Home | History | Annotate | Line # | Download | only in arm32
pmap.c revision 1.166
      1 /*	$NetBSD: pmap.c,v 1.166 2007/10/10 17:44:39 ad Exp $	*/
      2 
      3 /*
      4  * Copyright 2003 Wasabi Systems, Inc.
      5  * All rights reserved.
      6  *
      7  * Written by Steve C. Woodford for Wasabi Systems, Inc.
      8  *
      9  * Redistribution and use in source and binary forms, with or without
     10  * modification, are permitted provided that the following conditions
     11  * are met:
     12  * 1. Redistributions of source code must retain the above copyright
     13  *    notice, this list of conditions and the following disclaimer.
     14  * 2. Redistributions in binary form must reproduce the above copyright
     15  *    notice, this list of conditions and the following disclaimer in the
     16  *    documentation and/or other materials provided with the distribution.
     17  * 3. All advertising materials mentioning features or use of this software
     18  *    must display the following acknowledgement:
     19  *      This product includes software developed for the NetBSD Project by
     20  *      Wasabi Systems, Inc.
     21  * 4. The name of Wasabi Systems, Inc. may not be used to endorse
     22  *    or promote products derived from this software without specific prior
     23  *    written permission.
     24  *
     25  * THIS SOFTWARE IS PROVIDED BY WASABI SYSTEMS, INC. ``AS IS'' AND
     26  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     27  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     28  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL WASABI SYSTEMS, INC
     29  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     30  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     31  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     32  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     33  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     34  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     35  * POSSIBILITY OF SUCH DAMAGE.
     36  */
     37 
     38 /*
     39  * Copyright (c) 2002-2003 Wasabi Systems, Inc.
     40  * Copyright (c) 2001 Richard Earnshaw
     41  * Copyright (c) 2001-2002 Christopher Gilbert
     42  * All rights reserved.
     43  *
     44  * 1. Redistributions of source code must retain the above copyright
     45  *    notice, this list of conditions and the following disclaimer.
     46  * 2. Redistributions in binary form must reproduce the above copyright
     47  *    notice, this list of conditions and the following disclaimer in the
     48  *    documentation and/or other materials provided with the distribution.
     49  * 3. The name of the company nor the name of the author may be used to
     50  *    endorse or promote products derived from this software without specific
     51  *    prior written permission.
     52  *
     53  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
     54  * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
     55  * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
     56  * IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
     57  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
     58  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
     59  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     60  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     61  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     62  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     63  * SUCH DAMAGE.
     64  */
     65 
     66 /*-
     67  * Copyright (c) 1999 The NetBSD Foundation, Inc.
     68  * All rights reserved.
     69  *
     70  * This code is derived from software contributed to The NetBSD Foundation
     71  * by Charles M. Hannum.
     72  *
     73  * Redistribution and use in source and binary forms, with or without
     74  * modification, are permitted provided that the following conditions
     75  * are met:
     76  * 1. Redistributions of source code must retain the above copyright
     77  *    notice, this list of conditions and the following disclaimer.
     78  * 2. Redistributions in binary form must reproduce the above copyright
     79  *    notice, this list of conditions and the following disclaimer in the
     80  *    documentation and/or other materials provided with the distribution.
     81  * 3. All advertising materials mentioning features or use of this software
     82  *    must display the following acknowledgement:
     83  *        This product includes software developed by the NetBSD
     84  *        Foundation, Inc. and its contributors.
     85  * 4. Neither the name of The NetBSD Foundation nor the names of its
     86  *    contributors may be used to endorse or promote products derived
     87  *    from this software without specific prior written permission.
     88  *
     89  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     90  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     91  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     92  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     93  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     94  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     95  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     96  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     97  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     98  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     99  * POSSIBILITY OF SUCH DAMAGE.
    100  */
    101 
    102 /*
    103  * Copyright (c) 1994-1998 Mark Brinicombe.
    104  * Copyright (c) 1994 Brini.
    105  * All rights reserved.
    106  *
    107  * This code is derived from software written for Brini by Mark Brinicombe
    108  *
    109  * Redistribution and use in source and binary forms, with or without
    110  * modification, are permitted provided that the following conditions
    111  * are met:
    112  * 1. Redistributions of source code must retain the above copyright
    113  *    notice, this list of conditions and the following disclaimer.
    114  * 2. Redistributions in binary form must reproduce the above copyright
    115  *    notice, this list of conditions and the following disclaimer in the
    116  *    documentation and/or other materials provided with the distribution.
    117  * 3. All advertising materials mentioning features or use of this software
    118  *    must display the following acknowledgement:
    119  *	This product includes software developed by Mark Brinicombe.
    120  * 4. The name of the author may not be used to endorse or promote products
    121  *    derived from this software without specific prior written permission.
    122  *
    123  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
    124  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
    125  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
    126  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
    127  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
    128  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
    129  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
    130  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
    131  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
    132  *
    133  * RiscBSD kernel project
    134  *
    135  * pmap.c
    136  *
    137  * Machine dependant vm stuff
    138  *
    139  * Created      : 20/09/94
    140  */
    141 
    142 /*
    143  * Performance improvements, UVM changes, overhauls and part-rewrites
    144  * were contributed by Neil A. Carson <neil (at) causality.com>.
    145  */
    146 
    147 /*
    148  * Overhauled again to speedup the pmap, use MMU Domains so that L1 tables
    149  * can be shared, and re-work the KVM layout, by Steve Woodford of Wasabi
    150  * Systems, Inc.
    151  *
    152  * There are still a few things outstanding at this time:
    153  *
    154  *   - There are some unresolved issues for MP systems:
    155  *
    156  *     o The L1 metadata needs a lock, or more specifically, some places
    157  *       need to acquire an exclusive lock when modifying L1 translation
    158  *       table entries.
    159  *
    160  *     o When one cpu modifies an L1 entry, and that L1 table is also
    161  *       being used by another cpu, then the latter will need to be told
    162  *       that a tlb invalidation may be necessary. (But only if the old
    163  *       domain number in the L1 entry being over-written is currently
    164  *       the active domain on that cpu). I guess there are lots more tlb
    165  *       shootdown issues too...
    166  *
    167  *     o If the vector_page is at 0x00000000 instead of 0xffff0000, then
    168  *       MP systems will lose big-time because of the MMU domain hack.
    169  *       The only way this can be solved (apart from moving the vector
    170  *       page to 0xffff0000) is to reserve the first 1MB of user address
    171  *       space for kernel use only. This would require re-linking all
    172  *       applications so that the text section starts above this 1MB
    173  *       boundary.
    174  *
    175  *     o Tracking which VM space is resident in the cache/tlb has not yet
    176  *       been implemented for MP systems.
    177  *
    178  *     o Finally, there is a pathological condition where two cpus running
    179  *       two separate processes (not lwps) which happen to share an L1
    180  *       can get into a fight over one or more L1 entries. This will result
    181  *       in a significant slow-down if both processes are in tight loops.
    182  */
    183 
    184 /*
    185  * Special compilation symbols
    186  * PMAP_DEBUG		- Build in pmap_debug_level code
    187  */
    188 
    189 /* Include header files */
    190 
    191 #include "opt_cpuoptions.h"
    192 #include "opt_pmap_debug.h"
    193 #include "opt_ddb.h"
    194 #include "opt_lockdebug.h"
    195 #include "opt_multiprocessor.h"
    196 
    197 #include <sys/types.h>
    198 #include <sys/param.h>
    199 #include <sys/kernel.h>
    200 #include <sys/systm.h>
    201 #include <sys/proc.h>
    202 #include <sys/malloc.h>
    203 #include <sys/user.h>
    204 #include <sys/pool.h>
    205 #include <sys/cdefs.h>
    206 
    207 #include <uvm/uvm.h>
    208 
    209 #include <machine/bus.h>
    210 #include <machine/pmap.h>
    211 #include <machine/pcb.h>
    212 #include <machine/param.h>
    213 #include <arm/arm32/katelib.h>
    214 
    215 __KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.166 2007/10/10 17:44:39 ad Exp $");
    216 
    217 #ifdef PMAP_DEBUG
    218 
    219 /* XXX need to get rid of all refs to this */
    220 int pmap_debug_level = 0;
    221 
    222 /*
    223  * for switching to potentially finer grained debugging
    224  */
    225 #define	PDB_FOLLOW	0x0001
    226 #define	PDB_INIT	0x0002
    227 #define	PDB_ENTER	0x0004
    228 #define	PDB_REMOVE	0x0008
    229 #define	PDB_CREATE	0x0010
    230 #define	PDB_PTPAGE	0x0020
    231 #define	PDB_GROWKERN	0x0040
    232 #define	PDB_BITS	0x0080
    233 #define	PDB_COLLECT	0x0100
    234 #define	PDB_PROTECT	0x0200
    235 #define	PDB_MAP_L1	0x0400
    236 #define	PDB_BOOTSTRAP	0x1000
    237 #define	PDB_PARANOIA	0x2000
    238 #define	PDB_WIRING	0x4000
    239 #define	PDB_PVDUMP	0x8000
    240 #define	PDB_VAC		0x10000
    241 #define	PDB_KENTER	0x20000
    242 #define	PDB_KREMOVE	0x40000
    243 
    244 int debugmap = 1;
    245 int pmapdebug = 0;
    246 #define	NPDEBUG(_lev_,_stat_) \
    247 	if (pmapdebug & (_lev_)) \
    248         	((_stat_))
    249 
    250 #else	/* PMAP_DEBUG */
    251 #define NPDEBUG(_lev_,_stat_) /* Nothing */
    252 #endif	/* PMAP_DEBUG */
    253 
    254 /*
    255  * pmap_kernel() points here
    256  */
    257 struct pmap     kernel_pmap_store;
    258 
    259 /*
    260  * Which pmap is currently 'live' in the cache
    261  *
    262  * XXXSCW: Fix for SMP ...
    263  */
    264 static pmap_t pmap_recent_user;
    265 
    266 /*
    267  * Pool and cache that pmap structures are allocated from.
    268  * We use a cache to avoid clearing the pm_l2[] array (1KB)
    269  * in pmap_create().
    270  */
    271 static struct pool pmap_pmap_pool;
    272 static struct pool_cache pmap_pmap_cache;
    273 static LIST_HEAD(, pmap) pmap_pmaps;
    274 
    275 /*
    276  * Pool of PV structures
    277  */
    278 static struct pool pmap_pv_pool;
    279 static void *pmap_bootstrap_pv_page_alloc(struct pool *, int);
    280 static void pmap_bootstrap_pv_page_free(struct pool *, void *);
    281 static struct pool_allocator pmap_bootstrap_pv_allocator = {
    282 	pmap_bootstrap_pv_page_alloc, pmap_bootstrap_pv_page_free
    283 };
    284 
    285 /*
    286  * Pool and cache of l2_dtable structures.
    287  * We use a cache to avoid clearing the structures when they're
    288  * allocated. (196 bytes)
    289  */
    290 static struct pool pmap_l2dtable_pool;
    291 static struct pool_cache pmap_l2dtable_cache;
    292 static vaddr_t pmap_kernel_l2dtable_kva;
    293 
    294 /*
    295  * Pool and cache of L2 page descriptors.
    296  * We use a cache to avoid clearing the descriptor table
    297  * when they're allocated. (1KB)
    298  */
    299 static struct pool pmap_l2ptp_pool;
    300 static struct pool_cache pmap_l2ptp_cache;
    301 static vaddr_t pmap_kernel_l2ptp_kva;
    302 static paddr_t pmap_kernel_l2ptp_phys;
    303 
    304 /*
    305  * pmap copy/zero page, and mem(5) hook point
    306  */
    307 static pt_entry_t *csrc_pte, *cdst_pte;
    308 static vaddr_t csrcp, cdstp;
    309 char *memhook;
    310 extern void *msgbufaddr;
    311 
    312 /*
    313  * Flag to indicate if pmap_init() has done its thing
    314  */
    315 bool pmap_initialized;
    316 
    317 /*
    318  * Misc. locking data structures
    319  */
    320 
    321 #if 0 /* defined(MULTIPROCESSOR) || defined(LOCKDEBUG) */
    322 static struct lock pmap_main_lock;
    323 
    324 #define PMAP_MAP_TO_HEAD_LOCK() \
    325      (void) spinlockmgr(&pmap_main_lock, LK_SHARED, NULL)
    326 #define PMAP_MAP_TO_HEAD_UNLOCK() \
    327      (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL)
    328 #define PMAP_HEAD_TO_MAP_LOCK() \
    329      (void) spinlockmgr(&pmap_main_lock, LK_EXCLUSIVE, NULL)
    330 #define PMAP_HEAD_TO_MAP_UNLOCK() \
    331      spinlockmgr(&pmap_main_lock, LK_RELEASE, (void *) 0)
    332 #else
    333 #define PMAP_MAP_TO_HEAD_LOCK()		/* null */
    334 #define PMAP_MAP_TO_HEAD_UNLOCK()	/* null */
    335 #define PMAP_HEAD_TO_MAP_LOCK()		/* null */
    336 #define PMAP_HEAD_TO_MAP_UNLOCK()	/* null */
    337 #endif
    338 
    339 #define	pmap_acquire_pmap_lock(pm)			\
    340 	do {						\
    341 		if ((pm) != pmap_kernel())		\
    342 			simple_lock(&(pm)->pm_lock);	\
    343 	} while (/*CONSTCOND*/0)
    344 
    345 #define	pmap_release_pmap_lock(pm)			\
    346 	do {						\
    347 		if ((pm) != pmap_kernel())		\
    348 			simple_unlock(&(pm)->pm_lock);	\
    349 	} while (/*CONSTCOND*/0)
    350 
    351 
    352 /*
    353  * Metadata for L1 translation tables.
    354  */
    355 struct l1_ttable {
    356 	/* Entry on the L1 Table list */
    357 	SLIST_ENTRY(l1_ttable) l1_link;
    358 
    359 	/* Entry on the L1 Least Recently Used list */
    360 	TAILQ_ENTRY(l1_ttable) l1_lru;
    361 
    362 	/* Track how many domains are allocated from this L1 */
    363 	volatile u_int l1_domain_use_count;
    364 
    365 	/*
    366 	 * A free-list of domain numbers for this L1.
    367 	 * We avoid using ffs() and a bitmap to track domains since ffs()
    368 	 * is slow on ARM.
    369 	 */
    370 	u_int8_t l1_domain_first;
    371 	u_int8_t l1_domain_free[PMAP_DOMAINS];
    372 
    373 	/* Physical address of this L1 page table */
    374 	paddr_t l1_physaddr;
    375 
    376 	/* KVA of this L1 page table */
    377 	pd_entry_t *l1_kva;
    378 };
    379 
    380 /*
    381  * Convert a virtual address into its L1 table index. That is, the
    382  * index used to locate the L2 descriptor table pointer in an L1 table.
    383  * This is basically used to index l1->l1_kva[].
    384  *
    385  * Each L2 descriptor table represents 1MB of VA space.
    386  */
    387 #define	L1_IDX(va)		(((vaddr_t)(va)) >> L1_S_SHIFT)
    388 
    389 /*
    390  * L1 Page Tables are tracked using a Least Recently Used list.
    391  *  - New L1s are allocated from the HEAD.
    392  *  - Freed L1s are added to the TAIl.
    393  *  - Recently accessed L1s (where an 'access' is some change to one of
    394  *    the userland pmaps which owns this L1) are moved to the TAIL.
    395  */
    396 static TAILQ_HEAD(, l1_ttable) l1_lru_list;
    397 static struct simplelock l1_lru_lock;
    398 
    399 /*
    400  * A list of all L1 tables
    401  */
    402 static SLIST_HEAD(, l1_ttable) l1_list;
    403 
    404 /*
    405  * The l2_dtable tracks L2_BUCKET_SIZE worth of L1 slots.
    406  *
    407  * This is normally 16MB worth L2 page descriptors for any given pmap.
    408  * Reference counts are maintained for L2 descriptors so they can be
    409  * freed when empty.
    410  */
    411 struct l2_dtable {
    412 	/* The number of L2 page descriptors allocated to this l2_dtable */
    413 	u_int l2_occupancy;
    414 
    415 	/* List of L2 page descriptors */
    416 	struct l2_bucket {
    417 		pt_entry_t *l2b_kva;	/* KVA of L2 Descriptor Table */
    418 		paddr_t l2b_phys;	/* Physical address of same */
    419 		u_short l2b_l1idx;	/* This L2 table's L1 index */
    420 		u_short l2b_occupancy;	/* How many active descriptors */
    421 	} l2_bucket[L2_BUCKET_SIZE];
    422 };
    423 
    424 /*
    425  * Given an L1 table index, calculate the corresponding l2_dtable index
    426  * and bucket index within the l2_dtable.
    427  */
    428 #define	L2_IDX(l1idx)		(((l1idx) >> L2_BUCKET_LOG2) & \
    429 				 (L2_SIZE - 1))
    430 #define	L2_BUCKET(l1idx)	((l1idx) & (L2_BUCKET_SIZE - 1))
    431 
    432 /*
    433  * Given a virtual address, this macro returns the
    434  * virtual address required to drop into the next L2 bucket.
    435  */
    436 #define	L2_NEXT_BUCKET(va)	(((va) & L1_S_FRAME) + L1_S_SIZE)
    437 
    438 /*
    439  * L2 allocation.
    440  */
    441 #define	pmap_alloc_l2_dtable()		\
    442 	    pool_cache_get(&pmap_l2dtable_cache, PR_NOWAIT)
    443 #define	pmap_free_l2_dtable(l2)		\
    444 	    pool_cache_put(&pmap_l2dtable_cache, (l2))
    445 #define pmap_alloc_l2_ptp(pap)		\
    446 	    ((pt_entry_t *)pool_cache_get_paddr(&pmap_l2ptp_cache,\
    447 	    PR_NOWAIT, (pap)))
    448 
    449 /*
    450  * We try to map the page tables write-through, if possible.  However, not
    451  * all CPUs have a write-through cache mode, so on those we have to sync
    452  * the cache when we frob page tables.
    453  *
    454  * We try to evaluate this at compile time, if possible.  However, it's
    455  * not always possible to do that, hence this run-time var.
    456  */
    457 int	pmap_needs_pte_sync;
    458 
    459 /*
    460  * Real definition of pv_entry.
    461  */
    462 struct pv_entry {
    463 	struct pv_entry *pv_next;       /* next pv_entry */
    464 	pmap_t		pv_pmap;        /* pmap where mapping lies */
    465 	vaddr_t		pv_va;          /* virtual address for mapping */
    466 	u_int		pv_flags;       /* flags */
    467 };
    468 
    469 /*
    470  * Macro to determine if a mapping might be resident in the
    471  * instruction cache and/or TLB
    472  */
    473 #define	PV_BEEN_EXECD(f)  (((f) & (PVF_REF | PVF_EXEC)) == (PVF_REF | PVF_EXEC))
    474 
    475 /*
    476  * Macro to determine if a mapping might be resident in the
    477  * data cache and/or TLB
    478  */
    479 #define	PV_BEEN_REFD(f)   (((f) & PVF_REF) != 0)
    480 
    481 /*
    482  * Local prototypes
    483  */
    484 static int		pmap_set_pt_cache_mode(pd_entry_t *, vaddr_t);
    485 static void		pmap_alloc_specials(vaddr_t *, int, vaddr_t *,
    486 			    pt_entry_t **);
    487 static bool		pmap_is_current(pmap_t);
    488 static bool		pmap_is_cached(pmap_t);
    489 static void		pmap_enter_pv(struct vm_page *, struct pv_entry *,
    490 			    pmap_t, vaddr_t, u_int);
    491 static struct pv_entry *pmap_find_pv(struct vm_page *, pmap_t, vaddr_t);
    492 static struct pv_entry *pmap_remove_pv(struct vm_page *, pmap_t, vaddr_t, int);
    493 static u_int		pmap_modify_pv(struct vm_page *, pmap_t, vaddr_t,
    494 			    u_int, u_int);
    495 
    496 static void		pmap_pinit(pmap_t);
    497 static int		pmap_pmap_ctor(void *, void *, int);
    498 
    499 static void		pmap_alloc_l1(pmap_t);
    500 static void		pmap_free_l1(pmap_t);
    501 static void		pmap_use_l1(pmap_t);
    502 
    503 static struct l2_bucket *pmap_get_l2_bucket(pmap_t, vaddr_t);
    504 static struct l2_bucket *pmap_alloc_l2_bucket(pmap_t, vaddr_t);
    505 static void		pmap_free_l2_bucket(pmap_t, struct l2_bucket *, u_int);
    506 static int		pmap_l2ptp_ctor(void *, void *, int);
    507 static int		pmap_l2dtable_ctor(void *, void *, int);
    508 
    509 static void		pmap_vac_me_harder(struct vm_page *, pmap_t, vaddr_t);
    510 static void		pmap_vac_me_kpmap(struct vm_page *, pmap_t, vaddr_t);
    511 static void		pmap_vac_me_user(struct vm_page *, pmap_t, vaddr_t);
    512 
    513 static void		pmap_clearbit(struct vm_page *, u_int);
    514 static int		pmap_clean_page(struct pv_entry *, bool);
    515 static void		pmap_page_remove(struct vm_page *);
    516 
    517 static void		pmap_init_l1(struct l1_ttable *, pd_entry_t *);
    518 static vaddr_t		kernel_pt_lookup(paddr_t);
    519 
    520 void pmap_switch(struct lwp *, struct lwp *);
    521 
    522 
    523 /*
    524  * External function prototypes
    525  */
    526 extern void bzero_page(vaddr_t);
    527 extern void bcopy_page(vaddr_t, vaddr_t);
    528 
    529 /*
    530  * Misc variables
    531  */
    532 vaddr_t virtual_avail;
    533 vaddr_t virtual_end;
    534 vaddr_t pmap_curmaxkvaddr;
    535 
    536 vaddr_t avail_start;
    537 vaddr_t avail_end;
    538 
    539 extern pv_addr_t systempage;
    540 
    541 /* Function to set the debug level of the pmap code */
    542 
    543 #ifdef PMAP_DEBUG
    544 void
    545 pmap_debug(int level)
    546 {
    547 	pmap_debug_level = level;
    548 	printf("pmap_debug: level=%d\n", pmap_debug_level);
    549 }
    550 #endif	/* PMAP_DEBUG */
    551 
    552 /*
    553  * A bunch of routines to conditionally flush the caches/TLB depending
    554  * on whether the specified pmap actually needs to be flushed at any
    555  * given time.
    556  */
    557 static inline void
    558 pmap_tlb_flushID_SE(pmap_t pm, vaddr_t va)
    559 {
    560 
    561 	if (pm->pm_cstate.cs_tlb_id)
    562 		cpu_tlb_flushID_SE(va);
    563 }
    564 
    565 static inline void
    566 pmap_tlb_flushD_SE(pmap_t pm, vaddr_t va)
    567 {
    568 
    569 	if (pm->pm_cstate.cs_tlb_d)
    570 		cpu_tlb_flushD_SE(va);
    571 }
    572 
    573 static inline void
    574 pmap_tlb_flushID(pmap_t pm)
    575 {
    576 
    577 	if (pm->pm_cstate.cs_tlb_id) {
    578 		cpu_tlb_flushID();
    579 		pm->pm_cstate.cs_tlb = 0;
    580 	}
    581 }
    582 
    583 static inline void
    584 pmap_tlb_flushD(pmap_t pm)
    585 {
    586 
    587 	if (pm->pm_cstate.cs_tlb_d) {
    588 		cpu_tlb_flushD();
    589 		pm->pm_cstate.cs_tlb_d = 0;
    590 	}
    591 }
    592 
    593 static inline void
    594 pmap_idcache_wbinv_range(pmap_t pm, vaddr_t va, vsize_t len)
    595 {
    596 
    597 	if (pm->pm_cstate.cs_cache_id)
    598 		cpu_idcache_wbinv_range(va, len);
    599 }
    600 
    601 static inline void
    602 pmap_dcache_wb_range(pmap_t pm, vaddr_t va, vsize_t len,
    603     bool do_inv, bool rd_only)
    604 {
    605 
    606 	if (pm->pm_cstate.cs_cache_d) {
    607 		if (do_inv) {
    608 			if (rd_only)
    609 				cpu_dcache_inv_range(va, len);
    610 			else
    611 				cpu_dcache_wbinv_range(va, len);
    612 		} else
    613 		if (!rd_only)
    614 			cpu_dcache_wb_range(va, len);
    615 	}
    616 }
    617 
    618 static inline void
    619 pmap_idcache_wbinv_all(pmap_t pm)
    620 {
    621 
    622 	if (pm->pm_cstate.cs_cache_id) {
    623 		cpu_idcache_wbinv_all();
    624 		pm->pm_cstate.cs_cache = 0;
    625 	}
    626 }
    627 
    628 static inline void
    629 pmap_dcache_wbinv_all(pmap_t pm)
    630 {
    631 
    632 	if (pm->pm_cstate.cs_cache_d) {
    633 		cpu_dcache_wbinv_all();
    634 		pm->pm_cstate.cs_cache_d = 0;
    635 	}
    636 }
    637 
    638 static inline bool
    639 pmap_is_current(pmap_t pm)
    640 {
    641 
    642 	if (pm == pmap_kernel() ||
    643 	    (curproc && curproc->p_vmspace->vm_map.pmap == pm))
    644 		return (true);
    645 
    646 	return (false);
    647 }
    648 
    649 static inline bool
    650 pmap_is_cached(pmap_t pm)
    651 {
    652 
    653 	if (pm == pmap_kernel() || pmap_recent_user == NULL ||
    654 	    pmap_recent_user == pm)
    655 		return (true);
    656 
    657 	return (false);
    658 }
    659 
    660 /*
    661  * PTE_SYNC_CURRENT:
    662  *
    663  *     Make sure the pte is written out to RAM.
    664  *     We need to do this for one of two cases:
    665  *       - We're dealing with the kernel pmap
    666  *       - There is no pmap active in the cache/tlb.
    667  *       - The specified pmap is 'active' in the cache/tlb.
    668  */
    669 #ifdef PMAP_INCLUDE_PTE_SYNC
    670 #define	PTE_SYNC_CURRENT(pm, ptep)	\
    671 do {					\
    672 	if (PMAP_NEEDS_PTE_SYNC && 	\
    673 	    pmap_is_cached(pm))		\
    674 		PTE_SYNC(ptep);		\
    675 } while (/*CONSTCOND*/0)
    676 #else
    677 #define	PTE_SYNC_CURRENT(pm, ptep)	/* nothing */
    678 #endif
    679 
    680 /*
    681  * main pv_entry manipulation functions:
    682  *   pmap_enter_pv: enter a mapping onto a vm_page list
    683  *   pmap_remove_pv: remove a mappiing from a vm_page list
    684  *
    685  * NOTE: pmap_enter_pv expects to lock the pvh itself
    686  *       pmap_remove_pv expects te caller to lock the pvh before calling
    687  */
    688 
    689 /*
    690  * pmap_enter_pv: enter a mapping onto a vm_page lst
    691  *
    692  * => caller should hold the proper lock on pmap_main_lock
    693  * => caller should have pmap locked
    694  * => we will gain the lock on the vm_page and allocate the new pv_entry
    695  * => caller should adjust ptp's wire_count before calling
    696  * => caller should not adjust pmap's wire_count
    697  */
    698 static void
    699 pmap_enter_pv(struct vm_page *pg, struct pv_entry *pve, pmap_t pm,
    700     vaddr_t va, u_int flags)
    701 {
    702 
    703 	NPDEBUG(PDB_PVDUMP,
    704 	    printf("pmap_enter_pv: pm %p, pg %p, flags 0x%x\n", pm, pg, flags));
    705 
    706 	pve->pv_pmap = pm;
    707 	pve->pv_va = va;
    708 	pve->pv_flags = flags;
    709 
    710 	simple_lock(&pg->mdpage.pvh_slock);	/* lock vm_page */
    711 	pve->pv_next = pg->mdpage.pvh_list;	/* add to ... */
    712 	pg->mdpage.pvh_list = pve;		/* ... locked list */
    713 	pg->mdpage.pvh_attrs |= flags & (PVF_REF | PVF_MOD);
    714 	if (pm == pmap_kernel()) {
    715 		if (flags & PVF_WRITE)
    716 			pg->mdpage.krw_mappings++;
    717 		else
    718 			pg->mdpage.kro_mappings++;
    719 	} else
    720 	if (flags & PVF_WRITE)
    721 		pg->mdpage.urw_mappings++;
    722 	else
    723 		pg->mdpage.uro_mappings++;
    724 	simple_unlock(&pg->mdpage.pvh_slock);	/* unlock, done! */
    725 
    726 	if (pve->pv_flags & PVF_WIRED)
    727 		++pm->pm_stats.wired_count;
    728 }
    729 
    730 /*
    731  *
    732  * pmap_find_pv: Find a pv entry
    733  *
    734  * => caller should hold lock on vm_page
    735  */
    736 static inline struct pv_entry *
    737 pmap_find_pv(struct vm_page *pg, pmap_t pm, vaddr_t va)
    738 {
    739 	struct pv_entry *pv;
    740 
    741 	for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
    742 		if (pm == pv->pv_pmap && va == pv->pv_va)
    743 			break;
    744 	}
    745 
    746 	return (pv);
    747 }
    748 
    749 /*
    750  * pmap_remove_pv: try to remove a mapping from a pv_list
    751  *
    752  * => caller should hold proper lock on pmap_main_lock
    753  * => pmap should be locked
    754  * => caller should hold lock on vm_page [so that attrs can be adjusted]
    755  * => caller should adjust ptp's wire_count and free PTP if needed
    756  * => caller should NOT adjust pmap's wire_count
    757  * => we return the removed pve
    758  */
    759 static struct pv_entry *
    760 pmap_remove_pv(struct vm_page *pg, pmap_t pm, vaddr_t va, int skip_wired)
    761 {
    762 	struct pv_entry *pve, **prevptr;
    763 
    764 	NPDEBUG(PDB_PVDUMP,
    765 	    printf("pmap_remove_pv: pm %p, pg %p, va 0x%08lx\n", pm, pg, va));
    766 
    767 	prevptr = &pg->mdpage.pvh_list;		/* previous pv_entry pointer */
    768 	pve = *prevptr;
    769 
    770 	while (pve) {
    771 		if (pve->pv_pmap == pm && pve->pv_va == va) {	/* match? */
    772 			NPDEBUG(PDB_PVDUMP, printf("pmap_remove_pv: pm %p, pg "
    773 			    "%p, flags 0x%x\n", pm, pg, pve->pv_flags));
    774 			if (pve->pv_flags & PVF_WIRED) {
    775 				if (skip_wired)
    776 					return (NULL);
    777 				--pm->pm_stats.wired_count;
    778 			}
    779 			*prevptr = pve->pv_next;		/* remove it! */
    780 			if (pm == pmap_kernel()) {
    781 				if (pve->pv_flags & PVF_WRITE)
    782 					pg->mdpage.krw_mappings--;
    783 				else
    784 					pg->mdpage.kro_mappings--;
    785 			} else
    786 			if (pve->pv_flags & PVF_WRITE)
    787 				pg->mdpage.urw_mappings--;
    788 			else
    789 				pg->mdpage.uro_mappings--;
    790 			break;
    791 		}
    792 		prevptr = &pve->pv_next;		/* previous pointer */
    793 		pve = pve->pv_next;			/* advance */
    794 	}
    795 
    796 	return(pve);				/* return removed pve */
    797 }
    798 
    799 /*
    800  *
    801  * pmap_modify_pv: Update pv flags
    802  *
    803  * => caller should hold lock on vm_page [so that attrs can be adjusted]
    804  * => caller should NOT adjust pmap's wire_count
    805  * => caller must call pmap_vac_me_harder() if writable status of a page
    806  *    may have changed.
    807  * => we return the old flags
    808  *
    809  * Modify a physical-virtual mapping in the pv table
    810  */
    811 static u_int
    812 pmap_modify_pv(struct vm_page *pg, pmap_t pm, vaddr_t va,
    813     u_int clr_mask, u_int set_mask)
    814 {
    815 	struct pv_entry *npv;
    816 	u_int flags, oflags;
    817 
    818 	if ((npv = pmap_find_pv(pg, pm, va)) == NULL)
    819 		return (0);
    820 
    821 	NPDEBUG(PDB_PVDUMP,
    822 	    printf("pmap_modify_pv: pm %p, pg %p, clr 0x%x, set 0x%x, flags 0x%x\n", pm, pg, clr_mask, set_mask, npv->pv_flags));
    823 
    824 	/*
    825 	 * There is at least one VA mapping this page.
    826 	 */
    827 
    828 	if (clr_mask & (PVF_REF | PVF_MOD))
    829 		pg->mdpage.pvh_attrs |= set_mask & (PVF_REF | PVF_MOD);
    830 
    831 	oflags = npv->pv_flags;
    832 	npv->pv_flags = flags = (oflags & ~clr_mask) | set_mask;
    833 
    834 	if ((flags ^ oflags) & PVF_WIRED) {
    835 		if (flags & PVF_WIRED)
    836 			++pm->pm_stats.wired_count;
    837 		else
    838 			--pm->pm_stats.wired_count;
    839 	}
    840 
    841 	if ((flags ^ oflags) & PVF_WRITE) {
    842 		if (pm == pmap_kernel()) {
    843 			if (flags & PVF_WRITE) {
    844 				pg->mdpage.krw_mappings++;
    845 				pg->mdpage.kro_mappings--;
    846 			} else {
    847 				pg->mdpage.kro_mappings++;
    848 				pg->mdpage.krw_mappings--;
    849 			}
    850 		} else
    851 		if (flags & PVF_WRITE) {
    852 			pg->mdpage.urw_mappings++;
    853 			pg->mdpage.uro_mappings--;
    854 		} else {
    855 			pg->mdpage.uro_mappings++;
    856 			pg->mdpage.urw_mappings--;
    857 		}
    858 	}
    859 
    860 	return (oflags);
    861 }
    862 
    863 /*
    864  * Allocate an L1 translation table for the specified pmap.
    865  * This is called at pmap creation time.
    866  */
    867 static void
    868 pmap_alloc_l1(pmap_t pm)
    869 {
    870 	struct l1_ttable *l1;
    871 	u_int8_t domain;
    872 
    873 	/*
    874 	 * Remove the L1 at the head of the LRU list
    875 	 */
    876 	simple_lock(&l1_lru_lock);
    877 	l1 = TAILQ_FIRST(&l1_lru_list);
    878 	KDASSERT(l1 != NULL);
    879 	TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
    880 
    881 	/*
    882 	 * Pick the first available domain number, and update
    883 	 * the link to the next number.
    884 	 */
    885 	domain = l1->l1_domain_first;
    886 	l1->l1_domain_first = l1->l1_domain_free[domain];
    887 
    888 	/*
    889 	 * If there are still free domain numbers in this L1,
    890 	 * put it back on the TAIL of the LRU list.
    891 	 */
    892 	if (++l1->l1_domain_use_count < PMAP_DOMAINS)
    893 		TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
    894 
    895 	simple_unlock(&l1_lru_lock);
    896 
    897 	/*
    898 	 * Fix up the relevant bits in the pmap structure
    899 	 */
    900 	pm->pm_l1 = l1;
    901 	pm->pm_domain = domain;
    902 }
    903 
    904 /*
    905  * Free an L1 translation table.
    906  * This is called at pmap destruction time.
    907  */
    908 static void
    909 pmap_free_l1(pmap_t pm)
    910 {
    911 	struct l1_ttable *l1 = pm->pm_l1;
    912 
    913 	simple_lock(&l1_lru_lock);
    914 
    915 	/*
    916 	 * If this L1 is currently on the LRU list, remove it.
    917 	 */
    918 	if (l1->l1_domain_use_count < PMAP_DOMAINS)
    919 		TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
    920 
    921 	/*
    922 	 * Free up the domain number which was allocated to the pmap
    923 	 */
    924 	l1->l1_domain_free[pm->pm_domain] = l1->l1_domain_first;
    925 	l1->l1_domain_first = pm->pm_domain;
    926 	l1->l1_domain_use_count--;
    927 
    928 	/*
    929 	 * The L1 now must have at least 1 free domain, so add
    930 	 * it back to the LRU list. If the use count is zero,
    931 	 * put it at the head of the list, otherwise it goes
    932 	 * to the tail.
    933 	 */
    934 	if (l1->l1_domain_use_count == 0)
    935 		TAILQ_INSERT_HEAD(&l1_lru_list, l1, l1_lru);
    936 	else
    937 		TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
    938 
    939 	simple_unlock(&l1_lru_lock);
    940 }
    941 
    942 static inline void
    943 pmap_use_l1(pmap_t pm)
    944 {
    945 	struct l1_ttable *l1;
    946 
    947 	/*
    948 	 * Do nothing if we're in interrupt context.
    949 	 * Access to an L1 by the kernel pmap must not affect
    950 	 * the LRU list.
    951 	 */
    952 	if (current_intr_depth || pm == pmap_kernel())
    953 		return;
    954 
    955 	l1 = pm->pm_l1;
    956 
    957 	/*
    958 	 * If the L1 is not currently on the LRU list, just return
    959 	 */
    960 	if (l1->l1_domain_use_count == PMAP_DOMAINS)
    961 		return;
    962 
    963 	simple_lock(&l1_lru_lock);
    964 
    965 	/*
    966 	 * Check the use count again, now that we've acquired the lock
    967 	 */
    968 	if (l1->l1_domain_use_count == PMAP_DOMAINS) {
    969 		simple_unlock(&l1_lru_lock);
    970 		return;
    971 	}
    972 
    973 	/*
    974 	 * Move the L1 to the back of the LRU list
    975 	 */
    976 	TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
    977 	TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
    978 
    979 	simple_unlock(&l1_lru_lock);
    980 }
    981 
    982 /*
    983  * void pmap_free_l2_ptp(pt_entry_t *, paddr_t *)
    984  *
    985  * Free an L2 descriptor table.
    986  */
    987 static inline void
    988 #ifndef PMAP_INCLUDE_PTE_SYNC
    989 pmap_free_l2_ptp(pt_entry_t *l2, paddr_t pa)
    990 #else
    991 pmap_free_l2_ptp(bool need_sync, pt_entry_t *l2, paddr_t pa)
    992 #endif
    993 {
    994 #ifdef PMAP_INCLUDE_PTE_SYNC
    995 	/*
    996 	 * Note: With a write-back cache, we may need to sync this
    997 	 * L2 table before re-using it.
    998 	 * This is because it may have belonged to a non-current
    999 	 * pmap, in which case the cache syncs would have been
   1000 	 * skipped when the pages were being unmapped. If the
   1001 	 * L2 table were then to be immediately re-allocated to
   1002 	 * the *current* pmap, it may well contain stale mappings
   1003 	 * which have not yet been cleared by a cache write-back
   1004 	 * and so would still be visible to the mmu.
   1005 	 */
   1006 	if (need_sync)
   1007 		PTE_SYNC_RANGE(l2, L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
   1008 #endif
   1009 	pool_cache_put_paddr(&pmap_l2ptp_cache, (void *)l2, pa);
   1010 }
   1011 
   1012 /*
   1013  * Returns a pointer to the L2 bucket associated with the specified pmap
   1014  * and VA, or NULL if no L2 bucket exists for the address.
   1015  */
   1016 static inline struct l2_bucket *
   1017 pmap_get_l2_bucket(pmap_t pm, vaddr_t va)
   1018 {
   1019 	struct l2_dtable *l2;
   1020 	struct l2_bucket *l2b;
   1021 	u_short l1idx;
   1022 
   1023 	l1idx = L1_IDX(va);
   1024 
   1025 	if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL ||
   1026 	    (l2b = &l2->l2_bucket[L2_BUCKET(l1idx)])->l2b_kva == NULL)
   1027 		return (NULL);
   1028 
   1029 	return (l2b);
   1030 }
   1031 
   1032 /*
   1033  * Returns a pointer to the L2 bucket associated with the specified pmap
   1034  * and VA.
   1035  *
   1036  * If no L2 bucket exists, perform the necessary allocations to put an L2
   1037  * bucket/page table in place.
   1038  *
   1039  * Note that if a new L2 bucket/page was allocated, the caller *must*
   1040  * increment the bucket occupancy counter appropriately *before*
   1041  * releasing the pmap's lock to ensure no other thread or cpu deallocates
   1042  * the bucket/page in the meantime.
   1043  */
   1044 static struct l2_bucket *
   1045 pmap_alloc_l2_bucket(pmap_t pm, vaddr_t va)
   1046 {
   1047 	struct l2_dtable *l2;
   1048 	struct l2_bucket *l2b;
   1049 	u_short l1idx;
   1050 
   1051 	l1idx = L1_IDX(va);
   1052 
   1053 	if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL) {
   1054 		/*
   1055 		 * No mapping at this address, as there is
   1056 		 * no entry in the L1 table.
   1057 		 * Need to allocate a new l2_dtable.
   1058 		 */
   1059 		if ((l2 = pmap_alloc_l2_dtable()) == NULL)
   1060 			return (NULL);
   1061 
   1062 		/*
   1063 		 * Link it into the parent pmap
   1064 		 */
   1065 		pm->pm_l2[L2_IDX(l1idx)] = l2;
   1066 	}
   1067 
   1068 	l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
   1069 
   1070 	/*
   1071 	 * Fetch pointer to the L2 page table associated with the address.
   1072 	 */
   1073 	if (l2b->l2b_kva == NULL) {
   1074 		pt_entry_t *ptep;
   1075 
   1076 		/*
   1077 		 * No L2 page table has been allocated. Chances are, this
   1078 		 * is because we just allocated the l2_dtable, above.
   1079 		 */
   1080 		if ((ptep = pmap_alloc_l2_ptp(&l2b->l2b_phys)) == NULL) {
   1081 			/*
   1082 			 * Oops, no more L2 page tables available at this
   1083 			 * time. We may need to deallocate the l2_dtable
   1084 			 * if we allocated a new one above.
   1085 			 */
   1086 			if (l2->l2_occupancy == 0) {
   1087 				pm->pm_l2[L2_IDX(l1idx)] = NULL;
   1088 				pmap_free_l2_dtable(l2);
   1089 			}
   1090 			return (NULL);
   1091 		}
   1092 
   1093 		l2->l2_occupancy++;
   1094 		l2b->l2b_kva = ptep;
   1095 		l2b->l2b_l1idx = l1idx;
   1096 	}
   1097 
   1098 	return (l2b);
   1099 }
   1100 
   1101 /*
   1102  * One or more mappings in the specified L2 descriptor table have just been
   1103  * invalidated.
   1104  *
   1105  * Garbage collect the metadata and descriptor table itself if necessary.
   1106  *
   1107  * The pmap lock must be acquired when this is called (not necessary
   1108  * for the kernel pmap).
   1109  */
   1110 static void
   1111 pmap_free_l2_bucket(pmap_t pm, struct l2_bucket *l2b, u_int count)
   1112 {
   1113 	struct l2_dtable *l2;
   1114 	pd_entry_t *pl1pd, l1pd;
   1115 	pt_entry_t *ptep;
   1116 	u_short l1idx;
   1117 
   1118 	KDASSERT(count <= l2b->l2b_occupancy);
   1119 
   1120 	/*
   1121 	 * Update the bucket's reference count according to how many
   1122 	 * PTEs the caller has just invalidated.
   1123 	 */
   1124 	l2b->l2b_occupancy -= count;
   1125 
   1126 	/*
   1127 	 * Note:
   1128 	 *
   1129 	 * Level 2 page tables allocated to the kernel pmap are never freed
   1130 	 * as that would require checking all Level 1 page tables and
   1131 	 * removing any references to the Level 2 page table. See also the
   1132 	 * comment elsewhere about never freeing bootstrap L2 descriptors.
   1133 	 *
   1134 	 * We make do with just invalidating the mapping in the L2 table.
   1135 	 *
   1136 	 * This isn't really a big deal in practice and, in fact, leads
   1137 	 * to a performance win over time as we don't need to continually
   1138 	 * alloc/free.
   1139 	 */
   1140 	if (l2b->l2b_occupancy > 0 || pm == pmap_kernel())
   1141 		return;
   1142 
   1143 	/*
   1144 	 * There are no more valid mappings in this level 2 page table.
   1145 	 * Go ahead and NULL-out the pointer in the bucket, then
   1146 	 * free the page table.
   1147 	 */
   1148 	l1idx = l2b->l2b_l1idx;
   1149 	ptep = l2b->l2b_kva;
   1150 	l2b->l2b_kva = NULL;
   1151 
   1152 	pl1pd = &pm->pm_l1->l1_kva[l1idx];
   1153 
   1154 	/*
   1155 	 * If the L1 slot matches the pmap's domain
   1156 	 * number, then invalidate it.
   1157 	 */
   1158 	l1pd = *pl1pd & (L1_TYPE_MASK | L1_C_DOM_MASK);
   1159 	if (l1pd == (L1_C_DOM(pm->pm_domain) | L1_TYPE_C)) {
   1160 		*pl1pd = 0;
   1161 		PTE_SYNC(pl1pd);
   1162 	}
   1163 
   1164 	/*
   1165 	 * Release the L2 descriptor table back to the pool cache.
   1166 	 */
   1167 #ifndef PMAP_INCLUDE_PTE_SYNC
   1168 	pmap_free_l2_ptp(ptep, l2b->l2b_phys);
   1169 #else
   1170 	pmap_free_l2_ptp(!pmap_is_cached(pm), ptep, l2b->l2b_phys);
   1171 #endif
   1172 
   1173 	/*
   1174 	 * Update the reference count in the associated l2_dtable
   1175 	 */
   1176 	l2 = pm->pm_l2[L2_IDX(l1idx)];
   1177 	if (--l2->l2_occupancy > 0)
   1178 		return;
   1179 
   1180 	/*
   1181 	 * There are no more valid mappings in any of the Level 1
   1182 	 * slots managed by this l2_dtable. Go ahead and NULL-out
   1183 	 * the pointer in the parent pmap and free the l2_dtable.
   1184 	 */
   1185 	pm->pm_l2[L2_IDX(l1idx)] = NULL;
   1186 	pmap_free_l2_dtable(l2);
   1187 }
   1188 
   1189 /*
   1190  * Pool cache constructors for L2 descriptor tables, metadata and pmap
   1191  * structures.
   1192  */
   1193 static int
   1194 pmap_l2ptp_ctor(void *arg, void *v, int flags)
   1195 {
   1196 #ifndef PMAP_INCLUDE_PTE_SYNC
   1197 	struct l2_bucket *l2b;
   1198 	pt_entry_t *ptep, pte;
   1199 	vaddr_t va = (vaddr_t)v & ~PGOFSET;
   1200 
   1201 	/*
   1202 	 * The mappings for these page tables were initially made using
   1203 	 * pmap_kenter_pa() by the pool subsystem. Therefore, the cache-
   1204 	 * mode will not be right for page table mappings. To avoid
   1205 	 * polluting the pmap_kenter_pa() code with a special case for
   1206 	 * page tables, we simply fix up the cache-mode here if it's not
   1207 	 * correct.
   1208 	 */
   1209 	l2b = pmap_get_l2_bucket(pmap_kernel(), va);
   1210 	KDASSERT(l2b != NULL);
   1211 	ptep = &l2b->l2b_kva[l2pte_index(va)];
   1212 	pte = *ptep;
   1213 
   1214 	if ((pte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) {
   1215 		/*
   1216 		 * Page tables must have the cache-mode set to Write-Thru.
   1217 		 */
   1218 		*ptep = (pte & ~L2_S_CACHE_MASK) | pte_l2_s_cache_mode_pt;
   1219 		PTE_SYNC(ptep);
   1220 		cpu_tlb_flushD_SE(va);
   1221 		cpu_cpwait();
   1222 	}
   1223 #endif
   1224 
   1225 	memset(v, 0, L2_TABLE_SIZE_REAL);
   1226 	PTE_SYNC_RANGE(v, L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
   1227 	return (0);
   1228 }
   1229 
   1230 static int
   1231 pmap_l2dtable_ctor(void *arg, void *v, int flags)
   1232 {
   1233 
   1234 	memset(v, 0, sizeof(struct l2_dtable));
   1235 	return (0);
   1236 }
   1237 
   1238 static int
   1239 pmap_pmap_ctor(void *arg, void *v, int flags)
   1240 {
   1241 
   1242 	memset(v, 0, sizeof(struct pmap));
   1243 	return (0);
   1244 }
   1245 
   1246 static void
   1247 pmap_pinit(pmap_t pm)
   1248 {
   1249 	struct l2_bucket *l2b;
   1250 
   1251 	if (vector_page < KERNEL_BASE) {
   1252 		/*
   1253 		 * Map the vector page.
   1254 		 */
   1255 		pmap_enter(pm, vector_page, systempage.pv_pa,
   1256 		    VM_PROT_READ, VM_PROT_READ | PMAP_WIRED);
   1257 		pmap_update(pm);
   1258 
   1259 		pm->pm_pl1vec = &pm->pm_l1->l1_kva[L1_IDX(vector_page)];
   1260 		l2b = pmap_get_l2_bucket(pm, vector_page);
   1261 		pm->pm_l1vec = l2b->l2b_phys | L1_C_PROTO |
   1262 		    L1_C_DOM(pm->pm_domain);
   1263 	} else
   1264 		pm->pm_pl1vec = NULL;
   1265 }
   1266 
   1267 /*
   1268  * Since we have a virtually indexed cache, we may need to inhibit caching if
   1269  * there is more than one mapping and at least one of them is writable.
   1270  * Since we purge the cache on every context switch, we only need to check for
   1271  * other mappings within the same pmap, or kernel_pmap.
   1272  * This function is also called when a page is unmapped, to possibly reenable
   1273  * caching on any remaining mappings.
   1274  *
   1275  * The code implements the following logic, where:
   1276  *
   1277  * KW = # of kernel read/write pages
   1278  * KR = # of kernel read only pages
   1279  * UW = # of user read/write pages
   1280  * UR = # of user read only pages
   1281  *
   1282  * KC = kernel mapping is cacheable
   1283  * UC = user mapping is cacheable
   1284  *
   1285  *               KW=0,KR=0  KW=0,KR>0  KW=1,KR=0  KW>1,KR>=0
   1286  *             +---------------------------------------------
   1287  * UW=0,UR=0   | ---        KC=1       KC=1       KC=0
   1288  * UW=0,UR>0   | UC=1       KC=1,UC=1  KC=0,UC=0  KC=0,UC=0
   1289  * UW=1,UR=0   | UC=1       KC=0,UC=0  KC=0,UC=0  KC=0,UC=0
   1290  * UW>1,UR>=0  | UC=0       KC=0,UC=0  KC=0,UC=0  KC=0,UC=0
   1291  */
   1292 
   1293 static const int pmap_vac_flags[4][4] = {
   1294 	{-1,		0,		0,		PVF_KNC},
   1295 	{0,		0,		PVF_NC,		PVF_NC},
   1296 	{0,		PVF_NC,		PVF_NC,		PVF_NC},
   1297 	{PVF_UNC,	PVF_NC,		PVF_NC,		PVF_NC}
   1298 };
   1299 
   1300 static inline int
   1301 pmap_get_vac_flags(const struct vm_page *pg)
   1302 {
   1303 	int kidx, uidx;
   1304 
   1305 	kidx = 0;
   1306 	if (pg->mdpage.kro_mappings || pg->mdpage.krw_mappings > 1)
   1307 		kidx |= 1;
   1308 	if (pg->mdpage.krw_mappings)
   1309 		kidx |= 2;
   1310 
   1311 	uidx = 0;
   1312 	if (pg->mdpage.uro_mappings || pg->mdpage.urw_mappings > 1)
   1313 		uidx |= 1;
   1314 	if (pg->mdpage.urw_mappings)
   1315 		uidx |= 2;
   1316 
   1317 	return (pmap_vac_flags[uidx][kidx]);
   1318 }
   1319 
   1320 static inline void
   1321 pmap_vac_me_harder(struct vm_page *pg, pmap_t pm, vaddr_t va)
   1322 {
   1323 	int nattr;
   1324 
   1325 	nattr = pmap_get_vac_flags(pg);
   1326 
   1327 	if (nattr < 0) {
   1328 		pg->mdpage.pvh_attrs &= ~PVF_NC;
   1329 		return;
   1330 	}
   1331 
   1332 	if (nattr == 0 && (pg->mdpage.pvh_attrs & PVF_NC) == 0)
   1333 		return;
   1334 
   1335 	if (pm == pmap_kernel())
   1336 		pmap_vac_me_kpmap(pg, pm, va);
   1337 	else
   1338 		pmap_vac_me_user(pg, pm, va);
   1339 
   1340 	pg->mdpage.pvh_attrs = (pg->mdpage.pvh_attrs & ~PVF_NC) | nattr;
   1341 }
   1342 
   1343 static void
   1344 pmap_vac_me_kpmap(struct vm_page *pg, pmap_t pm, vaddr_t va)
   1345 {
   1346 	u_int u_cacheable, u_entries;
   1347 	struct pv_entry *pv;
   1348 	pmap_t last_pmap = pm;
   1349 
   1350 	/*
   1351 	 * Pass one, see if there are both kernel and user pmaps for
   1352 	 * this page.  Calculate whether there are user-writable or
   1353 	 * kernel-writable pages.
   1354 	 */
   1355 	u_cacheable = 0;
   1356 	for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
   1357 		if (pv->pv_pmap != pm && (pv->pv_flags & PVF_NC) == 0)
   1358 			u_cacheable++;
   1359 	}
   1360 
   1361 	u_entries = pg->mdpage.urw_mappings + pg->mdpage.uro_mappings;
   1362 
   1363 	/*
   1364 	 * We know we have just been updating a kernel entry, so if
   1365 	 * all user pages are already cacheable, then there is nothing
   1366 	 * further to do.
   1367 	 */
   1368 	if (pg->mdpage.k_mappings == 0 && u_cacheable == u_entries)
   1369 		return;
   1370 
   1371 	if (u_entries) {
   1372 		/*
   1373 		 * Scan over the list again, for each entry, if it
   1374 		 * might not be set correctly, call pmap_vac_me_user
   1375 		 * to recalculate the settings.
   1376 		 */
   1377 		for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
   1378 			/*
   1379 			 * We know kernel mappings will get set
   1380 			 * correctly in other calls.  We also know
   1381 			 * that if the pmap is the same as last_pmap
   1382 			 * then we've just handled this entry.
   1383 			 */
   1384 			if (pv->pv_pmap == pm || pv->pv_pmap == last_pmap)
   1385 				continue;
   1386 
   1387 			/*
   1388 			 * If there are kernel entries and this page
   1389 			 * is writable but non-cacheable, then we can
   1390 			 * skip this entry also.
   1391 			 */
   1392 			if (pg->mdpage.k_mappings &&
   1393 			    (pv->pv_flags & (PVF_NC | PVF_WRITE)) ==
   1394 			    (PVF_NC | PVF_WRITE))
   1395 				continue;
   1396 
   1397 			/*
   1398 			 * Similarly if there are no kernel-writable
   1399 			 * entries and the page is already
   1400 			 * read-only/cacheable.
   1401 			 */
   1402 			if (pg->mdpage.krw_mappings == 0 &&
   1403 			    (pv->pv_flags & (PVF_NC | PVF_WRITE)) == 0)
   1404 				continue;
   1405 
   1406 			/*
   1407 			 * For some of the remaining cases, we know
   1408 			 * that we must recalculate, but for others we
   1409 			 * can't tell if they are correct or not, so
   1410 			 * we recalculate anyway.
   1411 			 */
   1412 			pmap_vac_me_user(pg, (last_pmap = pv->pv_pmap), 0);
   1413 		}
   1414 
   1415 		if (pg->mdpage.k_mappings == 0)
   1416 			return;
   1417 	}
   1418 
   1419 	pmap_vac_me_user(pg, pm, va);
   1420 }
   1421 
   1422 static void
   1423 pmap_vac_me_user(struct vm_page *pg, pmap_t pm, vaddr_t va)
   1424 {
   1425 	pmap_t kpmap = pmap_kernel();
   1426 	struct pv_entry *pv, *npv;
   1427 	struct l2_bucket *l2b;
   1428 	pt_entry_t *ptep, pte;
   1429 	u_int entries = 0;
   1430 	u_int writable = 0;
   1431 	u_int cacheable_entries = 0;
   1432 	u_int kern_cacheable = 0;
   1433 	u_int other_writable = 0;
   1434 
   1435 	/*
   1436 	 * Count mappings and writable mappings in this pmap.
   1437 	 * Include kernel mappings as part of our own.
   1438 	 * Keep a pointer to the first one.
   1439 	 */
   1440 	for (pv = npv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
   1441 		/* Count mappings in the same pmap */
   1442 		if (pm == pv->pv_pmap || kpmap == pv->pv_pmap) {
   1443 			if (entries++ == 0)
   1444 				npv = pv;
   1445 
   1446 			/* Cacheable mappings */
   1447 			if ((pv->pv_flags & PVF_NC) == 0) {
   1448 				cacheable_entries++;
   1449 				if (kpmap == pv->pv_pmap)
   1450 					kern_cacheable++;
   1451 			}
   1452 
   1453 			/* Writable mappings */
   1454 			if (pv->pv_flags & PVF_WRITE)
   1455 				++writable;
   1456 		} else
   1457 		if (pv->pv_flags & PVF_WRITE)
   1458 			other_writable = 1;
   1459 	}
   1460 
   1461 	/*
   1462 	 * Enable or disable caching as necessary.
   1463 	 * Note: the first entry might be part of the kernel pmap,
   1464 	 * so we can't assume this is indicative of the state of the
   1465 	 * other (maybe non-kpmap) entries.
   1466 	 */
   1467 	if ((entries > 1 && writable) ||
   1468 	    (entries > 0 && pm == kpmap && other_writable)) {
   1469 		if (cacheable_entries == 0)
   1470 			return;
   1471 
   1472 		for (pv = npv; pv; pv = pv->pv_next) {
   1473 			if ((pm != pv->pv_pmap && kpmap != pv->pv_pmap) ||
   1474 			    (pv->pv_flags & PVF_NC))
   1475 				continue;
   1476 
   1477 			pv->pv_flags |= PVF_NC;
   1478 
   1479 			l2b = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va);
   1480 			ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
   1481 			pte = *ptep & ~L2_S_CACHE_MASK;
   1482 
   1483 			if ((va != pv->pv_va || pm != pv->pv_pmap) &&
   1484 			    l2pte_valid(pte)) {
   1485 				if (PV_BEEN_EXECD(pv->pv_flags)) {
   1486 					pmap_idcache_wbinv_range(pv->pv_pmap,
   1487 					    pv->pv_va, PAGE_SIZE);
   1488 					pmap_tlb_flushID_SE(pv->pv_pmap,
   1489 					    pv->pv_va);
   1490 				} else
   1491 				if (PV_BEEN_REFD(pv->pv_flags)) {
   1492 					pmap_dcache_wb_range(pv->pv_pmap,
   1493 					    pv->pv_va, PAGE_SIZE, true,
   1494 					    (pv->pv_flags & PVF_WRITE) == 0);
   1495 					pmap_tlb_flushD_SE(pv->pv_pmap,
   1496 					    pv->pv_va);
   1497 				}
   1498 			}
   1499 
   1500 			*ptep = pte;
   1501 			PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
   1502 		}
   1503 		cpu_cpwait();
   1504 	} else
   1505 	if (entries > cacheable_entries) {
   1506 		/*
   1507 		 * Turn cacheing back on for some pages.  If it is a kernel
   1508 		 * page, only do so if there are no other writable pages.
   1509 		 */
   1510 		for (pv = npv; pv; pv = pv->pv_next) {
   1511 			if (!(pv->pv_flags & PVF_NC) || (pm != pv->pv_pmap &&
   1512 			    (kpmap != pv->pv_pmap || other_writable)))
   1513 				continue;
   1514 
   1515 			pv->pv_flags &= ~PVF_NC;
   1516 
   1517 			l2b = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va);
   1518 			ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
   1519 			pte = (*ptep & ~L2_S_CACHE_MASK) | pte_l2_s_cache_mode;
   1520 
   1521 			if (l2pte_valid(pte)) {
   1522 				if (PV_BEEN_EXECD(pv->pv_flags)) {
   1523 					pmap_tlb_flushID_SE(pv->pv_pmap,
   1524 					    pv->pv_va);
   1525 				} else
   1526 				if (PV_BEEN_REFD(pv->pv_flags)) {
   1527 					pmap_tlb_flushD_SE(pv->pv_pmap,
   1528 					    pv->pv_va);
   1529 				}
   1530 			}
   1531 
   1532 			*ptep = pte;
   1533 			PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
   1534 		}
   1535 	}
   1536 }
   1537 
   1538 /*
   1539  * Modify pte bits for all ptes corresponding to the given physical address.
   1540  * We use `maskbits' rather than `clearbits' because we're always passing
   1541  * constants and the latter would require an extra inversion at run-time.
   1542  */
   1543 static void
   1544 pmap_clearbit(struct vm_page *pg, u_int maskbits)
   1545 {
   1546 	struct l2_bucket *l2b;
   1547 	struct pv_entry *pv;
   1548 	pt_entry_t *ptep, npte, opte;
   1549 	pmap_t pm;
   1550 	vaddr_t va;
   1551 	u_int oflags;
   1552 
   1553 	NPDEBUG(PDB_BITS,
   1554 	    printf("pmap_clearbit: pg %p (0x%08lx) mask 0x%x\n",
   1555 	    pg, VM_PAGE_TO_PHYS(pg), maskbits));
   1556 
   1557 	PMAP_HEAD_TO_MAP_LOCK();
   1558 	simple_lock(&pg->mdpage.pvh_slock);
   1559 
   1560 	/*
   1561 	 * Clear saved attributes (modify, reference)
   1562 	 */
   1563 	pg->mdpage.pvh_attrs &= ~(maskbits & (PVF_MOD | PVF_REF));
   1564 
   1565 	if (pg->mdpage.pvh_list == NULL) {
   1566 		simple_unlock(&pg->mdpage.pvh_slock);
   1567 		PMAP_HEAD_TO_MAP_UNLOCK();
   1568 		return;
   1569 	}
   1570 
   1571 	/*
   1572 	 * Loop over all current mappings setting/clearing as appropos
   1573 	 */
   1574 	for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
   1575 		va = pv->pv_va;
   1576 		pm = pv->pv_pmap;
   1577 		oflags = pv->pv_flags;
   1578 		pv->pv_flags &= ~maskbits;
   1579 
   1580 		pmap_acquire_pmap_lock(pm);
   1581 
   1582 		l2b = pmap_get_l2_bucket(pm, va);
   1583 		KDASSERT(l2b != NULL);
   1584 
   1585 		ptep = &l2b->l2b_kva[l2pte_index(va)];
   1586 		npte = opte = *ptep;
   1587 
   1588 		NPDEBUG(PDB_BITS,
   1589 		    printf(
   1590 		    "pmap_clearbit: pv %p, pm %p, va 0x%08lx, flag 0x%x\n",
   1591 		    pv, pv->pv_pmap, pv->pv_va, oflags));
   1592 
   1593 		if (maskbits & (PVF_WRITE|PVF_MOD)) {
   1594 			if ((pv->pv_flags & PVF_NC)) {
   1595 				/*
   1596 				 * Entry is not cacheable:
   1597 				 *
   1598 				 * Don't turn caching on again if this is a
   1599 				 * modified emulation. This would be
   1600 				 * inconsitent with the settings created by
   1601 				 * pmap_vac_me_harder(). Otherwise, it's safe
   1602 				 * to re-enable cacheing.
   1603 				 *
   1604 				 * There's no need to call pmap_vac_me_harder()
   1605 				 * here: all pages are losing their write
   1606 				 * permission.
   1607 				 */
   1608 				if (maskbits & PVF_WRITE) {
   1609 					npte |= pte_l2_s_cache_mode;
   1610 					pv->pv_flags &= ~PVF_NC;
   1611 				}
   1612 			} else
   1613 			if (opte & L2_S_PROT_W) {
   1614 				/*
   1615 				 * Entry is writable/cacheable: check if pmap
   1616 				 * is current if it is flush it, otherwise it
   1617 				 * won't be in the cache
   1618 				 */
   1619 				if (PV_BEEN_EXECD(oflags))
   1620 					pmap_idcache_wbinv_range(pm, pv->pv_va,
   1621 					    PAGE_SIZE);
   1622 				else
   1623 				if (PV_BEEN_REFD(oflags))
   1624 					pmap_dcache_wb_range(pm, pv->pv_va,
   1625 					    PAGE_SIZE,
   1626 					    (maskbits & PVF_REF) ? true : false,
   1627 					    false);
   1628 			}
   1629 
   1630 			/* make the pte read only */
   1631 			npte &= ~L2_S_PROT_W;
   1632 
   1633 			if (maskbits & PVF_WRITE) {
   1634 				/*
   1635 				 * Keep alias accounting up to date
   1636 				 */
   1637 				if (pv->pv_pmap == pmap_kernel()) {
   1638 					if (oflags & PVF_WRITE) {
   1639 						pg->mdpage.krw_mappings--;
   1640 						pg->mdpage.kro_mappings++;
   1641 					}
   1642 				} else
   1643 				if (oflags & PVF_WRITE) {
   1644 					pg->mdpage.urw_mappings--;
   1645 					pg->mdpage.uro_mappings++;
   1646 				}
   1647 			}
   1648 		}
   1649 
   1650 		if (maskbits & PVF_REF) {
   1651 			if ((pv->pv_flags & PVF_NC) == 0 &&
   1652 			    (maskbits & (PVF_WRITE|PVF_MOD)) == 0) {
   1653 				/*
   1654 				 * Check npte here; we may have already
   1655 				 * done the wbinv above, and the validity
   1656 				 * of the PTE is the same for opte and
   1657 				 * npte.
   1658 				 */
   1659 				if (npte & L2_S_PROT_W) {
   1660 					if (PV_BEEN_EXECD(oflags))
   1661 						pmap_idcache_wbinv_range(pm,
   1662 						    pv->pv_va, PAGE_SIZE);
   1663 					else
   1664 					if (PV_BEEN_REFD(oflags))
   1665 						pmap_dcache_wb_range(pm,
   1666 						    pv->pv_va, PAGE_SIZE,
   1667 						    true, false);
   1668 				} else
   1669 				if ((npte & L2_TYPE_MASK) != L2_TYPE_INV) {
   1670 					/* XXXJRT need idcache_inv_range */
   1671 					if (PV_BEEN_EXECD(oflags))
   1672 						pmap_idcache_wbinv_range(pm,
   1673 						    pv->pv_va, PAGE_SIZE);
   1674 					else
   1675 					if (PV_BEEN_REFD(oflags))
   1676 						pmap_dcache_wb_range(pm,
   1677 						    pv->pv_va, PAGE_SIZE,
   1678 						    true, true);
   1679 				}
   1680 			}
   1681 
   1682 			/*
   1683 			 * Make the PTE invalid so that we will take a
   1684 			 * page fault the next time the mapping is
   1685 			 * referenced.
   1686 			 */
   1687 			npte &= ~L2_TYPE_MASK;
   1688 			npte |= L2_TYPE_INV;
   1689 		}
   1690 
   1691 		if (npte != opte) {
   1692 			*ptep = npte;
   1693 			PTE_SYNC(ptep);
   1694 			/* Flush the TLB entry if a current pmap. */
   1695 			if (PV_BEEN_EXECD(oflags))
   1696 				pmap_tlb_flushID_SE(pm, pv->pv_va);
   1697 			else
   1698 			if (PV_BEEN_REFD(oflags))
   1699 				pmap_tlb_flushD_SE(pm, pv->pv_va);
   1700 		}
   1701 
   1702 		pmap_release_pmap_lock(pm);
   1703 
   1704 		NPDEBUG(PDB_BITS,
   1705 		    printf("pmap_clearbit: pm %p va 0x%lx opte 0x%08x npte 0x%08x\n",
   1706 		    pm, va, opte, npte));
   1707 	}
   1708 
   1709 	simple_unlock(&pg->mdpage.pvh_slock);
   1710 	PMAP_HEAD_TO_MAP_UNLOCK();
   1711 }
   1712 
   1713 /*
   1714  * pmap_clean_page()
   1715  *
   1716  * This is a local function used to work out the best strategy to clean
   1717  * a single page referenced by its entry in the PV table. It's used by
   1718  * pmap_copy_page, pmap_zero page and maybe some others later on.
   1719  *
   1720  * Its policy is effectively:
   1721  *  o If there are no mappings, we don't bother doing anything with the cache.
   1722  *  o If there is one mapping, we clean just that page.
   1723  *  o If there are multiple mappings, we clean the entire cache.
   1724  *
   1725  * So that some functions can be further optimised, it returns 0 if it didn't
   1726  * clean the entire cache, or 1 if it did.
   1727  *
   1728  * XXX One bug in this routine is that if the pv_entry has a single page
   1729  * mapped at 0x00000000 a whole cache clean will be performed rather than
   1730  * just the 1 page. Since this should not occur in everyday use and if it does
   1731  * it will just result in not the most efficient clean for the page.
   1732  */
   1733 static int
   1734 pmap_clean_page(struct pv_entry *pv, bool is_src)
   1735 {
   1736 	pmap_t pm, pm_to_clean = NULL;
   1737 	struct pv_entry *npv;
   1738 	u_int cache_needs_cleaning = 0;
   1739 	u_int flags = 0;
   1740 	vaddr_t page_to_clean = 0;
   1741 
   1742 	if (pv == NULL) {
   1743 		/* nothing mapped in so nothing to flush */
   1744 		return (0);
   1745 	}
   1746 
   1747 	/*
   1748 	 * Since we flush the cache each time we change to a different
   1749 	 * user vmspace, we only need to flush the page if it is in the
   1750 	 * current pmap.
   1751 	 */
   1752 	if (curproc)
   1753 		pm = curproc->p_vmspace->vm_map.pmap;
   1754 	else
   1755 		pm = pmap_kernel();
   1756 
   1757 	for (npv = pv; npv; npv = npv->pv_next) {
   1758 		if (npv->pv_pmap == pmap_kernel() || npv->pv_pmap == pm) {
   1759 			flags |= npv->pv_flags;
   1760 			/*
   1761 			 * The page is mapped non-cacheable in
   1762 			 * this map.  No need to flush the cache.
   1763 			 */
   1764 			if (npv->pv_flags & PVF_NC) {
   1765 #ifdef DIAGNOSTIC
   1766 				if (cache_needs_cleaning)
   1767 					panic("pmap_clean_page: "
   1768 					    "cache inconsistency");
   1769 #endif
   1770 				break;
   1771 			} else if (is_src && (npv->pv_flags & PVF_WRITE) == 0)
   1772 				continue;
   1773 			if (cache_needs_cleaning) {
   1774 				page_to_clean = 0;
   1775 				break;
   1776 			} else {
   1777 				page_to_clean = npv->pv_va;
   1778 				pm_to_clean = npv->pv_pmap;
   1779 			}
   1780 			cache_needs_cleaning = 1;
   1781 		}
   1782 	}
   1783 
   1784 	if (page_to_clean) {
   1785 		if (PV_BEEN_EXECD(flags))
   1786 			pmap_idcache_wbinv_range(pm_to_clean, page_to_clean,
   1787 			    PAGE_SIZE);
   1788 		else
   1789 			pmap_dcache_wb_range(pm_to_clean, page_to_clean,
   1790 			    PAGE_SIZE, !is_src, (flags & PVF_WRITE) == 0);
   1791 	} else if (cache_needs_cleaning) {
   1792 		if (PV_BEEN_EXECD(flags))
   1793 			pmap_idcache_wbinv_all(pm);
   1794 		else
   1795 			pmap_dcache_wbinv_all(pm);
   1796 		return (1);
   1797 	}
   1798 	return (0);
   1799 }
   1800 
   1801 /*
   1802  * Routine:	pmap_page_remove
   1803  * Function:
   1804  *		Removes this physical page from
   1805  *		all physical maps in which it resides.
   1806  *		Reflects back modify bits to the pager.
   1807  */
   1808 static void
   1809 pmap_page_remove(struct vm_page *pg)
   1810 {
   1811 	struct l2_bucket *l2b;
   1812 	struct pv_entry *pv, *npv;
   1813 	pmap_t pm, curpm;
   1814 	pt_entry_t *ptep, pte;
   1815 	bool flush;
   1816 	u_int flags;
   1817 
   1818 	NPDEBUG(PDB_FOLLOW,
   1819 	    printf("pmap_page_remove: pg %p (0x%08lx)\n", pg,
   1820 	    VM_PAGE_TO_PHYS(pg)));
   1821 
   1822 	PMAP_HEAD_TO_MAP_LOCK();
   1823 	simple_lock(&pg->mdpage.pvh_slock);
   1824 
   1825 	pv = pg->mdpage.pvh_list;
   1826 	if (pv == NULL) {
   1827 		simple_unlock(&pg->mdpage.pvh_slock);
   1828 		PMAP_HEAD_TO_MAP_UNLOCK();
   1829 		return;
   1830 	}
   1831 
   1832 	/*
   1833 	 * Clear alias counts
   1834 	 */
   1835 	pg->mdpage.k_mappings = 0;
   1836 	pg->mdpage.urw_mappings = pg->mdpage.uro_mappings = 0;
   1837 
   1838 	flush = false;
   1839 	flags = 0;
   1840 	if (curproc)
   1841 		curpm = curproc->p_vmspace->vm_map.pmap;
   1842 	else
   1843 		curpm = pmap_kernel();
   1844 
   1845 	pmap_clean_page(pv, false);
   1846 
   1847 	while (pv) {
   1848 		pm = pv->pv_pmap;
   1849 		if (flush == false && (pm == curpm || pm == pmap_kernel()))
   1850 			flush = true;
   1851 
   1852 		pmap_acquire_pmap_lock(pm);
   1853 
   1854 		l2b = pmap_get_l2_bucket(pm, pv->pv_va);
   1855 		KDASSERT(l2b != NULL);
   1856 
   1857 		ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
   1858 		pte = *ptep;
   1859 
   1860 		/*
   1861 		 * Update statistics
   1862 		 */
   1863 		--pm->pm_stats.resident_count;
   1864 
   1865 		/* Wired bit */
   1866 		if (pv->pv_flags & PVF_WIRED)
   1867 			--pm->pm_stats.wired_count;
   1868 
   1869 		flags |= pv->pv_flags;
   1870 
   1871 		/*
   1872 		 * Invalidate the PTEs.
   1873 		 */
   1874 		*ptep = 0;
   1875 		PTE_SYNC_CURRENT(pm, ptep);
   1876 		pmap_free_l2_bucket(pm, l2b, 1);
   1877 
   1878 		npv = pv->pv_next;
   1879 		pool_put(&pmap_pv_pool, pv);
   1880 		pv = npv;
   1881 		pmap_release_pmap_lock(pm);
   1882 	}
   1883 	pg->mdpage.pvh_list = NULL;
   1884 	simple_unlock(&pg->mdpage.pvh_slock);
   1885 	PMAP_HEAD_TO_MAP_UNLOCK();
   1886 
   1887 	if (flush) {
   1888 		/*
   1889 		 * Note: We can't use pmap_tlb_flush{I,}D() here since that
   1890 		 * would need a subsequent call to pmap_update() to ensure
   1891 		 * curpm->pm_cstate.cs_all is reset. Our callers are not
   1892 		 * required to do that (see pmap(9)), so we can't modify
   1893 		 * the current pmap's state.
   1894 		 */
   1895 		if (PV_BEEN_EXECD(flags))
   1896 			cpu_tlb_flushID();
   1897 		else
   1898 			cpu_tlb_flushD();
   1899 	}
   1900 	cpu_cpwait();
   1901 }
   1902 
   1903 /*
   1904  * pmap_t pmap_create(void)
   1905  *
   1906  *      Create a new pmap structure from scratch.
   1907  */
   1908 pmap_t
   1909 pmap_create(void)
   1910 {
   1911 	pmap_t pm;
   1912 
   1913 	pm = pool_cache_get(&pmap_pmap_cache, PR_WAITOK);
   1914 
   1915 	simple_lock_init(&pm->pm_lock);
   1916 	pm->pm_obj.pgops = NULL;	/* currently not a mappable object */
   1917 	TAILQ_INIT(&pm->pm_obj.memq);
   1918 	pm->pm_obj.uo_npages = 0;
   1919 	pm->pm_obj.uo_refs = 1;
   1920 	pm->pm_stats.wired_count = 0;
   1921 	pm->pm_stats.resident_count = 1;
   1922 	pm->pm_cstate.cs_all = 0;
   1923 	pmap_alloc_l1(pm);
   1924 
   1925 	/*
   1926 	 * Note: The pool cache ensures that the pm_l2[] array is already
   1927 	 * initialised to zero.
   1928 	 */
   1929 
   1930 	pmap_pinit(pm);
   1931 
   1932 	LIST_INSERT_HEAD(&pmap_pmaps, pm, pm_list);
   1933 
   1934 	return (pm);
   1935 }
   1936 
   1937 /*
   1938  * void pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
   1939  *     int flags)
   1940  *
   1941  *      Insert the given physical page (p) at
   1942  *      the specified virtual address (v) in the
   1943  *      target physical map with the protection requested.
   1944  *
   1945  *      NB:  This is the only routine which MAY NOT lazy-evaluate
   1946  *      or lose information.  That is, this routine must actually
   1947  *      insert this page into the given map NOW.
   1948  */
   1949 int
   1950 pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot, int flags)
   1951 {
   1952 	struct l2_bucket *l2b;
   1953 	struct vm_page *pg, *opg;
   1954 	struct pv_entry *pve;
   1955 	pt_entry_t *ptep, npte, opte;
   1956 	u_int nflags;
   1957 	u_int oflags;
   1958 
   1959 	NPDEBUG(PDB_ENTER, printf("pmap_enter: pm %p va 0x%lx pa 0x%lx prot %x flag %x\n", pm, va, pa, prot, flags));
   1960 
   1961 	KDASSERT((flags & PMAP_WIRED) == 0 || (flags & VM_PROT_ALL) != 0);
   1962 	KDASSERT(((va | pa) & PGOFSET) == 0);
   1963 
   1964 	/*
   1965 	 * Get a pointer to the page.  Later on in this function, we
   1966 	 * test for a managed page by checking pg != NULL.
   1967 	 */
   1968 	pg = pmap_initialized ? PHYS_TO_VM_PAGE(pa) : NULL;
   1969 
   1970 	nflags = 0;
   1971 	if (prot & VM_PROT_WRITE)
   1972 		nflags |= PVF_WRITE;
   1973 	if (prot & VM_PROT_EXECUTE)
   1974 		nflags |= PVF_EXEC;
   1975 	if (flags & PMAP_WIRED)
   1976 		nflags |= PVF_WIRED;
   1977 
   1978 	PMAP_MAP_TO_HEAD_LOCK();
   1979 	pmap_acquire_pmap_lock(pm);
   1980 
   1981 	/*
   1982 	 * Fetch the L2 bucket which maps this page, allocating one if
   1983 	 * necessary for user pmaps.
   1984 	 */
   1985 	if (pm == pmap_kernel())
   1986 		l2b = pmap_get_l2_bucket(pm, va);
   1987 	else
   1988 		l2b = pmap_alloc_l2_bucket(pm, va);
   1989 	if (l2b == NULL) {
   1990 		if (flags & PMAP_CANFAIL) {
   1991 			pmap_release_pmap_lock(pm);
   1992 			PMAP_MAP_TO_HEAD_UNLOCK();
   1993 			return (ENOMEM);
   1994 		}
   1995 		panic("pmap_enter: failed to allocate L2 bucket");
   1996 	}
   1997 	ptep = &l2b->l2b_kva[l2pte_index(va)];
   1998 	opte = *ptep;
   1999 	npte = pa;
   2000 	oflags = 0;
   2001 
   2002 	if (opte) {
   2003 		/*
   2004 		 * There is already a mapping at this address.
   2005 		 * If the physical address is different, lookup the
   2006 		 * vm_page.
   2007 		 */
   2008 		if (l2pte_pa(opte) != pa)
   2009 			opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
   2010 		else
   2011 			opg = pg;
   2012 	} else
   2013 		opg = NULL;
   2014 
   2015 	if (pg) {
   2016 		/*
   2017 		 * This is to be a managed mapping.
   2018 		 */
   2019 		if ((flags & VM_PROT_ALL) ||
   2020 		    (pg->mdpage.pvh_attrs & PVF_REF)) {
   2021 			/*
   2022 			 * - The access type indicates that we don't need
   2023 			 *   to do referenced emulation.
   2024 			 * OR
   2025 			 * - The physical page has already been referenced
   2026 			 *   so no need to re-do referenced emulation here.
   2027 			 */
   2028 			npte |= L2_S_PROTO;
   2029 
   2030 			nflags |= PVF_REF;
   2031 
   2032 			if ((prot & VM_PROT_WRITE) != 0 &&
   2033 			    ((flags & VM_PROT_WRITE) != 0 ||
   2034 			     (pg->mdpage.pvh_attrs & PVF_MOD) != 0)) {
   2035 				/*
   2036 				 * This is a writable mapping, and the
   2037 				 * page's mod state indicates it has
   2038 				 * already been modified. Make it
   2039 				 * writable from the outset.
   2040 				 */
   2041 				npte |= L2_S_PROT_W;
   2042 				nflags |= PVF_MOD;
   2043 			}
   2044 		} else {
   2045 			/*
   2046 			 * Need to do page referenced emulation.
   2047 			 */
   2048 			npte |= L2_TYPE_INV;
   2049 		}
   2050 
   2051 		npte |= pte_l2_s_cache_mode;
   2052 
   2053 		if (pg == opg) {
   2054 			/*
   2055 			 * We're changing the attrs of an existing mapping.
   2056 			 */
   2057 			simple_lock(&pg->mdpage.pvh_slock);
   2058 			oflags = pmap_modify_pv(pg, pm, va,
   2059 			    PVF_WRITE | PVF_EXEC | PVF_WIRED |
   2060 			    PVF_MOD | PVF_REF, nflags);
   2061 			simple_unlock(&pg->mdpage.pvh_slock);
   2062 
   2063 			/*
   2064 			 * We may need to flush the cache if we're
   2065 			 * doing rw-ro...
   2066 			 */
   2067 			if (pm->pm_cstate.cs_cache_d &&
   2068 			    (oflags & PVF_NC) == 0 &&
   2069 			    (opte & L2_S_PROT_W) != 0 &&
   2070 			    (prot & VM_PROT_WRITE) == 0)
   2071 				cpu_dcache_wb_range(va, PAGE_SIZE);
   2072 		} else {
   2073 			/*
   2074 			 * New mapping, or changing the backing page
   2075 			 * of an existing mapping.
   2076 			 */
   2077 			if (opg) {
   2078 				/*
   2079 				 * Replacing an existing mapping with a new one.
   2080 				 * It is part of our managed memory so we
   2081 				 * must remove it from the PV list
   2082 				 */
   2083 				simple_lock(&opg->mdpage.pvh_slock);
   2084 				pve = pmap_remove_pv(opg, pm, va, 0);
   2085 				pmap_vac_me_harder(opg, pm, 0);
   2086 				simple_unlock(&opg->mdpage.pvh_slock);
   2087 				oflags = pve->pv_flags;
   2088 
   2089 				/*
   2090 				 * If the old mapping was valid (ref/mod
   2091 				 * emulation creates 'invalid' mappings
   2092 				 * initially) then make sure to frob
   2093 				 * the cache.
   2094 				 */
   2095 				if ((oflags & PVF_NC) == 0 &&
   2096 				    l2pte_valid(opte)) {
   2097 					if (PV_BEEN_EXECD(oflags)) {
   2098 						pmap_idcache_wbinv_range(pm, va,
   2099 						    PAGE_SIZE);
   2100 					} else
   2101 					if (PV_BEEN_REFD(oflags)) {
   2102 						pmap_dcache_wb_range(pm, va,
   2103 						    PAGE_SIZE, true,
   2104 						    (oflags & PVF_WRITE) == 0);
   2105 					}
   2106 				}
   2107 			} else
   2108 			if ((pve = pool_get(&pmap_pv_pool, PR_NOWAIT)) == NULL){
   2109 				if ((flags & PMAP_CANFAIL) == 0)
   2110 					panic("pmap_enter: no pv entries");
   2111 
   2112 				if (pm != pmap_kernel())
   2113 					pmap_free_l2_bucket(pm, l2b, 0);
   2114 				pmap_release_pmap_lock(pm);
   2115 				PMAP_MAP_TO_HEAD_UNLOCK();
   2116 				NPDEBUG(PDB_ENTER,
   2117 				    printf("pmap_enter: ENOMEM\n"));
   2118 				return (ENOMEM);
   2119 			}
   2120 
   2121 			pmap_enter_pv(pg, pve, pm, va, nflags);
   2122 		}
   2123 	} else {
   2124 		/*
   2125 		 * We're mapping an unmanaged page.
   2126 		 * These are always readable, and possibly writable, from
   2127 		 * the get go as we don't need to track ref/mod status.
   2128 		 */
   2129 		npte |= L2_S_PROTO;
   2130 		if (prot & VM_PROT_WRITE)
   2131 			npte |= L2_S_PROT_W;
   2132 
   2133 		/*
   2134 		 * Make sure the vector table is mapped cacheable
   2135 		 */
   2136 		if (pm != pmap_kernel() && va == vector_page)
   2137 			npte |= pte_l2_s_cache_mode;
   2138 
   2139 		if (opg) {
   2140 			/*
   2141 			 * Looks like there's an existing 'managed' mapping
   2142 			 * at this address.
   2143 			 */
   2144 			simple_lock(&opg->mdpage.pvh_slock);
   2145 			pve = pmap_remove_pv(opg, pm, va, 0);
   2146 			pmap_vac_me_harder(opg, pm, 0);
   2147 			simple_unlock(&opg->mdpage.pvh_slock);
   2148 			oflags = pve->pv_flags;
   2149 
   2150 			if ((oflags & PVF_NC) == 0 && l2pte_valid(opte)) {
   2151 				if (PV_BEEN_EXECD(oflags))
   2152 					pmap_idcache_wbinv_range(pm, va,
   2153 					    PAGE_SIZE);
   2154 				else
   2155 				if (PV_BEEN_REFD(oflags))
   2156 					pmap_dcache_wb_range(pm, va, PAGE_SIZE,
   2157 					    true, (oflags & PVF_WRITE) == 0);
   2158 			}
   2159 			pool_put(&pmap_pv_pool, pve);
   2160 		}
   2161 	}
   2162 
   2163 	/*
   2164 	 * Make sure userland mappings get the right permissions
   2165 	 */
   2166 	if (pm != pmap_kernel() && va != vector_page)
   2167 		npte |= L2_S_PROT_U;
   2168 
   2169 	/*
   2170 	 * Keep the stats up to date
   2171 	 */
   2172 	if (opte == 0) {
   2173 		l2b->l2b_occupancy++;
   2174 		pm->pm_stats.resident_count++;
   2175 	}
   2176 
   2177 	NPDEBUG(PDB_ENTER,
   2178 	    printf("pmap_enter: opte 0x%08x npte 0x%08x\n", opte, npte));
   2179 
   2180 	/*
   2181 	 * If this is just a wiring change, the two PTEs will be
   2182 	 * identical, so there's no need to update the page table.
   2183 	 */
   2184 	if (npte != opte) {
   2185 		bool is_cached = pmap_is_cached(pm);
   2186 
   2187 		*ptep = npte;
   2188 		if (is_cached) {
   2189 			/*
   2190 			 * We only need to frob the cache/tlb if this pmap
   2191 			 * is current
   2192 			 */
   2193 			PTE_SYNC(ptep);
   2194 			if (va != vector_page && l2pte_valid(npte)) {
   2195 				/*
   2196 				 * This mapping is likely to be accessed as
   2197 				 * soon as we return to userland. Fix up the
   2198 				 * L1 entry to avoid taking another
   2199 				 * page/domain fault.
   2200 				 */
   2201 				pd_entry_t *pl1pd, l1pd;
   2202 
   2203 				pl1pd = &pm->pm_l1->l1_kva[L1_IDX(va)];
   2204 				l1pd = l2b->l2b_phys | L1_C_DOM(pm->pm_domain) |
   2205 				    L1_C_PROTO;
   2206 				if (*pl1pd != l1pd) {
   2207 					*pl1pd = l1pd;
   2208 					PTE_SYNC(pl1pd);
   2209 				}
   2210 			}
   2211 		}
   2212 
   2213 		if (PV_BEEN_EXECD(oflags))
   2214 			pmap_tlb_flushID_SE(pm, va);
   2215 		else
   2216 		if (PV_BEEN_REFD(oflags))
   2217 			pmap_tlb_flushD_SE(pm, va);
   2218 
   2219 		NPDEBUG(PDB_ENTER,
   2220 		    printf("pmap_enter: is_cached %d cs 0x%08x\n",
   2221 		    is_cached, pm->pm_cstate.cs_all));
   2222 
   2223 		if (pg != NULL) {
   2224 			simple_lock(&pg->mdpage.pvh_slock);
   2225 			pmap_vac_me_harder(pg, pm, va);
   2226 			simple_unlock(&pg->mdpage.pvh_slock);
   2227 		}
   2228 	}
   2229 
   2230 	pmap_release_pmap_lock(pm);
   2231 	PMAP_MAP_TO_HEAD_UNLOCK();
   2232 
   2233 	return (0);
   2234 }
   2235 
   2236 /*
   2237  * pmap_remove()
   2238  *
   2239  * pmap_remove is responsible for nuking a number of mappings for a range
   2240  * of virtual address space in the current pmap. To do this efficiently
   2241  * is interesting, because in a number of cases a wide virtual address
   2242  * range may be supplied that contains few actual mappings. So, the
   2243  * optimisations are:
   2244  *  1. Skip over hunks of address space for which no L1 or L2 entry exists.
   2245  *  2. Build up a list of pages we've hit, up to a maximum, so we can
   2246  *     maybe do just a partial cache clean. This path of execution is
   2247  *     complicated by the fact that the cache must be flushed _before_
   2248  *     the PTE is nuked, being a VAC :-)
   2249  *  3. If we're called after UVM calls pmap_remove_all(), we can defer
   2250  *     all invalidations until pmap_update(), since pmap_remove_all() has
   2251  *     already flushed the cache.
   2252  *  4. Maybe later fast-case a single page, but I don't think this is
   2253  *     going to make _that_ much difference overall.
   2254  */
   2255 
   2256 #define	PMAP_REMOVE_CLEAN_LIST_SIZE	3
   2257 
   2258 void
   2259 pmap_do_remove(pmap_t pm, vaddr_t sva, vaddr_t eva, int skip_wired)
   2260 {
   2261 	struct l2_bucket *l2b;
   2262 	vaddr_t next_bucket;
   2263 	pt_entry_t *ptep;
   2264 	u_int cleanlist_idx, total, cnt;
   2265 	struct {
   2266 		vaddr_t va;
   2267 		pt_entry_t *pte;
   2268 	} cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE];
   2269 	u_int mappings, is_exec, is_refd;
   2270 
   2271 	NPDEBUG(PDB_REMOVE, printf("pmap_do_remove: pmap=%p sva=%08lx "
   2272 	    "eva=%08lx\n", pm, sva, eva));
   2273 
   2274 	/*
   2275 	 * we lock in the pmap => pv_head direction
   2276 	 */
   2277 	PMAP_MAP_TO_HEAD_LOCK();
   2278 	pmap_acquire_pmap_lock(pm);
   2279 
   2280 	if (pm->pm_remove_all || !pmap_is_cached(pm)) {
   2281 		cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1;
   2282 		if (pm->pm_cstate.cs_tlb == 0)
   2283 			pm->pm_remove_all = true;
   2284 	} else
   2285 		cleanlist_idx = 0;
   2286 
   2287 	total = 0;
   2288 
   2289 	while (sva < eva) {
   2290 		/*
   2291 		 * Do one L2 bucket's worth at a time.
   2292 		 */
   2293 		next_bucket = L2_NEXT_BUCKET(sva);
   2294 		if (next_bucket > eva)
   2295 			next_bucket = eva;
   2296 
   2297 		l2b = pmap_get_l2_bucket(pm, sva);
   2298 		if (l2b == NULL) {
   2299 			sva = next_bucket;
   2300 			continue;
   2301 		}
   2302 
   2303 		ptep = &l2b->l2b_kva[l2pte_index(sva)];
   2304 
   2305 		for (mappings = 0; sva < next_bucket; sva += PAGE_SIZE, ptep++){
   2306 			struct vm_page *pg;
   2307 			pt_entry_t pte;
   2308 			paddr_t pa;
   2309 
   2310 			pte = *ptep;
   2311 
   2312 			if (pte == 0) {
   2313 				/* Nothing here, move along */
   2314 				continue;
   2315 			}
   2316 
   2317 			pa = l2pte_pa(pte);
   2318 			is_exec = 0;
   2319 			is_refd = 1;
   2320 
   2321 			/*
   2322 			 * Update flags. In a number of circumstances,
   2323 			 * we could cluster a lot of these and do a
   2324 			 * number of sequential pages in one go.
   2325 			 */
   2326 			if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
   2327 				struct pv_entry *pve;
   2328 				simple_lock(&pg->mdpage.pvh_slock);
   2329 				pve = pmap_remove_pv(pg, pm, sva, skip_wired);
   2330 				pmap_vac_me_harder(pg, pm, 0);
   2331 				simple_unlock(&pg->mdpage.pvh_slock);
   2332 				if (pve != NULL) {
   2333 					if (pm->pm_remove_all == false) {
   2334 						is_exec =
   2335 						   PV_BEEN_EXECD(pve->pv_flags);
   2336 						is_refd =
   2337 						   PV_BEEN_REFD(pve->pv_flags);
   2338 					}
   2339 					pool_put(&pmap_pv_pool, pve);
   2340 				} else
   2341 				if (skip_wired) {
   2342 					/* The mapping is wired. Skip it */
   2343 					continue;
   2344 				}
   2345 			} else
   2346 			if (skip_wired) {
   2347 				/* Unmanaged pages are always wired. */
   2348 				continue;
   2349 			}
   2350 
   2351 			mappings++;
   2352 
   2353 			if (!l2pte_valid(pte)) {
   2354 				/*
   2355 				 * Ref/Mod emulation is still active for this
   2356 				 * mapping, therefore it is has not yet been
   2357 				 * accessed. No need to frob the cache/tlb.
   2358 				 */
   2359 				*ptep = 0;
   2360 				PTE_SYNC_CURRENT(pm, ptep);
   2361 				continue;
   2362 			}
   2363 
   2364 			if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) {
   2365 				/* Add to the clean list. */
   2366 				cleanlist[cleanlist_idx].pte = ptep;
   2367 				cleanlist[cleanlist_idx].va =
   2368 				    sva | (is_exec & 1);
   2369 				cleanlist_idx++;
   2370 			} else
   2371 			if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) {
   2372 				/* Nuke everything if needed. */
   2373 				pmap_idcache_wbinv_all(pm);
   2374 				pmap_tlb_flushID(pm);
   2375 
   2376 				/*
   2377 				 * Roll back the previous PTE list,
   2378 				 * and zero out the current PTE.
   2379 				 */
   2380 				for (cnt = 0;
   2381 				     cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) {
   2382 					*cleanlist[cnt].pte = 0;
   2383 				}
   2384 				*ptep = 0;
   2385 				PTE_SYNC(ptep);
   2386 				cleanlist_idx++;
   2387 				pm->pm_remove_all = true;
   2388 			} else {
   2389 				*ptep = 0;
   2390 				PTE_SYNC(ptep);
   2391 				if (pm->pm_remove_all == false) {
   2392 					if (is_exec)
   2393 						pmap_tlb_flushID_SE(pm, sva);
   2394 					else
   2395 					if (is_refd)
   2396 						pmap_tlb_flushD_SE(pm, sva);
   2397 				}
   2398 			}
   2399 		}
   2400 
   2401 		/*
   2402 		 * Deal with any left overs
   2403 		 */
   2404 		if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) {
   2405 			total += cleanlist_idx;
   2406 			for (cnt = 0; cnt < cleanlist_idx; cnt++) {
   2407 				if (pm->pm_cstate.cs_all != 0) {
   2408 					vaddr_t clva = cleanlist[cnt].va & ~1;
   2409 					if (cleanlist[cnt].va & 1) {
   2410 						pmap_idcache_wbinv_range(pm,
   2411 						    clva, PAGE_SIZE);
   2412 						pmap_tlb_flushID_SE(pm, clva);
   2413 					} else {
   2414 						pmap_dcache_wb_range(pm,
   2415 						    clva, PAGE_SIZE, true,
   2416 						    false);
   2417 						pmap_tlb_flushD_SE(pm, clva);
   2418 					}
   2419 				}
   2420 				*cleanlist[cnt].pte = 0;
   2421 				PTE_SYNC_CURRENT(pm, cleanlist[cnt].pte);
   2422 			}
   2423 
   2424 			/*
   2425 			 * If it looks like we're removing a whole bunch
   2426 			 * of mappings, it's faster to just write-back
   2427 			 * the whole cache now and defer TLB flushes until
   2428 			 * pmap_update() is called.
   2429 			 */
   2430 			if (total <= PMAP_REMOVE_CLEAN_LIST_SIZE)
   2431 				cleanlist_idx = 0;
   2432 			else {
   2433 				cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1;
   2434 				pmap_idcache_wbinv_all(pm);
   2435 				pm->pm_remove_all = true;
   2436 			}
   2437 		}
   2438 
   2439 		pmap_free_l2_bucket(pm, l2b, mappings);
   2440 		pm->pm_stats.resident_count -= mappings;
   2441 	}
   2442 
   2443 	pmap_release_pmap_lock(pm);
   2444 	PMAP_MAP_TO_HEAD_UNLOCK();
   2445 }
   2446 
   2447 /*
   2448  * pmap_kenter_pa: enter an unmanaged, wired kernel mapping
   2449  *
   2450  * We assume there is already sufficient KVM space available
   2451  * to do this, as we can't allocate L2 descriptor tables/metadata
   2452  * from here.
   2453  */
   2454 void
   2455 pmap_kenter_pa(vaddr_t va, paddr_t pa, vm_prot_t prot)
   2456 {
   2457 	struct l2_bucket *l2b;
   2458 	pt_entry_t *ptep, opte;
   2459 
   2460 	NPDEBUG(PDB_KENTER,
   2461 	    printf("pmap_kenter_pa: va 0x%08lx, pa 0x%08lx, prot 0x%x\n",
   2462 	    va, pa, prot));
   2463 
   2464 	l2b = pmap_get_l2_bucket(pmap_kernel(), va);
   2465 	KDASSERT(l2b != NULL);
   2466 
   2467 	ptep = &l2b->l2b_kva[l2pte_index(va)];
   2468 	opte = *ptep;
   2469 
   2470 	if (l2pte_valid(opte)) {
   2471 		cpu_dcache_wbinv_range(va, PAGE_SIZE);
   2472 		cpu_tlb_flushD_SE(va);
   2473 		cpu_cpwait();
   2474 	} else
   2475 	if (opte == 0)
   2476 		l2b->l2b_occupancy++;
   2477 
   2478 	*ptep = L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) |
   2479 	    pte_l2_s_cache_mode;
   2480 	PTE_SYNC(ptep);
   2481 }
   2482 
   2483 void
   2484 pmap_kremove(vaddr_t va, vsize_t len)
   2485 {
   2486 	struct l2_bucket *l2b;
   2487 	pt_entry_t *ptep, *sptep, opte;
   2488 	vaddr_t next_bucket, eva;
   2489 	u_int mappings;
   2490 
   2491 	NPDEBUG(PDB_KREMOVE, printf("pmap_kremove: va 0x%08lx, len 0x%08lx\n",
   2492 	    va, len));
   2493 
   2494 	eva = va + len;
   2495 
   2496 	while (va < eva) {
   2497 		next_bucket = L2_NEXT_BUCKET(va);
   2498 		if (next_bucket > eva)
   2499 			next_bucket = eva;
   2500 
   2501 		l2b = pmap_get_l2_bucket(pmap_kernel(), va);
   2502 		KDASSERT(l2b != NULL);
   2503 
   2504 		sptep = ptep = &l2b->l2b_kva[l2pte_index(va)];
   2505 		mappings = 0;
   2506 
   2507 		while (va < next_bucket) {
   2508 			opte = *ptep;
   2509 			if (l2pte_valid(opte)) {
   2510 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
   2511 				cpu_tlb_flushD_SE(va);
   2512 			}
   2513 			if (opte) {
   2514 				*ptep = 0;
   2515 				mappings++;
   2516 			}
   2517 			va += PAGE_SIZE;
   2518 			ptep++;
   2519 		}
   2520 		KDASSERT(mappings <= l2b->l2b_occupancy);
   2521 		l2b->l2b_occupancy -= mappings;
   2522 		PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep));
   2523 	}
   2524 	cpu_cpwait();
   2525 }
   2526 
   2527 bool
   2528 pmap_extract(pmap_t pm, vaddr_t va, paddr_t *pap)
   2529 {
   2530 	struct l2_dtable *l2;
   2531 	pd_entry_t *pl1pd, l1pd;
   2532 	pt_entry_t *ptep, pte;
   2533 	paddr_t pa;
   2534 	u_int l1idx;
   2535 
   2536 	pmap_acquire_pmap_lock(pm);
   2537 
   2538 	l1idx = L1_IDX(va);
   2539 	pl1pd = &pm->pm_l1->l1_kva[l1idx];
   2540 	l1pd = *pl1pd;
   2541 
   2542 	if (l1pte_section_p(l1pd)) {
   2543 		/*
   2544 		 * These should only happen for pmap_kernel()
   2545 		 */
   2546 		KDASSERT(pm == pmap_kernel());
   2547 		pmap_release_pmap_lock(pm);
   2548 		pa = (l1pd & L1_S_FRAME) | (va & L1_S_OFFSET);
   2549 	} else {
   2550 		/*
   2551 		 * Note that we can't rely on the validity of the L1
   2552 		 * descriptor as an indication that a mapping exists.
   2553 		 * We have to look it up in the L2 dtable.
   2554 		 */
   2555 		l2 = pm->pm_l2[L2_IDX(l1idx)];
   2556 
   2557 		if (l2 == NULL ||
   2558 		    (ptep = l2->l2_bucket[L2_BUCKET(l1idx)].l2b_kva) == NULL) {
   2559 			pmap_release_pmap_lock(pm);
   2560 			return (false);
   2561 		}
   2562 
   2563 		ptep = &ptep[l2pte_index(va)];
   2564 		pte = *ptep;
   2565 		pmap_release_pmap_lock(pm);
   2566 
   2567 		if (pte == 0)
   2568 			return (false);
   2569 
   2570 		switch (pte & L2_TYPE_MASK) {
   2571 		case L2_TYPE_L:
   2572 			pa = (pte & L2_L_FRAME) | (va & L2_L_OFFSET);
   2573 			break;
   2574 
   2575 		default:
   2576 			pa = (pte & L2_S_FRAME) | (va & L2_S_OFFSET);
   2577 			break;
   2578 		}
   2579 	}
   2580 
   2581 	if (pap != NULL)
   2582 		*pap = pa;
   2583 
   2584 	return (true);
   2585 }
   2586 
   2587 void
   2588 pmap_protect(pmap_t pm, vaddr_t sva, vaddr_t eva, vm_prot_t prot)
   2589 {
   2590 	struct l2_bucket *l2b;
   2591 	pt_entry_t *ptep, pte;
   2592 	vaddr_t next_bucket;
   2593 	u_int flags;
   2594 	int flush;
   2595 
   2596 	NPDEBUG(PDB_PROTECT,
   2597 	    printf("pmap_protect: pm %p sva 0x%lx eva 0x%lx prot 0x%x\n",
   2598 	    pm, sva, eva, prot));
   2599 
   2600 	if ((prot & VM_PROT_READ) == 0) {
   2601 		pmap_remove(pm, sva, eva);
   2602 		return;
   2603 	}
   2604 
   2605 	if (prot & VM_PROT_WRITE) {
   2606 		/*
   2607 		 * If this is a read->write transition, just ignore it and let
   2608 		 * uvm_fault() take care of it later.
   2609 		 */
   2610 		return;
   2611 	}
   2612 
   2613 	PMAP_MAP_TO_HEAD_LOCK();
   2614 	pmap_acquire_pmap_lock(pm);
   2615 
   2616 	/*
   2617 	 * OK, at this point, we know we're doing write-protect operation.
   2618 	 * If the pmap is active, write-back the range.
   2619 	 */
   2620 	pmap_dcache_wb_range(pm, sva, eva - sva, false, false);
   2621 
   2622 	flush = ((eva - sva) >= (PAGE_SIZE * 4)) ? 0 : -1;
   2623 	flags = 0;
   2624 
   2625 	while (sva < eva) {
   2626 		next_bucket = L2_NEXT_BUCKET(sva);
   2627 		if (next_bucket > eva)
   2628 			next_bucket = eva;
   2629 
   2630 		l2b = pmap_get_l2_bucket(pm, sva);
   2631 		if (l2b == NULL) {
   2632 			sva = next_bucket;
   2633 			continue;
   2634 		}
   2635 
   2636 		ptep = &l2b->l2b_kva[l2pte_index(sva)];
   2637 
   2638 		while (sva < next_bucket) {
   2639 			if ((pte = *ptep) != 0 && (pte & L2_S_PROT_W) != 0) {
   2640 				struct vm_page *pg;
   2641 				u_int f;
   2642 
   2643 				pg = PHYS_TO_VM_PAGE(l2pte_pa(pte));
   2644 				pte &= ~L2_S_PROT_W;
   2645 				*ptep = pte;
   2646 				PTE_SYNC(ptep);
   2647 
   2648 				if (pg != NULL) {
   2649 					simple_lock(&pg->mdpage.pvh_slock);
   2650 					f = pmap_modify_pv(pg, pm, sva,
   2651 					    PVF_WRITE, 0);
   2652 					pmap_vac_me_harder(pg, pm, sva);
   2653 					simple_unlock(&pg->mdpage.pvh_slock);
   2654 				} else
   2655 					f = PVF_REF | PVF_EXEC;
   2656 
   2657 				if (flush >= 0) {
   2658 					flush++;
   2659 					flags |= f;
   2660 				} else
   2661 				if (PV_BEEN_EXECD(f))
   2662 					pmap_tlb_flushID_SE(pm, sva);
   2663 				else
   2664 				if (PV_BEEN_REFD(f))
   2665 					pmap_tlb_flushD_SE(pm, sva);
   2666 			}
   2667 
   2668 			sva += PAGE_SIZE;
   2669 			ptep++;
   2670 		}
   2671 	}
   2672 
   2673 	pmap_release_pmap_lock(pm);
   2674 	PMAP_MAP_TO_HEAD_UNLOCK();
   2675 
   2676 	if (flush) {
   2677 		if (PV_BEEN_EXECD(flags))
   2678 			pmap_tlb_flushID(pm);
   2679 		else
   2680 		if (PV_BEEN_REFD(flags))
   2681 			pmap_tlb_flushD(pm);
   2682 	}
   2683 }
   2684 
   2685 void
   2686 pmap_page_protect(struct vm_page *pg, vm_prot_t prot)
   2687 {
   2688 
   2689 	NPDEBUG(PDB_PROTECT,
   2690 	    printf("pmap_page_protect: pg %p (0x%08lx), prot 0x%x\n",
   2691 	    pg, VM_PAGE_TO_PHYS(pg), prot));
   2692 
   2693 	switch(prot) {
   2694 	case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE:
   2695 	case VM_PROT_READ|VM_PROT_WRITE:
   2696 		return;
   2697 
   2698 	case VM_PROT_READ:
   2699 	case VM_PROT_READ|VM_PROT_EXECUTE:
   2700 		pmap_clearbit(pg, PVF_WRITE);
   2701 		break;
   2702 
   2703 	default:
   2704 		pmap_page_remove(pg);
   2705 		break;
   2706 	}
   2707 }
   2708 
   2709 /*
   2710  * pmap_clear_modify:
   2711  *
   2712  *	Clear the "modified" attribute for a page.
   2713  */
   2714 bool
   2715 pmap_clear_modify(struct vm_page *pg)
   2716 {
   2717 	bool rv;
   2718 
   2719 	if (pg->mdpage.pvh_attrs & PVF_MOD) {
   2720 		rv = true;
   2721 		pmap_clearbit(pg, PVF_MOD);
   2722 	} else
   2723 		rv = false;
   2724 
   2725 	return (rv);
   2726 }
   2727 
   2728 /*
   2729  * pmap_clear_reference:
   2730  *
   2731  *	Clear the "referenced" attribute for a page.
   2732  */
   2733 bool
   2734 pmap_clear_reference(struct vm_page *pg)
   2735 {
   2736 	bool rv;
   2737 
   2738 	if (pg->mdpage.pvh_attrs & PVF_REF) {
   2739 		rv = true;
   2740 		pmap_clearbit(pg, PVF_REF);
   2741 	} else
   2742 		rv = false;
   2743 
   2744 	return (rv);
   2745 }
   2746 
   2747 /*
   2748  * pmap_is_modified:
   2749  *
   2750  *	Test if a page has the "modified" attribute.
   2751  */
   2752 /* See <arm/arm32/pmap.h> */
   2753 
   2754 /*
   2755  * pmap_is_referenced:
   2756  *
   2757  *	Test if a page has the "referenced" attribute.
   2758  */
   2759 /* See <arm/arm32/pmap.h> */
   2760 
   2761 int
   2762 pmap_fault_fixup(pmap_t pm, vaddr_t va, vm_prot_t ftype, int user)
   2763 {
   2764 	struct l2_dtable *l2;
   2765 	struct l2_bucket *l2b;
   2766 	pd_entry_t *pl1pd, l1pd;
   2767 	pt_entry_t *ptep, pte;
   2768 	paddr_t pa;
   2769 	u_int l1idx;
   2770 	int rv = 0;
   2771 
   2772 	PMAP_MAP_TO_HEAD_LOCK();
   2773 	pmap_acquire_pmap_lock(pm);
   2774 
   2775 	l1idx = L1_IDX(va);
   2776 
   2777 	/*
   2778 	 * If there is no l2_dtable for this address, then the process
   2779 	 * has no business accessing it.
   2780 	 *
   2781 	 * Note: This will catch userland processes trying to access
   2782 	 * kernel addresses.
   2783 	 */
   2784 	l2 = pm->pm_l2[L2_IDX(l1idx)];
   2785 	if (l2 == NULL)
   2786 		goto out;
   2787 
   2788 	/*
   2789 	 * Likewise if there is no L2 descriptor table
   2790 	 */
   2791 	l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
   2792 	if (l2b->l2b_kva == NULL)
   2793 		goto out;
   2794 
   2795 	/*
   2796 	 * Check the PTE itself.
   2797 	 */
   2798 	ptep = &l2b->l2b_kva[l2pte_index(va)];
   2799 	pte = *ptep;
   2800 	if (pte == 0)
   2801 		goto out;
   2802 
   2803 	/*
   2804 	 * Catch a userland access to the vector page mapped at 0x0
   2805 	 */
   2806 	if (user && (pte & L2_S_PROT_U) == 0)
   2807 		goto out;
   2808 
   2809 	pa = l2pte_pa(pte);
   2810 
   2811 	if ((ftype & VM_PROT_WRITE) && (pte & L2_S_PROT_W) == 0) {
   2812 		/*
   2813 		 * This looks like a good candidate for "page modified"
   2814 		 * emulation...
   2815 		 */
   2816 		struct pv_entry *pv;
   2817 		struct vm_page *pg;
   2818 
   2819 		/* Extract the physical address of the page */
   2820 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
   2821 			goto out;
   2822 
   2823 		/* Get the current flags for this page. */
   2824 		simple_lock(&pg->mdpage.pvh_slock);
   2825 
   2826 		pv = pmap_find_pv(pg, pm, va);
   2827 		if (pv == NULL) {
   2828 	    		simple_unlock(&pg->mdpage.pvh_slock);
   2829 			goto out;
   2830 		}
   2831 
   2832 		/*
   2833 		 * Do the flags say this page is writable? If not then it
   2834 		 * is a genuine write fault. If yes then the write fault is
   2835 		 * our fault as we did not reflect the write access in the
   2836 		 * PTE. Now we know a write has occurred we can correct this
   2837 		 * and also set the modified bit
   2838 		 */
   2839 		if ((pv->pv_flags & PVF_WRITE) == 0) {
   2840 		    	simple_unlock(&pg->mdpage.pvh_slock);
   2841 			goto out;
   2842 		}
   2843 
   2844 		NPDEBUG(PDB_FOLLOW,
   2845 		    printf("pmap_fault_fixup: mod emul. pm %p, va 0x%08lx, pa 0x%08lx\n",
   2846 		    pm, va, VM_PAGE_TO_PHYS(pg)));
   2847 
   2848 		pg->mdpage.pvh_attrs |= PVF_REF | PVF_MOD;
   2849 		pv->pv_flags |= PVF_REF | PVF_MOD;
   2850 		simple_unlock(&pg->mdpage.pvh_slock);
   2851 
   2852 		/*
   2853 		 * Re-enable write permissions for the page.  No need to call
   2854 		 * pmap_vac_me_harder(), since this is just a
   2855 		 * modified-emulation fault, and the PVF_WRITE bit isn't
   2856 		 * changing. We've already set the cacheable bits based on
   2857 		 * the assumption that we can write to this page.
   2858 		 */
   2859 		*ptep = (pte & ~L2_TYPE_MASK) | L2_S_PROTO | L2_S_PROT_W;
   2860 		PTE_SYNC(ptep);
   2861 		rv = 1;
   2862 	} else
   2863 	if ((pte & L2_TYPE_MASK) == L2_TYPE_INV) {
   2864 		/*
   2865 		 * This looks like a good candidate for "page referenced"
   2866 		 * emulation.
   2867 		 */
   2868 		struct pv_entry *pv;
   2869 		struct vm_page *pg;
   2870 
   2871 		/* Extract the physical address of the page */
   2872 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
   2873 			goto out;
   2874 
   2875 		/* Get the current flags for this page. */
   2876 		simple_lock(&pg->mdpage.pvh_slock);
   2877 
   2878 		pv = pmap_find_pv(pg, pm, va);
   2879 		if (pv == NULL) {
   2880 	    		simple_unlock(&pg->mdpage.pvh_slock);
   2881 			goto out;
   2882 		}
   2883 
   2884 		pg->mdpage.pvh_attrs |= PVF_REF;
   2885 		pv->pv_flags |= PVF_REF;
   2886 		simple_unlock(&pg->mdpage.pvh_slock);
   2887 
   2888 		NPDEBUG(PDB_FOLLOW,
   2889 		    printf("pmap_fault_fixup: ref emul. pm %p, va 0x%08lx, pa 0x%08lx\n",
   2890 		    pm, va, VM_PAGE_TO_PHYS(pg)));
   2891 
   2892 		*ptep = (pte & ~L2_TYPE_MASK) | L2_S_PROTO;
   2893 		PTE_SYNC(ptep);
   2894 		rv = 1;
   2895 	}
   2896 
   2897 	/*
   2898 	 * We know there is a valid mapping here, so simply
   2899 	 * fix up the L1 if necessary.
   2900 	 */
   2901 	pl1pd = &pm->pm_l1->l1_kva[l1idx];
   2902 	l1pd = l2b->l2b_phys | L1_C_DOM(pm->pm_domain) | L1_C_PROTO;
   2903 	if (*pl1pd != l1pd) {
   2904 		*pl1pd = l1pd;
   2905 		PTE_SYNC(pl1pd);
   2906 		rv = 1;
   2907 	}
   2908 
   2909 #ifdef CPU_SA110
   2910 	/*
   2911 	 * There are bugs in the rev K SA110.  This is a check for one
   2912 	 * of them.
   2913 	 */
   2914 	if (rv == 0 && curcpu()->ci_arm_cputype == CPU_ID_SA110 &&
   2915 	    curcpu()->ci_arm_cpurev < 3) {
   2916 		/* Always current pmap */
   2917 		if (l2pte_valid(pte)) {
   2918 			extern int kernel_debug;
   2919 			if (kernel_debug & 1) {
   2920 				struct proc *p = curlwp->l_proc;
   2921 				printf("prefetch_abort: page is already "
   2922 				    "mapped - pte=%p *pte=%08x\n", ptep, pte);
   2923 				printf("prefetch_abort: pc=%08lx proc=%p "
   2924 				    "process=%s\n", va, p, p->p_comm);
   2925 				printf("prefetch_abort: far=%08x fs=%x\n",
   2926 				    cpu_faultaddress(), cpu_faultstatus());
   2927 			}
   2928 #ifdef DDB
   2929 			if (kernel_debug & 2)
   2930 				Debugger();
   2931 #endif
   2932 			rv = 1;
   2933 		}
   2934 	}
   2935 #endif /* CPU_SA110 */
   2936 
   2937 #ifdef DEBUG
   2938 	/*
   2939 	 * If 'rv == 0' at this point, it generally indicates that there is a
   2940 	 * stale TLB entry for the faulting address. This happens when two or
   2941 	 * more processes are sharing an L1. Since we don't flush the TLB on
   2942 	 * a context switch between such processes, we can take domain faults
   2943 	 * for mappings which exist at the same VA in both processes. EVEN IF
   2944 	 * WE'VE RECENTLY FIXED UP THE CORRESPONDING L1 in pmap_enter(), for
   2945 	 * example.
   2946 	 *
   2947 	 * This is extremely likely to happen if pmap_enter() updated the L1
   2948 	 * entry for a recently entered mapping. In this case, the TLB is
   2949 	 * flushed for the new mapping, but there may still be TLB entries for
   2950 	 * other mappings belonging to other processes in the 1MB range
   2951 	 * covered by the L1 entry.
   2952 	 *
   2953 	 * Since 'rv == 0', we know that the L1 already contains the correct
   2954 	 * value, so the fault must be due to a stale TLB entry.
   2955 	 *
   2956 	 * Since we always need to flush the TLB anyway in the case where we
   2957 	 * fixed up the L1, or frobbed the L2 PTE, we effectively deal with
   2958 	 * stale TLB entries dynamically.
   2959 	 *
   2960 	 * However, the above condition can ONLY happen if the current L1 is
   2961 	 * being shared. If it happens when the L1 is unshared, it indicates
   2962 	 * that other parts of the pmap are not doing their job WRT managing
   2963 	 * the TLB.
   2964 	 */
   2965 	if (rv == 0 && pm->pm_l1->l1_domain_use_count == 1) {
   2966 		extern int last_fault_code;
   2967 		printf("fixup: pm %p, va 0x%lx, ftype %d - nothing to do!\n",
   2968 		    pm, va, ftype);
   2969 		printf("fixup: l2 %p, l2b %p, ptep %p, pl1pd %p\n",
   2970 		    l2, l2b, ptep, pl1pd);
   2971 		printf("fixup: pte 0x%x, l1pd 0x%x, last code 0x%x\n",
   2972 		    pte, l1pd, last_fault_code);
   2973 #ifdef DDB
   2974 		Debugger();
   2975 #endif
   2976 	}
   2977 #endif
   2978 
   2979 	cpu_tlb_flushID_SE(va);
   2980 	cpu_cpwait();
   2981 
   2982 	rv = 1;
   2983 
   2984 out:
   2985 	pmap_release_pmap_lock(pm);
   2986 	PMAP_MAP_TO_HEAD_UNLOCK();
   2987 
   2988 	return (rv);
   2989 }
   2990 
   2991 /*
   2992  * pmap_collect: free resources held by a pmap
   2993  *
   2994  * => optional function.
   2995  * => called when a process is swapped out to free memory.
   2996  */
   2997 void
   2998 pmap_collect(pmap_t pm)
   2999 {
   3000 
   3001 	pmap_idcache_wbinv_all(pm);
   3002 	pm->pm_remove_all = true;
   3003 	pmap_do_remove(pm, VM_MIN_ADDRESS, VM_MAX_ADDRESS, 1);
   3004 	pmap_update(pm);
   3005 }
   3006 
   3007 /*
   3008  * Routine:	pmap_procwr
   3009  *
   3010  * Function:
   3011  *	Synchronize caches corresponding to [addr, addr+len) in p.
   3012  *
   3013  */
   3014 void
   3015 pmap_procwr(struct proc *p, vaddr_t va, int len)
   3016 {
   3017 	/* We only need to do anything if it is the current process. */
   3018 	if (p == curproc)
   3019 		cpu_icache_sync_range(va, len);
   3020 }
   3021 
   3022 /*
   3023  * Routine:	pmap_unwire
   3024  * Function:	Clear the wired attribute for a map/virtual-address pair.
   3025  *
   3026  * In/out conditions:
   3027  *		The mapping must already exist in the pmap.
   3028  */
   3029 void
   3030 pmap_unwire(pmap_t pm, vaddr_t va)
   3031 {
   3032 	struct l2_bucket *l2b;
   3033 	pt_entry_t *ptep, pte;
   3034 	struct vm_page *pg;
   3035 	paddr_t pa;
   3036 
   3037 	NPDEBUG(PDB_WIRING, printf("pmap_unwire: pm %p, va 0x%08lx\n", pm, va));
   3038 
   3039 	PMAP_MAP_TO_HEAD_LOCK();
   3040 	pmap_acquire_pmap_lock(pm);
   3041 
   3042 	l2b = pmap_get_l2_bucket(pm, va);
   3043 	KDASSERT(l2b != NULL);
   3044 
   3045 	ptep = &l2b->l2b_kva[l2pte_index(va)];
   3046 	pte = *ptep;
   3047 
   3048 	/* Extract the physical address of the page */
   3049 	pa = l2pte_pa(pte);
   3050 
   3051 	if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
   3052 		/* Update the wired bit in the pv entry for this page. */
   3053 		simple_lock(&pg->mdpage.pvh_slock);
   3054 		(void) pmap_modify_pv(pg, pm, va, PVF_WIRED, 0);
   3055 		simple_unlock(&pg->mdpage.pvh_slock);
   3056 	}
   3057 
   3058 	pmap_release_pmap_lock(pm);
   3059 	PMAP_MAP_TO_HEAD_UNLOCK();
   3060 }
   3061 
   3062 void
   3063 pmap_switch(struct lwp *olwp, struct lwp *nlwp)
   3064 {
   3065 	extern int block_userspace_access;
   3066 	pmap_t opm, npm, rpm;
   3067 	uint32_t odacr, ndacr;
   3068 	int oldirqstate;
   3069 
   3070 	npm = nlwp->l_proc->p_vmspace->vm_map.pmap;
   3071 	ndacr = (DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL * 2)) |
   3072 	    (DOMAIN_CLIENT << (npm->pm_domain * 2));
   3073 
   3074 	/*
   3075 	 * If TTB and DACR are unchanged, short-circuit all the
   3076 	 * TLB/cache management stuff.
   3077 	 */
   3078 	if (olwp != NULL) {
   3079 		opm = olwp->l_proc->p_vmspace->vm_map.pmap;
   3080 		odacr = (DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL * 2)) |
   3081 		    (DOMAIN_CLIENT << (opm->pm_domain * 2));
   3082 
   3083 		if (opm->pm_l1 == npm->pm_l1 && odacr == ndacr)
   3084 			goto all_done;
   3085 	} else
   3086 		opm = NULL;
   3087 
   3088 	block_userspace_access = 1;
   3089 
   3090 	/*
   3091 	 * If switching to a user vmspace which is different to the
   3092 	 * most recent one, and the most recent one is potentially
   3093 	 * live in the cache, we must write-back and invalidate the
   3094 	 * entire cache.
   3095 	 */
   3096 	rpm = pmap_recent_user;
   3097 	if (npm != pmap_kernel() && rpm && npm != rpm &&
   3098 	    rpm->pm_cstate.cs_cache) {
   3099 		rpm->pm_cstate.cs_cache = 0;
   3100 		cpu_idcache_wbinv_all();
   3101 	}
   3102 
   3103 	/* No interrupts while we frob the TTB/DACR */
   3104 	oldirqstate = disable_interrupts(I32_bit | F32_bit);
   3105 
   3106 	/*
   3107 	 * For ARM_VECTORS_LOW, we MUST, I repeat, MUST fix up the L1
   3108 	 * entry corresponding to 'vector_page' in the incoming L1 table
   3109 	 * before switching to it otherwise subsequent interrupts/exceptions
   3110 	 * (including domain faults!) will jump into hyperspace.
   3111 	 */
   3112 	if (npm->pm_pl1vec != NULL) {
   3113 		cpu_tlb_flushID_SE((u_int)vector_page);
   3114 		cpu_cpwait();
   3115 		*npm->pm_pl1vec = npm->pm_l1vec;
   3116 		PTE_SYNC(npm->pm_pl1vec);
   3117 	}
   3118 
   3119 	cpu_domains(ndacr);
   3120 
   3121 	if (npm == pmap_kernel() || npm == rpm) {
   3122 		/*
   3123 		 * Switching to a kernel thread, or back to the
   3124 		 * same user vmspace as before... Simply update
   3125 		 * the TTB (no TLB flush required)
   3126 		 */
   3127 		__asm volatile("mcr p15, 0, %0, c2, c0, 0" ::
   3128 		    "r"(npm->pm_l1->l1_physaddr));
   3129 		cpu_cpwait();
   3130 	} else {
   3131 		/*
   3132 		 * Otherwise, update TTB and flush TLB
   3133 		 */
   3134 		cpu_context_switch(npm->pm_l1->l1_physaddr);
   3135 		if (rpm != NULL)
   3136 			rpm->pm_cstate.cs_tlb = 0;
   3137 	}
   3138 
   3139 	restore_interrupts(oldirqstate);
   3140 
   3141 	block_userspace_access = 0;
   3142 
   3143  all_done:
   3144 	/*
   3145 	 * The new pmap is resident. Make sure it's marked
   3146 	 * as resident in the cache/TLB.
   3147 	 */
   3148 	npm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
   3149 	if (npm != pmap_kernel())
   3150 		pmap_recent_user = npm;
   3151 
   3152 	/* The old pmap is not longer active */
   3153 	if (opm != NULL)
   3154 		opm->pm_activated = false;
   3155 
   3156 	/* But the new one is */
   3157 	npm->pm_activated = true;
   3158 }
   3159 
   3160 void
   3161 pmap_activate(struct lwp *l)
   3162 {
   3163 
   3164 	if (l == curlwp &&
   3165 	    l->l_proc->p_vmspace->vm_map.pmap->pm_activated == false)
   3166 		pmap_switch(NULL, l);
   3167 }
   3168 
   3169 void
   3170 pmap_deactivate(struct lwp *l)
   3171 {
   3172 
   3173 	l->l_proc->p_vmspace->vm_map.pmap->pm_activated = false;
   3174 }
   3175 
   3176 void
   3177 pmap_update(pmap_t pm)
   3178 {
   3179 
   3180 	if (pm->pm_remove_all) {
   3181 		/*
   3182 		 * Finish up the pmap_remove_all() optimisation by flushing
   3183 		 * the TLB.
   3184 		 */
   3185 		pmap_tlb_flushID(pm);
   3186 		pm->pm_remove_all = false;
   3187 	}
   3188 
   3189 	if (pmap_is_current(pm)) {
   3190 		/*
   3191 		 * If we're dealing with a current userland pmap, move its L1
   3192 		 * to the end of the LRU.
   3193 		 */
   3194 		if (pm != pmap_kernel())
   3195 			pmap_use_l1(pm);
   3196 
   3197 		/*
   3198 		 * We can assume we're done with frobbing the cache/tlb for
   3199 		 * now. Make sure any future pmap ops don't skip cache/tlb
   3200 		 * flushes.
   3201 		 */
   3202 		pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
   3203 	}
   3204 
   3205 	/*
   3206 	 * make sure TLB/cache operations have completed.
   3207 	 */
   3208 	cpu_cpwait();
   3209 }
   3210 
   3211 void
   3212 pmap_remove_all(pmap_t pm)
   3213 {
   3214 
   3215 	/*
   3216 	 * The vmspace described by this pmap is about to be torn down.
   3217 	 * Until pmap_update() is called, UVM will only make calls
   3218 	 * to pmap_remove(). We can make life much simpler by flushing
   3219 	 * the cache now, and deferring TLB invalidation to pmap_update().
   3220 	 */
   3221 	pmap_idcache_wbinv_all(pm);
   3222 	pm->pm_remove_all = true;
   3223 }
   3224 
   3225 /*
   3226  * Retire the given physical map from service.
   3227  * Should only be called if the map contains no valid mappings.
   3228  */
   3229 void
   3230 pmap_destroy(pmap_t pm)
   3231 {
   3232 	u_int count;
   3233 
   3234 	if (pm == NULL)
   3235 		return;
   3236 
   3237 	if (pm->pm_remove_all) {
   3238 		pmap_tlb_flushID(pm);
   3239 		pm->pm_remove_all = false;
   3240 	}
   3241 
   3242 	/*
   3243 	 * Drop reference count
   3244 	 */
   3245 	simple_lock(&pm->pm_lock);
   3246 	count = --pm->pm_obj.uo_refs;
   3247 	simple_unlock(&pm->pm_lock);
   3248 	if (count > 0) {
   3249 		if (pmap_is_current(pm)) {
   3250 			if (pm != pmap_kernel())
   3251 				pmap_use_l1(pm);
   3252 			pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
   3253 		}
   3254 		return;
   3255 	}
   3256 
   3257 	/*
   3258 	 * reference count is zero, free pmap resources and then free pmap.
   3259 	 */
   3260 
   3261 	if (vector_page < KERNEL_BASE) {
   3262 		KDASSERT(!pmap_is_current(pm));
   3263 
   3264 		/* Remove the vector page mapping */
   3265 		pmap_remove(pm, vector_page, vector_page + PAGE_SIZE);
   3266 		pmap_update(pm);
   3267 	}
   3268 
   3269 	LIST_REMOVE(pm, pm_list);
   3270 
   3271 	pmap_free_l1(pm);
   3272 
   3273 	if (pmap_recent_user == pm)
   3274 		pmap_recent_user = NULL;
   3275 
   3276 	/* return the pmap to the pool */
   3277 	pool_cache_put(&pmap_pmap_cache, pm);
   3278 }
   3279 
   3280 
   3281 /*
   3282  * void pmap_reference(pmap_t pm)
   3283  *
   3284  * Add a reference to the specified pmap.
   3285  */
   3286 void
   3287 pmap_reference(pmap_t pm)
   3288 {
   3289 
   3290 	if (pm == NULL)
   3291 		return;
   3292 
   3293 	pmap_use_l1(pm);
   3294 
   3295 	simple_lock(&pm->pm_lock);
   3296 	pm->pm_obj.uo_refs++;
   3297 	simple_unlock(&pm->pm_lock);
   3298 }
   3299 
   3300 /*
   3301  * pmap_zero_page()
   3302  *
   3303  * Zero a given physical page by mapping it at a page hook point.
   3304  * In doing the zero page op, the page we zero is mapped cachable, as with
   3305  * StrongARM accesses to non-cached pages are non-burst making writing
   3306  * _any_ bulk data very slow.
   3307  */
   3308 #if (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0
   3309 void
   3310 pmap_zero_page_generic(paddr_t phys)
   3311 {
   3312 #ifdef DEBUG
   3313 	struct vm_page *pg = PHYS_TO_VM_PAGE(phys);
   3314 
   3315 	if (pg->mdpage.pvh_list != NULL)
   3316 		panic("pmap_zero_page: page has mappings");
   3317 #endif
   3318 
   3319 	KDASSERT((phys & PGOFSET) == 0);
   3320 
   3321 	/*
   3322 	 * Hook in the page, zero it, and purge the cache for that
   3323 	 * zeroed page. Invalidate the TLB as needed.
   3324 	 */
   3325 	*cdst_pte = L2_S_PROTO | phys |
   3326 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
   3327 	PTE_SYNC(cdst_pte);
   3328 	cpu_tlb_flushD_SE(cdstp);
   3329 	cpu_cpwait();
   3330 	bzero_page(cdstp);
   3331 	cpu_dcache_wbinv_range(cdstp, PAGE_SIZE);
   3332 }
   3333 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0 */
   3334 
   3335 #if ARM_MMU_XSCALE == 1
   3336 void
   3337 pmap_zero_page_xscale(paddr_t phys)
   3338 {
   3339 #ifdef DEBUG
   3340 	struct vm_page *pg = PHYS_TO_VM_PAGE(phys);
   3341 
   3342 	if (pg->mdpage.pvh_list != NULL)
   3343 		panic("pmap_zero_page: page has mappings");
   3344 #endif
   3345 
   3346 	KDASSERT((phys & PGOFSET) == 0);
   3347 
   3348 	/*
   3349 	 * Hook in the page, zero it, and purge the cache for that
   3350 	 * zeroed page. Invalidate the TLB as needed.
   3351 	 */
   3352 	*cdst_pte = L2_S_PROTO | phys |
   3353 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
   3354 	    L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X);	/* mini-data */
   3355 	PTE_SYNC(cdst_pte);
   3356 	cpu_tlb_flushD_SE(cdstp);
   3357 	cpu_cpwait();
   3358 	bzero_page(cdstp);
   3359 	xscale_cache_clean_minidata();
   3360 }
   3361 #endif /* ARM_MMU_XSCALE == 1 */
   3362 
   3363 /* pmap_pageidlezero()
   3364  *
   3365  * The same as above, except that we assume that the page is not
   3366  * mapped.  This means we never have to flush the cache first.  Called
   3367  * from the idle loop.
   3368  */
   3369 bool
   3370 pmap_pageidlezero(paddr_t phys)
   3371 {
   3372 	unsigned int i;
   3373 	int *ptr;
   3374 	bool rv = true;
   3375 #ifdef DEBUG
   3376 	struct vm_page *pg;
   3377 
   3378 	pg = PHYS_TO_VM_PAGE(phys);
   3379 	if (pg->mdpage.pvh_list != NULL)
   3380 		panic("pmap_pageidlezero: page has mappings");
   3381 #endif
   3382 
   3383 	KDASSERT((phys & PGOFSET) == 0);
   3384 
   3385 	/*
   3386 	 * Hook in the page, zero it, and purge the cache for that
   3387 	 * zeroed page. Invalidate the TLB as needed.
   3388 	 */
   3389 	*cdst_pte = L2_S_PROTO | phys |
   3390 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
   3391 	PTE_SYNC(cdst_pte);
   3392 	cpu_tlb_flushD_SE(cdstp);
   3393 	cpu_cpwait();
   3394 
   3395 	for (i = 0, ptr = (int *)cdstp;
   3396 			i < (PAGE_SIZE / sizeof(int)); i++) {
   3397 		if (sched_curcpu_runnable_p()) {
   3398 			/*
   3399 			 * A process has become ready.  Abort now,
   3400 			 * so we don't keep it waiting while we
   3401 			 * do slow memory access to finish this
   3402 			 * page.
   3403 			 */
   3404 			rv = false;
   3405 			break;
   3406 		}
   3407 		*ptr++ = 0;
   3408 	}
   3409 
   3410 	if (rv)
   3411 		/*
   3412 		 * if we aborted we'll rezero this page again later so don't
   3413 		 * purge it unless we finished it
   3414 		 */
   3415 		cpu_dcache_wbinv_range(cdstp, PAGE_SIZE);
   3416 
   3417 	return (rv);
   3418 }
   3419 
   3420 /*
   3421  * pmap_copy_page()
   3422  *
   3423  * Copy one physical page into another, by mapping the pages into
   3424  * hook points. The same comment regarding cachability as in
   3425  * pmap_zero_page also applies here.
   3426  */
   3427 #if (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0
   3428 void
   3429 pmap_copy_page_generic(paddr_t src, paddr_t dst)
   3430 {
   3431 	struct vm_page *src_pg = PHYS_TO_VM_PAGE(src);
   3432 #ifdef DEBUG
   3433 	struct vm_page *dst_pg = PHYS_TO_VM_PAGE(dst);
   3434 
   3435 	if (dst_pg->mdpage.pvh_list != NULL)
   3436 		panic("pmap_copy_page: dst page has mappings");
   3437 #endif
   3438 
   3439 	KDASSERT((src & PGOFSET) == 0);
   3440 	KDASSERT((dst & PGOFSET) == 0);
   3441 
   3442 	/*
   3443 	 * Clean the source page.  Hold the source page's lock for
   3444 	 * the duration of the copy so that no other mappings can
   3445 	 * be created while we have a potentially aliased mapping.
   3446 	 */
   3447 	simple_lock(&src_pg->mdpage.pvh_slock);
   3448 	(void) pmap_clean_page(src_pg->mdpage.pvh_list, true);
   3449 
   3450 	/*
   3451 	 * Map the pages into the page hook points, copy them, and purge
   3452 	 * the cache for the appropriate page. Invalidate the TLB
   3453 	 * as required.
   3454 	 */
   3455 	*csrc_pte = L2_S_PROTO | src |
   3456 	    L2_S_PROT(PTE_KERNEL, VM_PROT_READ) | pte_l2_s_cache_mode;
   3457 	PTE_SYNC(csrc_pte);
   3458 	*cdst_pte = L2_S_PROTO | dst |
   3459 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
   3460 	PTE_SYNC(cdst_pte);
   3461 	cpu_tlb_flushD_SE(csrcp);
   3462 	cpu_tlb_flushD_SE(cdstp);
   3463 	cpu_cpwait();
   3464 	bcopy_page(csrcp, cdstp);
   3465 	cpu_dcache_inv_range(csrcp, PAGE_SIZE);
   3466 	simple_unlock(&src_pg->mdpage.pvh_slock); /* cache is safe again */
   3467 	cpu_dcache_wbinv_range(cdstp, PAGE_SIZE);
   3468 }
   3469 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0 */
   3470 
   3471 #if ARM_MMU_XSCALE == 1
   3472 void
   3473 pmap_copy_page_xscale(paddr_t src, paddr_t dst)
   3474 {
   3475 	struct vm_page *src_pg = PHYS_TO_VM_PAGE(src);
   3476 #ifdef DEBUG
   3477 	struct vm_page *dst_pg = PHYS_TO_VM_PAGE(dst);
   3478 
   3479 	if (dst_pg->mdpage.pvh_list != NULL)
   3480 		panic("pmap_copy_page: dst page has mappings");
   3481 #endif
   3482 
   3483 	KDASSERT((src & PGOFSET) == 0);
   3484 	KDASSERT((dst & PGOFSET) == 0);
   3485 
   3486 	/*
   3487 	 * Clean the source page.  Hold the source page's lock for
   3488 	 * the duration of the copy so that no other mappings can
   3489 	 * be created while we have a potentially aliased mapping.
   3490 	 */
   3491 	simple_lock(&src_pg->mdpage.pvh_slock);
   3492 	(void) pmap_clean_page(src_pg->mdpage.pvh_list, true);
   3493 
   3494 	/*
   3495 	 * Map the pages into the page hook points, copy them, and purge
   3496 	 * the cache for the appropriate page. Invalidate the TLB
   3497 	 * as required.
   3498 	 */
   3499 	*csrc_pte = L2_S_PROTO | src |
   3500 	    L2_S_PROT(PTE_KERNEL, VM_PROT_READ) |
   3501 	    L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X);	/* mini-data */
   3502 	PTE_SYNC(csrc_pte);
   3503 	*cdst_pte = L2_S_PROTO | dst |
   3504 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
   3505 	    L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X);	/* mini-data */
   3506 	PTE_SYNC(cdst_pte);
   3507 	cpu_tlb_flushD_SE(csrcp);
   3508 	cpu_tlb_flushD_SE(cdstp);
   3509 	cpu_cpwait();
   3510 	bcopy_page(csrcp, cdstp);
   3511 	simple_unlock(&src_pg->mdpage.pvh_slock); /* cache is safe again */
   3512 	xscale_cache_clean_minidata();
   3513 }
   3514 #endif /* ARM_MMU_XSCALE == 1 */
   3515 
   3516 /*
   3517  * void pmap_virtual_space(vaddr_t *start, vaddr_t *end)
   3518  *
   3519  * Return the start and end addresses of the kernel's virtual space.
   3520  * These values are setup in pmap_bootstrap and are updated as pages
   3521  * are allocated.
   3522  */
   3523 void
   3524 pmap_virtual_space(vaddr_t *start, vaddr_t *end)
   3525 {
   3526 	*start = virtual_avail;
   3527 	*end = virtual_end;
   3528 }
   3529 
   3530 /*
   3531  * Helper function for pmap_grow_l2_bucket()
   3532  */
   3533 static inline int
   3534 pmap_grow_map(vaddr_t va, pt_entry_t cache_mode, paddr_t *pap)
   3535 {
   3536 	struct l2_bucket *l2b;
   3537 	pt_entry_t *ptep;
   3538 	paddr_t pa;
   3539 
   3540 	if (uvm.page_init_done == false) {
   3541 		if (uvm_page_physget(&pa) == false)
   3542 			return (1);
   3543 	} else {
   3544 		struct vm_page *pg;
   3545 		pg = uvm_pagealloc(NULL, 0, NULL, UVM_PGA_USERESERVE);
   3546 		if (pg == NULL)
   3547 			return (1);
   3548 		pa = VM_PAGE_TO_PHYS(pg);
   3549 	}
   3550 
   3551 	if (pap)
   3552 		*pap = pa;
   3553 
   3554 	l2b = pmap_get_l2_bucket(pmap_kernel(), va);
   3555 	KDASSERT(l2b != NULL);
   3556 
   3557 	ptep = &l2b->l2b_kva[l2pte_index(va)];
   3558 	*ptep = L2_S_PROTO | pa | cache_mode |
   3559 	    L2_S_PROT(PTE_KERNEL, VM_PROT_READ | VM_PROT_WRITE);
   3560 	PTE_SYNC(ptep);
   3561 	memset((void *)va, 0, PAGE_SIZE);
   3562 	return (0);
   3563 }
   3564 
   3565 /*
   3566  * This is the same as pmap_alloc_l2_bucket(), except that it is only
   3567  * used by pmap_growkernel().
   3568  */
   3569 static inline struct l2_bucket *
   3570 pmap_grow_l2_bucket(pmap_t pm, vaddr_t va)
   3571 {
   3572 	struct l2_dtable *l2;
   3573 	struct l2_bucket *l2b;
   3574 	u_short l1idx;
   3575 	vaddr_t nva;
   3576 
   3577 	l1idx = L1_IDX(va);
   3578 
   3579 	if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL) {
   3580 		/*
   3581 		 * No mapping at this address, as there is
   3582 		 * no entry in the L1 table.
   3583 		 * Need to allocate a new l2_dtable.
   3584 		 */
   3585 		nva = pmap_kernel_l2dtable_kva;
   3586 		if ((nva & PGOFSET) == 0) {
   3587 			/*
   3588 			 * Need to allocate a backing page
   3589 			 */
   3590 			if (pmap_grow_map(nva, pte_l2_s_cache_mode, NULL))
   3591 				return (NULL);
   3592 		}
   3593 
   3594 		l2 = (struct l2_dtable *)nva;
   3595 		nva += sizeof(struct l2_dtable);
   3596 
   3597 		if ((nva & PGOFSET) < (pmap_kernel_l2dtable_kva & PGOFSET)) {
   3598 			/*
   3599 			 * The new l2_dtable straddles a page boundary.
   3600 			 * Map in another page to cover it.
   3601 			 */
   3602 			if (pmap_grow_map(nva, pte_l2_s_cache_mode, NULL))
   3603 				return (NULL);
   3604 		}
   3605 
   3606 		pmap_kernel_l2dtable_kva = nva;
   3607 
   3608 		/*
   3609 		 * Link it into the parent pmap
   3610 		 */
   3611 		pm->pm_l2[L2_IDX(l1idx)] = l2;
   3612 	}
   3613 
   3614 	l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
   3615 
   3616 	/*
   3617 	 * Fetch pointer to the L2 page table associated with the address.
   3618 	 */
   3619 	if (l2b->l2b_kva == NULL) {
   3620 		pt_entry_t *ptep;
   3621 
   3622 		/*
   3623 		 * No L2 page table has been allocated. Chances are, this
   3624 		 * is because we just allocated the l2_dtable, above.
   3625 		 */
   3626 		nva = pmap_kernel_l2ptp_kva;
   3627 		ptep = (pt_entry_t *)nva;
   3628 		if ((nva & PGOFSET) == 0) {
   3629 			/*
   3630 			 * Need to allocate a backing page
   3631 			 */
   3632 			if (pmap_grow_map(nva, pte_l2_s_cache_mode_pt,
   3633 			    &pmap_kernel_l2ptp_phys))
   3634 				return (NULL);
   3635 			PTE_SYNC_RANGE(ptep, PAGE_SIZE / sizeof(pt_entry_t));
   3636 		}
   3637 
   3638 		l2->l2_occupancy++;
   3639 		l2b->l2b_kva = ptep;
   3640 		l2b->l2b_l1idx = l1idx;
   3641 		l2b->l2b_phys = pmap_kernel_l2ptp_phys;
   3642 
   3643 		pmap_kernel_l2ptp_kva += L2_TABLE_SIZE_REAL;
   3644 		pmap_kernel_l2ptp_phys += L2_TABLE_SIZE_REAL;
   3645 	}
   3646 
   3647 	return (l2b);
   3648 }
   3649 
   3650 vaddr_t
   3651 pmap_growkernel(vaddr_t maxkvaddr)
   3652 {
   3653 	pmap_t kpm = pmap_kernel();
   3654 	struct l1_ttable *l1;
   3655 	struct l2_bucket *l2b;
   3656 	pd_entry_t *pl1pd;
   3657 	int s;
   3658 
   3659 	if (maxkvaddr <= pmap_curmaxkvaddr)
   3660 		goto out;		/* we are OK */
   3661 
   3662 	NPDEBUG(PDB_GROWKERN,
   3663 	    printf("pmap_growkernel: growing kernel from 0x%lx to 0x%lx\n",
   3664 	    pmap_curmaxkvaddr, maxkvaddr));
   3665 
   3666 	KDASSERT(maxkvaddr <= virtual_end);
   3667 
   3668 	/*
   3669 	 * whoops!   we need to add kernel PTPs
   3670 	 */
   3671 
   3672 	s = splhigh();	/* to be safe */
   3673 	simple_lock(&kpm->pm_lock);
   3674 
   3675 	/* Map 1MB at a time */
   3676 	for (; pmap_curmaxkvaddr < maxkvaddr; pmap_curmaxkvaddr += L1_S_SIZE) {
   3677 
   3678 		l2b = pmap_grow_l2_bucket(kpm, pmap_curmaxkvaddr);
   3679 		KDASSERT(l2b != NULL);
   3680 
   3681 		/* Distribute new L1 entry to all other L1s */
   3682 		SLIST_FOREACH(l1, &l1_list, l1_link) {
   3683 			pl1pd = &l1->l1_kva[L1_IDX(pmap_curmaxkvaddr)];
   3684 			*pl1pd = l2b->l2b_phys | L1_C_DOM(PMAP_DOMAIN_KERNEL) |
   3685 			    L1_C_PROTO;
   3686 			PTE_SYNC(pl1pd);
   3687 		}
   3688 	}
   3689 
   3690 	/*
   3691 	 * flush out the cache, expensive but growkernel will happen so
   3692 	 * rarely
   3693 	 */
   3694 	cpu_dcache_wbinv_all();
   3695 	cpu_tlb_flushD();
   3696 	cpu_cpwait();
   3697 
   3698 	simple_unlock(&kpm->pm_lock);
   3699 	splx(s);
   3700 
   3701 out:
   3702 	return (pmap_curmaxkvaddr);
   3703 }
   3704 
   3705 /************************ Utility routines ****************************/
   3706 
   3707 /*
   3708  * vector_page_setprot:
   3709  *
   3710  *	Manipulate the protection of the vector page.
   3711  */
   3712 void
   3713 vector_page_setprot(int prot)
   3714 {
   3715 	struct l2_bucket *l2b;
   3716 	pt_entry_t *ptep;
   3717 
   3718 	l2b = pmap_get_l2_bucket(pmap_kernel(), vector_page);
   3719 	KDASSERT(l2b != NULL);
   3720 
   3721 	ptep = &l2b->l2b_kva[l2pte_index(vector_page)];
   3722 
   3723 	*ptep = (*ptep & ~L1_S_PROT_MASK) | L2_S_PROT(PTE_KERNEL, prot);
   3724 	PTE_SYNC(ptep);
   3725 	cpu_tlb_flushD_SE(vector_page);
   3726 	cpu_cpwait();
   3727 }
   3728 
   3729 /*
   3730  * Fetch pointers to the PDE/PTE for the given pmap/VA pair.
   3731  * Returns true if the mapping exists, else false.
   3732  *
   3733  * NOTE: This function is only used by a couple of arm-specific modules.
   3734  * It is not safe to take any pmap locks here, since we could be right
   3735  * in the middle of debugging the pmap anyway...
   3736  *
   3737  * It is possible for this routine to return false even though a valid
   3738  * mapping does exist. This is because we don't lock, so the metadata
   3739  * state may be inconsistent.
   3740  *
   3741  * NOTE: We can return a NULL *ptp in the case where the L1 pde is
   3742  * a "section" mapping.
   3743  */
   3744 bool
   3745 pmap_get_pde_pte(pmap_t pm, vaddr_t va, pd_entry_t **pdp, pt_entry_t **ptp)
   3746 {
   3747 	struct l2_dtable *l2;
   3748 	pd_entry_t *pl1pd, l1pd;
   3749 	pt_entry_t *ptep;
   3750 	u_short l1idx;
   3751 
   3752 	if (pm->pm_l1 == NULL)
   3753 		return (false);
   3754 
   3755 	l1idx = L1_IDX(va);
   3756 	*pdp = pl1pd = &pm->pm_l1->l1_kva[l1idx];
   3757 	l1pd = *pl1pd;
   3758 
   3759 	if (l1pte_section_p(l1pd)) {
   3760 		*ptp = NULL;
   3761 		return (true);
   3762 	}
   3763 
   3764 	if (pm->pm_l2 == NULL)
   3765 		return (false);
   3766 
   3767 	l2 = pm->pm_l2[L2_IDX(l1idx)];
   3768 
   3769 	if (l2 == NULL ||
   3770 	    (ptep = l2->l2_bucket[L2_BUCKET(l1idx)].l2b_kva) == NULL) {
   3771 		return (false);
   3772 	}
   3773 
   3774 	*ptp = &ptep[l2pte_index(va)];
   3775 	return (true);
   3776 }
   3777 
   3778 bool
   3779 pmap_get_pde(pmap_t pm, vaddr_t va, pd_entry_t **pdp)
   3780 {
   3781 	u_short l1idx;
   3782 
   3783 	if (pm->pm_l1 == NULL)
   3784 		return (false);
   3785 
   3786 	l1idx = L1_IDX(va);
   3787 	*pdp = &pm->pm_l1->l1_kva[l1idx];
   3788 
   3789 	return (true);
   3790 }
   3791 
   3792 /************************ Bootstrapping routines ****************************/
   3793 
   3794 static void
   3795 pmap_init_l1(struct l1_ttable *l1, pd_entry_t *l1pt)
   3796 {
   3797 	int i;
   3798 
   3799 	l1->l1_kva = l1pt;
   3800 	l1->l1_domain_use_count = 0;
   3801 	l1->l1_domain_first = 0;
   3802 
   3803 	for (i = 0; i < PMAP_DOMAINS; i++)
   3804 		l1->l1_domain_free[i] = i + 1;
   3805 
   3806 	/*
   3807 	 * Copy the kernel's L1 entries to each new L1.
   3808 	 */
   3809 	if (pmap_initialized)
   3810 		memcpy(l1pt, pmap_kernel()->pm_l1->l1_kva, L1_TABLE_SIZE);
   3811 
   3812 	if (pmap_extract(pmap_kernel(), (vaddr_t)l1pt,
   3813 	    &l1->l1_physaddr) == false)
   3814 		panic("pmap_init_l1: can't get PA of L1 at %p", l1pt);
   3815 
   3816 	SLIST_INSERT_HEAD(&l1_list, l1, l1_link);
   3817 	TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
   3818 }
   3819 
   3820 /*
   3821  * pmap_bootstrap() is called from the board-specific initarm() routine
   3822  * once the kernel L1/L2 descriptors tables have been set up.
   3823  *
   3824  * This is a somewhat convoluted process since pmap bootstrap is, effectively,
   3825  * spread over a number of disparate files/functions.
   3826  *
   3827  * We are passed the following parameters
   3828  *  - kernel_l1pt
   3829  *    This is a pointer to the base of the kernel's L1 translation table.
   3830  *  - vstart
   3831  *    1MB-aligned start of managed kernel virtual memory.
   3832  *  - vend
   3833  *    1MB-aligned end of managed kernel virtual memory.
   3834  *
   3835  * We use the first parameter to build the metadata (struct l1_ttable and
   3836  * struct l2_dtable) necessary to track kernel mappings.
   3837  */
   3838 #define	PMAP_STATIC_L2_SIZE 16
   3839 void
   3840 pmap_bootstrap(pd_entry_t *kernel_l1pt, vaddr_t vstart, vaddr_t vend)
   3841 {
   3842 	static struct l1_ttable static_l1;
   3843 	static struct l2_dtable static_l2[PMAP_STATIC_L2_SIZE];
   3844 	struct l1_ttable *l1 = &static_l1;
   3845 	struct l2_dtable *l2;
   3846 	struct l2_bucket *l2b;
   3847 	pmap_t pm = pmap_kernel();
   3848 	pd_entry_t pde;
   3849 	pt_entry_t *ptep;
   3850 	paddr_t pa;
   3851 	vaddr_t va;
   3852 	vsize_t size;
   3853 	int l1idx, l2idx, l2next = 0;
   3854 
   3855 	/*
   3856 	 * Initialise the kernel pmap object
   3857 	 */
   3858 	pm->pm_l1 = l1;
   3859 	pm->pm_domain = PMAP_DOMAIN_KERNEL;
   3860 	pm->pm_activated = true;
   3861 	pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
   3862 	simple_lock_init(&pm->pm_lock);
   3863 	pm->pm_obj.pgops = NULL;
   3864 	TAILQ_INIT(&pm->pm_obj.memq);
   3865 	pm->pm_obj.uo_npages = 0;
   3866 	pm->pm_obj.uo_refs = 1;
   3867 
   3868 	/*
   3869 	 * Scan the L1 translation table created by initarm() and create
   3870 	 * the required metadata for all valid mappings found in it.
   3871 	 */
   3872 	for (l1idx = 0; l1idx < (L1_TABLE_SIZE / sizeof(pd_entry_t)); l1idx++) {
   3873 		pde = kernel_l1pt[l1idx];
   3874 
   3875 		/*
   3876 		 * We're only interested in Coarse mappings.
   3877 		 * pmap_extract() can deal with section mappings without
   3878 		 * recourse to checking L2 metadata.
   3879 		 */
   3880 		if ((pde & L1_TYPE_MASK) != L1_TYPE_C)
   3881 			continue;
   3882 
   3883 		/*
   3884 		 * Lookup the KVA of this L2 descriptor table
   3885 		 */
   3886 		pa = (paddr_t)(pde & L1_C_ADDR_MASK);
   3887 		ptep = (pt_entry_t *)kernel_pt_lookup(pa);
   3888 		if (ptep == NULL) {
   3889 			panic("pmap_bootstrap: No L2 for va 0x%x, pa 0x%lx",
   3890 			    (u_int)l1idx << L1_S_SHIFT, pa);
   3891 		}
   3892 
   3893 		/*
   3894 		 * Fetch the associated L2 metadata structure.
   3895 		 * Allocate a new one if necessary.
   3896 		 */
   3897 		if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL) {
   3898 			if (l2next == PMAP_STATIC_L2_SIZE)
   3899 				panic("pmap_bootstrap: out of static L2s");
   3900 			pm->pm_l2[L2_IDX(l1idx)] = l2 = &static_l2[l2next++];
   3901 		}
   3902 
   3903 		/*
   3904 		 * One more L1 slot tracked...
   3905 		 */
   3906 		l2->l2_occupancy++;
   3907 
   3908 		/*
   3909 		 * Fill in the details of the L2 descriptor in the
   3910 		 * appropriate bucket.
   3911 		 */
   3912 		l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
   3913 		l2b->l2b_kva = ptep;
   3914 		l2b->l2b_phys = pa;
   3915 		l2b->l2b_l1idx = l1idx;
   3916 
   3917 		/*
   3918 		 * Establish an initial occupancy count for this descriptor
   3919 		 */
   3920 		for (l2idx = 0;
   3921 		    l2idx < (L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
   3922 		    l2idx++) {
   3923 			if ((ptep[l2idx] & L2_TYPE_MASK) != L2_TYPE_INV) {
   3924 				l2b->l2b_occupancy++;
   3925 			}
   3926 		}
   3927 
   3928 		/*
   3929 		 * Make sure the descriptor itself has the correct cache mode.
   3930 		 * If not, fix it, but whine about the problem. Port-meisters
   3931 		 * should consider this a clue to fix up their initarm()
   3932 		 * function. :)
   3933 		 */
   3934 		if (pmap_set_pt_cache_mode(kernel_l1pt, (vaddr_t)ptep)) {
   3935 			printf("pmap_bootstrap: WARNING! wrong cache mode for "
   3936 			    "L2 pte @ %p\n", ptep);
   3937 		}
   3938 	}
   3939 
   3940 	/*
   3941 	 * Ensure the primary (kernel) L1 has the correct cache mode for
   3942 	 * a page table. Bitch if it is not correctly set.
   3943 	 */
   3944 	for (va = (vaddr_t)kernel_l1pt;
   3945 	    va < ((vaddr_t)kernel_l1pt + L1_TABLE_SIZE); va += PAGE_SIZE) {
   3946 		if (pmap_set_pt_cache_mode(kernel_l1pt, va))
   3947 			printf("pmap_bootstrap: WARNING! wrong cache mode for "
   3948 			    "primary L1 @ 0x%lx\n", va);
   3949 	}
   3950 
   3951 	cpu_dcache_wbinv_all();
   3952 	cpu_tlb_flushID();
   3953 	cpu_cpwait();
   3954 
   3955 	/*
   3956 	 * now we allocate the "special" VAs which are used for tmp mappings
   3957 	 * by the pmap (and other modules).  we allocate the VAs by advancing
   3958 	 * virtual_avail (note that there are no pages mapped at these VAs).
   3959 	 *
   3960 	 * Managed KVM space start from wherever initarm() tells us.
   3961 	 */
   3962 	virtual_avail = vstart;
   3963 	virtual_end = vend;
   3964 
   3965 	pmap_alloc_specials(&virtual_avail, 1, &csrcp, &csrc_pte);
   3966 	pmap_set_pt_cache_mode(kernel_l1pt, (vaddr_t)csrc_pte);
   3967 	pmap_alloc_specials(&virtual_avail, 1, &cdstp, &cdst_pte);
   3968 	pmap_set_pt_cache_mode(kernel_l1pt, (vaddr_t)cdst_pte);
   3969 	pmap_alloc_specials(&virtual_avail, 1, (void *)&memhook, NULL);
   3970 	pmap_alloc_specials(&virtual_avail, round_page(MSGBUFSIZE) / PAGE_SIZE,
   3971 	    (void *)&msgbufaddr, NULL);
   3972 
   3973 	/*
   3974 	 * Allocate a range of kernel virtual address space to be used
   3975 	 * for L2 descriptor tables and metadata allocation in
   3976 	 * pmap_growkernel().
   3977 	 */
   3978 	size = ((virtual_end - pmap_curmaxkvaddr) + L1_S_OFFSET) / L1_S_SIZE;
   3979 	pmap_alloc_specials(&virtual_avail,
   3980 	    round_page(size * L2_TABLE_SIZE_REAL) / PAGE_SIZE,
   3981 	    &pmap_kernel_l2ptp_kva, NULL);
   3982 
   3983 	size = (size + (L2_BUCKET_SIZE - 1)) / L2_BUCKET_SIZE;
   3984 	pmap_alloc_specials(&virtual_avail,
   3985 	    round_page(size * sizeof(struct l2_dtable)) / PAGE_SIZE,
   3986 	    &pmap_kernel_l2dtable_kva, NULL);
   3987 
   3988 	/*
   3989 	 * init the static-global locks and global pmap list.
   3990 	 */
   3991 	/* spinlockinit(&pmap_main_lock, "pmaplk", 0); */
   3992 
   3993 	/*
   3994 	 * We can now initialise the first L1's metadata.
   3995 	 */
   3996 	SLIST_INIT(&l1_list);
   3997 	TAILQ_INIT(&l1_lru_list);
   3998 	simple_lock_init(&l1_lru_lock);
   3999 	pmap_init_l1(l1, kernel_l1pt);
   4000 
   4001 	/* Set up vector page L1 details, if necessary */
   4002 	if (vector_page < KERNEL_BASE) {
   4003 		pm->pm_pl1vec = &pm->pm_l1->l1_kva[L1_IDX(vector_page)];
   4004 		l2b = pmap_get_l2_bucket(pm, vector_page);
   4005 		pm->pm_l1vec = l2b->l2b_phys | L1_C_PROTO |
   4006 		    L1_C_DOM(pm->pm_domain);
   4007 	} else
   4008 		pm->pm_pl1vec = NULL;
   4009 
   4010 	/*
   4011 	 * Initialize the pmap pool and cache
   4012 	 */
   4013 	pool_init(&pmap_pmap_pool, sizeof(struct pmap), 0, 0, 0, "pmappl",
   4014 	    &pool_allocator_nointr, IPL_NONE);
   4015 	pool_cache_init(&pmap_pmap_cache, &pmap_pmap_pool,
   4016 	    pmap_pmap_ctor, NULL, NULL);
   4017 	LIST_INIT(&pmap_pmaps);
   4018 	LIST_INSERT_HEAD(&pmap_pmaps, pm, pm_list);
   4019 
   4020 	/*
   4021 	 * Initialize the pv pool.
   4022 	 */
   4023 	pool_init(&pmap_pv_pool, sizeof(struct pv_entry), 0, 0, 0, "pvepl",
   4024 	    &pmap_bootstrap_pv_allocator, IPL_NONE);
   4025 
   4026 	/*
   4027 	 * Initialize the L2 dtable pool and cache.
   4028 	 */
   4029 	pool_init(&pmap_l2dtable_pool, sizeof(struct l2_dtable), 0, 0, 0,
   4030 	    "l2dtblpl", NULL, IPL_NONE);
   4031 	pool_cache_init(&pmap_l2dtable_cache, &pmap_l2dtable_pool,
   4032 	    pmap_l2dtable_ctor, NULL, NULL);
   4033 
   4034 	/*
   4035 	 * Initialise the L2 descriptor table pool and cache
   4036 	 */
   4037 	pool_init(&pmap_l2ptp_pool, L2_TABLE_SIZE_REAL, 0, L2_TABLE_SIZE_REAL,
   4038 	    0, "l2ptppl", NULL, IPL_NONE);
   4039 	pool_cache_init(&pmap_l2ptp_cache, &pmap_l2ptp_pool,
   4040 	    pmap_l2ptp_ctor, NULL, NULL);
   4041 
   4042 	cpu_dcache_wbinv_all();
   4043 }
   4044 
   4045 static int
   4046 pmap_set_pt_cache_mode(pd_entry_t *kl1, vaddr_t va)
   4047 {
   4048 	pd_entry_t *pdep, pde;
   4049 	pt_entry_t *ptep, pte;
   4050 	vaddr_t pa;
   4051 	int rv = 0;
   4052 
   4053 	/*
   4054 	 * Make sure the descriptor itself has the correct cache mode
   4055 	 */
   4056 	pdep = &kl1[L1_IDX(va)];
   4057 	pde = *pdep;
   4058 
   4059 	if (l1pte_section_p(pde)) {
   4060 		if ((pde & L1_S_CACHE_MASK) != pte_l1_s_cache_mode_pt) {
   4061 			*pdep = (pde & ~L1_S_CACHE_MASK) |
   4062 			    pte_l1_s_cache_mode_pt;
   4063 			PTE_SYNC(pdep);
   4064 			cpu_dcache_wbinv_range((vaddr_t)pdep, sizeof(*pdep));
   4065 			rv = 1;
   4066 		}
   4067 	} else {
   4068 		pa = (paddr_t)(pde & L1_C_ADDR_MASK);
   4069 		ptep = (pt_entry_t *)kernel_pt_lookup(pa);
   4070 		if (ptep == NULL)
   4071 			panic("pmap_bootstrap: No L2 for L2 @ va %p\n", ptep);
   4072 
   4073 		ptep = &ptep[l2pte_index(va)];
   4074 		pte = *ptep;
   4075 		if ((pte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) {
   4076 			*ptep = (pte & ~L2_S_CACHE_MASK) |
   4077 			    pte_l2_s_cache_mode_pt;
   4078 			PTE_SYNC(ptep);
   4079 			cpu_dcache_wbinv_range((vaddr_t)ptep, sizeof(*ptep));
   4080 			rv = 1;
   4081 		}
   4082 	}
   4083 
   4084 	return (rv);
   4085 }
   4086 
   4087 static void
   4088 pmap_alloc_specials(vaddr_t *availp, int pages, vaddr_t *vap, pt_entry_t **ptep)
   4089 {
   4090 	vaddr_t va = *availp;
   4091 	struct l2_bucket *l2b;
   4092 
   4093 	if (ptep) {
   4094 		l2b = pmap_get_l2_bucket(pmap_kernel(), va);
   4095 		if (l2b == NULL)
   4096 			panic("pmap_alloc_specials: no l2b for 0x%lx", va);
   4097 
   4098 		if (ptep)
   4099 			*ptep = &l2b->l2b_kva[l2pte_index(va)];
   4100 	}
   4101 
   4102 	*vap = va;
   4103 	*availp = va + (PAGE_SIZE * pages);
   4104 }
   4105 
   4106 void
   4107 pmap_init(void)
   4108 {
   4109 	extern int physmem;
   4110 
   4111 	/*
   4112 	 * Set the available memory vars - These do not map to real memory
   4113 	 * addresses and cannot as the physical memory is fragmented.
   4114 	 * They are used by ps for %mem calculations.
   4115 	 * One could argue whether this should be the entire memory or just
   4116 	 * the memory that is useable in a user process.
   4117 	 */
   4118 	avail_start = 0;
   4119 	avail_end = physmem * PAGE_SIZE;
   4120 
   4121 	/*
   4122 	 * Now we need to free enough pv_entry structures to allow us to get
   4123 	 * the kmem_map/kmem_object allocated and inited (done after this
   4124 	 * function is finished).  to do this we allocate one bootstrap page out
   4125 	 * of kernel_map and use it to provide an initial pool of pv_entry
   4126 	 * structures.   we never free this page.
   4127 	 */
   4128 	pool_setlowat(&pmap_pv_pool,
   4129 	    (PAGE_SIZE / sizeof(struct pv_entry)) * 2);
   4130 
   4131 	pmap_initialized = true;
   4132 }
   4133 
   4134 static vaddr_t last_bootstrap_page = 0;
   4135 static void *free_bootstrap_pages = NULL;
   4136 
   4137 static void *
   4138 pmap_bootstrap_pv_page_alloc(struct pool *pp, int flags)
   4139 {
   4140 	extern void *pool_page_alloc(struct pool *, int);
   4141 	vaddr_t new_page;
   4142 	void *rv;
   4143 
   4144 	if (pmap_initialized)
   4145 		return (pool_page_alloc(pp, flags));
   4146 
   4147 	if (free_bootstrap_pages) {
   4148 		rv = free_bootstrap_pages;
   4149 		free_bootstrap_pages = *((void **)rv);
   4150 		return (rv);
   4151 	}
   4152 
   4153 	new_page = uvm_km_alloc(kernel_map, PAGE_SIZE, 0,
   4154 	    UVM_KMF_WIRED | ((flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT));
   4155 
   4156 	KASSERT(new_page > last_bootstrap_page);
   4157 	last_bootstrap_page = new_page;
   4158 	return ((void *)new_page);
   4159 }
   4160 
   4161 static void
   4162 pmap_bootstrap_pv_page_free(struct pool *pp, void *v)
   4163 {
   4164 	extern void pool_page_free(struct pool *, void *);
   4165 
   4166 	if ((vaddr_t)v <= last_bootstrap_page) {
   4167 		*((void **)v) = free_bootstrap_pages;
   4168 		free_bootstrap_pages = v;
   4169 		return;
   4170 	}
   4171 
   4172 	if (pmap_initialized) {
   4173 		pool_page_free(pp, v);
   4174 		return;
   4175 	}
   4176 }
   4177 
   4178 /*
   4179  * pmap_postinit()
   4180  *
   4181  * This routine is called after the vm and kmem subsystems have been
   4182  * initialised. This allows the pmap code to perform any initialisation
   4183  * that can only be done one the memory allocation is in place.
   4184  */
   4185 void
   4186 pmap_postinit(void)
   4187 {
   4188 	extern paddr_t physical_start, physical_end;
   4189 	struct l2_bucket *l2b;
   4190 	struct l1_ttable *l1;
   4191 	struct pglist plist;
   4192 	struct vm_page *m;
   4193 	pd_entry_t *pl1pt;
   4194 	pt_entry_t *ptep, pte;
   4195 	vaddr_t va, eva;
   4196 	u_int loop, needed;
   4197 	int error;
   4198 
   4199 	pool_setlowat(&pmap_l2ptp_pool,
   4200 	    (PAGE_SIZE / L2_TABLE_SIZE_REAL) * 4);
   4201 	pool_setlowat(&pmap_l2dtable_pool,
   4202 	    (PAGE_SIZE / sizeof(struct l2_dtable)) * 2);
   4203 
   4204 	needed = (maxproc / PMAP_DOMAINS) + ((maxproc % PMAP_DOMAINS) ? 1 : 0);
   4205 	needed -= 1;
   4206 
   4207 	l1 = malloc(sizeof(*l1) * needed, M_VMPMAP, M_WAITOK);
   4208 
   4209 	for (loop = 0; loop < needed; loop++, l1++) {
   4210 		/* Allocate a L1 page table */
   4211 		va = uvm_km_alloc(kernel_map, L1_TABLE_SIZE, 0, UVM_KMF_VAONLY);
   4212 		if (va == 0)
   4213 			panic("Cannot allocate L1 KVM");
   4214 
   4215 		error = uvm_pglistalloc(L1_TABLE_SIZE, physical_start,
   4216 		    physical_end, L1_TABLE_SIZE, 0, &plist, 1, M_WAITOK);
   4217 		if (error)
   4218 			panic("Cannot allocate L1 physical pages");
   4219 
   4220 		m = TAILQ_FIRST(&plist);
   4221 		eva = va + L1_TABLE_SIZE;
   4222 		pl1pt = (pd_entry_t *)va;
   4223 
   4224 		while (m && va < eva) {
   4225 			paddr_t pa = VM_PAGE_TO_PHYS(m);
   4226 
   4227 			pmap_kenter_pa(va, pa, VM_PROT_READ | VM_PROT_WRITE);
   4228 
   4229 			/*
   4230 			 * Make sure the L1 descriptor table is mapped
   4231 			 * with the cache-mode set to write-through.
   4232 			 */
   4233 			l2b = pmap_get_l2_bucket(pmap_kernel(), va);
   4234 			ptep = &l2b->l2b_kva[l2pte_index(va)];
   4235 			pte = *ptep;
   4236 			pte = (pte & ~L2_S_CACHE_MASK) | pte_l2_s_cache_mode_pt;
   4237 			*ptep = pte;
   4238 			PTE_SYNC(ptep);
   4239 			cpu_tlb_flushD_SE(va);
   4240 
   4241 			va += PAGE_SIZE;
   4242 			m = TAILQ_NEXT(m, pageq);
   4243 		}
   4244 
   4245 #ifdef DIAGNOSTIC
   4246 		if (m)
   4247 			panic("pmap_alloc_l1pt: pglist not empty");
   4248 #endif	/* DIAGNOSTIC */
   4249 
   4250 		pmap_init_l1(l1, pl1pt);
   4251 	}
   4252 
   4253 #ifdef DEBUG
   4254 	printf("pmap_postinit: Allocated %d static L1 descriptor tables\n",
   4255 	    needed);
   4256 #endif
   4257 }
   4258 
   4259 /*
   4260  * Note that the following routines are used by board-specific initialisation
   4261  * code to configure the initial kernel page tables.
   4262  *
   4263  * If ARM32_NEW_VM_LAYOUT is *not* defined, they operate on the assumption that
   4264  * L2 page-table pages are 4KB in size and use 4 L1 slots. This mimics the
   4265  * behaviour of the old pmap, and provides an easy migration path for
   4266  * initial bring-up of the new pmap on existing ports. Fortunately,
   4267  * pmap_bootstrap() compensates for this hackery. This is only a stop-gap and
   4268  * will be deprecated.
   4269  *
   4270  * If ARM32_NEW_VM_LAYOUT *is* defined, these functions deal with 1KB L2 page
   4271  * tables.
   4272  */
   4273 
   4274 /*
   4275  * This list exists for the benefit of pmap_map_chunk().  It keeps track
   4276  * of the kernel L2 tables during bootstrap, so that pmap_map_chunk() can
   4277  * find them as necessary.
   4278  *
   4279  * Note that the data on this list MUST remain valid after initarm() returns,
   4280  * as pmap_bootstrap() uses it to contruct L2 table metadata.
   4281  */
   4282 SLIST_HEAD(, pv_addr) kernel_pt_list = SLIST_HEAD_INITIALIZER(kernel_pt_list);
   4283 
   4284 static vaddr_t
   4285 kernel_pt_lookup(paddr_t pa)
   4286 {
   4287 	pv_addr_t *pv;
   4288 
   4289 	SLIST_FOREACH(pv, &kernel_pt_list, pv_list) {
   4290 #ifndef ARM32_NEW_VM_LAYOUT
   4291 		if (pv->pv_pa == (pa & ~PGOFSET))
   4292 			return (pv->pv_va | (pa & PGOFSET));
   4293 #else
   4294 		if (pv->pv_pa == pa)
   4295 			return (pv->pv_va);
   4296 #endif
   4297 	}
   4298 	return (0);
   4299 }
   4300 
   4301 /*
   4302  * pmap_map_section:
   4303  *
   4304  *	Create a single section mapping.
   4305  */
   4306 void
   4307 pmap_map_section(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
   4308 {
   4309 	pd_entry_t *pde = (pd_entry_t *) l1pt;
   4310 	pd_entry_t fl;
   4311 
   4312 	KASSERT(((va | pa) & L1_S_OFFSET) == 0);
   4313 
   4314 	switch (cache) {
   4315 	case PTE_NOCACHE:
   4316 	default:
   4317 		fl = 0;
   4318 		break;
   4319 
   4320 	case PTE_CACHE:
   4321 		fl = pte_l1_s_cache_mode;
   4322 		break;
   4323 
   4324 	case PTE_PAGETABLE:
   4325 		fl = pte_l1_s_cache_mode_pt;
   4326 		break;
   4327 	}
   4328 
   4329 	pde[va >> L1_S_SHIFT] = L1_S_PROTO | pa |
   4330 	    L1_S_PROT(PTE_KERNEL, prot) | fl | L1_S_DOM(PMAP_DOMAIN_KERNEL);
   4331 	PTE_SYNC(&pde[va >> L1_S_SHIFT]);
   4332 }
   4333 
   4334 /*
   4335  * pmap_map_entry:
   4336  *
   4337  *	Create a single page mapping.
   4338  */
   4339 void
   4340 pmap_map_entry(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
   4341 {
   4342 	pd_entry_t *pde = (pd_entry_t *) l1pt;
   4343 	pt_entry_t fl;
   4344 	pt_entry_t *pte;
   4345 
   4346 	KASSERT(((va | pa) & PGOFSET) == 0);
   4347 
   4348 	switch (cache) {
   4349 	case PTE_NOCACHE:
   4350 	default:
   4351 		fl = 0;
   4352 		break;
   4353 
   4354 	case PTE_CACHE:
   4355 		fl = pte_l2_s_cache_mode;
   4356 		break;
   4357 
   4358 	case PTE_PAGETABLE:
   4359 		fl = pte_l2_s_cache_mode_pt;
   4360 		break;
   4361 	}
   4362 
   4363 	if ((pde[va >> L1_S_SHIFT] & L1_TYPE_MASK) != L1_TYPE_C)
   4364 		panic("pmap_map_entry: no L2 table for VA 0x%08lx", va);
   4365 
   4366 #ifndef ARM32_NEW_VM_LAYOUT
   4367 	pte = (pt_entry_t *)
   4368 	    kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
   4369 #else
   4370 	pte = (pt_entry_t *) kernel_pt_lookup(pde[L1_IDX(va)] & L1_C_ADDR_MASK);
   4371 #endif
   4372 	if (pte == NULL)
   4373 		panic("pmap_map_entry: can't find L2 table for VA 0x%08lx", va);
   4374 
   4375 #ifndef ARM32_NEW_VM_LAYOUT
   4376 	pte[(va >> PGSHIFT) & 0x3ff] =
   4377 	    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | fl;
   4378 	PTE_SYNC(&pte[(va >> PGSHIFT) & 0x3ff]);
   4379 #else
   4380 	pte[l2pte_index(va)] =
   4381 	    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | fl;
   4382 	PTE_SYNC(&pte[l2pte_index(va)]);
   4383 #endif
   4384 }
   4385 
   4386 /*
   4387  * pmap_link_l2pt:
   4388  *
   4389  *	Link the L2 page table specified by "l2pv" into the L1
   4390  *	page table at the slot for "va".
   4391  */
   4392 void
   4393 pmap_link_l2pt(vaddr_t l1pt, vaddr_t va, pv_addr_t *l2pv)
   4394 {
   4395 	pd_entry_t *pde = (pd_entry_t *) l1pt, proto;
   4396 	u_int slot = va >> L1_S_SHIFT;
   4397 
   4398 #ifndef ARM32_NEW_VM_LAYOUT
   4399 	KASSERT((va & ((L1_S_SIZE * 4) - 1)) == 0);
   4400 	KASSERT((l2pv->pv_pa & PGOFSET) == 0);
   4401 #endif
   4402 
   4403 	proto = L1_S_DOM(PMAP_DOMAIN_KERNEL) | L1_C_PROTO;
   4404 
   4405 	pde[slot + 0] = proto | (l2pv->pv_pa + 0x000);
   4406 #ifdef ARM32_NEW_VM_LAYOUT
   4407 	PTE_SYNC(&pde[slot]);
   4408 #else
   4409 	pde[slot + 1] = proto | (l2pv->pv_pa + 0x400);
   4410 	pde[slot + 2] = proto | (l2pv->pv_pa + 0x800);
   4411 	pde[slot + 3] = proto | (l2pv->pv_pa + 0xc00);
   4412 	PTE_SYNC_RANGE(&pde[slot + 0], 4);
   4413 #endif
   4414 
   4415 	SLIST_INSERT_HEAD(&kernel_pt_list, l2pv, pv_list);
   4416 }
   4417 
   4418 /*
   4419  * pmap_map_chunk:
   4420  *
   4421  *	Map a chunk of memory using the most efficient mappings
   4422  *	possible (section, large page, small page) into the
   4423  *	provided L1 and L2 tables at the specified virtual address.
   4424  */
   4425 vsize_t
   4426 pmap_map_chunk(vaddr_t l1pt, vaddr_t va, paddr_t pa, vsize_t size,
   4427     int prot, int cache)
   4428 {
   4429 	pd_entry_t *pde = (pd_entry_t *) l1pt;
   4430 	pt_entry_t *pte, f1, f2s, f2l;
   4431 	vsize_t resid;
   4432 	int i;
   4433 
   4434 	resid = (size + (PAGE_SIZE - 1)) & ~(PAGE_SIZE - 1);
   4435 
   4436 	if (l1pt == 0)
   4437 		panic("pmap_map_chunk: no L1 table provided");
   4438 
   4439 #ifdef VERBOSE_INIT_ARM
   4440 	printf("pmap_map_chunk: pa=0x%lx va=0x%lx size=0x%lx resid=0x%lx "
   4441 	    "prot=0x%x cache=%d\n", pa, va, size, resid, prot, cache);
   4442 #endif
   4443 
   4444 	switch (cache) {
   4445 	case PTE_NOCACHE:
   4446 	default:
   4447 		f1 = 0;
   4448 		f2l = 0;
   4449 		f2s = 0;
   4450 		break;
   4451 
   4452 	case PTE_CACHE:
   4453 		f1 = pte_l1_s_cache_mode;
   4454 		f2l = pte_l2_l_cache_mode;
   4455 		f2s = pte_l2_s_cache_mode;
   4456 		break;
   4457 
   4458 	case PTE_PAGETABLE:
   4459 		f1 = pte_l1_s_cache_mode_pt;
   4460 		f2l = pte_l2_l_cache_mode_pt;
   4461 		f2s = pte_l2_s_cache_mode_pt;
   4462 		break;
   4463 	}
   4464 
   4465 	size = resid;
   4466 
   4467 	while (resid > 0) {
   4468 		/* See if we can use a section mapping. */
   4469 		if (L1_S_MAPPABLE_P(va, pa, resid)) {
   4470 #ifdef VERBOSE_INIT_ARM
   4471 			printf("S");
   4472 #endif
   4473 			pde[va >> L1_S_SHIFT] = L1_S_PROTO | pa |
   4474 			    L1_S_PROT(PTE_KERNEL, prot) | f1 |
   4475 			    L1_S_DOM(PMAP_DOMAIN_KERNEL);
   4476 			PTE_SYNC(&pde[va >> L1_S_SHIFT]);
   4477 			va += L1_S_SIZE;
   4478 			pa += L1_S_SIZE;
   4479 			resid -= L1_S_SIZE;
   4480 			continue;
   4481 		}
   4482 
   4483 		/*
   4484 		 * Ok, we're going to use an L2 table.  Make sure
   4485 		 * one is actually in the corresponding L1 slot
   4486 		 * for the current VA.
   4487 		 */
   4488 		if ((pde[va >> L1_S_SHIFT] & L1_TYPE_MASK) != L1_TYPE_C)
   4489 			panic("pmap_map_chunk: no L2 table for VA 0x%08lx", va);
   4490 
   4491 #ifndef ARM32_NEW_VM_LAYOUT
   4492 		pte = (pt_entry_t *)
   4493 		    kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
   4494 #else
   4495 		pte = (pt_entry_t *) kernel_pt_lookup(
   4496 		    pde[L1_IDX(va)] & L1_C_ADDR_MASK);
   4497 #endif
   4498 		if (pte == NULL)
   4499 			panic("pmap_map_chunk: can't find L2 table for VA"
   4500 			    "0x%08lx", va);
   4501 
   4502 		/* See if we can use a L2 large page mapping. */
   4503 		if (L2_L_MAPPABLE_P(va, pa, resid)) {
   4504 #ifdef VERBOSE_INIT_ARM
   4505 			printf("L");
   4506 #endif
   4507 			for (i = 0; i < 16; i++) {
   4508 #ifndef ARM32_NEW_VM_LAYOUT
   4509 				pte[((va >> PGSHIFT) & 0x3f0) + i] =
   4510 				    L2_L_PROTO | pa |
   4511 				    L2_L_PROT(PTE_KERNEL, prot) | f2l;
   4512 				PTE_SYNC(&pte[((va >> PGSHIFT) & 0x3f0) + i]);
   4513 #else
   4514 				pte[l2pte_index(va) + i] =
   4515 				    L2_L_PROTO | pa |
   4516 				    L2_L_PROT(PTE_KERNEL, prot) | f2l;
   4517 				PTE_SYNC(&pte[l2pte_index(va) + i]);
   4518 #endif
   4519 			}
   4520 			va += L2_L_SIZE;
   4521 			pa += L2_L_SIZE;
   4522 			resid -= L2_L_SIZE;
   4523 			continue;
   4524 		}
   4525 
   4526 		/* Use a small page mapping. */
   4527 #ifdef VERBOSE_INIT_ARM
   4528 		printf("P");
   4529 #endif
   4530 #ifndef ARM32_NEW_VM_LAYOUT
   4531 		pte[(va >> PGSHIFT) & 0x3ff] =
   4532 		    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | f2s;
   4533 		PTE_SYNC(&pte[(va >> PGSHIFT) & 0x3ff]);
   4534 #else
   4535 		pte[l2pte_index(va)] =
   4536 		    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | f2s;
   4537 		PTE_SYNC(&pte[l2pte_index(va)]);
   4538 #endif
   4539 		va += PAGE_SIZE;
   4540 		pa += PAGE_SIZE;
   4541 		resid -= PAGE_SIZE;
   4542 	}
   4543 #ifdef VERBOSE_INIT_ARM
   4544 	printf("\n");
   4545 #endif
   4546 	return (size);
   4547 }
   4548 
   4549 /********************** Static device map routines ***************************/
   4550 
   4551 static const struct pmap_devmap *pmap_devmap_table;
   4552 
   4553 /*
   4554  * Register the devmap table.  This is provided in case early console
   4555  * initialization needs to register mappings created by bootstrap code
   4556  * before pmap_devmap_bootstrap() is called.
   4557  */
   4558 void
   4559 pmap_devmap_register(const struct pmap_devmap *table)
   4560 {
   4561 
   4562 	pmap_devmap_table = table;
   4563 }
   4564 
   4565 /*
   4566  * Map all of the static regions in the devmap table, and remember
   4567  * the devmap table so other parts of the kernel can look up entries
   4568  * later.
   4569  */
   4570 void
   4571 pmap_devmap_bootstrap(vaddr_t l1pt, const struct pmap_devmap *table)
   4572 {
   4573 	int i;
   4574 
   4575 	pmap_devmap_table = table;
   4576 
   4577 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
   4578 #ifdef VERBOSE_INIT_ARM
   4579 		printf("devmap: %08lx -> %08lx @ %08lx\n",
   4580 		    pmap_devmap_table[i].pd_pa,
   4581 		    pmap_devmap_table[i].pd_pa +
   4582 			pmap_devmap_table[i].pd_size - 1,
   4583 		    pmap_devmap_table[i].pd_va);
   4584 #endif
   4585 		pmap_map_chunk(l1pt, pmap_devmap_table[i].pd_va,
   4586 		    pmap_devmap_table[i].pd_pa,
   4587 		    pmap_devmap_table[i].pd_size,
   4588 		    pmap_devmap_table[i].pd_prot,
   4589 		    pmap_devmap_table[i].pd_cache);
   4590 	}
   4591 }
   4592 
   4593 const struct pmap_devmap *
   4594 pmap_devmap_find_pa(paddr_t pa, psize_t size)
   4595 {
   4596 	uint64_t endpa;
   4597 	int i;
   4598 
   4599 	if (pmap_devmap_table == NULL)
   4600 		return (NULL);
   4601 
   4602 	endpa = (uint64_t)pa + (uint64_t)(size - 1);
   4603 
   4604 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
   4605 		if (pa >= pmap_devmap_table[i].pd_pa &&
   4606 		    endpa <= (uint64_t)pmap_devmap_table[i].pd_pa +
   4607 			     (uint64_t)(pmap_devmap_table[i].pd_size - 1))
   4608 			return (&pmap_devmap_table[i]);
   4609 	}
   4610 
   4611 	return (NULL);
   4612 }
   4613 
   4614 const struct pmap_devmap *
   4615 pmap_devmap_find_va(vaddr_t va, vsize_t size)
   4616 {
   4617 	int i;
   4618 
   4619 	if (pmap_devmap_table == NULL)
   4620 		return (NULL);
   4621 
   4622 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
   4623 		if (va >= pmap_devmap_table[i].pd_va &&
   4624 		    va + size - 1 <= pmap_devmap_table[i].pd_va +
   4625 				     pmap_devmap_table[i].pd_size - 1)
   4626 			return (&pmap_devmap_table[i]);
   4627 	}
   4628 
   4629 	return (NULL);
   4630 }
   4631 
   4632 /********************** PTE initialization routines **************************/
   4633 
   4634 /*
   4635  * These routines are called when the CPU type is identified to set up
   4636  * the PTE prototypes, cache modes, etc.
   4637  *
   4638  * The variables are always here, just in case LKMs need to reference
   4639  * them (though, they shouldn't).
   4640  */
   4641 
   4642 pt_entry_t	pte_l1_s_cache_mode;
   4643 pt_entry_t	pte_l1_s_cache_mode_pt;
   4644 pt_entry_t	pte_l1_s_cache_mask;
   4645 
   4646 pt_entry_t	pte_l2_l_cache_mode;
   4647 pt_entry_t	pte_l2_l_cache_mode_pt;
   4648 pt_entry_t	pte_l2_l_cache_mask;
   4649 
   4650 pt_entry_t	pte_l2_s_cache_mode;
   4651 pt_entry_t	pte_l2_s_cache_mode_pt;
   4652 pt_entry_t	pte_l2_s_cache_mask;
   4653 
   4654 pt_entry_t	pte_l2_s_prot_u;
   4655 pt_entry_t	pte_l2_s_prot_w;
   4656 pt_entry_t	pte_l2_s_prot_mask;
   4657 
   4658 pt_entry_t	pte_l1_s_proto;
   4659 pt_entry_t	pte_l1_c_proto;
   4660 pt_entry_t	pte_l2_s_proto;
   4661 
   4662 void		(*pmap_copy_page_func)(paddr_t, paddr_t);
   4663 void		(*pmap_zero_page_func)(paddr_t);
   4664 
   4665 #if (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0
   4666 void
   4667 pmap_pte_init_generic(void)
   4668 {
   4669 
   4670 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
   4671 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_generic;
   4672 
   4673 	pte_l2_l_cache_mode = L2_B|L2_C;
   4674 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_generic;
   4675 
   4676 	pte_l2_s_cache_mode = L2_B|L2_C;
   4677 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_generic;
   4678 
   4679 	/*
   4680 	 * If we have a write-through cache, set B and C.  If
   4681 	 * we have a write-back cache, then we assume setting
   4682 	 * only C will make those pages write-through.
   4683 	 */
   4684 	if (cpufuncs.cf_dcache_wb_range == (void *) cpufunc_nullop) {
   4685 		pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C;
   4686 		pte_l2_l_cache_mode_pt = L2_B|L2_C;
   4687 		pte_l2_s_cache_mode_pt = L2_B|L2_C;
   4688 	} else {
   4689 		pte_l1_s_cache_mode_pt = L1_S_C;
   4690 		pte_l2_l_cache_mode_pt = L2_C;
   4691 		pte_l2_s_cache_mode_pt = L2_C;
   4692 	}
   4693 
   4694 	pte_l2_s_prot_u = L2_S_PROT_U_generic;
   4695 	pte_l2_s_prot_w = L2_S_PROT_W_generic;
   4696 	pte_l2_s_prot_mask = L2_S_PROT_MASK_generic;
   4697 
   4698 	pte_l1_s_proto = L1_S_PROTO_generic;
   4699 	pte_l1_c_proto = L1_C_PROTO_generic;
   4700 	pte_l2_s_proto = L2_S_PROTO_generic;
   4701 
   4702 	pmap_copy_page_func = pmap_copy_page_generic;
   4703 	pmap_zero_page_func = pmap_zero_page_generic;
   4704 }
   4705 
   4706 #if defined(CPU_ARM8)
   4707 void
   4708 pmap_pte_init_arm8(void)
   4709 {
   4710 
   4711 	/*
   4712 	 * ARM8 is compatible with generic, but we need to use
   4713 	 * the page tables uncached.
   4714 	 */
   4715 	pmap_pte_init_generic();
   4716 
   4717 	pte_l1_s_cache_mode_pt = 0;
   4718 	pte_l2_l_cache_mode_pt = 0;
   4719 	pte_l2_s_cache_mode_pt = 0;
   4720 }
   4721 #endif /* CPU_ARM8 */
   4722 
   4723 #if defined(CPU_ARM9) && defined(ARM9_CACHE_WRITE_THROUGH)
   4724 void
   4725 pmap_pte_init_arm9(void)
   4726 {
   4727 
   4728 	/*
   4729 	 * ARM9 is compatible with generic, but we want to use
   4730 	 * write-through caching for now.
   4731 	 */
   4732 	pmap_pte_init_generic();
   4733 
   4734 	pte_l1_s_cache_mode = L1_S_C;
   4735 	pte_l2_l_cache_mode = L2_C;
   4736 	pte_l2_s_cache_mode = L2_C;
   4737 
   4738 	pte_l1_s_cache_mode_pt = L1_S_C;
   4739 	pte_l2_l_cache_mode_pt = L2_C;
   4740 	pte_l2_s_cache_mode_pt = L2_C;
   4741 }
   4742 #endif /* CPU_ARM9 */
   4743 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0 */
   4744 
   4745 #if defined(CPU_ARM10)
   4746 void
   4747 pmap_pte_init_arm10(void)
   4748 {
   4749 
   4750 	/*
   4751 	 * ARM10 is compatible with generic, but we want to use
   4752 	 * write-through caching for now.
   4753 	 */
   4754 	pmap_pte_init_generic();
   4755 
   4756 	pte_l1_s_cache_mode = L1_S_B | L1_S_C;
   4757 	pte_l2_l_cache_mode = L2_B | L2_C;
   4758 	pte_l2_s_cache_mode = L2_B | L2_C;
   4759 
   4760 	pte_l1_s_cache_mode_pt = L1_S_C;
   4761 	pte_l2_l_cache_mode_pt = L2_C;
   4762 	pte_l2_s_cache_mode_pt = L2_C;
   4763 
   4764 }
   4765 #endif /* CPU_ARM10 */
   4766 
   4767 #if ARM_MMU_SA1 == 1
   4768 void
   4769 pmap_pte_init_sa1(void)
   4770 {
   4771 
   4772 	/*
   4773 	 * The StrongARM SA-1 cache does not have a write-through
   4774 	 * mode.  So, do the generic initialization, then reset
   4775 	 * the page table cache mode to B=1,C=1, and note that
   4776 	 * the PTEs need to be sync'd.
   4777 	 */
   4778 	pmap_pte_init_generic();
   4779 
   4780 	pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C;
   4781 	pte_l2_l_cache_mode_pt = L2_B|L2_C;
   4782 	pte_l2_s_cache_mode_pt = L2_B|L2_C;
   4783 
   4784 	pmap_needs_pte_sync = 1;
   4785 }
   4786 #endif /* ARM_MMU_SA1 == 1*/
   4787 
   4788 #if ARM_MMU_XSCALE == 1
   4789 #if (ARM_NMMUS > 1)
   4790 static u_int xscale_use_minidata;
   4791 #endif
   4792 
   4793 void
   4794 pmap_pte_init_xscale(void)
   4795 {
   4796 	uint32_t auxctl;
   4797 	int write_through = 0;
   4798 
   4799 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
   4800 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_xscale;
   4801 
   4802 	pte_l2_l_cache_mode = L2_B|L2_C;
   4803 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_xscale;
   4804 
   4805 	pte_l2_s_cache_mode = L2_B|L2_C;
   4806 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_xscale;
   4807 
   4808 	pte_l1_s_cache_mode_pt = L1_S_C;
   4809 	pte_l2_l_cache_mode_pt = L2_C;
   4810 	pte_l2_s_cache_mode_pt = L2_C;
   4811 
   4812 #ifdef XSCALE_CACHE_READ_WRITE_ALLOCATE
   4813 	/*
   4814 	 * The XScale core has an enhanced mode where writes that
   4815 	 * miss the cache cause a cache line to be allocated.  This
   4816 	 * is significantly faster than the traditional, write-through
   4817 	 * behavior of this case.
   4818 	 */
   4819 	pte_l1_s_cache_mode |= L1_S_XSCALE_TEX(TEX_XSCALE_X);
   4820 	pte_l2_l_cache_mode |= L2_XSCALE_L_TEX(TEX_XSCALE_X);
   4821 	pte_l2_s_cache_mode |= L2_XSCALE_T_TEX(TEX_XSCALE_X);
   4822 #endif /* XSCALE_CACHE_READ_WRITE_ALLOCATE */
   4823 
   4824 #ifdef XSCALE_CACHE_WRITE_THROUGH
   4825 	/*
   4826 	 * Some versions of the XScale core have various bugs in
   4827 	 * their cache units, the work-around for which is to run
   4828 	 * the cache in write-through mode.  Unfortunately, this
   4829 	 * has a major (negative) impact on performance.  So, we
   4830 	 * go ahead and run fast-and-loose, in the hopes that we
   4831 	 * don't line up the planets in a way that will trip the
   4832 	 * bugs.
   4833 	 *
   4834 	 * However, we give you the option to be slow-but-correct.
   4835 	 */
   4836 	write_through = 1;
   4837 #elif defined(XSCALE_CACHE_WRITE_BACK)
   4838 	/* force write back cache mode */
   4839 	write_through = 0;
   4840 #elif defined(CPU_XSCALE_PXA250) || defined(CPU_XSCALE_PXA270)
   4841 	/*
   4842 	 * Intel PXA2[15]0 processors are known to have a bug in
   4843 	 * write-back cache on revision 4 and earlier (stepping
   4844 	 * A[01] and B[012]).  Fixed for C0 and later.
   4845 	 */
   4846 	{
   4847 		uint32_t id, type;
   4848 
   4849 		id = cpufunc_id();
   4850 		type = id & ~(CPU_ID_XSCALE_COREREV_MASK|CPU_ID_REVISION_MASK);
   4851 
   4852 		if (type == CPU_ID_PXA250 || type == CPU_ID_PXA210) {
   4853 			if ((id & CPU_ID_REVISION_MASK) < 5) {
   4854 				/* write through for stepping A0-1 and B0-2 */
   4855 				write_through = 1;
   4856 			}
   4857 		}
   4858 	}
   4859 #endif /* XSCALE_CACHE_WRITE_THROUGH */
   4860 
   4861 	if (write_through) {
   4862 		pte_l1_s_cache_mode = L1_S_C;
   4863 		pte_l2_l_cache_mode = L2_C;
   4864 		pte_l2_s_cache_mode = L2_C;
   4865 	}
   4866 
   4867 #if (ARM_NMMUS > 1)
   4868 	xscale_use_minidata = 1;
   4869 #endif
   4870 
   4871 	pte_l2_s_prot_u = L2_S_PROT_U_xscale;
   4872 	pte_l2_s_prot_w = L2_S_PROT_W_xscale;
   4873 	pte_l2_s_prot_mask = L2_S_PROT_MASK_xscale;
   4874 
   4875 	pte_l1_s_proto = L1_S_PROTO_xscale;
   4876 	pte_l1_c_proto = L1_C_PROTO_xscale;
   4877 	pte_l2_s_proto = L2_S_PROTO_xscale;
   4878 
   4879 	pmap_copy_page_func = pmap_copy_page_xscale;
   4880 	pmap_zero_page_func = pmap_zero_page_xscale;
   4881 
   4882 	/*
   4883 	 * Disable ECC protection of page table access, for now.
   4884 	 */
   4885 	__asm volatile("mrc p15, 0, %0, c1, c0, 1" : "=r" (auxctl));
   4886 	auxctl &= ~XSCALE_AUXCTL_P;
   4887 	__asm volatile("mcr p15, 0, %0, c1, c0, 1" : : "r" (auxctl));
   4888 }
   4889 
   4890 /*
   4891  * xscale_setup_minidata:
   4892  *
   4893  *	Set up the mini-data cache clean area.  We require the
   4894  *	caller to allocate the right amount of physically and
   4895  *	virtually contiguous space.
   4896  */
   4897 void
   4898 xscale_setup_minidata(vaddr_t l1pt, vaddr_t va, paddr_t pa)
   4899 {
   4900 	extern vaddr_t xscale_minidata_clean_addr;
   4901 	extern vsize_t xscale_minidata_clean_size; /* already initialized */
   4902 	pd_entry_t *pde = (pd_entry_t *) l1pt;
   4903 	pt_entry_t *pte;
   4904 	vsize_t size;
   4905 	uint32_t auxctl;
   4906 
   4907 	xscale_minidata_clean_addr = va;
   4908 
   4909 	/* Round it to page size. */
   4910 	size = (xscale_minidata_clean_size + L2_S_OFFSET) & L2_S_FRAME;
   4911 
   4912 	for (; size != 0;
   4913 	     va += L2_S_SIZE, pa += L2_S_SIZE, size -= L2_S_SIZE) {
   4914 #ifndef ARM32_NEW_VM_LAYOUT
   4915 		pte = (pt_entry_t *)
   4916 		    kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
   4917 #else
   4918 		pte = (pt_entry_t *) kernel_pt_lookup(
   4919 		    pde[L1_IDX(va)] & L1_C_ADDR_MASK);
   4920 #endif
   4921 		if (pte == NULL)
   4922 			panic("xscale_setup_minidata: can't find L2 table for "
   4923 			    "VA 0x%08lx", va);
   4924 #ifndef ARM32_NEW_VM_LAYOUT
   4925 		pte[(va >> PGSHIFT) & 0x3ff] =
   4926 #else
   4927 		pte[l2pte_index(va)] =
   4928 #endif
   4929 		    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, VM_PROT_READ) |
   4930 		    L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X);
   4931 	}
   4932 
   4933 	/*
   4934 	 * Configure the mini-data cache for write-back with
   4935 	 * read/write-allocate.
   4936 	 *
   4937 	 * NOTE: In order to reconfigure the mini-data cache, we must
   4938 	 * make sure it contains no valid data!  In order to do that,
   4939 	 * we must issue a global data cache invalidate command!
   4940 	 *
   4941 	 * WE ASSUME WE ARE RUNNING UN-CACHED WHEN THIS ROUTINE IS CALLED!
   4942 	 * THIS IS VERY IMPORTANT!
   4943 	 */
   4944 
   4945 	/* Invalidate data and mini-data. */
   4946 	__asm volatile("mcr p15, 0, %0, c7, c6, 0" : : "r" (0));
   4947 	__asm volatile("mrc p15, 0, %0, c1, c0, 1" : "=r" (auxctl));
   4948 	auxctl = (auxctl & ~XSCALE_AUXCTL_MD_MASK) | XSCALE_AUXCTL_MD_WB_RWA;
   4949 	__asm volatile("mcr p15, 0, %0, c1, c0, 1" : : "r" (auxctl));
   4950 }
   4951 
   4952 /*
   4953  * Change the PTEs for the specified kernel mappings such that they
   4954  * will use the mini data cache instead of the main data cache.
   4955  */
   4956 void
   4957 pmap_uarea(vaddr_t va)
   4958 {
   4959 	struct l2_bucket *l2b;
   4960 	pt_entry_t *ptep, *sptep, pte;
   4961 	vaddr_t next_bucket, eva;
   4962 
   4963 #if (ARM_NMMUS > 1)
   4964 	if (xscale_use_minidata == 0)
   4965 		return;
   4966 #endif
   4967 
   4968 	eva = va + USPACE;
   4969 
   4970 	while (va < eva) {
   4971 		next_bucket = L2_NEXT_BUCKET(va);
   4972 		if (next_bucket > eva)
   4973 			next_bucket = eva;
   4974 
   4975 		l2b = pmap_get_l2_bucket(pmap_kernel(), va);
   4976 		KDASSERT(l2b != NULL);
   4977 
   4978 		sptep = ptep = &l2b->l2b_kva[l2pte_index(va)];
   4979 
   4980 		while (va < next_bucket) {
   4981 			pte = *ptep;
   4982 			if (!l2pte_minidata(pte)) {
   4983 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
   4984 				cpu_tlb_flushD_SE(va);
   4985 				*ptep = pte & ~L2_B;
   4986 			}
   4987 			ptep++;
   4988 			va += PAGE_SIZE;
   4989 		}
   4990 		PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep));
   4991 	}
   4992 	cpu_cpwait();
   4993 }
   4994 #endif /* ARM_MMU_XSCALE == 1 */
   4995 
   4996 #if defined(DDB)
   4997 /*
   4998  * A couple of ddb-callable functions for dumping pmaps
   4999  */
   5000 void pmap_dump_all(void);
   5001 void pmap_dump(pmap_t);
   5002 
   5003 void
   5004 pmap_dump_all(void)
   5005 {
   5006 	pmap_t pm;
   5007 
   5008 	LIST_FOREACH(pm, &pmap_pmaps, pm_list) {
   5009 		if (pm == pmap_kernel())
   5010 			continue;
   5011 		pmap_dump(pm);
   5012 		printf("\n");
   5013 	}
   5014 }
   5015 
   5016 static pt_entry_t ncptes[64];
   5017 static void pmap_dump_ncpg(pmap_t);
   5018 
   5019 void
   5020 pmap_dump(pmap_t pm)
   5021 {
   5022 	struct l2_dtable *l2;
   5023 	struct l2_bucket *l2b;
   5024 	pt_entry_t *ptep, pte;
   5025 	vaddr_t l2_va, l2b_va, va;
   5026 	int i, j, k, occ, rows = 0;
   5027 
   5028 	if (pm == pmap_kernel())
   5029 		printf("pmap_kernel (%p): ", pm);
   5030 	else
   5031 		printf("user pmap (%p): ", pm);
   5032 
   5033 	printf("domain %d, l1 at %p\n", pm->pm_domain, pm->pm_l1->l1_kva);
   5034 
   5035 	l2_va = 0;
   5036 	for (i = 0; i < L2_SIZE; i++, l2_va += 0x01000000) {
   5037 		l2 = pm->pm_l2[i];
   5038 
   5039 		if (l2 == NULL || l2->l2_occupancy == 0)
   5040 			continue;
   5041 
   5042 		l2b_va = l2_va;
   5043 		for (j = 0; j < L2_BUCKET_SIZE; j++, l2b_va += 0x00100000) {
   5044 			l2b = &l2->l2_bucket[j];
   5045 
   5046 			if (l2b->l2b_occupancy == 0 || l2b->l2b_kva == NULL)
   5047 				continue;
   5048 
   5049 			ptep = l2b->l2b_kva;
   5050 
   5051 			for (k = 0; k < 256 && ptep[k] == 0; k++)
   5052 				;
   5053 
   5054 			k &= ~63;
   5055 			occ = l2b->l2b_occupancy;
   5056 			va = l2b_va + (k * 4096);
   5057 			for (; k < 256; k++, va += 0x1000) {
   5058 				char ch = ' ';
   5059 				if ((k % 64) == 0) {
   5060 					if ((rows % 8) == 0) {
   5061 						printf(
   5062 "          |0000   |8000   |10000  |18000  |20000  |28000  |30000  |38000\n");
   5063 					}
   5064 					printf("%08lx: ", va);
   5065 				}
   5066 
   5067 				ncptes[k & 63] = 0;
   5068 				pte = ptep[k];
   5069 				if (pte == 0) {
   5070 					ch = '.';
   5071 				} else {
   5072 					occ--;
   5073 					switch (pte & 0x0c) {
   5074 					case 0x00:
   5075 						ch = 'D'; /* No cache No buff */
   5076 						break;
   5077 					case 0x04:
   5078 						ch = 'B'; /* No cache buff */
   5079 						break;
   5080 					case 0x08:
   5081 						if (pte & 0x40)
   5082 							ch = 'm';
   5083 						else
   5084 						   ch = 'C'; /* Cache No buff */
   5085 						break;
   5086 					case 0x0c:
   5087 						ch = 'F'; /* Cache Buff */
   5088 						break;
   5089 					}
   5090 
   5091 					if ((pte & L2_S_PROT_U) == L2_S_PROT_U)
   5092 						ch += 0x20;
   5093 
   5094 					if ((pte & 0xc) == 0)
   5095 						ncptes[k & 63] = pte;
   5096 				}
   5097 
   5098 				if ((k % 64) == 63) {
   5099 					rows++;
   5100 					printf("%c\n", ch);
   5101 					pmap_dump_ncpg(pm);
   5102 					if (occ == 0)
   5103 						break;
   5104 				} else
   5105 					printf("%c", ch);
   5106 			}
   5107 		}
   5108 	}
   5109 }
   5110 
   5111 static void
   5112 pmap_dump_ncpg(pmap_t pm)
   5113 {
   5114 	struct vm_page *pg;
   5115 	struct pv_entry *pv;
   5116 	int i;
   5117 
   5118 	for (i = 0; i < 63; i++) {
   5119 		if (ncptes[i] == 0)
   5120 			continue;
   5121 
   5122 		pg = PHYS_TO_VM_PAGE(l2pte_pa(ncptes[i]));
   5123 		if (pg == NULL)
   5124 			continue;
   5125 
   5126 		printf(" pa 0x%08lx: krw %d kro %d urw %d uro %d\n",
   5127 		    VM_PAGE_TO_PHYS(pg),
   5128 		    pg->mdpage.krw_mappings, pg->mdpage.kro_mappings,
   5129 		    pg->mdpage.urw_mappings, pg->mdpage.uro_mappings);
   5130 
   5131 		for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
   5132 			printf("   %c va 0x%08lx, flags 0x%x\n",
   5133 			    (pm == pv->pv_pmap) ? '*' : ' ',
   5134 			    pv->pv_va, pv->pv_flags);
   5135 		}
   5136 	}
   5137 }
   5138 #endif
   5139