Sumestatic int pmap_unuse_pt(pmap_t, vm_offset_t, vm_page_t); 25778064Sumestatic vm_offset_t pmap_kmem_choose(vm_offset_t addr); 25878064Sumestatic void *pmap_pv_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 25978064Sume#ifdef PAE 26052904Sshinstatic void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait); 26152904Sshin#endif 26252904Sshin 26362587SitojunCTASSERT(1 << PDESHIFT == sizeof(pd_entry_t)); 26478064SumeCTASSERT(1 << PTESHIFT == sizeof(pt_entry_t)); 26578064Sume 26678064Sume/* 26778064Sume * Move the kernel virtual free pointer to the next 26852904Sshin * 4MB. This is used to help improve performance 26952904Sshin * by using a large (4MB) page for much of the kernel 27052904Sshin * (.text, .data, .bss) 27152904Sshin */ 27262587Sitojunstatic vm_offset_t 27378064Sumepmap_kmem_choose(vm_offset_t addr) 27478064Sume{ 27578064Sume vm_offset_t newaddr = addr; 27678064Sume 27752904Sshin#ifndef DISABLE_PSE 27852904Sshin if (cpu_feature & CPUID_PSE) 27952904Sshin newaddr = (addr + PDRMASK) & ~PDRMASK; 28052904Sshin#endif 28162587Sitojun return newaddr; 28278064Sume} 28378064Sume 28478064Sume/* 28578064Sume * Bootstrap the system enough to run with virtual memory. 28678064Sume * 28752904Sshin * On the i386 this is called after mapping has already been enabled 28852904Sshin * and just syncs the pmap module with what has already been done. 28952904Sshin * [We can't call it easily with mapping off since the kernel is not 29052904Sshin * mapped with PA == VA, hence we would have to relocate every address 29162587Sitojun * from the linked base (virtual) address "KERNBASE" to the actual 29278064Sume * (physical) address starting relative to 0] 29378064Sume */ 29478064Sumevoid 29552904Sshinpmap_bootstrap(firstaddr, loadaddr) 29652904Sshin vm_paddr_t firstaddr; 29752904Sshin vm_paddr_t loadaddr; 29852904Sshin{ 29952904Sshin vm_offset_t va; 30095023Ssuz pt_entry_t *pte; 30162587Sitojun int i; 302121315Sume 30362587Sitojun avail_start = firstaddr; 30462587Sitojun 30562587Sitojun /* 30662587Sitojun * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too 30752904Sshin * large. It should instead be correctly calculated in locore.s and 30862587Sitojun * not based on 'first' (which is a physical address, not a virtual 309121315Sume * address, for the start of unused physical memory). The kernel 31062587Sitojun * page tables are NOT double mapped and thus should not be included 31162587Sitojun * in this calculation. 31262587Sitojun */ 31362587Sitojun virtual_avail = (vm_offset_t) KERNBASE + firstaddr; 31452904Sshin virtual_avail = pmap_kmem_choose(virtual_avail); 31552904Sshin 31652904Sshin virtual_end = VM_MAX_KERNEL_ADDRESS; 31752904Sshin 31852904Sshin /* 31952904Sshin * Initialize protection array. 32062587Sitojun */ 32152904Sshin i386_protection_init(); 32262587Sitojun 32352904Sshin /* 32452904Sshin * Initialize the kernel pmap (which is statically allocated). 32552904Sshin */ 32652904Sshin kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); 32752904Sshin#ifdef PAE 32852904Sshin kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT); 32952904Sshin#endif 33095023Ssuz kernel_pmap->pm_active = -1; /* don't allow deactivation */ 33162587Sitojun TAILQ_INIT(&kernel_pmap->pm_pvlist); 33252904Sshin LIST_INIT(&allpmaps); 33362587Sitojun#ifdef SMP 33452904Sshin mtx_init(&lazypmap_lock, "lazypmap", NULL, MTX_SPIN); 33552904Sshin#endif 33652904Sshin mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN); 33752904Sshin mtx_lock_spin(&allpmaps_lock); 33852904Sshin LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); 33995023Ssuz mtx_unlock_spin(&allpmaps_lock); 34062587Sitojun nkpt = NKPT; 34152904Sshin 34252904Sshin /* 343121315Sume * Reserve some special page table entries/VA space for temporary 344121315Sume * mapping of pages. 345121315Sume */ 34662587Sitojun#define SYSMAP(c, p, v, n) \ 34752904Sshin v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); 34852904Sshin 34962587Sitojun va = virtual_avail; 35052904Sshin pte = vtopte(va); 35152904Sshin 35262587Sitojun /* 35352904Sshin * CMAP1/CMAP2 are used for zeroing and copying pages. 35452904Sshin * CMAP3 is used for the idle process page zeroing. 35562587Sitojun */ 35652904Sshin SYSMAP(caddr_t, CMAP1, CADDR1, 1) 35752904Sshin SYSMAP(caddr_t, CMAP2, CADDR2, 1) 35852904Sshin SYSMAP(caddr_t, CMAP3, CADDR3, 1) 35962587Sitojun *CMAP3 = 0; 36052904Sshin 36152904Sshin mtx_init(&CMAPCADDR12_lock, "CMAPCADDR12", NULL, MTX_DEF); 36262587Sitojun 36352904Sshin /* 36452904Sshin * Crashdump maps. 36562587Sitojun */ 36652904Sshin SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS); 36752904Sshin 36862587Sitojun /* 36952904Sshin * ptvmmap is used for reading arbitrary physical pages via /dev/mem. 37052904Sshin * XXX ptmmap is not used. 37162587Sitojun */ 37252904Sshin SYSMAP(caddr_t, ptmmap, ptvmmap, 1) 37352904Sshin 37452904Sshin /* 37552904Sshin * msgbufp is used to map the system message buffer. 37695023Ssuz * XXX msgbufmap is not used. 37778064Sume */ 37852904Sshin SYSMAP(struct msgbuf *, msgbufmap, msgbufp, 37952904Sshin atop(round_page(MSGBUF_SIZE))) 38062587Sitojun 38152904Sshin /* 38252904Sshin * ptemap is used for pmap_pte_quick 38352904Sshin */ 38478064Sume SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1); 38578064Sume 38678064Sume virtual_avail = va; 38778064Sume 38878064Sume *CMAP1 = *CMAP2 = 0; 38978064Sume for (i = 0; i < NKPT; i++) 39078064Sume PTD[i] = 0; 39178064Sume 39252904Sshin /* Turn on PG_G on kernel page(s) */ 39352904Sshin pmap_set_pg(); 39452904Sshin 39597181Smike#ifdef SMP 39652904Sshin if (cpu_apic_address == 0) 39752904Sshin panic("pmap_bootstrap: no local apic! (non-SMP hardware?)"); 39852904Sshin 39952904Sshin /* local apic is mapped on last page */ 40052904Sshin SMPpt[NPTEPG - 1] = (pt_entry_t)(PG_V | PG_RW | PG_N | pgeflag | 40152904Sshin (cpu_apic_address & PG_FRAME)); 40252904Sshin#endif 40352904Sshin invltlb(); 40452904Sshin} 40552904Sshin 40652904Sshin/* 40762587Sitojun * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on. 40862587Sitojun */ 40962587Sitojunvoid 41062587Sitojunpmap_set_pg(void) 41162587Sitojun{ 41262587Sitojun pd_entry_t pdir; 41362587Sitojun pt_entry_t *pte; 41462587Sitojun vm_offset_t va, endva; 415121472Sume int i; 416121472Sume 417121472Sume if (pgeflag == 0) 418121472Sume return; 419121472Sume 420121472Sume i = KERNLOAD/NBPDR; 42162587Sitojun endva = KERNBASE + KERNend; 42262587Sitojun 42378064Sume if (pseflag) { 424121472Sume va = KERNBASE + KERNLOAD; 425121472Sume while (va < endva) { 426121472Sume pdir = kernel_pmap->pm_pdir[KPTDI+i]; 427121472Sume pdir |= pgeflag; 428121472Sume kernel_pmap->pm_pdir[KPTDI+i] = PTD[KPTDI+i] = pdir; 429121472Sume invltlb(); /* Play it safe, invltlb() every time */ 430121472Sume i++; 431121472Sume va += NBPDR; 432121472Sume } 43378064Sume } else { 43462587Sitojun va = (vm_offset_t)btext; 435121472Sume while (va < endva) { 43678064Sume pte = vtopte(va); 43778064Sume if (*pte) 43878064Sume *pte |= pgeflag; 43952904Sshin invltlb(); /* Play it safe, invltlb() every time */ 44095023Ssuz va += PAGE_SIZE; 44162587Sitojun } 44278064Sume } 44362587Sitojun} 44452904Sshin 44595023Ssuzstatic void * 44662587Sitojunpmap_pv_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 44762587Sitojun{ 44862587Sitojun *flags = UMA_SLAB_PRIV; 44962587Sitojun return (void *)kmem_alloc(kernel_map, bytes); 45062587Sitojun} 45178064Sume 45278064Sume#ifdef PAE 453121472Sumestatic void * 454121472Sumepmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 455121472Sume{ 456121472Sume *flags = UMA_SLAB_PRIV; 457121472Sume return (contigmalloc(PAGE_SIZE, NULL, 0, 0x0ULL, 0xffffffffULL, 1, 0)); 458121472Sume} 459121472Sume#endif 460121472Sume 461121472Sume/* 462121472Sume * Initialize the pmap module. 463121472Sume * Called by vm_init, to initialize any structures that the pmap 464121472Sume * system needs to map virtual memory. 465121472Sume * pmap_init has been enhanced to support in a fairly consistant 466121472Sume * way, discontiguous physical memory. 467121472Sume */ 468121472Sumevoid 469121472Sumepmap_init(phys_start, phys_end) 470121472Sume vm_paddr_t phys_start, phys_end; 471121472Sume{ 472121472Sume int i; 473121472Sume int initial_pvs; 474121472Sume 475121472Sume /* 476121472Sume * Allocate memory for random pmap data structures. Includes the 477121472Sume * pv_head_table. 478121472Sume */ 479121472Sume 480121472Sume for(i = 0; i < vm_page_array_size; i++) { 481121472Sume vm_page_t m; 482121472Sume 483121472Sume m = &vm_page_array[i]; 484121472Sume TAILQ_INIT(&m->md.pv_list); 485121472Sume m->md.pv_list_count = 0; 486121472Sume } 487121472Sume 488121472Sume /* 489121472Sume * init the pv free list 490121472Sume */ 491121472Sume initial_pvs = vm_page_array_size; 492121472Sume if (initial_pvs < MINPV) 493121472Sume initial_pvs = MINPV; 494121472Sume pvzone = uma_zcreate("PV ENTRY", sizeof (struct pv_entry), NULL, NULL, 495121472Sume NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE); 496121472Sume uma_zone_set_allocf(pvzone, pmap_pv_allocf); 497121472Sume uma_prealloc(pvzone, initial_pvs); 49865124Sitojun 49952904Sshin#ifdef PAE 50062587Sitojun pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL, 50162587Sitojun NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1, 50262587Sitojun UMA_ZONE_VM | UMA_ZONE_NOFREE); 50352904Sshin uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf); 50452904Sshin#endif 50552904Sshin 50652904Sshin /* 50795023Ssuz * Now it is safe to enable pv_table recording. 50895023Ssuz */ 50952904Sshin pmap_initialized = TRUE; 51052904Sshin} 51152904Sshin 51252904Sshin/* 51352904Sshin * Initialize the address space (zone) for the pv_entries. Set a 51462587Sitojun * high water mark so that the system can recover from excessive 51562587Sitojun * numbers of pv entries. 51652904Sshin */ 51752904Sshinvoid 51852904Sshinpmap_init2() 51952904Sshin{ 52052904Sshin int shpgperproc = PMAP_SHPGPERPROC; 52152904Sshin 52262587Sitojun TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); 52362587Sitojun pv_entry_max = shpgperproc * maxproc + vm_page_array_size; 52452904Sshin TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); 52552904Sshin pv_entry_high_water = 9 * (pv_entry_max / 10); 52652904Sshin uma_zone_set_obj(pvzone, &pvzone_obj, pv_entry_max); 527121472Sume} 528121472Sume 529121472Sume 530121472Sume/*************************************************** 531121569Sume * Low level helper routines..... 532121472Sume ***************************************************/ 533121472Sume 534121472Sume#if defined(PMAP_DIAGNOSTIC) 53552904Sshin 53652904Sshin/* 53752904Sshin * This code checks for non-writeable/modified pages. 53852904Sshin * This should be an invalid condition. 53952904Sshin */ 54052904Sshinstatic int 54152904Sshinpmap_nw_modified(pt_entry_t ptea) 542121472Sume{ 54352904Sshin int pte; 54452904Sshin 54552904Sshin pte = (int) ptea; 54652904Sshin 54752904Sshin if ((pte & (PG_M|PG_RW)) == PG_M) 54852904Sshin return 1; 54962587Sitojun else 55052904Sshin return 0; 55152904Sshin} 55252904Sshin#endif 55352904Sshin 55462587Sitojun 55562587Sitojun/* 55662587Sitojun * this routine defines the region(s) of memory that should 55752904Sshin * not be tested for the modified bit. 55862587Sitojun */ 55952904Sshinstatic PMAP_INLINE int 56062587Sitojunpmap_track_modified(vm_offset_t va) 56162587Sitojun{ 56262587Sitojun if ((va < kmi.clean_sva) || (va >= kmi.clean_eva)) 56362587Sitojun return 1; 56462587Sitojun else 56562587Sitojun return 0; 56662587Sitojun} 56762587Sitojun 56862587Sitojun#ifdef I386_CPU 56962587Sitojun/* 57062587Sitojun * i386 only has "invalidate everything" and no SMP to worry about. 57162587Sitojun */ 57262587SitojunPMAP_INLINE void 57362587Sitojunpmap_invalidate_page(pmap_t pmap, vm_offset_t va) 57462587Sitojun{ 57562587Sitojun 57662587Sitojun if (pmap == kernel_pmap || pmap->pm_active) 57762587Sitojun invltlb(); 57895023Ssuz} 57962587Sitojun 58062604SitojunPMAP_INLINE void 58178064Sumepmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 58262604Sitojun{ 58362604Sitojun 58462604Sitojun if (pmap == kernel_pmap || pmap->pm_active) 58578064Sume invltlb(); 58678064Sume} 58778064Sume 58878064SumePMAP_INLINE void 58978064Sumepmap_invalidate_all(pmap_t pmap) 59078064Sume{ 591122077Sume 592121742Sume if (pmap == kernel_pmap || pmap->pm_active) 593121742Sume invltlb(); 594121345Sume} 595121345Sume#else /* !I386_CPU */ 59652904Sshin#ifdef SMP 59765124Sitojun/* 598121345Sume * For SMP, these functions have to use the IPI mechanism for coherence. 599121472Sume */ 60078064Sumevoid 60153541Sshinpmap_invalidate_page(pmap_t pmap, vm_offset_t va) 60253541Sshin{ 60353541Sshin u_int cpumask; 60478064Sume u_int other_cpus; 60578064Sume 60678064Sume critical_enter(); 60778064Sume /* 60853541Sshin * We need to disable interrupt preemption but MUST NOT have 60952904Sshin * interrupts disabled here. 61078064Sume * XXX we may need to hold schedlock to get a coherent pm_active 61152904Sshin */ 61262587Sitojun if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 61352904Sshin invlpg(va); 61452904Sshin smp_invlpg(va); 61552904Sshin } else { 61678064Sume cpumask = PCPU_GET(cpumask); 61778064Sume other_cpus = PCPU_GET(other_cpus); 61892700Sdarrenr if (pmap->pm_active & cpumask) 61952904Sshin invlpg(va); 62052904Sshin if (pmap->pm_active & other_cpus) 62152904Sshin smp_masked_invlpg(pmap->pm_active & other_cpus, va); 62252904Sshin } 62352904Sshin critical_exit(); 62452904Sshin} 62552904Sshin 626121742Sumevoid 62752904Sshinpmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 62878064Sume{ 62978064Sume u_int cpumask; 63078064Sume u_int other_cpus; 63183934Sbrooks vm_offset_t addr; 63283934Sbrooks 63352904Sshin critical_enter(); 63452904Sshin /* 635102227Smike * We need to disable interrupt preemption but MUST NOT have 636102227Smike * interrupts disabled here. 637102227Smike * XXX we may need to hold schedlock to get a coherent pm_active 63897181Smike */ 63997181Smike if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 640121498Sume for (addr = sva; addr < eva; addr += PAGE_SIZE) 641121498Sume invlpg(addr); 642121498Sume smp_invlpg_range(sva, eva); 643121498Sume } else { 644121498Sume cpumask = PCPU_GET(cpumask); 645121472Sume other_cpus = PCPU_GET(other_cpus); 646121472Sume if (pmap->pm_active & cpumask) 64752904Sshin for (addr = sva; addr < eva; addr += PAGE_SIZE) 64852904Sshin invlpg(addr); 64952904Sshin if (pmap->pm_active & other_cpus) 65078064Sume smp_masked_invlpg_range(pmap->pm_active & other_cpus, 65178064Sume sva, eva); 65297181Smike } 65378064Sume critical_exit(); 654121499Sume} 655121499Sume 656121499Sumevoid 65752904Sshinpmap_invalidate_all(pmap_t pmap) 65878064Sume{ 65978064Sume u_int cpumask; 66078064Sume u_int other_cpus; 661121472Sume 66278064Sume critical_enter(); 66378064Sume /* 66478064Sume * We need to disable interrupt preemption but MUST NOT have 66578064Sume * interrupts disabled here. 66678064Sume * XXX we may need to hold schedlock to get a coherent pm_active 66778064Sume */ 66878064Sume if (pmap == kernel_pmap || pmap->pm_active == all_cpus) { 66978064Sume invltlb(); 670121472Sume smp_invltlb(); 671121499Sume } else { 672121499Sume cpumask = PCPU_GET(cpumask); 673121472Sume other_cpus = PCPU_GET(other_cpus); 674121472Sume if (pmap->pm_active & cpumask) 67578064Sume invltlb(); 676121499Sume if (pmap->pm_active & other_cpus) 677121472Sume smp_masked_invltlb(pmap->pm_active & other_cpus); 678121499Sume } 679121472Sume critical_exit(); 680121472Sume} 681121472Sume#else /* !SMP */ 682121472Sume/* 68378064Sume * Normal, non-SMP, 486+ invalidation functions. 68478064Sume * We inline these within pmap.c for speed. 68578064Sume */ 68678064SumePMAP_INLINE void 68752904Sshinpmap_invalidate_page(pmap_t pmap, vm_offset_t va) 68852904Sshin{ 68997181Smike 69097181Smike if (pmap == kernel_pmap || pmap->pm_active) 69152904Sshin invlpg(va); 692} 693 694PMAP_INLINE void 695pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 696{ 697 vm_offset_t addr; 698 699 if (pmap == kernel_pmap || pmap->pm_active) 700 for (addr = sva; addr < eva; addr += PAGE_SIZE) 701 invlpg(addr); 702} 703 704PMAP_INLINE void 705pmap_invalidate_all(pmap_t pmap) 706{ 707 708 if (pmap == kernel_pmap || pmap->pm_active) 709 invltlb(); 710} 711#endif /* !SMP */ 712#endif /* !I386_CPU */ 713 714/* 715 * Are we current address space or kernel? N.B. We return FALSE when 716 * a pmap's page table is in use because a kernel thread is borrowing 717 * it. The borrowed page table can change spontaneously, making any 718 * dependence on its continued use subject to a race condition. 719 */ 720static int 721pmap_is_current(pmap_t pmap) 722{ 723 return (pmap == kernel_pmap || 724 (pmap == vmspace_pmap(curthread->td_proc->p_vmspace) && 725 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME))); 726} 727 728/* 729 * Super fast pmap_pte routine best used when scanning 730 * the pv lists. This eliminates many coarse-grained 731 * invltlb calls. Note that many of the pv list 732 * scans are across different pmaps. It is very wasteful 733 * to do an entire invltlb for checking a single mapping. 734 */ 735pt_entry_t * 736pmap_pte_quick(pmap, va) 737 register pmap_t pmap; 738 vm_offset_t va; 739{ 740 pd_entry_t newpf; 741 pd_entry_t *pde; 742 743 pde = pmap_pde(pmap, va); 744 if (*pde & PG_PS) 745 return (pde); 746 if (*pde != 0) { 747 /* are we current address space or kernel? */ 748 if (pmap_is_current(pmap)) 749 return vtopte(va); 750 newpf = *pde & PG_FRAME; 751 if (((*PMAP1) & PG_FRAME) != newpf) { 752 *PMAP1 = newpf | PG_RW | PG_V; 753 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR1); 754 } 755 return PADDR1 + (i386_btop(va) & (NPTEPG - 1)); 756 } 757 return (0); 758} 759 760/* 761 * Routine: pmap_extract 762 * Function: 763 * Extract the physical page address associated 764 * with the given map/virtual_address pair. 765 */ 766vm_paddr_t 767pmap_extract(pmap, va) 768 register pmap_t pmap; 769 vm_offset_t va; 770{ 771 vm_paddr_t rtval; 772 pt_entry_t *pte; 773 pd_entry_t pde; 774 775 if (pmap == 0) 776 return 0; 777 pde = pmap->pm_pdir[va >> PDRSHIFT]; 778 if (pde != 0) { 779 if ((pde & PG_PS) != 0) { 780 rtval = (pde & ~PDRMASK) | (va & PDRMASK); 781 return rtval; 782 } 783 pte = pmap_pte_quick(pmap, va); 784 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK)); 785 return rtval; 786 } 787 return 0; 788 789} 790 791/* 792 * Routine: pmap_extract_and_hold 793 * Function: 794 * Atomically extract and hold the physical page 795 * with the given pmap and virtual address pair 796 * if that mapping permits the given protection. 797 */ 798vm_page_t 799pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) 800{ 801 vm_paddr_t pa; 802 vm_page_t m; 803 804 m = NULL; 805 mtx_lock(&Giant); 806 if ((pa = pmap_extract(pmap, va)) != 0) { 807 m = PHYS_TO_VM_PAGE(pa); 808 vm_page_lock_queues(); 809 vm_page_hold(m); 810 vm_page_unlock_queues(); 811 } 812 mtx_unlock(&Giant); 813 return (m); 814} 815 816/*************************************************** 817 * Low level mapping routines..... 818 ***************************************************/ 819 820/* 821 * Add a wired page to the kva. 822 * Note: not SMP coherent. 823 */ 824PMAP_INLINE void 825pmap_kenter(vm_offset_t va, vm_paddr_t pa) 826{ 827 pt_entry_t *pte; 828 829 pte = vtopte(va); 830 pte_store(pte, pa | PG_RW | PG_V | pgeflag); 831} 832 833/* 834 * Remove a page from the kernel pagetables. 835 * Note: not SMP coherent. 836 */ 837PMAP_INLINE void 838pmap_kremove(vm_offset_t va) 839{ 840 pt_entry_t *pte; 841 842 pte = vtopte(va); 843 pte_clear(pte); 844} 845 846/* 847 * Used to map a range of physical addresses into kernel 848 * virtual address space. 849 * 850 * The value passed in '*virt' is a suggested virtual address for 851 * the mapping. Architectures which can support a direct-mapped 852 * physical to virtual region can return the appropriate address 853 * within that region, leaving '*virt' unchanged. Other 854 * architectures should map the pages starting at '*virt' and 855 * update '*virt' with the first usable address after the mapped 856 * region. 857 */ 858vm_offset_t 859pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) 860{ 861 vm_offset_t va, sva; 862 863 va = sva = *virt; 864 while (start < end) { 865 pmap_kenter(va, start); 866 va += PAGE_SIZE; 867 start += PAGE_SIZE; 868 } 869 pmap_invalidate_range(kernel_pmap, sva, va); 870 *virt = va; 871 return (sva); 872} 873 874 875/* 876 * Add a list of wired pages to the kva 877 * this routine is only used for temporary 878 * kernel mappings that do not need to have 879 * page modification or references recorded. 880 * Note that old mappings are simply written 881 * over. The page *must* be wired. 882 * Note: SMP coherent. Uses a ranged shootdown IPI. 883 */ 884void 885pmap_qenter(vm_offset_t sva, vm_page_t *m, int count) 886{ 887 vm_offset_t va; 888 889 va = sva; 890 while (count-- > 0) { 891 pmap_kenter(va, VM_PAGE_TO_PHYS(*m)); 892 va += PAGE_SIZE; 893 m++; 894 } 895 pmap_invalidate_range(kernel_pmap, sva, va); 896} 897 898/* 899 * This routine tears out page mappings from the 900 * kernel -- it is meant only for temporary mappings. 901 * Note: SMP coherent. Uses a ranged shootdown IPI. 902 */ 903void 904pmap_qremove(vm_offset_t sva, int count) 905{ 906 vm_offset_t va; 907 908 va = sva; 909 while (count-- > 0) { 910 pmap_kremove(va); 911 va += PAGE_SIZE; 912 } 913 pmap_invalidate_range(kernel_pmap, sva, va); 914} 915 916/*************************************************** 917 * Page table page management routines..... 918 ***************************************************/ 919 920/* 921 * This routine unholds page table pages, and if the hold count 922 * drops to zero, then it decrements the wire count. 923 */ 924static int 925_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 926{ 927 928 while (vm_page_sleep_if_busy(m, FALSE, "pmuwpt")) 929 vm_page_lock_queues(); 930 931 if (m->hold_count == 0) { 932 vm_offset_t pteva; 933 /* 934 * unmap the page table page 935 */ 936 pmap->pm_pdir[m->pindex] = 0; 937 --pmap->pm_stats.resident_count; 938 /* 939 * We never unwire a kernel page table page, making a 940 * check for the kernel_pmap unnecessary. 941 */ 942 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME)) { 943 /* 944 * Do an invltlb to make the invalidated mapping 945 * take effect immediately. 946 */ 947 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); 948 pmap_invalidate_page(pmap, pteva); 949 } 950 951 /* 952 * If the page is finally unwired, simply free it. 953 */ 954 --m->wire_count; 955 if (m->wire_count == 0) { 956 vm_page_busy(m); 957 vm_page_free_zero(m); 958 atomic_subtract_int(&cnt.v_wire_count, 1); 959 } 960 return 1; 961 } 962 return 0; 963} 964 965static PMAP_INLINE int 966pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m) 967{ 968 vm_page_unhold(m); 969 if (m->hold_count == 0) 970 return _pmap_unwire_pte_hold(pmap, m); 971 else 972 return 0; 973} 974 975/* 976 * After removing a page table entry, this routine is used to 977 * conditionally free the page, and manage the hold/wire counts. 978 */ 979static int 980pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte) 981{ 982 983 if (va >= VM_MAXUSER_ADDRESS) 984 return 0; 985 986 return pmap_unwire_pte_hold(pmap, mpte); 987} 988 989void 990pmap_pinit0(pmap) 991 struct pmap *pmap; 992{ 993 994 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD); 995#ifdef PAE 996 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT); 997#endif 998 pmap->pm_active = 0; 999 PCPU_SET(curpmap, pmap); 1000 TAILQ_INIT(&pmap->pm_pvlist); 1001 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1002 mtx_lock_spin(&allpmaps_lock); 1003 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1004 mtx_unlock_spin(&allpmaps_lock); 1005} 1006 1007/* 1008 * Initialize a preallocated and zeroed pmap structure, 1009 * such as one in a vmspace structure. 1010 */ 1011void 1012pmap_pinit(pmap) 1013 register struct pmap *pmap; 1014{ 1015 vm_page_t m, ptdpg[NPGPTD]; 1016 vm_paddr_t pa; 1017 static int color; 1018 int i; 1019 1020 /* 1021 * No need to allocate page table space yet but we do need a valid 1022 * page directory table. 1023 */ 1024 if (pmap->pm_pdir == NULL) { 1025 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map, 1026 NBPTD); 1027#ifdef PAE 1028 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO); 1029 KASSERT(((vm_offset_t)pmap->pm_pdpt & 1030 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0, 1031 ("pmap_pinit: pdpt misaligned")); 1032 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30), 1033 ("pmap_pinit: pdpt above 4g")); 1034#endif 1035 } 1036 1037 /* 1038 * allocate the page directory page(s) 1039 */ 1040 for (i = 0; i < NPGPTD;) { 1041 m = vm_page_alloc(NULL, color++, 1042 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | 1043 VM_ALLOC_ZERO); 1044 if (m == NULL) 1045 VM_WAIT; 1046 else { 1047 vm_page_lock_queues(); 1048 vm_page_flag_clear(m, PG_BUSY); 1049 vm_page_unlock_queues(); 1050 ptdpg[i++] = m; 1051 } 1052 } 1053 1054 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD); 1055 1056 for (i = 0; i < NPGPTD; i++) { 1057 if ((ptdpg[i]->flags & PG_ZERO) == 0) 1058 bzero(pmap->pm_pdir + (i * NPDEPG), PAGE_SIZE); 1059 } 1060 1061 mtx_lock_spin(&allpmaps_lock); 1062 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); 1063 mtx_unlock_spin(&allpmaps_lock); 1064 /* Wire in kernel global address entries. */ 1065 /* XXX copies current process, does not fill in MPPTDI */ 1066 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t)); 1067#ifdef SMP 1068 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI]; 1069#endif 1070 1071 /* install self-referential address mapping entry(s) */ 1072 for (i = 0; i < NPGPTD; i++) { 1073 pa = VM_PAGE_TO_PHYS(ptdpg[i]); 1074 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M; 1075#ifdef PAE 1076 pmap->pm_pdpt[i] = pa | PG_V; 1077#endif 1078 } 1079 1080 pmap->pm_active = 0; 1081 TAILQ_INIT(&pmap->pm_pvlist); 1082 bzero(&pmap->pm_stats, sizeof pmap->pm_stats); 1083} 1084 1085/* 1086 * Wire in kernel global address entries. To avoid a race condition 1087 * between pmap initialization and pmap_growkernel, this procedure 1088 * should be called after the vmspace is attached to the process 1089 * but before this pmap is activated. 1090 */ 1091void 1092pmap_pinit2(pmap) 1093 struct pmap *pmap; 1094{ 1095 /* XXX: Remove this stub when no longer called */ 1096} 1097 1098/* 1099 * this routine is called if the page table page is not 1100 * mapped correctly. 1101 */ 1102static vm_page_t 1103_pmap_allocpte(pmap, ptepindex) 1104 pmap_t pmap; 1105 unsigned ptepindex; 1106{ 1107 vm_paddr_t ptepa; 1108 vm_page_t m; 1109 1110 /* 1111 * Allocate a page table page. 1112 */ 1113 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ | 1114 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { 1115 VM_WAIT; 1116 /* 1117 * Indicate the need to retry. While waiting, the page table 1118 * page may have been allocated. 1119 */ 1120 return (NULL); 1121 } 1122 if ((m->flags & PG_ZERO) == 0) 1123 pmap_zero_page(m); 1124 1125 KASSERT(m->queue == PQ_NONE, 1126 ("_pmap_allocpte: %p->queue != PQ_NONE", m)); 1127 1128 /* 1129 * Increment the hold count for the page table page 1130 * (denoting a new mapping.) 1131 */ 1132 m->hold_count++; 1133 1134 /* 1135 * Map the pagetable page into the process address space, if 1136 * it isn't already there. 1137 */ 1138 1139 pmap->pm_stats.resident_count++; 1140 1141 ptepa = VM_PAGE_TO_PHYS(m); 1142 pmap->pm_pdir[ptepindex] = 1143 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); 1144 1145 vm_page_lock_queues(); 1146 vm_page_flag_clear(m, PG_ZERO); 1147 vm_page_wakeup(m); 1148 vm_page_unlock_queues(); 1149 1150 return m; 1151} 1152 1153static vm_page_t 1154pmap_allocpte(pmap_t pmap, vm_offset_t va) 1155{ 1156 unsigned ptepindex; 1157 pd_entry_t ptepa; 1158 vm_page_t m; 1159 1160 /* 1161 * Calculate pagetable page index 1162 */ 1163 ptepindex = va >> PDRSHIFT; 1164retry: 1165 /* 1166 * Get the page directory entry 1167 */ 1168 ptepa = pmap->pm_pdir[ptepindex]; 1169 1170 /* 1171 * This supports switching from a 4MB page to a 1172 * normal 4K page. 1173 */ 1174 if (ptepa & PG_PS) { 1175 pmap->pm_pdir[ptepindex] = 0; 1176 ptepa = 0; 1177 pmap_invalidate_all(kernel_pmap); 1178 } 1179 1180 /* 1181 * If the page table page is mapped, we just increment the 1182 * hold count, and activate it. 1183 */ 1184 if (ptepa) { 1185 m = PHYS_TO_VM_PAGE(ptepa); 1186 m->hold_count++; 1187 } else { 1188 /* 1189 * Here if the pte page isn't mapped, or if it has 1190 * been deallocated. 1191 */ 1192 m = _pmap_allocpte(pmap, ptepindex); 1193 if (m == NULL) 1194 goto retry; 1195 } 1196 return (m); 1197} 1198 1199 1200/*************************************************** 1201* Pmap allocation/deallocation routines. 1202 ***************************************************/ 1203 1204#ifdef SMP 1205/* 1206 * Deal with a SMP shootdown of other users of the pmap that we are 1207 * trying to dispose of. This can be a bit hairy. 1208 */ 1209static u_int *lazymask; 1210static u_int lazyptd; 1211static volatile u_int lazywait; 1212 1213void pmap_lazyfix_action(void); 1214 1215void 1216pmap_lazyfix_action(void) 1217{ 1218 u_int mymask = PCPU_GET(cpumask); 1219 1220 if (rcr3() == lazyptd) 1221 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1222 atomic_clear_int(lazymask, mymask); 1223 atomic_store_rel_int(&lazywait, 1); 1224} 1225 1226static void 1227pmap_lazyfix_self(u_int mymask) 1228{ 1229 1230 if (rcr3() == lazyptd) 1231 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1232 atomic_clear_int(lazymask, mymask); 1233} 1234 1235 1236static void 1237pmap_lazyfix(pmap_t pmap) 1238{ 1239 u_int mymask = PCPU_GET(cpumask); 1240 u_int mask; 1241 register u_int spins; 1242 1243 while ((mask = pmap->pm_active) != 0) { 1244 spins = 50000000; 1245 mask = mask & -mask; /* Find least significant set bit */ 1246 mtx_lock_spin(&lazypmap_lock); 1247#ifdef PAE 1248 lazyptd = vtophys(pmap->pm_pdpt); 1249#else 1250 lazyptd = vtophys(pmap->pm_pdir); 1251#endif 1252 if (mask == mymask) { 1253 lazymask = &pmap->pm_active; 1254 pmap_lazyfix_self(mymask); 1255 } else { 1256 atomic_store_rel_int((u_int *)&lazymask, 1257 (u_int)&pmap->pm_active); 1258 atomic_store_rel_int(&lazywait, 0); 1259 ipi_selected(mask, IPI_LAZYPMAP); 1260 while (lazywait == 0) { 1261 ia32_pause(); 1262 if (--spins == 0) 1263 break; 1264 } 1265 } 1266 mtx_unlock_spin(&lazypmap_lock); 1267 if (spins == 0) 1268 printf("pmap_lazyfix: spun for 50000000\n"); 1269 } 1270} 1271 1272#else /* SMP */ 1273 1274/* 1275 * Cleaning up on uniprocessor is easy. For various reasons, we're 1276 * unlikely to have to even execute this code, including the fact 1277 * that the cleanup is deferred until the parent does a wait(2), which 1278 * means that another userland process has run. 1279 */ 1280static void 1281pmap_lazyfix(pmap_t pmap) 1282{ 1283 u_int cr3; 1284 1285 cr3 = vtophys(pmap->pm_pdir); 1286 if (cr3 == rcr3()) { 1287 load_cr3(PCPU_GET(curpcb)->pcb_cr3); 1288 pmap->pm_active &= ~(PCPU_GET(cpumask)); 1289 } 1290} 1291#endif /* SMP */ 1292 1293/* 1294 * Release any resources held by the given physical map. 1295 * Called when a pmap initialized by pmap_pinit is being released. 1296 * Should only be called if the map contains no valid mappings. 1297 */ 1298void 1299pmap_release(pmap_t pmap) 1300{ 1301 vm_page_t m, ptdpg[NPGPTD]; 1302 int i; 1303 1304 KASSERT(pmap->pm_stats.resident_count == 0, 1305 ("pmap_release: pmap resident count %ld != 0", 1306 pmap->pm_stats.resident_count)); 1307 1308 pmap_lazyfix(pmap); 1309 mtx_lock_spin(&allpmaps_lock); 1310 LIST_REMOVE(pmap, pm_list); 1311 mtx_unlock_spin(&allpmaps_lock); 1312 1313 for (i = 0; i < NPGPTD; i++) 1314 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i]); 1315 1316 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) * 1317 sizeof(*pmap->pm_pdir)); 1318#ifdef SMP 1319 pmap->pm_pdir[MPPTDI] = 0; 1320#endif 1321 1322 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD); 1323 1324 vm_page_lock_queues(); 1325 for (i = 0; i < NPGPTD; i++) { 1326 m = ptdpg[i]; 1327#ifdef PAE 1328 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME), 1329 ("pmap_release: got wrong ptd page")); 1330#endif 1331 m->wire_count--; 1332 atomic_subtract_int(&cnt.v_wire_count, 1); 1333 vm_page_busy(m); 1334 vm_page_free_zero(m); 1335 } 1336 vm_page_unlock_queues(); 1337} 1338 1339static int 1340kvm_size(SYSCTL_HANDLER_ARGS) 1341{ 1342 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; 1343 1344 return sysctl_handle_long(oidp, &ksize, 0, req); 1345} 1346SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 1347 0, 0, kvm_size, "IU", "Size of KVM"); 1348 1349static int 1350kvm_free(SYSCTL_HANDLER_ARGS) 1351{ 1352 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; 1353 1354 return sysctl_handle_long(oidp, &kfree, 0, req); 1355} 1356SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 1357 0, 0, kvm_free, "IU", "Amount of KVM free"); 1358 1359/* 1360 * grow the number of kernel page table entries, if needed 1361 */ 1362void 1363pmap_growkernel(vm_offset_t addr) 1364{ 1365 struct pmap *pmap; 1366 int s; 1367 vm_paddr_t ptppaddr; 1368 vm_page_t nkpg; 1369 pd_entry_t newpdir; 1370 pt_entry_t *pde; 1371 1372 s = splhigh(); 1373 mtx_assert(&kernel_map->system_mtx, MA_OWNED); 1374 if (kernel_vm_end == 0) { 1375 kernel_vm_end = KERNBASE; 1376 nkpt = 0; 1377 while (pdir_pde(PTD, kernel_vm_end)) { 1378 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1379 nkpt++; 1380 } 1381 } 1382 addr = roundup2(addr, PAGE_SIZE * NPTEPG); 1383 while (kernel_vm_end < addr) { 1384 if (pdir_pde(PTD, kernel_vm_end)) { 1385 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1386 continue; 1387 } 1388 1389 /* 1390 * This index is bogus, but out of the way 1391 */ 1392 nkpg = vm_page_alloc(NULL, nkpt, 1393 VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 1394 if (!nkpg) 1395 panic("pmap_growkernel: no memory to grow kernel"); 1396 1397 nkpt++; 1398 1399 pmap_zero_page(nkpg); 1400 ptppaddr = VM_PAGE_TO_PHYS(nkpg); 1401 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); 1402 pdir_pde(PTD, kernel_vm_end) = newpdir; 1403 1404 mtx_lock_spin(&allpmaps_lock); 1405 LIST_FOREACH(pmap, &allpmaps, pm_list) { 1406 pde = pmap_pde(pmap, kernel_vm_end); 1407 pde_store(pde, newpdir); 1408 } 1409 mtx_unlock_spin(&allpmaps_lock); 1410 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1); 1411 } 1412 splx(s); 1413} 1414 1415 1416/*************************************************** 1417 * page management routines. 1418 ***************************************************/ 1419 1420/* 1421 * free the pv_entry back to the free list 1422 */ 1423static PMAP_INLINE void 1424free_pv_entry(pv_entry_t pv) 1425{ 1426 pv_entry_count--; 1427 uma_zfree(pvzone, pv); 1428} 1429 1430/* 1431 * get a new pv_entry, allocating a block from the system 1432 * when needed. 1433 * the memory allocation is performed bypassing the malloc code 1434 * because of the possibility of allocations at interrupt time. 1435 */ 1436static pv_entry_t 1437get_pv_entry(void) 1438{ 1439 pv_entry_count++; 1440 if (pv_entry_high_water && 1441 (pv_entry_count > pv_entry_high_water) && 1442 (pmap_pagedaemon_waken == 0)) { 1443 pmap_pagedaemon_waken = 1; 1444 wakeup (&vm_pages_needed); 1445 } 1446 return uma_zalloc(pvzone, M_NOWAIT); 1447} 1448 1449/* 1450 * If it is the first entry on the list, it is actually 1451 * in the header and we must copy the following entry up 1452 * to the header. Otherwise we must search the list for 1453 * the entry. In either case we free the now unused entry. 1454 */ 1455 1456static int 1457pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) 1458{ 1459 pv_entry_t pv; 1460 int rtval; 1461 int s; 1462 1463 s = splvm(); 1464 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1465 if (m->md.pv_list_count < pmap->pm_stats.resident_count) { 1466 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 1467 if (pmap == pv->pv_pmap && va == pv->pv_va) 1468 break; 1469 } 1470 } else { 1471 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) { 1472 if (va == pv->pv_va) 1473 break; 1474 } 1475 } 1476 1477 rtval = 0; 1478 if (pv) { 1479 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem); 1480 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1481 m->md.pv_list_count--; 1482 if (TAILQ_FIRST(&m->md.pv_list) == NULL) 1483 vm_page_flag_clear(m, PG_WRITEABLE); 1484 1485 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist); 1486 free_pv_entry(pv); 1487 } 1488 1489 splx(s); 1490 return rtval; 1491} 1492 1493/* 1494 * Create a pv entry for page at pa for 1495 * (pmap, va). 1496 */ 1497static void 1498pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m) 1499{ 1500 1501 int s; 1502 pv_entry_t pv; 1503 1504 s = splvm(); 1505 pv = get_pv_entry(); 1506 pv->pv_va = va; 1507 pv->pv_pmap = pmap; 1508 pv->pv_ptem = mpte; 1509 1510 vm_page_lock_queues(); 1511 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist); 1512 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 1513 m->md.pv_list_count++; 1514 1515 vm_page_unlock_queues(); 1516 splx(s); 1517} 1518 1519/* 1520 * pmap_remove_pte: do the things to unmap a page in a process 1521 */ 1522static int 1523pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va) 1524{ 1525 pt_entry_t oldpte; 1526 vm_page_t m, mpte; 1527 1528 oldpte = pte_load_clear(ptq); 1529 if (oldpte & PG_W) 1530 pmap->pm_stats.wired_count -= 1; 1531 /* 1532 * Machines that don't support invlpg, also don't support 1533 * PG_G. 1534 */ 1535 if (oldpte & PG_G) 1536 pmap_invalidate_page(kernel_pmap, va); 1537 pmap->pm_stats.resident_count -= 1; 1538 if (oldpte & PG_MANAGED) { 1539 m = PHYS_TO_VM_PAGE(oldpte); 1540 if (oldpte & PG_M) { 1541#if defined(PMAP_DIAGNOSTIC) 1542 if (pmap_nw_modified((pt_entry_t) oldpte)) { 1543 printf( 1544 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n", 1545 va, oldpte); 1546 } 1547#endif 1548 if (pmap_track_modified(va)) 1549 vm_page_dirty(m); 1550 } 1551 if (oldpte & PG_A) 1552 vm_page_flag_set(m, PG_REFERENCED); 1553 return pmap_remove_entry(pmap, m, va); 1554 } else { 1555 mpte = PHYS_TO_VM_PAGE(*pmap_pde(pmap, va)); 1556 return pmap_unuse_pt(pmap, va, mpte); 1557 } 1558} 1559 1560/* 1561 * Remove a single page from a process address space 1562 */ 1563static void 1564pmap_remove_page(pmap_t pmap, vm_offset_t va) 1565{ 1566 pt_entry_t *pte; 1567 1568 if ((pte = pmap_pte_quick(pmap, va)) == NULL || *pte == 0) 1569 return; 1570 pmap_remove_pte(pmap, pte, va); 1571 pmap_invalidate_page(pmap, va); 1572} 1573 1574/* 1575 * Remove the given range of addresses from the specified map. 1576 * 1577 * It is assumed that the start and end are properly 1578 * rounded to the page size. 1579 */ 1580void 1581pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) 1582{ 1583 vm_offset_t pdnxt; 1584 pd_entry_t ptpaddr; 1585 pt_entry_t *pte; 1586 int anyvalid; 1587 1588 if (pmap == NULL) 1589 return; 1590 1591 if (pmap->pm_stats.resident_count == 0) 1592 return; 1593 1594 /* 1595 * special handling of removing one page. a very 1596 * common operation and easy to short circuit some 1597 * code. 1598 */ 1599 if ((sva + PAGE_SIZE == eva) && 1600 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { 1601 pmap_remove_page(pmap, sva); 1602 return; 1603 } 1604 1605 anyvalid = 0; 1606 1607 for (; sva < eva; sva = pdnxt) { 1608 unsigned pdirindex; 1609 1610 /* 1611 * Calculate index for next page table. 1612 */ 1613 pdnxt = (sva + NBPDR) & ~PDRMASK; 1614 if (pmap->pm_stats.resident_count == 0) 1615 break; 1616 1617 pdirindex = sva >> PDRSHIFT; 1618 ptpaddr = pmap->pm_pdir[pdirindex]; 1619 1620 /* 1621 * Weed out invalid mappings. Note: we assume that the page 1622 * directory table is always allocated, and in kernel virtual. 1623 */ 1624 if (ptpaddr == 0) 1625 continue; 1626 1627 /* 1628 * Check for large page. 1629 */ 1630 if ((ptpaddr & PG_PS) != 0) { 1631 pmap->pm_pdir[pdirindex] = 0; 1632 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1633 anyvalid = 1; 1634 continue; 1635 } 1636 1637 /* 1638 * Limit our scan to either the end of the va represented 1639 * by the current page table page, or to the end of the 1640 * range being removed. 1641 */ 1642 if (pdnxt > eva) 1643 pdnxt = eva; 1644 1645 for (; sva != pdnxt; sva += PAGE_SIZE) { 1646 if ((pte = pmap_pte_quick(pmap, sva)) == NULL || 1647 *pte == 0) 1648 continue; 1649 anyvalid = 1; 1650 if (pmap_remove_pte(pmap, pte, sva)) 1651 break; 1652 } 1653 } 1654 1655 if (anyvalid) 1656 pmap_invalidate_all(pmap); 1657} 1658 1659/* 1660 * Routine: pmap_remove_all 1661 * Function: 1662 * Removes this physical page from 1663 * all physical maps in which it resides. 1664 * Reflects back modify bits to the pager. 1665 * 1666 * Notes: 1667 * Original versions of this routine were very 1668 * inefficient because they iteratively called 1669 * pmap_remove (slow...) 1670 */ 1671 1672void 1673pmap_remove_all(vm_page_t m) 1674{ 1675 register pv_entry_t pv; 1676 pt_entry_t *pte, tpte; 1677 int s; 1678 1679#if defined(PMAP_DIAGNOSTIC) 1680 /* 1681 * XXX This makes pmap_remove_all() illegal for non-managed pages! 1682 */ 1683 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) { 1684 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x", 1685 VM_PAGE_TO_PHYS(m)); 1686 } 1687#endif 1688 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1689 s = splvm(); 1690 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 1691 pv->pv_pmap->pm_stats.resident_count--; 1692 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 1693 tpte = pte_load_clear(pte); 1694 if (tpte & PG_W) 1695 pv->pv_pmap->pm_stats.wired_count--; 1696 if (tpte & PG_A) 1697 vm_page_flag_set(m, PG_REFERENCED); 1698 1699 /* 1700 * Update the vm_page_t clean and reference bits. 1701 */ 1702 if (tpte & PG_M) { 1703#if defined(PMAP_DIAGNOSTIC) 1704 if (pmap_nw_modified((pt_entry_t) tpte)) { 1705 printf( 1706 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n", 1707 pv->pv_va, tpte); 1708 } 1709#endif 1710 if (pmap_track_modified(pv->pv_va)) 1711 vm_page_dirty(m); 1712 } 1713 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 1714 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 1715 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 1716 m->md.pv_list_count--; 1717 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 1718 free_pv_entry(pv); 1719 } 1720 vm_page_flag_clear(m, PG_WRITEABLE); 1721 splx(s); 1722} 1723 1724/* 1725 * Set the physical protection on the 1726 * specified range of this map as requested. 1727 */ 1728void 1729pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) 1730{ 1731 vm_offset_t pdnxt; 1732 pd_entry_t ptpaddr; 1733 int anychanged; 1734 1735 if (pmap == NULL) 1736 return; 1737 1738 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1739 pmap_remove(pmap, sva, eva); 1740 return; 1741 } 1742 1743 if (prot & VM_PROT_WRITE) 1744 return; 1745 1746 anychanged = 0; 1747 1748 for (; sva < eva; sva = pdnxt) { 1749 unsigned pdirindex; 1750 1751 pdnxt = (sva + NBPDR) & ~PDRMASK; 1752 1753 pdirindex = sva >> PDRSHIFT; 1754 ptpaddr = pmap->pm_pdir[pdirindex]; 1755 1756 /* 1757 * Weed out invalid mappings. Note: we assume that the page 1758 * directory table is always allocated, and in kernel virtual. 1759 */ 1760 if (ptpaddr == 0) 1761 continue; 1762 1763 /* 1764 * Check for large page. 1765 */ 1766 if ((ptpaddr & PG_PS) != 0) { 1767 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW); 1768 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; 1769 anychanged = 1; 1770 continue; 1771 } 1772 1773 if (pdnxt > eva) 1774 pdnxt = eva; 1775 1776 for (; sva != pdnxt; sva += PAGE_SIZE) { 1777 pt_entry_t pbits; 1778 pt_entry_t *pte; 1779 vm_page_t m; 1780 1781 if ((pte = pmap_pte_quick(pmap, sva)) == NULL) 1782 continue; 1783 pbits = *pte; 1784 if (pbits & PG_MANAGED) { 1785 m = NULL; 1786 if (pbits & PG_A) { 1787 m = PHYS_TO_VM_PAGE(pbits); 1788 vm_page_flag_set(m, PG_REFERENCED); 1789 pbits &= ~PG_A; 1790 } 1791 if ((pbits & PG_M) != 0 && 1792 pmap_track_modified(sva)) { 1793 if (m == NULL) 1794 m = PHYS_TO_VM_PAGE(pbits); 1795 vm_page_dirty(m); 1796 pbits &= ~PG_M; 1797 } 1798 } 1799 1800 pbits &= ~PG_RW; 1801 1802 if (pbits != *pte) { 1803 pte_store(pte, pbits); 1804 anychanged = 1; 1805 } 1806 } 1807 } 1808 if (anychanged) 1809 pmap_invalidate_all(pmap); 1810} 1811 1812/* 1813 * Insert the given physical page (p) at 1814 * the specified virtual address (v) in the 1815 * target physical map with the protection requested. 1816 * 1817 * If specified, the page will be wired down, meaning 1818 * that the related pte can not be reclaimed. 1819 * 1820 * NB: This is the only routine which MAY NOT lazy-evaluate 1821 * or lose information. That is, this routine must actually 1822 * insert this page into the given map NOW. 1823 */ 1824void 1825pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 1826 boolean_t wired) 1827{ 1828 vm_paddr_t pa; 1829 register pt_entry_t *pte; 1830 vm_paddr_t opa; 1831 pt_entry_t origpte, newpte; 1832 vm_page_t mpte; 1833 1834 if (pmap == NULL) 1835 return; 1836 1837 va &= PG_FRAME; 1838#ifdef PMAP_DIAGNOSTIC 1839 if (va > VM_MAX_KERNEL_ADDRESS) 1840 panic("pmap_enter: toobig"); 1841 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) 1842 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va); 1843#endif 1844 1845 mpte = NULL; 1846 /* 1847 * In the case that a page table page is not 1848 * resident, we are creating it here. 1849 */ 1850 if (va < VM_MAXUSER_ADDRESS) { 1851 mpte = pmap_allocpte(pmap, va); 1852 } 1853#if 0 && defined(PMAP_DIAGNOSTIC) 1854 else { 1855 pd_entry_t *pdeaddr = pmap_pde(pmap, va); 1856 origpte = *pdeaddr; 1857 if ((origpte & PG_V) == 0) { 1858 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n", 1859 pmap->pm_pdir[PTDPTDI], origpte, va); 1860 } 1861 } 1862#endif 1863 1864 pte = pmap_pte_quick(pmap, va); 1865 1866 /* 1867 * Page Directory table entry not valid, we need a new PT page 1868 */ 1869 if (pte == NULL) { 1870 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x\n", 1871 (uintmax_t)pmap->pm_pdir[PTDPTDI], va); 1872 } 1873 1874 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME; 1875 origpte = *pte; 1876 opa = origpte & PG_FRAME; 1877 1878 if (origpte & PG_PS) { 1879 /* 1880 * Yes, I know this will truncate upper address bits for PAE, 1881 * but I'm actually more interested in the lower bits 1882 */ 1883 printf("pmap_enter: va %p, pte %p, origpte %p\n", 1884 (void *)va, (void *)pte, (void *)(uintptr_t)origpte); 1885 panic("pmap_enter: attempted pmap_enter on 4MB page"); 1886 } 1887 1888 /* 1889 * Mapping has not changed, must be protection or wiring change. 1890 */ 1891 if (origpte && (opa == pa)) { 1892 /* 1893 * Wiring change, just update stats. We don't worry about 1894 * wiring PT pages as they remain resident as long as there 1895 * are valid mappings in them. Hence, if a user page is wired, 1896 * the PT page will be also. 1897 */ 1898 if (wired && ((origpte & PG_W) == 0)) 1899 pmap->pm_stats.wired_count++; 1900 else if (!wired && (origpte & PG_W)) 1901 pmap->pm_stats.wired_count--; 1902 1903#if defined(PMAP_DIAGNOSTIC) 1904 if (pmap_nw_modified((pt_entry_t) origpte)) { 1905 printf( 1906 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n", 1907 va, origpte); 1908 } 1909#endif 1910 1911 /* 1912 * Remove extra pte reference 1913 */ 1914 if (mpte) 1915 mpte->hold_count--; 1916 1917 if ((prot & VM_PROT_WRITE) && (origpte & PG_V)) { 1918 if ((origpte & PG_RW) == 0) { 1919 pte_store(pte, origpte | PG_RW); 1920 pmap_invalidate_page(pmap, va); 1921 } 1922 return; 1923 } 1924 1925 /* 1926 * We might be turning off write access to the page, 1927 * so we go ahead and sense modify status. 1928 */ 1929 if (origpte & PG_MANAGED) { 1930 if ((origpte & PG_M) && pmap_track_modified(va)) { 1931 vm_page_t om; 1932 om = PHYS_TO_VM_PAGE(opa); 1933 vm_page_dirty(om); 1934 } 1935 pa |= PG_MANAGED; 1936 } 1937 goto validate; 1938 } 1939 /* 1940 * Mapping has changed, invalidate old range and fall through to 1941 * handle validating new mapping. 1942 */ 1943 if (opa) { 1944 int err; 1945 vm_page_lock_queues(); 1946 err = pmap_remove_pte(pmap, pte, va); 1947 vm_page_unlock_queues(); 1948 if (err) 1949 panic("pmap_enter: pte vanished, va: 0x%x", va); 1950 } 1951 1952 /* 1953 * Enter on the PV list if part of our managed memory. Note that we 1954 * raise IPL while manipulating pv_table since pmap_enter can be 1955 * called at interrupt time. 1956 */ 1957 if (pmap_initialized && 1958 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) { 1959 pmap_insert_entry(pmap, va, mpte, m); 1960 pa |= PG_MANAGED; 1961 } 1962 1963 /* 1964 * Increment counters 1965 */ 1966 pmap->pm_stats.resident_count++; 1967 if (wired) 1968 pmap->pm_stats.wired_count++; 1969 1970validate: 1971 /* 1972 * Now validate mapping with desired protection/wiring. 1973 */ 1974 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) | PG_V); 1975 1976 if (wired) 1977 newpte |= PG_W; 1978 if (va < VM_MAXUSER_ADDRESS) 1979 newpte |= PG_U; 1980 if (pmap == kernel_pmap) 1981 newpte |= pgeflag; 1982 1983 /* 1984 * if the mapping or permission bits are different, we need 1985 * to update the pte. 1986 */ 1987 if ((origpte & ~(PG_M|PG_A)) != newpte) { 1988 pte_store(pte, newpte | PG_A); 1989 /*if (origpte)*/ { 1990 pmap_invalidate_page(pmap, va); 1991 } 1992 } 1993} 1994 1995/* 1996 * this code makes some *MAJOR* assumptions: 1997 * 1. Current pmap & pmap exists. 1998 * 2. Not wired. 1999 * 3. Read access. 2000 * 4. No page table pages. 2001 * 5. Tlbflush is deferred to calling procedure. 2002 * 6. Page IS managed. 2003 * but is *MUCH* faster than pmap_enter... 2004 */ 2005 2006vm_page_t 2007pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte) 2008{ 2009 pt_entry_t *pte; 2010 vm_paddr_t pa; 2011 2012 /* 2013 * In the case that a page table page is not 2014 * resident, we are creating it here. 2015 */ 2016 if (va < VM_MAXUSER_ADDRESS) { 2017 unsigned ptepindex; 2018 pd_entry_t ptepa; 2019 2020 /* 2021 * Calculate pagetable page index 2022 */ 2023 ptepindex = va >> PDRSHIFT; 2024 if (mpte && (mpte->pindex == ptepindex)) { 2025 mpte->hold_count++; 2026 } else { 2027retry: 2028 /* 2029 * Get the page directory entry 2030 */ 2031 ptepa = pmap->pm_pdir[ptepindex]; 2032 2033 /* 2034 * If the page table page is mapped, we just increment 2035 * the hold count, and activate it. 2036 */ 2037 if (ptepa) { 2038 if (ptepa & PG_PS) 2039 panic("pmap_enter_quick: unexpected mapping into 4MB page"); 2040 mpte = PHYS_TO_VM_PAGE(ptepa); 2041 mpte->hold_count++; 2042 } else { 2043 mpte = _pmap_allocpte(pmap, ptepindex); 2044 if (mpte == NULL) 2045 goto retry; 2046 } 2047 } 2048 } else { 2049 mpte = NULL; 2050 } 2051 2052 /* 2053 * This call to vtopte makes the assumption that we are 2054 * entering the page into the current pmap. In order to support 2055 * quick entry into any pmap, one would likely use pmap_pte_quick. 2056 * But that isn't as quick as vtopte. 2057 */ 2058 pte = vtopte(va); 2059 if (*pte) { 2060 if (mpte != NULL) { 2061 vm_page_lock_queues(); 2062 pmap_unwire_pte_hold(pmap, mpte); 2063 vm_page_unlock_queues(); 2064 } 2065 return 0; 2066 } 2067 2068 /* 2069 * Enter on the PV list if part of our managed memory. Note that we 2070 * raise IPL while manipulating pv_table since pmap_enter can be 2071 * called at interrupt time. 2072 */ 2073 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) 2074 pmap_insert_entry(pmap, va, mpte, m); 2075 2076 /* 2077 * Increment counters 2078 */ 2079 pmap->pm_stats.resident_count++; 2080 2081 pa = VM_PAGE_TO_PHYS(m); 2082 2083 /* 2084 * Now validate mapping with RO protection 2085 */ 2086 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) 2087 pte_store(pte, pa | PG_V | PG_U); 2088 else 2089 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED); 2090 2091 return mpte; 2092} 2093 2094/* 2095 * Make a temporary mapping for a physical address. This is only intended 2096 * to be used for panic dumps. 2097 */ 2098void * 2099pmap_kenter_temporary(vm_offset_t pa, int i) 2100{ 2101 vm_offset_t va; 2102 2103 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); 2104 pmap_kenter(va, pa); 2105#ifndef I386_CPU 2106 invlpg(va); 2107#else 2108 invltlb(); 2109#endif 2110 return ((void *)crashdumpmap); 2111} 2112 2113/* 2114 * This code maps large physical mmap regions into the 2115 * processor address space. Note that some shortcuts 2116 * are taken, but the code works. 2117 */ 2118void 2119pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, 2120 vm_object_t object, vm_pindex_t pindex, 2121 vm_size_t size) 2122{ 2123 vm_page_t p; 2124 2125 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2126 KASSERT(object->type == OBJT_DEVICE, 2127 ("pmap_object_init_pt: non-device object")); 2128 if (pseflag && 2129 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) { 2130 int i; 2131 vm_page_t m[1]; 2132 unsigned int ptepindex; 2133 int npdes; 2134 pd_entry_t ptepa; 2135 2136 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)]) 2137 return; 2138retry: 2139 p = vm_page_lookup(object, pindex); 2140 if (p != NULL) { 2141 vm_page_lock_queues(); 2142 if (vm_page_sleep_if_busy(p, FALSE, "init4p")) 2143 goto retry; 2144 } else { 2145 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 2146 if (p == NULL) 2147 return; 2148 m[0] = p; 2149 2150 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) { 2151 vm_page_lock_queues(); 2152 vm_page_free(p); 2153 vm_page_unlock_queues(); 2154 return; 2155 } 2156 2157 p = vm_page_lookup(object, pindex); 2158 vm_page_lock_queues(); 2159 vm_page_wakeup(p); 2160 } 2161 vm_page_unlock_queues(); 2162 2163 ptepa = VM_PAGE_TO_PHYS(p); 2164 if (ptepa & (NBPDR - 1)) 2165 return; 2166 2167 p->valid = VM_PAGE_BITS_ALL; 2168 2169 pmap->pm_stats.resident_count += size >> PAGE_SHIFT; 2170 npdes = size >> PDRSHIFT; 2171 for(i = 0; i < npdes; i++) { 2172 pde_store(&pmap->pm_pdir[ptepindex], 2173 ptepa | PG_U | PG_RW | PG_V | PG_PS); 2174 ptepa += NBPDR; 2175 ptepindex += 1; 2176 } 2177 pmap_invalidate_all(pmap); 2178 } 2179} 2180 2181/* 2182 * Routine: pmap_change_wiring 2183 * Function: Change the wiring attribute for a map/virtual-address 2184 * pair. 2185 * In/out conditions: 2186 * The mapping must already exist in the pmap. 2187 */ 2188void 2189pmap_change_wiring(pmap, va, wired) 2190 register pmap_t pmap; 2191 vm_offset_t va; 2192 boolean_t wired; 2193{ 2194 register pt_entry_t *pte; 2195 2196 if (pmap == NULL) 2197 return; 2198 2199 pte = pmap_pte_quick(pmap, va); 2200 2201 if (wired && !pmap_pte_w(pte)) 2202 pmap->pm_stats.wired_count++; 2203 else if (!wired && pmap_pte_w(pte)) 2204 pmap->pm_stats.wired_count--; 2205 2206 /* 2207 * Wiring is not a hardware characteristic so there is no need to 2208 * invalidate TLB. 2209 */ 2210 pmap_pte_set_w(pte, wired); 2211} 2212 2213 2214 2215/* 2216 * Copy the range specified by src_addr/len 2217 * from the source map to the range dst_addr/len 2218 * in the destination map. 2219 * 2220 * This routine is only advisory and need not do anything. 2221 */ 2222 2223void 2224pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, 2225 vm_offset_t src_addr) 2226{ 2227 vm_offset_t addr; 2228 vm_offset_t end_addr = src_addr + len; 2229 vm_offset_t pdnxt; 2230 vm_page_t m; 2231 2232 if (dst_addr != src_addr) 2233 return; 2234 2235 if (!pmap_is_current(src_pmap)) 2236 return; 2237 2238 for (addr = src_addr; addr < end_addr; addr = pdnxt) { 2239 pt_entry_t *src_pte, *dst_pte; 2240 vm_page_t dstmpte, srcmpte; 2241 pd_entry_t srcptepaddr; 2242 unsigned ptepindex; 2243 2244 if (addr >= UPT_MIN_ADDRESS) 2245 panic("pmap_copy: invalid to pmap_copy page tables\n"); 2246 2247 /* 2248 * Don't let optional prefaulting of pages make us go 2249 * way below the low water mark of free pages or way 2250 * above high water mark of used pv entries. 2251 */ 2252 if (cnt.v_free_count < cnt.v_free_reserved || 2253 pv_entry_count > pv_entry_high_water) 2254 break; 2255 2256 pdnxt = (addr + NBPDR) & ~PDRMASK; 2257 ptepindex = addr >> PDRSHIFT; 2258 2259 srcptepaddr = src_pmap->pm_pdir[ptepindex]; 2260 if (srcptepaddr == 0) 2261 continue; 2262 2263 if (srcptepaddr & PG_PS) { 2264 if (dst_pmap->pm_pdir[ptepindex] == 0) { 2265 dst_pmap->pm_pdir[ptepindex] = srcptepaddr; 2266 dst_pmap->pm_stats.resident_count += 2267 NBPDR / PAGE_SIZE; 2268 } 2269 continue; 2270 } 2271 2272 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr); 2273 if (srcmpte->hold_count == 0 || (srcmpte->flags & PG_BUSY)) 2274 continue; 2275 2276 if (pdnxt > end_addr) 2277 pdnxt = end_addr; 2278 2279 src_pte = vtopte(addr); 2280 while (addr < pdnxt) { 2281 pt_entry_t ptetemp; 2282 ptetemp = *src_pte; 2283 /* 2284 * we only virtual copy managed pages 2285 */ 2286 if ((ptetemp & PG_MANAGED) != 0) { 2287 /* 2288 * We have to check after allocpte for the 2289 * pte still being around... allocpte can 2290 * block. 2291 */ 2292 dstmpte = pmap_allocpte(dst_pmap, addr); 2293 dst_pte = pmap_pte_quick(dst_pmap, addr); 2294 if ((*dst_pte == 0) && (ptetemp = *src_pte)) { 2295 /* 2296 * Clear the modified and 2297 * accessed (referenced) bits 2298 * during the copy. 2299 */ 2300 m = PHYS_TO_VM_PAGE(ptetemp); 2301 *dst_pte = ptetemp & ~(PG_M | PG_A); 2302 dst_pmap->pm_stats.resident_count++; 2303 pmap_insert_entry(dst_pmap, addr, 2304 dstmpte, m); 2305 } else { 2306 vm_page_lock_queues(); 2307 pmap_unwire_pte_hold(dst_pmap, dstmpte); 2308 vm_page_unlock_queues(); 2309 } 2310 if (dstmpte->hold_count >= srcmpte->hold_count) 2311 break; 2312 } 2313 addr += PAGE_SIZE; 2314 src_pte++; 2315 } 2316 } 2317} 2318 2319#ifdef SMP 2320 2321/* 2322 * pmap_zpi_switchout*() 2323 * 2324 * These functions allow us to avoid doing IPIs alltogether in certain 2325 * temporary page-mapping situations (page zeroing). Instead to deal 2326 * with being preempted and moved onto a different cpu we invalidate 2327 * the page when the scheduler switches us in. This does not occur 2328 * very often so we remain relatively optimal with very little effort. 2329 */ 2330static void 2331pmap_zpi_switchout12(void) 2332{ 2333 invlpg((u_int)CADDR1); 2334 invlpg((u_int)CADDR2); 2335} 2336 2337static void 2338pmap_zpi_switchout2(void) 2339{ 2340 invlpg((u_int)CADDR2); 2341} 2342 2343static void 2344pmap_zpi_switchout3(void) 2345{ 2346 invlpg((u_int)CADDR3); 2347} 2348 2349#endif 2350 2351static __inline void 2352pagezero(void *page) 2353{ 2354#if defined(I686_CPU) 2355 if (cpu_class == CPUCLASS_686) { 2356#if defined(CPU_ENABLE_SSE) 2357 if (cpu_feature & CPUID_SSE2) 2358 sse2_pagezero(page); 2359 else 2360#endif 2361 i686_pagezero(page); 2362 } else 2363#endif 2364 bzero(page, PAGE_SIZE); 2365} 2366 2367static __inline void 2368invlcaddr(void *caddr) 2369{ 2370#ifdef I386_CPU 2371 invltlb(); 2372#else 2373 invlpg((u_int)caddr); 2374#endif 2375} 2376 2377/* 2378 * pmap_zero_page zeros the specified hardware page by mapping 2379 * the page into KVM and using bzero to clear its contents. 2380 */ 2381void 2382pmap_zero_page(vm_page_t m) 2383{ 2384 2385 mtx_lock(&CMAPCADDR12_lock); 2386 if (*CMAP2) 2387 panic("pmap_zero_page: CMAP2 busy"); 2388#ifdef SMP 2389 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout2; 2390#endif 2391 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2392#ifdef SMP 2393 invlpg((u_int)CADDR2); 2394#endif 2395 pagezero(CADDR2); 2396 *CMAP2 = 0; 2397 invlcaddr(CADDR2); 2398#ifdef SMP 2399 curthread->td_pcb->pcb_switchout = NULL; 2400#endif 2401 mtx_unlock(&CMAPCADDR12_lock); 2402} 2403 2404/* 2405 * pmap_zero_page_area zeros the specified hardware page by mapping 2406 * the page into KVM and using bzero to clear its contents. 2407 * 2408 * off and size may not cover an area beyond a single hardware page. 2409 */ 2410void 2411pmap_zero_page_area(vm_page_t m, int off, int size) 2412{ 2413 2414 mtx_lock(&CMAPCADDR12_lock); 2415 if (*CMAP2) 2416 panic("pmap_zero_page: CMAP2 busy"); 2417#ifdef SMP 2418 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout2; 2419#endif 2420 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2421#ifdef SMP 2422 invlpg((u_int)CADDR2); 2423#endif 2424 if (off == 0 && size == PAGE_SIZE) 2425 pagezero(CADDR2); 2426 else 2427 bzero((char *)CADDR2 + off, size); 2428 *CMAP2 = 0; 2429 invlcaddr(CADDR2); 2430#ifdef SMP 2431 curthread->td_pcb->pcb_switchout = NULL; 2432#endif 2433 mtx_unlock(&CMAPCADDR12_lock); 2434} 2435 2436/* 2437 * pmap_zero_page_idle zeros the specified hardware page by mapping 2438 * the page into KVM and using bzero to clear its contents. This 2439 * is intended to be called from the vm_pagezero process only and 2440 * outside of Giant. 2441 */ 2442void 2443pmap_zero_page_idle(vm_page_t m) 2444{ 2445 2446 if (*CMAP3) 2447 panic("pmap_zero_page: CMAP3 busy"); 2448#ifdef SMP 2449 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout3; 2450#endif 2451 *CMAP3 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M; 2452#ifdef SMP 2453 invlpg((u_int)CADDR3); 2454#endif 2455 pagezero(CADDR3); 2456 *CMAP3 = 0; 2457 invlcaddr(CADDR3); 2458#ifdef SMP 2459 curthread->td_pcb->pcb_switchout = NULL; 2460#endif 2461} 2462 2463/* 2464 * pmap_copy_page copies the specified (machine independent) 2465 * page by mapping the page into virtual memory and using 2466 * bcopy to copy the page, one machine dependent page at a 2467 * time. 2468 */ 2469void 2470pmap_copy_page(vm_page_t src, vm_page_t dst) 2471{ 2472 2473 mtx_lock(&CMAPCADDR12_lock); 2474 if (*CMAP1) 2475 panic("pmap_copy_page: CMAP1 busy"); 2476 if (*CMAP2) 2477 panic("pmap_copy_page: CMAP2 busy"); 2478#ifdef SMP 2479 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout12; 2480#endif 2481 *CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A; 2482 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M; 2483#ifdef SMP 2484 invlpg((u_int)CADDR1); 2485 invlpg((u_int)CADDR2); 2486#endif 2487 bcopy(CADDR1, CADDR2, PAGE_SIZE); 2488 *CMAP1 = 0; 2489 *CMAP2 = 0; 2490#ifdef I386_CPU 2491 invltlb(); 2492#else 2493 invlpg((u_int)CADDR1); 2494 invlpg((u_int)CADDR2); 2495#endif 2496#ifdef SMP 2497 curthread->td_pcb->pcb_switchout = NULL; 2498#endif 2499 mtx_unlock(&CMAPCADDR12_lock); 2500} 2501 2502/* 2503 * Returns true if the pmap's pv is one of the first 2504 * 16 pvs linked to from this page. This count may 2505 * be changed upwards or downwards in the future; it 2506 * is only necessary that true be returned for a small 2507 * subset of pmaps for proper page aging. 2508 */ 2509boolean_t 2510pmap_page_exists_quick(pmap, m) 2511 pmap_t pmap; 2512 vm_page_t m; 2513{ 2514 pv_entry_t pv; 2515 int loops = 0; 2516 int s; 2517 2518 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2519 return FALSE; 2520 2521 s = splvm(); 2522 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2523 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2524 if (pv->pv_pmap == pmap) { 2525 splx(s); 2526 return TRUE; 2527 } 2528 loops++; 2529 if (loops >= 16) 2530 break; 2531 } 2532 splx(s); 2533 return (FALSE); 2534} 2535 2536#define PMAP_REMOVE_PAGES_CURPROC_ONLY 2537/* 2538 * Remove all pages from specified address space 2539 * this aids process exit speeds. Also, this code 2540 * is special cased for current process only, but 2541 * can have the more generic (and slightly slower) 2542 * mode enabled. This is much faster than pmap_remove 2543 * in the case of running down an entire address space. 2544 */ 2545void 2546pmap_remove_pages(pmap, sva, eva) 2547 pmap_t pmap; 2548 vm_offset_t sva, eva; 2549{ 2550 pt_entry_t *pte, tpte; 2551 vm_page_t m; 2552 pv_entry_t pv, npv; 2553 int s; 2554 2555#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2556 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) { 2557 printf("warning: pmap_remove_pages called with non-current pmap\n"); 2558 return; 2559 } 2560#endif 2561 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2562 s = splvm(); 2563 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) { 2564 2565 if (pv->pv_va >= eva || pv->pv_va < sva) { 2566 npv = TAILQ_NEXT(pv, pv_plist); 2567 continue; 2568 } 2569 2570#ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY 2571 pte = vtopte(pv->pv_va); 2572#else 2573 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2574#endif 2575 tpte = *pte; 2576 2577 if (tpte == 0) { 2578 printf("TPTE at %p IS ZERO @ VA %08x\n", 2579 pte, pv->pv_va); 2580 panic("bad pte"); 2581 } 2582 2583/* 2584 * We cannot remove wired pages from a process' mapping at this time 2585 */ 2586 if (tpte & PG_W) { 2587 npv = TAILQ_NEXT(pv, pv_plist); 2588 continue; 2589 } 2590 2591 m = PHYS_TO_VM_PAGE(tpte); 2592 KASSERT(m->phys_addr == (tpte & PG_FRAME), 2593 ("vm_page_t %p phys_addr mismatch %016jx %016jx", 2594 m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); 2595 2596 KASSERT(m < &vm_page_array[vm_page_array_size], 2597 ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte)); 2598 2599 pv->pv_pmap->pm_stats.resident_count--; 2600 2601 pte_clear(pte); 2602 2603 /* 2604 * Update the vm_page_t clean and reference bits. 2605 */ 2606 if (tpte & PG_M) { 2607 vm_page_dirty(m); 2608 } 2609 2610 npv = TAILQ_NEXT(pv, pv_plist); 2611 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist); 2612 2613 m->md.pv_list_count--; 2614 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2615 if (TAILQ_FIRST(&m->md.pv_list) == NULL) { 2616 vm_page_flag_clear(m, PG_WRITEABLE); 2617 } 2618 2619 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem); 2620 free_pv_entry(pv); 2621 } 2622 splx(s); 2623 pmap_invalidate_all(pmap); 2624} 2625 2626/* 2627 * pmap_is_modified: 2628 * 2629 * Return whether or not the specified physical page was modified 2630 * in any physical maps. 2631 */ 2632boolean_t 2633pmap_is_modified(vm_page_t m) 2634{ 2635 pv_entry_t pv; 2636 pt_entry_t *pte; 2637 int s; 2638 2639 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2640 return FALSE; 2641 2642 s = splvm(); 2643 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2644 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2645 /* 2646 * if the bit being tested is the modified bit, then 2647 * mark clean_map and ptes as never 2648 * modified. 2649 */ 2650 if (!pmap_track_modified(pv->pv_va)) 2651 continue; 2652#if defined(PMAP_DIAGNOSTIC) 2653 if (!pv->pv_pmap) { 2654 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va); 2655 continue; 2656 } 2657#endif 2658 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2659 if (*pte & PG_M) { 2660 splx(s); 2661 return TRUE; 2662 } 2663 } 2664 splx(s); 2665 return (FALSE); 2666} 2667 2668/* 2669 * pmap_is_prefaultable: 2670 * 2671 * Return whether or not the specified virtual address is elgible 2672 * for prefault. 2673 */ 2674boolean_t 2675pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) 2676{ 2677 pt_entry_t *pte; 2678 2679 if ((*pmap_pde(pmap, addr)) == 0) 2680 return (FALSE); 2681 pte = vtopte(addr); 2682 if (*pte) 2683 return (FALSE); 2684 return (TRUE); 2685} 2686 2687/* 2688 * Clear the given bit in each of the given page's ptes. 2689 */ 2690static __inline void 2691pmap_clear_ptes(vm_page_t m, int bit) 2692{ 2693 register pv_entry_t pv; 2694 pt_entry_t pbits, *pte; 2695 int s; 2696 2697 if (!pmap_initialized || (m->flags & PG_FICTITIOUS) || 2698 (bit == PG_RW && (m->flags & PG_WRITEABLE) == 0)) 2699 return; 2700 2701 s = splvm(); 2702 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2703 /* 2704 * Loop over all current mappings setting/clearing as appropos If 2705 * setting RO do we need to clear the VAC? 2706 */ 2707 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 2708 /* 2709 * don't write protect pager mappings 2710 */ 2711 if (bit == PG_RW) { 2712 if (!pmap_track_modified(pv->pv_va)) 2713 continue; 2714 } 2715 2716#if defined(PMAP_DIAGNOSTIC) 2717 if (!pv->pv_pmap) { 2718 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va); 2719 continue; 2720 } 2721#endif 2722 2723 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2724 pbits = *pte; 2725 if (pbits & bit) { 2726 if (bit == PG_RW) { 2727 if (pbits & PG_M) { 2728 vm_page_dirty(m); 2729 } 2730 pte_store(pte, pbits & ~(PG_M|PG_RW)); 2731 } else { 2732 pte_store(pte, pbits & ~bit); 2733 } 2734 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2735 } 2736 } 2737 if (bit == PG_RW) 2738 vm_page_flag_clear(m, PG_WRITEABLE); 2739 splx(s); 2740} 2741 2742/* 2743 * pmap_page_protect: 2744 * 2745 * Lower the permission for all mappings to a given page. 2746 */ 2747void 2748pmap_page_protect(vm_page_t m, vm_prot_t prot) 2749{ 2750 if ((prot & VM_PROT_WRITE) == 0) { 2751 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) { 2752 pmap_clear_ptes(m, PG_RW); 2753 } else { 2754 pmap_remove_all(m); 2755 } 2756 } 2757} 2758 2759/* 2760 * pmap_ts_referenced: 2761 * 2762 * Return a count of reference bits for a page, clearing those bits. 2763 * It is not necessary for every reference bit to be cleared, but it 2764 * is necessary that 0 only be returned when there are truly no 2765 * reference bits set. 2766 * 2767 * XXX: The exact number of bits to check and clear is a matter that 2768 * should be tested and standardized at some point in the future for 2769 * optimal aging of shared pages. 2770 */ 2771int 2772pmap_ts_referenced(vm_page_t m) 2773{ 2774 register pv_entry_t pv, pvf, pvn; 2775 pt_entry_t *pte; 2776 pt_entry_t v; 2777 int s; 2778 int rtval = 0; 2779 2780 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) 2781 return (rtval); 2782 2783 s = splvm(); 2784 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2785 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { 2786 2787 pvf = pv; 2788 2789 do { 2790 pvn = TAILQ_NEXT(pv, pv_list); 2791 2792 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list); 2793 2794 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list); 2795 2796 if (!pmap_track_modified(pv->pv_va)) 2797 continue; 2798 2799 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va); 2800 2801 if (pte && ((v = pte_load(pte)) & PG_A) != 0) { 2802 pte_store(pte, v & ~PG_A); 2803 pmap_invalidate_page(pv->pv_pmap, pv->pv_va); 2804 2805 rtval++; 2806 if (rtval > 4) { 2807 break; 2808 } 2809 } 2810 } while ((pv = pvn) != NULL && pv != pvf); 2811 } 2812 splx(s); 2813 2814 return (rtval); 2815} 2816 2817/* 2818 * Clear the modify bits on the specified physical page. 2819 */ 2820void 2821pmap_clear_modify(vm_page_t m) 2822{ 2823 pmap_clear_ptes(m, PG_M); 2824} 2825 2826/* 2827 * pmap_clear_reference: 2828 * 2829 * Clear the reference bit on the specified physical page. 2830 */ 2831void 2832pmap_clear_reference(vm_page_t m) 2833{ 2834 pmap_clear_ptes(m, PG_A); 2835} 2836 2837/* 2838 * Miscellaneous support routines follow 2839 */ 2840 2841static void 2842i386_protection_init() 2843{ 2844 register int *kp, prot; 2845 2846 kp = protection_codes; 2847 for (prot = 0; prot < 8; prot++) { 2848 switch (prot) { 2849 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE: 2850 /* 2851 * Read access is also 0. There isn't any execute bit, 2852 * so just make it readable. 2853 */ 2854 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE: 2855 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE: 2856 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE: 2857 *kp++ = 0; 2858 break; 2859 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE: 2860 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE: 2861 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE: 2862 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE: 2863 *kp++ = PG_RW; 2864 break; 2865 } 2866 } 2867} 2868 2869/* 2870 * Map a set of physical memory pages into the kernel virtual 2871 * address space. Return a pointer to where it is mapped. This 2872 * routine is intended to be used for mapping device memory, 2873 * NOT real memory. 2874 */ 2875void * 2876pmap_mapdev(pa, size) 2877 vm_paddr_t pa; 2878 vm_size_t size; 2879{ 2880 vm_offset_t va, tmpva, offset; 2881 2882 offset = pa & PAGE_MASK; 2883 size = roundup(offset + size, PAGE_SIZE); 2884 pa = pa & PG_FRAME; 2885 2886 if (pa < KERNLOAD && pa + size <= KERNLOAD) 2887 va = KERNBASE + pa; 2888 else 2889 va = kmem_alloc_nofault(kernel_map, size); 2890 if (!va) 2891 panic("pmap_mapdev: Couldn't alloc kernel virtual memory"); 2892 2893 for (tmpva = va; size > 0; ) { 2894 pmap_kenter(tmpva, pa); 2895 size -= PAGE_SIZE; 2896 tmpva += PAGE_SIZE; 2897 pa += PAGE_SIZE; 2898 } 2899 pmap_invalidate_range(kernel_pmap, va, tmpva); 2900 return ((void *)(va + offset)); 2901} 2902 2903void 2904pmap_unmapdev(va, size) 2905 vm_offset_t va; 2906 vm_size_t size; 2907{ 2908 vm_offset_t base, offset, tmpva; 2909 2910 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD) 2911 return; 2912 base = va & PG_FRAME; 2913 offset = va & PAGE_MASK; 2914 size = roundup(offset + size, PAGE_SIZE); 2915 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) 2916 pmap_kremove(tmpva); 2917 pmap_invalidate_range(kernel_pmap, va, tmpva); 2918 kmem_free(kernel_map, base, size); 2919} 2920 2921/* 2922 * perform the pmap work for mincore 2923 */ 2924int 2925pmap_mincore(pmap, addr) 2926 pmap_t pmap; 2927 vm_offset_t addr; 2928{ 2929 pt_entry_t *ptep, pte; 2930 vm_page_t m; 2931 int val = 0; 2932 2933 ptep = pmap_pte_quick(pmap, addr); 2934 if (ptep == 0) { 2935 return 0; 2936 } 2937 2938 if ((pte = *ptep) != 0) { 2939 vm_paddr_t pa; 2940 2941 val = MINCORE_INCORE; 2942 if ((pte & PG_MANAGED) == 0) 2943 return val; 2944 2945 pa = pte & PG_FRAME; 2946 2947 m = PHYS_TO_VM_PAGE(pa); 2948 2949 /* 2950 * Modified by us 2951 */ 2952 if (pte & PG_M) 2953 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER; 2954 else { 2955 /* 2956 * Modified by someone else 2957 */ 2958 vm_page_lock_queues(); 2959 if (m->dirty || pmap_is_modified(m)) 2960 val |= MINCORE_MODIFIED_OTHER; 2961 vm_page_unlock_queues(); 2962 } 2963 /* 2964 * Referenced by us 2965 */ 2966 if (pte & PG_A) 2967 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER; 2968 else { 2969 /* 2970 * Referenced by someone else 2971 */ 2972 vm_page_lock_queues(); 2973 if ((m->flags & PG_REFERENCED) || 2974 pmap_ts_referenced(m)) { 2975 val |= MINCORE_REFERENCED_OTHER; 2976 vm_page_flag_set(m, PG_REFERENCED); 2977 } 2978 vm_page_unlock_queues(); 2979 } 2980 } 2981 return val; 2982} 2983 2984void 2985pmap_activate(struct thread *td) 2986{ 2987 struct proc *p = td->td_proc; 2988 pmap_t pmap, oldpmap; 2989 u_int32_t cr3; 2990 2991 critical_enter(); 2992 pmap = vmspace_pmap(td->td_proc->p_vmspace); 2993 oldpmap = PCPU_GET(curpmap); 2994#if defined(SMP) 2995 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask)); 2996 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask)); 2997#else 2998 oldpmap->pm_active &= ~1; 2999 pmap->pm_active |= 1; 3000#endif 3001#ifdef PAE 3002 cr3 = vtophys(pmap->pm_pdpt); 3003#else 3004 cr3 = vtophys(pmap->pm_pdir); 3005#endif 3006 /* XXXKSE this is wrong. 3007 * pmap_activate is for the current thread on the current cpu 3008 */ 3009 if (p->p_flag & P_SA) { 3010 /* Make sure all other cr3 entries are updated. */ 3011 /* what if they are running? XXXKSE (maybe abort them) */ 3012 FOREACH_THREAD_IN_PROC(p, td) { 3013 td->td_pcb->pcb_cr3 = cr3; 3014 } 3015 } else { 3016 td->td_pcb->pcb_cr3 = cr3; 3017 } 3018 load_cr3(cr3); 3019 PCPU_SET(curpmap, pmap); 3020 critical_exit(); 3021} 3022 3023vm_offset_t 3024pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size) 3025{ 3026 3027 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) { 3028 return addr; 3029 } 3030 3031 addr = (addr + PDRMASK) & ~PDRMASK; 3032 return addr; 3033} 3034 3035 3036#if defined(PMAP_DEBUG) 3037pmap_pid_dump(int pid) 3038{ 3039 pmap_t pmap; 3040 struct proc *p; 3041 int npte = 0; 3042 int index; 3043 3044 sx_slock(&allproc_lock); 3045 LIST_FOREACH(p, &allproc, p_list) { 3046 if (p->p_pid != pid) 3047 continue; 3048 3049 if (p->p_vmspace) { 3050 int i,j; 3051 index = 0; 3052 pmap = vmspace_pmap(p->p_vmspace); 3053 for (i = 0; i < NPDEPTD; i++) { 3054 pd_entry_t *pde; 3055 pt_entry_t *pte; 3056 vm_offset_t base = i << PDRSHIFT; 3057 3058 pde = &pmap->pm_pdir[i]; 3059 if (pde && pmap_pde_v(pde)) { 3060 for (j = 0; j < NPTEPG; j++) { 3061 vm_offset_t va = base + (j << PAGE_SHIFT); 3062 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { 3063 if (index) { 3064 index = 0; 3065 printf("\n"); 3066 } 3067 sx_sunlock(&allproc_lock); 3068 return npte; 3069 } 3070 pte = pmap_pte_quick(pmap, va); 3071 if (pte && pmap_pte_v(pte)) { 3072 pt_entry_t pa; 3073 vm_page_t m; 3074 pa = *pte; 3075 m = PHYS_TO_VM_PAGE(pa); 3076 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", 3077 va, pa, m->hold_count, m->wire_count, m->flags); 3078 npte++; 3079 index++; 3080 if (index >= 2) { 3081 index = 0; 3082 printf("\n"); 3083 } else { 3084 printf(" "); 3085 } 3086 } 3087 } 3088 } 3089 } 3090 } 3091 } 3092 sx_sunlock(&allproc_lock); 3093 return npte; 3094} 3095#endif 3096 3097#if defined(DEBUG) 3098 3099static void pads(pmap_t pm); 3100void pmap_pvdump(vm_offset_t pa); 3101 3102/* print address space of pmap*/ 3103static void 3104pads(pm) 3105 pmap_t pm; 3106{ 3107 int i, j; 3108 vm_paddr_t va; 3109 pt_entry_t *ptep; 3110 3111 if (pm == kernel_pmap) 3112 return; 3113 for (i = 0; i < NPDEPTD; i++) 3114 if (pm->pm_pdir[i]) 3115 for (j = 0; j < NPTEPG; j++) { 3116 va = (i << PDRSHIFT) + (j << PAGE_SHIFT); 3117 if (pm == kernel_pmap && va < KERNBASE) 3118 continue; 3119 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS) 3120 continue; 3121 ptep = pmap_pte_quick(pm, va); 3122 if (pmap_pte_v(ptep)) 3123 printf("%x:%x ", va, *ptep); 3124 }; 3125 3126} 3127 3128void 3129pmap_pvdump(pa) 3130 vm_paddr_t pa; 3131{ 3132 pv_entry_t pv; 3133 vm_page_t m; 3134 3135 printf("pa %x", pa); 3136 m = PHYS_TO_VM_PAGE(pa); 3137 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) { 3138 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va); 3139 pads(pv->pv_pmap); 3140 } 3141 printf(" "); 3142} 3143#endif 3144