xref: /freebsd/sys/arm/arm/pmap-v6.c (revision c7142afec42c4f0a3aa4da845f4c4e15b5e3f018)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause AND BSD-2-Clause
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * Copyright (c) 1994 John S. Dyson
6  * Copyright (c) 1994 David Greenman
7  * Copyright (c) 2005-2010 Alan L. Cox <alc@cs.rice.edu>
8  * Copyright (c) 2014-2016 Svatopluk Kraus <skra@FreeBSD.org>
9  * Copyright (c) 2014-2016 Michal Meloun <mmel@FreeBSD.org>
10  * All rights reserved.
11  *
12  * This code is derived from software contributed to Berkeley by
13  * the Systems Programming Group of the University of Utah Computer
14  * Science Department and William Jolitz of UUNET Technologies Inc.
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  * 1. Redistributions of source code must retain the above copyright
20  *    notice, this list of conditions and the following disclaimer.
21  * 2. Redistributions in binary form must reproduce the above copyright
22  *    notice, this list of conditions and the following disclaimer in the
23  *    documentation and/or other materials provided with the distribution.
24  * 3. Neither the name of the University nor the names of its contributors
25  *    may be used to endorse or promote products derived from this software
26  *    without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38  * SUCH DAMAGE.
39  */
40 /*-
41  * Copyright (c) 2003 Networks Associates Technology, Inc.
42  * All rights reserved.
43  *
44  * This software was developed for the FreeBSD Project by Jake Burkholder,
45  * Safeport Network Services, and Network Associates Laboratories, the
46  * Security Research Division of Network Associates, Inc. under
47  * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA
48  * CHATS research program.
49  *
50  * Redistribution and use in source and binary forms, with or without
51  * modification, are permitted provided that the following conditions
52  * are met:
53  * 1. Redistributions of source code must retain the above copyright
54  *    notice, this list of conditions and the following disclaimer.
55  * 2. Redistributions in binary form must reproduce the above copyright
56  *    notice, this list of conditions and the following disclaimer in the
57  *    documentation and/or other materials provided with the distribution.
58  *
59  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
60  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
61  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
62  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
63  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
64  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
65  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
66  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
67  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
68  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
69  * SUCH DAMAGE.
70  */
71 
72 #include <sys/cdefs.h>
73 /*
74  *	Manages physical address maps.
75  *
76  *	Since the information managed by this module is
77  *	also stored by the logical address mapping module,
78  *	this module may throw away valid virtual-to-physical
79  *	mappings at almost any time.  However, invalidations
80  *	of virtual-to-physical mappings must be done as
81  *	requested.
82  *
83  *	In order to cope with hardware architectures which
84  *	make virtual-to-physical map invalidates expensive,
85  *	this module may delay invalidate or reduced protection
86  *	operations until such time as they are actually
87  *	necessary.  This module is given full information as
88  *	to which processors are currently using which maps,
89  *	and to when physical maps must be made correct.
90  */
91 
92 #include "opt_vm.h"
93 #include "opt_pmap.h"
94 #include "opt_ddb.h"
95 
96 #include <sys/param.h>
97 #include <sys/systm.h>
98 #include <sys/kernel.h>
99 #include <sys/ktr.h>
100 #include <sys/lock.h>
101 #include <sys/proc.h>
102 #include <sys/rwlock.h>
103 #include <sys/malloc.h>
104 #include <sys/vmmeter.h>
105 #include <sys/malloc.h>
106 #include <sys/mman.h>
107 #include <sys/sf_buf.h>
108 #include <sys/smp.h>
109 #include <sys/sched.h>
110 #include <sys/sysctl.h>
111 
112 #ifdef DDB
113 #include <ddb/ddb.h>
114 #endif
115 
116 #include <vm/vm.h>
117 #include <vm/uma.h>
118 #include <vm/pmap.h>
119 #include <vm/vm_param.h>
120 #include <vm/vm_kern.h>
121 #include <vm/vm_object.h>
122 #include <vm/vm_map.h>
123 #include <vm/vm_page.h>
124 #include <vm/vm_pageout.h>
125 #include <vm/vm_phys.h>
126 #include <vm/vm_extern.h>
127 #include <vm/vm_reserv.h>
128 #include <sys/lock.h>
129 #include <sys/mutex.h>
130 
131 #include <machine/md_var.h>
132 #include <machine/pmap_var.h>
133 #include <machine/cpu.h>
134 #include <machine/pcb.h>
135 #include <machine/sf_buf.h>
136 #ifdef SMP
137 #include <machine/smp.h>
138 #endif
139 #ifndef PMAP_SHPGPERPROC
140 #define PMAP_SHPGPERPROC 200
141 #endif
142 
143 #ifndef DIAGNOSTIC
144 #define PMAP_INLINE	__inline
145 #else
146 #define PMAP_INLINE
147 #endif
148 
149 #ifdef PMAP_DEBUG
150 static void pmap_zero_page_check(vm_page_t m);
151 void pmap_debug(int level);
152 int pmap_pid_dump(int pid);
153 
154 #define PDEBUG(_lev_,_stat_) \
155 	if (pmap_debug_level >= (_lev_)) \
156 		((_stat_))
157 #define dprintf printf
158 int pmap_debug_level = 1;
159 #else   /* PMAP_DEBUG */
160 #define PDEBUG(_lev_,_stat_) /* Nothing */
161 #define dprintf(x, arg...)
162 #endif  /* PMAP_DEBUG */
163 
164 /*
165  *  Level 2 page tables map definion ('max' is excluded).
166  */
167 
168 #define PT2V_MIN_ADDRESS	((vm_offset_t)PT2MAP)
169 #define PT2V_MAX_ADDRESS	((vm_offset_t)PT2MAP + PT2MAP_SIZE)
170 
171 #define UPT2V_MIN_ADDRESS	((vm_offset_t)PT2MAP)
172 #define UPT2V_MAX_ADDRESS \
173     ((vm_offset_t)(PT2MAP + (KERNBASE >> PT2MAP_SHIFT)))
174 
175 /*
176  *  Promotion to a 1MB (PTE1) page mapping requires that the corresponding
177  *  4KB (PTE2) page mappings have identical settings for the following fields:
178  */
179 #define PTE2_PROMOTE	(PTE2_V | PTE2_A | PTE2_NM | PTE2_S | PTE2_NG |	\
180 			 PTE2_NX | PTE2_RO | PTE2_U | PTE2_W |		\
181 			 PTE2_ATTR_MASK)
182 
183 #define PTE1_PROMOTE	(PTE1_V | PTE1_A | PTE1_NM | PTE1_S | PTE1_NG |	\
184 			 PTE1_NX | PTE1_RO | PTE1_U | PTE1_W |		\
185 			 PTE1_ATTR_MASK)
186 
187 #define ATTR_TO_L1(l2_attr)	((((l2_attr) & L2_TEX0) ? L1_S_TEX0 : 0) | \
188 				 (((l2_attr) & L2_C)    ? L1_S_C    : 0) | \
189 				 (((l2_attr) & L2_B)    ? L1_S_B    : 0) | \
190 				 (((l2_attr) & PTE2_A)  ? PTE1_A    : 0) | \
191 				 (((l2_attr) & PTE2_NM) ? PTE1_NM   : 0) | \
192 				 (((l2_attr) & PTE2_S)  ? PTE1_S    : 0) | \
193 				 (((l2_attr) & PTE2_NG) ? PTE1_NG   : 0) | \
194 				 (((l2_attr) & PTE2_NX) ? PTE1_NX   : 0) | \
195 				 (((l2_attr) & PTE2_RO) ? PTE1_RO   : 0) | \
196 				 (((l2_attr) & PTE2_U)  ? PTE1_U    : 0) | \
197 				 (((l2_attr) & PTE2_W)  ? PTE1_W    : 0))
198 
199 #define ATTR_TO_L2(l1_attr)	((((l1_attr) & L1_S_TEX0) ? L2_TEX0 : 0) | \
200 				 (((l1_attr) & L1_S_C)    ? L2_C    : 0) | \
201 				 (((l1_attr) & L1_S_B)    ? L2_B    : 0) | \
202 				 (((l1_attr) & PTE1_A)    ? PTE2_A  : 0) | \
203 				 (((l1_attr) & PTE1_NM)   ? PTE2_NM : 0) | \
204 				 (((l1_attr) & PTE1_S)    ? PTE2_S  : 0) | \
205 				 (((l1_attr) & PTE1_NG)   ? PTE2_NG : 0) | \
206 				 (((l1_attr) & PTE1_NX)   ? PTE2_NX : 0) | \
207 				 (((l1_attr) & PTE1_RO)   ? PTE2_RO : 0) | \
208 				 (((l1_attr) & PTE1_U)    ? PTE2_U  : 0) | \
209 				 (((l1_attr) & PTE1_W)    ? PTE2_W  : 0))
210 
211 /*
212  *  PTE2 descriptors creation macros.
213  */
214 #define PTE2_ATTR_DEFAULT	vm_memattr_to_pte2(VM_MEMATTR_DEFAULT)
215 #define PTE2_ATTR_PT		vm_memattr_to_pte2(pt_memattr)
216 
217 #define PTE2_KPT(pa)	PTE2_KERN(pa, PTE2_AP_KRW, PTE2_ATTR_PT)
218 #define PTE2_KPT_NG(pa)	PTE2_KERN_NG(pa, PTE2_AP_KRW, PTE2_ATTR_PT)
219 
220 #define PTE2_KRW(pa)	PTE2_KERN(pa, PTE2_AP_KRW, PTE2_ATTR_DEFAULT)
221 #define PTE2_KRO(pa)	PTE2_KERN(pa, PTE2_AP_KR, PTE2_ATTR_DEFAULT)
222 
223 #define PV_STATS
224 #ifdef PV_STATS
225 #define PV_STAT(x)	do { x ; } while (0)
226 #else
227 #define PV_STAT(x)	do { } while (0)
228 #endif
229 
230 /*
231  *  The boot_pt1 is used temporary in very early boot stage as L1 page table.
232  *  We can init many things with no memory allocation thanks to its static
233  *  allocation and this brings two main advantages:
234  *  (1) other cores can be started very simply,
235  *  (2) various boot loaders can be supported as its arguments can be processed
236  *      in virtual address space and can be moved to safe location before
237  *      first allocation happened.
238  *  Only disadvantage is that boot_pt1 is used only in very early boot stage.
239  *  However, the table is uninitialized and so lays in bss. Therefore kernel
240  *  image size is not influenced.
241  *
242  *  QQQ: In the future, maybe, boot_pt1 can be used for soft reset and
243  *       CPU suspend/resume game.
244  */
245 extern pt1_entry_t boot_pt1[];
246 
247 vm_paddr_t base_pt1;
248 pt1_entry_t *kern_pt1;
249 pt2_entry_t *kern_pt2tab;
250 pt2_entry_t *PT2MAP;
251 
252 static uint32_t ttb_flags;
253 static vm_memattr_t pt_memattr;
254 ttb_entry_t pmap_kern_ttb;
255 
256 struct pmap kernel_pmap_store;
257 LIST_HEAD(pmaplist, pmap);
258 static struct pmaplist allpmaps;
259 static struct mtx allpmaps_lock;
260 
261 vm_offset_t virtual_avail;	/* VA of first avail page (after kernel bss) */
262 vm_offset_t virtual_end;	/* VA of last avail page (end of kernel AS) */
263 
264 static vm_offset_t kernel_vm_end_new;
265 vm_offset_t kernel_vm_end = KERNBASE + NKPT2PG * NPT2_IN_PG * PTE1_SIZE;
266 vm_offset_t vm_max_kernel_address;
267 vm_paddr_t kernel_l1pa;
268 
269 static struct rwlock __aligned(CACHE_LINE_SIZE) pvh_global_lock;
270 
271 /*
272  *  Data for the pv entry allocation mechanism
273  */
274 static TAILQ_HEAD(pch, pv_chunk) pv_chunks = TAILQ_HEAD_INITIALIZER(pv_chunks);
275 static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0;
276 static struct md_page *pv_table; /* XXX: Is it used only the list in md_page? */
277 static int shpgperproc = PMAP_SHPGPERPROC;
278 
279 struct pv_chunk *pv_chunkbase;		/* KVA block for pv_chunks */
280 int pv_maxchunks;			/* How many chunks we have KVA for */
281 vm_offset_t pv_vafree;			/* freelist stored in the PTE */
282 
283 vm_paddr_t first_managed_pa;
284 #define	pa_to_pvh(pa)	(&pv_table[pte1_index(pa - first_managed_pa)])
285 
286 /*
287  *  All those kernel PT submaps that BSD is so fond of
288  */
289 caddr_t _tmppt = 0;
290 
291 /*
292  *  Crashdump maps.
293  */
294 static caddr_t crashdumpmap;
295 
296 static pt2_entry_t *PMAP1 = NULL, *PMAP2;
297 static pt2_entry_t *PADDR1 = NULL, *PADDR2;
298 #ifdef DDB
299 static pt2_entry_t *PMAP3;
300 static pt2_entry_t *PADDR3;
301 static int PMAP3cpu __unused; /* for SMP only */
302 #endif
303 #ifdef SMP
304 static int PMAP1cpu;
305 static int PMAP1changedcpu;
306 SYSCTL_INT(_debug, OID_AUTO, PMAP1changedcpu, CTLFLAG_RD,
307     &PMAP1changedcpu, 0,
308     "Number of times pmap_pte2_quick changed CPU with same PMAP1");
309 #endif
310 static int PMAP1changed;
311 SYSCTL_INT(_debug, OID_AUTO, PMAP1changed, CTLFLAG_RD,
312     &PMAP1changed, 0,
313     "Number of times pmap_pte2_quick changed PMAP1");
314 static int PMAP1unchanged;
315 SYSCTL_INT(_debug, OID_AUTO, PMAP1unchanged, CTLFLAG_RD,
316     &PMAP1unchanged, 0,
317     "Number of times pmap_pte2_quick didn't change PMAP1");
318 static struct mtx PMAP2mutex;
319 
320 /*
321  * Internal flags for pmap_enter()'s helper functions.
322  */
323 #define	PMAP_ENTER_NORECLAIM	0x1000000	/* Don't reclaim PV entries. */
324 #define	PMAP_ENTER_NOREPLACE	0x2000000	/* Don't replace mappings. */
325 
326 static __inline void pt2_wirecount_init(vm_page_t m);
327 static bool pmap_demote_pte1(pmap_t pmap, pt1_entry_t *pte1p,
328     vm_offset_t va);
329 static int pmap_enter_pte1(pmap_t pmap, vm_offset_t va, pt1_entry_t pte1,
330     u_int flags, vm_page_t m);
331 void cache_icache_sync_fresh(vm_offset_t va, vm_paddr_t pa, vm_size_t size);
332 
333 /*
334  *  Function to set the debug level of the pmap code.
335  */
336 #ifdef PMAP_DEBUG
337 void
pmap_debug(int level)338 pmap_debug(int level)
339 {
340 
341 	pmap_debug_level = level;
342 	dprintf("pmap_debug: level=%d\n", pmap_debug_level);
343 }
344 #endif /* PMAP_DEBUG */
345 
346 /*
347  *  This table must corespond with memory attribute configuration in vm.h.
348  *  First entry is used for normal system mapping.
349  *
350  *  Device memory is always marked as shared.
351  *  Normal memory is shared only in SMP .
352  *  Not outer shareable bits are not used yet.
353  *  Class 6 cannot be used on ARM11.
354  */
355 #define TEXDEF_TYPE_SHIFT	0
356 #define TEXDEF_TYPE_MASK	0x3
357 #define TEXDEF_INNER_SHIFT	2
358 #define TEXDEF_INNER_MASK	0x3
359 #define TEXDEF_OUTER_SHIFT	4
360 #define TEXDEF_OUTER_MASK	0x3
361 #define TEXDEF_NOS_SHIFT	6
362 #define TEXDEF_NOS_MASK		0x1
363 
364 #define TEX(t, i, o, s) 			\
365 		((t) << TEXDEF_TYPE_SHIFT) |	\
366 		((i) << TEXDEF_INNER_SHIFT) |	\
367 		((o) << TEXDEF_OUTER_SHIFT | 	\
368 		((s) << TEXDEF_NOS_SHIFT))
369 
370 static uint32_t tex_class[8] = {
371 /*	    type      inner cache outer cache */
372 	TEX(PRRR_MEM, NMRR_WB_WA, NMRR_WB_WA, 0),  /* 0 - ATTR_WB_WA	*/
373 	TEX(PRRR_MEM, NMRR_NC,	  NMRR_NC,    0),  /* 1 - ATTR_NOCACHE	*/
374 	TEX(PRRR_DEV, NMRR_NC,	  NMRR_NC,    0),  /* 2 - ATTR_DEVICE	*/
375 	TEX(PRRR_SO,  NMRR_NC,	  NMRR_NC,    0),  /* 3 - ATTR_SO	*/
376 	TEX(PRRR_MEM, NMRR_WT,	  NMRR_WT,    0),  /* 4 - ATTR_WT	*/
377 	TEX(PRRR_MEM, NMRR_NC,	  NMRR_NC,    0),  /* 5 - NOT USED YET	*/
378 	TEX(PRRR_MEM, NMRR_NC,	  NMRR_NC,    0),  /* 6 - NOT USED YET	*/
379 	TEX(PRRR_MEM, NMRR_NC,	  NMRR_NC,    0),  /* 7 - NOT USED YET	*/
380 };
381 #undef TEX
382 
383 static uint32_t pte2_attr_tab[8] = {
384 	PTE2_ATTR_WB_WA,	/* 0 - VM_MEMATTR_WB_WA */
385 	PTE2_ATTR_NOCACHE,	/* 1 - VM_MEMATTR_NOCACHE */
386 	PTE2_ATTR_DEVICE,	/* 2 - VM_MEMATTR_DEVICE */
387 	PTE2_ATTR_SO,		/* 3 - VM_MEMATTR_SO */
388 	PTE2_ATTR_WT,		/* 4 - VM_MEMATTR_WRITE_THROUGH */
389 	0,			/* 5 - NOT USED YET */
390 	0,			/* 6 - NOT USED YET */
391 	0			/* 7 - NOT USED YET */
392 };
393 CTASSERT(VM_MEMATTR_WB_WA == 0);
394 CTASSERT(VM_MEMATTR_NOCACHE == 1);
395 CTASSERT(VM_MEMATTR_DEVICE == 2);
396 CTASSERT(VM_MEMATTR_SO == 3);
397 CTASSERT(VM_MEMATTR_WRITE_THROUGH == 4);
398 #define	VM_MEMATTR_END	(VM_MEMATTR_WRITE_THROUGH + 1)
399 
400 bool
pmap_is_valid_memattr(pmap_t pmap __unused,vm_memattr_t mode)401 pmap_is_valid_memattr(pmap_t pmap __unused, vm_memattr_t mode)
402 {
403 
404 	return (mode >= 0 && mode < VM_MEMATTR_END);
405 }
406 
407 static inline uint32_t
vm_memattr_to_pte2(vm_memattr_t ma)408 vm_memattr_to_pte2(vm_memattr_t ma)
409 {
410 
411 	KASSERT((u_int)ma < VM_MEMATTR_END,
412 	    ("%s: bad vm_memattr_t %d", __func__, ma));
413 	return (pte2_attr_tab[(u_int)ma]);
414 }
415 
416 static inline uint32_t
vm_page_pte2_attr(vm_page_t m)417 vm_page_pte2_attr(vm_page_t m)
418 {
419 
420 	return (vm_memattr_to_pte2(m->md.pat_mode));
421 }
422 
423 /*
424  * Convert TEX definition entry to TTB flags.
425  */
426 static uint32_t
encode_ttb_flags(int idx)427 encode_ttb_flags(int idx)
428 {
429 	uint32_t inner, outer, nos, reg;
430 
431 	inner = (tex_class[idx] >> TEXDEF_INNER_SHIFT) &
432 		TEXDEF_INNER_MASK;
433 	outer = (tex_class[idx] >> TEXDEF_OUTER_SHIFT) &
434 		TEXDEF_OUTER_MASK;
435 	nos = (tex_class[idx] >> TEXDEF_NOS_SHIFT) &
436 		TEXDEF_NOS_MASK;
437 
438 	reg = nos << 5;
439 	reg |= outer << 3;
440 	if (cpuinfo.coherent_walk)
441 		reg |= (inner & 0x1) << 6;
442 	reg |= (inner & 0x2) >> 1;
443 #ifdef SMP
444 	ARM_SMP_UP(
445 		reg |= 1 << 1,
446 	);
447 #endif
448 	return reg;
449 }
450 
451 /*
452  *  Set TEX remapping registers in current CPU.
453  */
454 void
pmap_set_tex(void)455 pmap_set_tex(void)
456 {
457 	uint32_t prrr, nmrr;
458 	uint32_t type, inner, outer, nos;
459 	int i;
460 
461 #ifdef PMAP_PTE_NOCACHE
462 	/* XXX fixme */
463 	if (cpuinfo.coherent_walk) {
464 		pt_memattr = VM_MEMATTR_WB_WA;
465 		ttb_flags = encode_ttb_flags(0);
466 	}
467 	else {
468 		pt_memattr = VM_MEMATTR_NOCACHE;
469 		ttb_flags = encode_ttb_flags(1);
470 	}
471 #else
472 	pt_memattr = VM_MEMATTR_WB_WA;
473 	ttb_flags = encode_ttb_flags(0);
474 #endif
475 
476 	prrr = 0;
477 	nmrr = 0;
478 
479 	/* Build remapping register from TEX classes. */
480 	for (i = 0; i < 8; i++) {
481 		type = (tex_class[i] >> TEXDEF_TYPE_SHIFT) &
482 			TEXDEF_TYPE_MASK;
483 		inner = (tex_class[i] >> TEXDEF_INNER_SHIFT) &
484 			TEXDEF_INNER_MASK;
485 		outer = (tex_class[i] >> TEXDEF_OUTER_SHIFT) &
486 			TEXDEF_OUTER_MASK;
487 		nos = (tex_class[i] >> TEXDEF_NOS_SHIFT) &
488 			TEXDEF_NOS_MASK;
489 
490 		prrr |= type  << (i * 2);
491 		prrr |= nos   << (i + 24);
492 		nmrr |= inner << (i * 2);
493 		nmrr |= outer << (i * 2 + 16);
494 	}
495 	/* Add shareable bits for device memory. */
496 	prrr |= PRRR_DS0 | PRRR_DS1;
497 
498 	/* Add shareable bits for normal memory in SMP case. */
499 #ifdef SMP
500 	ARM_SMP_UP(
501 		prrr |= PRRR_NS1,
502 	);
503 #endif
504 	cp15_prrr_set(prrr);
505 	cp15_nmrr_set(nmrr);
506 
507 	/* Caches are disabled, so full TLB flush should be enough. */
508 	tlb_flush_all_local();
509 }
510 
511 /*
512  * Remap one vm_meattr class to another one. This can be useful as
513  * workaround for SOC errata, e.g. if devices must be accessed using
514  * SO memory class.
515  *
516  * !!! Please note that this function is absolutely last resort thing.
517  * It should not be used under normal circumstances. !!!
518  *
519  * Usage rules:
520  * - it shall be called after pmap_bootstrap_prepare() and before
521  *   cpu_mp_start() (thus only on boot CPU). In practice, it's expected
522  *   to be called from platform_attach() or platform_late_init().
523  *
524  * - if remapping doesn't change caching mode, or until uncached class
525  *   is remapped to any kind of cached one, then no other restriction exists.
526  *
527  * - if pmap_remap_vm_attr() changes caching mode, but both (original and
528  *   remapped) remain cached, then caller is resposible for calling
529  *   of dcache_wbinv_poc_all().
530  *
531  * - remapping of any kind of cached class to uncached is not permitted.
532  */
533 void
pmap_remap_vm_attr(vm_memattr_t old_attr,vm_memattr_t new_attr)534 pmap_remap_vm_attr(vm_memattr_t old_attr, vm_memattr_t new_attr)
535 {
536 	int old_idx, new_idx;
537 
538 	/* Map VM memattrs to indexes to tex_class table. */
539 	old_idx = PTE2_ATTR2IDX(pte2_attr_tab[(int)old_attr]);
540 	new_idx = PTE2_ATTR2IDX(pte2_attr_tab[(int)new_attr]);
541 
542 	/* Replace TEX attribute and apply it. */
543 	tex_class[old_idx] = tex_class[new_idx];
544 	pmap_set_tex();
545 }
546 
547 /*
548  * KERNBASE must be multiple of NPT2_IN_PG * PTE1_SIZE. In other words,
549  * KERNBASE is mapped by first L2 page table in L2 page table page. It
550  * meets same constrain due to PT2MAP being placed just under KERNBASE.
551  */
552 CTASSERT((KERNBASE & (NPT2_IN_PG * PTE1_SIZE - 1)) == 0);
553 CTASSERT((KERNBASE - VM_MAXUSER_ADDRESS) >= PT2MAP_SIZE);
554 
555 /*
556  *  In crazy dreams, PAGE_SIZE could be a multiple of PTE2_SIZE in general.
557  *  For now, anyhow, the following check must be fulfilled.
558  */
559 CTASSERT(PAGE_SIZE == PTE2_SIZE);
560 /*
561  *  We don't want to mess up MI code with all MMU and PMAP definitions,
562  *  so some things, which depend on other ones, are defined independently.
563  *  Now, it is time to check that we don't screw up something.
564  */
565 CTASSERT(PDRSHIFT == PTE1_SHIFT);
566 /*
567  *  Check L1 and L2 page table entries definitions consistency.
568  */
569 CTASSERT(NB_IN_PT1 == (sizeof(pt1_entry_t) * NPTE1_IN_PT1));
570 CTASSERT(NB_IN_PT2 == (sizeof(pt2_entry_t) * NPTE2_IN_PT2));
571 /*
572  *  Check L2 page tables page consistency.
573  */
574 CTASSERT(PAGE_SIZE == (NPT2_IN_PG * NB_IN_PT2));
575 CTASSERT((1 << PT2PG_SHIFT) == NPT2_IN_PG);
576 /*
577  *  Check PT2TAB consistency.
578  *  PT2TAB_ENTRIES is defined as a division of NPTE1_IN_PT1 by NPT2_IN_PG.
579  *  This should be done without remainder.
580  */
581 CTASSERT(NPTE1_IN_PT1 == (PT2TAB_ENTRIES * NPT2_IN_PG));
582 
583 /*
584  *	A PT2MAP magic.
585  *
586  *  All level 2 page tables (PT2s) are mapped continuously and accordingly
587  *  into PT2MAP address space. As PT2 size is less than PAGE_SIZE, this can
588  *  be done only if PAGE_SIZE is a multiple of PT2 size. All PT2s in one page
589  *  must be used together, but not necessary at once. The first PT2 in a page
590  *  must map things on correctly aligned address and the others must follow
591  *  in right order.
592  */
593 #define NB_IN_PT2TAB	(PT2TAB_ENTRIES * sizeof(pt2_entry_t))
594 #define NPT2_IN_PT2TAB	(NB_IN_PT2TAB / NB_IN_PT2)
595 #define NPG_IN_PT2TAB	(NB_IN_PT2TAB / PAGE_SIZE)
596 
597 /*
598  *  Check PT2TAB consistency.
599  *  NPT2_IN_PT2TAB is defined as a division of NB_IN_PT2TAB by NB_IN_PT2.
600  *  NPG_IN_PT2TAB is defined as a division of NB_IN_PT2TAB by PAGE_SIZE.
601  *  The both should be done without remainder.
602  */
603 CTASSERT(NB_IN_PT2TAB == (NPT2_IN_PT2TAB * NB_IN_PT2));
604 CTASSERT(NB_IN_PT2TAB == (NPG_IN_PT2TAB * PAGE_SIZE));
605 /*
606  *  The implementation was made general, however, with the assumption
607  *  bellow in mind. In case of another value of NPG_IN_PT2TAB,
608  *  the code should be once more rechecked.
609  */
610 CTASSERT(NPG_IN_PT2TAB == 1);
611 
612 /*
613  *  Get offset of PT2 in a page
614  *  associated with given PT1 index.
615  */
616 static __inline u_int
page_pt2off(u_int pt1_idx)617 page_pt2off(u_int pt1_idx)
618 {
619 
620 	return ((pt1_idx & PT2PG_MASK) * NB_IN_PT2);
621 }
622 
623 /*
624  *  Get physical address of PT2
625  *  associated with given PT2s page and PT1 index.
626  */
627 static __inline vm_paddr_t
page_pt2pa(vm_paddr_t pgpa,u_int pt1_idx)628 page_pt2pa(vm_paddr_t pgpa, u_int pt1_idx)
629 {
630 
631 	return (pgpa + page_pt2off(pt1_idx));
632 }
633 
634 /*
635  *  Get first entry of PT2
636  *  associated with given PT2s page and PT1 index.
637  */
638 static __inline pt2_entry_t *
page_pt2(vm_offset_t pgva,u_int pt1_idx)639 page_pt2(vm_offset_t pgva, u_int pt1_idx)
640 {
641 
642 	return ((pt2_entry_t *)(pgva + page_pt2off(pt1_idx)));
643 }
644 
645 /*
646  *  Get virtual address of PT2s page (mapped in PT2MAP)
647  *  which holds PT2 which holds entry which maps given virtual address.
648  */
649 static __inline vm_offset_t
pt2map_pt2pg(vm_offset_t va)650 pt2map_pt2pg(vm_offset_t va)
651 {
652 
653 	va &= ~(NPT2_IN_PG * PTE1_SIZE - 1);
654 	return ((vm_offset_t)pt2map_entry(va));
655 }
656 
657 /*****************************************************************************
658  *
659  *     THREE pmap initialization milestones exist:
660  *
661  *  locore.S
662  *    -> fundamental init (including MMU) in ASM
663  *
664  *  initarm()
665  *    -> fundamental init continues in C
666  *    -> first available physical address is known
667  *
668  *    pmap_bootstrap_prepare() -> FIRST PMAP MILESTONE (first epoch begins)
669  *      -> basic (safe) interface for physical address allocation is made
670  *      -> basic (safe) interface for virtual mapping is made
671  *      -> limited not SMP coherent work is possible
672  *
673  *    -> more fundamental init continues in C
674  *    -> locks and some more things are available
675  *    -> all fundamental allocations and mappings are done
676  *
677  *    pmap_bootstrap() -> SECOND PMAP MILESTONE (second epoch begins)
678  *      -> phys_avail[] and virtual_avail is set
679  *      -> control is passed to vm subsystem
680  *      -> physical and virtual address allocation are off limit
681  *      -> low level mapping functions, some SMP coherent,
682  *         are available, which cannot be used before vm subsystem
683  *         is being inited
684  *
685  *  mi_startup()
686  *    -> vm subsystem is being inited
687  *
688  *      pmap_init() -> THIRD PMAP MILESTONE (third epoch begins)
689  *        -> pmap is fully inited
690  *
691  *****************************************************************************/
692 
693 /*****************************************************************************
694  *
695  *	PMAP first stage initialization and utility functions
696  *	for pre-bootstrap epoch.
697  *
698  *  After pmap_bootstrap_prepare() is called, the following functions
699  *  can be used:
700  *
701  *  (1) strictly only for this stage functions for physical page allocations,
702  *      virtual space allocations, and mappings:
703  *
704  *  vm_paddr_t pmap_preboot_get_pages(u_int num);
705  *  void pmap_preboot_map_pages(vm_paddr_t pa, vm_offset_t va, u_int num);
706  *  vm_offset_t pmap_preboot_reserve_pages(u_int num);
707  *  vm_offset_t pmap_preboot_get_vpages(u_int num);
708  *  void pmap_preboot_map_attr(vm_paddr_t pa, vm_offset_t va, vm_size_t size,
709  *      vm_prot_t prot, vm_memattr_t attr);
710  *
711  *  (2) for all stages:
712  *
713  *  vm_paddr_t pmap_kextract(vm_offset_t va);
714  *
715  *  NOTE: This is not SMP coherent stage.
716  *
717  *****************************************************************************/
718 
719 #define KERNEL_P2V(pa) \
720     ((vm_offset_t)((pa) - arm_physmem_kernaddr + KERNVIRTADDR))
721 #define KERNEL_V2P(va) \
722     ((vm_paddr_t)((va) - KERNVIRTADDR + arm_physmem_kernaddr))
723 
724 static vm_paddr_t last_paddr;
725 
726 /*
727  *  Pre-bootstrap epoch page allocator.
728  */
729 vm_paddr_t
pmap_preboot_get_pages(u_int num)730 pmap_preboot_get_pages(u_int num)
731 {
732 	vm_paddr_t ret;
733 
734 	ret = last_paddr;
735 	last_paddr += num * PAGE_SIZE;
736 
737 	return (ret);
738 }
739 
740 /*
741  *	The fundamental initialization of PMAP stuff.
742  *
743  *  Some things already happened in locore.S and some things could happen
744  *  before pmap_bootstrap_prepare() is called, so let's recall what is done:
745  *  1. Caches are disabled.
746  *  2. We are running on virtual addresses already with 'boot_pt1'
747  *     as L1 page table.
748  *  3. So far, all virtual addresses can be converted to physical ones and
749  *     vice versa by the following macros:
750  *       KERNEL_P2V(pa) .... physical to virtual ones,
751  *       KERNEL_V2P(va) .... virtual to physical ones.
752  *
753  *  What is done herein:
754  *  1. The 'boot_pt1' is replaced by real kernel L1 page table 'kern_pt1'.
755  *  2. PT2MAP magic is brought to live.
756  *  3. Basic preboot functions for page allocations and mappings can be used.
757  *  4. Everything is prepared for L1 cache enabling.
758  *
759  *  Variations:
760  *  1. To use second TTB register, so kernel and users page tables will be
761  *     separated. This way process forking - pmap_pinit() - could be faster,
762  *     it saves physical pages and KVA per a process, and it's simple change.
763  *     However, it will lead, due to hardware matter, to the following:
764  *     (a) 2G space for kernel and 2G space for users.
765  *     (b) 1G space for kernel in low addresses and 3G for users above it.
766  *     A question is: Is the case (b) really an option? Note that case (b)
767  *     does save neither physical memory and KVA.
768  */
769 void
pmap_bootstrap_prepare(vm_paddr_t last)770 pmap_bootstrap_prepare(vm_paddr_t last)
771 {
772 	vm_paddr_t pt2pg_pa, pt2tab_pa, pa, size;
773 	vm_offset_t pt2pg_va;
774 	pt1_entry_t *pte1p;
775 	pt2_entry_t *pte2p;
776 	u_int i;
777 	uint32_t l1_attr;
778 
779 	/*
780 	 * Now, we are going to make real kernel mapping. Note that we are
781 	 * already running on some mapping made in locore.S and we expect
782 	 * that it's large enough to ensure nofault access to physical memory
783 	 * allocated herein before switch.
784 	 *
785 	 * As kernel image and everything needed before are and will be mapped
786 	 * by section mappings, we align last physical address to PTE1_SIZE.
787 	 */
788 	last_paddr = pte1_roundup(last);
789 
790 	/*
791 	 * Allocate and zero page(s) for kernel L1 page table.
792 	 *
793 	 * Note that it's first allocation on space which was PTE1_SIZE
794 	 * aligned and as such base_pt1 is aligned to NB_IN_PT1 too.
795 	 */
796 	base_pt1 = pmap_preboot_get_pages(NPG_IN_PT1);
797 	kern_pt1 = (pt1_entry_t *)KERNEL_P2V(base_pt1);
798 	bzero((void*)kern_pt1, NB_IN_PT1);
799 	pte1_sync_range(kern_pt1, NB_IN_PT1);
800 
801 	/* Allocate and zero page(s) for kernel PT2TAB. */
802 	pt2tab_pa = pmap_preboot_get_pages(NPG_IN_PT2TAB);
803 	kern_pt2tab = (pt2_entry_t *)KERNEL_P2V(pt2tab_pa);
804 	bzero(kern_pt2tab, NB_IN_PT2TAB);
805 	pte2_sync_range(kern_pt2tab, NB_IN_PT2TAB);
806 
807 	/* Allocate and zero page(s) for kernel L2 page tables. */
808 	pt2pg_pa = pmap_preboot_get_pages(NKPT2PG);
809 	pt2pg_va = KERNEL_P2V(pt2pg_pa);
810 	size = NKPT2PG * PAGE_SIZE;
811 	bzero((void*)pt2pg_va, size);
812 	pte2_sync_range((pt2_entry_t *)pt2pg_va, size);
813 
814 	/*
815 	 * Add a physical memory segment (vm_phys_seg) corresponding to the
816 	 * preallocated pages for kernel L2 page tables so that vm_page
817 	 * structures representing these pages will be created. The vm_page
818 	 * structures are required for promotion of the corresponding kernel
819 	 * virtual addresses to section mappings.
820 	 */
821 	vm_phys_add_seg(pt2tab_pa, pmap_preboot_get_pages(0));
822 
823 	/*
824 	 * Insert allocated L2 page table pages to PT2TAB and make
825 	 * link to all PT2s in L1 page table. See how kernel_vm_end
826 	 * is initialized.
827 	 *
828 	 * We play simple and safe. So every KVA will have underlaying
829 	 * L2 page table, even kernel image mapped by sections.
830 	 */
831 	pte2p = kern_pt2tab_entry(KERNBASE);
832 	for (pa = pt2pg_pa; pa < pt2pg_pa + size; pa += PTE2_SIZE)
833 		pt2tab_store(pte2p++, PTE2_KPT(pa));
834 
835 	pte1p = kern_pte1(KERNBASE);
836 	for (pa = pt2pg_pa; pa < pt2pg_pa + size; pa += NB_IN_PT2)
837 		pte1_store(pte1p++, PTE1_LINK(pa));
838 
839 	/* Make section mappings for kernel. */
840 	l1_attr = ATTR_TO_L1(PTE2_ATTR_DEFAULT);
841 	pte1p = kern_pte1(KERNBASE);
842 	for (pa = KERNEL_V2P(KERNBASE); pa < last; pa += PTE1_SIZE)
843 		pte1_store(pte1p++, PTE1_KERN(pa, PTE1_AP_KRW, l1_attr));
844 
845 	/*
846 	 * Get free and aligned space for PT2MAP and make L1 page table links
847 	 * to L2 page tables held in PT2TAB.
848 	 *
849 	 * Note that pages holding PT2s are stored in PT2TAB as pt2_entry_t
850 	 * descriptors and PT2TAB page(s) itself is(are) used as PT2s. Thus
851 	 * each entry in PT2TAB maps all PT2s in a page. This implies that
852 	 * virtual address of PT2MAP must be aligned to NPT2_IN_PG * PTE1_SIZE.
853 	 */
854 	PT2MAP = (pt2_entry_t *)(KERNBASE - PT2MAP_SIZE);
855 	pte1p = kern_pte1((vm_offset_t)PT2MAP);
856 	for (pa = pt2tab_pa, i = 0; i < NPT2_IN_PT2TAB; i++, pa += NB_IN_PT2) {
857 		pte1_store(pte1p++, PTE1_LINK(pa));
858 	}
859 
860 	/*
861 	 * Store PT2TAB in PT2TAB itself, i.e. self reference mapping.
862 	 * Each pmap will hold own PT2TAB, so the mapping should be not global.
863 	 */
864 	pte2p = kern_pt2tab_entry((vm_offset_t)PT2MAP);
865 	for (pa = pt2tab_pa, i = 0; i < NPG_IN_PT2TAB; i++, pa += PTE2_SIZE) {
866 		pt2tab_store(pte2p++, PTE2_KPT_NG(pa));
867 	}
868 
869 	/*
870 	 * Choose correct L2 page table and make mappings for allocations
871 	 * made herein which replaces temporary locore.S mappings after a while.
872 	 * Note that PT2MAP cannot be used until we switch to kern_pt1.
873 	 *
874 	 * Note, that these allocations started aligned on 1M section and
875 	 * kernel PT1 was allocated first. Making of mappings must follow
876 	 * order of physical allocations as we've used KERNEL_P2V() macro
877 	 * for virtual addresses resolution.
878 	 */
879 	pte2p = kern_pt2tab_entry((vm_offset_t)kern_pt1);
880 	pt2pg_va = KERNEL_P2V(pte2_pa(pte2_load(pte2p)));
881 
882 	pte2p = page_pt2(pt2pg_va, pte1_index((vm_offset_t)kern_pt1));
883 
884 	/* Make mapping for kernel L1 page table. */
885 	for (pa = base_pt1, i = 0; i < NPG_IN_PT1; i++, pa += PTE2_SIZE)
886 		pte2_store(pte2p++, PTE2_KPT(pa));
887 
888 	/* Make mapping for kernel PT2TAB. */
889 	for (pa = pt2tab_pa, i = 0; i < NPG_IN_PT2TAB; i++, pa += PTE2_SIZE)
890 		pte2_store(pte2p++, PTE2_KPT(pa));
891 
892 	/* Finally, switch from 'boot_pt1' to 'kern_pt1'. */
893 	pmap_kern_ttb = base_pt1 | ttb_flags;
894 	cpuinfo_reinit_mmu(pmap_kern_ttb);
895 	/*
896 	 * Initialize the first available KVA. As kernel image is mapped by
897 	 * sections, we are leaving some gap behind.
898 	 */
899 	virtual_avail = (vm_offset_t)kern_pt2tab + NPG_IN_PT2TAB * PAGE_SIZE;
900 }
901 
902 /*
903  *  Setup L2 page table page for given KVA.
904  *  Used in pre-bootstrap epoch.
905  *
906  *  Note that we have allocated NKPT2PG pages for L2 page tables in advance
907  *  and used them for mapping KVA starting from KERNBASE. However, this is not
908  *  enough. Vectors and devices need L2 page tables too. Note that they are
909  *  even above VM_MAX_KERNEL_ADDRESS.
910  */
911 static __inline vm_paddr_t
pmap_preboot_pt2pg_setup(vm_offset_t va)912 pmap_preboot_pt2pg_setup(vm_offset_t va)
913 {
914 	pt2_entry_t *pte2p, pte2;
915 	vm_paddr_t pt2pg_pa;
916 
917 	/* Get associated entry in PT2TAB. */
918 	pte2p = kern_pt2tab_entry(va);
919 
920 	/* Just return, if PT2s page exists already. */
921 	pte2 = pt2tab_load(pte2p);
922 	if (pte2_is_valid(pte2))
923 		return (pte2_pa(pte2));
924 
925 	KASSERT(va >= VM_MAX_KERNEL_ADDRESS,
926 	    ("%s: NKPT2PG too small", __func__));
927 
928 	/*
929 	 * Allocate page for PT2s and insert it to PT2TAB.
930 	 * In other words, map it into PT2MAP space.
931 	 */
932 	pt2pg_pa = pmap_preboot_get_pages(1);
933 	pt2tab_store(pte2p, PTE2_KPT(pt2pg_pa));
934 
935 	/* Zero all PT2s in allocated page. */
936 	bzero((void*)pt2map_pt2pg(va), PAGE_SIZE);
937 	pte2_sync_range((pt2_entry_t *)pt2map_pt2pg(va), PAGE_SIZE);
938 
939 	return (pt2pg_pa);
940 }
941 
942 /*
943  *  Setup L2 page table for given KVA.
944  *  Used in pre-bootstrap epoch.
945  */
946 static void
pmap_preboot_pt2_setup(vm_offset_t va)947 pmap_preboot_pt2_setup(vm_offset_t va)
948 {
949 	pt1_entry_t *pte1p;
950 	vm_paddr_t pt2pg_pa, pt2_pa;
951 
952 	/* Setup PT2's page. */
953 	pt2pg_pa = pmap_preboot_pt2pg_setup(va);
954 	pt2_pa = page_pt2pa(pt2pg_pa, pte1_index(va));
955 
956 	/* Insert PT2 to PT1. */
957 	pte1p = kern_pte1(va);
958 	pte1_store(pte1p, PTE1_LINK(pt2_pa));
959 }
960 
961 /*
962  *  Get L2 page entry associated with given KVA.
963  *  Used in pre-bootstrap epoch.
964  */
965 static __inline pt2_entry_t*
pmap_preboot_vtopte2(vm_offset_t va)966 pmap_preboot_vtopte2(vm_offset_t va)
967 {
968 	pt1_entry_t *pte1p;
969 
970 	/* Setup PT2 if needed. */
971 	pte1p = kern_pte1(va);
972 	if (!pte1_is_valid(pte1_load(pte1p))) /* XXX - sections ?! */
973 		pmap_preboot_pt2_setup(va);
974 
975 	return (pt2map_entry(va));
976 }
977 
978 /*
979  *  Pre-bootstrap epoch page(s) mapping(s).
980  */
981 void
pmap_preboot_map_pages(vm_paddr_t pa,vm_offset_t va,u_int num)982 pmap_preboot_map_pages(vm_paddr_t pa, vm_offset_t va, u_int num)
983 {
984 	u_int i;
985 	pt2_entry_t *pte2p;
986 
987 	/* Map all the pages. */
988 	for (i = 0; i < num; i++) {
989 		pte2p = pmap_preboot_vtopte2(va);
990 		pte2_store(pte2p, PTE2_KRW(pa));
991 		va += PAGE_SIZE;
992 		pa += PAGE_SIZE;
993 	}
994 }
995 
996 /*
997  *  Pre-bootstrap epoch virtual space alocator.
998  */
999 vm_offset_t
pmap_preboot_reserve_pages(u_int num)1000 pmap_preboot_reserve_pages(u_int num)
1001 {
1002 	u_int i;
1003 	vm_offset_t start, va;
1004 	pt2_entry_t *pte2p;
1005 
1006 	/* Allocate virtual space. */
1007 	start = va = virtual_avail;
1008 	virtual_avail += num * PAGE_SIZE;
1009 
1010 	/* Zero the mapping. */
1011 	for (i = 0; i < num; i++) {
1012 		pte2p = pmap_preboot_vtopte2(va);
1013 		pte2_store(pte2p, 0);
1014 		va += PAGE_SIZE;
1015 	}
1016 
1017 	return (start);
1018 }
1019 
1020 /*
1021  *  Pre-bootstrap epoch page(s) allocation and mapping(s).
1022  */
1023 vm_offset_t
pmap_preboot_get_vpages(u_int num)1024 pmap_preboot_get_vpages(u_int num)
1025 {
1026 	vm_paddr_t  pa;
1027 	vm_offset_t va;
1028 
1029 	/* Allocate physical page(s). */
1030 	pa = pmap_preboot_get_pages(num);
1031 
1032 	/* Allocate virtual space. */
1033 	va = virtual_avail;
1034 	virtual_avail += num * PAGE_SIZE;
1035 
1036 	/* Map and zero all. */
1037 	pmap_preboot_map_pages(pa, va, num);
1038 	bzero((void *)va, num * PAGE_SIZE);
1039 
1040 	return (va);
1041 }
1042 
1043 /*
1044  *  Pre-bootstrap epoch page mapping(s) with attributes.
1045  */
1046 void
pmap_preboot_map_attr(vm_paddr_t pa,vm_offset_t va,vm_size_t size,vm_prot_t prot,vm_memattr_t attr)1047 pmap_preboot_map_attr(vm_paddr_t pa, vm_offset_t va, vm_size_t size,
1048     vm_prot_t prot, vm_memattr_t attr)
1049 {
1050 	u_int num;
1051 	u_int l1_attr, l1_prot, l2_prot, l2_attr;
1052 	pt1_entry_t *pte1p;
1053 	pt2_entry_t *pte2p;
1054 
1055 	l2_prot = prot & VM_PROT_WRITE ? PTE2_AP_KRW : PTE2_AP_KR;
1056 	l2_prot |= (prot & VM_PROT_EXECUTE) ? PTE2_X : PTE2_NX;
1057 	l2_attr = vm_memattr_to_pte2(attr);
1058 	l1_prot = ATTR_TO_L1(l2_prot);
1059 	l1_attr = ATTR_TO_L1(l2_attr);
1060 
1061 	/* Map all the pages. */
1062 	num = round_page(size);
1063 	while (num > 0) {
1064 		if ((((va | pa) & PTE1_OFFSET) == 0) && (num >= PTE1_SIZE)) {
1065 			pte1p = kern_pte1(va);
1066 			pte1_store(pte1p, PTE1_KERN(pa, l1_prot, l1_attr));
1067 			va += PTE1_SIZE;
1068 			pa += PTE1_SIZE;
1069 			num -= PTE1_SIZE;
1070 		} else {
1071 			pte2p = pmap_preboot_vtopte2(va);
1072 			pte2_store(pte2p, PTE2_KERN(pa, l2_prot, l2_attr));
1073 			va += PAGE_SIZE;
1074 			pa += PAGE_SIZE;
1075 			num -= PAGE_SIZE;
1076 		}
1077 	}
1078 }
1079 
1080 /*
1081  *  Extract from the kernel page table the physical address
1082  *  that is mapped by the given virtual address "va".
1083  */
1084 vm_paddr_t
pmap_kextract(vm_offset_t va)1085 pmap_kextract(vm_offset_t va)
1086 {
1087 	vm_paddr_t pa;
1088 	pt1_entry_t pte1;
1089 	pt2_entry_t pte2;
1090 
1091 	pte1 = pte1_load(kern_pte1(va));
1092 	if (pte1_is_section(pte1)) {
1093 		pa = pte1_pa(pte1) | (va & PTE1_OFFSET);
1094 	} else if (pte1_is_link(pte1)) {
1095 		/*
1096 		 * We should beware of concurrent promotion that changes
1097 		 * pte1 at this point. However, it's not a problem as PT2
1098 		 * page is preserved by promotion in PT2TAB. So even if
1099 		 * it happens, using of PT2MAP is still safe.
1100 		 *
1101 		 * QQQ: However, concurrent removing is a problem which
1102 		 *      ends in abort on PT2MAP space. Locking must be used
1103 		 *      to deal with this.
1104 		 */
1105 		pte2 = pte2_load(pt2map_entry(va));
1106 		pa = pte2_pa(pte2) | (va & PTE2_OFFSET);
1107 	}
1108 	else {
1109 		panic("%s: va %#x pte1 %#x", __func__, va, pte1);
1110 	}
1111 	return (pa);
1112 }
1113 
1114 /*
1115  *  Extract from the kernel page table the physical address
1116  *  that is mapped by the given virtual address "va". Also
1117  *  return L2 page table entry which maps the address.
1118  *
1119  *  This is only intended to be used for panic dumps.
1120  */
1121 vm_paddr_t
pmap_dump_kextract(vm_offset_t va,pt2_entry_t * pte2p)1122 pmap_dump_kextract(vm_offset_t va, pt2_entry_t *pte2p)
1123 {
1124 	vm_paddr_t pa;
1125 	pt1_entry_t pte1;
1126 	pt2_entry_t pte2;
1127 
1128 	pte1 = pte1_load(kern_pte1(va));
1129 	if (pte1_is_section(pte1)) {
1130 		pa = pte1_pa(pte1) | (va & PTE1_OFFSET);
1131 		pte2 = pa | ATTR_TO_L2(pte1) | PTE2_V;
1132 	} else if (pte1_is_link(pte1)) {
1133 		pte2 = pte2_load(pt2map_entry(va));
1134 		pa = pte2_pa(pte2);
1135 	} else {
1136 		pte2 = 0;
1137 		pa = 0;
1138 	}
1139 	if (pte2p != NULL)
1140 		*pte2p = pte2;
1141 	return (pa);
1142 }
1143 
1144 /*****************************************************************************
1145  *
1146  *	PMAP second stage initialization and utility functions
1147  *	for bootstrap epoch.
1148  *
1149  *  After pmap_bootstrap() is called, the following functions for
1150  *  mappings can be used:
1151  *
1152  *  void pmap_kenter(vm_offset_t va, vm_paddr_t pa);
1153  *  void pmap_kremove(vm_offset_t va);
1154  *  vm_offset_t pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end,
1155  *      int prot);
1156  *
1157  *  NOTE: This is not SMP coherent stage. And physical page allocation is not
1158  *        allowed during this stage.
1159  *
1160  *****************************************************************************/
1161 
1162 /*
1163  *  Initialize kernel PMAP locks and lists, kernel_pmap itself, and
1164  *  reserve various virtual spaces for temporary mappings.
1165  */
1166 void
pmap_bootstrap(vm_offset_t firstaddr)1167 pmap_bootstrap(vm_offset_t firstaddr)
1168 {
1169 	pt2_entry_t *unused __unused;
1170 	struct pcpu *pc;
1171 
1172 	/*
1173 	 * Initialize the kernel pmap (which is statically allocated).
1174 	 */
1175 	PMAP_LOCK_INIT(kernel_pmap);
1176 	kernel_l1pa = (vm_paddr_t)kern_pt1;  /* for libkvm */
1177 	kernel_pmap->pm_pt1 = kern_pt1;
1178 	kernel_pmap->pm_pt2tab = kern_pt2tab;
1179 	CPU_FILL(&kernel_pmap->pm_active);  /* don't allow deactivation */
1180 	TAILQ_INIT(&kernel_pmap->pm_pvchunk);
1181 
1182 	/*
1183 	 * Initialize the global pv list lock.
1184 	 */
1185 	rw_init(&pvh_global_lock, "pmap pv global");
1186 
1187 	LIST_INIT(&allpmaps);
1188 
1189 	/*
1190 	 * Request a spin mutex so that changes to allpmaps cannot be
1191 	 * preempted by smp_rendezvous_cpus().
1192 	 */
1193 	mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN);
1194 	mtx_lock_spin(&allpmaps_lock);
1195 	LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list);
1196 	mtx_unlock_spin(&allpmaps_lock);
1197 
1198 	/*
1199 	 * Reserve some special page table entries/VA space for temporary
1200 	 * mapping of pages.
1201 	 */
1202 #define	SYSMAP(c, p, v, n)  do {		\
1203 	v = (c)pmap_preboot_reserve_pages(n);	\
1204 	p = pt2map_entry((vm_offset_t)v);	\
1205 	} while (0)
1206 
1207 	/*
1208 	 * Local CMAP1/CMAP2 are used for zeroing and copying pages.
1209 	 * Local CMAP2 is also used for data cache cleaning.
1210 	 */
1211 	pc = get_pcpu();
1212 	mtx_init(&pc->pc_cmap_lock, "SYSMAPS", NULL, MTX_DEF);
1213 	SYSMAP(caddr_t, pc->pc_cmap1_pte2p, pc->pc_cmap1_addr, 1);
1214 	SYSMAP(caddr_t, pc->pc_cmap2_pte2p, pc->pc_cmap2_addr, 1);
1215 	SYSMAP(vm_offset_t, pc->pc_qmap_pte2p, pc->pc_qmap_addr, 1);
1216 
1217 	/*
1218 	 * Crashdump maps.
1219 	 */
1220 	SYSMAP(caddr_t, unused, crashdumpmap, MAXDUMPPGS);
1221 
1222 	/*
1223 	 * _tmppt is used for reading arbitrary physical pages via /dev/mem.
1224 	 */
1225 	SYSMAP(caddr_t, unused, _tmppt, 1);
1226 
1227 	/*
1228 	 * PADDR1 and PADDR2 are used by pmap_pte2_quick() and pmap_pte2(),
1229 	 * respectively. PADDR3 is used by pmap_pte2_ddb().
1230 	 */
1231 	SYSMAP(pt2_entry_t *, PMAP1, PADDR1, 1);
1232 	SYSMAP(pt2_entry_t *, PMAP2, PADDR2, 1);
1233 #ifdef DDB
1234 	SYSMAP(pt2_entry_t *, PMAP3, PADDR3, 1);
1235 #endif
1236 	mtx_init(&PMAP2mutex, "PMAP2", NULL, MTX_DEF);
1237 
1238 	/*
1239 	 * Note that in very short time in initarm(), we are going to
1240 	 * initialize phys_avail[] array and no further page allocation
1241 	 * can happen after that until vm subsystem will be initialized.
1242 	 */
1243 	kernel_vm_end_new = kernel_vm_end;
1244 	virtual_end = vm_max_kernel_address;
1245 }
1246 
1247 static void
pmap_init_reserved_pages(void)1248 pmap_init_reserved_pages(void)
1249 {
1250 	struct pcpu *pc;
1251 	vm_offset_t pages;
1252 	int i;
1253 
1254 	CPU_FOREACH(i) {
1255 		pc = pcpu_find(i);
1256 		/*
1257 		 * Skip if the mapping has already been initialized,
1258 		 * i.e. this is the BSP.
1259 		 */
1260 		if (pc->pc_cmap1_addr != 0)
1261 			continue;
1262 		mtx_init(&pc->pc_cmap_lock, "SYSMAPS", NULL, MTX_DEF);
1263 		pages = kva_alloc(PAGE_SIZE * 3);
1264 		if (pages == 0)
1265 			panic("%s: unable to allocate KVA", __func__);
1266 		pc->pc_cmap1_pte2p = pt2map_entry(pages);
1267 		pc->pc_cmap2_pte2p = pt2map_entry(pages + PAGE_SIZE);
1268 		pc->pc_qmap_pte2p = pt2map_entry(pages + (PAGE_SIZE * 2));
1269 		pc->pc_cmap1_addr = (caddr_t)pages;
1270 		pc->pc_cmap2_addr = (caddr_t)(pages + PAGE_SIZE);
1271 		pc->pc_qmap_addr = pages + (PAGE_SIZE * 2);
1272 	}
1273 }
1274 SYSINIT(rpages_init, SI_SUB_CPU, SI_ORDER_ANY, pmap_init_reserved_pages, NULL);
1275 
1276 /*
1277  *  The function can already be use in second initialization stage.
1278  *  As such, the function DOES NOT call pmap_growkernel() where PT2
1279  *  allocation can happen. So if used, be sure that PT2 for given
1280  *  virtual address is allocated already!
1281  *
1282  *  Add a wired page to the kva.
1283  *  Note: not SMP coherent.
1284  */
1285 static __inline void
pmap_kenter_prot_attr(vm_offset_t va,vm_paddr_t pa,uint32_t prot,uint32_t attr)1286 pmap_kenter_prot_attr(vm_offset_t va, vm_paddr_t pa, uint32_t prot,
1287     uint32_t attr)
1288 {
1289 	pt1_entry_t *pte1p;
1290 	pt2_entry_t *pte2p;
1291 
1292 	pte1p = kern_pte1(va);
1293 	if (!pte1_is_valid(pte1_load(pte1p))) { /* XXX - sections ?! */
1294 		/*
1295 		 * This is a very low level function, so PT2 and particularly
1296 		 * PT2PG associated with given virtual address must be already
1297 		 * allocated. It's a pain mainly during pmap initialization
1298 		 * stage. However, called after pmap initialization with
1299 		 * virtual address not under kernel_vm_end will lead to
1300 		 * the same misery.
1301 		 */
1302 		if (!pte2_is_valid(pte2_load(kern_pt2tab_entry(va))))
1303 			panic("%s: kernel PT2 not allocated!", __func__);
1304 	}
1305 
1306 	pte2p = pt2map_entry(va);
1307 	pte2_store(pte2p, PTE2_KERN(pa, prot, attr));
1308 }
1309 
1310 PMAP_INLINE void
pmap_kenter(vm_offset_t va,vm_paddr_t pa)1311 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1312 {
1313 
1314 	pmap_kenter_prot_attr(va, pa, PTE2_AP_KRW, PTE2_ATTR_DEFAULT);
1315 }
1316 
1317 /*
1318  *  Remove a page from the kernel pagetables.
1319  *  Note: not SMP coherent.
1320  */
1321 PMAP_INLINE void
pmap_kremove(vm_offset_t va)1322 pmap_kremove(vm_offset_t va)
1323 {
1324 	pt1_entry_t *pte1p;
1325 	pt2_entry_t *pte2p;
1326 
1327 	pte1p = kern_pte1(va);
1328 	if (pte1_is_section(pte1_load(pte1p))) {
1329 		pte1_clear(pte1p);
1330 	} else {
1331 		pte2p = pt2map_entry(va);
1332 		pte2_clear(pte2p);
1333 	}
1334 }
1335 
1336 /*
1337  *  Share new kernel PT2PG with all pmaps.
1338  *  The caller is responsible for maintaining TLB consistency.
1339  */
1340 static void
pmap_kenter_pt2tab(vm_offset_t va,pt2_entry_t npte2)1341 pmap_kenter_pt2tab(vm_offset_t va, pt2_entry_t npte2)
1342 {
1343 	pmap_t pmap;
1344 	pt2_entry_t *pte2p;
1345 
1346 	mtx_lock_spin(&allpmaps_lock);
1347 	LIST_FOREACH(pmap, &allpmaps, pm_list) {
1348 		pte2p = pmap_pt2tab_entry(pmap, va);
1349 		pt2tab_store(pte2p, npte2);
1350 	}
1351 	mtx_unlock_spin(&allpmaps_lock);
1352 }
1353 
1354 /*
1355  *  Share new kernel PTE1 with all pmaps.
1356  *  The caller is responsible for maintaining TLB consistency.
1357  */
1358 static void
pmap_kenter_pte1(vm_offset_t va,pt1_entry_t npte1)1359 pmap_kenter_pte1(vm_offset_t va, pt1_entry_t npte1)
1360 {
1361 	pmap_t pmap;
1362 	pt1_entry_t *pte1p;
1363 
1364 	mtx_lock_spin(&allpmaps_lock);
1365 	LIST_FOREACH(pmap, &allpmaps, pm_list) {
1366 		pte1p = pmap_pte1(pmap, va);
1367 		pte1_store(pte1p, npte1);
1368 	}
1369 	mtx_unlock_spin(&allpmaps_lock);
1370 }
1371 
1372 /*
1373  *  Used to map a range of physical addresses into kernel
1374  *  virtual address space.
1375  *
1376  *  The value passed in '*virt' is a suggested virtual address for
1377  *  the mapping. Architectures which can support a direct-mapped
1378  *  physical to virtual region can return the appropriate address
1379  *  within that region, leaving '*virt' unchanged. Other
1380  *  architectures should map the pages starting at '*virt' and
1381  *  update '*virt' with the first usable address after the mapped
1382  *  region.
1383  *
1384  *  NOTE: Read the comments above pmap_kenter_prot_attr() as
1385  *        the function is used herein!
1386  */
1387 vm_offset_t
pmap_map(vm_offset_t * virt,vm_paddr_t start,vm_paddr_t end,int prot)1388 pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot)
1389 {
1390 	vm_offset_t va, sva;
1391 	vm_paddr_t pte1_offset;
1392 	pt1_entry_t npte1;
1393 	uint32_t l1prot, l2prot;
1394 	uint32_t l1attr, l2attr;
1395 
1396 	PDEBUG(1, printf("%s: virt = %#x, start = %#x, end = %#x (size = %#x),"
1397 	    " prot = %d\n", __func__, *virt, start, end, end - start,  prot));
1398 
1399 	l2prot = (prot & VM_PROT_WRITE) ? PTE2_AP_KRW : PTE2_AP_KR;
1400 	l2prot |= (prot & VM_PROT_EXECUTE) ? PTE2_X : PTE2_NX;
1401 	l1prot = ATTR_TO_L1(l2prot);
1402 
1403 	l2attr = PTE2_ATTR_DEFAULT;
1404 	l1attr = ATTR_TO_L1(l2attr);
1405 
1406 	va = *virt;
1407 	/*
1408 	 * Does the physical address range's size and alignment permit at
1409 	 * least one section mapping to be created?
1410 	 */
1411 	pte1_offset = start & PTE1_OFFSET;
1412 	if ((end - start) - ((PTE1_SIZE - pte1_offset) & PTE1_OFFSET) >=
1413 	    PTE1_SIZE) {
1414 		/*
1415 		 * Increase the starting virtual address so that its alignment
1416 		 * does not preclude the use of section mappings.
1417 		 */
1418 		if ((va & PTE1_OFFSET) < pte1_offset)
1419 			va = pte1_trunc(va) + pte1_offset;
1420 		else if ((va & PTE1_OFFSET) > pte1_offset)
1421 			va = pte1_roundup(va) + pte1_offset;
1422 	}
1423 	sva = va;
1424 	while (start < end) {
1425 		if ((start & PTE1_OFFSET) == 0 && end - start >= PTE1_SIZE) {
1426 			KASSERT((va & PTE1_OFFSET) == 0,
1427 			    ("%s: misaligned va %#x", __func__, va));
1428 			npte1 = PTE1_KERN(start, l1prot, l1attr);
1429 			pmap_kenter_pte1(va, npte1);
1430 			va += PTE1_SIZE;
1431 			start += PTE1_SIZE;
1432 		} else {
1433 			pmap_kenter_prot_attr(va, start, l2prot, l2attr);
1434 			va += PAGE_SIZE;
1435 			start += PAGE_SIZE;
1436 		}
1437 	}
1438 	tlb_flush_range(sva, va - sva);
1439 	*virt = va;
1440 	return (sva);
1441 }
1442 
1443 /*
1444  *  Make a temporary mapping for a physical address.
1445  *  This is only intended to be used for panic dumps.
1446  */
1447 void *
pmap_kenter_temporary(vm_paddr_t pa,int i)1448 pmap_kenter_temporary(vm_paddr_t pa, int i)
1449 {
1450 	vm_offset_t va;
1451 
1452 	/* QQQ: 'i' should be less or equal to MAXDUMPPGS. */
1453 
1454 	va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE);
1455 	pmap_kenter(va, pa);
1456 	tlb_flush_local(va);
1457 	return ((void *)crashdumpmap);
1458 }
1459 
1460 /*************************************
1461  *
1462  *  TLB & cache maintenance routines.
1463  *
1464  *************************************/
1465 
1466 /*
1467  *  We inline these within pmap.c for speed.
1468  */
1469 PMAP_INLINE void
pmap_tlb_flush(pmap_t pmap,vm_offset_t va)1470 pmap_tlb_flush(pmap_t pmap, vm_offset_t va)
1471 {
1472 
1473 	if (pmap == kernel_pmap || !CPU_EMPTY(&pmap->pm_active))
1474 		tlb_flush(va);
1475 }
1476 
1477 PMAP_INLINE void
pmap_tlb_flush_range(pmap_t pmap,vm_offset_t sva,vm_size_t size)1478 pmap_tlb_flush_range(pmap_t pmap, vm_offset_t sva, vm_size_t size)
1479 {
1480 
1481 	if (pmap == kernel_pmap || !CPU_EMPTY(&pmap->pm_active))
1482 		tlb_flush_range(sva, size);
1483 }
1484 
1485 /*
1486  *  Abuse the pte2 nodes for unmapped kva to thread a kva freelist through.
1487  *  Requirements:
1488  *   - Must deal with pages in order to ensure that none of the PTE2_* bits
1489  *     are ever set, PTE2_V in particular.
1490  *   - Assumes we can write to pte2s without pte2_store() atomic ops.
1491  *   - Assumes nothing will ever test these addresses for 0 to indicate
1492  *     no mapping instead of correctly checking PTE2_V.
1493  *   - Assumes a vm_offset_t will fit in a pte2 (true for arm).
1494  *  Because PTE2_V is never set, there can be no mappings to invalidate.
1495  */
1496 static vm_offset_t
pmap_pte2list_alloc(vm_offset_t * head)1497 pmap_pte2list_alloc(vm_offset_t *head)
1498 {
1499 	pt2_entry_t *pte2p;
1500 	vm_offset_t va;
1501 
1502 	va = *head;
1503 	if (va == 0)
1504 		panic("pmap_ptelist_alloc: exhausted ptelist KVA");
1505 	pte2p = pt2map_entry(va);
1506 	*head = *pte2p;
1507 	if (*head & PTE2_V)
1508 		panic("%s: va with PTE2_V set!", __func__);
1509 	*pte2p = 0;
1510 	return (va);
1511 }
1512 
1513 static void
pmap_pte2list_free(vm_offset_t * head,vm_offset_t va)1514 pmap_pte2list_free(vm_offset_t *head, vm_offset_t va)
1515 {
1516 	pt2_entry_t *pte2p;
1517 
1518 	if (va & PTE2_V)
1519 		panic("%s: freeing va with PTE2_V set!", __func__);
1520 	pte2p = pt2map_entry(va);
1521 	*pte2p = *head;		/* virtual! PTE2_V is 0 though */
1522 	*head = va;
1523 }
1524 
1525 static void
pmap_pte2list_init(vm_offset_t * head,void * base,int npages)1526 pmap_pte2list_init(vm_offset_t *head, void *base, int npages)
1527 {
1528 	int i;
1529 	vm_offset_t va;
1530 
1531 	*head = 0;
1532 	for (i = npages - 1; i >= 0; i--) {
1533 		va = (vm_offset_t)base + i * PAGE_SIZE;
1534 		pmap_pte2list_free(head, va);
1535 	}
1536 }
1537 
1538 /*****************************************************************************
1539  *
1540  *	PMAP third and final stage initialization.
1541  *
1542  *  After pmap_init() is called, PMAP subsystem is fully initialized.
1543  *
1544  *****************************************************************************/
1545 
1546 SYSCTL_NODE(_vm, OID_AUTO, pmap, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
1547     "VM/pmap parameters");
1548 
1549 SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_max, CTLFLAG_RD, &pv_entry_max, 0,
1550     "Max number of PV entries");
1551 SYSCTL_INT(_vm_pmap, OID_AUTO, shpgperproc, CTLFLAG_RD, &shpgperproc, 0,
1552     "Page share factor per proc");
1553 
1554 static u_long nkpt2pg = NKPT2PG;
1555 SYSCTL_ULONG(_vm_pmap, OID_AUTO, nkpt2pg, CTLFLAG_RD,
1556     &nkpt2pg, 0, "Pre-allocated pages for kernel PT2s");
1557 
1558 static int sp_enabled = 1;
1559 SYSCTL_INT(_vm_pmap, OID_AUTO, sp_enabled, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
1560     &sp_enabled, 0, "Are large page mappings enabled?");
1561 
1562 bool
pmap_ps_enabled(pmap_t pmap __unused)1563 pmap_ps_enabled(pmap_t pmap __unused)
1564 {
1565 
1566 	return (sp_enabled != 0);
1567 }
1568 
1569 static SYSCTL_NODE(_vm_pmap, OID_AUTO, pte1, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
1570     "1MB page mapping counters");
1571 
1572 static u_long pmap_pte1_demotions;
1573 SYSCTL_ULONG(_vm_pmap_pte1, OID_AUTO, demotions, CTLFLAG_RD,
1574     &pmap_pte1_demotions, 0, "1MB page demotions");
1575 
1576 static u_long pmap_pte1_mappings;
1577 SYSCTL_ULONG(_vm_pmap_pte1, OID_AUTO, mappings, CTLFLAG_RD,
1578     &pmap_pte1_mappings, 0, "1MB page mappings");
1579 
1580 static u_long pmap_pte1_p_failures;
1581 SYSCTL_ULONG(_vm_pmap_pte1, OID_AUTO, p_failures, CTLFLAG_RD,
1582     &pmap_pte1_p_failures, 0, "1MB page promotion failures");
1583 
1584 static u_long pmap_pte1_promotions;
1585 SYSCTL_ULONG(_vm_pmap_pte1, OID_AUTO, promotions, CTLFLAG_RD,
1586     &pmap_pte1_promotions, 0, "1MB page promotions");
1587 
1588 static u_long pmap_pte1_kern_demotions;
1589 SYSCTL_ULONG(_vm_pmap_pte1, OID_AUTO, kern_demotions, CTLFLAG_RD,
1590     &pmap_pte1_kern_demotions, 0, "1MB page kernel demotions");
1591 
1592 static u_long pmap_pte1_kern_promotions;
1593 SYSCTL_ULONG(_vm_pmap_pte1, OID_AUTO, kern_promotions, CTLFLAG_RD,
1594     &pmap_pte1_kern_promotions, 0, "1MB page kernel promotions");
1595 
1596 static __inline ttb_entry_t
pmap_ttb_get(pmap_t pmap)1597 pmap_ttb_get(pmap_t pmap)
1598 {
1599 
1600 	return (vtophys(pmap->pm_pt1) | ttb_flags);
1601 }
1602 
1603 /*
1604  *  Initialize a vm_page's machine-dependent fields.
1605  *
1606  *  Variations:
1607  *  1. Pages for L2 page tables are always not managed. So, pv_list and
1608  *     pt2_wirecount can share same physical space. However, proper
1609  *     initialization on a page alloc for page tables and reinitialization
1610  *     on the page free must be ensured.
1611  */
1612 void
pmap_page_init(vm_page_t m)1613 pmap_page_init(vm_page_t m)
1614 {
1615 
1616 	TAILQ_INIT(&m->md.pv_list);
1617 	pt2_wirecount_init(m);
1618 	m->md.pat_mode = VM_MEMATTR_DEFAULT;
1619 }
1620 
1621 /*
1622  *  Virtualization for faster way how to zero whole page.
1623  */
1624 static __inline void
pagezero(void * page)1625 pagezero(void *page)
1626 {
1627 
1628 	bzero(page, PAGE_SIZE);
1629 }
1630 
1631 /*
1632  *  Zero L2 page table page.
1633  *  Use same KVA as in pmap_zero_page().
1634  */
1635 static __inline vm_paddr_t
pmap_pt2pg_zero(vm_page_t m)1636 pmap_pt2pg_zero(vm_page_t m)
1637 {
1638 	pt2_entry_t *cmap2_pte2p;
1639 	vm_paddr_t pa;
1640 	struct pcpu *pc;
1641 
1642 	pa = VM_PAGE_TO_PHYS(m);
1643 
1644 	/*
1645 	 * XXX: For now, we map whole page even if it's already zero,
1646 	 *      to sync it even if the sync is only DSB.
1647 	 */
1648 	sched_pin();
1649 	pc = get_pcpu();
1650 	cmap2_pte2p = pc->pc_cmap2_pte2p;
1651 	mtx_lock(&pc->pc_cmap_lock);
1652 	if (pte2_load(cmap2_pte2p) != 0)
1653 		panic("%s: CMAP2 busy", __func__);
1654 	pte2_store(cmap2_pte2p, PTE2_KERN_NG(pa, PTE2_AP_KRW,
1655 	    vm_page_pte2_attr(m)));
1656 	/*  Even VM_ALLOC_ZERO request is only advisory. */
1657 	if ((m->flags & PG_ZERO) == 0)
1658 		pagezero(pc->pc_cmap2_addr);
1659 	pte2_sync_range((pt2_entry_t *)pc->pc_cmap2_addr, PAGE_SIZE);
1660 	pte2_clear(cmap2_pte2p);
1661 	tlb_flush((vm_offset_t)pc->pc_cmap2_addr);
1662 
1663 	/*
1664 	 * Unpin the thread before releasing the lock.  Otherwise the thread
1665 	 * could be rescheduled while still bound to the current CPU, only
1666 	 * to unpin itself immediately upon resuming execution.
1667 	 */
1668 	sched_unpin();
1669 	mtx_unlock(&pc->pc_cmap_lock);
1670 
1671 	return (pa);
1672 }
1673 
1674 /*
1675  *  Init just allocated page as L2 page table(s) holder
1676  *  and return its physical address.
1677  */
1678 static __inline vm_paddr_t
pmap_pt2pg_init(pmap_t pmap,vm_offset_t va,vm_page_t m)1679 pmap_pt2pg_init(pmap_t pmap, vm_offset_t va, vm_page_t m)
1680 {
1681 	vm_paddr_t pa;
1682 	pt2_entry_t *pte2p;
1683 
1684 	/* Check page attributes. */
1685 	if (m->md.pat_mode != pt_memattr)
1686 		pmap_page_set_memattr(m, pt_memattr);
1687 
1688 	/* Zero page and init wire counts. */
1689 	pa = pmap_pt2pg_zero(m);
1690 	pt2_wirecount_init(m);
1691 
1692 	/*
1693 	 * Map page to PT2MAP address space for given pmap.
1694 	 * Note that PT2MAP space is shared with all pmaps.
1695 	 */
1696 	if (pmap == kernel_pmap)
1697 		pmap_kenter_pt2tab(va, PTE2_KPT(pa));
1698 	else {
1699 		pte2p = pmap_pt2tab_entry(pmap, va);
1700 		pt2tab_store(pte2p, PTE2_KPT_NG(pa));
1701 	}
1702 
1703 	return (pa);
1704 }
1705 
1706 /*
1707  *  Initialize the pmap module.
1708  *
1709  *  Called by vm_mem_init(), to initialize any structures that the pmap system
1710  *  needs to map virtual memory.
1711  */
1712 void
pmap_init(void)1713 pmap_init(void)
1714 {
1715 	vm_size_t s;
1716 	pt2_entry_t *pte2p, pte2;
1717 	u_int i, pte1_idx, pv_npg;
1718 
1719 	/*
1720 	 * Initialize the vm page array entries for kernel pmap's
1721 	 * L2 page table pages allocated in advance.
1722 	 */
1723 	pte1_idx = pte1_index(KERNBASE - PT2MAP_SIZE);
1724 	pte2p = kern_pt2tab_entry(KERNBASE - PT2MAP_SIZE);
1725 	for (i = 0; i < nkpt2pg + NPG_IN_PT2TAB; i++, pte2p++) {
1726 		vm_paddr_t pa;
1727 		vm_page_t m;
1728 
1729 		pte2 = pte2_load(pte2p);
1730 		KASSERT(pte2_is_valid(pte2), ("%s: no valid entry", __func__));
1731 
1732 		pa = pte2_pa(pte2);
1733 		m = PHYS_TO_VM_PAGE(pa);
1734 		KASSERT(m >= vm_page_array &&
1735 		    m < &vm_page_array[vm_page_array_size],
1736 		    ("%s: L2 page table page is out of range", __func__));
1737 
1738 		m->pindex = pte1_idx;
1739 		m->phys_addr = pa;
1740 		pte1_idx += NPT2_IN_PG;
1741 	}
1742 
1743 	/*
1744 	 * Initialize the address space (zone) for the pv entries.  Set a
1745 	 * high water mark so that the system can recover from excessive
1746 	 * numbers of pv entries.
1747 	 */
1748 	TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1749 	pv_entry_max = shpgperproc * maxproc + vm_cnt.v_page_count;
1750 	TUNABLE_INT_FETCH("vm.pmap.pv_entry_max", &pv_entry_max);
1751 	pv_entry_max = roundup(pv_entry_max, _NPCPV);
1752 	pv_entry_high_water = 9 * (pv_entry_max / 10);
1753 
1754 	/*
1755 	 * Are large page mappings enabled?
1756 	 */
1757 	TUNABLE_INT_FETCH("vm.pmap.sp_enabled", &sp_enabled);
1758 	if (sp_enabled) {
1759 		KASSERT(MAXPAGESIZES > 1 && pagesizes[1] == 0,
1760 		    ("%s: can't assign to pagesizes[1]", __func__));
1761 		pagesizes[1] = PTE1_SIZE;
1762 	}
1763 
1764 	/*
1765 	 * Calculate the size of the pv head table for sections.
1766 	 * Handle the possibility that "vm_phys_segs[...].end" is zero.
1767 	 * Note that the table is only for sections which could be promoted.
1768 	 */
1769 	first_managed_pa = pte1_trunc(vm_phys_segs[0].start);
1770 	pv_npg = (pte1_trunc(vm_phys_segs[vm_phys_nsegs - 1].end - PAGE_SIZE)
1771 	    - first_managed_pa) / PTE1_SIZE + 1;
1772 
1773 	/*
1774 	 * Allocate memory for the pv head table for sections.
1775 	 */
1776 	s = (vm_size_t)(pv_npg * sizeof(struct md_page));
1777 	s = round_page(s);
1778 	pv_table = kmem_malloc(s, M_WAITOK | M_ZERO);
1779 	for (i = 0; i < pv_npg; i++)
1780 		TAILQ_INIT(&pv_table[i].pv_list);
1781 
1782 	pv_maxchunks = MAX(pv_entry_max / _NPCPV, maxproc);
1783 	pv_chunkbase = (struct pv_chunk *)kva_alloc(PAGE_SIZE * pv_maxchunks);
1784 	if (pv_chunkbase == NULL)
1785 		panic("%s: not enough kvm for pv chunks", __func__);
1786 	pmap_pte2list_init(&pv_vafree, pv_chunkbase, pv_maxchunks);
1787 }
1788 
1789 /*
1790  *  Add a list of wired pages to the kva
1791  *  this routine is only used for temporary
1792  *  kernel mappings that do not need to have
1793  *  page modification or references recorded.
1794  *  Note that old mappings are simply written
1795  *  over.  The page *must* be wired.
1796  *  Note: SMP coherent.  Uses a ranged shootdown IPI.
1797  */
1798 void
pmap_qenter(vm_offset_t sva,vm_page_t * ma,int count)1799 pmap_qenter(vm_offset_t sva, vm_page_t *ma, int count)
1800 {
1801 	u_int anychanged;
1802 	pt2_entry_t *epte2p, *pte2p, pte2;
1803 	vm_page_t m;
1804 	vm_paddr_t pa;
1805 
1806 	anychanged = 0;
1807 	pte2p = pt2map_entry(sva);
1808 	epte2p = pte2p + count;
1809 	while (pte2p < epte2p) {
1810 		m = *ma++;
1811 		pa = VM_PAGE_TO_PHYS(m);
1812 		pte2 = pte2_load(pte2p);
1813 		if ((pte2_pa(pte2) != pa) ||
1814 		    (pte2_attr(pte2) != vm_page_pte2_attr(m))) {
1815 			anychanged++;
1816 			pte2_store(pte2p, PTE2_KERN(pa, PTE2_AP_KRW,
1817 			    vm_page_pte2_attr(m)));
1818 		}
1819 		pte2p++;
1820 	}
1821 	if (__predict_false(anychanged))
1822 		tlb_flush_range(sva, count * PAGE_SIZE);
1823 }
1824 
1825 /*
1826  *  This routine tears out page mappings from the
1827  *  kernel -- it is meant only for temporary mappings.
1828  *  Note: SMP coherent.  Uses a ranged shootdown IPI.
1829  */
1830 void
pmap_qremove(vm_offset_t sva,int count)1831 pmap_qremove(vm_offset_t sva, int count)
1832 {
1833 	vm_offset_t va;
1834 
1835 	va = sva;
1836 	while (count-- > 0) {
1837 		pmap_kremove(va);
1838 		va += PAGE_SIZE;
1839 	}
1840 	tlb_flush_range(sva, va - sva);
1841 }
1842 
1843 /*
1844  *  Are we current address space or kernel?
1845  */
1846 static __inline int
pmap_is_current(pmap_t pmap)1847 pmap_is_current(pmap_t pmap)
1848 {
1849 
1850 	return (pmap == kernel_pmap ||
1851 		(pmap == vmspace_pmap(curthread->td_proc->p_vmspace)));
1852 }
1853 
1854 /*
1855  *  If the given pmap is not the current or kernel pmap, the returned
1856  *  pte2 must be released by passing it to pmap_pte2_release().
1857  */
1858 static pt2_entry_t *
pmap_pte2(pmap_t pmap,vm_offset_t va)1859 pmap_pte2(pmap_t pmap, vm_offset_t va)
1860 {
1861 	pt1_entry_t pte1;
1862 	vm_paddr_t pt2pg_pa;
1863 
1864 	pte1 = pte1_load(pmap_pte1(pmap, va));
1865 	if (pte1_is_section(pte1))
1866 		panic("%s: attempt to map PTE1", __func__);
1867 	if (pte1_is_link(pte1)) {
1868 		/* Are we current address space or kernel? */
1869 		if (pmap_is_current(pmap))
1870 			return (pt2map_entry(va));
1871 		/* Note that L2 page table size is not equal to PAGE_SIZE. */
1872 		pt2pg_pa = trunc_page(pte1_link_pa(pte1));
1873 		mtx_lock(&PMAP2mutex);
1874 		if (pte2_pa(pte2_load(PMAP2)) != pt2pg_pa) {
1875 			pte2_store(PMAP2, PTE2_KPT(pt2pg_pa));
1876 			tlb_flush((vm_offset_t)PADDR2);
1877 		}
1878 		return (PADDR2 + (arm32_btop(va) & (NPTE2_IN_PG - 1)));
1879 	}
1880 	return (NULL);
1881 }
1882 
1883 /*
1884  *  Releases a pte2 that was obtained from pmap_pte2().
1885  *  Be prepared for the pte2p being NULL.
1886  */
1887 static __inline void
pmap_pte2_release(pt2_entry_t * pte2p)1888 pmap_pte2_release(pt2_entry_t *pte2p)
1889 {
1890 
1891 	if ((pt2_entry_t *)(trunc_page((vm_offset_t)pte2p)) == PADDR2) {
1892 		mtx_unlock(&PMAP2mutex);
1893 	}
1894 }
1895 
1896 /*
1897  *  Super fast pmap_pte2 routine best used when scanning
1898  *  the pv lists.  This eliminates many coarse-grained
1899  *  invltlb calls.  Note that many of the pv list
1900  *  scans are across different pmaps.  It is very wasteful
1901  *  to do an entire tlb flush for checking a single mapping.
1902  *
1903  *  If the given pmap is not the current pmap, pvh_global_lock
1904  *  must be held and curthread pinned to a CPU.
1905  */
1906 static pt2_entry_t *
pmap_pte2_quick(pmap_t pmap,vm_offset_t va)1907 pmap_pte2_quick(pmap_t pmap, vm_offset_t va)
1908 {
1909 	pt1_entry_t pte1;
1910 	vm_paddr_t pt2pg_pa;
1911 
1912 	pte1 = pte1_load(pmap_pte1(pmap, va));
1913 	if (pte1_is_section(pte1))
1914 		panic("%s: attempt to map PTE1", __func__);
1915 	if (pte1_is_link(pte1)) {
1916 		/* Are we current address space or kernel? */
1917 		if (pmap_is_current(pmap))
1918 			return (pt2map_entry(va));
1919 		rw_assert(&pvh_global_lock, RA_WLOCKED);
1920 		KASSERT(curthread->td_pinned > 0,
1921 		    ("%s: curthread not pinned", __func__));
1922 		/* Note that L2 page table size is not equal to PAGE_SIZE. */
1923 		pt2pg_pa = trunc_page(pte1_link_pa(pte1));
1924 		if (pte2_pa(pte2_load(PMAP1)) != pt2pg_pa) {
1925 			pte2_store(PMAP1, PTE2_KPT(pt2pg_pa));
1926 #ifdef SMP
1927 			PMAP1cpu = PCPU_GET(cpuid);
1928 #endif
1929 			tlb_flush_local((vm_offset_t)PADDR1);
1930 			PMAP1changed++;
1931 		} else
1932 #ifdef SMP
1933 		if (PMAP1cpu != PCPU_GET(cpuid)) {
1934 			PMAP1cpu = PCPU_GET(cpuid);
1935 			tlb_flush_local((vm_offset_t)PADDR1);
1936 			PMAP1changedcpu++;
1937 		} else
1938 #endif
1939 			PMAP1unchanged++;
1940 		return (PADDR1 + (arm32_btop(va) & (NPTE2_IN_PG - 1)));
1941 	}
1942 	return (NULL);
1943 }
1944 
1945 /*
1946  *  Routine: pmap_extract
1947  *  Function:
1948  * 	Extract the physical page address associated
1949  *	with the given map/virtual_address pair.
1950  */
1951 vm_paddr_t
pmap_extract(pmap_t pmap,vm_offset_t va)1952 pmap_extract(pmap_t pmap, vm_offset_t va)
1953 {
1954 	vm_paddr_t pa;
1955 	pt1_entry_t pte1;
1956 	pt2_entry_t *pte2p;
1957 
1958 	PMAP_LOCK(pmap);
1959 	pte1 = pte1_load(pmap_pte1(pmap, va));
1960 	if (pte1_is_section(pte1))
1961 		pa = pte1_pa(pte1) | (va & PTE1_OFFSET);
1962 	else if (pte1_is_link(pte1)) {
1963 		pte2p = pmap_pte2(pmap, va);
1964 		pa = pte2_pa(pte2_load(pte2p)) | (va & PTE2_OFFSET);
1965 		pmap_pte2_release(pte2p);
1966 	} else
1967 		pa = 0;
1968 	PMAP_UNLOCK(pmap);
1969 	return (pa);
1970 }
1971 
1972 /*
1973  *  Routine: pmap_extract_and_hold
1974  *  Function:
1975  *	Atomically extract and hold the physical page
1976  *	with the given pmap and virtual address pair
1977  *	if that mapping permits the given protection.
1978  */
1979 vm_page_t
pmap_extract_and_hold(pmap_t pmap,vm_offset_t va,vm_prot_t prot)1980 pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1981 {
1982 	vm_paddr_t pa;
1983 	pt1_entry_t pte1;
1984 	pt2_entry_t pte2, *pte2p;
1985 	vm_page_t m;
1986 
1987 	m = NULL;
1988 	PMAP_LOCK(pmap);
1989 	pte1 = pte1_load(pmap_pte1(pmap, va));
1990 	if (pte1_is_section(pte1)) {
1991 		if (!(pte1 & PTE1_RO) || !(prot & VM_PROT_WRITE)) {
1992 			pa = pte1_pa(pte1) | (va & PTE1_OFFSET);
1993 			m = PHYS_TO_VM_PAGE(pa);
1994 			if (!vm_page_wire_mapped(m))
1995 				m = NULL;
1996 		}
1997 	} else if (pte1_is_link(pte1)) {
1998 		pte2p = pmap_pte2(pmap, va);
1999 		pte2 = pte2_load(pte2p);
2000 		pmap_pte2_release(pte2p);
2001 		if (pte2_is_valid(pte2) &&
2002 		    (!(pte2 & PTE2_RO) || !(prot & VM_PROT_WRITE))) {
2003 			pa = pte2_pa(pte2);
2004 			m = PHYS_TO_VM_PAGE(pa);
2005 			if (!vm_page_wire_mapped(m))
2006 				m = NULL;
2007 		}
2008 	}
2009 	PMAP_UNLOCK(pmap);
2010 	return (m);
2011 }
2012 
2013 /*
2014  *  Grow the number of kernel L2 page table entries, if needed.
2015  */
2016 void
pmap_growkernel(vm_offset_t addr)2017 pmap_growkernel(vm_offset_t addr)
2018 {
2019 	vm_page_t m;
2020 	vm_paddr_t pt2pg_pa, pt2_pa;
2021 	pt1_entry_t pte1;
2022 	pt2_entry_t pte2;
2023 
2024 	PDEBUG(1, printf("%s: addr = %#x\n", __func__, addr));
2025 	/*
2026 	 * All the time kernel_vm_end is first KVA for which underlying
2027 	 * L2 page table is either not allocated or linked from L1 page table
2028 	 * (not considering sections). Except for two possible cases:
2029 	 *
2030 	 *   (1) in the very beginning as long as pmap_growkernel() was
2031 	 *       not called, it could be first unused KVA (which is not
2032 	 *       rounded up to PTE1_SIZE),
2033 	 *
2034 	 *   (2) when all KVA space is mapped and vm_map_max(kernel_map)
2035 	 *       address is not rounded up to PTE1_SIZE. (For example,
2036 	 *       it could be 0xFFFFFFFF.)
2037 	 */
2038 	kernel_vm_end = pte1_roundup(kernel_vm_end);
2039 	mtx_assert(&kernel_map->system_mtx, MA_OWNED);
2040 	addr = roundup2(addr, PTE1_SIZE);
2041 	if (addr - 1 >= vm_map_max(kernel_map))
2042 		addr = vm_map_max(kernel_map);
2043 	while (kernel_vm_end < addr) {
2044 		pte1 = pte1_load(kern_pte1(kernel_vm_end));
2045 		if (pte1_is_valid(pte1)) {
2046 			kernel_vm_end += PTE1_SIZE;
2047 			if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) {
2048 				kernel_vm_end = vm_map_max(kernel_map);
2049 				break;
2050 			}
2051 			continue;
2052 		}
2053 
2054 		/*
2055 		 * kernel_vm_end_new is used in pmap_pinit() when kernel
2056 		 * mappings are entered to new pmap all at once to avoid race
2057 		 * between pmap_kenter_pte1() and kernel_vm_end increase.
2058 		 * The same aplies to pmap_kenter_pt2tab().
2059 		 */
2060 		kernel_vm_end_new = kernel_vm_end + PTE1_SIZE;
2061 
2062 		pte2 = pt2tab_load(kern_pt2tab_entry(kernel_vm_end));
2063 		if (!pte2_is_valid(pte2)) {
2064 			/*
2065 			 * Install new PT2s page into kernel PT2TAB.
2066 			 */
2067 			m = vm_page_alloc_noobj(VM_ALLOC_INTERRUPT |
2068 			    VM_ALLOC_NOFREE | VM_ALLOC_WIRED | VM_ALLOC_ZERO);
2069 			if (m == NULL)
2070 				panic("%s: no memory to grow kernel", __func__);
2071 			m->pindex = pte1_index(kernel_vm_end) & ~PT2PG_MASK;
2072 
2073 			/*
2074 			 * QQQ: To link all new L2 page tables from L1 page
2075 			 *      table now and so pmap_kenter_pte1() them
2076 			 *      at once together with pmap_kenter_pt2tab()
2077 			 *      could be nice speed up. However,
2078 			 *      pmap_growkernel() does not happen so often...
2079 			 * QQQ: The other TTBR is another option.
2080 			 */
2081 			pt2pg_pa = pmap_pt2pg_init(kernel_pmap, kernel_vm_end,
2082 			    m);
2083 		} else
2084 			pt2pg_pa = pte2_pa(pte2);
2085 
2086 		pt2_pa = page_pt2pa(pt2pg_pa, pte1_index(kernel_vm_end));
2087 		pmap_kenter_pte1(kernel_vm_end, PTE1_LINK(pt2_pa));
2088 
2089 		kernel_vm_end = kernel_vm_end_new;
2090 		if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) {
2091 			kernel_vm_end = vm_map_max(kernel_map);
2092 			break;
2093 		}
2094 	}
2095 }
2096 
2097 static int
kvm_size(SYSCTL_HANDLER_ARGS)2098 kvm_size(SYSCTL_HANDLER_ARGS)
2099 {
2100 	unsigned long ksize = vm_max_kernel_address - KERNBASE;
2101 
2102 	return (sysctl_handle_long(oidp, &ksize, 0, req));
2103 }
2104 SYSCTL_PROC(_vm, OID_AUTO, kvm_size,
2105     CTLTYPE_LONG | CTLFLAG_RD | CTLFLAG_NEEDGIANT, 0, 0, kvm_size, "IU",
2106     "Size of KVM");
2107 
2108 static int
kvm_free(SYSCTL_HANDLER_ARGS)2109 kvm_free(SYSCTL_HANDLER_ARGS)
2110 {
2111 	unsigned long kfree = vm_max_kernel_address - kernel_vm_end;
2112 
2113 	return (sysctl_handle_long(oidp, &kfree, 0, req));
2114 }
2115 SYSCTL_PROC(_vm, OID_AUTO, kvm_free,
2116     CTLTYPE_LONG | CTLFLAG_RD | CTLFLAG_NEEDGIANT, 0, 0, kvm_free, "IU",
2117     "Amount of KVM free");
2118 
2119 /***********************************************
2120  *
2121  *  Pmap allocation/deallocation routines.
2122  *
2123  ***********************************************/
2124 
2125 /*
2126  *  Initialize the pmap for the swapper process.
2127  */
2128 void
pmap_pinit0(pmap_t pmap)2129 pmap_pinit0(pmap_t pmap)
2130 {
2131 	PDEBUG(1, printf("%s: pmap = %p\n", __func__, pmap));
2132 
2133 	PMAP_LOCK_INIT(pmap);
2134 
2135 	/*
2136 	 * Kernel page table directory and pmap stuff around is already
2137 	 * initialized, we are using it right now and here. So, finish
2138 	 * only PMAP structures initialization for process0 ...
2139 	 *
2140 	 * Since the L1 page table and PT2TAB is shared with the kernel pmap,
2141 	 * which is already included in the list "allpmaps", this pmap does
2142 	 * not need to be inserted into that list.
2143 	 */
2144 	pmap->pm_pt1 = kern_pt1;
2145 	pmap->pm_pt2tab = kern_pt2tab;
2146 	CPU_ZERO(&pmap->pm_active);
2147 	PCPU_SET(curpmap, pmap);
2148 	TAILQ_INIT(&pmap->pm_pvchunk);
2149 	bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
2150 	CPU_SET(0, &pmap->pm_active);
2151 }
2152 
2153 static __inline void
pte1_copy_nosync(pt1_entry_t * spte1p,pt1_entry_t * dpte1p,vm_offset_t sva,vm_offset_t eva)2154 pte1_copy_nosync(pt1_entry_t *spte1p, pt1_entry_t *dpte1p, vm_offset_t sva,
2155     vm_offset_t eva)
2156 {
2157 	u_int idx, count;
2158 
2159 	idx = pte1_index(sva);
2160 	count = (pte1_index(eva) - idx + 1) * sizeof(pt1_entry_t);
2161 	bcopy(spte1p + idx, dpte1p + idx, count);
2162 }
2163 
2164 static __inline void
pt2tab_copy_nosync(pt2_entry_t * spte2p,pt2_entry_t * dpte2p,vm_offset_t sva,vm_offset_t eva)2165 pt2tab_copy_nosync(pt2_entry_t *spte2p, pt2_entry_t *dpte2p, vm_offset_t sva,
2166     vm_offset_t eva)
2167 {
2168 	u_int idx, count;
2169 
2170 	idx = pt2tab_index(sva);
2171 	count = (pt2tab_index(eva) - idx + 1) * sizeof(pt2_entry_t);
2172 	bcopy(spte2p + idx, dpte2p + idx, count);
2173 }
2174 
2175 /*
2176  *  Initialize a preallocated and zeroed pmap structure,
2177  *  such as one in a vmspace structure.
2178  */
2179 int
pmap_pinit(pmap_t pmap)2180 pmap_pinit(pmap_t pmap)
2181 {
2182 	pt1_entry_t *pte1p;
2183 	pt2_entry_t *pte2p;
2184 	vm_paddr_t pa, pt2tab_pa;
2185 	u_int i;
2186 
2187 	PDEBUG(6, printf("%s: pmap = %p, pm_pt1 = %p\n", __func__, pmap,
2188 	    pmap->pm_pt1));
2189 
2190 	/*
2191 	 * No need to allocate L2 page table space yet but we do need
2192 	 * a valid L1 page table and PT2TAB table.
2193 	 *
2194 	 * Install shared kernel mappings to these tables. It's a little
2195 	 * tricky as some parts of KVA are reserved for vectors, devices,
2196 	 * and whatever else. These parts are supposed to be above
2197 	 * vm_max_kernel_address. Thus two regions should be installed:
2198 	 *
2199 	 *   (1) <KERNBASE, kernel_vm_end),
2200 	 *   (2) <vm_max_kernel_address, 0xFFFFFFFF>.
2201 	 *
2202 	 * QQQ: The second region should be stable enough to be installed
2203 	 *      only once in time when the tables are allocated.
2204 	 * QQQ: Maybe copy of both regions at once could be faster ...
2205 	 * QQQ: Maybe the other TTBR is an option.
2206 	 *
2207 	 * Finally, install own PT2TAB table to these tables.
2208 	 */
2209 
2210 	if (pmap->pm_pt1 == NULL) {
2211 		pmap->pm_pt1 = kmem_alloc_contig(NB_IN_PT1,
2212 		    M_NOWAIT | M_ZERO, 0, -1UL, NB_IN_PT1, 0, pt_memattr);
2213 		if (pmap->pm_pt1 == NULL)
2214 			return (0);
2215 	}
2216 	if (pmap->pm_pt2tab == NULL) {
2217 		/*
2218 		 * QQQ: (1) PT2TAB must be contiguous. If PT2TAB is one page
2219 		 *      only, what should be the only size for 32 bit systems,
2220 		 *      then we could allocate it with vm_page_alloc() and all
2221 		 *      the stuff needed as other L2 page table pages.
2222 		 *      (2) Note that a process PT2TAB is special L2 page table
2223 		 *      page. Its mapping in kernel_arena is permanent and can
2224 		 *      be used no matter which process is current. Its mapping
2225 		 *      in PT2MAP can be used only for current process.
2226 		 */
2227 		pmap->pm_pt2tab = kmem_alloc_attr(NB_IN_PT2TAB,
2228 		    M_NOWAIT | M_ZERO, 0, -1UL, pt_memattr);
2229 		if (pmap->pm_pt2tab == NULL) {
2230 			/*
2231 			 * QQQ: As struct pmap is allocated from UMA with
2232 			 *      UMA_ZONE_NOFREE flag, it's important to leave
2233 			 *      no allocation in pmap if initialization failed.
2234 			 */
2235 			kmem_free(pmap->pm_pt1, NB_IN_PT1);
2236 			pmap->pm_pt1 = NULL;
2237 			return (0);
2238 		}
2239 		/*
2240 		 * QQQ: Each L2 page table page vm_page_t has pindex set to
2241 		 *      pte1 index of virtual address mapped by this page.
2242 		 *      It's not valid for non kernel PT2TABs themselves.
2243 		 *      The pindex of these pages can not be altered because
2244 		 *      of the way how they are allocated now. However, it
2245 		 *      should not be a problem.
2246 		 */
2247 	}
2248 
2249 	mtx_lock_spin(&allpmaps_lock);
2250 	/*
2251 	 * To avoid race with pmap_kenter_pte1() and pmap_kenter_pt2tab(),
2252 	 * kernel_vm_end_new is used here instead of kernel_vm_end.
2253 	 */
2254 	pte1_copy_nosync(kern_pt1, pmap->pm_pt1, KERNBASE,
2255 	    kernel_vm_end_new - 1);
2256 	pte1_copy_nosync(kern_pt1, pmap->pm_pt1, vm_max_kernel_address,
2257 	    0xFFFFFFFF);
2258 	pt2tab_copy_nosync(kern_pt2tab, pmap->pm_pt2tab, KERNBASE,
2259 	    kernel_vm_end_new - 1);
2260 	pt2tab_copy_nosync(kern_pt2tab, pmap->pm_pt2tab, vm_max_kernel_address,
2261 	    0xFFFFFFFF);
2262 	LIST_INSERT_HEAD(&allpmaps, pmap, pm_list);
2263 	mtx_unlock_spin(&allpmaps_lock);
2264 
2265 	/*
2266 	 * Store PT2MAP PT2 pages (a.k.a. PT2TAB) in PT2TAB itself.
2267 	 * I.e. self reference mapping.  The PT2TAB is private, however mapped
2268 	 * into shared PT2MAP space, so the mapping should be not global.
2269 	 */
2270 	pt2tab_pa = vtophys(pmap->pm_pt2tab);
2271 	pte2p = pmap_pt2tab_entry(pmap, (vm_offset_t)PT2MAP);
2272 	for (pa = pt2tab_pa, i = 0; i < NPG_IN_PT2TAB; i++, pa += PTE2_SIZE) {
2273 		pt2tab_store(pte2p++, PTE2_KPT_NG(pa));
2274 	}
2275 
2276 	/* Insert PT2MAP PT2s into pmap PT1. */
2277 	pte1p = pmap_pte1(pmap, (vm_offset_t)PT2MAP);
2278 	for (pa = pt2tab_pa, i = 0; i < NPT2_IN_PT2TAB; i++, pa += NB_IN_PT2) {
2279 		pte1_store(pte1p++, PTE1_LINK(pa));
2280 	}
2281 
2282 	/*
2283 	 * Now synchronize new mapping which was made above.
2284 	 */
2285 	pte1_sync_range(pmap->pm_pt1, NB_IN_PT1);
2286 	pte2_sync_range(pmap->pm_pt2tab, NB_IN_PT2TAB);
2287 
2288 	CPU_ZERO(&pmap->pm_active);
2289 	TAILQ_INIT(&pmap->pm_pvchunk);
2290 	bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
2291 
2292 	return (1);
2293 }
2294 
2295 #ifdef INVARIANTS
2296 static bool
pt2tab_user_is_empty(pt2_entry_t * tab)2297 pt2tab_user_is_empty(pt2_entry_t *tab)
2298 {
2299 	u_int i, end;
2300 
2301 	end = pt2tab_index(VM_MAXUSER_ADDRESS);
2302 	for (i = 0; i < end; i++)
2303 		if (tab[i] != 0) return (false);
2304 	return (true);
2305 }
2306 #endif
2307 /*
2308  *  Release any resources held by the given physical map.
2309  *  Called when a pmap initialized by pmap_pinit is being released.
2310  *  Should only be called if the map contains no valid mappings.
2311  */
2312 void
pmap_release(pmap_t pmap)2313 pmap_release(pmap_t pmap)
2314 {
2315 #ifdef INVARIANTS
2316 	vm_offset_t start, end;
2317 #endif
2318 	KASSERT(pmap->pm_stats.resident_count == 0,
2319 	    ("%s: pmap resident count %ld != 0", __func__,
2320 	    pmap->pm_stats.resident_count));
2321 	KASSERT(pt2tab_user_is_empty(pmap->pm_pt2tab),
2322 	    ("%s: has allocated user PT2(s)", __func__));
2323 	KASSERT(CPU_EMPTY(&pmap->pm_active),
2324 	    ("%s: pmap %p is active on some CPU(s)", __func__, pmap));
2325 
2326 	mtx_lock_spin(&allpmaps_lock);
2327 	LIST_REMOVE(pmap, pm_list);
2328 	mtx_unlock_spin(&allpmaps_lock);
2329 
2330 #ifdef INVARIANTS
2331 	start = pte1_index(KERNBASE) * sizeof(pt1_entry_t);
2332 	end = (pte1_index(0xFFFFFFFF) + 1) * sizeof(pt1_entry_t);
2333 	bzero((char *)pmap->pm_pt1 + start, end - start);
2334 
2335 	start = pt2tab_index(KERNBASE) * sizeof(pt2_entry_t);
2336 	end = (pt2tab_index(0xFFFFFFFF) + 1) * sizeof(pt2_entry_t);
2337 	bzero((char *)pmap->pm_pt2tab + start, end - start);
2338 #endif
2339 	/*
2340 	 * We are leaving PT1 and PT2TAB allocated on released pmap,
2341 	 * so hopefully UMA vmspace_zone will always be inited with
2342 	 * UMA_ZONE_NOFREE flag.
2343 	 */
2344 }
2345 
2346 /*********************************************************
2347  *
2348  *  L2 table pages and their pages management routines.
2349  *
2350  *********************************************************/
2351 
2352 /*
2353  *  Virtual interface for L2 page table wire counting.
2354  *
2355  *  Each L2 page table in a page has own counter which counts a number of
2356  *  valid mappings in a table. Global page counter counts mappings in all
2357  *  tables in a page plus a single itself mapping in PT2TAB.
2358  *
2359  *  During a promotion we leave the associated L2 page table counter
2360  *  untouched, so the table (strictly speaking a page which holds it)
2361  *  is never freed if promoted.
2362  *
2363  *  If a page m->ref_count == 1 then no valid mappings exist in any L2 page
2364  *  table in the page and the page itself is only mapped in PT2TAB.
2365  */
2366 
2367 static __inline void
pt2_wirecount_init(vm_page_t m)2368 pt2_wirecount_init(vm_page_t m)
2369 {
2370 	u_int i;
2371 
2372 	/*
2373 	 * Note: A page m is allocated with VM_ALLOC_WIRED flag and
2374 	 *       m->ref_count should be already set correctly.
2375 	 *       So, there is no need to set it again herein.
2376 	 */
2377 	for (i = 0; i < NPT2_IN_PG; i++)
2378 		m->md.pt2_wirecount[i] = 0;
2379 }
2380 
2381 static __inline void
pt2_wirecount_inc(vm_page_t m,uint32_t pte1_idx)2382 pt2_wirecount_inc(vm_page_t m, uint32_t pte1_idx)
2383 {
2384 
2385 	/*
2386 	 * Note: A just modificated pte2 (i.e. already allocated)
2387 	 *       is acquiring one extra reference which must be
2388 	 *       explicitly cleared. It influences the KASSERTs herein.
2389 	 *       All L2 page tables in a page always belong to the same
2390 	 *       pmap, so we allow only one extra reference for the page.
2391 	 */
2392 	KASSERT(m->md.pt2_wirecount[pte1_idx & PT2PG_MASK] < (NPTE2_IN_PT2 + 1),
2393 	    ("%s: PT2 is overflowing ...", __func__));
2394 	KASSERT(m->ref_count <= (NPTE2_IN_PG + 1),
2395 	    ("%s: PT2PG is overflowing ...", __func__));
2396 
2397 	m->ref_count++;
2398 	m->md.pt2_wirecount[pte1_idx & PT2PG_MASK]++;
2399 }
2400 
2401 static __inline void
pt2_wirecount_dec(vm_page_t m,uint32_t pte1_idx)2402 pt2_wirecount_dec(vm_page_t m, uint32_t pte1_idx)
2403 {
2404 
2405 	KASSERT(m->md.pt2_wirecount[pte1_idx & PT2PG_MASK] != 0,
2406 	    ("%s: PT2 is underflowing ...", __func__));
2407 	KASSERT(m->ref_count > 1,
2408 	    ("%s: PT2PG is underflowing ...", __func__));
2409 
2410 	m->ref_count--;
2411 	m->md.pt2_wirecount[pte1_idx & PT2PG_MASK]--;
2412 }
2413 
2414 static __inline void
pt2_wirecount_set(vm_page_t m,uint32_t pte1_idx,uint16_t count)2415 pt2_wirecount_set(vm_page_t m, uint32_t pte1_idx, uint16_t count)
2416 {
2417 
2418 	KASSERT(count <= NPTE2_IN_PT2,
2419 	    ("%s: invalid count %u", __func__, count));
2420 	KASSERT(m->ref_count >  m->md.pt2_wirecount[pte1_idx & PT2PG_MASK],
2421 	    ("%s: PT2PG corrupting (%u, %u) ...", __func__, m->ref_count,
2422 	    m->md.pt2_wirecount[pte1_idx & PT2PG_MASK]));
2423 
2424 	m->ref_count -= m->md.pt2_wirecount[pte1_idx & PT2PG_MASK];
2425 	m->ref_count += count;
2426 	m->md.pt2_wirecount[pte1_idx & PT2PG_MASK] = count;
2427 
2428 	KASSERT(m->ref_count <= (NPTE2_IN_PG + 1),
2429 	    ("%s: PT2PG is overflowed (%u) ...", __func__, m->ref_count));
2430 }
2431 
2432 static __inline uint32_t
pt2_wirecount_get(vm_page_t m,uint32_t pte1_idx)2433 pt2_wirecount_get(vm_page_t m, uint32_t pte1_idx)
2434 {
2435 
2436 	return (m->md.pt2_wirecount[pte1_idx & PT2PG_MASK]);
2437 }
2438 
2439 static __inline bool
pt2_is_empty(vm_page_t m,vm_offset_t va)2440 pt2_is_empty(vm_page_t m, vm_offset_t va)
2441 {
2442 
2443 	return (m->md.pt2_wirecount[pte1_index(va) & PT2PG_MASK] == 0);
2444 }
2445 
2446 static __inline bool
pt2_is_full(vm_page_t m,vm_offset_t va)2447 pt2_is_full(vm_page_t m, vm_offset_t va)
2448 {
2449 
2450 	return (m->md.pt2_wirecount[pte1_index(va) & PT2PG_MASK] ==
2451 	    NPTE2_IN_PT2);
2452 }
2453 
2454 static __inline bool
pt2pg_is_empty(vm_page_t m)2455 pt2pg_is_empty(vm_page_t m)
2456 {
2457 
2458 	return (m->ref_count == 1);
2459 }
2460 
2461 /*
2462  *  This routine is called if the L2 page table
2463  *  is not mapped correctly.
2464  */
2465 static vm_page_t
_pmap_allocpte2(pmap_t pmap,vm_offset_t va,u_int flags)2466 _pmap_allocpte2(pmap_t pmap, vm_offset_t va, u_int flags)
2467 {
2468 	uint32_t pte1_idx;
2469 	pt1_entry_t *pte1p;
2470 	pt2_entry_t pte2;
2471 	vm_page_t  m;
2472 	vm_paddr_t pt2pg_pa, pt2_pa;
2473 
2474 	pte1_idx = pte1_index(va);
2475 	pte1p = pmap->pm_pt1 + pte1_idx;
2476 
2477 	KASSERT(pte1_load(pte1p) == 0,
2478 	    ("%s: pm_pt1[%#x] is not zero: %#x", __func__, pte1_idx,
2479 	    pte1_load(pte1p)));
2480 
2481 	pte2 = pt2tab_load(pmap_pt2tab_entry(pmap, va));
2482 	if (!pte2_is_valid(pte2)) {
2483 		/*
2484 		 * Install new PT2s page into pmap PT2TAB.
2485 		 */
2486 		m = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_ZERO);
2487 		if (m == NULL) {
2488 			if ((flags & PMAP_ENTER_NOSLEEP) == 0) {
2489 				PMAP_UNLOCK(pmap);
2490 				rw_wunlock(&pvh_global_lock);
2491 				vm_wait(NULL);
2492 				rw_wlock(&pvh_global_lock);
2493 				PMAP_LOCK(pmap);
2494 			}
2495 
2496 			/*
2497 			 * Indicate the need to retry.  While waiting,
2498 			 * the L2 page table page may have been allocated.
2499 			 */
2500 			return (NULL);
2501 		}
2502 		m->pindex = pte1_idx & ~PT2PG_MASK;
2503 		pmap->pm_stats.resident_count++;
2504 		pt2pg_pa = pmap_pt2pg_init(pmap, va, m);
2505 	} else {
2506 		pt2pg_pa = pte2_pa(pte2);
2507 		m = PHYS_TO_VM_PAGE(pt2pg_pa);
2508 	}
2509 
2510 	pt2_wirecount_inc(m, pte1_idx);
2511 	pt2_pa = page_pt2pa(pt2pg_pa, pte1_idx);
2512 	pte1_store(pte1p, PTE1_LINK(pt2_pa));
2513 
2514 	return (m);
2515 }
2516 
2517 static vm_page_t
pmap_allocpte2(pmap_t pmap,vm_offset_t va,u_int flags)2518 pmap_allocpte2(pmap_t pmap, vm_offset_t va, u_int flags)
2519 {
2520 	u_int pte1_idx;
2521 	pt1_entry_t *pte1p, pte1;
2522 	vm_page_t m;
2523 
2524 	pte1_idx = pte1_index(va);
2525 retry:
2526 	pte1p = pmap->pm_pt1 + pte1_idx;
2527 	pte1 = pte1_load(pte1p);
2528 
2529 	/*
2530 	 * This supports switching from a 1MB page to a
2531 	 * normal 4K page.
2532 	 */
2533 	if (pte1_is_section(pte1)) {
2534 		(void)pmap_demote_pte1(pmap, pte1p, va);
2535 		/*
2536 		 * Reload pte1 after demotion.
2537 		 *
2538 		 * Note: Demotion can even fail as either PT2 is not find for
2539 		 *       the virtual address or PT2PG can not be allocated.
2540 		 */
2541 		pte1 = pte1_load(pte1p);
2542 	}
2543 
2544 	/*
2545 	 * If the L2 page table page is mapped, we just increment the
2546 	 * hold count, and activate it.
2547 	 */
2548 	if (pte1_is_link(pte1)) {
2549 		m = PHYS_TO_VM_PAGE(pte1_link_pa(pte1));
2550 		pt2_wirecount_inc(m, pte1_idx);
2551 	} else  {
2552 		/*
2553 		 * Here if the PT2 isn't mapped, or if it has
2554 		 * been deallocated.
2555 		 */
2556 		m = _pmap_allocpte2(pmap, va, flags);
2557 		if (m == NULL && (flags & PMAP_ENTER_NOSLEEP) == 0)
2558 			goto retry;
2559 	}
2560 
2561 	return (m);
2562 }
2563 
2564 /*
2565  *  Schedule the specified unused L2 page table page to be freed. Specifically,
2566  *  add the page to the specified list of pages that will be released to the
2567  *  physical memory manager after the TLB has been updated.
2568  */
2569 static __inline void
pmap_add_delayed_free_list(vm_page_t m,struct spglist * free)2570 pmap_add_delayed_free_list(vm_page_t m, struct spglist *free)
2571 {
2572 
2573 	/*
2574 	 * Put page on a list so that it is released after
2575 	 * *ALL* TLB shootdown is done
2576 	 */
2577 #ifdef PMAP_DEBUG
2578 	pmap_zero_page_check(m);
2579 #endif
2580 	m->flags |= PG_ZERO;
2581 	SLIST_INSERT_HEAD(free, m, plinks.s.ss);
2582 }
2583 
2584 /*
2585  *  Unwire L2 page tables page.
2586  */
2587 static void
pmap_unwire_pt2pg(pmap_t pmap,vm_offset_t va,vm_page_t m)2588 pmap_unwire_pt2pg(pmap_t pmap, vm_offset_t va, vm_page_t m)
2589 {
2590 	pt1_entry_t *pte1p, opte1 __unused;
2591 	pt2_entry_t *pte2p;
2592 	uint32_t i;
2593 
2594 	KASSERT(pt2pg_is_empty(m),
2595 	    ("%s: pmap %p PT2PG %p wired", __func__, pmap, m));
2596 
2597 	/*
2598 	 * Unmap all L2 page tables in the page from L1 page table.
2599 	 *
2600 	 * QQQ: Individual L2 page tables (except the last one) can be unmapped
2601 	 * earlier. However, we are doing that this way.
2602 	 */
2603 	KASSERT(m->pindex == (pte1_index(va) & ~PT2PG_MASK),
2604 	    ("%s: pmap %p va %#x PT2PG %p bad index", __func__, pmap, va, m));
2605 	pte1p = pmap->pm_pt1 + m->pindex;
2606 	for (i = 0; i < NPT2_IN_PG; i++, pte1p++) {
2607 		KASSERT(m->md.pt2_wirecount[i] == 0,
2608 		    ("%s: pmap %p PT2 %u (PG %p) wired", __func__, pmap, i, m));
2609 		opte1 = pte1_load(pte1p);
2610 		if (pte1_is_link(opte1)) {
2611 			pte1_clear(pte1p);
2612 			/*
2613 			 * Flush intermediate TLB cache.
2614 			 */
2615 			pmap_tlb_flush(pmap, (m->pindex + i) << PTE1_SHIFT);
2616 		}
2617 #ifdef INVARIANTS
2618 		else
2619 			KASSERT((opte1 == 0) || pte1_is_section(opte1),
2620 			    ("%s: pmap %p va %#x bad pte1 %x at %u", __func__,
2621 			    pmap, va, opte1, i));
2622 #endif
2623 	}
2624 
2625 	/*
2626 	 * Unmap the page from PT2TAB.
2627 	 */
2628 	pte2p = pmap_pt2tab_entry(pmap, va);
2629 	(void)pt2tab_load_clear(pte2p);
2630 	pmap_tlb_flush(pmap, pt2map_pt2pg(va));
2631 
2632 	m->ref_count = 0;
2633 	pmap->pm_stats.resident_count--;
2634 
2635 	/*
2636 	 * This barrier is so that the ordinary store unmapping
2637 	 * the L2 page table page is globally performed before TLB shoot-
2638 	 * down is begun.
2639 	 */
2640 	wmb();
2641 	vm_wire_sub(1);
2642 }
2643 
2644 /*
2645  *  Decrements a L2 page table page's wire count, which is used to record the
2646  *  number of valid page table entries within the page.  If the wire count
2647  *  drops to zero, then the page table page is unmapped.  Returns true if the
2648  *  page table page was unmapped and false otherwise.
2649  */
2650 static __inline bool
pmap_unwire_pt2(pmap_t pmap,vm_offset_t va,vm_page_t m,struct spglist * free)2651 pmap_unwire_pt2(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free)
2652 {
2653 	pt2_wirecount_dec(m, pte1_index(va));
2654 	if (pt2pg_is_empty(m)) {
2655 		/*
2656 		 * QQQ: Wire count is zero, so whole page should be zero and
2657 		 *      we can set PG_ZERO flag to it.
2658 		 *      Note that when promotion is enabled, it takes some
2659 		 *      more efforts. See pmap_unwire_pt2_all() below.
2660 		 */
2661 		pmap_unwire_pt2pg(pmap, va, m);
2662 		pmap_add_delayed_free_list(m, free);
2663 		return (true);
2664 	} else
2665 		return (false);
2666 }
2667 
2668 /*
2669  *  Drop a L2 page table page's wire count at once, which is used to record
2670  *  the number of valid L2 page table entries within the page. If the wire
2671  *  count drops to zero, then the L2 page table page is unmapped.
2672  */
2673 static __inline void
pmap_unwire_pt2_all(pmap_t pmap,vm_offset_t va,vm_page_t m,struct spglist * free)2674 pmap_unwire_pt2_all(pmap_t pmap, vm_offset_t va, vm_page_t m,
2675     struct spglist *free)
2676 {
2677 	u_int pte1_idx = pte1_index(va);
2678 
2679 	KASSERT(m->pindex == (pte1_idx & ~PT2PG_MASK),
2680 		("%s: PT2 page's pindex is wrong", __func__));
2681 	KASSERT(m->ref_count > pt2_wirecount_get(m, pte1_idx),
2682 	    ("%s: bad pt2 wire count %u > %u", __func__, m->ref_count,
2683 	    pt2_wirecount_get(m, pte1_idx)));
2684 
2685 	/*
2686 	 * It's possible that the L2 page table was never used.
2687 	 * It happened in case that a section was created without promotion.
2688 	 */
2689 	if (pt2_is_full(m, va)) {
2690 		pt2_wirecount_set(m, pte1_idx, 0);
2691 
2692 		/*
2693 		 * QQQ: We clear L2 page table now, so when L2 page table page
2694 		 *      is going to be freed, we can set it PG_ZERO flag ...
2695 		 *      This function is called only on section mappings, so
2696 		 *      hopefully it's not to big overload.
2697 		 *
2698 		 * XXX: If pmap is current, existing PT2MAP mapping could be
2699 		 *      used for zeroing.
2700 		 */
2701 		pmap_zero_page_area(m, page_pt2off(pte1_idx), NB_IN_PT2);
2702 	}
2703 #ifdef INVARIANTS
2704 	else
2705 		KASSERT(pt2_is_empty(m, va), ("%s: PT2 is not empty (%u)",
2706 		    __func__, pt2_wirecount_get(m, pte1_idx)));
2707 #endif
2708 	if (pt2pg_is_empty(m)) {
2709 		pmap_unwire_pt2pg(pmap, va, m);
2710 		pmap_add_delayed_free_list(m, free);
2711 	}
2712 }
2713 
2714 /*
2715  *  After removing a L2 page table entry, this routine is used to
2716  *  conditionally free the page, and manage the hold/wire counts.
2717  */
2718 static bool
pmap_unuse_pt2(pmap_t pmap,vm_offset_t va,struct spglist * free)2719 pmap_unuse_pt2(pmap_t pmap, vm_offset_t va, struct spglist *free)
2720 {
2721 	pt1_entry_t pte1;
2722 	vm_page_t mpte;
2723 
2724 	if (va >= VM_MAXUSER_ADDRESS)
2725 		return (false);
2726 	pte1 = pte1_load(pmap_pte1(pmap, va));
2727 	mpte = PHYS_TO_VM_PAGE(pte1_link_pa(pte1));
2728 	return (pmap_unwire_pt2(pmap, va, mpte, free));
2729 }
2730 
2731 /*************************************
2732  *
2733  *  Page management routines.
2734  *
2735  *************************************/
2736 
2737 static const uint32_t pc_freemask[_NPCM] = {
2738 	[0 ... _NPCM - 2] = PC_FREEN,
2739 	[_NPCM - 1] = PC_FREEL
2740 };
2741 
2742 SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_count, CTLFLAG_RD, &pv_entry_count, 0,
2743 	"Current number of pv entries");
2744 
2745 #ifdef PV_STATS
2746 static int pc_chunk_count, pc_chunk_allocs, pc_chunk_frees, pc_chunk_tryfail;
2747 
2748 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_count, CTLFLAG_RD, &pc_chunk_count, 0,
2749     "Current number of pv entry chunks");
2750 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_allocs, CTLFLAG_RD, &pc_chunk_allocs, 0,
2751     "Current number of pv entry chunks allocated");
2752 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_frees, CTLFLAG_RD, &pc_chunk_frees, 0,
2753     "Current number of pv entry chunks frees");
2754 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_tryfail, CTLFLAG_RD, &pc_chunk_tryfail,
2755     0, "Number of times tried to get a chunk page but failed.");
2756 
2757 static long pv_entry_frees, pv_entry_allocs;
2758 static int pv_entry_spare;
2759 
2760 SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_frees, CTLFLAG_RD, &pv_entry_frees, 0,
2761     "Current number of pv entry frees");
2762 SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_allocs, CTLFLAG_RD, &pv_entry_allocs,
2763     0, "Current number of pv entry allocs");
2764 SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_spare, CTLFLAG_RD, &pv_entry_spare, 0,
2765     "Current number of spare pv entries");
2766 #endif
2767 
2768 /*
2769  *  Is given page managed?
2770  */
2771 static __inline bool
is_managed(vm_paddr_t pa)2772 is_managed(vm_paddr_t pa)
2773 {
2774 	vm_page_t m;
2775 
2776 	m = PHYS_TO_VM_PAGE(pa);
2777 	if (m == NULL)
2778 		return (false);
2779 	return ((m->oflags & VPO_UNMANAGED) == 0);
2780 }
2781 
2782 static __inline bool
pte1_is_managed(pt1_entry_t pte1)2783 pte1_is_managed(pt1_entry_t pte1)
2784 {
2785 
2786 	return (is_managed(pte1_pa(pte1)));
2787 }
2788 
2789 static __inline bool
pte2_is_managed(pt2_entry_t pte2)2790 pte2_is_managed(pt2_entry_t pte2)
2791 {
2792 
2793 	return (is_managed(pte2_pa(pte2)));
2794 }
2795 
2796 /*
2797  *  We are in a serious low memory condition.  Resort to
2798  *  drastic measures to free some pages so we can allocate
2799  *  another pv entry chunk.
2800  */
2801 static vm_page_t
pmap_pv_reclaim(pmap_t locked_pmap)2802 pmap_pv_reclaim(pmap_t locked_pmap)
2803 {
2804 	struct pch newtail;
2805 	struct pv_chunk *pc;
2806 	struct md_page *pvh;
2807 	pt1_entry_t *pte1p;
2808 	pmap_t pmap;
2809 	pt2_entry_t *pte2p, tpte2;
2810 	pv_entry_t pv;
2811 	vm_offset_t va;
2812 	vm_page_t m, m_pc;
2813 	struct spglist free;
2814 	uint32_t inuse;
2815 	int bit, field, freed;
2816 
2817 	PMAP_LOCK_ASSERT(locked_pmap, MA_OWNED);
2818 	pmap = NULL;
2819 	m_pc = NULL;
2820 	SLIST_INIT(&free);
2821 	TAILQ_INIT(&newtail);
2822 	while ((pc = TAILQ_FIRST(&pv_chunks)) != NULL && (pv_vafree == 0 ||
2823 	    SLIST_EMPTY(&free))) {
2824 		TAILQ_REMOVE(&pv_chunks, pc, pc_lru);
2825 		if (pmap != pc->pc_pmap) {
2826 			if (pmap != NULL) {
2827 				if (pmap != locked_pmap)
2828 					PMAP_UNLOCK(pmap);
2829 			}
2830 			pmap = pc->pc_pmap;
2831 			/* Avoid deadlock and lock recursion. */
2832 			if (pmap > locked_pmap)
2833 				PMAP_LOCK(pmap);
2834 			else if (pmap != locked_pmap && !PMAP_TRYLOCK(pmap)) {
2835 				pmap = NULL;
2836 				TAILQ_INSERT_TAIL(&newtail, pc, pc_lru);
2837 				continue;
2838 			}
2839 		}
2840 
2841 		/*
2842 		 * Destroy every non-wired, 4 KB page mapping in the chunk.
2843 		 */
2844 		freed = 0;
2845 		for (field = 0; field < _NPCM; field++) {
2846 			for (inuse = ~pc->pc_map[field] & pc_freemask[field];
2847 			    inuse != 0; inuse &= ~(1UL << bit)) {
2848 				bit = ffs(inuse) - 1;
2849 				pv = &pc->pc_pventry[field * 32 + bit];
2850 				va = pv->pv_va;
2851 				pte1p = pmap_pte1(pmap, va);
2852 				if (pte1_is_section(pte1_load(pte1p)))
2853 					continue;
2854 				pte2p = pmap_pte2(pmap, va);
2855 				tpte2 = pte2_load(pte2p);
2856 				if ((tpte2 & PTE2_W) == 0)
2857 					tpte2 = pte2_load_clear(pte2p);
2858 				pmap_pte2_release(pte2p);
2859 				if ((tpte2 & PTE2_W) != 0)
2860 					continue;
2861 				KASSERT(tpte2 != 0,
2862 				    ("pmap_pv_reclaim: pmap %p va %#x zero pte",
2863 				    pmap, va));
2864 				pmap_tlb_flush(pmap, va);
2865 				m = PHYS_TO_VM_PAGE(pte2_pa(tpte2));
2866 				if (pte2_is_dirty(tpte2))
2867 					vm_page_dirty(m);
2868 				if ((tpte2 & PTE2_A) != 0)
2869 					vm_page_aflag_set(m, PGA_REFERENCED);
2870 				TAILQ_REMOVE(&m->md.pv_list, pv, pv_next);
2871 				if (TAILQ_EMPTY(&m->md.pv_list) &&
2872 				    (m->flags & PG_FICTITIOUS) == 0) {
2873 					pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
2874 					if (TAILQ_EMPTY(&pvh->pv_list)) {
2875 						vm_page_aflag_clear(m,
2876 						    PGA_WRITEABLE);
2877 					}
2878 				}
2879 				pc->pc_map[field] |= 1UL << bit;
2880 				pmap_unuse_pt2(pmap, va, &free);
2881 				freed++;
2882 			}
2883 		}
2884 		if (freed == 0) {
2885 			TAILQ_INSERT_TAIL(&newtail, pc, pc_lru);
2886 			continue;
2887 		}
2888 		/* Every freed mapping is for a 4 KB page. */
2889 		pmap->pm_stats.resident_count -= freed;
2890 		PV_STAT(pv_entry_frees += freed);
2891 		PV_STAT(pv_entry_spare += freed);
2892 		pv_entry_count -= freed;
2893 		TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
2894 		for (field = 0; field < _NPCM; field++)
2895 			if (pc->pc_map[field] != pc_freemask[field]) {
2896 				TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc,
2897 				    pc_list);
2898 				TAILQ_INSERT_TAIL(&newtail, pc, pc_lru);
2899 
2900 				/*
2901 				 * One freed pv entry in locked_pmap is
2902 				 * sufficient.
2903 				 */
2904 				if (pmap == locked_pmap)
2905 					goto out;
2906 				break;
2907 			}
2908 		if (field == _NPCM) {
2909 			PV_STAT(pv_entry_spare -= _NPCPV);
2910 			PV_STAT(pc_chunk_count--);
2911 			PV_STAT(pc_chunk_frees++);
2912 			/* Entire chunk is free; return it. */
2913 			m_pc = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)pc));
2914 			pmap_qremove((vm_offset_t)pc, 1);
2915 			pmap_pte2list_free(&pv_vafree, (vm_offset_t)pc);
2916 			break;
2917 		}
2918 	}
2919 out:
2920 	TAILQ_CONCAT(&pv_chunks, &newtail, pc_lru);
2921 	if (pmap != NULL) {
2922 		if (pmap != locked_pmap)
2923 			PMAP_UNLOCK(pmap);
2924 	}
2925 	if (m_pc == NULL && pv_vafree != 0 && SLIST_EMPTY(&free)) {
2926 		m_pc = SLIST_FIRST(&free);
2927 		SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2928 		/* Recycle a freed page table page. */
2929 		m_pc->ref_count = 1;
2930 		vm_wire_add(1);
2931 	}
2932 	vm_page_free_pages_toq(&free, false);
2933 	return (m_pc);
2934 }
2935 
2936 static void
free_pv_chunk(struct pv_chunk * pc)2937 free_pv_chunk(struct pv_chunk *pc)
2938 {
2939 	vm_page_t m;
2940 
2941 	TAILQ_REMOVE(&pv_chunks, pc, pc_lru);
2942 	PV_STAT(pv_entry_spare -= _NPCPV);
2943 	PV_STAT(pc_chunk_count--);
2944 	PV_STAT(pc_chunk_frees++);
2945 	/* entire chunk is free, return it */
2946 	m = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)pc));
2947 	pmap_qremove((vm_offset_t)pc, 1);
2948 	vm_page_unwire_noq(m);
2949 	vm_page_free(m);
2950 	pmap_pte2list_free(&pv_vafree, (vm_offset_t)pc);
2951 }
2952 
2953 /*
2954  *  Free the pv_entry back to the free list.
2955  */
2956 static void
free_pv_entry(pmap_t pmap,pv_entry_t pv)2957 free_pv_entry(pmap_t pmap, pv_entry_t pv)
2958 {
2959 	struct pv_chunk *pc;
2960 	int idx, field, bit;
2961 
2962 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2963 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2964 	PV_STAT(pv_entry_frees++);
2965 	PV_STAT(pv_entry_spare++);
2966 	pv_entry_count--;
2967 	pc = pv_to_chunk(pv);
2968 	idx = pv - &pc->pc_pventry[0];
2969 	field = idx / 32;
2970 	bit = idx % 32;
2971 	pc->pc_map[field] |= 1ul << bit;
2972 	for (idx = 0; idx < _NPCM; idx++)
2973 		if (pc->pc_map[idx] != pc_freemask[idx]) {
2974 			/*
2975 			 * 98% of the time, pc is already at the head of the
2976 			 * list.  If it isn't already, move it to the head.
2977 			 */
2978 			if (__predict_false(TAILQ_FIRST(&pmap->pm_pvchunk) !=
2979 			    pc)) {
2980 				TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
2981 				TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc,
2982 				    pc_list);
2983 			}
2984 			return;
2985 		}
2986 	TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
2987 	free_pv_chunk(pc);
2988 }
2989 
2990 /*
2991  *  Get a new pv_entry, allocating a block from the system
2992  *  when needed.
2993  */
2994 static pv_entry_t
get_pv_entry(pmap_t pmap,bool try)2995 get_pv_entry(pmap_t pmap, bool try)
2996 {
2997 	static const struct timeval printinterval = { 60, 0 };
2998 	static struct timeval lastprint;
2999 	int bit, field;
3000 	pv_entry_t pv;
3001 	struct pv_chunk *pc;
3002 	vm_page_t m;
3003 
3004 	rw_assert(&pvh_global_lock, RA_WLOCKED);
3005 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3006 	PV_STAT(pv_entry_allocs++);
3007 	pv_entry_count++;
3008 	if (pv_entry_count > pv_entry_high_water)
3009 		if (ratecheck(&lastprint, &printinterval))
3010 			printf("Approaching the limit on PV entries, consider "
3011 			    "increasing either the vm.pmap.shpgperproc or the "
3012 			    "vm.pmap.pv_entry_max tunable.\n");
3013 retry:
3014 	pc = TAILQ_FIRST(&pmap->pm_pvchunk);
3015 	if (pc != NULL) {
3016 		for (field = 0; field < _NPCM; field++) {
3017 			if (pc->pc_map[field]) {
3018 				bit = ffs(pc->pc_map[field]) - 1;
3019 				break;
3020 			}
3021 		}
3022 		if (field < _NPCM) {
3023 			pv = &pc->pc_pventry[field * 32 + bit];
3024 			pc->pc_map[field] &= ~(1ul << bit);
3025 			/* If this was the last item, move it to tail */
3026 			for (field = 0; field < _NPCM; field++)
3027 				if (pc->pc_map[field] != 0) {
3028 					PV_STAT(pv_entry_spare--);
3029 					return (pv);	/* not full, return */
3030 				}
3031 			TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
3032 			TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list);
3033 			PV_STAT(pv_entry_spare--);
3034 			return (pv);
3035 		}
3036 	}
3037 	/*
3038 	 * Access to the pte2list "pv_vafree" is synchronized by the pvh
3039 	 * global lock.  If "pv_vafree" is currently non-empty, it will
3040 	 * remain non-empty until pmap_pte2list_alloc() completes.
3041 	 */
3042 	if (pv_vafree == 0 ||
3043 	    (m = vm_page_alloc_noobj(VM_ALLOC_WIRED)) == NULL) {
3044 		if (try) {
3045 			pv_entry_count--;
3046 			PV_STAT(pc_chunk_tryfail++);
3047 			return (NULL);
3048 		}
3049 		m = pmap_pv_reclaim(pmap);
3050 		if (m == NULL)
3051 			goto retry;
3052 	}
3053 	PV_STAT(pc_chunk_count++);
3054 	PV_STAT(pc_chunk_allocs++);
3055 	pc = (struct pv_chunk *)pmap_pte2list_alloc(&pv_vafree);
3056 	pmap_qenter((vm_offset_t)pc, &m, 1);
3057 	pc->pc_pmap = pmap;
3058 	pc->pc_map[0] = pc_freemask[0] & ~1ul;	/* preallocated bit 0 */
3059 	for (field = 1; field < _NPCM; field++)
3060 		pc->pc_map[field] = pc_freemask[field];
3061 	TAILQ_INSERT_TAIL(&pv_chunks, pc, pc_lru);
3062 	pv = &pc->pc_pventry[0];
3063 	TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list);
3064 	PV_STAT(pv_entry_spare += _NPCPV - 1);
3065 	return (pv);
3066 }
3067 
3068 /*
3069  *  Create a pv entry for page at pa for
3070  *  (pmap, va).
3071  */
3072 static void
pmap_insert_entry(pmap_t pmap,vm_offset_t va,vm_page_t m)3073 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m)
3074 {
3075 	pv_entry_t pv;
3076 
3077 	rw_assert(&pvh_global_lock, RA_WLOCKED);
3078 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3079 	pv = get_pv_entry(pmap, false);
3080 	pv->pv_va = va;
3081 	TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
3082 }
3083 
3084 static __inline pv_entry_t
pmap_pvh_remove(struct md_page * pvh,pmap_t pmap,vm_offset_t va)3085 pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va)
3086 {
3087 	pv_entry_t pv;
3088 
3089 	rw_assert(&pvh_global_lock, RA_WLOCKED);
3090 	TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) {
3091 		if (pmap == PV_PMAP(pv) && va == pv->pv_va) {
3092 			TAILQ_REMOVE(&pvh->pv_list, pv, pv_next);
3093 			break;
3094 		}
3095 	}
3096 	return (pv);
3097 }
3098 
3099 static void
pmap_pvh_free(struct md_page * pvh,pmap_t pmap,vm_offset_t va)3100 pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va)
3101 {
3102 	pv_entry_t pv;
3103 
3104 	pv = pmap_pvh_remove(pvh, pmap, va);
3105 	KASSERT(pv != NULL, ("pmap_pvh_free: pv not found"));
3106 	free_pv_entry(pmap, pv);
3107 }
3108 
3109 static void
pmap_remove_entry(pmap_t pmap,vm_page_t m,vm_offset_t va)3110 pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va)
3111 {
3112 	struct md_page *pvh;
3113 
3114 	rw_assert(&pvh_global_lock, RA_WLOCKED);
3115 	pmap_pvh_free(&m->md, pmap, va);
3116 	if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) {
3117 		pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
3118 		if (TAILQ_EMPTY(&pvh->pv_list))
3119 			vm_page_aflag_clear(m, PGA_WRITEABLE);
3120 	}
3121 }
3122 
3123 static void
pmap_pv_demote_pte1(pmap_t pmap,vm_offset_t va,vm_paddr_t pa)3124 pmap_pv_demote_pte1(pmap_t pmap, vm_offset_t va, vm_paddr_t pa)
3125 {
3126 	struct md_page *pvh;
3127 	pv_entry_t pv;
3128 	vm_offset_t va_last;
3129 	vm_page_t m;
3130 
3131 	rw_assert(&pvh_global_lock, RA_WLOCKED);
3132 	KASSERT((pa & PTE1_OFFSET) == 0,
3133 	    ("pmap_pv_demote_pte1: pa is not 1mpage aligned"));
3134 
3135 	/*
3136 	 * Transfer the 1mpage's pv entry for this mapping to the first
3137 	 * page's pv list.
3138 	 */
3139 	pvh = pa_to_pvh(pa);
3140 	va = pte1_trunc(va);
3141 	pv = pmap_pvh_remove(pvh, pmap, va);
3142 	KASSERT(pv != NULL, ("pmap_pv_demote_pte1: pv not found"));
3143 	m = PHYS_TO_VM_PAGE(pa);
3144 	TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
3145 	/* Instantiate the remaining NPTE2_IN_PT2 - 1 pv entries. */
3146 	va_last = va + PTE1_SIZE - PAGE_SIZE;
3147 	do {
3148 		m++;
3149 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3150 		    ("pmap_pv_demote_pte1: page %p is not managed", m));
3151 		va += PAGE_SIZE;
3152 		pmap_insert_entry(pmap, va, m);
3153 	} while (va < va_last);
3154 }
3155 
3156 #if VM_NRESERVLEVEL > 0
3157 static void
pmap_pv_promote_pte1(pmap_t pmap,vm_offset_t va,vm_paddr_t pa)3158 pmap_pv_promote_pte1(pmap_t pmap, vm_offset_t va, vm_paddr_t pa)
3159 {
3160 	struct md_page *pvh;
3161 	pv_entry_t pv;
3162 	vm_offset_t va_last;
3163 	vm_page_t m;
3164 
3165 	rw_assert(&pvh_global_lock, RA_WLOCKED);
3166 	KASSERT((pa & PTE1_OFFSET) == 0,
3167 	    ("pmap_pv_promote_pte1: pa is not 1mpage aligned"));
3168 
3169 	/*
3170 	 * Transfer the first page's pv entry for this mapping to the
3171 	 * 1mpage's pv list.  Aside from avoiding the cost of a call
3172 	 * to get_pv_entry(), a transfer avoids the possibility that
3173 	 * get_pv_entry() calls pmap_pv_reclaim() and that pmap_pv_reclaim()
3174 	 * removes one of the mappings that is being promoted.
3175 	 */
3176 	m = PHYS_TO_VM_PAGE(pa);
3177 	va = pte1_trunc(va);
3178 	pv = pmap_pvh_remove(&m->md, pmap, va);
3179 	KASSERT(pv != NULL, ("pmap_pv_promote_pte1: pv not found"));
3180 	pvh = pa_to_pvh(pa);
3181 	TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next);
3182 	/* Free the remaining NPTE2_IN_PT2 - 1 pv entries. */
3183 	va_last = va + PTE1_SIZE - PAGE_SIZE;
3184 	do {
3185 		m++;
3186 		va += PAGE_SIZE;
3187 		pmap_pvh_free(&m->md, pmap, va);
3188 	} while (va < va_last);
3189 }
3190 #endif
3191 
3192 /*
3193  *  Conditionally create a pv entry.
3194  */
3195 static bool
pmap_try_insert_pv_entry(pmap_t pmap,vm_offset_t va,vm_page_t m)3196 pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m)
3197 {
3198 	pv_entry_t pv;
3199 
3200 	rw_assert(&pvh_global_lock, RA_WLOCKED);
3201 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3202 	if (pv_entry_count < pv_entry_high_water &&
3203 	    (pv = get_pv_entry(pmap, true)) != NULL) {
3204 		pv->pv_va = va;
3205 		TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
3206 		return (true);
3207 	} else
3208 		return (false);
3209 }
3210 
3211 /*
3212  *  Create the pv entries for each of the pages within a section.
3213  */
3214 static bool
pmap_pv_insert_pte1(pmap_t pmap,vm_offset_t va,pt1_entry_t pte1,u_int flags)3215 pmap_pv_insert_pte1(pmap_t pmap, vm_offset_t va, pt1_entry_t pte1, u_int flags)
3216 {
3217 	struct md_page *pvh;
3218 	pv_entry_t pv;
3219 	bool noreclaim;
3220 
3221 	rw_assert(&pvh_global_lock, RA_WLOCKED);
3222 	noreclaim = (flags & PMAP_ENTER_NORECLAIM) != 0;
3223 	if ((noreclaim && pv_entry_count >= pv_entry_high_water) ||
3224 	    (pv = get_pv_entry(pmap, noreclaim)) == NULL)
3225 		return (false);
3226 	pv->pv_va = va;
3227 	pvh = pa_to_pvh(pte1_pa(pte1));
3228 	TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next);
3229 	return (true);
3230 }
3231 
3232 static inline void
pmap_tlb_flush_pte1(pmap_t pmap,vm_offset_t va,pt1_entry_t npte1)3233 pmap_tlb_flush_pte1(pmap_t pmap, vm_offset_t va, pt1_entry_t npte1)
3234 {
3235 
3236 	/* Kill all the small mappings or the big one only. */
3237 	if (pte1_is_section(npte1))
3238 		pmap_tlb_flush_range(pmap, pte1_trunc(va), PTE1_SIZE);
3239 	else
3240 		pmap_tlb_flush(pmap, pte1_trunc(va));
3241 }
3242 
3243 /*
3244  *  Update kernel pte1 on all pmaps.
3245  *
3246  *  The following function is called only on one cpu with disabled interrupts.
3247  *  In SMP case, smp_rendezvous_cpus() is used to stop other cpus. This way
3248  *  nobody can invoke explicit hardware table walk during the update of pte1.
3249  *  Unsolicited hardware table walk can still happen, invoked by speculative
3250  *  data or instruction prefetch or even by speculative hardware table walk.
3251  *
3252  *  The break-before-make approach should be implemented here. However, it's
3253  *  not so easy to do that for kernel mappings as it would be unhappy to unmap
3254  *  itself unexpectedly but voluntarily.
3255  */
3256 static void
pmap_update_pte1_kernel(vm_offset_t va,pt1_entry_t npte1)3257 pmap_update_pte1_kernel(vm_offset_t va, pt1_entry_t npte1)
3258 {
3259 	pmap_t pmap;
3260 	pt1_entry_t *pte1p;
3261 
3262 	/*
3263 	 * Get current pmap. Interrupts should be disabled here
3264 	 * so PCPU_GET() is done atomically.
3265 	 */
3266 	pmap = PCPU_GET(curpmap);
3267 	if (pmap == NULL)
3268 		pmap = kernel_pmap;
3269 
3270 	/*
3271 	 * (1) Change pte1 on current pmap.
3272 	 * (2) Flush all obsolete TLB entries on current CPU.
3273 	 * (3) Change pte1 on all pmaps.
3274 	 * (4) Flush all obsolete TLB entries on all CPUs in SMP case.
3275 	 */
3276 
3277 	pte1p = pmap_pte1(pmap, va);
3278 	pte1_store(pte1p, npte1);
3279 
3280 	/* Kill all the small mappings or the big one only. */
3281 	if (pte1_is_section(npte1)) {
3282 		pmap_pte1_kern_promotions++;
3283 		tlb_flush_range_local(pte1_trunc(va), PTE1_SIZE);
3284 	} else {
3285 		pmap_pte1_kern_demotions++;
3286 		tlb_flush_local(pte1_trunc(va));
3287 	}
3288 
3289 	/*
3290 	 * In SMP case, this function is called when all cpus are at smp
3291 	 * rendezvous, so there is no need to use 'allpmaps_lock' lock here.
3292 	 * In UP case, the function is called with this lock locked.
3293 	 */
3294 	LIST_FOREACH(pmap, &allpmaps, pm_list) {
3295 		pte1p = pmap_pte1(pmap, va);
3296 		pte1_store(pte1p, npte1);
3297 	}
3298 
3299 #ifdef SMP
3300 	/* Kill all the small mappings or the big one only. */
3301 	if (pte1_is_section(npte1))
3302 		tlb_flush_range(pte1_trunc(va), PTE1_SIZE);
3303 	else
3304 		tlb_flush(pte1_trunc(va));
3305 #endif
3306 }
3307 
3308 #ifdef SMP
3309 struct pte1_action {
3310 	vm_offset_t va;
3311 	pt1_entry_t npte1;
3312 	u_int update;		/* CPU that updates the PTE1 */
3313 };
3314 
3315 static void
pmap_update_pte1_action(void * arg)3316 pmap_update_pte1_action(void *arg)
3317 {
3318 	struct pte1_action *act = arg;
3319 
3320 	if (act->update == PCPU_GET(cpuid))
3321 		pmap_update_pte1_kernel(act->va, act->npte1);
3322 }
3323 
3324 /*
3325  *  Change pte1 on current pmap.
3326  *  Note that kernel pte1 must be changed on all pmaps.
3327  *
3328  *  According to the architecture reference manual published by ARM,
3329  *  the behaviour is UNPREDICTABLE when two or more TLB entries map the same VA.
3330  *  According to this manual, UNPREDICTABLE behaviours must never happen in
3331  *  a viable system. In contrast, on x86 processors, it is not specified which
3332  *  TLB entry mapping the virtual address will be used, but the MMU doesn't
3333  *  generate a bogus translation the way it does on Cortex-A8 rev 2 (Beaglebone
3334  *  Black).
3335  *
3336  *  It's a problem when either promotion or demotion is being done. The pte1
3337  *  update and appropriate TLB flush must be done atomically in general.
3338  */
3339 static void
pmap_change_pte1(pmap_t pmap,pt1_entry_t * pte1p,vm_offset_t va,pt1_entry_t npte1)3340 pmap_change_pte1(pmap_t pmap, pt1_entry_t *pte1p, vm_offset_t va,
3341     pt1_entry_t npte1)
3342 {
3343 
3344 	if (pmap == kernel_pmap) {
3345 		struct pte1_action act;
3346 
3347 		sched_pin();
3348 		act.va = va;
3349 		act.npte1 = npte1;
3350 		act.update = PCPU_GET(cpuid);
3351 		smp_rendezvous_cpus(all_cpus, smp_no_rendezvous_barrier,
3352 		    pmap_update_pte1_action, NULL, &act);
3353 		sched_unpin();
3354 	} else {
3355 		register_t cspr;
3356 
3357 		/*
3358 		 * Use break-before-make approach for changing userland
3359 		 * mappings. It can cause L1 translation aborts on other
3360 		 * cores in SMP case. So, special treatment is implemented
3361 		 * in pmap_fault(). To reduce the likelihood that another core
3362 		 * will be affected by the broken mapping, disable interrupts
3363 		 * until the mapping change is completed.
3364 		 */
3365 		cspr = disable_interrupts(PSR_I);
3366 		pte1_clear(pte1p);
3367 		pmap_tlb_flush_pte1(pmap, va, npte1);
3368 		pte1_store(pte1p, npte1);
3369 		restore_interrupts(cspr);
3370 	}
3371 }
3372 #else
3373 static void
pmap_change_pte1(pmap_t pmap,pt1_entry_t * pte1p,vm_offset_t va,pt1_entry_t npte1)3374 pmap_change_pte1(pmap_t pmap, pt1_entry_t *pte1p, vm_offset_t va,
3375     pt1_entry_t npte1)
3376 {
3377 
3378 	if (pmap == kernel_pmap) {
3379 		mtx_lock_spin(&allpmaps_lock);
3380 		pmap_update_pte1_kernel(va, npte1);
3381 		mtx_unlock_spin(&allpmaps_lock);
3382 	} else {
3383 		register_t cspr;
3384 
3385 		/*
3386 		 * Use break-before-make approach for changing userland
3387 		 * mappings. It's absolutely safe in UP case when interrupts
3388 		 * are disabled.
3389 		 */
3390 		cspr = disable_interrupts(PSR_I);
3391 		pte1_clear(pte1p);
3392 		pmap_tlb_flush_pte1(pmap, va, npte1);
3393 		pte1_store(pte1p, npte1);
3394 		restore_interrupts(cspr);
3395 	}
3396 }
3397 #endif
3398 
3399 #if VM_NRESERVLEVEL > 0
3400 /*
3401  *  Tries to promote the NPTE2_IN_PT2, contiguous 4KB page mappings that are
3402  *  within a single page table page (PT2) to a single 1MB page mapping.
3403  *  For promotion to occur, two conditions must be met: (1) the 4KB page
3404  *  mappings must map aligned, contiguous physical memory and (2) the 4KB page
3405  *  mappings must have identical characteristics.
3406  *
3407  *  Managed (PG_MANAGED) mappings within the kernel address space are not
3408  *  promoted.  The reason is that kernel PTE1s are replicated in each pmap but
3409  *  pmap_remove_write(), pmap_clear_modify(), and pmap_clear_reference() only
3410  *  read the PTE1 from the kernel pmap.
3411  */
3412 static void
pmap_promote_pte1(pmap_t pmap,pt1_entry_t * pte1p,vm_offset_t va)3413 pmap_promote_pte1(pmap_t pmap, pt1_entry_t *pte1p, vm_offset_t va)
3414 {
3415 	pt1_entry_t npte1;
3416 	pt2_entry_t *fpte2p, fpte2, fpte2_fav;
3417 	pt2_entry_t *pte2p, pte2;
3418 	vm_offset_t pteva __unused;
3419 	vm_page_t m __unused;
3420 
3421 	PDEBUG(6, printf("%s(%p): try for va %#x pte1 %#x at %p\n", __func__,
3422 	    pmap, va, pte1_load(pte1p), pte1p));
3423 
3424 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3425 
3426 	/*
3427 	 * Examine the first PTE2 in the specified PT2. Abort if this PTE2 is
3428 	 * either invalid, unused, or does not map the first 4KB physical page
3429 	 * within a 1MB page.
3430 	 */
3431 	fpte2p = pmap_pte2_quick(pmap, pte1_trunc(va));
3432 	fpte2 = pte2_load(fpte2p);
3433 	if ((fpte2 & ((PTE2_FRAME & PTE1_OFFSET) | PTE2_A | PTE2_V)) !=
3434 	    (PTE2_A | PTE2_V)) {
3435 		pmap_pte1_p_failures++;
3436 		CTR3(KTR_PMAP, "%s: failure(1) for va %#x in pmap %p",
3437 		    __func__, va, pmap);
3438 		return;
3439 	}
3440 	if (pte2_is_managed(fpte2) && pmap == kernel_pmap) {
3441 		pmap_pte1_p_failures++;
3442 		CTR3(KTR_PMAP, "%s: failure(2) for va %#x in pmap %p",
3443 		    __func__, va, pmap);
3444 		return;
3445 	}
3446 	if ((fpte2 & (PTE2_NM | PTE2_RO)) == PTE2_NM) {
3447 		/*
3448 		 * When page is not modified, PTE2_RO can be set without
3449 		 * a TLB invalidation.
3450 		 */
3451 		fpte2 |= PTE2_RO;
3452 		pte2_store(fpte2p, fpte2);
3453 	}
3454 
3455 	/*
3456 	 * Examine each of the other PTE2s in the specified PT2. Abort if this
3457 	 * PTE2 maps an unexpected 4KB physical page or does not have identical
3458 	 * characteristics to the first PTE2.
3459 	 */
3460 	fpte2_fav = (fpte2 & (PTE2_FRAME | PTE2_A | PTE2_V));
3461 	fpte2_fav += PTE1_SIZE - PTE2_SIZE; /* examine from the end */
3462 	for (pte2p = fpte2p + NPTE2_IN_PT2 - 1; pte2p > fpte2p; pte2p--) {
3463 		pte2 = pte2_load(pte2p);
3464 		if ((pte2 & (PTE2_FRAME | PTE2_A | PTE2_V)) != fpte2_fav) {
3465 			pmap_pte1_p_failures++;
3466 			CTR3(KTR_PMAP, "%s: failure(3) for va %#x in pmap %p",
3467 			    __func__, va, pmap);
3468 			return;
3469 		}
3470 		if ((pte2 & (PTE2_NM | PTE2_RO)) == PTE2_NM) {
3471 			/*
3472 			 * When page is not modified, PTE2_RO can be set
3473 			 * without a TLB invalidation. See note above.
3474 			 */
3475 			pte2 |= PTE2_RO;
3476 			pte2_store(pte2p, pte2);
3477 			pteva = pte1_trunc(va) | (pte2 & PTE1_OFFSET &
3478 			    PTE2_FRAME);
3479 			CTR3(KTR_PMAP, "%s: protect for va %#x in pmap %p",
3480 			    __func__, pteva, pmap);
3481 		}
3482 		if ((pte2 & PTE2_PROMOTE) != (fpte2 & PTE2_PROMOTE)) {
3483 			pmap_pte1_p_failures++;
3484 			CTR3(KTR_PMAP, "%s: failure(4) for va %#x in pmap %p",
3485 			    __func__, va, pmap);
3486 			return;
3487 		}
3488 
3489 		fpte2_fav -= PTE2_SIZE;
3490 	}
3491 	/*
3492 	 * The page table page in its current state will stay in PT2TAB
3493 	 * until the PTE1 mapping the section is demoted by pmap_demote_pte1()
3494 	 * or destroyed by pmap_remove_pte1().
3495 	 *
3496 	 * Note that L2 page table size is not equal to PAGE_SIZE.
3497 	 */
3498 	m = PHYS_TO_VM_PAGE(trunc_page(pte1_link_pa(pte1_load(pte1p))));
3499 	KASSERT(m >= vm_page_array && m < &vm_page_array[vm_page_array_size],
3500 	    ("%s: PT2 page is out of range", __func__));
3501 	KASSERT(m->pindex == (pte1_index(va) & ~PT2PG_MASK),
3502 	    ("%s: PT2 page's pindex is wrong", __func__));
3503 
3504 	/*
3505 	 * Get pte1 from pte2 format.
3506 	 */
3507 	npte1 = (fpte2 & PTE1_FRAME) | ATTR_TO_L1(fpte2) | PTE1_V;
3508 
3509 	/*
3510 	 * Promote the pv entries.
3511 	 */
3512 	if (pte2_is_managed(fpte2))
3513 		pmap_pv_promote_pte1(pmap, va, pte1_pa(npte1));
3514 
3515 	/*
3516 	 * Promote the mappings.
3517 	 */
3518 	pmap_change_pte1(pmap, pte1p, va, npte1);
3519 
3520 	pmap_pte1_promotions++;
3521 	CTR3(KTR_PMAP, "%s: success for va %#x in pmap %p",
3522 	    __func__, va, pmap);
3523 
3524 	PDEBUG(6, printf("%s(%p): success for va %#x pte1 %#x(%#x) at %p\n",
3525 	    __func__, pmap, va, npte1, pte1_load(pte1p), pte1p));
3526 }
3527 #endif /* VM_NRESERVLEVEL > 0 */
3528 
3529 /*
3530  *  Zero L2 page table page.
3531  */
3532 static __inline void
pmap_clear_pt2(pt2_entry_t * fpte2p)3533 pmap_clear_pt2(pt2_entry_t *fpte2p)
3534 {
3535 	pt2_entry_t *pte2p;
3536 
3537 	for (pte2p = fpte2p; pte2p < fpte2p + NPTE2_IN_PT2; pte2p++)
3538 		pte2_clear(pte2p);
3539 
3540 }
3541 
3542 /*
3543  *  Removes a 1MB page mapping from the kernel pmap.
3544  */
3545 static void
pmap_remove_kernel_pte1(pmap_t pmap,pt1_entry_t * pte1p,vm_offset_t va)3546 pmap_remove_kernel_pte1(pmap_t pmap, pt1_entry_t *pte1p, vm_offset_t va)
3547 {
3548 	vm_page_t m;
3549 	uint32_t pte1_idx;
3550 	pt2_entry_t *fpte2p;
3551 	vm_paddr_t pt2_pa;
3552 
3553 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3554 	m = pmap_pt2_page(pmap, va);
3555 	if (m == NULL)
3556 		/*
3557 		 * QQQ: Is this function called only on promoted pte1?
3558 		 *      We certainly do section mappings directly
3559 		 *      (without promotion) in kernel !!!
3560 		 */
3561 		panic("%s: missing pt2 page", __func__);
3562 
3563 	pte1_idx = pte1_index(va);
3564 
3565 	/*
3566 	 * Initialize the L2 page table.
3567 	 */
3568 	fpte2p = page_pt2(pt2map_pt2pg(va), pte1_idx);
3569 	pmap_clear_pt2(fpte2p);
3570 
3571 	/*
3572 	 * Remove the mapping.
3573 	 */
3574 	pt2_pa = page_pt2pa(VM_PAGE_TO_PHYS(m), pte1_idx);
3575 	pmap_kenter_pte1(va, PTE1_LINK(pt2_pa));
3576 
3577 	/*
3578 	 * QQQ: We do not need to invalidate PT2MAP mapping
3579 	 * as we did not change it. I.e. the L2 page table page
3580 	 * was and still is mapped the same way.
3581 	 */
3582 }
3583 
3584 /*
3585  *  Do the things to unmap a section in a process
3586  */
3587 static void
pmap_remove_pte1(pmap_t pmap,pt1_entry_t * pte1p,vm_offset_t sva,struct spglist * free)3588 pmap_remove_pte1(pmap_t pmap, pt1_entry_t *pte1p, vm_offset_t sva,
3589     struct spglist *free)
3590 {
3591 	pt1_entry_t opte1;
3592 	struct md_page *pvh;
3593 	vm_offset_t eva, va;
3594 	vm_page_t m;
3595 
3596 	PDEBUG(6, printf("%s(%p): va %#x pte1 %#x at %p\n", __func__, pmap, sva,
3597 	    pte1_load(pte1p), pte1p));
3598 
3599 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3600 	KASSERT((sva & PTE1_OFFSET) == 0,
3601 	    ("%s: sva is not 1mpage aligned", __func__));
3602 
3603 	/*
3604 	 * Clear and invalidate the mapping. It should occupy one and only TLB
3605 	 * entry. So, pmap_tlb_flush() called with aligned address should be
3606 	 * sufficient.
3607 	 */
3608 	opte1 = pte1_load_clear(pte1p);
3609 	pmap_tlb_flush(pmap, sva);
3610 
3611 	if (pte1_is_wired(opte1))
3612 		pmap->pm_stats.wired_count -= PTE1_SIZE / PAGE_SIZE;
3613 	pmap->pm_stats.resident_count -= PTE1_SIZE / PAGE_SIZE;
3614 	if (pte1_is_managed(opte1)) {
3615 		pvh = pa_to_pvh(pte1_pa(opte1));
3616 		pmap_pvh_free(pvh, pmap, sva);
3617 		eva = sva + PTE1_SIZE;
3618 		for (va = sva, m = PHYS_TO_VM_PAGE(pte1_pa(opte1));
3619 		    va < eva; va += PAGE_SIZE, m++) {
3620 			if (pte1_is_dirty(opte1))
3621 				vm_page_dirty(m);
3622 			if (opte1 & PTE1_A)
3623 				vm_page_aflag_set(m, PGA_REFERENCED);
3624 			if (TAILQ_EMPTY(&m->md.pv_list) &&
3625 			    TAILQ_EMPTY(&pvh->pv_list))
3626 				vm_page_aflag_clear(m, PGA_WRITEABLE);
3627 		}
3628 	}
3629 	if (pmap == kernel_pmap) {
3630 		/*
3631 		 * L2 page table(s) can't be removed from kernel map as
3632 		 * kernel counts on it (stuff around pmap_growkernel()).
3633 		 */
3634 		 pmap_remove_kernel_pte1(pmap, pte1p, sva);
3635 	} else {
3636 		/*
3637 		 * Get associated L2 page table page.
3638 		 * It's possible that the page was never allocated.
3639 		 */
3640 		m = pmap_pt2_page(pmap, sva);
3641 		if (m != NULL)
3642 			pmap_unwire_pt2_all(pmap, sva, m, free);
3643 	}
3644 }
3645 
3646 /*
3647  *  Fills L2 page table page with mappings to consecutive physical pages.
3648  */
3649 static __inline void
pmap_fill_pt2(pt2_entry_t * fpte2p,pt2_entry_t npte2)3650 pmap_fill_pt2(pt2_entry_t *fpte2p, pt2_entry_t npte2)
3651 {
3652 	pt2_entry_t *pte2p;
3653 
3654 	for (pte2p = fpte2p; pte2p < fpte2p + NPTE2_IN_PT2; pte2p++) {
3655 		pte2_store(pte2p, npte2);
3656 		npte2 += PTE2_SIZE;
3657 	}
3658 }
3659 
3660 /*
3661  *  Tries to demote a 1MB page mapping. If demotion fails, the
3662  *  1MB page mapping is invalidated.
3663  */
3664 static bool
pmap_demote_pte1(pmap_t pmap,pt1_entry_t * pte1p,vm_offset_t va)3665 pmap_demote_pte1(pmap_t pmap, pt1_entry_t *pte1p, vm_offset_t va)
3666 {
3667 	pt1_entry_t opte1, npte1;
3668 	pt2_entry_t *fpte2p, npte2;
3669 	vm_paddr_t pt2pg_pa, pt2_pa;
3670 	vm_page_t m;
3671 	struct spglist free;
3672 	uint32_t pte1_idx, isnew = 0;
3673 
3674 	PDEBUG(6, printf("%s(%p): try for va %#x pte1 %#x at %p\n", __func__,
3675 	    pmap, va, pte1_load(pte1p), pte1p));
3676 
3677 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3678 
3679 	opte1 = pte1_load(pte1p);
3680 	KASSERT(pte1_is_section(opte1), ("%s: opte1 not a section", __func__));
3681 
3682 	if ((opte1 & PTE1_A) == 0 || (m = pmap_pt2_page(pmap, va)) == NULL) {
3683 		KASSERT(!pte1_is_wired(opte1),
3684 		    ("%s: PT2 page for a wired mapping is missing", __func__));
3685 
3686 		/*
3687 		 * Invalidate the 1MB page mapping and return
3688 		 * "failure" if the mapping was never accessed or the
3689 		 * allocation of the new page table page fails.
3690 		 */
3691 		if ((opte1 & PTE1_A) == 0 ||
3692 		    (m = vm_page_alloc_noobj(VM_ALLOC_WIRED)) == NULL) {
3693 			SLIST_INIT(&free);
3694 			pmap_remove_pte1(pmap, pte1p, pte1_trunc(va), &free);
3695 			vm_page_free_pages_toq(&free, false);
3696 			CTR3(KTR_PMAP, "%s: failure for va %#x in pmap %p",
3697 			    __func__, va, pmap);
3698 			return (false);
3699 		}
3700 		m->pindex = pte1_index(va) & ~PT2PG_MASK;
3701 		if (va < VM_MAXUSER_ADDRESS)
3702 			pmap->pm_stats.resident_count++;
3703 
3704 		isnew = 1;
3705 
3706 		/*
3707 		 * We init all L2 page tables in the page even if
3708 		 * we are going to change everything for one L2 page
3709 		 * table in a while.
3710 		 */
3711 		pt2pg_pa = pmap_pt2pg_init(pmap, va, m);
3712 	} else {
3713 		if (va < VM_MAXUSER_ADDRESS) {
3714 			if (pt2_is_empty(m, va))
3715 				isnew = 1; /* Demoting section w/o promotion. */
3716 #ifdef INVARIANTS
3717 			else
3718 				KASSERT(pt2_is_full(m, va), ("%s: bad PT2 wire"
3719 				    " count %u", __func__,
3720 				    pt2_wirecount_get(m, pte1_index(va))));
3721 #endif
3722 		}
3723 	}
3724 
3725 	pt2pg_pa = VM_PAGE_TO_PHYS(m);
3726 	pte1_idx = pte1_index(va);
3727 	/*
3728 	 * If the pmap is current, then the PT2MAP can provide access to
3729 	 * the page table page (promoted L2 page tables are not unmapped).
3730 	 * Otherwise, temporarily map the L2 page table page (m) into
3731 	 * the kernel's address space at either PADDR1 or PADDR2.
3732 	 *
3733 	 * Note that L2 page table size is not equal to PAGE_SIZE.
3734 	 */
3735 	if (pmap_is_current(pmap))
3736 		fpte2p = page_pt2(pt2map_pt2pg(va), pte1_idx);
3737 	else if (curthread->td_pinned > 0 && rw_wowned(&pvh_global_lock)) {
3738 		if (pte2_pa(pte2_load(PMAP1)) != pt2pg_pa) {
3739 			pte2_store(PMAP1, PTE2_KPT(pt2pg_pa));
3740 #ifdef SMP
3741 			PMAP1cpu = PCPU_GET(cpuid);
3742 #endif
3743 			tlb_flush_local((vm_offset_t)PADDR1);
3744 			PMAP1changed++;
3745 		} else
3746 #ifdef SMP
3747 		if (PMAP1cpu != PCPU_GET(cpuid)) {
3748 			PMAP1cpu = PCPU_GET(cpuid);
3749 			tlb_flush_local((vm_offset_t)PADDR1);
3750 			PMAP1changedcpu++;
3751 		} else
3752 #endif
3753 			PMAP1unchanged++;
3754 		fpte2p = page_pt2((vm_offset_t)PADDR1, pte1_idx);
3755 	} else {
3756 		mtx_lock(&PMAP2mutex);
3757 		if (pte2_pa(pte2_load(PMAP2)) != pt2pg_pa) {
3758 			pte2_store(PMAP2, PTE2_KPT(pt2pg_pa));
3759 			tlb_flush((vm_offset_t)PADDR2);
3760 		}
3761 		fpte2p = page_pt2((vm_offset_t)PADDR2, pte1_idx);
3762 	}
3763 	pt2_pa = page_pt2pa(pt2pg_pa, pte1_idx);
3764 	npte1 = PTE1_LINK(pt2_pa);
3765 
3766 	KASSERT((opte1 & PTE1_A) != 0,
3767 	    ("%s: opte1 is missing PTE1_A", __func__));
3768 	KASSERT((opte1 & (PTE1_NM | PTE1_RO)) != PTE1_NM,
3769 	    ("%s: opte1 has PTE1_NM", __func__));
3770 
3771 	/*
3772 	 *  Get pte2 from pte1 format.
3773 	*/
3774 	npte2 = pte1_pa(opte1) | ATTR_TO_L2(opte1) | PTE2_V;
3775 
3776 	/*
3777 	 * If the L2 page table page is new, initialize it. If the mapping
3778 	 * has changed attributes, update the page table entries.
3779 	 */
3780 	if (isnew != 0) {
3781 		pt2_wirecount_set(m, pte1_idx, NPTE2_IN_PT2);
3782 		pmap_fill_pt2(fpte2p, npte2);
3783 	} else if ((pte2_load(fpte2p) & PTE2_PROMOTE) !=
3784 		    (npte2 & PTE2_PROMOTE))
3785 		pmap_fill_pt2(fpte2p, npte2);
3786 
3787 	KASSERT(pte2_pa(pte2_load(fpte2p)) == pte2_pa(npte2),
3788 	    ("%s: fpte2p and npte2 map different physical addresses",
3789 	    __func__));
3790 
3791 	if (fpte2p == PADDR2)
3792 		mtx_unlock(&PMAP2mutex);
3793 
3794 	/*
3795 	 * Demote the mapping. This pmap is locked. The old PTE1 has
3796 	 * PTE1_A set. If the old PTE1 has not PTE1_RO set, it also
3797 	 * has not PTE1_NM set. Thus, there is no danger of a race with
3798 	 * another processor changing the setting of PTE1_A and/or PTE1_NM
3799 	 * between the read above and the store below.
3800 	 */
3801 	pmap_change_pte1(pmap, pte1p, va, npte1);
3802 
3803 	/*
3804 	 * Demote the pv entry. This depends on the earlier demotion
3805 	 * of the mapping. Specifically, the (re)creation of a per-
3806 	 * page pv entry might trigger the execution of pmap_pv_reclaim(),
3807 	 * which might reclaim a newly (re)created per-page pv entry
3808 	 * and destroy the associated mapping. In order to destroy
3809 	 * the mapping, the PTE1 must have already changed from mapping
3810 	 * the 1mpage to referencing the page table page.
3811 	 */
3812 	if (pte1_is_managed(opte1))
3813 		pmap_pv_demote_pte1(pmap, va, pte1_pa(opte1));
3814 
3815 	pmap_pte1_demotions++;
3816 	CTR3(KTR_PMAP, "%s: success for va %#x in pmap %p",
3817 	    __func__, va, pmap);
3818 
3819 	PDEBUG(6, printf("%s(%p): success for va %#x pte1 %#x(%#x) at %p\n",
3820 	    __func__, pmap, va, npte1, pte1_load(pte1p), pte1p));
3821 	return (true);
3822 }
3823 
3824 /*
3825  *	Insert the given physical page (p) at
3826  *	the specified virtual address (v) in the
3827  *	target physical map with the protection requested.
3828  *
3829  *	If specified, the page will be wired down, meaning
3830  *	that the related pte can not be reclaimed.
3831  *
3832  *	NB:  This is the only routine which MAY NOT lazy-evaluate
3833  *	or lose information.  That is, this routine must actually
3834  *	insert this page into the given map NOW.
3835  */
3836 int
pmap_enter(pmap_t pmap,vm_offset_t va,vm_page_t m,vm_prot_t prot,u_int flags,int8_t psind)3837 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
3838     u_int flags, int8_t psind)
3839 {
3840 	pt1_entry_t *pte1p;
3841 	pt2_entry_t *pte2p;
3842 	pt2_entry_t npte2, opte2;
3843 	pv_entry_t pv;
3844 	vm_paddr_t opa, pa;
3845 	vm_page_t mpte2, om;
3846 	int rv;
3847 
3848 	va = trunc_page(va);
3849 	KASSERT(va <= vm_max_kernel_address, ("%s: toobig", __func__));
3850 	KASSERT(va < UPT2V_MIN_ADDRESS || va >= UPT2V_MAX_ADDRESS,
3851 	    ("%s: invalid to pmap_enter page table pages (va: 0x%x)", __func__,
3852 	    va));
3853 	KASSERT((m->oflags & VPO_UNMANAGED) != 0 || !VA_IS_CLEANMAP(va),
3854 	    ("%s: managed mapping within the clean submap", __func__));
3855 	if ((m->oflags & VPO_UNMANAGED) == 0)
3856 		VM_PAGE_OBJECT_BUSY_ASSERT(m);
3857 	KASSERT((flags & PMAP_ENTER_RESERVED) == 0,
3858 	    ("%s: flags %u has reserved bits set", __func__, flags));
3859 	pa = VM_PAGE_TO_PHYS(m);
3860 	npte2 = PTE2(pa, PTE2_A, vm_page_pte2_attr(m));
3861 	if ((flags & VM_PROT_WRITE) == 0)
3862 		npte2 |= PTE2_NM;
3863 	if ((prot & VM_PROT_WRITE) == 0)
3864 		npte2 |= PTE2_RO;
3865 	KASSERT((npte2 & (PTE2_NM | PTE2_RO)) != PTE2_RO,
3866 	    ("%s: flags includes VM_PROT_WRITE but prot doesn't", __func__));
3867 	if ((prot & VM_PROT_EXECUTE) == 0)
3868 		npte2 |= PTE2_NX;
3869 	if ((flags & PMAP_ENTER_WIRED) != 0)
3870 		npte2 |= PTE2_W;
3871 	if (va < VM_MAXUSER_ADDRESS)
3872 		npte2 |= PTE2_U;
3873 	if (pmap != kernel_pmap)
3874 		npte2 |= PTE2_NG;
3875 
3876 	rw_wlock(&pvh_global_lock);
3877 	PMAP_LOCK(pmap);
3878 	sched_pin();
3879 	if (psind == 1) {
3880 		/* Assert the required virtual and physical alignment. */
3881 		KASSERT((va & PTE1_OFFSET) == 0,
3882 		    ("%s: va unaligned", __func__));
3883 		KASSERT(m->psind > 0, ("%s: m->psind < psind", __func__));
3884 		rv = pmap_enter_pte1(pmap, va, PTE1_PA(pa) | ATTR_TO_L1(npte2) |
3885 		    PTE1_V, flags, m);
3886 		goto out;
3887 	}
3888 
3889 	/*
3890 	 * In the case that a page table page is not
3891 	 * resident, we are creating it here.
3892 	 */
3893 	if (va < VM_MAXUSER_ADDRESS) {
3894 		mpte2 = pmap_allocpte2(pmap, va, flags);
3895 		if (mpte2 == NULL) {
3896 			KASSERT((flags & PMAP_ENTER_NOSLEEP) != 0,
3897 			    ("pmap_allocpte2 failed with sleep allowed"));
3898 			rv = KERN_RESOURCE_SHORTAGE;
3899 			goto out;
3900 		}
3901 	} else
3902 		mpte2 = NULL;
3903 	pte1p = pmap_pte1(pmap, va);
3904 	if (pte1_is_section(pte1_load(pte1p)))
3905 		panic("%s: attempted on 1MB page", __func__);
3906 	pte2p = pmap_pte2_quick(pmap, va);
3907 	if (pte2p == NULL)
3908 		panic("%s: invalid L1 page table entry va=%#x", __func__, va);
3909 
3910 	om = NULL;
3911 	opte2 = pte2_load(pte2p);
3912 	opa = pte2_pa(opte2);
3913 	/*
3914 	 * Mapping has not changed, must be protection or wiring change.
3915 	 */
3916 	if (pte2_is_valid(opte2) && (opa == pa)) {
3917 		/*
3918 		 * Wiring change, just update stats. We don't worry about
3919 		 * wiring PT2 pages as they remain resident as long as there
3920 		 * are valid mappings in them. Hence, if a user page is wired,
3921 		 * the PT2 page will be also.
3922 		 */
3923 		if (pte2_is_wired(npte2) && !pte2_is_wired(opte2))
3924 			pmap->pm_stats.wired_count++;
3925 		else if (!pte2_is_wired(npte2) && pte2_is_wired(opte2))
3926 			pmap->pm_stats.wired_count--;
3927 
3928 		/*
3929 		 * Remove extra pte2 reference
3930 		 */
3931 		if (mpte2)
3932 			pt2_wirecount_dec(mpte2, pte1_index(va));
3933 		if ((m->oflags & VPO_UNMANAGED) == 0)
3934 			om = m;
3935 		goto validate;
3936 	}
3937 
3938 	/*
3939 	 * QQQ: We think that changing physical address on writeable mapping
3940 	 *      is not safe. Well, maybe on kernel address space with correct
3941 	 *      locking, it can make a sense. However, we have no idea why
3942 	 *      anyone should do that on user address space. Are we wrong?
3943 	 */
3944 	KASSERT((opa == 0) || (opa == pa) ||
3945 	    !pte2_is_valid(opte2) || ((opte2 & PTE2_RO) != 0),
3946 	    ("%s: pmap %p va %#x(%#x) opa %#x pa %#x - gotcha %#x %#x!",
3947 	    __func__, pmap, va, opte2, opa, pa, flags, prot));
3948 
3949 	pv = NULL;
3950 
3951 	/*
3952 	 * Mapping has changed, invalidate old range and fall through to
3953 	 * handle validating new mapping.
3954 	 */
3955 	if (opa) {
3956 		if (pte2_is_wired(opte2))
3957 			pmap->pm_stats.wired_count--;
3958 		om = PHYS_TO_VM_PAGE(opa);
3959 		if (om != NULL && (om->oflags & VPO_UNMANAGED) != 0)
3960 			om = NULL;
3961 		if (om != NULL)
3962 			pv = pmap_pvh_remove(&om->md, pmap, va);
3963 
3964 		/*
3965 		 * Remove extra pte2 reference
3966 		 */
3967 		if (mpte2 != NULL)
3968 			pt2_wirecount_dec(mpte2, va >> PTE1_SHIFT);
3969 	} else
3970 		pmap->pm_stats.resident_count++;
3971 
3972 	/*
3973 	 * Enter on the PV list if part of our managed memory.
3974 	 */
3975 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3976 		if (pv == NULL) {
3977 			pv = get_pv_entry(pmap, false);
3978 			pv->pv_va = va;
3979 		}
3980 		TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
3981 	} else if (pv != NULL)
3982 		free_pv_entry(pmap, pv);
3983 
3984 	/*
3985 	 * Increment counters
3986 	 */
3987 	if (pte2_is_wired(npte2))
3988 		pmap->pm_stats.wired_count++;
3989 
3990 validate:
3991 	/*
3992 	 * Now validate mapping with desired protection/wiring.
3993 	 */
3994 	if (prot & VM_PROT_WRITE) {
3995 		if ((m->oflags & VPO_UNMANAGED) == 0)
3996 			vm_page_aflag_set(m, PGA_WRITEABLE);
3997 	}
3998 
3999 	/*
4000 	 * If the mapping or permission bits are different, we need
4001 	 * to update the pte2.
4002 	 *
4003 	 * QQQ: Think again and again what to do
4004 	 *      if the mapping is going to be changed!
4005 	 */
4006 	if ((opte2 & ~(PTE2_NM | PTE2_A)) != (npte2 & ~(PTE2_NM | PTE2_A))) {
4007 		/*
4008 		 * Sync icache if exec permission and attribute VM_MEMATTR_WB_WA
4009 		 * is set. Do it now, before the mapping is stored and made
4010 		 * valid for hardware table walk. If done later, there is a race
4011 		 * for other threads of current process in lazy loading case.
4012 		 * Don't do it for kernel memory which is mapped with exec
4013 		 * permission even if the memory isn't going to hold executable
4014 		 * code. The only time when icache sync is needed is after
4015 		 * kernel module is loaded and the relocation info is processed.
4016 		 * And it's done in elf_cpu_load_file().
4017 		 *
4018 		 * QQQ: (1) Does it exist any better way where
4019 		 *          or how to sync icache?
4020 		 *      (2) Now, we do it on a page basis.
4021 		 */
4022 		if ((prot & VM_PROT_EXECUTE) && pmap != kernel_pmap &&
4023 		    m->md.pat_mode == VM_MEMATTR_WB_WA &&
4024 		    (opa != pa || (opte2 & PTE2_NX)))
4025 			cache_icache_sync_fresh(va, pa, PAGE_SIZE);
4026 
4027 		if (opte2 & PTE2_V) {
4028 			/* Change mapping with break-before-make approach. */
4029 			opte2 = pte2_load_clear(pte2p);
4030 			pmap_tlb_flush(pmap, va);
4031 			pte2_store(pte2p, npte2);
4032 			if (om != NULL) {
4033 				KASSERT((om->oflags & VPO_UNMANAGED) == 0,
4034 				    ("%s: om %p unmanaged", __func__, om));
4035 				if ((opte2 & PTE2_A) != 0)
4036 					vm_page_aflag_set(om, PGA_REFERENCED);
4037 				if (pte2_is_dirty(opte2))
4038 					vm_page_dirty(om);
4039 				if (TAILQ_EMPTY(&om->md.pv_list) &&
4040 				    ((om->flags & PG_FICTITIOUS) != 0 ||
4041 				    TAILQ_EMPTY(&pa_to_pvh(opa)->pv_list)))
4042 					vm_page_aflag_clear(om, PGA_WRITEABLE);
4043 			}
4044 		} else
4045 			pte2_store(pte2p, npte2);
4046 	}
4047 #if 0
4048 	else {
4049 		/*
4050 		 * QQQ: In time when both access and not mofified bits are
4051 		 *      emulated by software, this should not happen. Some
4052 		 *      analysis is need, if this really happen. Missing
4053 		 *      tlb flush somewhere could be the reason.
4054 		 */
4055 		panic("%s: pmap %p va %#x opte2 %x npte2 %x !!", __func__, pmap,
4056 		    va, opte2, npte2);
4057 	}
4058 #endif
4059 
4060 #if VM_NRESERVLEVEL > 0
4061 	/*
4062 	 * If both the L2 page table page and the reservation are fully
4063 	 * populated, then attempt promotion.
4064 	 */
4065 	if ((mpte2 == NULL || pt2_is_full(mpte2, va)) &&
4066 	    sp_enabled && (m->flags & PG_FICTITIOUS) == 0 &&
4067 	    vm_reserv_level_iffullpop(m) == 0)
4068 		pmap_promote_pte1(pmap, pte1p, va);
4069 #endif
4070 
4071 	rv = KERN_SUCCESS;
4072 out:
4073 	sched_unpin();
4074 	rw_wunlock(&pvh_global_lock);
4075 	PMAP_UNLOCK(pmap);
4076 	return (rv);
4077 }
4078 
4079 /*
4080  *  Do the things to unmap a page in a process.
4081  */
4082 static int
pmap_remove_pte2(pmap_t pmap,pt2_entry_t * pte2p,vm_offset_t va,struct spglist * free)4083 pmap_remove_pte2(pmap_t pmap, pt2_entry_t *pte2p, vm_offset_t va,
4084     struct spglist *free)
4085 {
4086 	pt2_entry_t opte2;
4087 	vm_page_t m;
4088 
4089 	rw_assert(&pvh_global_lock, RA_WLOCKED);
4090 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4091 
4092 	/* Clear and invalidate the mapping. */
4093 	opte2 = pte2_load_clear(pte2p);
4094 	pmap_tlb_flush(pmap, va);
4095 
4096 	KASSERT(pte2_is_valid(opte2), ("%s: pmap %p va %#x not link pte2 %#x",
4097 	    __func__, pmap, va, opte2));
4098 
4099 	if (opte2 & PTE2_W)
4100 		pmap->pm_stats.wired_count -= 1;
4101 	pmap->pm_stats.resident_count -= 1;
4102 	if (pte2_is_managed(opte2)) {
4103 		m = PHYS_TO_VM_PAGE(pte2_pa(opte2));
4104 		if (pte2_is_dirty(opte2))
4105 			vm_page_dirty(m);
4106 		if (opte2 & PTE2_A)
4107 			vm_page_aflag_set(m, PGA_REFERENCED);
4108 		pmap_remove_entry(pmap, m, va);
4109 	}
4110 	return (pmap_unuse_pt2(pmap, va, free));
4111 }
4112 
4113 /*
4114  *  Remove a single page from a process address space.
4115  */
4116 static void
pmap_remove_page(pmap_t pmap,vm_offset_t va,struct spglist * free)4117 pmap_remove_page(pmap_t pmap, vm_offset_t va, struct spglist *free)
4118 {
4119 	pt2_entry_t *pte2p;
4120 
4121 	rw_assert(&pvh_global_lock, RA_WLOCKED);
4122 	KASSERT(curthread->td_pinned > 0,
4123 	    ("%s: curthread not pinned", __func__));
4124 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4125 	if ((pte2p = pmap_pte2_quick(pmap, va)) == NULL ||
4126 	    !pte2_is_valid(pte2_load(pte2p)))
4127 		return;
4128 	pmap_remove_pte2(pmap, pte2p, va, free);
4129 }
4130 
4131 /*
4132  *  Remove the given range of addresses from the specified map.
4133  *
4134  *  It is assumed that the start and end are properly
4135  *  rounded to the page size.
4136  */
4137 void
pmap_remove(pmap_t pmap,vm_offset_t sva,vm_offset_t eva)4138 pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
4139 {
4140 	vm_offset_t nextva;
4141 	pt1_entry_t *pte1p, pte1;
4142 	pt2_entry_t *pte2p, pte2;
4143 	struct spglist free;
4144 
4145 	/*
4146 	 * Perform an unsynchronized read. This is, however, safe.
4147 	 */
4148 	if (pmap->pm_stats.resident_count == 0)
4149 		return;
4150 
4151 	SLIST_INIT(&free);
4152 
4153 	rw_wlock(&pvh_global_lock);
4154 	sched_pin();
4155 	PMAP_LOCK(pmap);
4156 
4157 	/*
4158 	 * Special handling of removing one page. A very common
4159 	 * operation and easy to short circuit some code.
4160 	 */
4161 	if (sva + PAGE_SIZE == eva) {
4162 		pte1 = pte1_load(pmap_pte1(pmap, sva));
4163 		if (pte1_is_link(pte1)) {
4164 			pmap_remove_page(pmap, sva, &free);
4165 			goto out;
4166 		}
4167 	}
4168 
4169 	for (; sva < eva; sva = nextva) {
4170 		/*
4171 		 * Calculate address for next L2 page table.
4172 		 */
4173 		nextva = pte1_trunc(sva + PTE1_SIZE);
4174 		if (nextva < sva)
4175 			nextva = eva;
4176 		if (pmap->pm_stats.resident_count == 0)
4177 			break;
4178 
4179 		pte1p = pmap_pte1(pmap, sva);
4180 		pte1 = pte1_load(pte1p);
4181 
4182 		/*
4183 		 * Weed out invalid mappings. Note: we assume that the L1 page
4184 		 * table is always allocated, and in kernel virtual.
4185 		 */
4186 		if (pte1 == 0)
4187 			continue;
4188 
4189 		if (pte1_is_section(pte1)) {
4190 			/*
4191 			 * Are we removing the entire large page?  If not,
4192 			 * demote the mapping and fall through.
4193 			 */
4194 			if (sva + PTE1_SIZE == nextva && eva >= nextva) {
4195 				pmap_remove_pte1(pmap, pte1p, sva, &free);
4196 				continue;
4197 			} else if (!pmap_demote_pte1(pmap, pte1p, sva)) {
4198 				/* The large page mapping was destroyed. */
4199 				continue;
4200 			}
4201 #ifdef INVARIANTS
4202 			else {
4203 				/* Update pte1 after demotion. */
4204 				pte1 = pte1_load(pte1p);
4205 			}
4206 #endif
4207 		}
4208 
4209 		KASSERT(pte1_is_link(pte1), ("%s: pmap %p va %#x pte1 %#x at %p"
4210 		    " is not link", __func__, pmap, sva, pte1, pte1p));
4211 
4212 		/*
4213 		 * Limit our scan to either the end of the va represented
4214 		 * by the current L2 page table page, or to the end of the
4215 		 * range being removed.
4216 		 */
4217 		if (nextva > eva)
4218 			nextva = eva;
4219 
4220 		for (pte2p = pmap_pte2_quick(pmap, sva); sva != nextva;
4221 		    pte2p++, sva += PAGE_SIZE) {
4222 			pte2 = pte2_load(pte2p);
4223 			if (!pte2_is_valid(pte2))
4224 				continue;
4225 			if (pmap_remove_pte2(pmap, pte2p, sva, &free))
4226 				break;
4227 		}
4228 	}
4229 out:
4230 	sched_unpin();
4231 	rw_wunlock(&pvh_global_lock);
4232 	PMAP_UNLOCK(pmap);
4233 	vm_page_free_pages_toq(&free, false);
4234 }
4235 
4236 /*
4237  *	Routine:	pmap_remove_all
4238  *	Function:
4239  *		Removes this physical page from
4240  *		all physical maps in which it resides.
4241  *		Reflects back modify bits to the pager.
4242  *
4243  *	Notes:
4244  *		Original versions of this routine were very
4245  *		inefficient because they iteratively called
4246  *		pmap_remove (slow...)
4247  */
4248 
4249 void
pmap_remove_all(vm_page_t m)4250 pmap_remove_all(vm_page_t m)
4251 {
4252 	struct md_page *pvh;
4253 	pv_entry_t pv;
4254 	pmap_t pmap;
4255 	pt2_entry_t *pte2p, opte2;
4256 	pt1_entry_t *pte1p;
4257 	vm_offset_t va;
4258 	struct spglist free;
4259 
4260 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4261 	    ("%s: page %p is not managed", __func__, m));
4262 	SLIST_INIT(&free);
4263 	rw_wlock(&pvh_global_lock);
4264 	sched_pin();
4265 	if ((m->flags & PG_FICTITIOUS) != 0)
4266 		goto small_mappings;
4267 	pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
4268 	while ((pv = TAILQ_FIRST(&pvh->pv_list)) != NULL) {
4269 		va = pv->pv_va;
4270 		pmap = PV_PMAP(pv);
4271 		PMAP_LOCK(pmap);
4272 		pte1p = pmap_pte1(pmap, va);
4273 		(void)pmap_demote_pte1(pmap, pte1p, va);
4274 		PMAP_UNLOCK(pmap);
4275 	}
4276 small_mappings:
4277 	while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
4278 		pmap = PV_PMAP(pv);
4279 		PMAP_LOCK(pmap);
4280 		pmap->pm_stats.resident_count--;
4281 		pte1p = pmap_pte1(pmap, pv->pv_va);
4282 		KASSERT(!pte1_is_section(pte1_load(pte1p)), ("%s: found "
4283 		    "a 1mpage in page %p's pv list", __func__, m));
4284 		pte2p = pmap_pte2_quick(pmap, pv->pv_va);
4285 		opte2 = pte2_load_clear(pte2p);
4286 		pmap_tlb_flush(pmap, pv->pv_va);
4287 		KASSERT(pte2_is_valid(opte2), ("%s: pmap %p va %x zero pte2",
4288 		    __func__, pmap, pv->pv_va));
4289 		if (pte2_is_wired(opte2))
4290 			pmap->pm_stats.wired_count--;
4291 		if (opte2 & PTE2_A)
4292 			vm_page_aflag_set(m, PGA_REFERENCED);
4293 
4294 		/*
4295 		 * Update the vm_page_t clean and reference bits.
4296 		 */
4297 		if (pte2_is_dirty(opte2))
4298 			vm_page_dirty(m);
4299 		pmap_unuse_pt2(pmap, pv->pv_va, &free);
4300 		TAILQ_REMOVE(&m->md.pv_list, pv, pv_next);
4301 		free_pv_entry(pmap, pv);
4302 		PMAP_UNLOCK(pmap);
4303 	}
4304 	vm_page_aflag_clear(m, PGA_WRITEABLE);
4305 	sched_unpin();
4306 	rw_wunlock(&pvh_global_lock);
4307 	vm_page_free_pages_toq(&free, false);
4308 }
4309 
4310 /*
4311  *  Just subroutine for pmap_remove_pages() to reasonably satisfy
4312  *  good coding style, a.k.a. 80 character line width limit hell.
4313  */
4314 static __inline void
pmap_remove_pte1_quick(pmap_t pmap,pt1_entry_t pte1,pv_entry_t pv,struct spglist * free)4315 pmap_remove_pte1_quick(pmap_t pmap, pt1_entry_t pte1, pv_entry_t pv,
4316     struct spglist *free)
4317 {
4318 	vm_paddr_t pa;
4319 	vm_page_t m, mt, mpt2pg;
4320 	struct md_page *pvh;
4321 
4322 	pa = pte1_pa(pte1);
4323 	m = PHYS_TO_VM_PAGE(pa);
4324 
4325 	KASSERT(m->phys_addr == pa, ("%s: vm_page_t %p addr mismatch %#x %#x",
4326 	    __func__, m, m->phys_addr, pa));
4327 	KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
4328 	    m < &vm_page_array[vm_page_array_size],
4329 	    ("%s: bad pte1 %#x", __func__, pte1));
4330 
4331 	if (pte1_is_dirty(pte1)) {
4332 		for (mt = m; mt < &m[PTE1_SIZE / PAGE_SIZE]; mt++)
4333 			vm_page_dirty(mt);
4334 	}
4335 
4336 	pmap->pm_stats.resident_count -= PTE1_SIZE / PAGE_SIZE;
4337 	pvh = pa_to_pvh(pa);
4338 	TAILQ_REMOVE(&pvh->pv_list, pv, pv_next);
4339 	if (TAILQ_EMPTY(&pvh->pv_list)) {
4340 		for (mt = m; mt < &m[PTE1_SIZE / PAGE_SIZE]; mt++)
4341 			if (TAILQ_EMPTY(&mt->md.pv_list))
4342 				vm_page_aflag_clear(mt, PGA_WRITEABLE);
4343 	}
4344 	mpt2pg = pmap_pt2_page(pmap, pv->pv_va);
4345 	if (mpt2pg != NULL)
4346 		pmap_unwire_pt2_all(pmap, pv->pv_va, mpt2pg, free);
4347 }
4348 
4349 /*
4350  *  Just subroutine for pmap_remove_pages() to reasonably satisfy
4351  *  good coding style, a.k.a. 80 character line width limit hell.
4352  */
4353 static __inline void
pmap_remove_pte2_quick(pmap_t pmap,pt2_entry_t pte2,pv_entry_t pv,struct spglist * free)4354 pmap_remove_pte2_quick(pmap_t pmap, pt2_entry_t pte2, pv_entry_t pv,
4355     struct spglist *free)
4356 {
4357 	vm_paddr_t pa;
4358 	vm_page_t m;
4359 	struct md_page *pvh;
4360 
4361 	pa = pte2_pa(pte2);
4362 	m = PHYS_TO_VM_PAGE(pa);
4363 
4364 	KASSERT(m->phys_addr == pa, ("%s: vm_page_t %p addr mismatch %#x %#x",
4365 	    __func__, m, m->phys_addr, pa));
4366 	KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
4367 	    m < &vm_page_array[vm_page_array_size],
4368 	    ("%s: bad pte2 %#x", __func__, pte2));
4369 
4370 	if (pte2_is_dirty(pte2))
4371 		vm_page_dirty(m);
4372 
4373 	pmap->pm_stats.resident_count--;
4374 	TAILQ_REMOVE(&m->md.pv_list, pv, pv_next);
4375 	if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) {
4376 		pvh = pa_to_pvh(pa);
4377 		if (TAILQ_EMPTY(&pvh->pv_list))
4378 			vm_page_aflag_clear(m, PGA_WRITEABLE);
4379 	}
4380 	pmap_unuse_pt2(pmap, pv->pv_va, free);
4381 }
4382 
4383 /*
4384  *  Remove all pages from specified address space this aids process
4385  *  exit speeds. Also, this code is special cased for current process
4386  *  only, but can have the more generic (and slightly slower) mode enabled.
4387  *  This is much faster than pmap_remove in the case of running down
4388  *  an entire address space.
4389  */
4390 void
pmap_remove_pages(pmap_t pmap)4391 pmap_remove_pages(pmap_t pmap)
4392 {
4393 	pt1_entry_t *pte1p, pte1;
4394 	pt2_entry_t *pte2p, pte2;
4395 	pv_entry_t pv;
4396 	struct pv_chunk *pc, *npc;
4397 	struct spglist free;
4398 	int field, idx;
4399 	int32_t bit;
4400 	uint32_t inuse, bitmask;
4401 	bool allfree;
4402 
4403 	/*
4404 	 * Assert that the given pmap is only active on the current
4405 	 * CPU.  Unfortunately, we cannot block another CPU from
4406 	 * activating the pmap while this function is executing.
4407 	 */
4408 	KASSERT(pmap == vmspace_pmap(curthread->td_proc->p_vmspace),
4409 	    ("%s: non-current pmap %p", __func__, pmap));
4410 #if defined(SMP) && defined(INVARIANTS)
4411 	{
4412 		cpuset_t other_cpus;
4413 
4414 		sched_pin();
4415 		other_cpus = pmap->pm_active;
4416 		CPU_CLR(PCPU_GET(cpuid), &other_cpus);
4417 		sched_unpin();
4418 		KASSERT(CPU_EMPTY(&other_cpus),
4419 		    ("%s: pmap %p active on other cpus", __func__, pmap));
4420 	}
4421 #endif
4422 	SLIST_INIT(&free);
4423 	rw_wlock(&pvh_global_lock);
4424 	PMAP_LOCK(pmap);
4425 	sched_pin();
4426 	TAILQ_FOREACH_SAFE(pc, &pmap->pm_pvchunk, pc_list, npc) {
4427 		KASSERT(pc->pc_pmap == pmap, ("%s: wrong pmap %p %p",
4428 		    __func__, pmap, pc->pc_pmap));
4429 		allfree = true;
4430 		for (field = 0; field < _NPCM; field++) {
4431 			inuse = (~(pc->pc_map[field])) & pc_freemask[field];
4432 			while (inuse != 0) {
4433 				bit = ffs(inuse) - 1;
4434 				bitmask = 1UL << bit;
4435 				idx = field * 32 + bit;
4436 				pv = &pc->pc_pventry[idx];
4437 				inuse &= ~bitmask;
4438 
4439 				/*
4440 				 * Note that we cannot remove wired pages
4441 				 * from a process' mapping at this time
4442 				 */
4443 				pte1p = pmap_pte1(pmap, pv->pv_va);
4444 				pte1 = pte1_load(pte1p);
4445 				if (pte1_is_section(pte1)) {
4446 					if (pte1_is_wired(pte1))  {
4447 						allfree = false;
4448 						continue;
4449 					}
4450 					pte1_clear(pte1p);
4451 					pmap_remove_pte1_quick(pmap, pte1, pv,
4452 					    &free);
4453 				}
4454 				else if (pte1_is_link(pte1)) {
4455 					pte2p = pt2map_entry(pv->pv_va);
4456 					pte2 = pte2_load(pte2p);
4457 
4458 					if (!pte2_is_valid(pte2)) {
4459 						printf("%s: pmap %p va %#x "
4460 						    "pte2 %#x\n", __func__,
4461 						    pmap, pv->pv_va, pte2);
4462 						panic("bad pte2");
4463 					}
4464 
4465 					if (pte2_is_wired(pte2))   {
4466 						allfree = false;
4467 						continue;
4468 					}
4469 					pte2_clear(pte2p);
4470 					pmap_remove_pte2_quick(pmap, pte2, pv,
4471 					    &free);
4472 				} else {
4473 					printf("%s: pmap %p va %#x pte1 %#x\n",
4474 					    __func__, pmap, pv->pv_va, pte1);
4475 					panic("bad pte1");
4476 				}
4477 
4478 				/* Mark free */
4479 				PV_STAT(pv_entry_frees++);
4480 				PV_STAT(pv_entry_spare++);
4481 				pv_entry_count--;
4482 				pc->pc_map[field] |= bitmask;
4483 			}
4484 		}
4485 		if (allfree) {
4486 			TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
4487 			free_pv_chunk(pc);
4488 		}
4489 	}
4490 	tlb_flush_all_ng_local();
4491 	sched_unpin();
4492 	rw_wunlock(&pvh_global_lock);
4493 	PMAP_UNLOCK(pmap);
4494 	vm_page_free_pages_toq(&free, false);
4495 }
4496 
4497 /*
4498  *  This code makes some *MAJOR* assumptions:
4499  *  1. Current pmap & pmap exists.
4500  *  2. Not wired.
4501  *  3. Read access.
4502  *  4. No L2 page table pages.
4503  *  but is *MUCH* faster than pmap_enter...
4504  */
4505 static vm_page_t
pmap_enter_quick_locked(pmap_t pmap,vm_offset_t va,vm_page_t m,vm_prot_t prot,vm_page_t mpt2pg)4506 pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m,
4507     vm_prot_t prot, vm_page_t mpt2pg)
4508 {
4509 	pt2_entry_t *pte2p, pte2;
4510 	vm_paddr_t pa;
4511 	struct spglist free;
4512 	uint32_t l2prot;
4513 
4514 	KASSERT(!VA_IS_CLEANMAP(va) ||
4515 	    (m->oflags & VPO_UNMANAGED) != 0,
4516 	    ("%s: managed mapping within the clean submap", __func__));
4517 	rw_assert(&pvh_global_lock, RA_WLOCKED);
4518 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4519 
4520 	/*
4521 	 * In the case that a L2 page table page is not
4522 	 * resident, we are creating it here.
4523 	 */
4524 	if (va < VM_MAXUSER_ADDRESS) {
4525 		u_int pte1_idx;
4526 		pt1_entry_t pte1, *pte1p;
4527 		vm_paddr_t pt2_pa;
4528 
4529 		/*
4530 		 * Get L1 page table things.
4531 		 */
4532 		pte1_idx = pte1_index(va);
4533 		pte1p = pmap_pte1(pmap, va);
4534 		pte1 = pte1_load(pte1p);
4535 
4536 		if (mpt2pg && (mpt2pg->pindex == (pte1_idx & ~PT2PG_MASK))) {
4537 			/*
4538 			 * Each of NPT2_IN_PG L2 page tables on the page can
4539 			 * come here. Make sure that associated L1 page table
4540 			 * link is established.
4541 			 *
4542 			 * QQQ: It comes that we don't establish all links to
4543 			 *      L2 page tables for newly allocated L2 page
4544 			 *      tables page.
4545 			 */
4546 			KASSERT(!pte1_is_section(pte1),
4547 			    ("%s: pte1 %#x is section", __func__, pte1));
4548 			if (!pte1_is_link(pte1)) {
4549 				pt2_pa = page_pt2pa(VM_PAGE_TO_PHYS(mpt2pg),
4550 				    pte1_idx);
4551 				pte1_store(pte1p, PTE1_LINK(pt2_pa));
4552 			}
4553 			pt2_wirecount_inc(mpt2pg, pte1_idx);
4554 		} else {
4555 			/*
4556 			 * If the L2 page table page is mapped, we just
4557 			 * increment the hold count, and activate it.
4558 			 */
4559 			if (pte1_is_section(pte1)) {
4560 				return (NULL);
4561 			} else if (pte1_is_link(pte1)) {
4562 				mpt2pg = PHYS_TO_VM_PAGE(pte1_link_pa(pte1));
4563 				pt2_wirecount_inc(mpt2pg, pte1_idx);
4564 			} else {
4565 				mpt2pg = _pmap_allocpte2(pmap, va,
4566 				    PMAP_ENTER_NOSLEEP);
4567 				if (mpt2pg == NULL)
4568 					return (NULL);
4569 			}
4570 		}
4571 	} else {
4572 		mpt2pg = NULL;
4573 	}
4574 
4575 	/*
4576 	 * This call to pt2map_entry() makes the assumption that we are
4577 	 * entering the page into the current pmap.  In order to support
4578 	 * quick entry into any pmap, one would likely use pmap_pte2_quick().
4579 	 * But that isn't as quick as pt2map_entry().
4580 	 */
4581 	pte2p = pt2map_entry(va);
4582 	pte2 = pte2_load(pte2p);
4583 	if (pte2_is_valid(pte2)) {
4584 		if (mpt2pg != NULL) {
4585 			/*
4586 			 * Remove extra pte2 reference
4587 			 */
4588 			pt2_wirecount_dec(mpt2pg, pte1_index(va));
4589 			mpt2pg = NULL;
4590 		}
4591 		return (NULL);
4592 	}
4593 
4594 	/*
4595 	 * Enter on the PV list if part of our managed memory.
4596 	 */
4597 	if ((m->oflags & VPO_UNMANAGED) == 0 &&
4598 	    !pmap_try_insert_pv_entry(pmap, va, m)) {
4599 		if (mpt2pg != NULL) {
4600 			SLIST_INIT(&free);
4601 			if (pmap_unwire_pt2(pmap, va, mpt2pg, &free)) {
4602 				pmap_tlb_flush(pmap, va);
4603 				vm_page_free_pages_toq(&free, false);
4604 			}
4605 
4606 			mpt2pg = NULL;
4607 		}
4608 		return (NULL);
4609 	}
4610 
4611 	/*
4612 	 * Increment counters
4613 	 */
4614 	pmap->pm_stats.resident_count++;
4615 
4616 	/*
4617 	 * Now validate mapping with RO protection
4618 	 */
4619 	pa = VM_PAGE_TO_PHYS(m);
4620 	l2prot = PTE2_RO | PTE2_NM;
4621 	if (va < VM_MAXUSER_ADDRESS)
4622 		l2prot |= PTE2_U | PTE2_NG;
4623 	if ((prot & VM_PROT_EXECUTE) == 0)
4624 		l2prot |= PTE2_NX;
4625 	else if (m->md.pat_mode == VM_MEMATTR_WB_WA && pmap != kernel_pmap) {
4626 		/*
4627 		 * Sync icache if exec permission and attribute VM_MEMATTR_WB_WA
4628 		 * is set. QQQ: For more info, see comments in pmap_enter().
4629 		 */
4630 		cache_icache_sync_fresh(va, pa, PAGE_SIZE);
4631 	}
4632 	pte2_store(pte2p, PTE2(pa, l2prot, vm_page_pte2_attr(m)));
4633 
4634 	return (mpt2pg);
4635 }
4636 
4637 void
pmap_enter_quick(pmap_t pmap,vm_offset_t va,vm_page_t m,vm_prot_t prot)4638 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot)
4639 {
4640 
4641 	rw_wlock(&pvh_global_lock);
4642 	PMAP_LOCK(pmap);
4643 	(void)pmap_enter_quick_locked(pmap, va, m, prot, NULL);
4644 	rw_wunlock(&pvh_global_lock);
4645 	PMAP_UNLOCK(pmap);
4646 }
4647 
4648 /*
4649  *  Tries to create a read- and/or execute-only 1 MB page mapping.  Returns
4650  *  true if successful.  Returns false if (1) a mapping already exists at the
4651  *  specified virtual address or (2) a PV entry cannot be allocated without
4652  *  reclaiming another PV entry.
4653  */
4654 static bool
pmap_enter_1mpage(pmap_t pmap,vm_offset_t va,vm_page_t m,vm_prot_t prot)4655 pmap_enter_1mpage(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot)
4656 {
4657 	pt1_entry_t pte1;
4658 	vm_paddr_t pa;
4659 
4660 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4661 	pa = VM_PAGE_TO_PHYS(m);
4662 	pte1 = PTE1(pa, PTE1_NM | PTE1_RO, ATTR_TO_L1(vm_page_pte2_attr(m)));
4663 	if ((prot & VM_PROT_EXECUTE) == 0)
4664 		pte1 |= PTE1_NX;
4665 	if (va < VM_MAXUSER_ADDRESS)
4666 		pte1 |= PTE1_U;
4667 	if (pmap != kernel_pmap)
4668 		pte1 |= PTE1_NG;
4669 	return (pmap_enter_pte1(pmap, va, pte1, PMAP_ENTER_NOSLEEP |
4670 	    PMAP_ENTER_NOREPLACE | PMAP_ENTER_NORECLAIM, m) == KERN_SUCCESS);
4671 }
4672 
4673 /*
4674  *  Tries to create the specified 1 MB page mapping.  Returns KERN_SUCCESS if
4675  *  the mapping was created, and either KERN_FAILURE or KERN_RESOURCE_SHORTAGE
4676  *  otherwise.  Returns KERN_FAILURE if PMAP_ENTER_NOREPLACE was specified and
4677  *  a mapping already exists at the specified virtual address.  Returns
4678  *  KERN_RESOURCE_SHORTAGE if PMAP_ENTER_NORECLAIM was specified and PV entry
4679  *  allocation failed.
4680  */
4681 static int
pmap_enter_pte1(pmap_t pmap,vm_offset_t va,pt1_entry_t pte1,u_int flags,vm_page_t m)4682 pmap_enter_pte1(pmap_t pmap, vm_offset_t va, pt1_entry_t pte1, u_int flags,
4683     vm_page_t m)
4684 {
4685 	struct spglist free;
4686 	pt1_entry_t opte1, *pte1p;
4687 	pt2_entry_t pte2, *pte2p;
4688 	vm_offset_t cur, end;
4689 	vm_page_t mt;
4690 
4691 	rw_assert(&pvh_global_lock, RA_WLOCKED);
4692 	KASSERT((pte1 & (PTE1_NM | PTE1_RO)) == 0 ||
4693 	    (pte1 & (PTE1_NM | PTE1_RO)) == (PTE1_NM | PTE1_RO),
4694 	    ("%s: pte1 has inconsistent NM and RO attributes", __func__));
4695 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4696 	pte1p = pmap_pte1(pmap, va);
4697 	opte1 = pte1_load(pte1p);
4698 	if (pte1_is_valid(opte1)) {
4699 		if ((flags & PMAP_ENTER_NOREPLACE) != 0) {
4700 			CTR3(KTR_PMAP, "%s: failure for va %#lx in pmap %p",
4701 			    __func__, va, pmap);
4702 			return (KERN_FAILURE);
4703 		}
4704 		/* Break the existing mapping(s). */
4705 		SLIST_INIT(&free);
4706 		if (pte1_is_section(opte1)) {
4707 			/*
4708 			 * If the section resulted from a promotion, then a
4709 			 * reserved PT page could be freed.
4710 			 */
4711 			pmap_remove_pte1(pmap, pte1p, va, &free);
4712 		} else {
4713 			sched_pin();
4714 			end = va + PTE1_SIZE;
4715 			for (cur = va, pte2p = pmap_pte2_quick(pmap, va);
4716 			    cur != end; cur += PAGE_SIZE, pte2p++) {
4717 				pte2 = pte2_load(pte2p);
4718 				if (!pte2_is_valid(pte2))
4719 					continue;
4720 				if (pmap_remove_pte2(pmap, pte2p, cur, &free))
4721 					break;
4722 			}
4723 			sched_unpin();
4724 		}
4725 		vm_page_free_pages_toq(&free, false);
4726 	}
4727 	if ((m->oflags & VPO_UNMANAGED) == 0) {
4728 		/*
4729 		 * Abort this mapping if its PV entry could not be created.
4730 		 */
4731 		if (!pmap_pv_insert_pte1(pmap, va, pte1, flags)) {
4732 			CTR3(KTR_PMAP, "%s: failure for va %#lx in pmap %p",
4733 			    __func__, va, pmap);
4734 			return (KERN_RESOURCE_SHORTAGE);
4735 		}
4736 		if ((pte1 & PTE1_RO) == 0) {
4737 			for (mt = m; mt < &m[PTE1_SIZE / PAGE_SIZE]; mt++)
4738 				vm_page_aflag_set(mt, PGA_WRITEABLE);
4739 		}
4740 	}
4741 
4742 	/*
4743 	 * Increment counters.
4744 	 */
4745 	if (pte1_is_wired(pte1))
4746 		pmap->pm_stats.wired_count += PTE1_SIZE / PAGE_SIZE;
4747 	pmap->pm_stats.resident_count += PTE1_SIZE / PAGE_SIZE;
4748 
4749 	/*
4750 	 * Sync icache if exec permission and attribute VM_MEMATTR_WB_WA
4751 	 * is set.  QQQ: For more info, see comments in pmap_enter().
4752 	 */
4753 	if ((pte1 & PTE1_NX) == 0 && m->md.pat_mode == VM_MEMATTR_WB_WA &&
4754 	    pmap != kernel_pmap && (!pte1_is_section(opte1) ||
4755 	    pte1_pa(opte1) != VM_PAGE_TO_PHYS(m) || (opte1 & PTE2_NX) != 0))
4756 		cache_icache_sync_fresh(va, VM_PAGE_TO_PHYS(m), PTE1_SIZE);
4757 
4758 	/*
4759 	 * Map the section.
4760 	 */
4761 	pte1_store(pte1p, pte1);
4762 
4763 	pmap_pte1_mappings++;
4764 	CTR3(KTR_PMAP, "%s: success for va %#lx in pmap %p", __func__, va,
4765 	    pmap);
4766 	return (KERN_SUCCESS);
4767 }
4768 
4769 /*
4770  *  Maps a sequence of resident pages belonging to the same object.
4771  *  The sequence begins with the given page m_start.  This page is
4772  *  mapped at the given virtual address start.  Each subsequent page is
4773  *  mapped at a virtual address that is offset from start by the same
4774  *  amount as the page is offset from m_start within the object.  The
4775  *  last page in the sequence is the page with the largest offset from
4776  *  m_start that can be mapped at a virtual address less than the given
4777  *  virtual address end.  Not every virtual page between start and end
4778  *  is mapped; only those for which a resident page exists with the
4779  *  corresponding offset from m_start are mapped.
4780  */
4781 void
pmap_enter_object(pmap_t pmap,vm_offset_t start,vm_offset_t end,vm_page_t m_start,vm_prot_t prot)4782 pmap_enter_object(pmap_t pmap, vm_offset_t start, vm_offset_t end,
4783     vm_page_t m_start, vm_prot_t prot)
4784 {
4785 	vm_offset_t va;
4786 	vm_page_t m, mpt2pg;
4787 	vm_pindex_t diff, psize;
4788 
4789 	PDEBUG(6, printf("%s: pmap %p start %#x end  %#x m %p prot %#x\n",
4790 	    __func__, pmap, start, end, m_start, prot));
4791 
4792 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
4793 	psize = atop(end - start);
4794 	mpt2pg = NULL;
4795 	m = m_start;
4796 	rw_wlock(&pvh_global_lock);
4797 	PMAP_LOCK(pmap);
4798 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
4799 		va = start + ptoa(diff);
4800 		if ((va & PTE1_OFFSET) == 0 && va + PTE1_SIZE <= end &&
4801 		    m->psind == 1 && sp_enabled &&
4802 		    pmap_enter_1mpage(pmap, va, m, prot))
4803 			m = &m[PTE1_SIZE / PAGE_SIZE - 1];
4804 		else
4805 			mpt2pg = pmap_enter_quick_locked(pmap, va, m, prot,
4806 			    mpt2pg);
4807 		m = TAILQ_NEXT(m, listq);
4808 	}
4809 	rw_wunlock(&pvh_global_lock);
4810 	PMAP_UNLOCK(pmap);
4811 }
4812 
4813 /*
4814  *  This code maps large physical mmap regions into the
4815  *  processor address space.  Note that some shortcuts
4816  *  are taken, but the code works.
4817  */
4818 void
pmap_object_init_pt(pmap_t pmap,vm_offset_t addr,vm_object_t object,vm_pindex_t pindex,vm_size_t size)4819 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_object_t object,
4820     vm_pindex_t pindex, vm_size_t size)
4821 {
4822 	pt1_entry_t *pte1p;
4823 	vm_paddr_t pa, pte2_pa;
4824 	vm_page_t p;
4825 	vm_memattr_t pat_mode;
4826 	u_int l1attr, l1prot;
4827 
4828 	VM_OBJECT_ASSERT_WLOCKED(object);
4829 	KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG,
4830 	    ("%s: non-device object", __func__));
4831 	if ((addr & PTE1_OFFSET) == 0 && (size & PTE1_OFFSET) == 0) {
4832 		if (!vm_object_populate(object, pindex, pindex + atop(size)))
4833 			return;
4834 		p = vm_page_lookup(object, pindex);
4835 		KASSERT(p->valid == VM_PAGE_BITS_ALL,
4836 		    ("%s: invalid page %p", __func__, p));
4837 		pat_mode = p->md.pat_mode;
4838 
4839 		/*
4840 		 * Abort the mapping if the first page is not physically
4841 		 * aligned to a 1MB page boundary.
4842 		 */
4843 		pte2_pa = VM_PAGE_TO_PHYS(p);
4844 		if (pte2_pa & PTE1_OFFSET)
4845 			return;
4846 
4847 		/*
4848 		 * Skip the first page. Abort the mapping if the rest of
4849 		 * the pages are not physically contiguous or have differing
4850 		 * memory attributes.
4851 		 */
4852 		p = TAILQ_NEXT(p, listq);
4853 		for (pa = pte2_pa + PAGE_SIZE; pa < pte2_pa + size;
4854 		    pa += PAGE_SIZE) {
4855 			KASSERT(p->valid == VM_PAGE_BITS_ALL,
4856 			    ("%s: invalid page %p", __func__, p));
4857 			if (pa != VM_PAGE_TO_PHYS(p) ||
4858 			    pat_mode != p->md.pat_mode)
4859 				return;
4860 			p = TAILQ_NEXT(p, listq);
4861 		}
4862 
4863 		/*
4864 		 * Map using 1MB pages.
4865 		 *
4866 		 * QQQ: Well, we are mapping a section, so same condition must
4867 		 * be hold like during promotion. It looks that only RW mapping
4868 		 * is done here, so readonly mapping must be done elsewhere.
4869 		 */
4870 		l1prot = PTE1_U | PTE1_NG | PTE1_RW | PTE1_M | PTE1_A;
4871 		l1attr = ATTR_TO_L1(vm_memattr_to_pte2(pat_mode));
4872 		PMAP_LOCK(pmap);
4873 		for (pa = pte2_pa; pa < pte2_pa + size; pa += PTE1_SIZE) {
4874 			pte1p = pmap_pte1(pmap, addr);
4875 			if (!pte1_is_valid(pte1_load(pte1p))) {
4876 				pte1_store(pte1p, PTE1(pa, l1prot, l1attr));
4877 				pmap->pm_stats.resident_count += PTE1_SIZE /
4878 				    PAGE_SIZE;
4879 				pmap_pte1_mappings++;
4880 			}
4881 			/* Else continue on if the PTE1 is already valid. */
4882 			addr += PTE1_SIZE;
4883 		}
4884 		PMAP_UNLOCK(pmap);
4885 	}
4886 }
4887 
4888 /*
4889  *  Do the things to protect a 1mpage in a process.
4890  */
4891 static void
pmap_protect_pte1(pmap_t pmap,pt1_entry_t * pte1p,vm_offset_t sva,vm_prot_t prot)4892 pmap_protect_pte1(pmap_t pmap, pt1_entry_t *pte1p, vm_offset_t sva,
4893     vm_prot_t prot)
4894 {
4895 	pt1_entry_t npte1, opte1;
4896 	vm_offset_t eva, va;
4897 	vm_page_t m;
4898 
4899 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4900 	KASSERT((sva & PTE1_OFFSET) == 0,
4901 	    ("%s: sva is not 1mpage aligned", __func__));
4902 
4903 	opte1 = npte1 = pte1_load(pte1p);
4904 	if (pte1_is_managed(opte1) && pte1_is_dirty(opte1)) {
4905 		eva = sva + PTE1_SIZE;
4906 		for (va = sva, m = PHYS_TO_VM_PAGE(pte1_pa(opte1));
4907 		    va < eva; va += PAGE_SIZE, m++)
4908 			vm_page_dirty(m);
4909 	}
4910 	if ((prot & VM_PROT_WRITE) == 0)
4911 		npte1 |= PTE1_RO | PTE1_NM;
4912 	if ((prot & VM_PROT_EXECUTE) == 0)
4913 		npte1 |= PTE1_NX;
4914 
4915 	/*
4916 	 * QQQ: Herein, execute permission is never set.
4917 	 *      It only can be cleared. So, no icache
4918 	 *      syncing is needed.
4919 	 */
4920 
4921 	if (npte1 != opte1) {
4922 		pte1_store(pte1p, npte1);
4923 		pmap_tlb_flush(pmap, sva);
4924 	}
4925 }
4926 
4927 /*
4928  *	Set the physical protection on the
4929  *	specified range of this map as requested.
4930  */
4931 void
pmap_protect(pmap_t pmap,vm_offset_t sva,vm_offset_t eva,vm_prot_t prot)4932 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
4933 {
4934 	bool pv_lists_locked;
4935 	vm_offset_t nextva;
4936 	pt1_entry_t *pte1p, pte1;
4937 	pt2_entry_t *pte2p, opte2, npte2;
4938 
4939 	KASSERT((prot & ~VM_PROT_ALL) == 0, ("invalid prot %x", prot));
4940 	if (prot == VM_PROT_NONE) {
4941 		pmap_remove(pmap, sva, eva);
4942 		return;
4943 	}
4944 
4945 	if ((prot & (VM_PROT_WRITE | VM_PROT_EXECUTE)) ==
4946 	    (VM_PROT_WRITE | VM_PROT_EXECUTE))
4947 		return;
4948 
4949 	if (pmap_is_current(pmap))
4950 		pv_lists_locked = false;
4951 	else {
4952 		pv_lists_locked = true;
4953 resume:
4954 		rw_wlock(&pvh_global_lock);
4955 		sched_pin();
4956 	}
4957 
4958 	PMAP_LOCK(pmap);
4959 	for (; sva < eva; sva = nextva) {
4960 		/*
4961 		 * Calculate address for next L2 page table.
4962 		 */
4963 		nextva = pte1_trunc(sva + PTE1_SIZE);
4964 		if (nextva < sva)
4965 			nextva = eva;
4966 
4967 		pte1p = pmap_pte1(pmap, sva);
4968 		pte1 = pte1_load(pte1p);
4969 
4970 		/*
4971 		 * Weed out invalid mappings. Note: we assume that L1 page
4972 		 * page table is always allocated, and in kernel virtual.
4973 		 */
4974 		if (pte1 == 0)
4975 			continue;
4976 
4977 		if (pte1_is_section(pte1)) {
4978 			/*
4979 			 * Are we protecting the entire large page?  If not,
4980 			 * demote the mapping and fall through.
4981 			 */
4982 			if (sva + PTE1_SIZE == nextva && eva >= nextva) {
4983 				pmap_protect_pte1(pmap, pte1p, sva, prot);
4984 				continue;
4985 			} else {
4986 				if (!pv_lists_locked) {
4987 					pv_lists_locked = true;
4988 					if (!rw_try_wlock(&pvh_global_lock)) {
4989 						PMAP_UNLOCK(pmap);
4990 						goto resume;
4991 					}
4992 					sched_pin();
4993 				}
4994 				if (!pmap_demote_pte1(pmap, pte1p, sva)) {
4995 					/*
4996 					 * The large page mapping
4997 					 * was destroyed.
4998 					 */
4999 					continue;
5000 				}
5001 #ifdef INVARIANTS
5002 				else {
5003 					/* Update pte1 after demotion */
5004 					pte1 = pte1_load(pte1p);
5005 				}
5006 #endif
5007 			}
5008 		}
5009 
5010 		KASSERT(pte1_is_link(pte1), ("%s: pmap %p va %#x pte1 %#x at %p"
5011 		    " is not link", __func__, pmap, sva, pte1, pte1p));
5012 
5013 		/*
5014 		 * Limit our scan to either the end of the va represented
5015 		 * by the current L2 page table page, or to the end of the
5016 		 * range being protected.
5017 		 */
5018 		if (nextva > eva)
5019 			nextva = eva;
5020 
5021 		for (pte2p = pmap_pte2_quick(pmap, sva); sva != nextva; pte2p++,
5022 		    sva += PAGE_SIZE) {
5023 			vm_page_t m;
5024 
5025 			opte2 = npte2 = pte2_load(pte2p);
5026 			if (!pte2_is_valid(opte2))
5027 				continue;
5028 
5029 			if ((prot & VM_PROT_WRITE) == 0) {
5030 				if (pte2_is_managed(opte2) &&
5031 				    pte2_is_dirty(opte2)) {
5032 					m = PHYS_TO_VM_PAGE(pte2_pa(opte2));
5033 					vm_page_dirty(m);
5034 				}
5035 				npte2 |= PTE2_RO | PTE2_NM;
5036 			}
5037 
5038 			if ((prot & VM_PROT_EXECUTE) == 0)
5039 				npte2 |= PTE2_NX;
5040 
5041 			/*
5042 			 * QQQ: Herein, execute permission is never set.
5043 			 *      It only can be cleared. So, no icache
5044 			 *      syncing is needed.
5045 			 */
5046 
5047 			if (npte2 != opte2) {
5048 				pte2_store(pte2p, npte2);
5049 				pmap_tlb_flush(pmap, sva);
5050 			}
5051 		}
5052 	}
5053 	if (pv_lists_locked) {
5054 		sched_unpin();
5055 		rw_wunlock(&pvh_global_lock);
5056 	}
5057 	PMAP_UNLOCK(pmap);
5058 }
5059 
5060 /*
5061  *	pmap_pvh_wired_mappings:
5062  *
5063  *	Return the updated number "count" of managed mappings that are wired.
5064  */
5065 static int
pmap_pvh_wired_mappings(struct md_page * pvh,int count)5066 pmap_pvh_wired_mappings(struct md_page *pvh, int count)
5067 {
5068 	pmap_t pmap;
5069 	pt1_entry_t pte1;
5070 	pt2_entry_t pte2;
5071 	pv_entry_t pv;
5072 
5073 	rw_assert(&pvh_global_lock, RA_WLOCKED);
5074 	sched_pin();
5075 	TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) {
5076 		pmap = PV_PMAP(pv);
5077 		PMAP_LOCK(pmap);
5078 		pte1 = pte1_load(pmap_pte1(pmap, pv->pv_va));
5079 		if (pte1_is_section(pte1)) {
5080 			if (pte1_is_wired(pte1))
5081 				count++;
5082 		} else {
5083 			KASSERT(pte1_is_link(pte1),
5084 			    ("%s: pte1 %#x is not link", __func__, pte1));
5085 			pte2 = pte2_load(pmap_pte2_quick(pmap, pv->pv_va));
5086 			if (pte2_is_wired(pte2))
5087 				count++;
5088 		}
5089 		PMAP_UNLOCK(pmap);
5090 	}
5091 	sched_unpin();
5092 	return (count);
5093 }
5094 
5095 /*
5096  *	pmap_page_wired_mappings:
5097  *
5098  *	Return the number of managed mappings to the given physical page
5099  *	that are wired.
5100  */
5101 int
pmap_page_wired_mappings(vm_page_t m)5102 pmap_page_wired_mappings(vm_page_t m)
5103 {
5104 	int count;
5105 
5106 	count = 0;
5107 	if ((m->oflags & VPO_UNMANAGED) != 0)
5108 		return (count);
5109 	rw_wlock(&pvh_global_lock);
5110 	count = pmap_pvh_wired_mappings(&m->md, count);
5111 	if ((m->flags & PG_FICTITIOUS) == 0) {
5112 		count = pmap_pvh_wired_mappings(pa_to_pvh(VM_PAGE_TO_PHYS(m)),
5113 		    count);
5114 	}
5115 	rw_wunlock(&pvh_global_lock);
5116 	return (count);
5117 }
5118 
5119 /*
5120  *  Returns true if any of the given mappings were used to modify
5121  *  physical memory.  Otherwise, returns false.  Both page and 1mpage
5122  *  mappings are supported.
5123  */
5124 static bool
pmap_is_modified_pvh(struct md_page * pvh)5125 pmap_is_modified_pvh(struct md_page *pvh)
5126 {
5127 	pv_entry_t pv;
5128 	pt1_entry_t pte1;
5129 	pt2_entry_t pte2;
5130 	pmap_t pmap;
5131 	bool rv;
5132 
5133 	rw_assert(&pvh_global_lock, RA_WLOCKED);
5134 	rv = false;
5135 	sched_pin();
5136 	TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) {
5137 		pmap = PV_PMAP(pv);
5138 		PMAP_LOCK(pmap);
5139 		pte1 = pte1_load(pmap_pte1(pmap, pv->pv_va));
5140 		if (pte1_is_section(pte1)) {
5141 			rv = pte1_is_dirty(pte1);
5142 		} else {
5143 			KASSERT(pte1_is_link(pte1),
5144 			    ("%s: pte1 %#x is not link", __func__, pte1));
5145 			pte2 = pte2_load(pmap_pte2_quick(pmap, pv->pv_va));
5146 			rv = pte2_is_dirty(pte2);
5147 		}
5148 		PMAP_UNLOCK(pmap);
5149 		if (rv)
5150 			break;
5151 	}
5152 	sched_unpin();
5153 	return (rv);
5154 }
5155 
5156 /*
5157  *	pmap_is_modified:
5158  *
5159  *	Return whether or not the specified physical page was modified
5160  *	in any physical maps.
5161  */
5162 bool
pmap_is_modified(vm_page_t m)5163 pmap_is_modified(vm_page_t m)
5164 {
5165 	bool rv;
5166 
5167 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5168 	    ("%s: page %p is not managed", __func__, m));
5169 
5170 	/*
5171 	 * If the page is not busied then this check is racy.
5172 	 */
5173 	if (!pmap_page_is_write_mapped(m))
5174 		return (false);
5175 	rw_wlock(&pvh_global_lock);
5176 	rv = pmap_is_modified_pvh(&m->md) ||
5177 	    ((m->flags & PG_FICTITIOUS) == 0 &&
5178 	    pmap_is_modified_pvh(pa_to_pvh(VM_PAGE_TO_PHYS(m))));
5179 	rw_wunlock(&pvh_global_lock);
5180 	return (rv);
5181 }
5182 
5183 /*
5184  *	pmap_is_prefaultable:
5185  *
5186  *	Return whether or not the specified virtual address is eligible
5187  *	for prefault.
5188  */
5189 bool
pmap_is_prefaultable(pmap_t pmap,vm_offset_t addr)5190 pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr)
5191 {
5192 	pt1_entry_t pte1;
5193 	pt2_entry_t pte2;
5194 	bool rv;
5195 
5196 	rv = false;
5197 	PMAP_LOCK(pmap);
5198 	pte1 = pte1_load(pmap_pte1(pmap, addr));
5199 	if (pte1_is_link(pte1)) {
5200 		pte2 = pte2_load(pt2map_entry(addr));
5201 		rv = !pte2_is_valid(pte2) ;
5202 	}
5203 	PMAP_UNLOCK(pmap);
5204 	return (rv);
5205 }
5206 
5207 /*
5208  *  Returns true if any of the given mappings were referenced and false
5209  *  otherwise. Both page and 1mpage mappings are supported.
5210  */
5211 static bool
pmap_is_referenced_pvh(struct md_page * pvh)5212 pmap_is_referenced_pvh(struct md_page *pvh)
5213 {
5214 
5215 	pv_entry_t pv;
5216 	pt1_entry_t pte1;
5217 	pt2_entry_t pte2;
5218 	pmap_t pmap;
5219 	bool rv;
5220 
5221 	rw_assert(&pvh_global_lock, RA_WLOCKED);
5222 	rv = false;
5223 	sched_pin();
5224 	TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) {
5225 		pmap = PV_PMAP(pv);
5226 		PMAP_LOCK(pmap);
5227 		pte1 = pte1_load(pmap_pte1(pmap, pv->pv_va));
5228 		if (pte1_is_section(pte1)) {
5229 			rv = (pte1 & (PTE1_A | PTE1_V)) == (PTE1_A | PTE1_V);
5230 		} else {
5231 			pte2 = pte2_load(pmap_pte2_quick(pmap, pv->pv_va));
5232 			rv = (pte2 & (PTE2_A | PTE2_V)) == (PTE2_A | PTE2_V);
5233 		}
5234 		PMAP_UNLOCK(pmap);
5235 		if (rv)
5236 			break;
5237 	}
5238 	sched_unpin();
5239 	return (rv);
5240 }
5241 
5242 /*
5243  *	pmap_is_referenced:
5244  *
5245  *	Return whether or not the specified physical page was referenced
5246  *	in any physical maps.
5247  */
5248 bool
pmap_is_referenced(vm_page_t m)5249 pmap_is_referenced(vm_page_t m)
5250 {
5251 	bool rv;
5252 
5253 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5254 	    ("%s: page %p is not managed", __func__, m));
5255 	rw_wlock(&pvh_global_lock);
5256 	rv = pmap_is_referenced_pvh(&m->md) ||
5257 	    ((m->flags & PG_FICTITIOUS) == 0 &&
5258 	    pmap_is_referenced_pvh(pa_to_pvh(VM_PAGE_TO_PHYS(m))));
5259 	rw_wunlock(&pvh_global_lock);
5260 	return (rv);
5261 }
5262 
5263 /*
5264  *	pmap_ts_referenced:
5265  *
5266  *	Return a count of reference bits for a page, clearing those bits.
5267  *	It is not necessary for every reference bit to be cleared, but it
5268  *	is necessary that 0 only be returned when there are truly no
5269  *	reference bits set.
5270  *
5271  *	As an optimization, update the page's dirty field if a modified bit is
5272  *	found while counting reference bits.  This opportunistic update can be
5273  *	performed at low cost and can eliminate the need for some future calls
5274  *	to pmap_is_modified().  However, since this function stops after
5275  *	finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some
5276  *	dirty pages.  Those dirty pages will only be detected by a future call
5277  *	to pmap_is_modified().
5278  */
5279 int
pmap_ts_referenced(vm_page_t m)5280 pmap_ts_referenced(vm_page_t m)
5281 {
5282 	struct md_page *pvh;
5283 	pv_entry_t pv, pvf;
5284 	pmap_t pmap;
5285 	pt1_entry_t  *pte1p, opte1;
5286 	pt2_entry_t *pte2p, opte2;
5287 	vm_paddr_t pa;
5288 	int rtval = 0;
5289 
5290 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5291 	    ("%s: page %p is not managed", __func__, m));
5292 	pa = VM_PAGE_TO_PHYS(m);
5293 	pvh = pa_to_pvh(pa);
5294 	rw_wlock(&pvh_global_lock);
5295 	sched_pin();
5296 	if ((m->flags & PG_FICTITIOUS) != 0 ||
5297 	    (pvf = TAILQ_FIRST(&pvh->pv_list)) == NULL)
5298 		goto small_mappings;
5299 	pv = pvf;
5300 	do {
5301 		pmap = PV_PMAP(pv);
5302 		PMAP_LOCK(pmap);
5303 		pte1p = pmap_pte1(pmap, pv->pv_va);
5304 		opte1 = pte1_load(pte1p);
5305 		if (pte1_is_dirty(opte1)) {
5306 			/*
5307 			 * Although "opte1" is mapping a 1MB page, because
5308 			 * this function is called at a 4KB page granularity,
5309 			 * we only update the 4KB page under test.
5310 			 */
5311 			vm_page_dirty(m);
5312 		}
5313 		if ((opte1 & PTE1_A) != 0) {
5314 			/*
5315 			 * Since this reference bit is shared by 256 4KB pages,
5316 			 * it should not be cleared every time it is tested.
5317 			 * Apply a simple "hash" function on the physical page
5318 			 * number, the virtual section number, and the pmap
5319 			 * address to select one 4KB page out of the 256
5320 			 * on which testing the reference bit will result
5321 			 * in clearing that bit. This function is designed
5322 			 * to avoid the selection of the same 4KB page
5323 			 * for every 1MB page mapping.
5324 			 *
5325 			 * On demotion, a mapping that hasn't been referenced
5326 			 * is simply destroyed.  To avoid the possibility of a
5327 			 * subsequent page fault on a demoted wired mapping,
5328 			 * always leave its reference bit set.  Moreover,
5329 			 * since the section is wired, the current state of
5330 			 * its reference bit won't affect page replacement.
5331 			 */
5332 			 if ((((pa >> PAGE_SHIFT) ^ (pv->pv_va >> PTE1_SHIFT) ^
5333 			    (uintptr_t)pmap) & (NPTE2_IN_PG - 1)) == 0 &&
5334 			    !pte1_is_wired(opte1)) {
5335 				pte1_clear_bit(pte1p, PTE1_A);
5336 				pmap_tlb_flush(pmap, pv->pv_va);
5337 			}
5338 			rtval++;
5339 		}
5340 		PMAP_UNLOCK(pmap);
5341 		/* Rotate the PV list if it has more than one entry. */
5342 		if (TAILQ_NEXT(pv, pv_next) != NULL) {
5343 			TAILQ_REMOVE(&pvh->pv_list, pv, pv_next);
5344 			TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next);
5345 		}
5346 		if (rtval >= PMAP_TS_REFERENCED_MAX)
5347 			goto out;
5348 	} while ((pv = TAILQ_FIRST(&pvh->pv_list)) != pvf);
5349 small_mappings:
5350 	if ((pvf = TAILQ_FIRST(&m->md.pv_list)) == NULL)
5351 		goto out;
5352 	pv = pvf;
5353 	do {
5354 		pmap = PV_PMAP(pv);
5355 		PMAP_LOCK(pmap);
5356 		pte1p = pmap_pte1(pmap, pv->pv_va);
5357 		KASSERT(pte1_is_link(pte1_load(pte1p)),
5358 		    ("%s: not found a link in page %p's pv list", __func__, m));
5359 
5360 		pte2p = pmap_pte2_quick(pmap, pv->pv_va);
5361 		opte2 = pte2_load(pte2p);
5362 		if (pte2_is_dirty(opte2))
5363 			vm_page_dirty(m);
5364 		if ((opte2 & PTE2_A) != 0) {
5365 			pte2_clear_bit(pte2p, PTE2_A);
5366 			pmap_tlb_flush(pmap, pv->pv_va);
5367 			rtval++;
5368 		}
5369 		PMAP_UNLOCK(pmap);
5370 		/* Rotate the PV list if it has more than one entry. */
5371 		if (TAILQ_NEXT(pv, pv_next) != NULL) {
5372 			TAILQ_REMOVE(&m->md.pv_list, pv, pv_next);
5373 			TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
5374 		}
5375 	} while ((pv = TAILQ_FIRST(&m->md.pv_list)) != pvf && rtval <
5376 	    PMAP_TS_REFERENCED_MAX);
5377 out:
5378 	sched_unpin();
5379 	rw_wunlock(&pvh_global_lock);
5380 	return (rtval);
5381 }
5382 
5383 /*
5384  *	Clear the wired attribute from the mappings for the specified range of
5385  *	addresses in the given pmap.  Every valid mapping within that range
5386  *	must have the wired attribute set.  In contrast, invalid mappings
5387  *	cannot have the wired attribute set, so they are ignored.
5388  *
5389  *	The wired attribute of the page table entry is not a hardware feature,
5390  *	so there is no need to invalidate any TLB entries.
5391  */
5392 void
pmap_unwire(pmap_t pmap,vm_offset_t sva,vm_offset_t eva)5393 pmap_unwire(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
5394 {
5395 	vm_offset_t nextva;
5396 	pt1_entry_t *pte1p, pte1;
5397 	pt2_entry_t *pte2p, pte2;
5398 	bool pv_lists_locked;
5399 
5400 	if (pmap_is_current(pmap))
5401 		pv_lists_locked = false;
5402 	else {
5403 		pv_lists_locked = true;
5404 resume:
5405 		rw_wlock(&pvh_global_lock);
5406 		sched_pin();
5407 	}
5408 	PMAP_LOCK(pmap);
5409 	for (; sva < eva; sva = nextva) {
5410 		nextva = pte1_trunc(sva + PTE1_SIZE);
5411 		if (nextva < sva)
5412 			nextva = eva;
5413 
5414 		pte1p = pmap_pte1(pmap, sva);
5415 		pte1 = pte1_load(pte1p);
5416 
5417 		/*
5418 		 * Weed out invalid mappings. Note: we assume that L1 page
5419 		 * page table is always allocated, and in kernel virtual.
5420 		 */
5421 		if (pte1 == 0)
5422 			continue;
5423 
5424 		if (pte1_is_section(pte1)) {
5425 			if (!pte1_is_wired(pte1))
5426 				panic("%s: pte1 %#x not wired", __func__, pte1);
5427 
5428 			/*
5429 			 * Are we unwiring the entire large page?  If not,
5430 			 * demote the mapping and fall through.
5431 			 */
5432 			if (sva + PTE1_SIZE == nextva && eva >= nextva) {
5433 				pte1_clear_bit(pte1p, PTE1_W);
5434 				pmap->pm_stats.wired_count -= PTE1_SIZE /
5435 				    PAGE_SIZE;
5436 				continue;
5437 			} else {
5438 				if (!pv_lists_locked) {
5439 					pv_lists_locked = true;
5440 					if (!rw_try_wlock(&pvh_global_lock)) {
5441 						PMAP_UNLOCK(pmap);
5442 						/* Repeat sva. */
5443 						goto resume;
5444 					}
5445 					sched_pin();
5446 				}
5447 				if (!pmap_demote_pte1(pmap, pte1p, sva))
5448 					panic("%s: demotion failed", __func__);
5449 #ifdef INVARIANTS
5450 				else {
5451 					/* Update pte1 after demotion */
5452 					pte1 = pte1_load(pte1p);
5453 				}
5454 #endif
5455 			}
5456 		}
5457 
5458 		KASSERT(pte1_is_link(pte1), ("%s: pmap %p va %#x pte1 %#x at %p"
5459 		    " is not link", __func__, pmap, sva, pte1, pte1p));
5460 
5461 		/*
5462 		 * Limit our scan to either the end of the va represented
5463 		 * by the current L2 page table page, or to the end of the
5464 		 * range being protected.
5465 		 */
5466 		if (nextva > eva)
5467 			nextva = eva;
5468 
5469 		for (pte2p = pmap_pte2_quick(pmap, sva); sva != nextva; pte2p++,
5470 		    sva += PAGE_SIZE) {
5471 			pte2 = pte2_load(pte2p);
5472 			if (!pte2_is_valid(pte2))
5473 				continue;
5474 			if (!pte2_is_wired(pte2))
5475 				panic("%s: pte2 %#x is missing PTE2_W",
5476 				    __func__, pte2);
5477 
5478 			/*
5479 			 * PTE2_W must be cleared atomically. Although the pmap
5480 			 * lock synchronizes access to PTE2_W, another processor
5481 			 * could be changing PTE2_NM and/or PTE2_A concurrently.
5482 			 */
5483 			pte2_clear_bit(pte2p, PTE2_W);
5484 			pmap->pm_stats.wired_count--;
5485 		}
5486 	}
5487 	if (pv_lists_locked) {
5488 		sched_unpin();
5489 		rw_wunlock(&pvh_global_lock);
5490 	}
5491 	PMAP_UNLOCK(pmap);
5492 }
5493 
5494 /*
5495  *  Clear the write and modified bits in each of the given page's mappings.
5496  */
5497 void
pmap_remove_write(vm_page_t m)5498 pmap_remove_write(vm_page_t m)
5499 {
5500 	struct md_page *pvh;
5501 	pv_entry_t next_pv, pv;
5502 	pmap_t pmap;
5503 	pt1_entry_t *pte1p;
5504 	pt2_entry_t *pte2p, opte2;
5505 	vm_offset_t va;
5506 
5507 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5508 	    ("%s: page %p is not managed", __func__, m));
5509 	vm_page_assert_busied(m);
5510 
5511 	if (!pmap_page_is_write_mapped(m))
5512 		return;
5513 	rw_wlock(&pvh_global_lock);
5514 	sched_pin();
5515 	if ((m->flags & PG_FICTITIOUS) != 0)
5516 		goto small_mappings;
5517 	pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
5518 	TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) {
5519 		va = pv->pv_va;
5520 		pmap = PV_PMAP(pv);
5521 		PMAP_LOCK(pmap);
5522 		pte1p = pmap_pte1(pmap, va);
5523 		if (!(pte1_load(pte1p) & PTE1_RO))
5524 			(void)pmap_demote_pte1(pmap, pte1p, va);
5525 		PMAP_UNLOCK(pmap);
5526 	}
5527 small_mappings:
5528 	TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) {
5529 		pmap = PV_PMAP(pv);
5530 		PMAP_LOCK(pmap);
5531 		pte1p = pmap_pte1(pmap, pv->pv_va);
5532 		KASSERT(!pte1_is_section(pte1_load(pte1p)), ("%s: found"
5533 		    " a section in page %p's pv list", __func__, m));
5534 		pte2p = pmap_pte2_quick(pmap, pv->pv_va);
5535 		opte2 = pte2_load(pte2p);
5536 		if (!(opte2 & PTE2_RO)) {
5537 			pte2_store(pte2p, opte2 | PTE2_RO | PTE2_NM);
5538 			if (pte2_is_dirty(opte2))
5539 				vm_page_dirty(m);
5540 			pmap_tlb_flush(pmap, pv->pv_va);
5541 		}
5542 		PMAP_UNLOCK(pmap);
5543 	}
5544 	vm_page_aflag_clear(m, PGA_WRITEABLE);
5545 	sched_unpin();
5546 	rw_wunlock(&pvh_global_lock);
5547 }
5548 
5549 /*
5550  *	Apply the given advice to the specified range of addresses within the
5551  *	given pmap.  Depending on the advice, clear the referenced and/or
5552  *	modified flags in each mapping and set the mapped page's dirty field.
5553  */
5554 void
pmap_advise(pmap_t pmap,vm_offset_t sva,vm_offset_t eva,int advice)5555 pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice)
5556 {
5557 	pt1_entry_t *pte1p, opte1;
5558 	pt2_entry_t *pte2p, pte2;
5559 	vm_offset_t pdnxt;
5560 	vm_page_t m;
5561 	bool pv_lists_locked;
5562 
5563 	if (advice != MADV_DONTNEED && advice != MADV_FREE)
5564 		return;
5565 	if (pmap_is_current(pmap))
5566 		pv_lists_locked = false;
5567 	else {
5568 		pv_lists_locked = true;
5569 resume:
5570 		rw_wlock(&pvh_global_lock);
5571 		sched_pin();
5572 	}
5573 	PMAP_LOCK(pmap);
5574 	for (; sva < eva; sva = pdnxt) {
5575 		pdnxt = pte1_trunc(sva + PTE1_SIZE);
5576 		if (pdnxt < sva)
5577 			pdnxt = eva;
5578 		pte1p = pmap_pte1(pmap, sva);
5579 		opte1 = pte1_load(pte1p);
5580 		if (!pte1_is_valid(opte1)) /* XXX */
5581 			continue;
5582 		else if (pte1_is_section(opte1)) {
5583 			if (!pte1_is_managed(opte1))
5584 				continue;
5585 			if (!pv_lists_locked) {
5586 				pv_lists_locked = true;
5587 				if (!rw_try_wlock(&pvh_global_lock)) {
5588 					PMAP_UNLOCK(pmap);
5589 					goto resume;
5590 				}
5591 				sched_pin();
5592 			}
5593 			if (!pmap_demote_pte1(pmap, pte1p, sva)) {
5594 				/*
5595 				 * The large page mapping was destroyed.
5596 				 */
5597 				continue;
5598 			}
5599 
5600 			/*
5601 			 * Unless the page mappings are wired, remove the
5602 			 * mapping to a single page so that a subsequent
5603 			 * access may repromote.  Since the underlying L2 page
5604 			 * table is fully populated, this removal never
5605 			 * frees a L2 page table page.
5606 			 */
5607 			if (!pte1_is_wired(opte1)) {
5608 				pte2p = pmap_pte2_quick(pmap, sva);
5609 				KASSERT(pte2_is_valid(pte2_load(pte2p)),
5610 				    ("%s: invalid PTE2", __func__));
5611 				pmap_remove_pte2(pmap, pte2p, sva, NULL);
5612 			}
5613 		}
5614 		if (pdnxt > eva)
5615 			pdnxt = eva;
5616 		for (pte2p = pmap_pte2_quick(pmap, sva); sva != pdnxt; pte2p++,
5617 		    sva += PAGE_SIZE) {
5618 			pte2 = pte2_load(pte2p);
5619 			if (!pte2_is_valid(pte2) || !pte2_is_managed(pte2))
5620 				continue;
5621 			else if (pte2_is_dirty(pte2)) {
5622 				if (advice == MADV_DONTNEED) {
5623 					/*
5624 					 * Future calls to pmap_is_modified()
5625 					 * can be avoided by making the page
5626 					 * dirty now.
5627 					 */
5628 					m = PHYS_TO_VM_PAGE(pte2_pa(pte2));
5629 					vm_page_dirty(m);
5630 				}
5631 				pte2_set_bit(pte2p, PTE2_NM);
5632 				pte2_clear_bit(pte2p, PTE2_A);
5633 			} else if ((pte2 & PTE2_A) != 0)
5634 				pte2_clear_bit(pte2p, PTE2_A);
5635 			else
5636 				continue;
5637 			pmap_tlb_flush(pmap, sva);
5638 		}
5639 	}
5640 	if (pv_lists_locked) {
5641 		sched_unpin();
5642 		rw_wunlock(&pvh_global_lock);
5643 	}
5644 	PMAP_UNLOCK(pmap);
5645 }
5646 
5647 /*
5648  *	Clear the modify bits on the specified physical page.
5649  */
5650 void
pmap_clear_modify(vm_page_t m)5651 pmap_clear_modify(vm_page_t m)
5652 {
5653 	struct md_page *pvh;
5654 	pv_entry_t next_pv, pv;
5655 	pmap_t pmap;
5656 	pt1_entry_t *pte1p, opte1;
5657 	pt2_entry_t *pte2p, opte2;
5658 	vm_offset_t va;
5659 
5660 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5661 	    ("%s: page %p is not managed", __func__, m));
5662 	vm_page_assert_busied(m);
5663 
5664 	if (!pmap_page_is_write_mapped(m))
5665 		return;
5666 	rw_wlock(&pvh_global_lock);
5667 	sched_pin();
5668 	if ((m->flags & PG_FICTITIOUS) != 0)
5669 		goto small_mappings;
5670 	pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
5671 	TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) {
5672 		va = pv->pv_va;
5673 		pmap = PV_PMAP(pv);
5674 		PMAP_LOCK(pmap);
5675 		pte1p = pmap_pte1(pmap, va);
5676 		opte1 = pte1_load(pte1p);
5677 		if (!(opte1 & PTE1_RO)) {
5678 			if (pmap_demote_pte1(pmap, pte1p, va) &&
5679 			    !pte1_is_wired(opte1)) {
5680 				/*
5681 				 * Write protect the mapping to a
5682 				 * single page so that a subsequent
5683 				 * write access may repromote.
5684 				 */
5685 				va += VM_PAGE_TO_PHYS(m) - pte1_pa(opte1);
5686 				pte2p = pmap_pte2_quick(pmap, va);
5687 				opte2 = pte2_load(pte2p);
5688 				if ((opte2 & PTE2_V)) {
5689 					pte2_set_bit(pte2p, PTE2_NM | PTE2_RO);
5690 					vm_page_dirty(m);
5691 					pmap_tlb_flush(pmap, va);
5692 				}
5693 			}
5694 		}
5695 		PMAP_UNLOCK(pmap);
5696 	}
5697 small_mappings:
5698 	TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) {
5699 		pmap = PV_PMAP(pv);
5700 		PMAP_LOCK(pmap);
5701 		pte1p = pmap_pte1(pmap, pv->pv_va);
5702 		KASSERT(!pte1_is_section(pte1_load(pte1p)), ("%s: found"
5703 		    " a section in page %p's pv list", __func__, m));
5704 		pte2p = pmap_pte2_quick(pmap, pv->pv_va);
5705 		if (pte2_is_dirty(pte2_load(pte2p))) {
5706 			pte2_set_bit(pte2p, PTE2_NM);
5707 			pmap_tlb_flush(pmap, pv->pv_va);
5708 		}
5709 		PMAP_UNLOCK(pmap);
5710 	}
5711 	sched_unpin();
5712 	rw_wunlock(&pvh_global_lock);
5713 }
5714 
5715 /*
5716  *  Sets the memory attribute for the specified page.
5717  */
5718 void
pmap_page_set_memattr(vm_page_t m,vm_memattr_t ma)5719 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5720 {
5721 	pt2_entry_t *cmap2_pte2p;
5722 	vm_memattr_t oma;
5723 	vm_paddr_t pa;
5724 	struct pcpu *pc;
5725 
5726 	oma = m->md.pat_mode;
5727 	m->md.pat_mode = ma;
5728 
5729 	CTR5(KTR_PMAP, "%s: page %p - 0x%08X oma: %d, ma: %d", __func__, m,
5730 	    VM_PAGE_TO_PHYS(m), oma, ma);
5731 	if ((m->flags & PG_FICTITIOUS) != 0)
5732 		return;
5733 #if 0
5734 	/*
5735 	 * If "m" is a normal page, flush it from the cache.
5736 	 *
5737 	 * First, try to find an existing mapping of the page by sf
5738 	 * buffer. sf_buf_invalidate_cache() modifies mapping and
5739 	 * flushes the cache.
5740 	 */
5741 	if (sf_buf_invalidate_cache(m, oma))
5742 		return;
5743 #endif
5744 	/*
5745 	 * If page is not mapped by sf buffer, map the page
5746 	 * transient and do invalidation.
5747 	 */
5748 	if (ma != oma) {
5749 		pa = VM_PAGE_TO_PHYS(m);
5750 		sched_pin();
5751 		pc = get_pcpu();
5752 		cmap2_pte2p = pc->pc_cmap2_pte2p;
5753 		mtx_lock(&pc->pc_cmap_lock);
5754 		if (pte2_load(cmap2_pte2p) != 0)
5755 			panic("%s: CMAP2 busy", __func__);
5756 		pte2_store(cmap2_pte2p, PTE2_KERN_NG(pa, PTE2_AP_KRW,
5757 		    vm_memattr_to_pte2(ma)));
5758 		dcache_wbinv_poc((vm_offset_t)pc->pc_cmap2_addr, pa, PAGE_SIZE);
5759 		pte2_clear(cmap2_pte2p);
5760 		tlb_flush((vm_offset_t)pc->pc_cmap2_addr);
5761 		sched_unpin();
5762 		mtx_unlock(&pc->pc_cmap_lock);
5763 	}
5764 }
5765 
5766 /*
5767  *  Miscellaneous support routines follow
5768  */
5769 
5770 /*
5771  *  Returns true if the given page is mapped individually or as part of
5772  *  a 1mpage.  Otherwise, returns false.
5773  */
5774 bool
pmap_page_is_mapped(vm_page_t m)5775 pmap_page_is_mapped(vm_page_t m)
5776 {
5777 	bool rv;
5778 
5779 	if ((m->oflags & VPO_UNMANAGED) != 0)
5780 		return (false);
5781 	rw_wlock(&pvh_global_lock);
5782 	rv = !TAILQ_EMPTY(&m->md.pv_list) ||
5783 	    ((m->flags & PG_FICTITIOUS) == 0 &&
5784 	    !TAILQ_EMPTY(&pa_to_pvh(VM_PAGE_TO_PHYS(m))->pv_list));
5785 	rw_wunlock(&pvh_global_lock);
5786 	return (rv);
5787 }
5788 
5789 /*
5790  *  Returns true if the pmap's pv is one of the first
5791  *  16 pvs linked to from this page.  This count may
5792  *  be changed upwards or downwards in the future; it
5793  *  is only necessary that true be returned for a small
5794  *  subset of pmaps for proper page aging.
5795  */
5796 bool
pmap_page_exists_quick(pmap_t pmap,vm_page_t m)5797 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
5798 {
5799 	struct md_page *pvh;
5800 	pv_entry_t pv;
5801 	int loops = 0;
5802 	bool rv;
5803 
5804 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5805 	    ("%s: page %p is not managed", __func__, m));
5806 	rv = false;
5807 	rw_wlock(&pvh_global_lock);
5808 	TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) {
5809 		if (PV_PMAP(pv) == pmap) {
5810 			rv = true;
5811 			break;
5812 		}
5813 		loops++;
5814 		if (loops >= 16)
5815 			break;
5816 	}
5817 	if (!rv && loops < 16 && (m->flags & PG_FICTITIOUS) == 0) {
5818 		pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
5819 		TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) {
5820 			if (PV_PMAP(pv) == pmap) {
5821 				rv = true;
5822 				break;
5823 			}
5824 			loops++;
5825 			if (loops >= 16)
5826 				break;
5827 		}
5828 	}
5829 	rw_wunlock(&pvh_global_lock);
5830 	return (rv);
5831 }
5832 
5833 /*
5834  *	pmap_zero_page zeros the specified hardware page by mapping
5835  *	the page into KVM and using bzero to clear its contents.
5836  */
5837 void
pmap_zero_page(vm_page_t m)5838 pmap_zero_page(vm_page_t m)
5839 {
5840 	pt2_entry_t *cmap2_pte2p;
5841 	struct pcpu *pc;
5842 
5843 	sched_pin();
5844 	pc = get_pcpu();
5845 	cmap2_pte2p = pc->pc_cmap2_pte2p;
5846 	mtx_lock(&pc->pc_cmap_lock);
5847 	if (pte2_load(cmap2_pte2p) != 0)
5848 		panic("%s: CMAP2 busy", __func__);
5849 	pte2_store(cmap2_pte2p, PTE2_KERN_NG(VM_PAGE_TO_PHYS(m), PTE2_AP_KRW,
5850 	    vm_page_pte2_attr(m)));
5851 	pagezero(pc->pc_cmap2_addr);
5852 	pte2_clear(cmap2_pte2p);
5853 	tlb_flush((vm_offset_t)pc->pc_cmap2_addr);
5854 	sched_unpin();
5855 	mtx_unlock(&pc->pc_cmap_lock);
5856 }
5857 
5858 /*
5859  *	pmap_zero_page_area zeros the specified hardware page by mapping
5860  *	the page into KVM and using bzero to clear its contents.
5861  *
5862  *	off and size may not cover an area beyond a single hardware page.
5863  */
5864 void
pmap_zero_page_area(vm_page_t m,int off,int size)5865 pmap_zero_page_area(vm_page_t m, int off, int size)
5866 {
5867 	pt2_entry_t *cmap2_pte2p;
5868 	struct pcpu *pc;
5869 
5870 	sched_pin();
5871 	pc = get_pcpu();
5872 	cmap2_pte2p = pc->pc_cmap2_pte2p;
5873 	mtx_lock(&pc->pc_cmap_lock);
5874 	if (pte2_load(cmap2_pte2p) != 0)
5875 		panic("%s: CMAP2 busy", __func__);
5876 	pte2_store(cmap2_pte2p, PTE2_KERN_NG(VM_PAGE_TO_PHYS(m), PTE2_AP_KRW,
5877 	    vm_page_pte2_attr(m)));
5878 	if (off == 0 && size == PAGE_SIZE)
5879 		pagezero(pc->pc_cmap2_addr);
5880 	else
5881 		bzero(pc->pc_cmap2_addr + off, size);
5882 	pte2_clear(cmap2_pte2p);
5883 	tlb_flush((vm_offset_t)pc->pc_cmap2_addr);
5884 	sched_unpin();
5885 	mtx_unlock(&pc->pc_cmap_lock);
5886 }
5887 
5888 /*
5889  *	pmap_copy_page copies the specified (machine independent)
5890  *	page by mapping the page into virtual memory and using
5891  *	bcopy to copy the page, one machine dependent page at a
5892  *	time.
5893  */
5894 void
pmap_copy_page(vm_page_t src,vm_page_t dst)5895 pmap_copy_page(vm_page_t src, vm_page_t dst)
5896 {
5897 	pt2_entry_t *cmap1_pte2p, *cmap2_pte2p;
5898 	struct pcpu *pc;
5899 
5900 	sched_pin();
5901 	pc = get_pcpu();
5902 	cmap1_pte2p = pc->pc_cmap1_pte2p;
5903 	cmap2_pte2p = pc->pc_cmap2_pte2p;
5904 	mtx_lock(&pc->pc_cmap_lock);
5905 	if (pte2_load(cmap1_pte2p) != 0)
5906 		panic("%s: CMAP1 busy", __func__);
5907 	if (pte2_load(cmap2_pte2p) != 0)
5908 		panic("%s: CMAP2 busy", __func__);
5909 	pte2_store(cmap1_pte2p, PTE2_KERN_NG(VM_PAGE_TO_PHYS(src),
5910 	    PTE2_AP_KR | PTE2_NM, vm_page_pte2_attr(src)));
5911 	pte2_store(cmap2_pte2p, PTE2_KERN_NG(VM_PAGE_TO_PHYS(dst),
5912 	    PTE2_AP_KRW, vm_page_pte2_attr(dst)));
5913 	bcopy(pc->pc_cmap1_addr, pc->pc_cmap2_addr, PAGE_SIZE);
5914 	pte2_clear(cmap1_pte2p);
5915 	tlb_flush((vm_offset_t)pc->pc_cmap1_addr);
5916 	pte2_clear(cmap2_pte2p);
5917 	tlb_flush((vm_offset_t)pc->pc_cmap2_addr);
5918 	sched_unpin();
5919 	mtx_unlock(&pc->pc_cmap_lock);
5920 }
5921 
5922 int unmapped_buf_allowed = 1;
5923 
5924 void
pmap_copy_pages(vm_page_t ma[],vm_offset_t a_offset,vm_page_t mb[],vm_offset_t b_offset,int xfersize)5925 pmap_copy_pages(vm_page_t ma[], vm_offset_t a_offset, vm_page_t mb[],
5926     vm_offset_t b_offset, int xfersize)
5927 {
5928 	pt2_entry_t *cmap1_pte2p, *cmap2_pte2p;
5929 	vm_page_t a_pg, b_pg;
5930 	char *a_cp, *b_cp;
5931 	vm_offset_t a_pg_offset, b_pg_offset;
5932 	struct pcpu *pc;
5933 	int cnt;
5934 
5935 	sched_pin();
5936 	pc = get_pcpu();
5937 	cmap1_pte2p = pc->pc_cmap1_pte2p;
5938 	cmap2_pte2p = pc->pc_cmap2_pte2p;
5939 	mtx_lock(&pc->pc_cmap_lock);
5940 	if (pte2_load(cmap1_pte2p) != 0)
5941 		panic("pmap_copy_pages: CMAP1 busy");
5942 	if (pte2_load(cmap2_pte2p) != 0)
5943 		panic("pmap_copy_pages: CMAP2 busy");
5944 	while (xfersize > 0) {
5945 		a_pg = ma[a_offset >> PAGE_SHIFT];
5946 		a_pg_offset = a_offset & PAGE_MASK;
5947 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
5948 		b_pg = mb[b_offset >> PAGE_SHIFT];
5949 		b_pg_offset = b_offset & PAGE_MASK;
5950 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
5951 		pte2_store(cmap1_pte2p, PTE2_KERN_NG(VM_PAGE_TO_PHYS(a_pg),
5952 		    PTE2_AP_KR | PTE2_NM, vm_page_pte2_attr(a_pg)));
5953 		tlb_flush_local((vm_offset_t)pc->pc_cmap1_addr);
5954 		pte2_store(cmap2_pte2p, PTE2_KERN_NG(VM_PAGE_TO_PHYS(b_pg),
5955 		    PTE2_AP_KRW, vm_page_pte2_attr(b_pg)));
5956 		tlb_flush_local((vm_offset_t)pc->pc_cmap2_addr);
5957 		a_cp = pc->pc_cmap1_addr + a_pg_offset;
5958 		b_cp = pc->pc_cmap2_addr + b_pg_offset;
5959 		bcopy(a_cp, b_cp, cnt);
5960 		a_offset += cnt;
5961 		b_offset += cnt;
5962 		xfersize -= cnt;
5963 	}
5964 	pte2_clear(cmap1_pte2p);
5965 	tlb_flush((vm_offset_t)pc->pc_cmap1_addr);
5966 	pte2_clear(cmap2_pte2p);
5967 	tlb_flush((vm_offset_t)pc->pc_cmap2_addr);
5968 	sched_unpin();
5969 	mtx_unlock(&pc->pc_cmap_lock);
5970 }
5971 
5972 vm_offset_t
pmap_quick_enter_page(vm_page_t m)5973 pmap_quick_enter_page(vm_page_t m)
5974 {
5975 	struct pcpu *pc;
5976 	pt2_entry_t *pte2p;
5977 
5978 	critical_enter();
5979 	pc = get_pcpu();
5980 	pte2p = pc->pc_qmap_pte2p;
5981 
5982 	KASSERT(pte2_load(pte2p) == 0, ("%s: PTE2 busy", __func__));
5983 
5984 	pte2_store(pte2p, PTE2_KERN_NG(VM_PAGE_TO_PHYS(m), PTE2_AP_KRW,
5985 	    vm_page_pte2_attr(m)));
5986 	return (pc->pc_qmap_addr);
5987 }
5988 
5989 void
pmap_quick_remove_page(vm_offset_t addr)5990 pmap_quick_remove_page(vm_offset_t addr)
5991 {
5992 	struct pcpu *pc;
5993 	pt2_entry_t *pte2p;
5994 
5995 	pc = get_pcpu();
5996 	pte2p = pc->pc_qmap_pte2p;
5997 
5998 	KASSERT(addr == pc->pc_qmap_addr, ("%s: invalid address", __func__));
5999 	KASSERT(pte2_load(pte2p) != 0, ("%s: PTE2 not in use", __func__));
6000 
6001 	pte2_clear(pte2p);
6002 	tlb_flush(pc->pc_qmap_addr);
6003 	critical_exit();
6004 }
6005 
6006 /*
6007  *	Copy the range specified by src_addr/len
6008  *	from the source map to the range dst_addr/len
6009  *	in the destination map.
6010  *
6011  *	This routine is only advisory and need not do anything.
6012  */
6013 void
pmap_copy(pmap_t dst_pmap,pmap_t src_pmap,vm_offset_t dst_addr,vm_size_t len,vm_offset_t src_addr)6014 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len,
6015     vm_offset_t src_addr)
6016 {
6017 	struct spglist free;
6018 	vm_offset_t addr;
6019 	vm_offset_t end_addr = src_addr + len;
6020 	vm_offset_t nextva;
6021 
6022 	if (dst_addr != src_addr)
6023 		return;
6024 
6025 	if (!pmap_is_current(src_pmap))
6026 		return;
6027 
6028 	rw_wlock(&pvh_global_lock);
6029 	if (dst_pmap < src_pmap) {
6030 		PMAP_LOCK(dst_pmap);
6031 		PMAP_LOCK(src_pmap);
6032 	} else {
6033 		PMAP_LOCK(src_pmap);
6034 		PMAP_LOCK(dst_pmap);
6035 	}
6036 	sched_pin();
6037 	for (addr = src_addr; addr < end_addr; addr = nextva) {
6038 		pt2_entry_t *src_pte2p, *dst_pte2p;
6039 		vm_page_t dst_mpt2pg, src_mpt2pg;
6040 		pt1_entry_t src_pte1;
6041 		u_int pte1_idx;
6042 
6043 		KASSERT(addr < VM_MAXUSER_ADDRESS,
6044 		    ("%s: invalid to pmap_copy page tables", __func__));
6045 
6046 		nextva = pte1_trunc(addr + PTE1_SIZE);
6047 		if (nextva < addr)
6048 			nextva = end_addr;
6049 
6050 		pte1_idx = pte1_index(addr);
6051 		src_pte1 = src_pmap->pm_pt1[pte1_idx];
6052 		if (pte1_is_section(src_pte1)) {
6053 			if ((addr & PTE1_OFFSET) != 0 ||
6054 			    (addr + PTE1_SIZE) > end_addr)
6055 				continue;
6056 			if (dst_pmap->pm_pt1[pte1_idx] == 0 &&
6057 			    (!pte1_is_managed(src_pte1) ||
6058 			    pmap_pv_insert_pte1(dst_pmap, addr, src_pte1,
6059 			    PMAP_ENTER_NORECLAIM))) {
6060 				dst_pmap->pm_pt1[pte1_idx] = src_pte1 &
6061 				    ~PTE1_W;
6062 				dst_pmap->pm_stats.resident_count +=
6063 				    PTE1_SIZE / PAGE_SIZE;
6064 				pmap_pte1_mappings++;
6065 			}
6066 			continue;
6067 		} else if (!pte1_is_link(src_pte1))
6068 			continue;
6069 
6070 		src_mpt2pg = PHYS_TO_VM_PAGE(pte1_link_pa(src_pte1));
6071 
6072 		/*
6073 		 * We leave PT2s to be linked from PT1 even if they are not
6074 		 * referenced until all PT2s in a page are without reference.
6075 		 *
6076 		 * QQQ: It could be changed ...
6077 		 */
6078 #if 0 /* single_pt2_link_is_cleared */
6079 		KASSERT(pt2_wirecount_get(src_mpt2pg, pte1_idx) > 0,
6080 		    ("%s: source page table page is unused", __func__));
6081 #else
6082 		if (pt2_wirecount_get(src_mpt2pg, pte1_idx) == 0)
6083 			continue;
6084 #endif
6085 		if (nextva > end_addr)
6086 			nextva = end_addr;
6087 
6088 		src_pte2p = pt2map_entry(addr);
6089 		while (addr < nextva) {
6090 			pt2_entry_t temp_pte2;
6091 			temp_pte2 = pte2_load(src_pte2p);
6092 			/*
6093 			 * we only virtual copy managed pages
6094 			 */
6095 			if (pte2_is_managed(temp_pte2)) {
6096 				dst_mpt2pg = pmap_allocpte2(dst_pmap, addr,
6097 				    PMAP_ENTER_NOSLEEP);
6098 				if (dst_mpt2pg == NULL)
6099 					goto out;
6100 				dst_pte2p = pmap_pte2_quick(dst_pmap, addr);
6101 				if (!pte2_is_valid(pte2_load(dst_pte2p)) &&
6102 				    pmap_try_insert_pv_entry(dst_pmap, addr,
6103 				    PHYS_TO_VM_PAGE(pte2_pa(temp_pte2)))) {
6104 					/*
6105 					 * Clear the wired, modified, and
6106 					 * accessed (referenced) bits
6107 					 * during the copy.
6108 					 */
6109 					temp_pte2 &=  ~(PTE2_W | PTE2_A);
6110 					temp_pte2 |= PTE2_NM;
6111 					pte2_store(dst_pte2p, temp_pte2);
6112 					dst_pmap->pm_stats.resident_count++;
6113 				} else {
6114 					SLIST_INIT(&free);
6115 					if (pmap_unwire_pt2(dst_pmap, addr,
6116 					    dst_mpt2pg, &free)) {
6117 						pmap_tlb_flush(dst_pmap, addr);
6118 						vm_page_free_pages_toq(&free,
6119 						    false);
6120 					}
6121 					goto out;
6122 				}
6123 				if (pt2_wirecount_get(dst_mpt2pg, pte1_idx) >=
6124 				    pt2_wirecount_get(src_mpt2pg, pte1_idx))
6125 					break;
6126 			}
6127 			addr += PAGE_SIZE;
6128 			src_pte2p++;
6129 		}
6130 	}
6131 out:
6132 	sched_unpin();
6133 	rw_wunlock(&pvh_global_lock);
6134 	PMAP_UNLOCK(src_pmap);
6135 	PMAP_UNLOCK(dst_pmap);
6136 }
6137 
6138 /*
6139  *	Increase the starting virtual address of the given mapping if a
6140  *	different alignment might result in more section mappings.
6141  */
6142 void
pmap_align_superpage(vm_object_t object,vm_ooffset_t offset,vm_offset_t * addr,vm_size_t size)6143 pmap_align_superpage(vm_object_t object, vm_ooffset_t offset,
6144     vm_offset_t *addr, vm_size_t size)
6145 {
6146 	vm_offset_t pte1_offset;
6147 
6148 	if (size < PTE1_SIZE)
6149 		return;
6150 	if (object != NULL && (object->flags & OBJ_COLORED) != 0)
6151 		offset += ptoa(object->pg_color);
6152 	pte1_offset = offset & PTE1_OFFSET;
6153 	if (size - ((PTE1_SIZE - pte1_offset) & PTE1_OFFSET) < PTE1_SIZE ||
6154 	    (*addr & PTE1_OFFSET) == pte1_offset)
6155 		return;
6156 	if ((*addr & PTE1_OFFSET) < pte1_offset)
6157 		*addr = pte1_trunc(*addr) + pte1_offset;
6158 	else
6159 		*addr = pte1_roundup(*addr) + pte1_offset;
6160 }
6161 
6162 void
pmap_activate(struct thread * td)6163 pmap_activate(struct thread *td)
6164 {
6165 	pmap_t pmap, oldpmap;
6166 	u_int cpuid, ttb;
6167 
6168 	PDEBUG(9, printf("%s: td = %08x\n", __func__, (uint32_t)td));
6169 
6170 	critical_enter();
6171 	pmap = vmspace_pmap(td->td_proc->p_vmspace);
6172 	oldpmap = PCPU_GET(curpmap);
6173 	cpuid = PCPU_GET(cpuid);
6174 
6175 #if defined(SMP)
6176 	CPU_CLR_ATOMIC(cpuid, &oldpmap->pm_active);
6177 	CPU_SET_ATOMIC(cpuid, &pmap->pm_active);
6178 #else
6179 	CPU_CLR(cpuid, &oldpmap->pm_active);
6180 	CPU_SET(cpuid, &pmap->pm_active);
6181 #endif
6182 
6183 	ttb = pmap_ttb_get(pmap);
6184 
6185 	/*
6186 	 * pmap_activate is for the current thread on the current cpu
6187 	 */
6188 	td->td_pcb->pcb_pagedir = ttb;
6189 	cp15_ttbr_set(ttb);
6190 	PCPU_SET(curpmap, pmap);
6191 	critical_exit();
6192 }
6193 
6194 void
pmap_active_cpus(pmap_t pmap,cpuset_t * res)6195 pmap_active_cpus(pmap_t pmap, cpuset_t *res)
6196 {
6197 	*res = pmap->pm_active;
6198 }
6199 
6200 /*
6201  * Perform the pmap work for mincore(2).  If the page is not both referenced and
6202  * modified by this pmap, returns its physical address so that the caller can
6203  * find other mappings.
6204  */
6205 int
pmap_mincore(pmap_t pmap,vm_offset_t addr,vm_paddr_t * pap)6206 pmap_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *pap)
6207 {
6208 	pt1_entry_t *pte1p, pte1;
6209 	pt2_entry_t *pte2p, pte2;
6210 	vm_paddr_t pa;
6211 	bool managed;
6212 	int val;
6213 
6214 	PMAP_LOCK(pmap);
6215 	pte1p = pmap_pte1(pmap, addr);
6216 	pte1 = pte1_load(pte1p);
6217 	if (pte1_is_section(pte1)) {
6218 		pa = trunc_page(pte1_pa(pte1) | (addr & PTE1_OFFSET));
6219 		managed = pte1_is_managed(pte1);
6220 		val = MINCORE_PSIND(1) | MINCORE_INCORE;
6221 		if (pte1_is_dirty(pte1))
6222 			val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER;
6223 		if (pte1 & PTE1_A)
6224 			val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER;
6225 	} else if (pte1_is_link(pte1)) {
6226 		pte2p = pmap_pte2(pmap, addr);
6227 		pte2 = pte2_load(pte2p);
6228 		pmap_pte2_release(pte2p);
6229 		pa = pte2_pa(pte2);
6230 		managed = pte2_is_managed(pte2);
6231 		val = MINCORE_INCORE;
6232 		if (pte2_is_dirty(pte2))
6233 			val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER;
6234 		if (pte2 & PTE2_A)
6235 			val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER;
6236 	} else {
6237 		managed = false;
6238 		val = 0;
6239 	}
6240 	if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) !=
6241 	    (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) && managed) {
6242 		*pap = pa;
6243 	}
6244 	PMAP_UNLOCK(pmap);
6245 	return (val);
6246 }
6247 
6248 void
pmap_kenter_device(vm_offset_t va,vm_size_t size,vm_paddr_t pa)6249 pmap_kenter_device(vm_offset_t va, vm_size_t size, vm_paddr_t pa)
6250 {
6251 	vm_offset_t sva;
6252 	uint32_t l2attr;
6253 
6254 	KASSERT((size & PAGE_MASK) == 0,
6255 	    ("%s: device mapping not page-sized", __func__));
6256 
6257 	sva = va;
6258 	l2attr = vm_memattr_to_pte2(VM_MEMATTR_DEVICE);
6259 	while (size != 0) {
6260 		pmap_kenter_prot_attr(va, pa, PTE2_AP_KRW, l2attr);
6261 		va += PAGE_SIZE;
6262 		pa += PAGE_SIZE;
6263 		size -= PAGE_SIZE;
6264 	}
6265 	tlb_flush_range(sva, va - sva);
6266 }
6267 
6268 void
pmap_kremove_device(vm_offset_t va,vm_size_t size)6269 pmap_kremove_device(vm_offset_t va, vm_size_t size)
6270 {
6271 	vm_offset_t sva;
6272 
6273 	KASSERT((size & PAGE_MASK) == 0,
6274 	    ("%s: device mapping not page-sized", __func__));
6275 
6276 	sva = va;
6277 	while (size != 0) {
6278 		pmap_kremove(va);
6279 		va += PAGE_SIZE;
6280 		size -= PAGE_SIZE;
6281 	}
6282 	tlb_flush_range(sva, va - sva);
6283 }
6284 
6285 void
pmap_set_pcb_pagedir(pmap_t pmap,struct pcb * pcb)6286 pmap_set_pcb_pagedir(pmap_t pmap, struct pcb *pcb)
6287 {
6288 
6289 	pcb->pcb_pagedir = pmap_ttb_get(pmap);
6290 }
6291 
6292 /*
6293  *  Clean L1 data cache range by physical address.
6294  *  The range must be within a single page.
6295  */
6296 static void
pmap_dcache_wb_pou(vm_paddr_t pa,vm_size_t size,uint32_t attr)6297 pmap_dcache_wb_pou(vm_paddr_t pa, vm_size_t size, uint32_t attr)
6298 {
6299 	pt2_entry_t *cmap2_pte2p;
6300 	struct pcpu *pc;
6301 
6302 	KASSERT(((pa & PAGE_MASK) + size) <= PAGE_SIZE,
6303 	    ("%s: not on single page", __func__));
6304 
6305 	sched_pin();
6306 	pc = get_pcpu();
6307 	cmap2_pte2p = pc->pc_cmap2_pte2p;
6308 	mtx_lock(&pc->pc_cmap_lock);
6309 	if (pte2_load(cmap2_pte2p) != 0)
6310 		panic("%s: CMAP2 busy", __func__);
6311 	pte2_store(cmap2_pte2p, PTE2_KERN_NG(pa, PTE2_AP_KRW, attr));
6312 	dcache_wb_pou((vm_offset_t)pc->pc_cmap2_addr + (pa & PAGE_MASK), size);
6313 	pte2_clear(cmap2_pte2p);
6314 	tlb_flush((vm_offset_t)pc->pc_cmap2_addr);
6315 	sched_unpin();
6316 	mtx_unlock(&pc->pc_cmap_lock);
6317 }
6318 
6319 /*
6320  *  Sync instruction cache range which is not mapped yet.
6321  */
6322 void
cache_icache_sync_fresh(vm_offset_t va,vm_paddr_t pa,vm_size_t size)6323 cache_icache_sync_fresh(vm_offset_t va, vm_paddr_t pa, vm_size_t size)
6324 {
6325 	uint32_t len, offset;
6326 	vm_page_t m;
6327 
6328 	/* Write back d-cache on given address range. */
6329 	offset = pa & PAGE_MASK;
6330 	for ( ; size != 0; size -= len, pa += len, offset = 0) {
6331 		len = min(PAGE_SIZE - offset, size);
6332 		m = PHYS_TO_VM_PAGE(pa);
6333 		KASSERT(m != NULL, ("%s: vm_page_t is null for %#x",
6334 		  __func__, pa));
6335 		pmap_dcache_wb_pou(pa, len, vm_page_pte2_attr(m));
6336 	}
6337 	/*
6338 	 * I-cache is VIPT. Only way how to flush all virtual mappings
6339 	 * on given physical address is to invalidate all i-cache.
6340 	 */
6341 	icache_inv_all();
6342 }
6343 
6344 void
pmap_sync_icache(pmap_t pmap,vm_offset_t va,vm_size_t size)6345 pmap_sync_icache(pmap_t pmap, vm_offset_t va, vm_size_t size)
6346 {
6347 
6348 	/* Write back d-cache on given address range. */
6349 	if (va >= VM_MIN_KERNEL_ADDRESS) {
6350 		dcache_wb_pou(va, size);
6351 	} else {
6352 		uint32_t len, offset;
6353 		vm_paddr_t pa;
6354 		vm_page_t m;
6355 
6356 		offset = va & PAGE_MASK;
6357 		for ( ; size != 0; size -= len, va += len, offset = 0) {
6358 			pa = pmap_extract(pmap, va); /* offset is preserved */
6359 			len = min(PAGE_SIZE - offset, size);
6360 			m = PHYS_TO_VM_PAGE(pa);
6361 			KASSERT(m != NULL, ("%s: vm_page_t is null for %#x",
6362 				__func__, pa));
6363 			pmap_dcache_wb_pou(pa, len, vm_page_pte2_attr(m));
6364 		}
6365 	}
6366 	/*
6367 	 * I-cache is VIPT. Only way how to flush all virtual mappings
6368 	 * on given physical address is to invalidate all i-cache.
6369 	 */
6370 	icache_inv_all();
6371 }
6372 
6373 /*
6374  *  The implementation of pmap_fault() uses IN_RANGE2() macro which
6375  *  depends on the fact that given range size is a power of 2.
6376  */
6377 CTASSERT(powerof2(NB_IN_PT1));
6378 CTASSERT(powerof2(PT2MAP_SIZE));
6379 
6380 #define IN_RANGE2(addr, start, size)	\
6381     ((vm_offset_t)(start) == ((vm_offset_t)(addr) & ~((size) - 1)))
6382 
6383 /*
6384  *  Handle access and R/W emulation faults.
6385  */
6386 int
pmap_fault(pmap_t pmap,vm_offset_t far,uint32_t fsr,int idx,bool usermode)6387 pmap_fault(pmap_t pmap, vm_offset_t far, uint32_t fsr, int idx, bool usermode)
6388 {
6389 	pt1_entry_t *pte1p, pte1;
6390 	pt2_entry_t *pte2p, pte2;
6391 
6392 	if (pmap == NULL)
6393 		pmap = kernel_pmap;
6394 
6395 	/*
6396 	 * In kernel, we should never get abort with FAR which is in range of
6397 	 * pmap->pm_pt1 or PT2MAP address spaces. If it happens, stop here
6398 	 * and print out a useful abort message and even get to the debugger
6399 	 * otherwise it likely ends with never ending loop of aborts.
6400 	 */
6401 	if (__predict_false(IN_RANGE2(far, pmap->pm_pt1, NB_IN_PT1))) {
6402 		/*
6403 		 * All L1 tables should always be mapped and present.
6404 		 * However, we check only current one herein. For user mode,
6405 		 * only permission abort from malicious user is not fatal.
6406 		 * And alignment abort as it may have higher priority.
6407 		 */
6408 		if (!usermode || (idx != FAULT_ALIGN && idx != FAULT_PERM_L2)) {
6409 			CTR4(KTR_PMAP, "%s: pmap %#x pm_pt1 %#x far %#x",
6410 			    __func__, pmap, pmap->pm_pt1, far);
6411 			panic("%s: pm_pt1 abort", __func__);
6412 		}
6413 		return (KERN_INVALID_ADDRESS);
6414 	}
6415 	if (__predict_false(IN_RANGE2(far, PT2MAP, PT2MAP_SIZE))) {
6416 		/*
6417 		 * PT2MAP should be always mapped and present in current
6418 		 * L1 table. However, only existing L2 tables are mapped
6419 		 * in PT2MAP. For user mode, only L2 translation abort and
6420 		 * permission abort from malicious user is not fatal.
6421 		 * And alignment abort as it may have higher priority.
6422 		 */
6423 		if (!usermode || (idx != FAULT_ALIGN &&
6424 		    idx != FAULT_TRAN_L2 && idx != FAULT_PERM_L2)) {
6425 			CTR4(KTR_PMAP, "%s: pmap %#x PT2MAP %#x far %#x",
6426 			    __func__, pmap, PT2MAP, far);
6427 			panic("%s: PT2MAP abort", __func__);
6428 		}
6429 		return (KERN_INVALID_ADDRESS);
6430 	}
6431 
6432 	/*
6433 	 * A pmap lock is used below for handling of access and R/W emulation
6434 	 * aborts. They were handled by atomic operations before so some
6435 	 * analysis of new situation is needed to answer the following question:
6436 	 * Is it safe to use the lock even for these aborts?
6437 	 *
6438 	 * There may happen two cases in general:
6439 	 *
6440 	 * (1) Aborts while the pmap lock is locked already - this should not
6441 	 * happen as pmap lock is not recursive. However, under pmap lock only
6442 	 * internal kernel data should be accessed and such data should be
6443 	 * mapped with A bit set and NM bit cleared. If double abort happens,
6444 	 * then a mapping of data which has caused it must be fixed. Further,
6445 	 * all new mappings are always made with A bit set and the bit can be
6446 	 * cleared only on managed mappings.
6447 	 *
6448 	 * (2) Aborts while another lock(s) is/are locked - this already can
6449 	 * happen. However, there is no difference here if it's either access or
6450 	 * R/W emulation abort, or if it's some other abort.
6451 	 */
6452 
6453 	PMAP_LOCK(pmap);
6454 #ifdef INVARIANTS
6455 	pte1 = pte1_load(pmap_pte1(pmap, far));
6456 	if (pte1_is_link(pte1)) {
6457 		/*
6458 		 * Check in advance that associated L2 page table is mapped into
6459 		 * PT2MAP space. Note that faulty access to not mapped L2 page
6460 		 * table is caught in more general check above where "far" is
6461 		 * checked that it does not lay in PT2MAP space. Note also that
6462 		 * L1 page table and PT2TAB always exist and are mapped.
6463 		 */
6464 		pte2 = pt2tab_load(pmap_pt2tab_entry(pmap, far));
6465 		if (!pte2_is_valid(pte2))
6466 			panic("%s: missing L2 page table (%p, %#x)",
6467 			    __func__, pmap, far);
6468 	}
6469 #endif
6470 #ifdef SMP
6471 	/*
6472 	 * Special treatment is due to break-before-make approach done when
6473 	 * pte1 is updated for userland mapping during section promotion or
6474 	 * demotion. If not caught here, pmap_enter() can find a section
6475 	 * mapping on faulting address. That is not allowed.
6476 	 */
6477 	if (idx == FAULT_TRAN_L1 && usermode && cp15_ats1cur_check(far) == 0) {
6478 		PMAP_UNLOCK(pmap);
6479 		return (KERN_SUCCESS);
6480 	}
6481 #endif
6482 	/*
6483 	 * Access bits for page and section. Note that the entry
6484 	 * is not in TLB yet, so TLB flush is not necessary.
6485 	 *
6486 	 * QQQ: This is hardware emulation, we do not call userret()
6487 	 *      for aborts from user mode.
6488 	 */
6489 	if (idx == FAULT_ACCESS_L2) {
6490 		pte1 = pte1_load(pmap_pte1(pmap, far));
6491 		if (pte1_is_link(pte1)) {
6492 			/* L2 page table should exist and be mapped. */
6493 			pte2p = pt2map_entry(far);
6494 			pte2 = pte2_load(pte2p);
6495 			if (pte2_is_valid(pte2)) {
6496 				pte2_store(pte2p, pte2 | PTE2_A);
6497 				PMAP_UNLOCK(pmap);
6498 				return (KERN_SUCCESS);
6499 			}
6500 		} else {
6501 			/*
6502 			 * We got L2 access fault but PTE1 is not a link.
6503 			 * Probably some race happened, do nothing.
6504 			 */
6505 			CTR3(KTR_PMAP, "%s: FAULT_ACCESS_L2 - pmap %#x far %#x",
6506 			    __func__, pmap, far);
6507 			PMAP_UNLOCK(pmap);
6508 			return (KERN_SUCCESS);
6509 		}
6510 	}
6511 	if (idx == FAULT_ACCESS_L1) {
6512 		pte1p = pmap_pte1(pmap, far);
6513 		pte1 = pte1_load(pte1p);
6514 		if (pte1_is_section(pte1)) {
6515 			pte1_store(pte1p, pte1 | PTE1_A);
6516 			PMAP_UNLOCK(pmap);
6517 			return (KERN_SUCCESS);
6518 		} else {
6519 			/*
6520 			 * We got L1 access fault but PTE1 is not section
6521 			 * mapping. Probably some race happened, do nothing.
6522 			 */
6523 			CTR3(KTR_PMAP, "%s: FAULT_ACCESS_L1 - pmap %#x far %#x",
6524 			    __func__, pmap, far);
6525 			PMAP_UNLOCK(pmap);
6526 			return (KERN_SUCCESS);
6527 		}
6528 	}
6529 
6530 	/*
6531 	 * Handle modify bits for page and section. Note that the modify
6532 	 * bit is emulated by software. So PTEx_RO is software read only
6533 	 * bit and PTEx_NM flag is real hardware read only bit.
6534 	 *
6535 	 * QQQ: This is hardware emulation, we do not call userret()
6536 	 *      for aborts from user mode.
6537 	 */
6538 	if ((fsr & FSR_WNR) && (idx == FAULT_PERM_L2)) {
6539 		pte1 = pte1_load(pmap_pte1(pmap, far));
6540 		if (pte1_is_link(pte1)) {
6541 			/* L2 page table should exist and be mapped. */
6542 			pte2p = pt2map_entry(far);
6543 			pte2 = pte2_load(pte2p);
6544 			if (pte2_is_valid(pte2) && !(pte2 & PTE2_RO) &&
6545 			    (pte2 & PTE2_NM)) {
6546 				pte2_store(pte2p, pte2 & ~PTE2_NM);
6547 				tlb_flush(trunc_page(far));
6548 				PMAP_UNLOCK(pmap);
6549 				return (KERN_SUCCESS);
6550 			}
6551 		} else {
6552 			/*
6553 			 * We got L2 permission fault but PTE1 is not a link.
6554 			 * Probably some race happened, do nothing.
6555 			 */
6556 			CTR3(KTR_PMAP, "%s: FAULT_PERM_L2 - pmap %#x far %#x",
6557 			    __func__, pmap, far);
6558 			PMAP_UNLOCK(pmap);
6559 			return (KERN_SUCCESS);
6560 		}
6561 	}
6562 	if ((fsr & FSR_WNR) && (idx == FAULT_PERM_L1)) {
6563 		pte1p = pmap_pte1(pmap, far);
6564 		pte1 = pte1_load(pte1p);
6565 		if (pte1_is_section(pte1)) {
6566 			if (!(pte1 & PTE1_RO) && (pte1 & PTE1_NM)) {
6567 				pte1_store(pte1p, pte1 & ~PTE1_NM);
6568 				tlb_flush(pte1_trunc(far));
6569 				PMAP_UNLOCK(pmap);
6570 				return (KERN_SUCCESS);
6571 			}
6572 		} else {
6573 			/*
6574 			 * We got L1 permission fault but PTE1 is not section
6575 			 * mapping. Probably some race happened, do nothing.
6576 			 */
6577 			CTR3(KTR_PMAP, "%s: FAULT_PERM_L1 - pmap %#x far %#x",
6578 			    __func__, pmap, far);
6579 			PMAP_UNLOCK(pmap);
6580 			return (KERN_SUCCESS);
6581 		}
6582 	}
6583 
6584 	/*
6585 	 * QQQ: The previous code, mainly fast handling of access and
6586 	 *      modify bits aborts, could be moved to ASM. Now we are
6587 	 *      starting to deal with not fast aborts.
6588 	 */
6589 	PMAP_UNLOCK(pmap);
6590 	return (KERN_FAILURE);
6591 }
6592 
6593 #if defined(PMAP_DEBUG)
6594 /*
6595  *  Reusing of KVA used in pmap_zero_page function !!!
6596  */
6597 static void
pmap_zero_page_check(vm_page_t m)6598 pmap_zero_page_check(vm_page_t m)
6599 {
6600 	pt2_entry_t *cmap2_pte2p;
6601 	uint32_t *p, *end;
6602 	struct pcpu *pc;
6603 
6604 	sched_pin();
6605 	pc = get_pcpu();
6606 	cmap2_pte2p = pc->pc_cmap2_pte2p;
6607 	mtx_lock(&pc->pc_cmap_lock);
6608 	if (pte2_load(cmap2_pte2p) != 0)
6609 		panic("%s: CMAP2 busy", __func__);
6610 	pte2_store(cmap2_pte2p, PTE2_KERN_NG(VM_PAGE_TO_PHYS(m), PTE2_AP_KRW,
6611 	    vm_page_pte2_attr(m)));
6612 	end = (uint32_t*)(pc->pc_cmap2_addr + PAGE_SIZE);
6613 	for (p = (uint32_t*)pc->pc_cmap2_addr; p < end; p++)
6614 		if (*p != 0)
6615 			panic("%s: page %p not zero, va: %p", __func__, m,
6616 			    pc->pc_cmap2_addr);
6617 	pte2_clear(cmap2_pte2p);
6618 	tlb_flush((vm_offset_t)pc->pc_cmap2_addr);
6619 	sched_unpin();
6620 	mtx_unlock(&pc->pc_cmap_lock);
6621 }
6622 
6623 int
pmap_pid_dump(int pid)6624 pmap_pid_dump(int pid)
6625 {
6626 	pmap_t pmap;
6627 	struct proc *p;
6628 	int npte2 = 0;
6629 	int i, j, index;
6630 
6631 	sx_slock(&allproc_lock);
6632 	FOREACH_PROC_IN_SYSTEM(p) {
6633 		if (p->p_pid != pid || p->p_vmspace == NULL)
6634 			continue;
6635 		index = 0;
6636 		pmap = vmspace_pmap(p->p_vmspace);
6637 		for (i = 0; i < NPTE1_IN_PT1; i++) {
6638 			pt1_entry_t pte1;
6639 			pt2_entry_t *pte2p, pte2;
6640 			vm_offset_t base, va;
6641 			vm_paddr_t pa;
6642 			vm_page_t m;
6643 
6644 			base = i << PTE1_SHIFT;
6645 			pte1 = pte1_load(&pmap->pm_pt1[i]);
6646 
6647 			if (pte1_is_section(pte1)) {
6648 				/*
6649 				 * QQQ: Do something here!
6650 				 */
6651 			} else if (pte1_is_link(pte1)) {
6652 				for (j = 0; j < NPTE2_IN_PT2; j++) {
6653 					va = base + (j << PAGE_SHIFT);
6654 					if (va >= VM_MIN_KERNEL_ADDRESS) {
6655 						if (index) {
6656 							index = 0;
6657 							printf("\n");
6658 						}
6659 						sx_sunlock(&allproc_lock);
6660 						return (npte2);
6661 					}
6662 					pte2p = pmap_pte2(pmap, va);
6663 					pte2 = pte2_load(pte2p);
6664 					pmap_pte2_release(pte2p);
6665 					if (!pte2_is_valid(pte2))
6666 						continue;
6667 
6668 					pa = pte2_pa(pte2);
6669 					m = PHYS_TO_VM_PAGE(pa);
6670 					printf("va: 0x%x, pa: 0x%x, w: %d, "
6671 					    "f: 0x%x", va, pa,
6672 					    m->ref_count, m->flags);
6673 					npte2++;
6674 					index++;
6675 					if (index >= 2) {
6676 						index = 0;
6677 						printf("\n");
6678 					} else {
6679 						printf(" ");
6680 					}
6681 				}
6682 			}
6683 		}
6684 	}
6685 	sx_sunlock(&allproc_lock);
6686 	return (npte2);
6687 }
6688 
6689 #endif
6690 
6691 #ifdef DDB
6692 static pt2_entry_t *
pmap_pte2_ddb(pmap_t pmap,vm_offset_t va)6693 pmap_pte2_ddb(pmap_t pmap, vm_offset_t va)
6694 {
6695 	pt1_entry_t pte1;
6696 	vm_paddr_t pt2pg_pa;
6697 
6698 	pte1 = pte1_load(pmap_pte1(pmap, va));
6699 	if (!pte1_is_link(pte1))
6700 		return (NULL);
6701 
6702 	if (pmap_is_current(pmap))
6703 		return (pt2map_entry(va));
6704 
6705 	/* Note that L2 page table size is not equal to PAGE_SIZE. */
6706 	pt2pg_pa = trunc_page(pte1_link_pa(pte1));
6707 	if (pte2_pa(pte2_load(PMAP3)) != pt2pg_pa) {
6708 		pte2_store(PMAP3, PTE2_KPT(pt2pg_pa));
6709 #ifdef SMP
6710 		PMAP3cpu = PCPU_GET(cpuid);
6711 #endif
6712 		tlb_flush_local((vm_offset_t)PADDR3);
6713 	}
6714 #ifdef SMP
6715 	else if (PMAP3cpu != PCPU_GET(cpuid)) {
6716 		PMAP3cpu = PCPU_GET(cpuid);
6717 		tlb_flush_local((vm_offset_t)PADDR3);
6718 	}
6719 #endif
6720 	return (PADDR3 + (arm32_btop(va) & (NPTE2_IN_PG - 1)));
6721 }
6722 
6723 static void
dump_pmap(pmap_t pmap)6724 dump_pmap(pmap_t pmap)
6725 {
6726 
6727 	printf("pmap %p\n", pmap);
6728 	printf("  pm_pt1: %p\n", pmap->pm_pt1);
6729 	printf("  pm_pt2tab: %p\n", pmap->pm_pt2tab);
6730 	printf("  pm_active: 0x%08lX\n", pmap->pm_active.__bits[0]);
6731 }
6732 
DB_SHOW_COMMAND(pmaps,pmap_list_pmaps)6733 DB_SHOW_COMMAND(pmaps, pmap_list_pmaps)
6734 {
6735 
6736 	pmap_t pmap;
6737 	LIST_FOREACH(pmap, &allpmaps, pm_list) {
6738 		dump_pmap(pmap);
6739 	}
6740 }
6741 
6742 static int
pte2_class(pt2_entry_t pte2)6743 pte2_class(pt2_entry_t pte2)
6744 {
6745 	int cls;
6746 
6747 	cls = (pte2 >> 2) & 0x03;
6748 	cls |= (pte2 >> 4) & 0x04;
6749 	return (cls);
6750 }
6751 
6752 static void
dump_section(pmap_t pmap,uint32_t pte1_idx)6753 dump_section(pmap_t pmap, uint32_t pte1_idx)
6754 {
6755 }
6756 
6757 static void
dump_link(pmap_t pmap,uint32_t pte1_idx,bool invalid_ok)6758 dump_link(pmap_t pmap, uint32_t pte1_idx, bool invalid_ok)
6759 {
6760 	uint32_t i;
6761 	vm_offset_t va;
6762 	pt2_entry_t *pte2p, pte2;
6763 	vm_page_t m;
6764 
6765 	va = pte1_idx << PTE1_SHIFT;
6766 	pte2p = pmap_pte2_ddb(pmap, va);
6767 	for (i = 0; i < NPTE2_IN_PT2; i++, pte2p++, va += PAGE_SIZE) {
6768 		pte2 = pte2_load(pte2p);
6769 		if (pte2 == 0)
6770 			continue;
6771 		if (!pte2_is_valid(pte2)) {
6772 			printf(" 0x%08X: 0x%08X", va, pte2);
6773 			if (!invalid_ok)
6774 				printf(" - not valid !!!");
6775 			printf("\n");
6776 			continue;
6777 		}
6778 		m = PHYS_TO_VM_PAGE(pte2_pa(pte2));
6779 		printf(" 0x%08X: 0x%08X, TEX%d, s:%d, g:%d, m:%p", va , pte2,
6780 		    pte2_class(pte2), !!(pte2 & PTE2_S), !(pte2 & PTE2_NG), m);
6781 		if (m != NULL) {
6782 			printf(" v:%d w:%d f:0x%04X\n", m->valid,
6783 			    m->ref_count, m->flags);
6784 		} else {
6785 			printf("\n");
6786 		}
6787 	}
6788 }
6789 
6790 static __inline bool
is_pv_chunk_space(vm_offset_t va)6791 is_pv_chunk_space(vm_offset_t va)
6792 {
6793 
6794 	if ((((vm_offset_t)pv_chunkbase) <= va) &&
6795 	    (va < ((vm_offset_t)pv_chunkbase + PAGE_SIZE * pv_maxchunks)))
6796 		return (true);
6797 	return (false);
6798 }
6799 
DB_SHOW_COMMAND(pmap,pmap_pmap_print)6800 DB_SHOW_COMMAND(pmap, pmap_pmap_print)
6801 {
6802 	/* XXX convert args. */
6803 	pmap_t pmap = (pmap_t)addr;
6804 	pt1_entry_t pte1;
6805 	pt2_entry_t pte2;
6806 	vm_offset_t va, eva;
6807 	vm_page_t m;
6808 	uint32_t i;
6809 	bool invalid_ok, dump_link_ok, dump_pv_chunk;
6810 
6811 	if (have_addr) {
6812 		pmap_t pm;
6813 
6814 		LIST_FOREACH(pm, &allpmaps, pm_list)
6815 			if (pm == pmap) break;
6816 		if (pm == NULL) {
6817 			printf("given pmap %p is not in allpmaps list\n", pmap);
6818 			return;
6819 		}
6820 	} else
6821 		pmap = PCPU_GET(curpmap);
6822 
6823 	eva = (modif[0] == 'u') ? VM_MAXUSER_ADDRESS : 0xFFFFFFFF;
6824 	dump_pv_chunk = false; /* XXX evaluate from modif[] */
6825 
6826 	printf("pmap: 0x%08X\n", (uint32_t)pmap);
6827 	printf("PT2MAP: 0x%08X\n", (uint32_t)PT2MAP);
6828 	printf("pt2tab: 0x%08X\n", (uint32_t)pmap->pm_pt2tab);
6829 
6830 	for(i = 0; i < NPTE1_IN_PT1; i++) {
6831 		pte1 = pte1_load(&pmap->pm_pt1[i]);
6832 		if (pte1 == 0)
6833 			continue;
6834 		va = i << PTE1_SHIFT;
6835 		if (va >= eva)
6836 			break;
6837 
6838 		if (pte1_is_section(pte1)) {
6839 			printf("0x%08X: Section 0x%08X, s:%d g:%d\n", va, pte1,
6840 			    !!(pte1 & PTE1_S), !(pte1 & PTE1_NG));
6841 			dump_section(pmap, i);
6842 		} else if (pte1_is_link(pte1)) {
6843 			dump_link_ok = true;
6844 			invalid_ok = false;
6845 			pte2 = pte2_load(pmap_pt2tab_entry(pmap, va));
6846 			m = PHYS_TO_VM_PAGE(pte1_link_pa(pte1));
6847 			printf("0x%08X: Link 0x%08X, pt2tab: 0x%08X m: %p",
6848 			    va, pte1, pte2, m);
6849 			if (is_pv_chunk_space(va)) {
6850 				printf(" - pv_chunk space");
6851 				if (dump_pv_chunk)
6852 					invalid_ok = true;
6853 				else
6854 					dump_link_ok = false;
6855 			}
6856 			else if (m != NULL)
6857 				printf(" w:%d w2:%u", m->ref_count,
6858 				    pt2_wirecount_get(m, pte1_index(va)));
6859 			if (pte2 == 0)
6860 				printf(" !!! pt2tab entry is ZERO");
6861 			else if (pte2_pa(pte1) != pte2_pa(pte2))
6862 				printf(" !!! pt2tab entry is DIFFERENT - m: %p",
6863 				    PHYS_TO_VM_PAGE(pte2_pa(pte2)));
6864 			printf("\n");
6865 			if (dump_link_ok)
6866 				dump_link(pmap, i, invalid_ok);
6867 		} else
6868 			printf("0x%08X: Invalid entry 0x%08X\n", va, pte1);
6869 	}
6870 }
6871 
6872 static void
dump_pt2tab(pmap_t pmap)6873 dump_pt2tab(pmap_t pmap)
6874 {
6875 	uint32_t i;
6876 	pt2_entry_t pte2;
6877 	vm_offset_t va;
6878 	vm_paddr_t pa;
6879 	vm_page_t m;
6880 
6881 	printf("PT2TAB:\n");
6882 	for (i = 0; i < PT2TAB_ENTRIES; i++) {
6883 		pte2 = pte2_load(&pmap->pm_pt2tab[i]);
6884 		if (!pte2_is_valid(pte2))
6885 			continue;
6886 		va = i << PT2TAB_SHIFT;
6887 		pa = pte2_pa(pte2);
6888 		m = PHYS_TO_VM_PAGE(pa);
6889 		printf(" 0x%08X: 0x%08X, TEX%d, s:%d, m:%p", va, pte2,
6890 		    pte2_class(pte2), !!(pte2 & PTE2_S), m);
6891 		if (m != NULL)
6892 			printf(" , w: %d, f: 0x%04X pidx: %lld",
6893 			    m->ref_count, m->flags, m->pindex);
6894 		printf("\n");
6895 	}
6896 }
6897 
DB_SHOW_COMMAND(pmap_pt2tab,pmap_pt2tab_print)6898 DB_SHOW_COMMAND(pmap_pt2tab, pmap_pt2tab_print)
6899 {
6900 	/* XXX convert args. */
6901 	pmap_t pmap = (pmap_t)addr;
6902 	pt1_entry_t pte1;
6903 	pt2_entry_t pte2;
6904 	vm_offset_t va;
6905 	uint32_t i, start;
6906 
6907 	if (have_addr) {
6908 		printf("supported only on current pmap\n");
6909 		return;
6910 	}
6911 
6912 	pmap = PCPU_GET(curpmap);
6913 	printf("curpmap: 0x%08X\n", (uint32_t)pmap);
6914 	printf("PT2MAP: 0x%08X\n", (uint32_t)PT2MAP);
6915 	printf("pt2tab: 0x%08X\n", (uint32_t)pmap->pm_pt2tab);
6916 
6917 	start = pte1_index((vm_offset_t)PT2MAP);
6918 	for (i = start; i < (start + NPT2_IN_PT2TAB); i++) {
6919 		pte1 = pte1_load(&pmap->pm_pt1[i]);
6920 		if (pte1 == 0)
6921 			continue;
6922 		va = i << PTE1_SHIFT;
6923 		if (pte1_is_section(pte1)) {
6924 			printf("0x%08X: Section 0x%08X, s:%d\n", va, pte1,
6925 			    !!(pte1 & PTE1_S));
6926 			dump_section(pmap, i);
6927 		} else if (pte1_is_link(pte1)) {
6928 			pte2 = pte2_load(pmap_pt2tab_entry(pmap, va));
6929 			printf("0x%08X: Link 0x%08X, pt2tab: 0x%08X\n", va,
6930 			    pte1, pte2);
6931 			if (pte2 == 0)
6932 				printf("  !!! pt2tab entry is ZERO\n");
6933 		} else
6934 			printf("0x%08X: Invalid entry 0x%08X\n", va, pte1);
6935 	}
6936 	dump_pt2tab(pmap);
6937 }
6938 #endif
6939