Home | History | Annotate | Line # | Download | only in arm32
pmap.c revision 1.96
      1 /*	$NetBSD: pmap.c,v 1.96 2002/04/24 17:35:10 thorpej Exp $	*/
      2 
      3 /*
      4  * Copyright (c) 2002 Wasabi Systems, Inc.
      5  * Copyright (c) 2001 Richard Earnshaw
      6  * Copyright (c) 2001 Christopher Gilbert
      7  * All rights reserved.
      8  *
      9  * 1. Redistributions of source code must retain the above copyright
     10  *    notice, this list of conditions and the following disclaimer.
     11  * 2. Redistributions in binary form must reproduce the above copyright
     12  *    notice, this list of conditions and the following disclaimer in the
     13  *    documentation and/or other materials provided with the distribution.
     14  * 3. The name of the company nor the name of the author may be used to
     15  *    endorse or promote products derived from this software without specific
     16  *    prior written permission.
     17  *
     18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
     19  * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
     20  * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
     21  * IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
     22  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
     23  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
     24  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     28  * SUCH DAMAGE.
     29  */
     30 
     31 /*-
     32  * Copyright (c) 1999 The NetBSD Foundation, Inc.
     33  * All rights reserved.
     34  *
     35  * This code is derived from software contributed to The NetBSD Foundation
     36  * by Charles M. Hannum.
     37  *
     38  * Redistribution and use in source and binary forms, with or without
     39  * modification, are permitted provided that the following conditions
     40  * are met:
     41  * 1. Redistributions of source code must retain the above copyright
     42  *    notice, this list of conditions and the following disclaimer.
     43  * 2. Redistributions in binary form must reproduce the above copyright
     44  *    notice, this list of conditions and the following disclaimer in the
     45  *    documentation and/or other materials provided with the distribution.
     46  * 3. All advertising materials mentioning features or use of this software
     47  *    must display the following acknowledgement:
     48  *        This product includes software developed by the NetBSD
     49  *        Foundation, Inc. and its contributors.
     50  * 4. Neither the name of The NetBSD Foundation nor the names of its
     51  *    contributors may be used to endorse or promote products derived
     52  *    from this software without specific prior written permission.
     53  *
     54  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     55  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     56  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     57  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     58  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     59  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     60  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     61  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     62  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     63  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     64  * POSSIBILITY OF SUCH DAMAGE.
     65  */
     66 
     67 /*
     68  * Copyright (c) 1994-1998 Mark Brinicombe.
     69  * Copyright (c) 1994 Brini.
     70  * All rights reserved.
     71  *
     72  * This code is derived from software written for Brini by Mark Brinicombe
     73  *
     74  * Redistribution and use in source and binary forms, with or without
     75  * modification, are permitted provided that the following conditions
     76  * are met:
     77  * 1. Redistributions of source code must retain the above copyright
     78  *    notice, this list of conditions and the following disclaimer.
     79  * 2. Redistributions in binary form must reproduce the above copyright
     80  *    notice, this list of conditions and the following disclaimer in the
     81  *    documentation and/or other materials provided with the distribution.
     82  * 3. All advertising materials mentioning features or use of this software
     83  *    must display the following acknowledgement:
     84  *	This product includes software developed by Mark Brinicombe.
     85  * 4. The name of the author may not be used to endorse or promote products
     86  *    derived from this software without specific prior written permission.
     87  *
     88  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
     89  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
     90  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
     91  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
     92  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
     93  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
     94  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
     95  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
     96  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
     97  *
     98  * RiscBSD kernel project
     99  *
    100  * pmap.c
    101  *
    102  * Machine dependant vm stuff
    103  *
    104  * Created      : 20/09/94
    105  */
    106 
    107 /*
    108  * Performance improvements, UVM changes, overhauls and part-rewrites
    109  * were contributed by Neil A. Carson <neil (at) causality.com>.
    110  */
    111 
    112 /*
    113  * The dram block info is currently referenced from the bootconfig.
    114  * This should be placed in a separate structure.
    115  */
    116 
    117 /*
    118  * Special compilation symbols
    119  * PMAP_DEBUG		- Build in pmap_debug_level code
    120  */
    121 
    122 /* Include header files */
    123 
    124 #include "opt_pmap_debug.h"
    125 #include "opt_ddb.h"
    126 
    127 #include <sys/types.h>
    128 #include <sys/param.h>
    129 #include <sys/kernel.h>
    130 #include <sys/systm.h>
    131 #include <sys/proc.h>
    132 #include <sys/malloc.h>
    133 #include <sys/user.h>
    134 #include <sys/pool.h>
    135 #include <sys/cdefs.h>
    136 
    137 #include <uvm/uvm.h>
    138 
    139 #include <machine/bootconfig.h>
    140 #include <machine/bus.h>
    141 #include <machine/pmap.h>
    142 #include <machine/pcb.h>
    143 #include <machine/param.h>
    144 #include <arm/arm32/katelib.h>
    145 
    146 __KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.96 2002/04/24 17:35:10 thorpej Exp $");
    147 #ifdef PMAP_DEBUG
    148 #define	PDEBUG(_lev_,_stat_) \
    149 	if (pmap_debug_level >= (_lev_)) \
    150         	((_stat_))
    151 int pmap_debug_level = -2;
    152 void pmap_dump_pvlist(vaddr_t phys, char *m);
    153 
    154 /*
    155  * for switching to potentially finer grained debugging
    156  */
    157 #define	PDB_FOLLOW	0x0001
    158 #define	PDB_INIT	0x0002
    159 #define	PDB_ENTER	0x0004
    160 #define	PDB_REMOVE	0x0008
    161 #define	PDB_CREATE	0x0010
    162 #define	PDB_PTPAGE	0x0020
    163 #define	PDB_GROWKERN	0x0040
    164 #define	PDB_BITS	0x0080
    165 #define	PDB_COLLECT	0x0100
    166 #define	PDB_PROTECT	0x0200
    167 #define	PDB_MAP_L1	0x0400
    168 #define	PDB_BOOTSTRAP	0x1000
    169 #define	PDB_PARANOIA	0x2000
    170 #define	PDB_WIRING	0x4000
    171 #define	PDB_PVDUMP	0x8000
    172 
    173 int debugmap = 0;
    174 int pmapdebug = PDB_PARANOIA | PDB_FOLLOW;
    175 #define	NPDEBUG(_lev_,_stat_) \
    176 	if (pmapdebug & (_lev_)) \
    177         	((_stat_))
    178 
    179 #else	/* PMAP_DEBUG */
    180 #define	PDEBUG(_lev_,_stat_) /* Nothing */
    181 #define NPDEBUG(_lev_,_stat_) /* Nothing */
    182 #endif	/* PMAP_DEBUG */
    183 
    184 struct pmap     kernel_pmap_store;
    185 
    186 /*
    187  * linked list of all non-kernel pmaps
    188  */
    189 
    190 static LIST_HEAD(, pmap) pmaps;
    191 
    192 /*
    193  * pool that pmap structures are allocated from
    194  */
    195 
    196 struct pool pmap_pmap_pool;
    197 
    198 static pt_entry_t *csrc_pte, *cdst_pte;
    199 static vaddr_t csrcp, cdstp;
    200 
    201 char *memhook;
    202 extern caddr_t msgbufaddr;
    203 
    204 boolean_t pmap_initialized = FALSE;	/* Has pmap_init completed? */
    205 /*
    206  * locking data structures
    207  */
    208 
    209 static struct lock pmap_main_lock;
    210 static struct simplelock pvalloc_lock;
    211 static struct simplelock pmaps_lock;
    212 #ifdef LOCKDEBUG
    213 #define PMAP_MAP_TO_HEAD_LOCK() \
    214      (void) spinlockmgr(&pmap_main_lock, LK_SHARED, NULL)
    215 #define PMAP_MAP_TO_HEAD_UNLOCK() \
    216      (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL)
    217 
    218 #define PMAP_HEAD_TO_MAP_LOCK() \
    219      (void) spinlockmgr(&pmap_main_lock, LK_EXCLUSIVE, NULL)
    220 #define PMAP_HEAD_TO_MAP_UNLOCK() \
    221      (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL)
    222 #else
    223 #define	PMAP_MAP_TO_HEAD_LOCK()		/* nothing */
    224 #define	PMAP_MAP_TO_HEAD_UNLOCK()	/* nothing */
    225 #define	PMAP_HEAD_TO_MAP_LOCK()		/* nothing */
    226 #define	PMAP_HEAD_TO_MAP_UNLOCK()	/* nothing */
    227 #endif /* LOCKDEBUG */
    228 
    229 /*
    230  * pv_page management structures: locked by pvalloc_lock
    231  */
    232 
    233 TAILQ_HEAD(pv_pagelist, pv_page);
    234 static struct pv_pagelist pv_freepages;	/* list of pv_pages with free entrys */
    235 static struct pv_pagelist pv_unusedpgs; /* list of unused pv_pages */
    236 static int pv_nfpvents;			/* # of free pv entries */
    237 static struct pv_page *pv_initpage;	/* bootstrap page from kernel_map */
    238 static vaddr_t pv_cachedva;		/* cached VA for later use */
    239 
    240 #define PVE_LOWAT (PVE_PER_PVPAGE / 2)	/* free pv_entry low water mark */
    241 #define PVE_HIWAT (PVE_LOWAT + (PVE_PER_PVPAGE * 2))
    242 					/* high water mark */
    243 
    244 /*
    245  * local prototypes
    246  */
    247 
    248 static struct pv_entry	*pmap_add_pvpage __P((struct pv_page *, boolean_t));
    249 static struct pv_entry	*pmap_alloc_pv __P((struct pmap *, int)); /* see codes below */
    250 #define ALLOCPV_NEED	0	/* need PV now */
    251 #define ALLOCPV_TRY	1	/* just try to allocate, don't steal */
    252 #define ALLOCPV_NONEED	2	/* don't need PV, just growing cache */
    253 static struct pv_entry	*pmap_alloc_pvpage __P((struct pmap *, int));
    254 static void		 pmap_enter_pv __P((struct vm_page *,
    255 					    struct pv_entry *, struct pmap *,
    256 					    vaddr_t, struct vm_page *, int));
    257 static void		 pmap_free_pv __P((struct pmap *, struct pv_entry *));
    258 static void		 pmap_free_pvs __P((struct pmap *, struct pv_entry *));
    259 static void		 pmap_free_pv_doit __P((struct pv_entry *));
    260 static void		 pmap_free_pvpage __P((void));
    261 static boolean_t	 pmap_is_curpmap __P((struct pmap *));
    262 static struct pv_entry	*pmap_remove_pv __P((struct vm_page *, struct pmap *,
    263 			vaddr_t));
    264 #define PMAP_REMOVE_ALL		0	/* remove all mappings */
    265 #define PMAP_REMOVE_SKIPWIRED	1	/* skip wired mappings */
    266 
    267 static u_int pmap_modify_pv __P((struct pmap *, vaddr_t, struct vm_page *,
    268 	u_int, u_int));
    269 
    270 /*
    271  * Structure that describes and L1 table.
    272  */
    273 struct l1pt {
    274 	SIMPLEQ_ENTRY(l1pt)	pt_queue;	/* Queue pointers */
    275 	struct pglist		pt_plist;	/* Allocated page list */
    276 	vaddr_t			pt_va;		/* Allocated virtual address */
    277 	int			pt_flags;	/* Flags */
    278 };
    279 #define	PTFLAG_STATIC		0x01		/* Statically allocated */
    280 #define	PTFLAG_KPT		0x02		/* Kernel pt's are mapped */
    281 #define	PTFLAG_CLEAN		0x04		/* L1 is clean */
    282 
    283 static void pmap_free_l1pt __P((struct l1pt *));
    284 static int pmap_allocpagedir __P((struct pmap *));
    285 static int pmap_clean_page __P((struct pv_entry *, boolean_t));
    286 static void pmap_remove_all __P((struct vm_page *));
    287 
    288 static int pmap_alloc_ptpt(struct pmap *);
    289 static void pmap_free_ptpt(struct pmap *);
    290 
    291 static struct vm_page	*pmap_alloc_ptp __P((struct pmap *, vaddr_t));
    292 static struct vm_page	*pmap_get_ptp __P((struct pmap *, vaddr_t));
    293 __inline static void pmap_clearbit __P((struct vm_page *, unsigned int));
    294 
    295 extern paddr_t physical_start;
    296 extern paddr_t physical_freestart;
    297 extern paddr_t physical_end;
    298 extern paddr_t physical_freeend;
    299 extern unsigned int free_pages;
    300 extern int max_processes;
    301 
    302 vaddr_t virtual_avail;
    303 vaddr_t virtual_end;
    304 vaddr_t pmap_curmaxkvaddr;
    305 
    306 vaddr_t avail_start;
    307 vaddr_t avail_end;
    308 
    309 extern pv_addr_t systempage;
    310 
    311 /* Variables used by the L1 page table queue code */
    312 SIMPLEQ_HEAD(l1pt_queue, l1pt);
    313 static struct l1pt_queue l1pt_static_queue; /* head of our static l1 queue */
    314 static int l1pt_static_queue_count;	    /* items in the static l1 queue */
    315 static int l1pt_static_create_count;	    /* static l1 items created */
    316 static struct l1pt_queue l1pt_queue;	    /* head of our l1 queue */
    317 static int l1pt_queue_count;		    /* items in the l1 queue */
    318 static int l1pt_create_count;		    /* stat - L1's create count */
    319 static int l1pt_reuse_count;		    /* stat - L1's reused count */
    320 
    321 /* Local function prototypes (not used outside this file) */
    322 void pmap_pinit __P((struct pmap *));
    323 void pmap_freepagedir __P((struct pmap *));
    324 
    325 /* Other function prototypes */
    326 extern void bzero_page __P((vaddr_t));
    327 extern void bcopy_page __P((vaddr_t, vaddr_t));
    328 
    329 struct l1pt *pmap_alloc_l1pt __P((void));
    330 static __inline void pmap_map_in_l1 __P((struct pmap *pmap, vaddr_t va,
    331      vaddr_t l2pa, boolean_t));
    332 
    333 static pt_entry_t *pmap_map_ptes __P((struct pmap *));
    334 static void pmap_unmap_ptes __P((struct pmap *));
    335 
    336 __inline static void pmap_vac_me_harder __P((struct pmap *, struct vm_page *,
    337     pt_entry_t *, boolean_t));
    338 static void pmap_vac_me_kpmap __P((struct pmap *, struct vm_page *,
    339     pt_entry_t *, boolean_t));
    340 static void pmap_vac_me_user __P((struct pmap *, struct vm_page *,
    341     pt_entry_t *, boolean_t));
    342 
    343 /*
    344  * real definition of pv_entry.
    345  */
    346 
    347 struct pv_entry {
    348 	struct pv_entry *pv_next;       /* next pv_entry */
    349 	struct pmap     *pv_pmap;        /* pmap where mapping lies */
    350 	vaddr_t         pv_va;          /* virtual address for mapping */
    351 	int             pv_flags;       /* flags */
    352 	struct vm_page	*pv_ptp;	/* vm_page for the ptp */
    353 };
    354 
    355 /*
    356  * pv_entrys are dynamically allocated in chunks from a single page.
    357  * we keep track of how many pv_entrys are in use for each page and
    358  * we can free pv_entry pages if needed.  there is one lock for the
    359  * entire allocation system.
    360  */
    361 
    362 struct pv_page_info {
    363 	TAILQ_ENTRY(pv_page) pvpi_list;
    364 	struct pv_entry *pvpi_pvfree;
    365 	int pvpi_nfree;
    366 };
    367 
    368 /*
    369  * number of pv_entry's in a pv_page
    370  * (note: won't work on systems where NPBG isn't a constant)
    371  */
    372 
    373 #define PVE_PER_PVPAGE ((NBPG - sizeof(struct pv_page_info)) / \
    374 			sizeof(struct pv_entry))
    375 
    376 /*
    377  * a pv_page: where pv_entrys are allocated from
    378  */
    379 
    380 struct pv_page {
    381 	struct pv_page_info pvinfo;
    382 	struct pv_entry pvents[PVE_PER_PVPAGE];
    383 };
    384 
    385 #ifdef MYCROFT_HACK
    386 int mycroft_hack = 0;
    387 #endif
    388 
    389 /* Function to set the debug level of the pmap code */
    390 
    391 #ifdef PMAP_DEBUG
    392 void
    393 pmap_debug(int level)
    394 {
    395 	pmap_debug_level = level;
    396 	printf("pmap_debug: level=%d\n", pmap_debug_level);
    397 }
    398 #endif	/* PMAP_DEBUG */
    399 
    400 __inline static boolean_t
    401 pmap_is_curpmap(struct pmap *pmap)
    402 {
    403 
    404 	if ((curproc && curproc->p_vmspace->vm_map.pmap == pmap) ||
    405 	    pmap == pmap_kernel())
    406 		return (TRUE);
    407 
    408 	return (FALSE);
    409 }
    410 
    411 #include "isadma.h"
    412 
    413 #if NISADMA > 0
    414 /*
    415  * Used to protect memory for ISA DMA bounce buffers.  If, when loading
    416  * pages into the system, memory intersects with any of these ranges,
    417  * the intersecting memory will be loaded into a lower-priority free list.
    418  */
    419 bus_dma_segment_t *pmap_isa_dma_ranges;
    420 int pmap_isa_dma_nranges;
    421 
    422 /*
    423  * Check if a memory range intersects with an ISA DMA range, and
    424  * return the page-rounded intersection if it does.  The intersection
    425  * will be placed on a lower-priority free list.
    426  */
    427 static boolean_t
    428 pmap_isa_dma_range_intersect(paddr_t pa, psize_t size, paddr_t *pap,
    429     psize_t *sizep)
    430 {
    431 	bus_dma_segment_t *ds;
    432 	int i;
    433 
    434 	if (pmap_isa_dma_ranges == NULL)
    435 		return (FALSE);
    436 
    437 	for (i = 0, ds = pmap_isa_dma_ranges;
    438 	     i < pmap_isa_dma_nranges; i++, ds++) {
    439 		if (ds->ds_addr <= pa && pa < (ds->ds_addr + ds->ds_len)) {
    440 			/*
    441 			 * Beginning of region intersects with this range.
    442 			 */
    443 			*pap = trunc_page(pa);
    444 			*sizep = round_page(min(pa + size,
    445 			    ds->ds_addr + ds->ds_len) - pa);
    446 			return (TRUE);
    447 		}
    448 		if (pa < ds->ds_addr && ds->ds_addr < (pa + size)) {
    449 			/*
    450 			 * End of region intersects with this range.
    451 			 */
    452 			*pap = trunc_page(ds->ds_addr);
    453 			*sizep = round_page(min((pa + size) - ds->ds_addr,
    454 			    ds->ds_len));
    455 			return (TRUE);
    456 		}
    457 	}
    458 
    459 	/*
    460 	 * No intersection found.
    461 	 */
    462 	return (FALSE);
    463 }
    464 #endif /* NISADMA > 0 */
    465 
    466 /*
    467  * p v _ e n t r y   f u n c t i o n s
    468  */
    469 
    470 /*
    471  * pv_entry allocation functions:
    472  *   the main pv_entry allocation functions are:
    473  *     pmap_alloc_pv: allocate a pv_entry structure
    474  *     pmap_free_pv: free one pv_entry
    475  *     pmap_free_pvs: free a list of pv_entrys
    476  *
    477  * the rest are helper functions
    478  */
    479 
    480 /*
    481  * pmap_alloc_pv: inline function to allocate a pv_entry structure
    482  * => we lock pvalloc_lock
    483  * => if we fail, we call out to pmap_alloc_pvpage
    484  * => 3 modes:
    485  *    ALLOCPV_NEED   = we really need a pv_entry, even if we have to steal it
    486  *    ALLOCPV_TRY    = we want a pv_entry, but not enough to steal
    487  *    ALLOCPV_NONEED = we are trying to grow our free list, don't really need
    488  *			one now
    489  *
    490  * "try" is for optional functions like pmap_copy().
    491  */
    492 
    493 __inline static struct pv_entry *
    494 pmap_alloc_pv(struct pmap *pmap, int mode)
    495 {
    496 	struct pv_page *pvpage;
    497 	struct pv_entry *pv;
    498 
    499 	simple_lock(&pvalloc_lock);
    500 
    501 	pvpage = TAILQ_FIRST(&pv_freepages);
    502 
    503 	if (pvpage != NULL) {
    504 		pvpage->pvinfo.pvpi_nfree--;
    505 		if (pvpage->pvinfo.pvpi_nfree == 0) {
    506 			/* nothing left in this one? */
    507 			TAILQ_REMOVE(&pv_freepages, pvpage, pvinfo.pvpi_list);
    508 		}
    509 		pv = pvpage->pvinfo.pvpi_pvfree;
    510 		KASSERT(pv);
    511 		pvpage->pvinfo.pvpi_pvfree = pv->pv_next;
    512 		pv_nfpvents--;  /* took one from pool */
    513 	} else {
    514 		pv = NULL;		/* need more of them */
    515 	}
    516 
    517 	/*
    518 	 * if below low water mark or we didn't get a pv_entry we try and
    519 	 * create more pv_entrys ...
    520 	 */
    521 
    522 	if (pv_nfpvents < PVE_LOWAT || pv == NULL) {
    523 		if (pv == NULL)
    524 			pv = pmap_alloc_pvpage(pmap, (mode == ALLOCPV_TRY) ?
    525 					       mode : ALLOCPV_NEED);
    526 		else
    527 			(void) pmap_alloc_pvpage(pmap, ALLOCPV_NONEED);
    528 	}
    529 
    530 	simple_unlock(&pvalloc_lock);
    531 	return(pv);
    532 }
    533 
    534 /*
    535  * pmap_alloc_pvpage: maybe allocate a new pvpage
    536  *
    537  * if need_entry is false: try and allocate a new pv_page
    538  * if need_entry is true: try and allocate a new pv_page and return a
    539  *	new pv_entry from it.   if we are unable to allocate a pv_page
    540  *	we make a last ditch effort to steal a pv_page from some other
    541  *	mapping.    if that fails, we panic...
    542  *
    543  * => we assume that the caller holds pvalloc_lock
    544  */
    545 
    546 static struct pv_entry *
    547 pmap_alloc_pvpage(struct pmap *pmap, int mode)
    548 {
    549 	struct vm_page *pg;
    550 	struct pv_page *pvpage;
    551 	struct pv_entry *pv;
    552 	int s;
    553 
    554 	/*
    555 	 * if we need_entry and we've got unused pv_pages, allocate from there
    556 	 */
    557 
    558 	pvpage = TAILQ_FIRST(&pv_unusedpgs);
    559 	if (mode != ALLOCPV_NONEED && pvpage != NULL) {
    560 
    561 		/* move it to pv_freepages list */
    562 		TAILQ_REMOVE(&pv_unusedpgs, pvpage, pvinfo.pvpi_list);
    563 		TAILQ_INSERT_HEAD(&pv_freepages, pvpage, pvinfo.pvpi_list);
    564 
    565 		/* allocate a pv_entry */
    566 		pvpage->pvinfo.pvpi_nfree--;	/* can't go to zero */
    567 		pv = pvpage->pvinfo.pvpi_pvfree;
    568 		KASSERT(pv);
    569 		pvpage->pvinfo.pvpi_pvfree = pv->pv_next;
    570 
    571 		pv_nfpvents--;  /* took one from pool */
    572 		return(pv);
    573 	}
    574 
    575 	/*
    576 	 *  see if we've got a cached unmapped VA that we can map a page in.
    577 	 * if not, try to allocate one.
    578 	 */
    579 
    580 
    581 	if (pv_cachedva == 0) {
    582 		s = splvm();
    583 		pv_cachedva = uvm_km_kmemalloc(kmem_map, NULL,
    584 		    PAGE_SIZE, UVM_KMF_TRYLOCK|UVM_KMF_VALLOC);
    585 		splx(s);
    586 		if (pv_cachedva == 0) {
    587 			return (NULL);
    588 		}
    589 	}
    590 
    591 	pg = uvm_pagealloc(NULL, pv_cachedva - vm_map_min(kernel_map), NULL,
    592 	    UVM_PGA_USERESERVE);
    593 
    594 	if (pg == NULL)
    595 		return (NULL);
    596 	pg->flags &= ~PG_BUSY;	/* never busy */
    597 
    598 	/*
    599 	 * add a mapping for our new pv_page and free its entrys (save one!)
    600 	 *
    601 	 * NOTE: If we are allocating a PV page for the kernel pmap, the
    602 	 * pmap is already locked!  (...but entering the mapping is safe...)
    603 	 */
    604 
    605 	pmap_kenter_pa(pv_cachedva, VM_PAGE_TO_PHYS(pg),
    606 		VM_PROT_READ|VM_PROT_WRITE);
    607 	pmap_update(pmap_kernel());
    608 	pvpage = (struct pv_page *) pv_cachedva;
    609 	pv_cachedva = 0;
    610 	return (pmap_add_pvpage(pvpage, mode != ALLOCPV_NONEED));
    611 }
    612 
    613 /*
    614  * pmap_add_pvpage: add a pv_page's pv_entrys to the free list
    615  *
    616  * => caller must hold pvalloc_lock
    617  * => if need_entry is true, we allocate and return one pv_entry
    618  */
    619 
    620 static struct pv_entry *
    621 pmap_add_pvpage(struct pv_page *pvp, boolean_t need_entry)
    622 {
    623 	int tofree, lcv;
    624 
    625 	/* do we need to return one? */
    626 	tofree = (need_entry) ? PVE_PER_PVPAGE - 1 : PVE_PER_PVPAGE;
    627 
    628 	pvp->pvinfo.pvpi_pvfree = NULL;
    629 	pvp->pvinfo.pvpi_nfree = tofree;
    630 	for (lcv = 0 ; lcv < tofree ; lcv++) {
    631 		pvp->pvents[lcv].pv_next = pvp->pvinfo.pvpi_pvfree;
    632 		pvp->pvinfo.pvpi_pvfree = &pvp->pvents[lcv];
    633 	}
    634 	if (need_entry)
    635 		TAILQ_INSERT_TAIL(&pv_freepages, pvp, pvinfo.pvpi_list);
    636 	else
    637 		TAILQ_INSERT_TAIL(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
    638 	pv_nfpvents += tofree;
    639 	return((need_entry) ? &pvp->pvents[lcv] : NULL);
    640 }
    641 
    642 /*
    643  * pmap_free_pv_doit: actually free a pv_entry
    644  *
    645  * => do not call this directly!  instead use either
    646  *    1. pmap_free_pv ==> free a single pv_entry
    647  *    2. pmap_free_pvs => free a list of pv_entrys
    648  * => we must be holding pvalloc_lock
    649  */
    650 
    651 __inline static void
    652 pmap_free_pv_doit(struct pv_entry *pv)
    653 {
    654 	struct pv_page *pvp;
    655 
    656 	pvp = (struct pv_page *) arm_trunc_page((vaddr_t)pv);
    657 	pv_nfpvents++;
    658 	pvp->pvinfo.pvpi_nfree++;
    659 
    660 	/* nfree == 1 => fully allocated page just became partly allocated */
    661 	if (pvp->pvinfo.pvpi_nfree == 1) {
    662 		TAILQ_INSERT_HEAD(&pv_freepages, pvp, pvinfo.pvpi_list);
    663 	}
    664 
    665 	/* free it */
    666 	pv->pv_next = pvp->pvinfo.pvpi_pvfree;
    667 	pvp->pvinfo.pvpi_pvfree = pv;
    668 
    669 	/*
    670 	 * are all pv_page's pv_entry's free?  move it to unused queue.
    671 	 */
    672 
    673 	if (pvp->pvinfo.pvpi_nfree == PVE_PER_PVPAGE) {
    674 		TAILQ_REMOVE(&pv_freepages, pvp, pvinfo.pvpi_list);
    675 		TAILQ_INSERT_HEAD(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
    676 	}
    677 }
    678 
    679 /*
    680  * pmap_free_pv: free a single pv_entry
    681  *
    682  * => we gain the pvalloc_lock
    683  */
    684 
    685 __inline static void
    686 pmap_free_pv(struct pmap *pmap, struct pv_entry *pv)
    687 {
    688 	simple_lock(&pvalloc_lock);
    689 	pmap_free_pv_doit(pv);
    690 
    691 	/*
    692 	 * Can't free the PV page if the PV entries were associated with
    693 	 * the kernel pmap; the pmap is already locked.
    694 	 */
    695 	if (pv_nfpvents > PVE_HIWAT && TAILQ_FIRST(&pv_unusedpgs) != NULL &&
    696 	    pmap != pmap_kernel())
    697 		pmap_free_pvpage();
    698 
    699 	simple_unlock(&pvalloc_lock);
    700 }
    701 
    702 /*
    703  * pmap_free_pvs: free a list of pv_entrys
    704  *
    705  * => we gain the pvalloc_lock
    706  */
    707 
    708 __inline static void
    709 pmap_free_pvs(struct pmap *pmap, struct pv_entry *pvs)
    710 {
    711 	struct pv_entry *nextpv;
    712 
    713 	simple_lock(&pvalloc_lock);
    714 
    715 	for ( /* null */ ; pvs != NULL ; pvs = nextpv) {
    716 		nextpv = pvs->pv_next;
    717 		pmap_free_pv_doit(pvs);
    718 	}
    719 
    720 	/*
    721 	 * Can't free the PV page if the PV entries were associated with
    722 	 * the kernel pmap; the pmap is already locked.
    723 	 */
    724 	if (pv_nfpvents > PVE_HIWAT && TAILQ_FIRST(&pv_unusedpgs) != NULL &&
    725 	    pmap != pmap_kernel())
    726 		pmap_free_pvpage();
    727 
    728 	simple_unlock(&pvalloc_lock);
    729 }
    730 
    731 
    732 /*
    733  * pmap_free_pvpage: try and free an unused pv_page structure
    734  *
    735  * => assume caller is holding the pvalloc_lock and that
    736  *	there is a page on the pv_unusedpgs list
    737  * => if we can't get a lock on the kmem_map we try again later
    738  */
    739 
    740 static void
    741 pmap_free_pvpage(void)
    742 {
    743 	int s;
    744 	struct vm_map *map;
    745 	struct vm_map_entry *dead_entries;
    746 	struct pv_page *pvp;
    747 
    748 	s = splvm(); /* protect kmem_map */
    749 
    750 	pvp = TAILQ_FIRST(&pv_unusedpgs);
    751 
    752 	/*
    753 	 * note: watch out for pv_initpage which is allocated out of
    754 	 * kernel_map rather than kmem_map.
    755 	 */
    756 	if (pvp == pv_initpage)
    757 		map = kernel_map;
    758 	else
    759 		map = kmem_map;
    760 	if (vm_map_lock_try(map)) {
    761 
    762 		/* remove pvp from pv_unusedpgs */
    763 		TAILQ_REMOVE(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
    764 
    765 		/* unmap the page */
    766 		dead_entries = NULL;
    767 		uvm_unmap_remove(map, (vaddr_t)pvp, ((vaddr_t)pvp) + PAGE_SIZE,
    768 		    &dead_entries);
    769 		vm_map_unlock(map);
    770 
    771 		if (dead_entries != NULL)
    772 			uvm_unmap_detach(dead_entries, 0);
    773 
    774 		pv_nfpvents -= PVE_PER_PVPAGE;  /* update free count */
    775 	}
    776 	if (pvp == pv_initpage)
    777 		/* no more initpage, we've freed it */
    778 		pv_initpage = NULL;
    779 
    780 	splx(s);
    781 }
    782 
    783 /*
    784  * main pv_entry manipulation functions:
    785  *   pmap_enter_pv: enter a mapping onto a vm_page list
    786  *   pmap_remove_pv: remove a mappiing from a vm_page list
    787  *
    788  * NOTE: pmap_enter_pv expects to lock the pvh itself
    789  *       pmap_remove_pv expects te caller to lock the pvh before calling
    790  */
    791 
    792 /*
    793  * pmap_enter_pv: enter a mapping onto a vm_page lst
    794  *
    795  * => caller should hold the proper lock on pmap_main_lock
    796  * => caller should have pmap locked
    797  * => we will gain the lock on the vm_page and allocate the new pv_entry
    798  * => caller should adjust ptp's wire_count before calling
    799  * => caller should not adjust pmap's wire_count
    800  */
    801 
    802 __inline static void
    803 pmap_enter_pv(struct vm_page *pg, struct pv_entry *pve, struct pmap *pmap,
    804     vaddr_t va, struct vm_page *ptp, int flags)
    805 {
    806 	pve->pv_pmap = pmap;
    807 	pve->pv_va = va;
    808 	pve->pv_ptp = ptp;			/* NULL for kernel pmap */
    809 	pve->pv_flags = flags;
    810 	simple_lock(&pg->mdpage.pvh_slock);	/* lock vm_page */
    811 	pve->pv_next = pg->mdpage.pvh_list;	/* add to ... */
    812 	pg->mdpage.pvh_list = pve;		/* ... locked list */
    813 	simple_unlock(&pg->mdpage.pvh_slock);	/* unlock, done! */
    814 	if (pve->pv_flags & PVF_WIRED)
    815 		++pmap->pm_stats.wired_count;
    816 }
    817 
    818 /*
    819  * pmap_remove_pv: try to remove a mapping from a pv_list
    820  *
    821  * => caller should hold proper lock on pmap_main_lock
    822  * => pmap should be locked
    823  * => caller should hold lock on vm_page [so that attrs can be adjusted]
    824  * => caller should adjust ptp's wire_count and free PTP if needed
    825  * => caller should NOT adjust pmap's wire_count
    826  * => we return the removed pve
    827  */
    828 
    829 __inline static struct pv_entry *
    830 pmap_remove_pv(struct vm_page *pg, struct pmap *pmap, vaddr_t va)
    831 {
    832 	struct pv_entry *pve, **prevptr;
    833 
    834 	prevptr = &pg->mdpage.pvh_list;		/* previous pv_entry pointer */
    835 	pve = *prevptr;
    836 	while (pve) {
    837 		if (pve->pv_pmap == pmap && pve->pv_va == va) {	/* match? */
    838 			*prevptr = pve->pv_next;		/* remove it! */
    839 			if (pve->pv_flags & PVF_WIRED)
    840 			    --pmap->pm_stats.wired_count;
    841 			break;
    842 		}
    843 		prevptr = &pve->pv_next;		/* previous pointer */
    844 		pve = pve->pv_next;			/* advance */
    845 	}
    846 	return(pve);				/* return removed pve */
    847 }
    848 
    849 /*
    850  *
    851  * pmap_modify_pv: Update pv flags
    852  *
    853  * => caller should hold lock on vm_page [so that attrs can be adjusted]
    854  * => caller should NOT adjust pmap's wire_count
    855  * => caller must call pmap_vac_me_harder() if writable status of a page
    856  *    may have changed.
    857  * => we return the old flags
    858  *
    859  * Modify a physical-virtual mapping in the pv table
    860  */
    861 
    862 static /* __inline */ u_int
    863 pmap_modify_pv(struct pmap *pmap, vaddr_t va, struct vm_page *pg,
    864     u_int bic_mask, u_int eor_mask)
    865 {
    866 	struct pv_entry *npv;
    867 	u_int flags, oflags;
    868 
    869 	/*
    870 	 * There is at least one VA mapping this page.
    871 	 */
    872 
    873 	for (npv = pg->mdpage.pvh_list; npv; npv = npv->pv_next) {
    874 		if (pmap == npv->pv_pmap && va == npv->pv_va) {
    875 			oflags = npv->pv_flags;
    876 			npv->pv_flags = flags =
    877 			    ((oflags & ~bic_mask) ^ eor_mask);
    878 			if ((flags ^ oflags) & PVF_WIRED) {
    879 				if (flags & PVF_WIRED)
    880 					++pmap->pm_stats.wired_count;
    881 				else
    882 					--pmap->pm_stats.wired_count;
    883 			}
    884 			return (oflags);
    885 		}
    886 	}
    887 	return (0);
    888 }
    889 
    890 /*
    891  * Map the specified level 2 pagetable into the level 1 page table for
    892  * the given pmap to cover a chunk of virtual address space starting from the
    893  * address specified.
    894  */
    895 static __inline void
    896 pmap_map_in_l1(struct pmap *pmap, vaddr_t va, paddr_t l2pa, boolean_t selfref)
    897 {
    898 	vaddr_t ptva;
    899 
    900 	/* Calculate the index into the L1 page table. */
    901 	ptva = (va >> L1_S_SHIFT) & ~3;
    902 
    903 	/* Map page table into the L1. */
    904 	pmap->pm_pdir[ptva + 0] = L1_C_PROTO | (l2pa + 0x000);
    905 	pmap->pm_pdir[ptva + 1] = L1_C_PROTO | (l2pa + 0x400);
    906 	pmap->pm_pdir[ptva + 2] = L1_C_PROTO | (l2pa + 0x800);
    907 	pmap->pm_pdir[ptva + 3] = L1_C_PROTO | (l2pa + 0xc00);
    908 
    909 	/* Map the page table into the page table area. */
    910 	if (selfref)
    911 		*((pt_entry_t *)(pmap->pm_vptpt + ptva)) = L2_S_PROTO | l2pa |
    912 		    L2_S_PROT(PTE_KERNEL, VM_PROT_READ|VM_PROT_WRITE);
    913 }
    914 
    915 #if 0
    916 static __inline void
    917 pmap_unmap_in_l1(struct pmap *pmap, vaddr_t va)
    918 {
    919 	vaddr_t ptva;
    920 
    921 	/* Calculate the index into the L1 page table. */
    922 	ptva = (va >> L1_S_SHIFT) & ~3;
    923 
    924 	/* Unmap page table from the L1. */
    925 	pmap->pm_pdir[ptva + 0] = 0;
    926 	pmap->pm_pdir[ptva + 1] = 0;
    927 	pmap->pm_pdir[ptva + 2] = 0;
    928 	pmap->pm_pdir[ptva + 3] = 0;
    929 
    930 	/* Unmap the page table from the page table area. */
    931 	*((pt_entry_t *)(pmap->pm_vptpt + ptva)) = 0;
    932 }
    933 #endif
    934 
    935 /*
    936  *	Used to map a range of physical addresses into kernel
    937  *	virtual address space.
    938  *
    939  *	For now, VM is already on, we only need to map the
    940  *	specified memory.
    941  */
    942 vaddr_t
    943 pmap_map(vaddr_t va, paddr_t spa, paddr_t epa, vm_prot_t prot)
    944 {
    945 	while (spa < epa) {
    946 		pmap_kenter_pa(va, spa, prot);
    947 		va += NBPG;
    948 		spa += NBPG;
    949 	}
    950 	pmap_update(pmap_kernel());
    951 	return(va);
    952 }
    953 
    954 
    955 /*
    956  * void pmap_bootstrap(pd_entry_t *kernel_l1pt, pv_addr_t kernel_ptpt)
    957  *
    958  * bootstrap the pmap system. This is called from initarm and allows
    959  * the pmap system to initailise any structures it requires.
    960  *
    961  * Currently this sets up the kernel_pmap that is statically allocated
    962  * and also allocated virtual addresses for certain page hooks.
    963  * Currently the only one page hook is allocated that is used
    964  * to zero physical pages of memory.
    965  * It also initialises the start and end address of the kernel data space.
    966  */
    967 extern paddr_t physical_freestart;
    968 extern paddr_t physical_freeend;
    969 
    970 char *boot_head;
    971 
    972 void
    973 pmap_bootstrap(pd_entry_t *kernel_l1pt, pv_addr_t kernel_ptpt)
    974 {
    975 	pt_entry_t *pte;
    976 	int loop;
    977 	paddr_t start, end;
    978 #if NISADMA > 0
    979 	paddr_t istart;
    980 	psize_t isize;
    981 #endif
    982 
    983 	pmap_kernel()->pm_pdir = kernel_l1pt;
    984 	pmap_kernel()->pm_pptpt = kernel_ptpt.pv_pa;
    985 	pmap_kernel()->pm_vptpt = kernel_ptpt.pv_va;
    986 	simple_lock_init(&pmap_kernel()->pm_lock);
    987 	pmap_kernel()->pm_obj.pgops = NULL;
    988 	TAILQ_INIT(&(pmap_kernel()->pm_obj.memq));
    989 	pmap_kernel()->pm_obj.uo_npages = 0;
    990 	pmap_kernel()->pm_obj.uo_refs = 1;
    991 
    992 	/*
    993 	 * Initialize PAGE_SIZE-dependent variables.
    994 	 */
    995 	uvm_setpagesize();
    996 
    997 	loop = 0;
    998 	while (loop < bootconfig.dramblocks) {
    999 		start = (paddr_t)bootconfig.dram[loop].address;
   1000 		end = start + (bootconfig.dram[loop].pages * NBPG);
   1001 		if (start < physical_freestart)
   1002 			start = physical_freestart;
   1003 		if (end > physical_freeend)
   1004 			end = physical_freeend;
   1005 #if 0
   1006 		printf("%d: %lx -> %lx\n", loop, start, end - 1);
   1007 #endif
   1008 #if NISADMA > 0
   1009 		if (pmap_isa_dma_range_intersect(start, end - start,
   1010 		    &istart, &isize)) {
   1011 			/*
   1012 			 * Place the pages that intersect with the
   1013 			 * ISA DMA range onto the ISA DMA free list.
   1014 			 */
   1015 #if 0
   1016 			printf("    ISADMA 0x%lx -> 0x%lx\n", istart,
   1017 			    istart + isize - 1);
   1018 #endif
   1019 			uvm_page_physload(atop(istart),
   1020 			    atop(istart + isize), atop(istart),
   1021 			    atop(istart + isize), VM_FREELIST_ISADMA);
   1022 
   1023 			/*
   1024 			 * Load the pieces that come before
   1025 			 * the intersection into the default
   1026 			 * free list.
   1027 			 */
   1028 			if (start < istart) {
   1029 #if 0
   1030 				printf("    BEFORE 0x%lx -> 0x%lx\n",
   1031 				    start, istart - 1);
   1032 #endif
   1033 				uvm_page_physload(atop(start),
   1034 				    atop(istart), atop(start),
   1035 				    atop(istart), VM_FREELIST_DEFAULT);
   1036 			}
   1037 
   1038 			/*
   1039 			 * Load the pieces that come after
   1040 			 * the intersection into the default
   1041 			 * free list.
   1042 			 */
   1043 			if ((istart + isize) < end) {
   1044 #if 0
   1045 				printf("     AFTER 0x%lx -> 0x%lx\n",
   1046 				    (istart + isize), end - 1);
   1047 #endif
   1048 				uvm_page_physload(atop(istart + isize),
   1049 				    atop(end), atop(istart + isize),
   1050 				    atop(end), VM_FREELIST_DEFAULT);
   1051 			}
   1052 		} else {
   1053 			uvm_page_physload(atop(start), atop(end),
   1054 			    atop(start), atop(end), VM_FREELIST_DEFAULT);
   1055 		}
   1056 #else	/* NISADMA > 0 */
   1057 		uvm_page_physload(atop(start), atop(end),
   1058 		    atop(start), atop(end), VM_FREELIST_DEFAULT);
   1059 #endif /* NISADMA > 0 */
   1060 		++loop;
   1061 	}
   1062 
   1063 	virtual_avail = KERNEL_VM_BASE;
   1064 	virtual_end = KERNEL_VM_BASE + KERNEL_VM_SIZE;
   1065 
   1066 	/*
   1067 	 * now we allocate the "special" VAs which are used for tmp mappings
   1068 	 * by the pmap (and other modules).  we allocate the VAs by advancing
   1069 	 * virtual_avail (note that there are no pages mapped at these VAs).
   1070 	 * we find the PTE that maps the allocated VA via the linear PTE
   1071 	 * mapping.
   1072 	 */
   1073 
   1074 	pte = ((pt_entry_t *) PTE_BASE) + atop(virtual_avail);
   1075 
   1076 	csrcp = virtual_avail; csrc_pte = pte;
   1077 	virtual_avail += PAGE_SIZE; pte++;
   1078 
   1079 	cdstp = virtual_avail; cdst_pte = pte;
   1080 	virtual_avail += PAGE_SIZE; pte++;
   1081 
   1082 	memhook = (char *) virtual_avail;	/* don't need pte */
   1083 	virtual_avail += PAGE_SIZE; pte++;
   1084 
   1085 	msgbufaddr = (caddr_t) virtual_avail;	/* don't need pte */
   1086 	virtual_avail += round_page(MSGBUFSIZE);
   1087 	pte += atop(round_page(MSGBUFSIZE));
   1088 
   1089 	/*
   1090 	 * init the static-global locks and global lists.
   1091 	 */
   1092 	spinlockinit(&pmap_main_lock, "pmaplk", 0);
   1093 	simple_lock_init(&pvalloc_lock);
   1094 	simple_lock_init(&pmaps_lock);
   1095 	LIST_INIT(&pmaps);
   1096 	TAILQ_INIT(&pv_freepages);
   1097 	TAILQ_INIT(&pv_unusedpgs);
   1098 
   1099 	/*
   1100 	 * initialize the pmap pool.
   1101 	 */
   1102 
   1103 	pool_init(&pmap_pmap_pool, sizeof(struct pmap), 0, 0, 0, "pmappl",
   1104 		  &pool_allocator_nointr);
   1105 
   1106 	cpu_dcache_wbinv_all();
   1107 }
   1108 
   1109 /*
   1110  * void pmap_init(void)
   1111  *
   1112  * Initialize the pmap module.
   1113  * Called by vm_init() in vm/vm_init.c in order to initialise
   1114  * any structures that the pmap system needs to map virtual memory.
   1115  */
   1116 
   1117 extern int physmem;
   1118 
   1119 void
   1120 pmap_init(void)
   1121 {
   1122 
   1123 	/*
   1124 	 * Set the available memory vars - These do not map to real memory
   1125 	 * addresses and cannot as the physical memory is fragmented.
   1126 	 * They are used by ps for %mem calculations.
   1127 	 * One could argue whether this should be the entire memory or just
   1128 	 * the memory that is useable in a user process.
   1129 	 */
   1130 	avail_start = 0;
   1131 	avail_end = physmem * NBPG;
   1132 
   1133 	/*
   1134 	 * now we need to free enough pv_entry structures to allow us to get
   1135 	 * the kmem_map/kmem_object allocated and inited (done after this
   1136 	 * function is finished).  to do this we allocate one bootstrap page out
   1137 	 * of kernel_map and use it to provide an initial pool of pv_entry
   1138 	 * structures.   we never free this page.
   1139 	 */
   1140 
   1141 	pv_initpage = (struct pv_page *) uvm_km_alloc(kernel_map, PAGE_SIZE);
   1142 	if (pv_initpage == NULL)
   1143 		panic("pmap_init: pv_initpage");
   1144 	pv_cachedva = 0;   /* a VA we have allocated but not used yet */
   1145 	pv_nfpvents = 0;
   1146 	(void) pmap_add_pvpage(pv_initpage, FALSE);
   1147 
   1148 	pmap_initialized = TRUE;
   1149 
   1150 	/* Initialise our L1 page table queues and counters */
   1151 	SIMPLEQ_INIT(&l1pt_static_queue);
   1152 	l1pt_static_queue_count = 0;
   1153 	l1pt_static_create_count = 0;
   1154 	SIMPLEQ_INIT(&l1pt_queue);
   1155 	l1pt_queue_count = 0;
   1156 	l1pt_create_count = 0;
   1157 	l1pt_reuse_count = 0;
   1158 }
   1159 
   1160 /*
   1161  * pmap_postinit()
   1162  *
   1163  * This routine is called after the vm and kmem subsystems have been
   1164  * initialised. This allows the pmap code to perform any initialisation
   1165  * that can only be done one the memory allocation is in place.
   1166  */
   1167 
   1168 void
   1169 pmap_postinit(void)
   1170 {
   1171 	int loop;
   1172 	struct l1pt *pt;
   1173 
   1174 #ifdef PMAP_STATIC_L1S
   1175 	for (loop = 0; loop < PMAP_STATIC_L1S; ++loop) {
   1176 #else	/* PMAP_STATIC_L1S */
   1177 	for (loop = 0; loop < max_processes; ++loop) {
   1178 #endif	/* PMAP_STATIC_L1S */
   1179 		/* Allocate a L1 page table */
   1180 		pt = pmap_alloc_l1pt();
   1181 		if (!pt)
   1182 			panic("Cannot allocate static L1 page tables\n");
   1183 
   1184 		/* Clean it */
   1185 		bzero((void *)pt->pt_va, L1_TABLE_SIZE);
   1186 		pt->pt_flags |= (PTFLAG_STATIC | PTFLAG_CLEAN);
   1187 		/* Add the page table to the queue */
   1188 		SIMPLEQ_INSERT_TAIL(&l1pt_static_queue, pt, pt_queue);
   1189 		++l1pt_static_queue_count;
   1190 		++l1pt_static_create_count;
   1191 	}
   1192 }
   1193 
   1194 
   1195 /*
   1196  * Create and return a physical map.
   1197  *
   1198  * If the size specified for the map is zero, the map is an actual physical
   1199  * map, and may be referenced by the hardware.
   1200  *
   1201  * If the size specified is non-zero, the map will be used in software only,
   1202  * and is bounded by that size.
   1203  */
   1204 
   1205 pmap_t
   1206 pmap_create(void)
   1207 {
   1208 	struct pmap *pmap;
   1209 
   1210 	/*
   1211 	 * Fetch pmap entry from the pool
   1212 	 */
   1213 
   1214 	pmap = pool_get(&pmap_pmap_pool, PR_WAITOK);
   1215 	/* XXX is this really needed! */
   1216 	memset(pmap, 0, sizeof(*pmap));
   1217 
   1218 	simple_lock_init(&pmap->pm_obj.vmobjlock);
   1219 	pmap->pm_obj.pgops = NULL;	/* currently not a mappable object */
   1220 	TAILQ_INIT(&pmap->pm_obj.memq);
   1221 	pmap->pm_obj.uo_npages = 0;
   1222 	pmap->pm_obj.uo_refs = 1;
   1223 	pmap->pm_stats.wired_count = 0;
   1224 	pmap->pm_stats.resident_count = 1;
   1225 	pmap->pm_ptphint = NULL;
   1226 
   1227 	/* Now init the machine part of the pmap */
   1228 	pmap_pinit(pmap);
   1229 	return(pmap);
   1230 }
   1231 
   1232 /*
   1233  * pmap_alloc_l1pt()
   1234  *
   1235  * This routine allocates physical and virtual memory for a L1 page table
   1236  * and wires it.
   1237  * A l1pt structure is returned to describe the allocated page table.
   1238  *
   1239  * This routine is allowed to fail if the required memory cannot be allocated.
   1240  * In this case NULL is returned.
   1241  */
   1242 
   1243 struct l1pt *
   1244 pmap_alloc_l1pt(void)
   1245 {
   1246 	paddr_t pa;
   1247 	vaddr_t va;
   1248 	struct l1pt *pt;
   1249 	int error;
   1250 	struct vm_page *m;
   1251 	pt_entry_t *pte;
   1252 
   1253 	/* Allocate virtual address space for the L1 page table */
   1254 	va = uvm_km_valloc(kernel_map, L1_TABLE_SIZE);
   1255 	if (va == 0) {
   1256 #ifdef DIAGNOSTIC
   1257 		PDEBUG(0,
   1258 		    printf("pmap: Cannot allocate pageable memory for L1\n"));
   1259 #endif	/* DIAGNOSTIC */
   1260 		return(NULL);
   1261 	}
   1262 
   1263 	/* Allocate memory for the l1pt structure */
   1264 	pt = (struct l1pt *)malloc(sizeof(struct l1pt), M_VMPMAP, M_WAITOK);
   1265 
   1266 	/*
   1267 	 * Allocate pages from the VM system.
   1268 	 */
   1269 	TAILQ_INIT(&pt->pt_plist);
   1270 	error = uvm_pglistalloc(L1_TABLE_SIZE, physical_start, physical_end,
   1271 	    L1_TABLE_SIZE, 0, &pt->pt_plist, 1, M_WAITOK);
   1272 	if (error) {
   1273 #ifdef DIAGNOSTIC
   1274 		PDEBUG(0,
   1275 		    printf("pmap: Cannot allocate physical mem for L1 (%d)\n",
   1276 		    error));
   1277 #endif	/* DIAGNOSTIC */
   1278 		/* Release the resources we already have claimed */
   1279 		free(pt, M_VMPMAP);
   1280 		uvm_km_free(kernel_map, va, L1_TABLE_SIZE);
   1281 		return(NULL);
   1282 	}
   1283 
   1284 	/* Map our physical pages into our virtual space */
   1285 	pt->pt_va = va;
   1286 	m = TAILQ_FIRST(&pt->pt_plist);
   1287 	while (m && va < (pt->pt_va + L1_TABLE_SIZE)) {
   1288 		pa = VM_PAGE_TO_PHYS(m);
   1289 
   1290 		pte = vtopte(va);
   1291 
   1292 		/*
   1293 		 * Assert that the PTE is invalid.  If it's invalid,
   1294 		 * then we are guaranteed that there won't be an entry
   1295 		 * for this VA in the TLB.
   1296 		 */
   1297 		KDASSERT(pmap_pte_v(pte) == 0);
   1298 
   1299 		*pte = L2_S_PROTO | VM_PAGE_TO_PHYS(m) |
   1300 		    L2_S_PROT(PTE_KERNEL, VM_PROT_READ|VM_PROT_WRITE);
   1301 
   1302 		va += NBPG;
   1303 		m = m->pageq.tqe_next;
   1304 	}
   1305 
   1306 #ifdef DIAGNOSTIC
   1307 	if (m)
   1308 		panic("pmap_alloc_l1pt: pglist not empty\n");
   1309 #endif	/* DIAGNOSTIC */
   1310 
   1311 	pt->pt_flags = 0;
   1312 	return(pt);
   1313 }
   1314 
   1315 /*
   1316  * Free a L1 page table previously allocated with pmap_alloc_l1pt().
   1317  */
   1318 static void
   1319 pmap_free_l1pt(struct l1pt *pt)
   1320 {
   1321 	/* Separate the physical memory for the virtual space */
   1322 	pmap_kremove(pt->pt_va, L1_TABLE_SIZE);
   1323 	pmap_update(pmap_kernel());
   1324 
   1325 	/* Return the physical memory */
   1326 	uvm_pglistfree(&pt->pt_plist);
   1327 
   1328 	/* Free the virtual space */
   1329 	uvm_km_free(kernel_map, pt->pt_va, L1_TABLE_SIZE);
   1330 
   1331 	/* Free the l1pt structure */
   1332 	free(pt, M_VMPMAP);
   1333 }
   1334 
   1335 /*
   1336  * pmap_alloc_ptpt:
   1337  *
   1338  *	Allocate the page table that maps the PTE array.
   1339  */
   1340 static int
   1341 pmap_alloc_ptpt(struct pmap *pmap)
   1342 {
   1343 	struct vm_page *pg;
   1344 	pt_entry_t *pte;
   1345 
   1346 	KASSERT(pmap->pm_vptpt == 0);
   1347 
   1348 	pmap->pm_vptpt = uvm_km_valloc(kernel_map, L2_TABLE_SIZE);
   1349 	if (pmap->pm_vptpt == 0) {
   1350 		PDEBUG(0,
   1351 		    printf("pmap_alloc_ptpt: no KVA for PTPT\n"));
   1352 		return (ENOMEM);
   1353 	}
   1354 
   1355 	for (;;) {
   1356 		pg = uvm_pagealloc(NULL, 0, NULL, UVM_PGA_ZERO);
   1357 		if (pg != NULL)
   1358 			break;
   1359 		uvm_wait("pmap_ptpt");
   1360 	}
   1361 
   1362 	pmap->pm_pptpt = VM_PAGE_TO_PHYS(pg);
   1363 
   1364 	pte = vtopte(pmap->pm_vptpt);
   1365 
   1366 	KDASSERT(pmap_pte_v(pte) == 0);
   1367 
   1368 	*pte = L2_S_PROTO | pmap->pm_pptpt |
   1369 	    L2_S_PROT(PTE_KERNEL, VM_PROT_READ|VM_PROT_WRITE);
   1370 
   1371 	return (0);
   1372 }
   1373 
   1374 /*
   1375  * pmap_free_ptpt:
   1376  *
   1377  *	Free the page table that maps the PTE array.
   1378  */
   1379 static void
   1380 pmap_free_ptpt(struct pmap *pmap)
   1381 {
   1382 
   1383 	pmap_kremove(pmap->pm_vptpt, L2_TABLE_SIZE);
   1384 	pmap_update(pmap_kernel());
   1385 
   1386 	uvm_pagefree(PHYS_TO_VM_PAGE(pmap->pm_pptpt));
   1387 
   1388 	uvm_km_free(kernel_map, pmap->pm_vptpt, L2_TABLE_SIZE);
   1389 }
   1390 
   1391 /*
   1392  * Allocate a page directory.
   1393  * This routine will either allocate a new page directory from the pool
   1394  * of L1 page tables currently held by the kernel or it will allocate
   1395  * a new one via pmap_alloc_l1pt().
   1396  * It will then initialise the l1 page table for use.
   1397  */
   1398 static int
   1399 pmap_allocpagedir(struct pmap *pmap)
   1400 {
   1401 	paddr_t pa;
   1402 	struct l1pt *pt;
   1403 	int error;
   1404 
   1405 	PDEBUG(0, printf("pmap_allocpagedir(%p)\n", pmap));
   1406 
   1407 	/* Do we have any spare L1's lying around ? */
   1408 	if (l1pt_static_queue_count) {
   1409 		--l1pt_static_queue_count;
   1410 		pt = l1pt_static_queue.sqh_first;
   1411 		SIMPLEQ_REMOVE_HEAD(&l1pt_static_queue, pt, pt_queue);
   1412 	} else if (l1pt_queue_count) {
   1413 		--l1pt_queue_count;
   1414 		pt = l1pt_queue.sqh_first;
   1415 		SIMPLEQ_REMOVE_HEAD(&l1pt_queue, pt, pt_queue);
   1416 		++l1pt_reuse_count;
   1417 	} else {
   1418 		pt = pmap_alloc_l1pt();
   1419 		if (!pt)
   1420 			return(ENOMEM);
   1421 		++l1pt_create_count;
   1422 	}
   1423 
   1424 	/* Store the pointer to the l1 descriptor in the pmap. */
   1425 	pmap->pm_l1pt = pt;
   1426 
   1427 	/* Get the physical address of the start of the l1 */
   1428 	pa = VM_PAGE_TO_PHYS(TAILQ_FIRST(&pt->pt_plist));
   1429 
   1430 	/* Store the virtual address of the l1 in the pmap. */
   1431 	pmap->pm_pdir = (pd_entry_t *)pt->pt_va;
   1432 
   1433 	/* Clean the L1 if it is dirty */
   1434 	if (!(pt->pt_flags & PTFLAG_CLEAN))
   1435 		bzero((void *)pmap->pm_pdir, (L1_TABLE_SIZE - KERNEL_PD_SIZE));
   1436 
   1437 	/* Allocate a page table to map all the page tables for this pmap */
   1438 	if ((error = pmap_alloc_ptpt(pmap)) != 0) {
   1439 		pmap_freepagedir(pmap);
   1440 		return (error);
   1441 	}
   1442 
   1443 	/* need to lock this all up for growkernel */
   1444 	simple_lock(&pmaps_lock);
   1445 
   1446 	/* Duplicate the kernel mappings. */
   1447 	bcopy((char *)pmap_kernel()->pm_pdir + (L1_TABLE_SIZE - KERNEL_PD_SIZE),
   1448 		(char *)pmap->pm_pdir + (L1_TABLE_SIZE - KERNEL_PD_SIZE),
   1449 		KERNEL_PD_SIZE);
   1450 
   1451 	/* Wire in this page table */
   1452 	pmap_map_in_l1(pmap, PTE_BASE, pmap->pm_pptpt, TRUE);
   1453 
   1454 	pt->pt_flags &= ~PTFLAG_CLEAN;	/* L1 is dirty now */
   1455 
   1456 	/*
   1457 	 * Map the kernel page tables into the new PT map.
   1458 	 */
   1459 	bcopy((char *)(PTE_BASE
   1460 	    + (PTE_BASE >> (PGSHIFT - 2))
   1461 	    + ((L1_TABLE_SIZE - KERNEL_PD_SIZE) >> 2)),
   1462 	    (char *)pmap->pm_vptpt + ((L1_TABLE_SIZE - KERNEL_PD_SIZE) >> 2),
   1463 	    (KERNEL_PD_SIZE >> 2));
   1464 
   1465 	LIST_INSERT_HEAD(&pmaps, pmap, pm_list);
   1466 	simple_unlock(&pmaps_lock);
   1467 
   1468 	return(0);
   1469 }
   1470 
   1471 
   1472 /*
   1473  * Initialize a preallocated and zeroed pmap structure,
   1474  * such as one in a vmspace structure.
   1475  */
   1476 
   1477 void
   1478 pmap_pinit(struct pmap *pmap)
   1479 {
   1480 	int backoff = 6;
   1481 	int retry = 10;
   1482 
   1483 	PDEBUG(0, printf("pmap_pinit(%p)\n", pmap));
   1484 
   1485 	/* Keep looping until we succeed in allocating a page directory */
   1486 	while (pmap_allocpagedir(pmap) != 0) {
   1487 		/*
   1488 		 * Ok we failed to allocate a suitable block of memory for an
   1489 		 * L1 page table. This means that either:
   1490 		 * 1. 16KB of virtual address space could not be allocated
   1491 		 * 2. 16KB of physically contiguous memory on a 16KB boundary
   1492 		 *    could not be allocated.
   1493 		 *
   1494 		 * Since we cannot fail we will sleep for a while and try
   1495 		 * again.
   1496 		 *
   1497 		 * Searching for a suitable L1 PT is expensive:
   1498 		 * to avoid hogging the system when memory is really
   1499 		 * scarce, use an exponential back-off so that
   1500 		 * eventually we won't retry more than once every 8
   1501 		 * seconds.  This should allow other processes to run
   1502 		 * to completion and free up resources.
   1503 		 */
   1504 		(void) ltsleep(&lbolt, PVM, "l1ptwait", (hz << 3) >> backoff,
   1505 		    NULL);
   1506 		if (--retry == 0) {
   1507 			retry = 10;
   1508 			if (backoff)
   1509 				--backoff;
   1510 		}
   1511 	}
   1512 
   1513 	if (vector_page < KERNEL_BASE) {
   1514 		/*
   1515 		 * Map the vector page.  This will also allocate and map
   1516 		 * an L2 table for it.
   1517 		 */
   1518 		pmap_enter(pmap, vector_page, systempage.pv_pa,
   1519 		    VM_PROT_READ, VM_PROT_READ | PMAP_WIRED);
   1520 		pmap_update(pmap);
   1521 	}
   1522 }
   1523 
   1524 
   1525 void
   1526 pmap_freepagedir(struct pmap *pmap)
   1527 {
   1528 	/* Free the memory used for the page table mapping */
   1529 	if (pmap->pm_vptpt != 0)
   1530 		pmap_free_ptpt(pmap);
   1531 
   1532 	/* junk the L1 page table */
   1533 	if (pmap->pm_l1pt->pt_flags & PTFLAG_STATIC) {
   1534 		/* Add the page table to the queue */
   1535 		SIMPLEQ_INSERT_TAIL(&l1pt_static_queue, pmap->pm_l1pt, pt_queue);
   1536 		++l1pt_static_queue_count;
   1537 	} else if (l1pt_queue_count < 8) {
   1538 		/* Add the page table to the queue */
   1539 		SIMPLEQ_INSERT_TAIL(&l1pt_queue, pmap->pm_l1pt, pt_queue);
   1540 		++l1pt_queue_count;
   1541 	} else
   1542 		pmap_free_l1pt(pmap->pm_l1pt);
   1543 }
   1544 
   1545 
   1546 /*
   1547  * Retire the given physical map from service.
   1548  * Should only be called if the map contains no valid mappings.
   1549  */
   1550 
   1551 void
   1552 pmap_destroy(struct pmap *pmap)
   1553 {
   1554 	struct vm_page *page;
   1555 	int count;
   1556 
   1557 	if (pmap == NULL)
   1558 		return;
   1559 
   1560 	PDEBUG(0, printf("pmap_destroy(%p)\n", pmap));
   1561 
   1562 	/*
   1563 	 * Drop reference count
   1564 	 */
   1565 	simple_lock(&pmap->pm_obj.vmobjlock);
   1566 	count = --pmap->pm_obj.uo_refs;
   1567 	simple_unlock(&pmap->pm_obj.vmobjlock);
   1568 	if (count > 0) {
   1569 		return;
   1570 	}
   1571 
   1572 	/*
   1573 	 * reference count is zero, free pmap resources and then free pmap.
   1574 	 */
   1575 
   1576 	/*
   1577 	 * remove it from global list of pmaps
   1578 	 */
   1579 
   1580 	simple_lock(&pmaps_lock);
   1581 	LIST_REMOVE(pmap, pm_list);
   1582 	simple_unlock(&pmaps_lock);
   1583 
   1584 	if (vector_page < KERNEL_BASE) {
   1585 		/* Remove the vector page mapping */
   1586 		pmap_remove(pmap, vector_page, vector_page + NBPG);
   1587 		pmap_update(pmap);
   1588 	}
   1589 
   1590 	/*
   1591 	 * Free any page tables still mapped
   1592 	 * This is only temporay until pmap_enter can count the number
   1593 	 * of mappings made in a page table. Then pmap_remove() can
   1594 	 * reduce the count and free the pagetable when the count
   1595 	 * reaches zero.  Note that entries in this list should match the
   1596 	 * contents of the ptpt, however this is faster than walking a 1024
   1597 	 * entries looking for pt's
   1598 	 * taken from i386 pmap.c
   1599 	 */
   1600 	while ((page = TAILQ_FIRST(&pmap->pm_obj.memq)) != NULL) {
   1601 		KASSERT((page->flags & PG_BUSY) == 0);
   1602 		page->wire_count = 0;
   1603 		uvm_pagefree(page);
   1604 	}
   1605 
   1606 	/* Free the page dir */
   1607 	pmap_freepagedir(pmap);
   1608 
   1609 	/* return the pmap to the pool */
   1610 	pool_put(&pmap_pmap_pool, pmap);
   1611 }
   1612 
   1613 
   1614 /*
   1615  * void pmap_reference(struct pmap *pmap)
   1616  *
   1617  * Add a reference to the specified pmap.
   1618  */
   1619 
   1620 void
   1621 pmap_reference(struct pmap *pmap)
   1622 {
   1623 	if (pmap == NULL)
   1624 		return;
   1625 
   1626 	simple_lock(&pmap->pm_lock);
   1627 	pmap->pm_obj.uo_refs++;
   1628 	simple_unlock(&pmap->pm_lock);
   1629 }
   1630 
   1631 /*
   1632  * void pmap_virtual_space(vaddr_t *start, vaddr_t *end)
   1633  *
   1634  * Return the start and end addresses of the kernel's virtual space.
   1635  * These values are setup in pmap_bootstrap and are updated as pages
   1636  * are allocated.
   1637  */
   1638 
   1639 void
   1640 pmap_virtual_space(vaddr_t *start, vaddr_t *end)
   1641 {
   1642 	*start = virtual_avail;
   1643 	*end = virtual_end;
   1644 }
   1645 
   1646 /*
   1647  * Activate the address space for the specified process.  If the process
   1648  * is the current process, load the new MMU context.
   1649  */
   1650 void
   1651 pmap_activate(struct proc *p)
   1652 {
   1653 	struct pmap *pmap = p->p_vmspace->vm_map.pmap;
   1654 	struct pcb *pcb = &p->p_addr->u_pcb;
   1655 
   1656 	(void) pmap_extract(pmap_kernel(), (vaddr_t)pmap->pm_pdir,
   1657 	    (paddr_t *)&pcb->pcb_pagedir);
   1658 
   1659 	PDEBUG(0, printf("pmap_activate: p=%p pmap=%p pcb=%p pdir=%p l1=%p\n",
   1660 	    p, pmap, pcb, pmap->pm_pdir, pcb->pcb_pagedir));
   1661 
   1662 	if (p == curproc) {
   1663 		PDEBUG(0, printf("pmap_activate: setting TTB\n"));
   1664 		setttb((u_int)pcb->pcb_pagedir);
   1665 	}
   1666 }
   1667 
   1668 /*
   1669  * Deactivate the address space of the specified process.
   1670  */
   1671 void
   1672 pmap_deactivate(struct proc *p)
   1673 {
   1674 }
   1675 
   1676 /*
   1677  * Perform any deferred pmap operations.
   1678  */
   1679 void
   1680 pmap_update(struct pmap *pmap)
   1681 {
   1682 
   1683 	/*
   1684 	 * We haven't deferred any pmap operations, but we do need to
   1685 	 * make sure TLB/cache operations have completed.
   1686 	 */
   1687 	cpu_cpwait();
   1688 }
   1689 
   1690 /*
   1691  * pmap_clean_page()
   1692  *
   1693  * This is a local function used to work out the best strategy to clean
   1694  * a single page referenced by its entry in the PV table. It's used by
   1695  * pmap_copy_page, pmap_zero page and maybe some others later on.
   1696  *
   1697  * Its policy is effectively:
   1698  *  o If there are no mappings, we don't bother doing anything with the cache.
   1699  *  o If there is one mapping, we clean just that page.
   1700  *  o If there are multiple mappings, we clean the entire cache.
   1701  *
   1702  * So that some functions can be further optimised, it returns 0 if it didn't
   1703  * clean the entire cache, or 1 if it did.
   1704  *
   1705  * XXX One bug in this routine is that if the pv_entry has a single page
   1706  * mapped at 0x00000000 a whole cache clean will be performed rather than
   1707  * just the 1 page. Since this should not occur in everyday use and if it does
   1708  * it will just result in not the most efficient clean for the page.
   1709  */
   1710 static int
   1711 pmap_clean_page(struct pv_entry *pv, boolean_t is_src)
   1712 {
   1713 	struct pmap *pmap;
   1714 	struct pv_entry *npv;
   1715 	int cache_needs_cleaning = 0;
   1716 	vaddr_t page_to_clean = 0;
   1717 
   1718 	if (pv == NULL)
   1719 		/* nothing mapped in so nothing to flush */
   1720 		return (0);
   1721 
   1722 	/* Since we flush the cache each time we change curproc, we
   1723 	 * only need to flush the page if it is in the current pmap.
   1724 	 */
   1725 	if (curproc)
   1726 		pmap = curproc->p_vmspace->vm_map.pmap;
   1727 	else
   1728 		pmap = pmap_kernel();
   1729 
   1730 	for (npv = pv; npv; npv = npv->pv_next) {
   1731 		if (npv->pv_pmap == pmap) {
   1732 			/* The page is mapped non-cacheable in
   1733 			 * this map.  No need to flush the cache.
   1734 			 */
   1735 			if (npv->pv_flags & PVF_NC) {
   1736 #ifdef DIAGNOSTIC
   1737 				if (cache_needs_cleaning)
   1738 					panic("pmap_clean_page: "
   1739 							"cache inconsistency");
   1740 #endif
   1741 				break;
   1742 			}
   1743 #if 0
   1744 			/*
   1745 			 * XXX Can't do this because pmap_protect doesn't
   1746 			 * XXX clean the page when it does a write-protect.
   1747 			 */
   1748 			else if (is_src && (npv->pv_flags & PVF_WRITE) == 0)
   1749 				continue;
   1750 #endif
   1751 			if (cache_needs_cleaning){
   1752 				page_to_clean = 0;
   1753 				break;
   1754 			}
   1755 			else
   1756 				page_to_clean = npv->pv_va;
   1757 			cache_needs_cleaning = 1;
   1758 		}
   1759 	}
   1760 
   1761 	if (page_to_clean)
   1762 		cpu_idcache_wbinv_range(page_to_clean, NBPG);
   1763 	else if (cache_needs_cleaning) {
   1764 		cpu_idcache_wbinv_all();
   1765 		return (1);
   1766 	}
   1767 	return (0);
   1768 }
   1769 
   1770 /*
   1771  * pmap_zero_page()
   1772  *
   1773  * Zero a given physical page by mapping it at a page hook point.
   1774  * In doing the zero page op, the page we zero is mapped cachable, as with
   1775  * StrongARM accesses to non-cached pages are non-burst making writing
   1776  * _any_ bulk data very slow.
   1777  */
   1778 #if ARM_MMU_GENERIC == 1
   1779 void
   1780 pmap_zero_page_generic(paddr_t phys)
   1781 {
   1782 #ifdef DEBUG
   1783 	struct vm_page *pg = PHYS_TO_VM_PAGE(phys);
   1784 
   1785 	if (pg->mdpage.pvh_list != NULL)
   1786 		panic("pmap_zero_page: page has mappings");
   1787 #endif
   1788 
   1789 	KDASSERT((phys & PGOFSET) == 0);
   1790 
   1791 	/*
   1792 	 * Hook in the page, zero it, and purge the cache for that
   1793 	 * zeroed page. Invalidate the TLB as needed.
   1794 	 */
   1795 	*cdst_pte = L2_S_PROTO | phys |
   1796 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
   1797 	cpu_tlb_flushD_SE(cdstp);
   1798 	cpu_cpwait();
   1799 	bzero_page(cdstp);
   1800 	cpu_dcache_wbinv_range(cdstp, NBPG);
   1801 }
   1802 #endif /* ARM_MMU_GENERIC == 1 */
   1803 
   1804 #if ARM_MMU_XSCALE == 1
   1805 void
   1806 pmap_zero_page_xscale(paddr_t phys)
   1807 {
   1808 #ifdef DEBUG
   1809 	struct vm_page *pg = PHYS_TO_VM_PAGE(phys);
   1810 
   1811 	if (pg->mdpage.pvh_list != NULL)
   1812 		panic("pmap_zero_page: page has mappings");
   1813 #endif
   1814 
   1815 	KDASSERT((phys & PGOFSET) == 0);
   1816 
   1817 	/*
   1818 	 * Hook in the page, zero it, and purge the cache for that
   1819 	 * zeroed page. Invalidate the TLB as needed.
   1820 	 */
   1821 	*cdst_pte = L2_S_PROTO | phys |
   1822 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
   1823 	    L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X);	/* mini-data */
   1824 	cpu_tlb_flushD_SE(cdstp);
   1825 	cpu_cpwait();
   1826 	bzero_page(cdstp);
   1827 	xscale_cache_clean_minidata();
   1828 }
   1829 #endif /* ARM_MMU_XSCALE == 1 */
   1830 
   1831 /* pmap_pageidlezero()
   1832  *
   1833  * The same as above, except that we assume that the page is not
   1834  * mapped.  This means we never have to flush the cache first.  Called
   1835  * from the idle loop.
   1836  */
   1837 boolean_t
   1838 pmap_pageidlezero(paddr_t phys)
   1839 {
   1840 	int i, *ptr;
   1841 	boolean_t rv = TRUE;
   1842 #ifdef DEBUG
   1843 	struct vm_page *pg;
   1844 
   1845 	pg = PHYS_TO_VM_PAGE(phys);
   1846 	if (pg->mdpage.pvh_list != NULL)
   1847 		panic("pmap_pageidlezero: page has mappings");
   1848 #endif
   1849 
   1850 	KDASSERT((phys & PGOFSET) == 0);
   1851 
   1852 	/*
   1853 	 * Hook in the page, zero it, and purge the cache for that
   1854 	 * zeroed page. Invalidate the TLB as needed.
   1855 	 */
   1856 	*cdst_pte = L2_S_PROTO | phys |
   1857 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
   1858 	cpu_tlb_flushD_SE(cdstp);
   1859 	cpu_cpwait();
   1860 
   1861 	for (i = 0, ptr = (int *)cdstp;
   1862 			i < (NBPG / sizeof(int)); i++) {
   1863 		if (sched_whichqs != 0) {
   1864 			/*
   1865 			 * A process has become ready.  Abort now,
   1866 			 * so we don't keep it waiting while we
   1867 			 * do slow memory access to finish this
   1868 			 * page.
   1869 			 */
   1870 			rv = FALSE;
   1871 			break;
   1872 		}
   1873 		*ptr++ = 0;
   1874 	}
   1875 
   1876 	if (rv)
   1877 		/*
   1878 		 * if we aborted we'll rezero this page again later so don't
   1879 		 * purge it unless we finished it
   1880 		 */
   1881 		cpu_dcache_wbinv_range(cdstp, NBPG);
   1882 	return (rv);
   1883 }
   1884 
   1885 /*
   1886  * pmap_copy_page()
   1887  *
   1888  * Copy one physical page into another, by mapping the pages into
   1889  * hook points. The same comment regarding cachability as in
   1890  * pmap_zero_page also applies here.
   1891  */
   1892 #if ARM_MMU_GENERIC == 1
   1893 void
   1894 pmap_copy_page_generic(paddr_t src, paddr_t dst)
   1895 {
   1896 	struct vm_page *src_pg = PHYS_TO_VM_PAGE(src);
   1897 #ifdef DEBUG
   1898 	struct vm_page *dst_pg = PHYS_TO_VM_PAGE(dst);
   1899 
   1900 	if (dst_pg->mdpage.pvh_list != NULL)
   1901 		panic("pmap_copy_page: dst page has mappings");
   1902 #endif
   1903 
   1904 	KDASSERT((src & PGOFSET) == 0);
   1905 	KDASSERT((dst & PGOFSET) == 0);
   1906 
   1907 	/*
   1908 	 * Clean the source page.  Hold the source page's lock for
   1909 	 * the duration of the copy so that no other mappings can
   1910 	 * be created while we have a potentially aliased mapping.
   1911 	 */
   1912 	simple_lock(&src_pg->mdpage.pvh_slock);
   1913 	(void) pmap_clean_page(src_pg->mdpage.pvh_list, TRUE);
   1914 
   1915 	/*
   1916 	 * Map the pages into the page hook points, copy them, and purge
   1917 	 * the cache for the appropriate page. Invalidate the TLB
   1918 	 * as required.
   1919 	 */
   1920 	*csrc_pte = L2_S_PROTO | src |
   1921 	    L2_S_PROT(PTE_KERNEL, VM_PROT_READ) | pte_l2_s_cache_mode;
   1922 	*cdst_pte = L2_S_PROTO | dst |
   1923 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
   1924 	cpu_tlb_flushD_SE(csrcp);
   1925 	cpu_tlb_flushD_SE(cdstp);
   1926 	cpu_cpwait();
   1927 	bcopy_page(csrcp, cdstp);
   1928 	cpu_dcache_inv_range(csrcp, NBPG);
   1929 	simple_unlock(&src_pg->mdpage.pvh_slock); /* cache is safe again */
   1930 	cpu_dcache_wbinv_range(cdstp, NBPG);
   1931 }
   1932 #endif /* ARM_MMU_GENERIC == 1 */
   1933 
   1934 #if ARM_MMU_XSCALE == 1
   1935 void
   1936 pmap_copy_page_xscale(paddr_t src, paddr_t dst)
   1937 {
   1938 	struct vm_page *src_pg = PHYS_TO_VM_PAGE(src);
   1939 #ifdef DEBUG
   1940 	struct vm_page *dst_pg = PHYS_TO_VM_PAGE(dst);
   1941 
   1942 	if (dst_pg->mdpage.pvh_list != NULL)
   1943 		panic("pmap_copy_page: dst page has mappings");
   1944 #endif
   1945 
   1946 	KDASSERT((src & PGOFSET) == 0);
   1947 	KDASSERT((dst & PGOFSET) == 0);
   1948 
   1949 	/*
   1950 	 * Clean the source page.  Hold the source page's lock for
   1951 	 * the duration of the copy so that no other mappings can
   1952 	 * be created while we have a potentially aliased mapping.
   1953 	 */
   1954 	simple_lock(&src_pg->mdpage.pvh_slock);
   1955 	(void) pmap_clean_page(src_pg->mdpage.pvh_list, TRUE);
   1956 
   1957 	/*
   1958 	 * Map the pages into the page hook points, copy them, and purge
   1959 	 * the cache for the appropriate page. Invalidate the TLB
   1960 	 * as required.
   1961 	 */
   1962 	*csrc_pte = L2_S_PROTO | src |
   1963 	    L2_S_PROT(PTE_KERNEL, VM_PROT_READ) |
   1964 	    L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X);	/* mini-data */
   1965 	*cdst_pte = L2_S_PROTO | dst |
   1966 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
   1967 	    L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X);	/* mini-data */
   1968 	cpu_tlb_flushD_SE(csrcp);
   1969 	cpu_tlb_flushD_SE(cdstp);
   1970 	cpu_cpwait();
   1971 	bcopy_page(csrcp, cdstp);
   1972 	simple_unlock(&src_pg->mdpage.pvh_slock); /* cache is safe again */
   1973 	xscale_cache_clean_minidata();
   1974 }
   1975 #endif /* ARM_MMU_XSCALE == 1 */
   1976 
   1977 #if 0
   1978 void
   1979 pmap_pte_addref(struct pmap *pmap, vaddr_t va)
   1980 {
   1981 	pd_entry_t *pde;
   1982 	paddr_t pa;
   1983 	struct vm_page *m;
   1984 
   1985 	if (pmap == pmap_kernel())
   1986 		return;
   1987 
   1988 	pde = pmap_pde(pmap, va & ~(3 << L1_S_SHIFT));
   1989 	pa = pmap_pte_pa(pde);
   1990 	m = PHYS_TO_VM_PAGE(pa);
   1991 	++m->wire_count;
   1992 #ifdef MYCROFT_HACK
   1993 	printf("addref pmap=%p va=%08lx pde=%p pa=%08lx m=%p wire=%d\n",
   1994 	    pmap, va, pde, pa, m, m->wire_count);
   1995 #endif
   1996 }
   1997 
   1998 void
   1999 pmap_pte_delref(struct pmap *pmap, vaddr_t va)
   2000 {
   2001 	pd_entry_t *pde;
   2002 	paddr_t pa;
   2003 	struct vm_page *m;
   2004 
   2005 	if (pmap == pmap_kernel())
   2006 		return;
   2007 
   2008 	pde = pmap_pde(pmap, va & ~(3 << L1_S_SHIFT));
   2009 	pa = pmap_pte_pa(pde);
   2010 	m = PHYS_TO_VM_PAGE(pa);
   2011 	--m->wire_count;
   2012 #ifdef MYCROFT_HACK
   2013 	printf("delref pmap=%p va=%08lx pde=%p pa=%08lx m=%p wire=%d\n",
   2014 	    pmap, va, pde, pa, m, m->wire_count);
   2015 #endif
   2016 	if (m->wire_count == 0) {
   2017 #ifdef MYCROFT_HACK
   2018 		printf("delref pmap=%p va=%08lx pde=%p pa=%08lx m=%p\n",
   2019 		    pmap, va, pde, pa, m);
   2020 #endif
   2021 		pmap_unmap_in_l1(pmap, va);
   2022 		uvm_pagefree(m);
   2023 		--pmap->pm_stats.resident_count;
   2024 	}
   2025 }
   2026 #else
   2027 #define	pmap_pte_addref(pmap, va)
   2028 #define	pmap_pte_delref(pmap, va)
   2029 #endif
   2030 
   2031 /*
   2032  * Since we have a virtually indexed cache, we may need to inhibit caching if
   2033  * there is more than one mapping and at least one of them is writable.
   2034  * Since we purge the cache on every context switch, we only need to check for
   2035  * other mappings within the same pmap, or kernel_pmap.
   2036  * This function is also called when a page is unmapped, to possibly reenable
   2037  * caching on any remaining mappings.
   2038  *
   2039  * The code implements the following logic, where:
   2040  *
   2041  * KW = # of kernel read/write pages
   2042  * KR = # of kernel read only pages
   2043  * UW = # of user read/write pages
   2044  * UR = # of user read only pages
   2045  * OW = # of user read/write pages in another pmap, then
   2046  *
   2047  * KC = kernel mapping is cacheable
   2048  * UC = user mapping is cacheable
   2049  *
   2050  *                     KW=0,KR=0  KW=0,KR>0  KW=1,KR=0  KW>1,KR>=0
   2051  *                   +---------------------------------------------
   2052  * UW=0,UR=0,OW=0    | ---        KC=1       KC=1       KC=0
   2053  * UW=0,UR>0,OW=0    | UC=1       KC=1,UC=1  KC=0,UC=0  KC=0,UC=0
   2054  * UW=0,UR>0,OW>0    | UC=1       KC=0,UC=1  KC=0,UC=0  KC=0,UC=0
   2055  * UW=1,UR=0,OW=0    | UC=1       KC=0,UC=0  KC=0,UC=0  KC=0,UC=0
   2056  * UW>1,UR>=0,OW>=0  | UC=0       KC=0,UC=0  KC=0,UC=0  KC=0,UC=0
   2057  *
   2058  * Note that the pmap must have it's ptes mapped in, and passed with ptes.
   2059  */
   2060 __inline static void
   2061 pmap_vac_me_harder(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes,
   2062 	boolean_t clear_cache)
   2063 {
   2064 	if (pmap == pmap_kernel())
   2065 		pmap_vac_me_kpmap(pmap, pg, ptes, clear_cache);
   2066 	else
   2067 		pmap_vac_me_user(pmap, pg, ptes, clear_cache);
   2068 }
   2069 
   2070 static void
   2071 pmap_vac_me_kpmap(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes,
   2072 	boolean_t clear_cache)
   2073 {
   2074 	int user_entries = 0;
   2075 	int user_writable = 0;
   2076 	int user_cacheable = 0;
   2077 	int kernel_entries = 0;
   2078 	int kernel_writable = 0;
   2079 	int kernel_cacheable = 0;
   2080 	struct pv_entry *pv;
   2081 	struct pmap *last_pmap = pmap;
   2082 
   2083 #ifdef DIAGNOSTIC
   2084 	if (pmap != pmap_kernel())
   2085 		panic("pmap_vac_me_kpmap: pmap != pmap_kernel()");
   2086 #endif
   2087 
   2088 	/*
   2089 	 * Pass one, see if there are both kernel and user pmaps for
   2090 	 * this page.  Calculate whether there are user-writable or
   2091 	 * kernel-writable pages.
   2092 	 */
   2093 	for (pv = pg->mdpage.pvh_list; pv != NULL; pv = pv->pv_next) {
   2094 		if (pv->pv_pmap != pmap) {
   2095 			user_entries++;
   2096 			if (pv->pv_flags & PVF_WRITE)
   2097 				user_writable++;
   2098 			if ((pv->pv_flags & PVF_NC) == 0)
   2099 				user_cacheable++;
   2100 		} else {
   2101 			kernel_entries++;
   2102 			if (pv->pv_flags & PVF_WRITE)
   2103 				kernel_writable++;
   2104 			if ((pv->pv_flags & PVF_NC) == 0)
   2105 				kernel_cacheable++;
   2106 		}
   2107 	}
   2108 
   2109 	/*
   2110 	 * We know we have just been updating a kernel entry, so if
   2111 	 * all user pages are already cacheable, then there is nothing
   2112 	 * further to do.
   2113 	 */
   2114 	if (kernel_entries == 0 &&
   2115 	    user_cacheable == user_entries)
   2116 		return;
   2117 
   2118 	if (user_entries) {
   2119 		/*
   2120 		 * Scan over the list again, for each entry, if it
   2121 		 * might not be set correctly, call pmap_vac_me_user
   2122 		 * to recalculate the settings.
   2123 		 */
   2124 		for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
   2125 			/*
   2126 			 * We know kernel mappings will get set
   2127 			 * correctly in other calls.  We also know
   2128 			 * that if the pmap is the same as last_pmap
   2129 			 * then we've just handled this entry.
   2130 			 */
   2131 			if (pv->pv_pmap == pmap || pv->pv_pmap == last_pmap)
   2132 				continue;
   2133 			/*
   2134 			 * If there are kernel entries and this page
   2135 			 * is writable but non-cacheable, then we can
   2136 			 * skip this entry also.
   2137 			 */
   2138 			if (kernel_entries > 0 &&
   2139 			    (pv->pv_flags & (PVF_NC | PVF_WRITE)) ==
   2140 			    (PVF_NC | PVF_WRITE))
   2141 				continue;
   2142 			/*
   2143 			 * Similarly if there are no kernel-writable
   2144 			 * entries and the page is already
   2145 			 * read-only/cacheable.
   2146 			 */
   2147 			if (kernel_writable == 0 &&
   2148 			    (pv->pv_flags & (PVF_NC | PVF_WRITE)) == 0)
   2149 				continue;
   2150 			/*
   2151 			 * For some of the remaining cases, we know
   2152 			 * that we must recalculate, but for others we
   2153 			 * can't tell if they are correct or not, so
   2154 			 * we recalculate anyway.
   2155 			 */
   2156 			pmap_unmap_ptes(last_pmap);
   2157 			last_pmap = pv->pv_pmap;
   2158 			ptes = pmap_map_ptes(last_pmap);
   2159 			pmap_vac_me_user(last_pmap, pg, ptes,
   2160 			    pmap_is_curpmap(last_pmap));
   2161 		}
   2162 		/* Restore the pte mapping that was passed to us.  */
   2163 		if (last_pmap != pmap) {
   2164 			pmap_unmap_ptes(last_pmap);
   2165 			ptes = pmap_map_ptes(pmap);
   2166 		}
   2167 		if (kernel_entries == 0)
   2168 			return;
   2169 	}
   2170 
   2171 	pmap_vac_me_user(pmap, pg, ptes, clear_cache);
   2172 	return;
   2173 }
   2174 
   2175 static void
   2176 pmap_vac_me_user(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes,
   2177 	boolean_t clear_cache)
   2178 {
   2179 	struct pmap *kpmap = pmap_kernel();
   2180 	struct pv_entry *pv, *npv;
   2181 	int entries = 0;
   2182 	int writable = 0;
   2183 	int cacheable_entries = 0;
   2184 	int kern_cacheable = 0;
   2185 	int other_writable = 0;
   2186 
   2187 	pv = pg->mdpage.pvh_list;
   2188 	KASSERT(ptes != NULL);
   2189 
   2190 	/*
   2191 	 * Count mappings and writable mappings in this pmap.
   2192 	 * Include kernel mappings as part of our own.
   2193 	 * Keep a pointer to the first one.
   2194 	 */
   2195 	for (npv = pv; npv; npv = npv->pv_next) {
   2196 		/* Count mappings in the same pmap */
   2197 		if (pmap == npv->pv_pmap ||
   2198 		    kpmap == npv->pv_pmap) {
   2199 			if (entries++ == 0)
   2200 				pv = npv;
   2201 			/* Cacheable mappings */
   2202 			if ((npv->pv_flags & PVF_NC) == 0) {
   2203 				cacheable_entries++;
   2204 				if (kpmap == npv->pv_pmap)
   2205 					kern_cacheable++;
   2206 			}
   2207 			/* Writable mappings */
   2208 			if (npv->pv_flags & PVF_WRITE)
   2209 				++writable;
   2210 		} else if (npv->pv_flags & PVF_WRITE)
   2211 			other_writable = 1;
   2212 	}
   2213 
   2214 	PDEBUG(3,printf("pmap_vac_me_harder: pmap %p Entries %d, "
   2215 		"writable %d cacheable %d %s\n", pmap, entries, writable,
   2216 	    	cacheable_entries, clear_cache ? "clean" : "no clean"));
   2217 
   2218 	/*
   2219 	 * Enable or disable caching as necessary.
   2220 	 * Note: the first entry might be part of the kernel pmap,
   2221 	 * so we can't assume this is indicative of the state of the
   2222 	 * other (maybe non-kpmap) entries.
   2223 	 */
   2224 	if ((entries > 1 && writable) ||
   2225 	    (entries > 0 && pmap == kpmap && other_writable)) {
   2226 		if (cacheable_entries == 0)
   2227 		    return;
   2228 		for (npv = pv; npv; npv = npv->pv_next) {
   2229 			if ((pmap == npv->pv_pmap
   2230 			    || kpmap == npv->pv_pmap) &&
   2231 			    (npv->pv_flags & PVF_NC) == 0) {
   2232 				ptes[arm_btop(npv->pv_va)] &= ~L2_S_CACHE_MASK;
   2233  				npv->pv_flags |= PVF_NC;
   2234 				/*
   2235 				 * If this page needs flushing from the
   2236 				 * cache, and we aren't going to do it
   2237 				 * below, do it now.
   2238 				 */
   2239 				if ((cacheable_entries < 4 &&
   2240 				    (clear_cache || npv->pv_pmap == kpmap)) ||
   2241 				    (npv->pv_pmap == kpmap &&
   2242 				    !clear_cache && kern_cacheable < 4)) {
   2243 					cpu_idcache_wbinv_range(npv->pv_va,
   2244 					    NBPG);
   2245 					cpu_tlb_flushID_SE(npv->pv_va);
   2246 				}
   2247 			}
   2248 		}
   2249 		if ((clear_cache && cacheable_entries >= 4) ||
   2250 		    kern_cacheable >= 4) {
   2251 			cpu_idcache_wbinv_all();
   2252 			cpu_tlb_flushID();
   2253 		}
   2254 		cpu_cpwait();
   2255 	} else if (entries > 0) {
   2256 		/*
   2257 		 * Turn cacheing back on for some pages.  If it is a kernel
   2258 		 * page, only do so if there are no other writable pages.
   2259 		 */
   2260 		for (npv = pv; npv; npv = npv->pv_next) {
   2261 			if ((pmap == npv->pv_pmap ||
   2262 			    (kpmap == npv->pv_pmap && other_writable == 0)) &&
   2263 			    (npv->pv_flags & PVF_NC)) {
   2264 				ptes[arm_btop(npv->pv_va)] |=
   2265 				    pte_l2_s_cache_mode;
   2266 				npv->pv_flags &= ~PVF_NC;
   2267 			}
   2268 		}
   2269 	}
   2270 }
   2271 
   2272 /*
   2273  * pmap_remove()
   2274  *
   2275  * pmap_remove is responsible for nuking a number of mappings for a range
   2276  * of virtual address space in the current pmap. To do this efficiently
   2277  * is interesting, because in a number of cases a wide virtual address
   2278  * range may be supplied that contains few actual mappings. So, the
   2279  * optimisations are:
   2280  *  1. Try and skip over hunks of address space for which an L1 entry
   2281  *     does not exist.
   2282  *  2. Build up a list of pages we've hit, up to a maximum, so we can
   2283  *     maybe do just a partial cache clean. This path of execution is
   2284  *     complicated by the fact that the cache must be flushed _before_
   2285  *     the PTE is nuked, being a VAC :-)
   2286  *  3. Maybe later fast-case a single page, but I don't think this is
   2287  *     going to make _that_ much difference overall.
   2288  */
   2289 
   2290 #define PMAP_REMOVE_CLEAN_LIST_SIZE	3
   2291 
   2292 void
   2293 pmap_remove(struct pmap *pmap, vaddr_t sva, vaddr_t eva)
   2294 {
   2295 	int cleanlist_idx = 0;
   2296 	struct pagelist {
   2297 		vaddr_t va;
   2298 		pt_entry_t *pte;
   2299 	} cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE];
   2300 	pt_entry_t *pte = 0, *ptes;
   2301 	paddr_t pa;
   2302 	int pmap_active;
   2303 	struct vm_page *pg;
   2304 
   2305 	/* Exit quick if there is no pmap */
   2306 	if (!pmap)
   2307 		return;
   2308 
   2309 	PDEBUG(0, printf("pmap_remove: pmap=%p sva=%08lx eva=%08lx\n",
   2310 	    pmap, sva, eva));
   2311 
   2312 	/*
   2313 	 * we lock in the pmap => vm_page direction
   2314 	 */
   2315 	PMAP_MAP_TO_HEAD_LOCK();
   2316 
   2317 	ptes = pmap_map_ptes(pmap);
   2318 	/* Get a page table pointer */
   2319 	while (sva < eva) {
   2320 		if (pmap_pde_page(pmap_pde(pmap, sva)))
   2321 			break;
   2322 		sva = (sva & L1_S_FRAME) + L1_S_SIZE;
   2323 	}
   2324 
   2325 	pte = &ptes[arm_btop(sva)];
   2326 	/* Note if the pmap is active thus require cache and tlb cleans */
   2327 	pmap_active = pmap_is_curpmap(pmap);
   2328 
   2329 	/* Now loop along */
   2330 	while (sva < eva) {
   2331 		/* Check if we can move to the next PDE (l1 chunk) */
   2332 		if (!(sva & L2_ADDR_BITS))
   2333 			if (!pmap_pde_page(pmap_pde(pmap, sva))) {
   2334 				sva += L1_S_SIZE;
   2335 				pte += arm_btop(L1_S_SIZE);
   2336 				continue;
   2337 			}
   2338 
   2339 		/* We've found a valid PTE, so this page of PTEs has to go. */
   2340 		if (pmap_pte_v(pte)) {
   2341 			/* Update statistics */
   2342 			--pmap->pm_stats.resident_count;
   2343 
   2344 			/*
   2345 			 * Add this page to our cache remove list, if we can.
   2346 			 * If, however the cache remove list is totally full,
   2347 			 * then do a complete cache invalidation taking note
   2348 			 * to backtrack the PTE table beforehand, and ignore
   2349 			 * the lists in future because there's no longer any
   2350 			 * point in bothering with them (we've paid the
   2351 			 * penalty, so will carry on unhindered). Otherwise,
   2352 			 * when we fall out, we just clean the list.
   2353 			 */
   2354 			PDEBUG(10, printf("remove: inv pte at %p(%x) ", pte, *pte));
   2355 			pa = pmap_pte_pa(pte);
   2356 
   2357 			if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) {
   2358 				/* Add to the clean list. */
   2359 				cleanlist[cleanlist_idx].pte = pte;
   2360 				cleanlist[cleanlist_idx].va = sva;
   2361 				cleanlist_idx++;
   2362 			} else if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) {
   2363 				int cnt;
   2364 
   2365 				/* Nuke everything if needed. */
   2366 				if (pmap_active) {
   2367 					cpu_idcache_wbinv_all();
   2368 					cpu_tlb_flushID();
   2369 				}
   2370 
   2371 				/*
   2372 				 * Roll back the previous PTE list,
   2373 				 * and zero out the current PTE.
   2374 				 */
   2375 				for (cnt = 0; cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) {
   2376 					*cleanlist[cnt].pte = 0;
   2377 					pmap_pte_delref(pmap, cleanlist[cnt].va);
   2378 				}
   2379 				*pte = 0;
   2380 				pmap_pte_delref(pmap, sva);
   2381 				cleanlist_idx++;
   2382 			} else {
   2383 				/*
   2384 				 * We've already nuked the cache and
   2385 				 * TLB, so just carry on regardless,
   2386 				 * and we won't need to do it again
   2387 				 */
   2388 				*pte = 0;
   2389 				pmap_pte_delref(pmap, sva);
   2390 			}
   2391 
   2392 			/*
   2393 			 * Update flags. In a number of circumstances,
   2394 			 * we could cluster a lot of these and do a
   2395 			 * number of sequential pages in one go.
   2396 			 */
   2397 			if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
   2398 				struct pv_entry *pve;
   2399 				simple_lock(&pg->mdpage.pvh_slock);
   2400 				pve = pmap_remove_pv(pg, pmap, sva);
   2401 				pmap_free_pv(pmap, pve);
   2402 				pmap_vac_me_harder(pmap, pg, ptes, FALSE);
   2403 				simple_unlock(&pg->mdpage.pvh_slock);
   2404 			}
   2405 		}
   2406 		sva += NBPG;
   2407 		pte++;
   2408 	}
   2409 
   2410 	pmap_unmap_ptes(pmap);
   2411 	/*
   2412 	 * Now, if we've fallen through down to here, chances are that there
   2413 	 * are less than PMAP_REMOVE_CLEAN_LIST_SIZE mappings left.
   2414 	 */
   2415 	if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) {
   2416 		u_int cnt;
   2417 
   2418 		for (cnt = 0; cnt < cleanlist_idx; cnt++) {
   2419 			if (pmap_active) {
   2420 				cpu_idcache_wbinv_range(cleanlist[cnt].va,
   2421 				    NBPG);
   2422 				*cleanlist[cnt].pte = 0;
   2423 				cpu_tlb_flushID_SE(cleanlist[cnt].va);
   2424 			} else
   2425 				*cleanlist[cnt].pte = 0;
   2426 			pmap_pte_delref(pmap, cleanlist[cnt].va);
   2427 		}
   2428 	}
   2429 	PMAP_MAP_TO_HEAD_UNLOCK();
   2430 }
   2431 
   2432 /*
   2433  * Routine:	pmap_remove_all
   2434  * Function:
   2435  *		Removes this physical page from
   2436  *		all physical maps in which it resides.
   2437  *		Reflects back modify bits to the pager.
   2438  */
   2439 
   2440 static void
   2441 pmap_remove_all(struct vm_page *pg)
   2442 {
   2443 	struct pv_entry *pv, *npv;
   2444 	struct pmap *pmap;
   2445 	pt_entry_t *pte, *ptes;
   2446 
   2447 	PDEBUG(0, printf("pmap_remove_all: pa=%lx ", VM_PAGE_TO_PHYS(pg)));
   2448 
   2449 	/* set vm_page => pmap locking */
   2450 	PMAP_HEAD_TO_MAP_LOCK();
   2451 
   2452 	simple_lock(&pg->mdpage.pvh_slock);
   2453 
   2454 	pv = pg->mdpage.pvh_list;
   2455 	if (pv == NULL) {
   2456 		PDEBUG(0, printf("free page\n"));
   2457 		simple_unlock(&pg->mdpage.pvh_slock);
   2458 		PMAP_HEAD_TO_MAP_UNLOCK();
   2459 		return;
   2460 	}
   2461 	pmap_clean_page(pv, FALSE);
   2462 
   2463 	while (pv) {
   2464 		pmap = pv->pv_pmap;
   2465 		ptes = pmap_map_ptes(pmap);
   2466 		pte = &ptes[arm_btop(pv->pv_va)];
   2467 
   2468 		PDEBUG(0, printf("[%p,%08x,%08lx,%08x] ", pmap, *pte,
   2469 		    pv->pv_va, pv->pv_flags));
   2470 #ifdef DEBUG
   2471 		if (pmap_pde_page(pmap_pde(pmap, pv->pv_va)) == 0 ||
   2472 		    pmap_pte_v(pte) == 0 ||
   2473 		    pmap_pte_pa(pte) != VM_PAGE_TO_PHYS(pg))
   2474 			panic("pmap_remove_all: bad mapping");
   2475 #endif	/* DEBUG */
   2476 
   2477 		/*
   2478 		 * Update statistics
   2479 		 */
   2480 		--pmap->pm_stats.resident_count;
   2481 
   2482 		/* Wired bit */
   2483 		if (pv->pv_flags & PVF_WIRED)
   2484 			--pmap->pm_stats.wired_count;
   2485 
   2486 		/*
   2487 		 * Invalidate the PTEs.
   2488 		 * XXX: should cluster them up and invalidate as many
   2489 		 * as possible at once.
   2490 		 */
   2491 
   2492 #ifdef needednotdone
   2493 reduce wiring count on page table pages as references drop
   2494 #endif
   2495 
   2496 		*pte = 0;
   2497 		pmap_pte_delref(pmap, pv->pv_va);
   2498 
   2499 		npv = pv->pv_next;
   2500 		pmap_free_pv(pmap, pv);
   2501 		pv = npv;
   2502 		pmap_unmap_ptes(pmap);
   2503 	}
   2504 	pg->mdpage.pvh_list = NULL;
   2505 	simple_unlock(&pg->mdpage.pvh_slock);
   2506 	PMAP_HEAD_TO_MAP_UNLOCK();
   2507 
   2508 	PDEBUG(0, printf("done\n"));
   2509 	cpu_tlb_flushID();
   2510 	cpu_cpwait();
   2511 }
   2512 
   2513 
   2514 /*
   2515  * Set the physical protection on the specified range of this map as requested.
   2516  */
   2517 
   2518 void
   2519 pmap_protect(struct pmap *pmap, vaddr_t sva, vaddr_t eva, vm_prot_t prot)
   2520 {
   2521 	pt_entry_t *pte = NULL, *ptes;
   2522 	struct vm_page *pg;
   2523 	int armprot;
   2524 	int flush = 0;
   2525 	paddr_t pa;
   2526 
   2527 	PDEBUG(0, printf("pmap_protect: pmap=%p %08lx->%08lx %x\n",
   2528 	    pmap, sva, eva, prot));
   2529 
   2530 	if (~prot & VM_PROT_READ) {
   2531 		/* Just remove the mappings. */
   2532 		pmap_remove(pmap, sva, eva);
   2533 		/* pmap_update not needed as it should be called by the caller
   2534 		 * of pmap_protect */
   2535 		return;
   2536 	}
   2537 	if (prot & VM_PROT_WRITE) {
   2538 		/*
   2539 		 * If this is a read->write transition, just ignore it and let
   2540 		 * uvm_fault() take care of it later.
   2541 		 */
   2542 		return;
   2543 	}
   2544 
   2545 	/* Need to lock map->head */
   2546 	PMAP_MAP_TO_HEAD_LOCK();
   2547 
   2548 	ptes = pmap_map_ptes(pmap);
   2549 
   2550 	/*
   2551 	 * OK, at this point, we know we're doing write-protect operation.
   2552 	 * If the pmap is active, write-back the range.
   2553 	 */
   2554 	if (pmap_is_curpmap(pmap))
   2555 		cpu_dcache_wb_range(sva, eva - sva);
   2556 
   2557 	/*
   2558 	 * We need to acquire a pointer to a page table page before entering
   2559 	 * the following loop.
   2560 	 */
   2561 	while (sva < eva) {
   2562 		if (pmap_pde_page(pmap_pde(pmap, sva)))
   2563 			break;
   2564 		sva = (sva & L1_S_FRAME) + L1_S_SIZE;
   2565 	}
   2566 
   2567 	pte = &ptes[arm_btop(sva)];
   2568 
   2569 	while (sva < eva) {
   2570 		/* only check once in a while */
   2571 		if ((sva & L2_ADDR_BITS) == 0) {
   2572 			if (!pmap_pde_page(pmap_pde(pmap, sva))) {
   2573 				/* We can race ahead here, to the next pde. */
   2574 				sva += L1_S_SIZE;
   2575 				pte += arm_btop(L1_S_SIZE);
   2576 				continue;
   2577 			}
   2578 		}
   2579 
   2580 		if (!pmap_pte_v(pte))
   2581 			goto next;
   2582 
   2583 		flush = 1;
   2584 
   2585 		armprot = 0;
   2586 		if (sva < VM_MAXUSER_ADDRESS)
   2587 			armprot |= L2_S_PROT_U;
   2588 		else if (sva < VM_MAX_ADDRESS)
   2589 			armprot |= L2_S_PROT_W;  /* XXX Ekk what is this ? */
   2590 		*pte = (*pte & 0xfffff00f) | armprot;
   2591 
   2592 		pa = pmap_pte_pa(pte);
   2593 
   2594 		/* Get the physical page index */
   2595 
   2596 		/* Clear write flag */
   2597 		if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
   2598 			simple_lock(&pg->mdpage.pvh_slock);
   2599 			(void) pmap_modify_pv(pmap, sva, pg, PVF_WRITE, 0);
   2600 			pmap_vac_me_harder(pmap, pg, ptes, FALSE);
   2601 			simple_unlock(&pg->mdpage.pvh_slock);
   2602 		}
   2603 
   2604 next:
   2605 		sva += NBPG;
   2606 		pte++;
   2607 	}
   2608 	pmap_unmap_ptes(pmap);
   2609 	PMAP_MAP_TO_HEAD_UNLOCK();
   2610 	if (flush)
   2611 		cpu_tlb_flushID();
   2612 }
   2613 
   2614 /*
   2615  * void pmap_enter(struct pmap *pmap, vaddr_t va, paddr_t pa, vm_prot_t prot,
   2616  * int flags)
   2617  *
   2618  *      Insert the given physical page (p) at
   2619  *      the specified virtual address (v) in the
   2620  *      target physical map with the protection requested.
   2621  *
   2622  *      If specified, the page will be wired down, meaning
   2623  *      that the related pte can not be reclaimed.
   2624  *
   2625  *      NB:  This is the only routine which MAY NOT lazy-evaluate
   2626  *      or lose information.  That is, this routine must actually
   2627  *      insert this page into the given map NOW.
   2628  */
   2629 
   2630 int
   2631 pmap_enter(struct pmap *pmap, vaddr_t va, paddr_t pa, vm_prot_t prot,
   2632     int flags)
   2633 {
   2634 	pt_entry_t *ptes, opte, npte;
   2635 	paddr_t opa;
   2636 	boolean_t wired = (flags & PMAP_WIRED) != 0;
   2637 	struct vm_page *pg;
   2638 	struct pv_entry *pve;
   2639 	int error, nflags;
   2640 
   2641 	PDEBUG(5, printf("pmap_enter: V%08lx P%08lx in pmap %p prot=%08x, wired = %d\n",
   2642 	    va, pa, pmap, prot, wired));
   2643 
   2644 #ifdef DIAGNOSTIC
   2645 	/* Valid address ? */
   2646 	if (va >= (pmap_curmaxkvaddr))
   2647 		panic("pmap_enter: too big");
   2648 	if (pmap != pmap_kernel() && va != 0) {
   2649 		if (va < VM_MIN_ADDRESS || va >= VM_MAXUSER_ADDRESS)
   2650 			panic("pmap_enter: kernel page in user map");
   2651 	} else {
   2652 		if (va >= VM_MIN_ADDRESS && va < VM_MAXUSER_ADDRESS)
   2653 			panic("pmap_enter: user page in kernel map");
   2654 		if (va >= VM_MAXUSER_ADDRESS && va < VM_MAX_ADDRESS)
   2655 			panic("pmap_enter: entering PT page");
   2656 	}
   2657 #endif
   2658 
   2659 	KDASSERT(((va | pa) & PGOFSET) == 0);
   2660 
   2661 	/*
   2662 	 * Get a pointer to the page.  Later on in this function, we
   2663 	 * test for a managed page by checking pg != NULL.
   2664 	 */
   2665 	pg = pmap_initialized ? PHYS_TO_VM_PAGE(pa) : NULL;
   2666 
   2667 	/* get lock */
   2668 	PMAP_MAP_TO_HEAD_LOCK();
   2669 
   2670 	/*
   2671 	 * map the ptes.  If there's not already an L2 table for this
   2672 	 * address, allocate one.
   2673 	 */
   2674 	ptes = pmap_map_ptes(pmap);		/* locks pmap */
   2675 	if (pmap_pde_v(pmap_pde(pmap, va)) == 0) {
   2676 		struct vm_page *ptp;
   2677 
   2678 		/* kernel should be pre-grown */
   2679 		KASSERT(pmap != pmap_kernel());
   2680 
   2681 		/* if failure is allowed then don't try too hard */
   2682 		ptp = pmap_get_ptp(pmap, va & L1_S_FRAME);
   2683 		if (ptp == NULL) {
   2684 			if (flags & PMAP_CANFAIL) {
   2685 				error = ENOMEM;
   2686 				goto out;
   2687 			}
   2688 			panic("pmap_enter: get ptp failed");
   2689 		}
   2690 	}
   2691 	opte = ptes[arm_btop(va)];
   2692 
   2693 	nflags = 0;
   2694 	if (prot & VM_PROT_WRITE)
   2695 		nflags |= PVF_WRITE;
   2696 	if (wired)
   2697 		nflags |= PVF_WIRED;
   2698 
   2699 	/* Is the pte valid ? If so then this page is already mapped */
   2700 	if (l2pte_valid(opte)) {
   2701 		/* Get the physical address of the current page mapped */
   2702 		opa = l2pte_pa(opte);
   2703 
   2704 		/* Are we mapping the same page ? */
   2705 		if (opa == pa) {
   2706 			/* Has the wiring changed ? */
   2707 			if (pg != NULL) {
   2708 				simple_lock(&pg->mdpage.pvh_slock);
   2709 				(void) pmap_modify_pv(pmap, va, pg,
   2710 				    PVF_WRITE | PVF_WIRED, nflags);
   2711 				simple_unlock(&pg->mdpage.pvh_slock);
   2712  			}
   2713 		} else {
   2714 			struct vm_page *opg;
   2715 
   2716 			/* We are replacing the page with a new one. */
   2717 			cpu_idcache_wbinv_range(va, NBPG);
   2718 
   2719 			/*
   2720 			 * If it is part of our managed memory then we
   2721 			 * must remove it from the PV list
   2722 			 */
   2723 			if ((opg = PHYS_TO_VM_PAGE(opa)) != NULL) {
   2724 				simple_lock(&opg->mdpage.pvh_slock);
   2725 				pve = pmap_remove_pv(opg, pmap, va);
   2726 				simple_unlock(&opg->mdpage.pvh_slock);
   2727 			} else {
   2728 				pve = NULL;
   2729 			}
   2730 
   2731 			goto enter;
   2732 		}
   2733 	} else {
   2734 		opa = 0;
   2735 		pve = NULL;
   2736 		pmap_pte_addref(pmap, va);
   2737 
   2738 		/* pte is not valid so we must be hooking in a new page */
   2739 		++pmap->pm_stats.resident_count;
   2740 
   2741 	enter:
   2742 		/*
   2743 		 * Enter on the PV list if part of our managed memory
   2744 		 */
   2745 		if (pg != NULL) {
   2746 			if (pve == NULL) {
   2747 				pve = pmap_alloc_pv(pmap, ALLOCPV_NEED);
   2748 				if (pve == NULL) {
   2749 					if (flags & PMAP_CANFAIL) {
   2750 						error = ENOMEM;
   2751 						goto out;
   2752 					}
   2753 					panic("pmap_enter: no pv entries "
   2754 					    "available");
   2755 				}
   2756 			}
   2757 			/* enter_pv locks pvh when adding */
   2758 			pmap_enter_pv(pg, pve, pmap, va, NULL, nflags);
   2759 		} else {
   2760 			if (pve != NULL)
   2761 				pmap_free_pv(pmap, pve);
   2762 		}
   2763 	}
   2764 
   2765 	/* Construct the pte, giving the correct access. */
   2766 	npte = pa;
   2767 
   2768 	/* VA 0 is magic. */
   2769 	if (pmap != pmap_kernel() && va != vector_page)
   2770 		npte |= L2_S_PROT_U;
   2771 
   2772 	if (pg != NULL) {
   2773 #ifdef DIAGNOSTIC
   2774 		if ((flags & VM_PROT_ALL) & ~prot)
   2775 			panic("pmap_enter: access_type exceeds prot");
   2776 #endif
   2777 		npte |= pte_l2_s_cache_mode;
   2778 		if (flags & VM_PROT_WRITE) {
   2779 			npte |= L2_S_PROTO | L2_S_PROT_W;
   2780 			pg->mdpage.pvh_attrs |= PVF_REF | PVF_MOD;
   2781 		} else if (flags & VM_PROT_ALL) {
   2782 			npte |= L2_S_PROTO;
   2783 			pg->mdpage.pvh_attrs |= PVF_REF;
   2784 		} else
   2785 			npte |= L2_TYPE_INV;
   2786 	} else {
   2787 		if (prot & VM_PROT_WRITE)
   2788 			npte |= L2_S_PROTO | L2_S_PROT_W;
   2789 		else if (prot & VM_PROT_ALL)
   2790 			npte |= L2_S_PROTO;
   2791 		else
   2792 			npte |= L2_TYPE_INV;
   2793 	}
   2794 
   2795 	ptes[arm_btop(va)] = npte;
   2796 
   2797 	if (pg != NULL) {
   2798 		simple_lock(&pg->mdpage.pvh_slock);
   2799  		pmap_vac_me_harder(pmap, pg, ptes, pmap_is_curpmap(pmap));
   2800 		simple_unlock(&pg->mdpage.pvh_slock);
   2801 	}
   2802 
   2803 	/* Better flush the TLB ... */
   2804 	cpu_tlb_flushID_SE(va);
   2805 	error = 0;
   2806 out:
   2807 	pmap_unmap_ptes(pmap);			/* unlocks pmap */
   2808 	PMAP_MAP_TO_HEAD_UNLOCK();
   2809 
   2810 	return error;
   2811 }
   2812 
   2813 /*
   2814  * pmap_kenter_pa: enter a kernel mapping
   2815  *
   2816  * => no need to lock anything assume va is already allocated
   2817  * => should be faster than normal pmap enter function
   2818  */
   2819 void
   2820 pmap_kenter_pa(vaddr_t va, paddr_t pa, vm_prot_t prot)
   2821 {
   2822 	pt_entry_t *pte;
   2823 
   2824 	pte = vtopte(va);
   2825 	KASSERT(!pmap_pte_v(pte));
   2826 
   2827 	*pte = L2_S_PROTO | pa |
   2828 	    L2_S_PROT(PTE_KERNEL, prot) | pte_l2_s_cache_mode;
   2829 }
   2830 
   2831 void
   2832 pmap_kremove(vaddr_t va, vsize_t len)
   2833 {
   2834 	pt_entry_t *pte;
   2835 
   2836 	for (len >>= PAGE_SHIFT; len > 0; len--, va += PAGE_SIZE) {
   2837 
   2838 		/*
   2839 		 * We assume that we will only be called with small
   2840 		 * regions of memory.
   2841 		 */
   2842 
   2843 		KASSERT(pmap_pde_page(pmap_pde(pmap_kernel(), va)));
   2844 		pte = vtopte(va);
   2845 		cpu_idcache_wbinv_range(va, PAGE_SIZE);
   2846 		*pte = 0;
   2847 		cpu_tlb_flushID_SE(va);
   2848 	}
   2849 }
   2850 
   2851 /*
   2852  * pmap_page_protect:
   2853  *
   2854  * Lower the permission for all mappings to a given page.
   2855  */
   2856 
   2857 void
   2858 pmap_page_protect(struct vm_page *pg, vm_prot_t prot)
   2859 {
   2860 
   2861 	PDEBUG(0, printf("pmap_page_protect(pa=%lx, prot=%d)\n",
   2862 	    VM_PAGE_TO_PHYS(pg), prot));
   2863 
   2864 	switch(prot) {
   2865 	case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE:
   2866 	case VM_PROT_READ|VM_PROT_WRITE:
   2867 		return;
   2868 
   2869 	case VM_PROT_READ:
   2870 	case VM_PROT_READ|VM_PROT_EXECUTE:
   2871 		pmap_clearbit(pg, PVF_WRITE);
   2872 		break;
   2873 
   2874 	default:
   2875 		pmap_remove_all(pg);
   2876 		break;
   2877 	}
   2878 }
   2879 
   2880 
   2881 /*
   2882  * Routine:	pmap_unwire
   2883  * Function:	Clear the wired attribute for a map/virtual-address
   2884  *		pair.
   2885  * In/out conditions:
   2886  *		The mapping must already exist in the pmap.
   2887  */
   2888 
   2889 void
   2890 pmap_unwire(struct pmap *pmap, vaddr_t va)
   2891 {
   2892 	pt_entry_t *ptes;
   2893 	struct vm_page *pg;
   2894 	paddr_t pa;
   2895 
   2896 	PMAP_MAP_TO_HEAD_LOCK();
   2897 	ptes = pmap_map_ptes(pmap);		/* locks pmap */
   2898 
   2899 	if (pmap_pde_v(pmap_pde(pmap, va))) {
   2900 #ifdef DIAGNOSTIC
   2901 		if (l2pte_valid(ptes[arm_btop(va)]) == 0)
   2902 			panic("pmap_unwire: invalid L2 PTE");
   2903 #endif
   2904 		/* Extract the physical address of the page */
   2905 		pa = l2pte_pa(ptes[arm_btop(va)]);
   2906 
   2907 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
   2908 			goto out;
   2909 
   2910 		/* Update the wired bit in the pv entry for this page. */
   2911 		simple_lock(&pg->mdpage.pvh_slock);
   2912 		(void) pmap_modify_pv(pmap, va, pg, PVF_WIRED, 0);
   2913 		simple_unlock(&pg->mdpage.pvh_slock);
   2914 	}
   2915 #ifdef DIAGNOSTIC
   2916 	else {
   2917 		panic("pmap_unwire: invalid L1 PTE");
   2918 	}
   2919 #endif
   2920  out:
   2921 	pmap_unmap_ptes(pmap);			/* unlocks pmap */
   2922 	PMAP_MAP_TO_HEAD_UNLOCK();
   2923 }
   2924 
   2925 /*
   2926  * Routine:  pmap_extract
   2927  * Function:
   2928  *           Extract the physical page address associated
   2929  *           with the given map/virtual_address pair.
   2930  */
   2931 boolean_t
   2932 pmap_extract(struct pmap *pmap, vaddr_t va, paddr_t *pap)
   2933 {
   2934 	pd_entry_t *pde;
   2935 	pt_entry_t *pte, *ptes;
   2936 	paddr_t pa;
   2937 
   2938 	PDEBUG(5, printf("pmap_extract: pmap=%p, va=0x%08lx -> ", pmap, va));
   2939 
   2940 	ptes = pmap_map_ptes(pmap);		/* locks pmap */
   2941 
   2942 	pde = pmap_pde(pmap, va);
   2943 	pte = &ptes[arm_btop(va)];
   2944 
   2945 	if (pmap_pde_section(pde)) {
   2946 		pa = (*pde & L1_S_FRAME) | (va & L1_S_OFFSET);
   2947 		PDEBUG(5, printf("section pa=0x%08lx\n", pa));
   2948 		goto out;
   2949 	} else if (pmap_pde_page(pde) == 0 || pmap_pte_v(pte) == 0) {
   2950 		PDEBUG(5, printf("no mapping\n"));
   2951 		goto failed;
   2952 	}
   2953 
   2954 	if ((*pte & L2_TYPE_MASK) == L2_TYPE_L) {
   2955 		pa = (*pte & L2_L_FRAME) | (va & L2_L_OFFSET);
   2956 		PDEBUG(5, printf("large page pa=0x%08lx\n", pa));
   2957 		goto out;
   2958 	}
   2959 
   2960 	pa = (*pte & L2_S_FRAME) | (va & L2_S_OFFSET);
   2961 	PDEBUG(5, printf("small page pa=0x%08lx\n", pa));
   2962 
   2963  out:
   2964 	if (pap != NULL)
   2965 		*pap = pa;
   2966 
   2967 	pmap_unmap_ptes(pmap);			/* unlocks pmap */
   2968 	return (TRUE);
   2969 
   2970  failed:
   2971 	pmap_unmap_ptes(pmap);			/* unlocks pmap */
   2972 	return (FALSE);
   2973 }
   2974 
   2975 
   2976 /*
   2977  * pmap_copy:
   2978  *
   2979  *	Copy the range specified by src_addr/len from the source map to the
   2980  *	range dst_addr/len in the destination map.
   2981  *
   2982  *	This routine is only advisory and need not do anything.
   2983  */
   2984 /* Call deleted in <arm/arm32/pmap.h> */
   2985 
   2986 #if defined(PMAP_DEBUG)
   2987 void
   2988 pmap_dump_pvlist(phys, m)
   2989 	vaddr_t phys;
   2990 	char *m;
   2991 {
   2992 	struct vm_page *pg;
   2993 	struct pv_entry *pv;
   2994 
   2995 	if ((pg = PHYS_TO_VM_PAGE(phys)) == NULL) {
   2996 		printf("INVALID PA\n");
   2997 		return;
   2998 	}
   2999 	simple_lock(&pg->mdpage.pvh_slock);
   3000 	printf("%s %08lx:", m, phys);
   3001 	if (pg->mdpage.pvh_list == NULL) {
   3002 		printf(" no mappings\n");
   3003 		return;
   3004 	}
   3005 
   3006 	for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next)
   3007 		printf(" pmap %p va %08lx flags %08x", pv->pv_pmap,
   3008 		    pv->pv_va, pv->pv_flags);
   3009 
   3010 	printf("\n");
   3011 	simple_unlock(&pg->mdpage.pvh_slock);
   3012 }
   3013 
   3014 #endif	/* PMAP_DEBUG */
   3015 
   3016 static pt_entry_t *
   3017 pmap_map_ptes(struct pmap *pmap)
   3018 {
   3019 	struct proc *p;
   3020 
   3021     	/* the kernel's pmap is always accessible */
   3022 	if (pmap == pmap_kernel()) {
   3023 		return (pt_entry_t *)PTE_BASE;
   3024 	}
   3025 
   3026 	if (pmap_is_curpmap(pmap)) {
   3027 		simple_lock(&pmap->pm_obj.vmobjlock);
   3028 		return (pt_entry_t *)PTE_BASE;
   3029 	}
   3030 
   3031 	p = curproc;
   3032 	KDASSERT(p != NULL);
   3033 
   3034 	/* need to lock both curpmap and pmap: use ordered locking */
   3035 	if ((vaddr_t) pmap < (vaddr_t) p->p_vmspace->vm_map.pmap) {
   3036 		simple_lock(&pmap->pm_obj.vmobjlock);
   3037 		simple_lock(&p->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
   3038 	} else {
   3039 		simple_lock(&p->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
   3040 		simple_lock(&pmap->pm_obj.vmobjlock);
   3041 	}
   3042 
   3043 	pmap_map_in_l1(p->p_vmspace->vm_map.pmap, APTE_BASE, pmap->pm_pptpt,
   3044 	    FALSE);
   3045 	cpu_tlb_flushD();
   3046 	cpu_cpwait();
   3047 	return (pt_entry_t *)APTE_BASE;
   3048 }
   3049 
   3050 /*
   3051  * pmap_unmap_ptes: unlock the PTE mapping of "pmap"
   3052  */
   3053 
   3054 static void
   3055 pmap_unmap_ptes(struct pmap *pmap)
   3056 {
   3057 
   3058 	if (pmap == pmap_kernel()) {
   3059 		return;
   3060 	}
   3061 	if (pmap_is_curpmap(pmap)) {
   3062 		simple_unlock(&pmap->pm_obj.vmobjlock);
   3063 	} else {
   3064 		KDASSERT(curproc != NULL);
   3065 		simple_unlock(&pmap->pm_obj.vmobjlock);
   3066 		simple_unlock(
   3067 		    &curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
   3068 	}
   3069 }
   3070 
   3071 /*
   3072  * Modify pte bits for all ptes corresponding to the given physical address.
   3073  * We use `maskbits' rather than `clearbits' because we're always passing
   3074  * constants and the latter would require an extra inversion at run-time.
   3075  */
   3076 
   3077 static void
   3078 pmap_clearbit(struct vm_page *pg, u_int maskbits)
   3079 {
   3080 	struct pv_entry *pv;
   3081 	pt_entry_t *ptes;
   3082 	vaddr_t va;
   3083 	int tlbentry;
   3084 
   3085 	PDEBUG(1, printf("pmap_clearbit: pa=%08lx mask=%08x\n",
   3086 	    VM_PAGE_TO_PHYS(pg), maskbits));
   3087 
   3088 	tlbentry = 0;
   3089 
   3090 	PMAP_HEAD_TO_MAP_LOCK();
   3091 	simple_lock(&pg->mdpage.pvh_slock);
   3092 
   3093 	/*
   3094 	 * Clear saved attributes (modify, reference)
   3095 	 */
   3096 	pg->mdpage.pvh_attrs &= ~maskbits;
   3097 
   3098 	if (pg->mdpage.pvh_list == NULL) {
   3099 		simple_unlock(&pg->mdpage.pvh_slock);
   3100 		PMAP_HEAD_TO_MAP_UNLOCK();
   3101 		return;
   3102 	}
   3103 
   3104 	/*
   3105 	 * Loop over all current mappings setting/clearing as appropos
   3106 	 */
   3107 	for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
   3108 		va = pv->pv_va;
   3109 		pv->pv_flags &= ~maskbits;
   3110 		ptes = pmap_map_ptes(pv->pv_pmap);	/* locks pmap */
   3111 		KASSERT(pmap_pde_v(pmap_pde(pv->pv_pmap, va)));
   3112 		if (maskbits & (PVF_WRITE|PVF_MOD)) {
   3113 			if ((pv->pv_flags & PVF_NC)) {
   3114 				/*
   3115 				 * Entry is not cacheable: reenable
   3116 				 * the cache, nothing to flush
   3117 				 *
   3118 				 * Don't turn caching on again if this
   3119 				 * is a modified emulation.  This
   3120 				 * would be inconsitent with the
   3121 				 * settings created by
   3122 				 * pmap_vac_me_harder().
   3123 				 *
   3124 				 * There's no need to call
   3125 				 * pmap_vac_me_harder() here: all
   3126 				 * pages are loosing their write
   3127 				 * permission.
   3128 				 *
   3129 				 */
   3130 				if (maskbits & PVF_WRITE) {
   3131 					ptes[arm_btop(va)] |=
   3132 					    pte_l2_s_cache_mode;
   3133 					pv->pv_flags &= ~PVF_NC;
   3134 				}
   3135 			} else if (pmap_is_curpmap(pv->pv_pmap)) {
   3136 				/*
   3137 				 * Entry is cacheable: check if pmap is
   3138 				 * current if it is flush it,
   3139 				 * otherwise it won't be in the cache
   3140 				 */
   3141 				cpu_idcache_wbinv_range(pv->pv_va, NBPG);
   3142 			}
   3143 
   3144 			/* make the pte read only */
   3145 			ptes[arm_btop(va)] &= ~L2_S_PROT_W;
   3146 		}
   3147 
   3148 		if (maskbits & PVF_REF)
   3149 			ptes[arm_btop(va)] =
   3150 			    (ptes[arm_btop(va)] & ~L2_TYPE_MASK) | L2_TYPE_INV;
   3151 
   3152 		if (pmap_is_curpmap(pv->pv_pmap)) {
   3153 			/*
   3154 			 * if we had cacheable pte's we'd clean the
   3155 			 * pte out to memory here
   3156 			 *
   3157 			 * flush tlb entry as it's in the current pmap
   3158 			 */
   3159 			cpu_tlb_flushID_SE(pv->pv_va);
   3160 		}
   3161 		pmap_unmap_ptes(pv->pv_pmap);		/* unlocks pmap */
   3162 	}
   3163 	cpu_cpwait();
   3164 
   3165 	simple_unlock(&pg->mdpage.pvh_slock);
   3166 	PMAP_HEAD_TO_MAP_UNLOCK();
   3167 }
   3168 
   3169 /*
   3170  * pmap_clear_modify:
   3171  *
   3172  *	Clear the "modified" attribute for a page.
   3173  */
   3174 boolean_t
   3175 pmap_clear_modify(struct vm_page *pg)
   3176 {
   3177 	boolean_t rv;
   3178 
   3179 	if (pg->mdpage.pvh_attrs & PVF_MOD) {
   3180 		rv = TRUE;
   3181 		pmap_clearbit(pg, PVF_MOD);
   3182 	} else
   3183 		rv = FALSE;
   3184 
   3185 	PDEBUG(0, printf("pmap_clear_modify pa=%08lx -> %d\n",
   3186 	    VM_PAGE_TO_PHYS(pg), rv));
   3187 
   3188 	return (rv);
   3189 }
   3190 
   3191 /*
   3192  * pmap_clear_reference:
   3193  *
   3194  *	Clear the "referenced" attribute for a page.
   3195  */
   3196 boolean_t
   3197 pmap_clear_reference(struct vm_page *pg)
   3198 {
   3199 	boolean_t rv;
   3200 
   3201 	if (pg->mdpage.pvh_attrs & PVF_REF) {
   3202 		rv = TRUE;
   3203 		pmap_clearbit(pg, PVF_REF);
   3204 	} else
   3205 		rv = FALSE;
   3206 
   3207 	PDEBUG(0, printf("pmap_clear_reference pa=%08lx -> %d\n",
   3208 	    VM_PAGE_TO_PHYS(pg), rv));
   3209 
   3210 	return (rv);
   3211 }
   3212 
   3213 /*
   3214  * pmap_is_modified:
   3215  *
   3216  *	Test if a page has the "modified" attribute.
   3217  */
   3218 /* See <arm/arm32/pmap.h> */
   3219 
   3220 /*
   3221  * pmap_is_referenced:
   3222  *
   3223  *	Test if a page has the "referenced" attribute.
   3224  */
   3225 /* See <arm/arm32/pmap.h> */
   3226 
   3227 int
   3228 pmap_modified_emulation(struct pmap *pmap, vaddr_t va)
   3229 {
   3230 	pt_entry_t *ptes;
   3231 	struct vm_page *pg;
   3232 	paddr_t pa;
   3233 	u_int flags;
   3234 	int rv = 0;
   3235 
   3236 	PDEBUG(2, printf("pmap_modified_emulation\n"));
   3237 
   3238 	PMAP_MAP_TO_HEAD_LOCK();
   3239 	ptes = pmap_map_ptes(pmap);		/* locks pmap */
   3240 
   3241 	if (pmap_pde_v(pmap_pde(pmap, va)) == 0) {
   3242 		PDEBUG(2, printf("L1 PTE invalid\n"));
   3243 		goto out;
   3244 	}
   3245 
   3246 	PDEBUG(1, printf("pte=%08x\n", ptes[arm_btop(va)]));
   3247 
   3248 	/* Check for a invalid pte */
   3249 	if (l2pte_valid(ptes[arm_btop(va)]) == 0)
   3250 		goto out;
   3251 
   3252 	/* This can happen if user code tries to access kernel memory. */
   3253 	if ((ptes[arm_btop(va)] & L2_S_PROT_W) != 0)
   3254 		goto out;
   3255 
   3256 	/* Extract the physical address of the page */
   3257 	pa = l2pte_pa(ptes[arm_btop(va)]);
   3258 	if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
   3259 		goto out;
   3260 
   3261 	/* Get the current flags for this page. */
   3262 	simple_lock(&pg->mdpage.pvh_slock);
   3263 
   3264 	flags = pmap_modify_pv(pmap, va, pg, 0, 0);
   3265 	PDEBUG(2, printf("pmap_modified_emulation: flags = %08x\n", flags));
   3266 
   3267 	/*
   3268 	 * Do the flags say this page is writable ? If not then it is a
   3269 	 * genuine write fault. If yes then the write fault is our fault
   3270 	 * as we did not reflect the write access in the PTE. Now we know
   3271 	 * a write has occurred we can correct this and also set the
   3272 	 * modified bit
   3273 	 */
   3274 	if (~flags & PVF_WRITE) {
   3275 	    	simple_unlock(&pg->mdpage.pvh_slock);
   3276 		goto out;
   3277 	}
   3278 
   3279 	PDEBUG(0,
   3280 	    printf("pmap_modified_emulation: Got a hit va=%08lx, pte = %08x\n",
   3281 	    va, ptes[arm_btop(va)]));
   3282 	pg->mdpage.pvh_attrs |= PVF_REF | PVF_MOD;
   3283 
   3284 	/*
   3285 	 * Re-enable write permissions for the page.  No need to call
   3286 	 * pmap_vac_me_harder(), since this is just a
   3287 	 * modified-emulation fault, and the PVF_WRITE bit isn't changing.
   3288 	 * We've already set the cacheable bits based on the assumption
   3289 	 * that we can write to this page.
   3290 	 */
   3291 	ptes[arm_btop(va)] =
   3292 	    (ptes[arm_btop(va)] & ~L2_TYPE_MASK) | L2_S_PROTO | L2_S_PROT_W;
   3293 	PDEBUG(0, printf("->(%08x)\n", ptes[arm_btop(va)]));
   3294 
   3295 	simple_unlock(&pg->mdpage.pvh_slock);
   3296 
   3297 	cpu_tlb_flushID_SE(va);
   3298 	cpu_cpwait();
   3299 	rv = 1;
   3300  out:
   3301 	pmap_unmap_ptes(pmap);			/* unlocks pmap */
   3302 	PMAP_MAP_TO_HEAD_UNLOCK();
   3303 	return (rv);
   3304 }
   3305 
   3306 int
   3307 pmap_handled_emulation(struct pmap *pmap, vaddr_t va)
   3308 {
   3309 	pt_entry_t *ptes;
   3310 	struct vm_page *pg;
   3311 	paddr_t pa;
   3312 	int rv = 0;
   3313 
   3314 	PDEBUG(2, printf("pmap_handled_emulation\n"));
   3315 
   3316 	PMAP_MAP_TO_HEAD_LOCK();
   3317 	ptes = pmap_map_ptes(pmap);		/* locks pmap */
   3318 
   3319 	if (pmap_pde_v(pmap_pde(pmap, va)) == 0) {
   3320 		PDEBUG(2, printf("L1 PTE invalid\n"));
   3321 		goto out;
   3322 	}
   3323 
   3324 	PDEBUG(1, printf("pte=%08x\n", ptes[arm_btop(va)]));
   3325 
   3326 	/* Check for invalid pte */
   3327 	if (l2pte_valid(ptes[arm_btop(va)]) == 0)
   3328 		goto out;
   3329 
   3330 	/* This can happen if user code tries to access kernel memory. */
   3331 	if ((ptes[arm_btop(va)] & L2_TYPE_MASK) != L2_TYPE_INV)
   3332 		goto out;
   3333 
   3334 	/* Extract the physical address of the page */
   3335 	pa = l2pte_pa(ptes[arm_btop(va)]);
   3336 	if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
   3337 		goto out;
   3338 
   3339 	simple_lock(&pg->mdpage.pvh_slock);
   3340 
   3341 	/*
   3342 	 * Ok we just enable the pte and mark the attibs as handled
   3343 	 * XXX Should we traverse the PV list and enable all PTEs?
   3344 	 */
   3345 	PDEBUG(0,
   3346 	    printf("pmap_handled_emulation: Got a hit va=%08lx pte = %08x\n",
   3347 	    va, ptes[arm_btop(va)]));
   3348 	pg->mdpage.pvh_attrs |= PVF_REF;
   3349 
   3350 	ptes[arm_btop(va)] = (ptes[arm_btop(va)] & ~L2_TYPE_MASK) | L2_S_PROTO;
   3351 	PDEBUG(0, printf("->(%08x)\n", ptes[arm_btop(va)]));
   3352 
   3353 	simple_unlock(&pg->mdpage.pvh_slock);
   3354 
   3355 	cpu_tlb_flushID_SE(va);
   3356 	cpu_cpwait();
   3357 	rv = 1;
   3358  out:
   3359 	pmap_unmap_ptes(pmap);			/* unlocks pmap */
   3360 	PMAP_MAP_TO_HEAD_UNLOCK();
   3361 	return (rv);
   3362 }
   3363 
   3364 /*
   3365  * pmap_collect: free resources held by a pmap
   3366  *
   3367  * => optional function.
   3368  * => called when a process is swapped out to free memory.
   3369  */
   3370 
   3371 void
   3372 pmap_collect(struct pmap *pmap)
   3373 {
   3374 }
   3375 
   3376 /*
   3377  * Routine:	pmap_procwr
   3378  *
   3379  * Function:
   3380  *	Synchronize caches corresponding to [addr, addr+len) in p.
   3381  *
   3382  */
   3383 void
   3384 pmap_procwr(struct proc *p, vaddr_t va, int len)
   3385 {
   3386 	/* We only need to do anything if it is the current process. */
   3387 	if (p == curproc)
   3388 		cpu_icache_sync_range(va, len);
   3389 }
   3390 /*
   3391  * PTP functions
   3392  */
   3393 
   3394 /*
   3395  * pmap_get_ptp: get a PTP (if there isn't one, allocate a new one)
   3396  *
   3397  * => pmap should NOT be pmap_kernel()
   3398  * => pmap should be locked
   3399  */
   3400 
   3401 static struct vm_page *
   3402 pmap_get_ptp(struct pmap *pmap, vaddr_t va)
   3403 {
   3404 	struct vm_page *ptp;
   3405 
   3406 	if (pmap_pde_page(pmap_pde(pmap, va))) {
   3407 
   3408 		/* valid... check hint (saves us a PA->PG lookup) */
   3409 		if (pmap->pm_ptphint &&
   3410 		    (pmap->pm_pdir[pmap_pdei(va)] & L2_S_FRAME) ==
   3411 		    VM_PAGE_TO_PHYS(pmap->pm_ptphint))
   3412 			return (pmap->pm_ptphint);
   3413 		ptp = uvm_pagelookup(&pmap->pm_obj, va);
   3414 #ifdef DIAGNOSTIC
   3415 		if (ptp == NULL)
   3416 			panic("pmap_get_ptp: unmanaged user PTP");
   3417 #endif
   3418 		pmap->pm_ptphint = ptp;
   3419 		return(ptp);
   3420 	}
   3421 
   3422 	/* allocate a new PTP (updates ptphint) */
   3423 	return(pmap_alloc_ptp(pmap, va));
   3424 }
   3425 
   3426 /*
   3427  * pmap_alloc_ptp: allocate a PTP for a PMAP
   3428  *
   3429  * => pmap should already be locked by caller
   3430  * => we use the ptp's wire_count to count the number of active mappings
   3431  *	in the PTP (we start it at one to prevent any chance this PTP
   3432  *	will ever leak onto the active/inactive queues)
   3433  */
   3434 
   3435 /*__inline */ static struct vm_page *
   3436 pmap_alloc_ptp(struct pmap *pmap, vaddr_t va)
   3437 {
   3438 	struct vm_page *ptp;
   3439 
   3440 	ptp = uvm_pagealloc(&pmap->pm_obj, va, NULL,
   3441 		UVM_PGA_USERESERVE|UVM_PGA_ZERO);
   3442 	if (ptp == NULL)
   3443 		return (NULL);
   3444 
   3445 	/* got one! */
   3446 	ptp->flags &= ~PG_BUSY;	/* never busy */
   3447 	ptp->wire_count = 1;	/* no mappings yet */
   3448 	pmap_map_in_l1(pmap, va, VM_PAGE_TO_PHYS(ptp), TRUE);
   3449 	pmap->pm_stats.resident_count++;	/* count PTP as resident */
   3450 	pmap->pm_ptphint = ptp;
   3451 	return (ptp);
   3452 }
   3453 
   3454 vaddr_t
   3455 pmap_growkernel(vaddr_t maxkvaddr)
   3456 {
   3457 	struct pmap *kpm = pmap_kernel(), *pm;
   3458 	int s;
   3459 	paddr_t ptaddr;
   3460 	struct vm_page *ptp;
   3461 
   3462 	if (maxkvaddr <= pmap_curmaxkvaddr)
   3463 		goto out;		/* we are OK */
   3464 	NPDEBUG(PDB_GROWKERN, printf("pmap_growkernel: growing kernel from %lx to %lx\n",
   3465 		    pmap_curmaxkvaddr, maxkvaddr));
   3466 
   3467 	/*
   3468 	 * whoops!   we need to add kernel PTPs
   3469 	 */
   3470 
   3471 	s = splhigh();	/* to be safe */
   3472 	simple_lock(&kpm->pm_obj.vmobjlock);
   3473 	/* due to the way the arm pmap works we map 4MB at a time */
   3474 	for (/*null*/ ; pmap_curmaxkvaddr < maxkvaddr;
   3475 	     pmap_curmaxkvaddr += 4 * L1_S_SIZE) {
   3476 
   3477 		if (uvm.page_init_done == FALSE) {
   3478 
   3479 			/*
   3480 			 * we're growing the kernel pmap early (from
   3481 			 * uvm_pageboot_alloc()).  this case must be
   3482 			 * handled a little differently.
   3483 			 */
   3484 
   3485 			if (uvm_page_physget(&ptaddr) == FALSE)
   3486 				panic("pmap_growkernel: out of memory");
   3487 			pmap_zero_page(ptaddr);
   3488 
   3489 			/* map this page in */
   3490 			pmap_map_in_l1(kpm, pmap_curmaxkvaddr, ptaddr, TRUE);
   3491 
   3492 			/* count PTP as resident */
   3493 			kpm->pm_stats.resident_count++;
   3494 			continue;
   3495 		}
   3496 
   3497 		/*
   3498 		 * THIS *MUST* BE CODED SO AS TO WORK IN THE
   3499 		 * pmap_initialized == FALSE CASE!  WE MAY BE
   3500 		 * INVOKED WHILE pmap_init() IS RUNNING!
   3501 		 */
   3502 
   3503 		if ((ptp = pmap_alloc_ptp(kpm, pmap_curmaxkvaddr)) == NULL)
   3504 			panic("pmap_growkernel: alloc ptp failed");
   3505 
   3506 		/* distribute new kernel PTP to all active pmaps */
   3507 		simple_lock(&pmaps_lock);
   3508 		LIST_FOREACH(pm, &pmaps, pm_list) {
   3509 			pmap_map_in_l1(pm, pmap_curmaxkvaddr,
   3510 			    VM_PAGE_TO_PHYS(ptp), TRUE);
   3511 		}
   3512 
   3513 		simple_unlock(&pmaps_lock);
   3514 	}
   3515 
   3516 	/*
   3517 	 * flush out the cache, expensive but growkernel will happen so
   3518 	 * rarely
   3519 	 */
   3520 	cpu_tlb_flushD();
   3521 	cpu_cpwait();
   3522 
   3523 	simple_unlock(&kpm->pm_obj.vmobjlock);
   3524 	splx(s);
   3525 
   3526 out:
   3527 	return (pmap_curmaxkvaddr);
   3528 }
   3529 
   3530 /************************ Utility routines ****************************/
   3531 
   3532 /*
   3533  * vector_page_setprot:
   3534  *
   3535  *	Manipulate the protection of the vector page.
   3536  */
   3537 void
   3538 vector_page_setprot(int prot)
   3539 {
   3540 	pt_entry_t *pte;
   3541 
   3542 	pte = vtopte(vector_page);
   3543 
   3544 	*pte = (*pte & ~L1_S_PROT_MASK) | L2_S_PROT(PTE_KERNEL, prot);
   3545 	cpu_tlb_flushD_SE(vector_page);
   3546 	cpu_cpwait();
   3547 }
   3548 
   3549 /************************ Bootstrapping routines ****************************/
   3550 
   3551 /*
   3552  * This list exists for the benefit of pmap_map_chunk().  It keeps track
   3553  * of the kernel L2 tables during bootstrap, so that pmap_map_chunk() can
   3554  * find them as necessary.
   3555  *
   3556  * Note that the data on this list is not valid after initarm() returns.
   3557  */
   3558 SLIST_HEAD(, pv_addr) kernel_pt_list = SLIST_HEAD_INITIALIZER(kernel_pt_list);
   3559 
   3560 static vaddr_t
   3561 kernel_pt_lookup(paddr_t pa)
   3562 {
   3563 	pv_addr_t *pv;
   3564 
   3565 	SLIST_FOREACH(pv, &kernel_pt_list, pv_list) {
   3566 		if (pv->pv_pa == pa)
   3567 			return (pv->pv_va);
   3568 	}
   3569 	return (0);
   3570 }
   3571 
   3572 /*
   3573  * pmap_map_section:
   3574  *
   3575  *	Create a single section mapping.
   3576  */
   3577 void
   3578 pmap_map_section(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
   3579 {
   3580 	pd_entry_t *pde = (pd_entry_t *) l1pt;
   3581 	pd_entry_t fl = (cache == PTE_CACHE) ? pte_l1_s_cache_mode : 0;
   3582 
   3583 	KASSERT(((va | pa) & L1_S_OFFSET) == 0);
   3584 
   3585 	pde[va >> L1_S_SHIFT] = L1_S_PROTO | pa |
   3586 	    L1_S_PROT(PTE_KERNEL, prot) | fl;
   3587 }
   3588 
   3589 /*
   3590  * pmap_map_entry:
   3591  *
   3592  *	Create a single page mapping.
   3593  */
   3594 void
   3595 pmap_map_entry(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
   3596 {
   3597 	pd_entry_t *pde = (pd_entry_t *) l1pt;
   3598 	pt_entry_t fl = (cache == PTE_CACHE) ? pte_l2_s_cache_mode : 0;
   3599 	pt_entry_t *pte;
   3600 
   3601 	KASSERT(((va | pa) & PGOFSET) == 0);
   3602 
   3603 	if ((pde[va >> L1_S_SHIFT] & L1_TYPE_MASK) != L1_TYPE_C)
   3604 		panic("pmap_map_entry: no L2 table for VA 0x%08lx", va);
   3605 
   3606 	pte = (pt_entry_t *)
   3607 	    kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
   3608 	if (pte == NULL)
   3609 		panic("pmap_map_entry: can't find L2 table for VA 0x%08lx", va);
   3610 
   3611 	pte[(va >> PGSHIFT) & 0x3ff] = L2_S_PROTO | pa |
   3612 	    L2_S_PROT(PTE_KERNEL, prot) | fl;
   3613 }
   3614 
   3615 /*
   3616  * pmap_link_l2pt:
   3617  *
   3618  *	Link the L2 page table specified by "pa" into the L1
   3619  *	page table at the slot for "va".
   3620  */
   3621 void
   3622 pmap_link_l2pt(vaddr_t l1pt, vaddr_t va, pv_addr_t *l2pv)
   3623 {
   3624 	pd_entry_t *pde = (pd_entry_t *) l1pt;
   3625 	u_int slot = va >> L1_S_SHIFT;
   3626 
   3627 	KASSERT((l2pv->pv_pa & PGOFSET) == 0);
   3628 
   3629 	pde[slot + 0] = L1_C_PROTO | (l2pv->pv_pa + 0x000);
   3630 	pde[slot + 1] = L1_C_PROTO | (l2pv->pv_pa + 0x400);
   3631 	pde[slot + 2] = L1_C_PROTO | (l2pv->pv_pa + 0x800);
   3632 	pde[slot + 3] = L1_C_PROTO | (l2pv->pv_pa + 0xc00);
   3633 
   3634 	SLIST_INSERT_HEAD(&kernel_pt_list, l2pv, pv_list);
   3635 }
   3636 
   3637 /*
   3638  * pmap_map_chunk:
   3639  *
   3640  *	Map a chunk of memory using the most efficient mappings
   3641  *	possible (section, large page, small page) into the
   3642  *	provided L1 and L2 tables at the specified virtual address.
   3643  */
   3644 vsize_t
   3645 pmap_map_chunk(vaddr_t l1pt, vaddr_t va, paddr_t pa, vsize_t size,
   3646     int prot, int cache)
   3647 {
   3648 	pd_entry_t *pde = (pd_entry_t *) l1pt;
   3649 	pt_entry_t *pte, fl;
   3650 	vsize_t resid;
   3651 	int i;
   3652 
   3653 	resid = (size + (NBPG - 1)) & ~(NBPG - 1);
   3654 
   3655 	if (l1pt == 0)
   3656 		panic("pmap_map_chunk: no L1 table provided");
   3657 
   3658 #ifdef VERBOSE_INIT_ARM
   3659 	printf("pmap_map_chunk: pa=0x%lx va=0x%lx size=0x%lx resid=0x%lx "
   3660 	    "prot=0x%x cache=%d\n", pa, va, size, resid, prot, cache);
   3661 #endif
   3662 
   3663 	size = resid;
   3664 
   3665 	while (resid > 0) {
   3666 		/* See if we can use a section mapping. */
   3667 		if (((pa | va) & L1_S_OFFSET) == 0 &&
   3668 		    resid >= L1_S_SIZE) {
   3669 			fl = (cache == PTE_CACHE) ? pte_l1_s_cache_mode : 0;
   3670 #ifdef VERBOSE_INIT_ARM
   3671 			printf("S");
   3672 #endif
   3673 			pde[va >> L1_S_SHIFT] = L1_S_PROTO | pa |
   3674 			    L1_S_PROT(PTE_KERNEL, prot) | fl;
   3675 			va += L1_S_SIZE;
   3676 			pa += L1_S_SIZE;
   3677 			resid -= L1_S_SIZE;
   3678 			continue;
   3679 		}
   3680 
   3681 		/*
   3682 		 * Ok, we're going to use an L2 table.  Make sure
   3683 		 * one is actually in the corresponding L1 slot
   3684 		 * for the current VA.
   3685 		 */
   3686 		if ((pde[va >> L1_S_SHIFT] & L1_TYPE_MASK) != L1_TYPE_C)
   3687 			panic("pmap_map_chunk: no L2 table for VA 0x%08lx", va);
   3688 
   3689 		pte = (pt_entry_t *)
   3690 		    kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
   3691 		if (pte == NULL)
   3692 			panic("pmap_map_chunk: can't find L2 table for VA"
   3693 			    "0x%08lx", va);
   3694 
   3695 		/* See if we can use a L2 large page mapping. */
   3696 		if (((pa | va) & L2_L_OFFSET) == 0 &&
   3697 		    resid >= L2_L_SIZE) {
   3698 			fl = (cache == PTE_CACHE) ? pte_l2_l_cache_mode : 0;
   3699 #ifdef VERBOSE_INIT_ARM
   3700 			printf("L");
   3701 #endif
   3702 			for (i = 0; i < 16; i++) {
   3703 				pte[((va >> PGSHIFT) & 0x3f0) + i] =
   3704 				    L2_L_PROTO | pa |
   3705 				    L2_L_PROT(PTE_KERNEL, prot) | fl;
   3706 			}
   3707 			va += L2_L_SIZE;
   3708 			pa += L2_L_SIZE;
   3709 			resid -= L2_L_SIZE;
   3710 			continue;
   3711 		}
   3712 
   3713 		/* Use a small page mapping. */
   3714 		fl = (cache == PTE_CACHE) ? pte_l2_s_cache_mode : 0;
   3715 #ifdef VERBOSE_INIT_ARM
   3716 		printf("P");
   3717 #endif
   3718 		pte[(va >> PGSHIFT) & 0x3ff] = L2_S_PROTO | pa |
   3719 		    L2_S_PROT(PTE_KERNEL, prot) | fl;
   3720 		va += NBPG;
   3721 		pa += NBPG;
   3722 		resid -= NBPG;
   3723 	}
   3724 #ifdef VERBOSE_INIT_ARM
   3725 	printf("\n");
   3726 #endif
   3727 	return (size);
   3728 }
   3729 
   3730 /********************** PTE initialization routines **************************/
   3731 
   3732 /*
   3733  * These routines are called when the CPU type is identified to set up
   3734  * the PTE prototypes, cache modes, etc.
   3735  *
   3736  * The variables are always here, just in case LKMs need to reference
   3737  * them (though, they shouldn't).
   3738  */
   3739 
   3740 pt_entry_t	pte_l1_s_cache_mode;
   3741 pt_entry_t	pte_l1_s_cache_mask;
   3742 
   3743 pt_entry_t	pte_l2_l_cache_mode;
   3744 pt_entry_t	pte_l2_l_cache_mask;
   3745 
   3746 pt_entry_t	pte_l2_s_cache_mode;
   3747 pt_entry_t	pte_l2_s_cache_mask;
   3748 
   3749 pt_entry_t	pte_l2_s_prot_u;
   3750 pt_entry_t	pte_l2_s_prot_w;
   3751 pt_entry_t	pte_l2_s_prot_mask;
   3752 
   3753 pt_entry_t	pte_l1_s_proto;
   3754 pt_entry_t	pte_l1_c_proto;
   3755 pt_entry_t	pte_l2_s_proto;
   3756 
   3757 void		(*pmap_copy_page_func)(paddr_t, paddr_t);
   3758 void		(*pmap_zero_page_func)(paddr_t);
   3759 
   3760 #if ARM_MMU_GENERIC == 1
   3761 void
   3762 pmap_pte_init_generic(void)
   3763 {
   3764 
   3765 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
   3766 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_generic;
   3767 
   3768 	pte_l2_l_cache_mode = L2_B|L2_C;
   3769 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_generic;
   3770 
   3771 	pte_l2_s_cache_mode = L2_B|L2_C;
   3772 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_generic;
   3773 
   3774 	pte_l2_s_prot_u = L2_S_PROT_U_generic;
   3775 	pte_l2_s_prot_w = L2_S_PROT_W_generic;
   3776 	pte_l2_s_prot_mask = L2_S_PROT_MASK_generic;
   3777 
   3778 	pte_l1_s_proto = L1_S_PROTO_generic;
   3779 	pte_l1_c_proto = L1_C_PROTO_generic;
   3780 	pte_l2_s_proto = L2_S_PROTO_generic;
   3781 
   3782 	pmap_copy_page_func = pmap_copy_page_generic;
   3783 	pmap_zero_page_func = pmap_zero_page_generic;
   3784 }
   3785 
   3786 #if defined(CPU_ARM9)
   3787 void
   3788 pmap_pte_init_arm9(void)
   3789 {
   3790 
   3791 	/*
   3792 	 * ARM9 is compatible with generic, but we want to use
   3793 	 * write-through caching for now.
   3794 	 */
   3795 	pmap_pte_init_generic();
   3796 
   3797 	pte_l1_s_cache_mode = L1_S_C;
   3798 	pte_l2_l_cache_mode = L2_C;
   3799 	pte_l2_s_cache_mode = L2_C;
   3800 }
   3801 #endif /* CPU_ARM9 */
   3802 #endif /* ARM_MMU_GENERIC == 1 */
   3803 
   3804 #if ARM_MMU_XSCALE == 1
   3805 void
   3806 pmap_pte_init_xscale(void)
   3807 {
   3808 	uint32_t auxctl;
   3809 
   3810 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
   3811 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_xscale;
   3812 
   3813 	pte_l2_l_cache_mode = L2_B|L2_C;
   3814 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_xscale;
   3815 
   3816 	pte_l2_s_cache_mode = L2_B|L2_C;
   3817 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_xscale;
   3818 
   3819 #ifdef XSCALE_CACHE_WRITE_THROUGH
   3820 	/*
   3821 	 * Some versions of the XScale core have various bugs in
   3822 	 * their cache units, the work-around for which is to run
   3823 	 * the cache in write-through mode.  Unfortunately, this
   3824 	 * has a major (negative) impact on performance.  So, we
   3825 	 * go ahead and run fast-and-loose, in the hopes that we
   3826 	 * don't line up the planets in a way that will trip the
   3827 	 * bugs.
   3828 	 *
   3829 	 * However, we give you the option to be slow-but-correct.
   3830 	 */
   3831 	pte_l1_s_cache_mode = L1_S_C;
   3832 	pte_l2_l_cache_mode = L2_C;
   3833 	pte_l2_s_cache_mode = L2_C;
   3834 #endif /* XSCALE_CACHE_WRITE_THROUGH */
   3835 
   3836 	pte_l2_s_prot_u = L2_S_PROT_U_xscale;
   3837 	pte_l2_s_prot_w = L2_S_PROT_W_xscale;
   3838 	pte_l2_s_prot_mask = L2_S_PROT_MASK_xscale;
   3839 
   3840 	pte_l1_s_proto = L1_S_PROTO_xscale;
   3841 	pte_l1_c_proto = L1_C_PROTO_xscale;
   3842 	pte_l2_s_proto = L2_S_PROTO_xscale;
   3843 
   3844 	pmap_copy_page_func = pmap_copy_page_xscale;
   3845 	pmap_zero_page_func = pmap_zero_page_xscale;
   3846 
   3847 	/*
   3848 	 * Disable ECC protection of page table access, for now.
   3849 	 */
   3850 	__asm __volatile("mrc p15, 0, %0, c1, c0, 1"
   3851 		: "=r" (auxctl));
   3852 	auxctl &= ~XSCALE_AUXCTL_P;
   3853 	__asm __volatile("mcr p15, 0, %0, c1, c0, 1"
   3854 		:
   3855 		: "r" (auxctl));
   3856 }
   3857 
   3858 /*
   3859  * xscale_setup_minidata:
   3860  *
   3861  *	Set up the mini-data cache clean area.  We require the
   3862  *	caller to allocate the right amount of physically and
   3863  *	virtually contiguous space.
   3864  */
   3865 void
   3866 xscale_setup_minidata(vaddr_t l1pt, vaddr_t va, paddr_t pa)
   3867 {
   3868 	extern vaddr_t xscale_minidata_clean_addr;
   3869 	extern vsize_t xscale_minidata_clean_size; /* already initialized */
   3870 	pd_entry_t *pde = (pd_entry_t *) l1pt;
   3871 	pt_entry_t *pte;
   3872 	vsize_t size;
   3873 	uint32_t auxctl;
   3874 
   3875 	xscale_minidata_clean_addr = va;
   3876 
   3877 	/* Round it to page size. */
   3878 	size = (xscale_minidata_clean_size + L2_S_OFFSET) & L2_S_FRAME;
   3879 
   3880 	for (; size != 0;
   3881 	     va += L2_S_SIZE, pa += L2_S_SIZE, size -= L2_S_SIZE) {
   3882 		pte = (pt_entry_t *)
   3883 		    kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
   3884 		if (pte == NULL)
   3885 			panic("xscale_setup_minidata: can't find L2 table for "
   3886 			    "VA 0x%08lx", va);
   3887 		pte[(va >> PGSHIFT) & 0x3ff] = L2_S_PROTO | pa |
   3888 		    L2_S_PROT(PTE_KERNEL, VM_PROT_READ) |
   3889 		    L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X);
   3890 	}
   3891 
   3892 	/*
   3893 	 * Configure the mini-data cache for write-back with
   3894 	 * read/write-allocate.
   3895 	 *
   3896 	 * NOTE: In order to reconfigure the mini-data cache, we must
   3897 	 * make sure it contains no valid data!  In order to do that,
   3898 	 * we must issue a global data cache invalidate command!
   3899 	 *
   3900 	 * WE ASSUME WE ARE RUNNING UN-CACHED WHEN THIS ROUTINE IS CALLED!
   3901 	 * THIS IS VERY IMPORTANT!
   3902 	 */
   3903 
   3904 	/* Invalidate data and mini-data. */
   3905 	__asm __volatile("mcr p15, 0, %0, c7, c6, 0"
   3906 		:
   3907 		: "r" (auxctl));
   3908 
   3909 
   3910 	__asm __volatile("mrc p15, 0, %0, c1, c0, 1"
   3911 		: "=r" (auxctl));
   3912 	auxctl = (auxctl & ~XSCALE_AUXCTL_MD_MASK) | XSCALE_AUXCTL_MD_WB_RWA;
   3913 	__asm __volatile("mcr p15, 0, %0, c1, c0, 1"
   3914 		:
   3915 		: "r" (auxctl));
   3916 }
   3917 #endif /* ARM_MMU_XSCALE == 1 */
   3918