xref: /titanic_41/usr/src/uts/sfmmu/vm/hat_sfmmu.h (revision 5aefb6555731130ca4fd295960123d71f2d21fe8)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*
28  * VM - Hardware Address Translation management.
29  *
30  * This file describes the contents of the sun-reference-mmu(sfmmu)-
31  * specific hat data structures and the sfmmu-specific hat procedures.
32  * The machine-independent interface is described in <vm/hat.h>.
33  */
34 
35 #ifndef	_VM_HAT_SFMMU_H
36 #define	_VM_HAT_SFMMU_H
37 
38 #pragma ident	"%Z%%M%	%I%	%E% SMI"
39 
40 #ifdef	__cplusplus
41 extern "C" {
42 #endif
43 
44 #ifndef _ASM
45 
46 #include <sys/types.h>
47 
48 #endif /* _ASM */
49 
50 #ifdef	_KERNEL
51 
52 #include <sys/pte.h>
53 #include <vm/mach_sfmmu.h>
54 #include <sys/mmu.h>
55 
56 /*
57  * Don't alter these without considering changes to ism_map_t.
58  */
59 #define	DEFAULT_ISM_PAGESIZE		MMU_PAGESIZE4M
60 #define	ISM_PG_SIZE(ism_vbshift)	(1 << ism_vbshift)
61 #define	ISM_SZ_MASK(ism_vbshift)	(ISM_PG_SIZE(ism_vbshift) - 1)
62 #define	ISM_MAP_SLOTS	8	/* Change this carefully. */
63 
64 #ifndef _ASM
65 
66 #include <sys/t_lock.h>
67 #include <vm/hat.h>
68 #include <vm/seg.h>
69 #include <sys/machparam.h>
70 #include <sys/systm.h>
71 #include <sys/x_call.h>
72 #include <vm/page.h>
73 #include <sys/ksynch.h>
74 
75 typedef struct hat sfmmu_t;
76 
77 /*
78  * SFMMU attributes for hat_memload/hat_devload
79  */
80 #define	SFMMU_UNCACHEPTTE	0x01000000	/* unencache in physical $ */
81 #define	SFMMU_UNCACHEVTTE	0x02000000	/* unencache in virtual $ */
82 #define	SFMMU_SIDEFFECT		0x04000000	/* set side effect bit */
83 #define	SFMMU_LOAD_ALLATTR	(HAT_PROT_MASK | HAT_ORDER_MASK |	\
84 		HAT_ENDIAN_MASK | HAT_NOFAULT | HAT_NOSYNC |		\
85 		SFMMU_UNCACHEPTTE | SFMMU_UNCACHEVTTE | SFMMU_SIDEFFECT)
86 
87 
88 /*
89  * sfmmu flags for hat_memload/hat_devload
90  */
91 #define	SFMMU_NO_TSBLOAD	0x08000000	/* do not preload tsb */
92 #define	SFMMU_LOAD_ALLFLAG	(HAT_LOAD | HAT_LOAD_LOCK |		\
93 		HAT_LOAD_ADV | HAT_LOAD_CONTIG | HAT_LOAD_NOCONSIST |	\
94 		HAT_LOAD_SHARE | HAT_LOAD_REMAP | SFMMU_NO_TSBLOAD |	\
95 		HAT_RELOAD_SHARE | HAT_NO_KALLOC | HAT_LOAD_TEXT)
96 
97 /*
98  * sfmmu internal flag to hat_pageunload that spares locked mappings
99  */
100 #define	SFMMU_KERNEL_RELOC	0x8000
101 
102 /*
103  * mode for sfmmu_chgattr
104  */
105 #define	SFMMU_SETATTR	0x0
106 #define	SFMMU_CLRATTR	0x1
107 #define	SFMMU_CHGATTR	0x2
108 
109 /*
110  * sfmmu specific flags for page_t
111  */
112 #define	P_PNC	0x8		/* non-caching is permanent bit */
113 #define	P_TNC	0x10		/* non-caching is temporary bit */
114 #define	P_KPMS	0x20		/* kpm mapped small (vac alias prevention) */
115 #define	P_KPMC	0x40		/* kpm conflict page (vac alias prevention) */
116 
117 #define	PP_GENERIC_ATTR(pp)	((pp)->p_nrm & (P_MOD | P_REF | P_RO))
118 #define	PP_ISMOD(pp)		((pp)->p_nrm & P_MOD)
119 #define	PP_ISREF(pp)		((pp)->p_nrm & P_REF)
120 #define	PP_ISRO(pp)		((pp)->p_nrm & P_RO)
121 #define	PP_ISNC(pp)		((pp)->p_nrm & (P_PNC|P_TNC))
122 #define	PP_ISPNC(pp)		((pp)->p_nrm & P_PNC)
123 #define	PP_ISTNC(pp)		((pp)->p_nrm & P_TNC)
124 #define	PP_ISKPMS(pp)		((pp)->p_nrm & P_KPMS)
125 #define	PP_ISKPMC(pp)		((pp)->p_nrm & P_KPMC)
126 
127 #define	PP_SETMOD(pp)		((pp)->p_nrm |= P_MOD)
128 #define	PP_SETREF(pp)		((pp)->p_nrm |= P_REF)
129 #define	PP_SETREFMOD(pp)	((pp)->p_nrm |= (P_REF|P_MOD))
130 #define	PP_SETRO(pp)		((pp)->p_nrm |= P_RO)
131 #define	PP_SETREFRO(pp)		((pp)->p_nrm |= (P_REF|P_RO))
132 #define	PP_SETPNC(pp)		((pp)->p_nrm |= P_PNC)
133 #define	PP_SETTNC(pp)		((pp)->p_nrm |= P_TNC)
134 #define	PP_SETKPMS(pp)		((pp)->p_nrm |= P_KPMS)
135 #define	PP_SETKPMC(pp)		((pp)->p_nrm |= P_KPMC)
136 
137 #define	PP_CLRMOD(pp)		((pp)->p_nrm &= ~P_MOD)
138 #define	PP_CLRREF(pp)		((pp)->p_nrm &= ~P_REF)
139 #define	PP_CLRREFMOD(pp)	((pp)->p_nrm &= ~(P_REF|P_MOD))
140 #define	PP_CLRRO(pp)		((pp)->p_nrm &= ~P_RO)
141 #define	PP_CLRPNC(pp)		((pp)->p_nrm &= ~P_PNC)
142 #define	PP_CLRTNC(pp)		((pp)->p_nrm &= ~P_TNC)
143 #define	PP_CLRKPMS(pp)		((pp)->p_nrm &= ~P_KPMS)
144 #define	PP_CLRKPMC(pp)		((pp)->p_nrm &= ~P_KPMC)
145 
146 /*
147  * All shared memory segments attached with the SHM_SHARE_MMU flag (ISM)
148  * will be constrained to a 4M, 32M or 256M alignment. Also since every newly-
149  * created ISM segment is created out of a new address space at base va
150  * of 0 we don't need to store it.
151  */
152 #define	ISM_ALIGN(shift)	(1 << shift)	/* base va aligned to <n>M  */
153 #define	ISM_ALIGNED(shift, va)	(((uintptr_t)va & (ISM_ALIGN(shift) - 1)) == 0)
154 #define	ISM_SHIFT(shift, x)	((uintptr_t)x >> (shift))
155 
156 /*
157  * Pad locks out to cache sub-block boundaries to prevent
158  * false sharing, so several processes don't contend for
159  * the same line if they aren't using the same lock.  Since
160  * this is a typedef we also have a bit of freedom in
161  * changing lock implementations later if we decide it
162  * is necessary.
163  */
164 typedef struct hat_lock {
165 	kmutex_t hl_mutex;
166 	uchar_t hl_pad[64 - sizeof (kmutex_t)];
167 } hatlock_t;
168 
169 #define	HATLOCK_MUTEXP(hatlockp)	(&((hatlockp)->hl_mutex))
170 
171 /*
172  * All segments mapped with ISM are guaranteed to be 4M, 32M or 256M aligned.
173  * Also size is guaranteed to be in 4M, 32M or 256M chunks.
174  * ism_seg consists of the following members:
175  * [XX..22] base address of ism segment. XX is 63 or 31 depending whether
176  *	caddr_t is 64 bits or 32 bits.
177  * [21..0] size of segment.
178  *
179  * NOTE: Don't alter this structure without changing defines above and
180  * the tsb_miss and protection handlers.
181  */
182 typedef struct ism_map {
183 	uintptr_t	imap_seg;  	/* base va + sz of ISM segment */
184 	ushort_t	imap_vb_shift;	/* mmu_pageshift for ism page size */
185 	ushort_t	imap_hatflags;	/* primary ism page size */
186 	uint_t		imap_sz_mask;	/* mmu_pagemask for ism page size */
187 	sfmmu_t		*imap_ismhat; 	/* hat id of dummy ISM as */
188 	struct ism_ment	*imap_ment;	/* pointer to mapping list entry */
189 } ism_map_t;
190 
191 #define	ism_start(map)	((caddr_t)((map).imap_seg & \
192 				~ISM_SZ_MASK((map).imap_vb_shift)))
193 #define	ism_size(map)	((map).imap_seg & ISM_SZ_MASK((map).imap_vb_shift))
194 #define	ism_end(map)	((caddr_t)(ism_start(map) + (ism_size(map) * \
195 				ISM_PG_SIZE((map).imap_vb_shift))))
196 /*
197  * ISM mapping entry. Used to link all hat's sharing a ism_hat.
198  * Same function as the p_mapping list for a page.
199  */
200 typedef struct ism_ment {
201 	sfmmu_t		*iment_hat;	/* back pointer to hat_share() hat */
202 	caddr_t		iment_base_va;	/* hat's va base for this ism seg */
203 	struct ism_ment	*iment_next;	/* next ism map entry */
204 	struct ism_ment	*iment_prev;	/* prev ism map entry */
205 } ism_ment_t;
206 
207 /*
208  * ISM segment block. One will be hung off the sfmmu structure if a
209  * a process uses ISM.  More will be linked using ismblk_next if more
210  * than ISM_MAP_SLOTS segments are attached to this proc.
211  *
212  * All modifications to fields in this structure will be protected
213  * by the hat mutex.  In order to avoid grabbing this lock in low level
214  * routines (tsb miss/protection handlers and vatopfn) while not
215  * introducing any race conditions with hat_unshare, we will set
216  * CTX_ISM_BUSY bit in the ctx struct. Any mmu traps that occur
217  * for this ctx while this bit is set will be handled in sfmmu_tsb_excption
218  * where it will synchronize behind the hat mutex.
219  */
220 typedef struct ism_blk {
221 	ism_map_t		iblk_maps[ISM_MAP_SLOTS];
222 	struct ism_blk		*iblk_next;
223 	uint64_t		iblk_nextpa;
224 } ism_blk_t;
225 
226 /*
227  * TSB access information.  All fields are protected by the process's
228  * hat lock.
229  */
230 
231 struct tsb_info {
232 	caddr_t		tsb_va;		/* tsb base virtual address */
233 	uint64_t	tsb_pa;		/* tsb base physical address */
234 	struct tsb_info	*tsb_next;	/* next tsb used by this process */
235 	uint16_t	tsb_szc;	/* tsb size code */
236 	uint16_t	tsb_flags;	/* flags for this tsb; see below */
237 	uint_t		tsb_ttesz_mask;	/* page size masks; see below */
238 
239 	tte_t		tsb_tte;	/* tte to lock into DTLB */
240 	sfmmu_t		*tsb_sfmmu;	/* sfmmu */
241 	kmem_cache_t	*tsb_cache;	/* cache from which mem allocated */
242 	vmem_t		*tsb_vmp;	/* vmem arena from which mem alloc'd */
243 };
244 
245 /*
246  * Values for "tsb_ttesz_mask" bitmask.
247  */
248 #define	TSB8K	(1 << TTE8K)
249 #define	TSB64K  (1 << TTE64K)
250 #define	TSB512K (1 << TTE512K)
251 #define	TSB4M   (1 << TTE4M)
252 #define	TSB32M  (1 << TTE32M)
253 #define	TSB256M (1 << TTE256M)
254 
255 /*
256  * Values for "tsb_flags" field.
257  */
258 #define	TSB_RELOC_FLAG		0x1
259 #define	TSB_FLUSH_NEEDED	0x2
260 #define	TSB_SWAPPED	0x4
261 
262 /*
263  * The platform dependent hat structure.
264  * tte counts should be protected by cas.
265  * cpuset is protected by cas.
266  *
267  * Note that sfmmu_xhat_provider MUST be the first element.
268  */
269 struct hat {
270 	void		*sfmmu_xhat_provider;	/* NULL for CPU hat */
271 	cpuset_t	sfmmu_cpusran;	/* cpu bit mask for efficient xcalls */
272 	struct	as	*sfmmu_as;	/* as this hat provides mapping for */
273 	ulong_t		sfmmu_ttecnt[MMU_PAGE_SIZES]; /* per sz tte counts */
274 	ulong_t		sfmmu_ismttecnt[MMU_PAGE_SIZES]; /* est. ism ttes */
275 	union _h_un {
276 		ism_blk_t	*sfmmu_iblkp;  /* maps to ismhat(s) */
277 		ism_ment_t	*sfmmu_imentp; /* ism hat's mapping list */
278 	} h_un;
279 	uint_t		sfmmu_free:1;	/* hat to be freed - set on as_free */
280 	uint_t		sfmmu_ismhat:1;	/* hat is dummy ism hatid */
281 	uint_t		sfmmu_ctxflushed:1;	/* ctx has been flushed */
282 	uchar_t		sfmmu_rmstat;	/* refmod stats refcnt */
283 	uchar_t		sfmmu_clrstart;	/* start color bin for page coloring */
284 	ushort_t	sfmmu_clrbin;	/* per as phys page coloring bin */
285 	short		sfmmu_cnum;	/* context number */
286 	ushort_t	sfmmu_flags;	/* flags */
287 	struct tsb_info	*sfmmu_tsb;	/* list of per as tsbs */
288 	uint64_t	sfmmu_ismblkpa; /* pa of sfmmu_iblkp, or -1 */
289 	kcondvar_t	sfmmu_tsb_cv;	/* signals TSB swapin or relocation */
290 	uchar_t		sfmmu_cext;	/* context page size encoding */
291 	uint8_t		sfmmu_pgsz[MMU_PAGE_SIZES];  /* ranking for MMU */
292 #ifdef sun4v
293 	struct hv_tsb_block sfmmu_hvblock;
294 #endif
295 };
296 
297 #define	sfmmu_iblk	h_un.sfmmu_iblkp
298 #define	sfmmu_iment	h_un.sfmmu_imentp
299 
300 /*
301  * bit mask for managing vac conflicts on large pages.
302  * bit 1 is for uncache flag.
303  * bits 2 through min(num of cache colors + 1,31) are
304  * for cache colors that have already been flushed.
305  */
306 #define	CACHE_UNCACHE		1
307 #define	CACHE_NUM_COLOR		(shm_alignment >> MMU_PAGESHIFT)
308 
309 #define	CACHE_VCOLOR_MASK(vcolor)	(2 << (vcolor & (CACHE_NUM_COLOR - 1)))
310 
311 #define	CacheColor_IsFlushed(flag, vcolor) \
312 					((flag) & CACHE_VCOLOR_MASK(vcolor))
313 
314 #define	CacheColor_SetFlushed(flag, vcolor) \
315 					((flag) |= CACHE_VCOLOR_MASK(vcolor))
316 /*
317  * Flags passed to sfmmu_page_cache to flush page from vac or not.
318  */
319 #define	CACHE_FLUSH	0
320 #define	CACHE_NO_FLUSH	1
321 
322 /*
323  * Flags passed to sfmmu_tlbcache_demap
324  */
325 #define	FLUSH_NECESSARY_CPUS	0
326 #define	FLUSH_ALL_CPUS		1
327 
328 /*
329  * Software context structure.  The size of this structure is currently
330  * hardwired into the tsb miss handlers in assembly code through the
331  * CTX_SZ_SHIFT define.  Since this define is used in a shift we should keep
332  * this structure a power of two.
333  *
334  * ctx_flags:
335  * Bit 0 : Free flag.
336  */
337 struct ctx {
338 	union _ctx_un {
339 		sfmmu_t *ctx_sfmmup;	/* back pointer to hat id */
340 		struct ctx *ctx_freep;	/* next ctx in freelist */
341 	} ctx_un;
342 	krwlock_t	ctx_rwlock;	/* protect context from stealer */
343 	uint32_t	ctx_flags;	/* flags */
344 	uint8_t		pad[12];
345 };
346 
347 #define	ctx_sfmmu	ctx_un.ctx_sfmmup
348 #define	ctx_free	ctx_un.ctx_freep
349 
350 #ifdef	DEBUG
351 /*
352  * For debugging purpose only. Maybe removed later.
353  */
354 struct ctx_trace {
355 	sfmmu_t		*sc_sfmmu_stolen;
356 	sfmmu_t		*sc_sfmmu_stealing;
357 	clock_t		sc_time;
358 	ushort_t	sc_type;
359 	ushort_t	sc_cnum;
360 };
361 #define	CTX_TRC_STEAL	0x1
362 #define	CTX_TRC_FREE	0x0
363 #define	TRSIZE	0x400
364 #define	NEXT_CTXTR(ptr)	(((ptr) >= ctx_trace_last) ? \
365 		ctx_trace_first : ((ptr) + 1))
366 #define	TRACE_CTXS(mutex, ptr, cnum, stolen_sfmmu, stealing_sfmmu, type) \
367 	mutex_enter(mutex);						\
368 	(ptr)->sc_sfmmu_stolen = (stolen_sfmmu);			\
369 	(ptr)->sc_sfmmu_stealing = (stealing_sfmmu);			\
370 	(ptr)->sc_cnum = (cnum);					\
371 	(ptr)->sc_type = (type);					\
372 	(ptr)->sc_time = lbolt;						\
373 	(ptr) = NEXT_CTXTR(ptr);					\
374 	num_ctx_stolen += (type);					\
375 	mutex_exit(mutex);
376 #else
377 
378 #define	TRACE_CTXS(mutex, ptr, cnum, stolen_sfmmu, stealing_sfmmu, type)
379 
380 #endif	/* DEBUG */
381 
382 #endif	/* !_ASM */
383 
384 /*
385  * Macros for sfmmup->sfmmu_flags access.  The macros that change the flags
386  * ASSERT() that we're holding the HAT lock before changing the flags;
387  * however callers that read the flags may do so without acquiring the lock
388  * in a fast path, and then recheck the flag after acquiring the lock in
389  * a slow path.
390  */
391 #define	SFMMU_FLAGS_ISSET(sfmmup, flags) \
392 	(((sfmmup)->sfmmu_flags & (flags)) == (flags))
393 
394 #define	SFMMU_FLAGS_CLEAR(sfmmup, flags) \
395 	(ASSERT(sfmmu_hat_lock_held((sfmmup))), \
396 	(sfmmup)->sfmmu_flags &= ~(flags))
397 
398 #define	SFMMU_FLAGS_SET(sfmmup, flags) \
399 	(ASSERT(sfmmu_hat_lock_held((sfmmup))), \
400 	(sfmmup)->sfmmu_flags |= (flags))
401 
402 /*
403  * sfmmu HAT flags
404  */
405 #define	HAT_64K_FLAG	0x01
406 #define	HAT_512K_FLAG	0x02
407 #define	HAT_4M_FLAG	0x04
408 #define	HAT_32M_FLAG	0x08
409 #define	HAT_256M_FLAG	0x10
410 #define	HAT_4MTEXT_FLAG	0x80
411 #define	HAT_SWAPPED	0x100	/* swapped out */
412 #define	HAT_SWAPIN	0x200	/* swapping in */
413 #define	HAT_BUSY	0x400	/* replacing TSB(s) */
414 #define	HAT_ISMBUSY	0x800	/* adding/removing/traversing ISM maps */
415 
416 #define	HAT_LGPG_FLAGS						\
417 	(HAT_64K_FLAG | HAT_512K_FLAG | HAT_4M_FLAG |		\
418 	    HAT_32M_FLAG | HAT_256M_FLAG)
419 
420 #define	HAT_FLAGS_MASK						\
421 	(HAT_LGPG_FLAGS | HAT_4MTEXT_FLAG | HAT_SWAPPED |	\
422 	    HAT_SWAPIN | HAT_BUSY | HAT_ISMBUSY)
423 
424 /*
425  * Context flags
426  */
427 #define	CTX_FREE_FLAG		0x1
428 #define	CTX_FLAGS_MASK		0x1
429 
430 #define	CTX_SET_FLAGS(ctx, flag)					\
431 {									\
432 	uint32_t old, new;						\
433 									\
434 	do {								\
435 		new = old = (ctx)->ctx_flags;				\
436 		new &= CTX_FLAGS_MASK;					\
437 		new |= flag;						\
438 		new = cas32(&(ctx)->ctx_flags, old, new);		\
439 	} while (new != old);						\
440 }
441 
442 #define	CTX_CLEAR_FLAGS(ctx, flag)					\
443 {									\
444 	uint32_t old, new;						\
445 									\
446 	do {								\
447 		new = old = (ctx)->ctx_flags;				\
448 		new &= CTX_FLAGS_MASK & ~(flag);			\
449 		new = cas32(&(ctx)->ctx_flags, old, new);		\
450 	} while (new != old);						\
451 }
452 
453 #define	ctxtoctxnum(ctx)	((ushort_t)((ctx) - ctxs))
454 
455 /*
456  * Defines needed for ctx stealing.
457  */
458 #define	GET_CTX_RETRY_CNT	100
459 
460 /*
461  * Starting with context 0, the first NUM_LOCKED_CTXS contexts
462  * are locked so that sfmmu_getctx can't steal any of these
463  * contexts.  At the time this software was being developed, the
464  * only context that needs to be locked is context 0 (the kernel
465  * context), and context 1 (reserved for stolen context). So this constant
466  * was originally defined to be 2.
467  */
468 #define	NUM_LOCKED_CTXS 2
469 #define	INVALID_CONTEXT	1
470 
471 #ifndef	_ASM
472 
473 /*
474  * Kernel page relocation stuff.
475  */
476 struct sfmmu_callback {
477 	int key;
478 	int (*prehandler)(caddr_t, uint_t, uint_t, void *);
479 	int (*posthandler)(caddr_t, uint_t, uint_t, void *, pfn_t);
480 	int (*errhandler)(caddr_t, uint_t, uint_t, void *);
481 	int capture_cpus;
482 };
483 
484 extern int sfmmu_max_cb_id;
485 extern struct sfmmu_callback *sfmmu_cb_table;
486 
487 extern int hat_kpr_enabled;
488 
489 struct pa_hment;
490 
491 /*
492  * RFE: With multihat gone we gain back an int.  We could use this to
493  * keep ref bits on a per cpu basis to eliminate xcalls.
494  */
495 struct sf_hment {
496 	tte_t hme_tte;			/* tte for this hment */
497 
498 	union {
499 		struct page *page;	/* what page this maps */
500 		struct pa_hment *data;	/* pa_hment */
501 	} sf_hment_un;
502 
503 	struct	sf_hment *hme_next;	/* next hment */
504 	struct	sf_hment *hme_prev;	/* prev hment */
505 };
506 
507 struct pa_hment {
508 	caddr_t		addr;		/* va */
509 	uint_t		len;		/* bytes */
510 	ushort_t	flags;		/* internal flags */
511 	ushort_t	refcnt;		/* reference count */
512 	id_t		cb_id;		/* callback id, table index */
513 	void		*pvt;		/* handler's private data */
514 	struct sf_hment	sfment;		/* corresponding dummy sf_hment */
515 };
516 
517 #define	hme_page		sf_hment_un.page
518 #define	hme_data		sf_hment_un.data
519 #define	hme_size(sfhmep)	((int)(TTE_CSZ(&(sfhmep)->hme_tte)))
520 #define	PAHME_SZ		(sizeof (struct pa_hment))
521 #define	SFHME_SZ		(sizeof (struct sf_hment))
522 
523 #define	IS_PAHME(hme)	((hme)->hme_tte.ll == 0)
524 
525 /*
526  * hmeblk_tag structure
527  * structure used to obtain a match on a hme_blk.  Currently consists of
528  * the address of the sfmmu struct (or hatid), the base page address of the
529  * hme_blk, and the rehash count.  The rehash count is actually only 2 bits
530  * and has the following meaning:
531  * 1 = 8k or 64k hash sequence.
532  * 2 = 512k hash sequence.
533  * 3 = 4M hash sequence.
534  * We require this count because we don't want to get a false hit on a 512K or
535  * 4M rehash with a base address corresponding to a 8k or 64k hmeblk.
536  * Note:  The ordering and size of the hmeblk_tag members are implictly known
537  * by the tsb miss handlers written in assembly.  Do not change this structure
538  * without checking those routines.  See HTAG_SFMMUPSZ define.
539  */
540 
541 typedef union {
542 	struct {
543 		uint64_t	hblk_basepg: 51, /* hme_blk base pg # */
544 				hblk_rehash: 13; /* rehash number */
545 		sfmmu_t		*sfmmup;
546 	} hblk_tag_un;
547 	uint64_t		htag_tag[2];
548 } hmeblk_tag;
549 
550 #define	htag_id		hblk_tag_un.sfmmup
551 #define	htag_bspage	hblk_tag_un.hblk_basepg
552 #define	htag_rehash	hblk_tag_un.hblk_rehash
553 
554 #define	HTAGS_EQ(tag1, tag2)	(((tag1.htag_tag[0] ^ tag2.htag_tag[0]) | \
555 				(tag1.htag_tag[1] ^ tag2.htag_tag[1])) == 0)
556 #define	HME_REHASH(sfmmup)						\
557 	((sfmmup)->sfmmu_ttecnt[TTE512K] != 0 ||			\
558 	(sfmmup)->sfmmu_ttecnt[TTE4M] != 0 ||				\
559 	(sfmmup)->sfmmu_ttecnt[TTE32M] != 0 ||				\
560 	(sfmmup)->sfmmu_ttecnt[TTE256M] != 0)
561 
562 #endif /* !_ASM */
563 
564 #define	NHMENTS		8		/* # of hments in an 8k hme_blk */
565 					/* needs to be multiple of 2 */
566 #ifndef	_ASM
567 
568 #ifdef	HBLK_TRACE
569 
570 #define	HBLK_LOCK		1
571 #define	HBLK_UNLOCK		0
572 #define	HBLK_STACK_DEPTH	6
573 #define	HBLK_AUDIT_CACHE_SIZE	16
574 #define	HBLK_LOCK_PATTERN	0xaaaaaaaa
575 #define	HBLK_UNLOCK_PATTERN	0xbbbbbbbb
576 
577 struct hblk_lockcnt_audit {
578 	int		flag;		/* lock or unlock */
579 	kthread_id_t	thread;
580 	int		depth;
581 	pc_t		stack[HBLK_STACK_DEPTH];
582 };
583 
584 #endif	/* HBLK_TRACE */
585 
586 
587 /*
588  * Hment block structure.
589  * The hme_blk is the node data structure which the hash structure
590  * mantains. An hme_blk can have 2 different sizes depending on the
591  * number of hments it implicitly contains.  When dealing with 64K, 512K,
592  * or 4M hments there is one hment per hme_blk.  When dealing with
593  * 8k hments we allocate an hme_blk plus an additional 7 hments to
594  * give us a total of 8 (NHMENTS) hments that can be referenced through a
595  * hme_blk.
596  *
597  * The hmeblk structure contains 2 tte reference counters used to determine if
598  * it is ok to free up the hmeblk.  Both counters have to be zero in order
599  * to be able to free up hmeblk.  They are protected by cas.
600  * hblk_hmecnt is the number of hments present on pp mapping lists.
601  * hblk_vcnt reflects number of valid ttes in hmeblk.
602  *
603  * The hmeblk now also has per tte lock cnts.  This is required because
604  * the counts can be high and there are not enough bits in the tte. When
605  * physio is fixed to not lock the translations we should be able to move
606  * the lock cnt back to the tte.  See bug id 1198554.
607  *
608  * Note that xhat_hme_blk's layout follows this structure: hme_blk_misc
609  * and sf_hment are at the same offsets in both structures. Whenever
610  * hme_blk is changed, xhat_hme_blk may need to be updated as well.
611  */
612 
613 struct hme_blk_misc {
614 	ushort_t locked_cnt;	/* HAT_LOAD_LOCK ref cnt */
615 	uint_t	notused:10;
616 	uint_t	xhat_bit:1;	/* set for an xhat hme_blk */
617 	uint_t	shadow_bit:1;	/* set for a shadow hme_blk */
618 	uint_t	nucleus_bit:1;	/* set for a nucleus hme_blk */
619 	uint_t	ttesize:3;	/* contains ttesz of hmeblk */
620 };
621 
622 struct hme_blk {
623 	uint64_t	hblk_nextpa;	/* physical address for hash list */
624 
625 	hmeblk_tag	hblk_tag;	/* tag used to obtain an hmeblk match */
626 
627 	struct hme_blk	*hblk_next;	/* on free list or on hash list */
628 					/* protected by hash lock */
629 
630 	struct hme_blk	*hblk_shadow;	/* pts to shadow hblk */
631 					/* protected by hash lock */
632 	uint_t		hblk_span;	/* span of memory hmeblk maps */
633 
634 	struct hme_blk_misc	hblk_misc;
635 
636 	union {
637 		struct {
638 			ushort_t hblk_hmecount;	/* hment on mlists counter */
639 			ushort_t hblk_validcnt;	/* valid tte reference count */
640 		} hblk_counts;
641 		uint_t		hblk_shadow_mask;
642 	} hblk_un;
643 
644 #ifdef	HBLK_TRACE
645 	kmutex_t	hblk_audit_lock;	/* lock to protect index */
646 	uint_t		hblk_audit_index;	/* index into audit_cache */
647 	struct	hblk_lockcnt_audit hblk_audit_cache[HBLK_AUDIT_CACHE_SIZE];
648 #endif	/* HBLK_AUDIT */
649 
650 	struct sf_hment hblk_hme[1];	/* hment array */
651 };
652 
653 #define	hblk_lckcnt	hblk_misc.locked_cnt
654 #define	hblk_xhat_bit   hblk_misc.xhat_bit
655 #define	hblk_shw_bit	hblk_misc.shadow_bit
656 #define	hblk_nuc_bit	hblk_misc.nucleus_bit
657 #define	hblk_ttesz	hblk_misc.ttesize
658 #define	hblk_hmecnt	hblk_un.hblk_counts.hblk_hmecount
659 #define	hblk_vcnt	hblk_un.hblk_counts.hblk_validcnt
660 #define	hblk_shw_mask	hblk_un.hblk_shadow_mask
661 
662 #define	MAX_HBLK_LCKCNT	0xFFFF
663 #define	HMEBLK_ALIGN	0x8		/* hmeblk has to be double aligned */
664 
665 #ifdef	HBLK_TRACE
666 
667 #define	HBLK_STACK_TRACE(hmeblkp, lock)					\
668 {									\
669 	int flag = lock;	/* to pacify lint */			\
670 	int audit_index;						\
671 									\
672 	mutex_enter(&hmeblkp->hblk_audit_lock);				\
673 	audit_index = hmeblkp->hblk_audit_index;			\
674 	hmeblkp->hblk_audit_index = ((hmeblkp->hblk_audit_index + 1) &	\
675 	    (HBLK_AUDIT_CACHE_SIZE - 1));				\
676 	mutex_exit(&hmeblkp->hblk_audit_lock);				\
677 									\
678 	if (flag)							\
679 		hmeblkp->hblk_audit_cache[audit_index].flag =		\
680 		    HBLK_LOCK_PATTERN;					\
681 	else								\
682 		hmeblkp->hblk_audit_cache[audit_index].flag =		\
683 		    HBLK_UNLOCK_PATTERN;				\
684 									\
685 	hmeblkp->hblk_audit_cache[audit_index].thread = curthread;	\
686 	hmeblkp->hblk_audit_cache[audit_index].depth =			\
687 	    getpcstack(hmeblkp->hblk_audit_cache[audit_index].stack,	\
688 	    HBLK_STACK_DEPTH);						\
689 }
690 
691 #else
692 
693 #define	HBLK_STACK_TRACE(hmeblkp, lock)
694 
695 #endif	/* HBLK_TRACE */
696 
697 #define	HMEHASH_FACTOR	16	/* used to calc # of buckets in hme hash */
698 
699 /*
700  * A maximum number of user hmeblks is defined in order to place an upper
701  * limit on how much nucleus memory is required and to avoid overflowing the
702  * tsbmiss uhashsz and khashsz data areas. The number below corresponds to
703  * the number of buckets required, for an average hash chain length of 4 on
704  * a 16TB machine.
705  */
706 
707 #define	MAX_UHME_BUCKETS	(0x1 << 30)
708 #define	MAX_KHME_BUCKETS	(0x1 << 30)
709 
710 /*
711  * The minimum number of kernel hash buckets.
712  */
713 #define	MIN_KHME_BUCKETS	0x800
714 
715 /*
716  * The number of hash buckets must be a power of 2. If the initial calculated
717  * value is less than USER_BUCKETS_THRESHOLD we round up to the next greater
718  * power of 2, otherwise we round down to avoid huge over allocations.
719  */
720 #define	USER_BUCKETS_THRESHOLD	(1<<22)
721 
722 #define	MAX_NUCUHME_BUCKETS	0x4000
723 #define	MAX_NUCKHME_BUCKETS	0x2000
724 
725 /*
726  * There are 2 locks in the hmehash bucket.  The hmehash_mutex is
727  * a regular mutex used to make sure operations on a hash link are only
728  * done by one thread.  Any operation which comes into the hat with
729  * a <vaddr, as> will grab the hmehash_mutex.  Normally one would expect
730  * the tsb miss handlers to grab the hash lock to make sure the hash list
731  * is consistent while we traverse it.  Unfortunately this can lead to
732  * deadlocks or recursive mutex enters since it is possible for
733  * someone holding the lock to take a tlb/tsb miss.
734  * To solve this problem we have added the hmehash_listlock.  This lock
735  * is only grabbed by the tsb miss handlers, vatopfn, and while
736  * adding/removing a hmeblk from the hash list. The code is written to
737  * guarantee we won't take a tlb miss while holding this lock.
738  */
739 struct hmehash_bucket {
740 	kmutex_t	hmehash_mutex;
741 	uint64_t	hmeh_nextpa;	/* physical address for hash list */
742 	struct hme_blk *hmeblkp;
743 	uint_t		hmeh_listlock;
744 };
745 
746 #endif /* !_ASM */
747 
748 
749 /*
750  * The tsb miss handlers written in assembly know that sfmmup
751  * is a 64 bit ptr.
752  *
753  * The bspage and re-hash part is 64 bits, with the sfmmup being another 64
754  * bits.
755  */
756 #define	HTAG_SFMMUPSZ		0	/* Not really used for LP64 */
757 #define	HTAG_REHASHSZ		13
758 
759 /*
760  * Assembly routines need to be able to get to ttesz
761  */
762 #define	HBLK_SZMASK		0x7
763 
764 #ifndef _ASM
765 
766 /*
767  * Returns the number of bytes that an hmeblk spans given its tte size
768  */
769 #define	get_hblk_span(hmeblkp) ((hmeblkp)->hblk_span)
770 #define	get_hblk_ttesz(hmeblkp)	((hmeblkp)->hblk_ttesz)
771 #define	get_hblk_cache(hmeblkp)	(((hmeblkp)->hblk_ttesz == TTE8K) ? \
772 	sfmmu8_cache : sfmmu1_cache)
773 #define	HMEBLK_SPAN(ttesz)						\
774 	((ttesz == TTE8K)? (TTEBYTES(ttesz) * NHMENTS) : TTEBYTES(ttesz))
775 
776 #define	set_hblk_sz(hmeblkp, ttesz)				\
777 	(hmeblkp)->hblk_ttesz = (ttesz);			\
778 	(hmeblkp)->hblk_span = HMEBLK_SPAN(ttesz)
779 
780 #define	get_hblk_base(hmeblkp)					\
781 	((uintptr_t)(hmeblkp)->hblk_tag.htag_bspage << MMU_PAGESHIFT)
782 
783 #define	get_hblk_endaddr(hmeblkp)				\
784 	((caddr_t)(get_hblk_base(hmeblkp) + get_hblk_span(hmeblkp)))
785 
786 #define	in_hblk_range(hmeblkp, vaddr)					\
787 	(((uintptr_t)(vaddr) >= get_hblk_base(hmeblkp)) &&		\
788 	((uintptr_t)(vaddr) < (get_hblk_base(hmeblkp) +			\
789 	get_hblk_span(hmeblkp))))
790 
791 #define	tte_to_vaddr(hmeblkp, tte)	((caddr_t)(get_hblk_base(hmeblkp) \
792 	+ (TTEBYTES(TTE_CSZ(&tte)) * (tte).tte_hmenum)))
793 
794 #define	vaddr_to_vshift(hblktag, vaddr, shwsz)				\
795 	((((uintptr_t)(vaddr) >> MMU_PAGESHIFT) - (hblktag.htag_bspage)) >>\
796 	TTE_BSZS_SHIFT((shwsz) - 1))
797 
798 #define	HME8BLK_SZ	(sizeof (struct hme_blk) + \
799 			(NHMENTS - 1) * sizeof (struct sf_hment))
800 #define	HME1BLK_SZ	(sizeof (struct hme_blk))
801 #define	H8TOH1		(MMU_PAGESIZE4M / MMU_PAGESIZE)
802 #define	H1MIN		(2 + MAX_BIGKTSB_TTES)	/* nucleus text+data, ktsb */
803 
804 /*
805  * Hme_blk hash structure
806  * Active mappings are kept in a hash structure of hme_blks.  The hash
807  * function is based on (ctx, vaddr) The size of the hash table size is a
808  * power of 2 such that the average hash chain lenth is HMENT_HASHAVELEN.
809  * The hash actually consists of 2 separate hashes.  One hash is for the user
810  * address space and the other hash is for the kernel address space.
811  * The number of buckets are calculated at boot time and stored in the global
812  * variables "uhmehash_num" and "khmehash_num".  By making the hash table size
813  * a power of 2 we can use a simply & function to derive an index instead of
814  * a divide.
815  *
816  * HME_HASH_FUNCTION(hatid, vaddr, shift) returns a pointer to a hme_hash
817  * bucket.
818  * An hme hash bucket contains a pointer to an hme_blk and the mutex that
819  * protects the link list.
820  * Spitfire supports 4 page sizes.  8k and 64K pages only need one hash.
821  * 512K pages need 2 hashes and 4M pages need 3 hashes.
822  * The 'shift' parameter controls how many bits the vaddr will be shifted in
823  * the hash function. It is calculated in the HME_HASH_SHIFT(ttesz) function
824  * and it varies depending on the page size as follows:
825  *	8k pages:  	HBLK_RANGE_SHIFT
826  *	64k pages:	MMU_PAGESHIFT64K
827  *	512K pages:	MMU_PAGESHIFT512K
828  *	4M pages:	MMU_PAGESHIFT4M
829  * An assembly version of the hash function exists in sfmmu_ktsb_miss(). All
830  * changes should be reflected in both versions.  This function and the TSB
831  * miss handlers are the only places which know about the two hashes.
832  *
833  * HBLK_RANGE_SHIFT controls range of virtual addresses that will fall
834  * into the same bucket for a particular process.  It is currently set to
835  * be equivalent to 64K range or one hme_blk.
836  *
837  * The hme_blks in the hash are protected by a per hash bucket mutex
838  * known as SFMMU_HASH_LOCK.
839  * You need to acquire this lock before traversing the hash bucket link
840  * list, while adding/removing a hme_blk to the list, and while
841  * modifying an hme_blk.  A possible optimization is to replace these
842  * mutexes by readers/writer lock but right now it is not clear whether
843  * this is a win or not.
844  *
845  * The HME_HASH_TABLE_SEARCH will search the hash table for the
846  * hme_blk that contains the hment that corresponds to the passed
847  * ctx and vaddr.  It assumed the SFMMU_HASH_LOCK is held.
848  */
849 
850 #endif /* ! _ASM */
851 
852 #define	KHATID			ksfmmup
853 #define	UHMEHASH_SZ		uhmehash_num
854 #define	KHMEHASH_SZ		khmehash_num
855 #define	HMENT_HASHAVELEN	4
856 #define	HBLK_RANGE_SHIFT	MMU_PAGESHIFT64K /* shift for HBLK_BS_MASK */
857 #define	MAX_HASHCNT		5
858 #define	DEFAULT_MAX_HASHCNT	3
859 
860 #ifndef _ASM
861 
862 #define	HASHADDR_MASK(hashno)	TTE_PAGEMASK(hashno)
863 
864 #define	HME_HASH_SHIFT(ttesz)						\
865 	((ttesz == TTE8K)? HBLK_RANGE_SHIFT : TTE_PAGE_SHIFT(ttesz))	\
866 
867 #define	HME_HASH_ADDR(vaddr, hmeshift)					\
868 	((caddr_t)(((uintptr_t)(vaddr) >> (hmeshift)) << (hmeshift)))
869 
870 #define	HME_HASH_BSPAGE(vaddr, hmeshift)				\
871 	(((uintptr_t)(vaddr) >> (hmeshift)) << ((hmeshift) - MMU_PAGESHIFT))
872 
873 #define	HME_HASH_REHASH(ttesz)						\
874 	(((ttesz) < TTE512K)? 1 : (ttesz))
875 
876 #define	HME_HASH_FUNCTION(hatid, vaddr, shift)				\
877 	((hatid != KHATID)?						\
878 	(&uhme_hash[ (((uintptr_t)(hatid) ^ ((uintptr_t)vaddr >> (shift))) & \
879 	    UHMEHASH_SZ) ]):					\
880 	(&khme_hash[ (((uintptr_t)(hatid) ^ ((uintptr_t)vaddr >> (shift))) & \
881 	    KHMEHASH_SZ) ]))
882 
883 /*
884  * This macro will traverse a hmeblk hash link list looking for an hme_blk
885  * that owns the specified vaddr and hatid.  If if doesn't find one , hmeblkp
886  * will be set to NULL, otherwise it will point to the correct hme_blk.
887  * This macro also cleans empty hblks.
888  */
889 #define	HME_HASH_SEARCH_PREV(hmebp, hblktag, hblkp, hblkpa,		\
890 	pr_hblk, prevpa, listp)						\
891 {									\
892 	struct hme_blk *nx_hblk;					\
893 	uint64_t 	nx_pa;						\
894 									\
895 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));				\
896 	hblkp = hmebp->hmeblkp;						\
897 	hblkpa = hmebp->hmeh_nextpa;					\
898 	prevpa = 0;							\
899 	pr_hblk = NULL;							\
900 	while (hblkp) {							\
901 		if (HTAGS_EQ(hblkp->hblk_tag, hblktag)) {		\
902 			/* found hme_blk */				\
903 			break;						\
904 		}							\
905 		nx_hblk = hblkp->hblk_next;				\
906 		nx_pa = hblkp->hblk_nextpa;				\
907 		if (!hblkp->hblk_vcnt && !hblkp->hblk_hmecnt) {		\
908 			sfmmu_hblk_hash_rm(hmebp, hblkp, prevpa, pr_hblk); \
909 			sfmmu_hblk_free(hmebp, hblkp, hblkpa, listp);   \
910 		} else {						\
911 			pr_hblk = hblkp;				\
912 			prevpa = hblkpa;				\
913 		}							\
914 		hblkp = nx_hblk;					\
915 		hblkpa = nx_pa;						\
916 	}								\
917 }
918 
919 #define	HME_HASH_SEARCH(hmebp, hblktag, hblkp, listp)			\
920 {									\
921 	struct hme_blk *pr_hblk;					\
922 	uint64_t hblkpa, prevpa;					\
923 									\
924 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hblkp, hblkpa, pr_hblk,	\
925 		prevpa, listp);						\
926 }
927 
928 /*
929  * This macro will traverse a hmeblk hash link list looking for an hme_blk
930  * that owns the specified vaddr and hatid.  If if doesn't find one , hmeblkp
931  * will be set to NULL, otherwise it will point to the correct hme_blk.
932  * It doesn't remove empty hblks.
933  */
934 #define	HME_HASH_FAST_SEARCH(hmebp, hblktag, hblkp)			\
935 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));				\
936 	for (hblkp = hmebp->hmeblkp; hblkp;				\
937 	    hblkp = hblkp->hblk_next) {					\
938 		if (HTAGS_EQ(hblkp->hblk_tag, hblktag)) {		\
939 			/* found hme_blk */				\
940 			break;						\
941 		}							\
942 	}								\
943 
944 
945 #define	SFMMU_HASH_LOCK(hmebp)						\
946 		(mutex_enter(&hmebp->hmehash_mutex))
947 
948 #define	SFMMU_HASH_UNLOCK(hmebp)					\
949 		(mutex_exit(&hmebp->hmehash_mutex))
950 
951 #define	SFMMU_HASH_LOCK_TRYENTER(hmebp)					\
952 		(mutex_tryenter(&hmebp->hmehash_mutex))
953 
954 #define	SFMMU_HASH_LOCK_ISHELD(hmebp)					\
955 		(mutex_owned(&hmebp->hmehash_mutex))
956 
957 #define	SFMMU_XCALL_STATS(ctxnum)					\
958 {									\
959 	if (ctxnum == KCONTEXT) {					\
960 		SFMMU_STAT(sf_kernel_xcalls);				\
961 	} else {							\
962 		SFMMU_STAT(sf_user_xcalls);				\
963 	}								\
964 }
965 
966 #define	astosfmmu(as)		((as)->a_hat)
967 #define	sfmmutoctxnum(sfmmup)	((sfmmup)->sfmmu_cnum)
968 #define	sfmmutoctx(sfmmup)	(&ctxs[sfmmutoctxnum(sfmmup)])
969 #define	hblktosfmmu(hmeblkp)	((sfmmu_t *)(hmeblkp)->hblk_tag.htag_id)
970 #define	sfmmutoas(sfmmup)	((sfmmup)->sfmmu_as)
971 #define	ctxnumtoctx(ctxnum)	(&ctxs[ctxnum])
972 /*
973  * We use the sfmmu data structure to keep the per as page coloring info.
974  */
975 #define	as_color_bin(as)	(astosfmmu(as)->sfmmu_clrbin)
976 #define	as_color_start(as)	(astosfmmu(as)->sfmmu_clrstart)
977 
978 typedef struct {
979 	char	h8[HME8BLK_SZ];
980 } hblk8_t;
981 
982 typedef struct {
983 	char	h1[HME1BLK_SZ];
984 } hblk1_t;
985 
986 typedef struct {
987 	ulong_t  	index;
988 	ulong_t  	len;
989 	hblk8_t		*list;
990 } nucleus_hblk8_info_t;
991 
992 typedef struct {
993 	ulong_t		index;
994 	ulong_t		len;
995 	hblk1_t		*list;
996 } nucleus_hblk1_info_t;
997 
998 /*
999  * This struct is used for accumlating information about a range
1000  * of pages that are unloading so that a single xcall can flush
1001  * the entire range from remote tlbs. A function that must demap
1002  * a range of virtual addresses declares one of these structures
1003  * and initializes using DEMP_RANGE_INIT(). It then passes a pointer to this
1004  * struct to the appropriate sfmmu_hblk_* level function which does
1005  * all the bookkeeping using the other macros. When the function has
1006  * finished the virtual address range, it needs to call DEMAP_RANGE_FLUSH()
1007  * macro to take care of any remaining unflushed mappings.
1008  *
1009  * The maximum range this struct can represent is the number of bits
1010  * in the dmr_bitvec field times the pagesize in dmr_pgsz. Currently, only
1011  * MMU_PAGESIZE pages are supported.
1012  *
1013  * Since there are now cases where it's no longer necessary to do
1014  * flushes (e.g. when the process isn't runnable because it's swapping
1015  * out or exiting) we allow these macros to take a NULL dmr input and do
1016  * nothing in that case.
1017  */
1018 typedef struct {
1019 	sfmmu_t		*dmr_sfmmup;	/* relevent hat */
1020 	caddr_t		dmr_addr;	/* beginning address */
1021 	caddr_t		dmr_endaddr;	/* ending  address */
1022 	ulong_t		dmr_bitvec;	/* valid pages found */
1023 	ulong_t		dmr_bit;	/* next page to examine */
1024 	ulong_t		dmr_maxbit;	/* highest page in range */
1025 	ulong_t		dmr_pgsz;	/* page size in range */
1026 } demap_range_t;
1027 
1028 #define	DMR_MAXBIT ((ulong_t)1<<63) /* dmr_bit high bit */
1029 
1030 #define	DEMAP_RANGE_INIT(sfmmup, dmrp) \
1031 	if ((dmrp) != NULL) { \
1032 	(dmrp)->dmr_sfmmup = (sfmmup); \
1033 	(dmrp)->dmr_bitvec = 0; \
1034 	(dmrp)->dmr_maxbit = sfmmu_dmr_maxbit; \
1035 	(dmrp)->dmr_pgsz = MMU_PAGESIZE; \
1036 	}
1037 
1038 #define	DEMAP_RANGE_PGSZ(dmrp) ((dmrp)? (dmrp)->dmr_pgsz : MMU_PAGESIZE)
1039 
1040 #define	DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr) \
1041 	if ((dmrp) != NULL) { \
1042 	if ((dmrp)->dmr_bitvec != 0 && (dmrp)->dmr_endaddr != (addr)) \
1043 		sfmmu_tlb_range_demap(dmrp); \
1044 	(dmrp)->dmr_endaddr = (endaddr); \
1045 	}
1046 
1047 #define	DEMAP_RANGE_FLUSH(dmrp) \
1048 	if ((dmrp) != NULL) { \
1049 		if ((dmrp)->dmr_bitvec != 0) \
1050 			sfmmu_tlb_range_demap(dmrp); \
1051 	}
1052 
1053 #define	DEMAP_RANGE_MARKPG(dmrp, addr) \
1054 	if ((dmrp) != NULL) { \
1055 		if ((dmrp)->dmr_bitvec == 0) { \
1056 			(dmrp)->dmr_addr = (addr); \
1057 			(dmrp)->dmr_bit = 1; \
1058 		} \
1059 		(dmrp)->dmr_bitvec |= (dmrp)->dmr_bit; \
1060 	}
1061 
1062 #define	DEMAP_RANGE_NEXTPG(dmrp) \
1063 	if ((dmrp) != NULL && (dmrp)->dmr_bitvec != 0) { \
1064 		if ((dmrp)->dmr_bit & (dmrp)->dmr_maxbit) { \
1065 			sfmmu_tlb_range_demap(dmrp); \
1066 		} else { \
1067 			(dmrp)->dmr_bit <<= 1; \
1068 		} \
1069 	}
1070 
1071 /*
1072  * TSB related structures
1073  *
1074  * The TSB is made up of tte entries.  Both the tag and data are present
1075  * in the TSB.  The TSB locking is managed as follows:
1076  * A software bit in the tsb tag is used to indicate that entry is locked.
1077  * If a cpu servicing a tsb miss reads a locked entry the tag compare will
1078  * fail forcing the cpu to go to the hat hash for the translation.
1079  * The cpu who holds the lock can then modify the data side, and the tag side.
1080  * The last write should be to the word containing the lock bit which will
1081  * clear the lock and allow the tsb entry to be read.  It is assumed that all
1082  * cpus reading the tsb will do so with atomic 128-bit loads.  An atomic 128
1083  * bit load is required to prevent the following from happening:
1084  *
1085  * cpu 0			cpu 1			comments
1086  *
1087  * ldx tag						tag unlocked
1088  *				ldstub lock		set lock
1089  *				stx data
1090  *				stx tag			unlock
1091  * ldx tag						incorrect tte!!!
1092  *
1093  * The software also maintains a bit in the tag to indicate an invalid
1094  * tsb entry.  The purpose of this bit is to allow the tsb invalidate code
1095  * to invalidate a tsb entry with a single cas.  See code for details.
1096  */
1097 
1098 union tsb_tag {
1099 	struct {
1100 		uint32_t	tag_res0:16;	/* reserved - context area */
1101 		uint32_t	tag_inv:1;	/* sw - invalid tsb entry */
1102 		uint32_t	tag_lock:1;	/* sw - locked tsb entry */
1103 		uint32_t	tag_res1:4;	/* reserved */
1104 		uint32_t	tag_va_hi:10;	/* va[63:54] */
1105 		uint32_t	tag_va_lo;	/* va[53:22] */
1106 	} tagbits;
1107 	struct tsb_tagints {
1108 		uint32_t	inthi;
1109 		uint32_t	intlo;
1110 	} tagints;
1111 };
1112 #define	tag_invalid		tagbits.tag_inv
1113 #define	tag_locked		tagbits.tag_lock
1114 #define	tag_vahi		tagbits.tag_va_hi
1115 #define	tag_valo		tagbits.tag_va_lo
1116 #define	tag_inthi		tagints.inthi
1117 #define	tag_intlo		tagints.intlo
1118 
1119 struct tsbe {
1120 	union tsb_tag	tte_tag;
1121 	tte_t		tte_data;
1122 };
1123 
1124 /*
1125  * A per cpu struct is kept that duplicates some info
1126  * used by the tl>0 tsb miss handlers plus it provides
1127  * a scratch area.  Its purpose is to minimize cache misses
1128  * in the tsb miss handler and is 128 bytes (2 e$ lines).
1129  *
1130  * There should be one allocated per cpu in nucleus memory
1131  * and should be aligned on an ecache line boundary.
1132  */
1133 struct tsbmiss {
1134 	sfmmu_t			*ksfmmup;	/* kernel hat id */
1135 	sfmmu_t			*usfmmup;	/* user hat id */
1136 	struct tsbe		*tsbptr;	/* hardware computed ptr */
1137 	struct tsbe		*tsbptr4m;	/* hardware computed ptr */
1138 	uint64_t		ismblkpa;
1139 	struct hmehash_bucket	*khashstart;
1140 	struct hmehash_bucket	*uhashstart;
1141 	uint_t			khashsz;
1142 	uint_t			uhashsz;
1143 	uint16_t 		dcache_line_mask; /* used to flush dcache */
1144 	uint16_t		hat_flags;
1145 	uint32_t		itlb_misses;
1146 	uint32_t		dtlb_misses;
1147 	uint32_t		utsb_misses;
1148 	uint32_t		ktsb_misses;
1149 	uint16_t		uprot_traps;
1150 	uint16_t		kprot_traps;
1151 
1152 	/*
1153 	 * scratch[0] -> TSB_TAGACC
1154 	 * scratch[1] -> TSBMISS_HMEBP
1155 	 * scratch[2] -> TSBMISS_HATID
1156 	 */
1157 	uintptr_t		scratch[3];
1158 	uint8_t			pad[0x10];
1159 };
1160 
1161 /*
1162  * A per cpu struct is kept for the use within the tl>0 kpm tsb
1163  * miss handler. Some members are duplicates of common data or
1164  * the physical addresses of common data. A few members are also
1165  * written by the tl>0 kpm tsb miss handler. Its purpose is to
1166  * minimize cache misses in the kpm tsb miss handler and occupies
1167  * one ecache line. There should be one allocated per cpu in
1168  * nucleus memory and it should be aligned on an ecache line
1169  * boundary. It is not merged w/ struct tsbmiss since there is
1170  * not much to share and the tsbmiss pathes are different, so
1171  * a kpm tlbmiss/tsbmiss only touches one cacheline, except for
1172  * (DEBUG || SFMMU_STAT_GATHER) where the dtlb_misses counter
1173  * of struct tsbmiss is used on every dtlb miss.
1174  */
1175 struct kpmtsbm {
1176 	caddr_t		vbase;		/* start of address kpm range */
1177 	caddr_t		vend;		/* end of address kpm range */
1178 	uchar_t		flags;		/* flags needed in TL tsbmiss handler */
1179 	uchar_t		sz_shift;	/* for single kpm window */
1180 	uchar_t		kpmp_shift;	/* hash lock shift */
1181 	uchar_t		kpmp2pshft;	/* kpm page to page shift */
1182 	uint_t		kpmp_table_sz;	/* size of kpmp_table or kpmp_stable */
1183 	uint64_t	kpmp_tablepa;	/* paddr of kpmp_table or kpmp_stable */
1184 	uint64_t	msegphashpa;	/* paddr of memseg_phash */
1185 	struct tsbe	*tsbptr;	/* saved ktsb pointer */
1186 	uint_t		kpm_dtlb_misses; /* kpm tlbmiss counter */
1187 	uint_t		kpm_tsb_misses;	/* kpm tsbmiss counter */
1188 	uintptr_t	pad[1];
1189 };
1190 
1191 extern uint_t  tsb_slab_size;
1192 extern uint_t  tsb_slab_shift;
1193 extern uint_t  tsb_slab_ttesz;
1194 extern uint_t  tsb_slab_pamask;
1195 
1196 #endif /* !_ASM */
1197 
1198 /*
1199  * Flags for TL kpm tsbmiss handler
1200  */
1201 #define	KPMTSBM_ENABLE_FLAG	0x01	/* bit copy of kpm_enable */
1202 #define	KPMTSBM_TLTSBM_FLAG	0x02	/* use TL tsbmiss handler */
1203 #define	KPMTSBM_TSBPHYS_FLAG	0x04	/* use ASI_MEM for TSB update */
1204 
1205 /*
1206  * The TSB
1207  * All TSB sizes supported by the hardware are now supported (8K - 1M).
1208  * For kernel TSBs we may go beyond the hardware supported sizes and support
1209  * larger TSBs via software.
1210  * All TTE sizes are supported in the TSB; the manner in which this is
1211  * done is cpu dependent.
1212  */
1213 #define	TSB_MIN_SZCODE		TSB_8K_SZCODE	/* min. supported TSB size */
1214 #define	TSB_MIN_OFFSET_MASK	(TSB_OFFSET_MASK(TSB_MIN_SZCODE))
1215 
1216 #define	UTSB_MAX_SZCODE		TSB_1M_SZCODE /* max. supported TSB size */
1217 #define	UTSB_MAX_OFFSET_MASK	(TSB_OFFSET_MASK(UTSB_MAX_SZCODE))
1218 
1219 #define	TSB_FREEMEM_MIN		0x1000		/* 32 mb */
1220 #define	TSB_FREEMEM_LARGE	0x10000		/* 512 mb */
1221 #define	TSB_8K_SZCODE		0		/* 512 entries */
1222 #define	TSB_16K_SZCODE		1		/* 1k entries */
1223 #define	TSB_32K_SZCODE		2		/* 2k entries */
1224 #define	TSB_64K_SZCODE		3		/* 4k entries */
1225 #define	TSB_128K_SZCODE		4		/* 8k entries */
1226 #define	TSB_256K_SZCODE		5		/* 16k entries */
1227 #define	TSB_512K_SZCODE		6		/* 32k entries */
1228 #define	TSB_1M_SZCODE		7		/* 64k entries */
1229 #define	TSB_2M_SZCODE		8		/* 128k entries */
1230 #define	TSB_4M_SZCODE		9		/* 256k entries */
1231 #define	TSB_ENTRY_SHIFT		4	/* each entry = 128 bits = 16 bytes */
1232 #define	TSB_ENTRY_SIZE		(1 << 4)
1233 #define	TSB_START_SIZE		9
1234 #define	TSB_ENTRIES(tsbsz)	(1 << (TSB_START_SIZE + tsbsz))
1235 #define	TSB_BYTES(tsbsz)	(TSB_ENTRIES(tsbsz) << TSB_ENTRY_SHIFT)
1236 #define	TSB_OFFSET_MASK(tsbsz)	(TSB_ENTRIES(tsbsz) - 1)
1237 #define	TSB_BASEADDR_MASK	((1 << 12) - 1)
1238 
1239 /*
1240  * sun4u platforms
1241  * ---------------
1242  * We now support two user TSBs with one TSB base register.
1243  * Hence the TSB base register is split up as follows:
1244  *
1245  * When only one TSB present:
1246  *   [63  62..42  41..13  12..4  3..0]
1247  *     ^   ^       ^       ^     ^
1248  *     |   |       |       |     |
1249  *     |   |       |       |     |_ TSB size code
1250  *     |   |       |       |
1251  *     |   |       |       |_ Reserved 0
1252  *     |   |       |
1253  *     |   |       |_ TSB VA[41..13]
1254  *     |   |
1255  *     |   |_ VA hole (Spitfire), zeros (Cheetah and beyond)
1256  *     |
1257  *     |_ 0
1258  *
1259  * When second TSB present:
1260  *   [63  62..42  41..33  32..29  28..22  21..13  12..4  3..0]
1261  *     ^   ^       ^       ^       ^       ^       ^     ^
1262  *     |   |       |       |       |       |       |     |
1263  *     |   |       |       |       |       |       |     |_ First TSB size code
1264  *     |   |       |       |       |       |       |
1265  *     |   |       |       |       |       |       |_ Reserved 0
1266  *     |   |       |       |       |       |
1267  *     |   |       |       |       |       |_ First TSB's VA[21..13]
1268  *     |   |       |       |       |
1269  *     |   |       |       |       |_ Reserved for future use
1270  *     |   |       |       |
1271  *     |   |       |       |_ Second TSB's size code
1272  *     |   |       |
1273  *     |   |       |_ Second TSB's VA[21..13]
1274  *     |   |
1275  *     |   |_ VA hole (Spitfire) / ones (Cheetah and beyond)
1276  *     |
1277  *     |_ 1
1278  *
1279  * Note that since we store 21..13 of each TSB's VA, TSBs and their slabs
1280  * may be up to 4M in size.  For now, only hardware supported TSB sizes
1281  * are supported, though the slabs are usually 4M in size.
1282  *
1283  * sun4v platforms
1284  * ---------------
1285  * On sun4v platforms, we use two dedicated scratchpad registers as pseudo
1286  * hardware TSB base registers to hold up to two different user TSBs.
1287  *
1288  * Each register contains TSB's physical base and size code information
1289  * as follows:
1290  *
1291  *   [63..56  55..13  12..4  3..0]
1292  *      ^       ^       ^     ^
1293  *      |       |       |     |
1294  *      |       |       |     |_ TSB size code
1295  *      |       |       |
1296  *      |       |       |_ Reserved 0
1297  *      |       |
1298  *      |       |_ TSB PA[55..13]
1299  *      |
1300  *      |
1301  *      |
1302  *      |_ 0 for valid TSB
1303  *
1304  * Absence of a user TSB (primarily the second user TSB) is indicated by
1305  * storing a negative value in the TSB base register. This allows us to
1306  * check for presence of a user TSB by simply checking bit# 63.
1307  */
1308 #define	TSBREG_MSB_SHIFT	32		/* set upper bits */
1309 #define	TSBREG_MSB_CONST	0xfffff800	/* set bits 63..43 */
1310 #define	TSBREG_FIRTSB_SHIFT	42		/* to clear bits 63:22 */
1311 #define	TSBREG_SECTSB_MKSHIFT	20		/* 21:13 --> 41:33 */
1312 #define	TSBREG_SECTSB_LSHIFT	22		/* to clear bits 63:42 */
1313 #define	TSBREG_SECTSB_RSHIFT	(TSBREG_SECTSB_MKSHIFT + TSBREG_SECTSB_LSHIFT)
1314 						/* sectsb va -> bits 21:13 */
1315 						/* after clearing upper bits */
1316 #define	TSBREG_SECSZ_SHIFT	29		/* to get sectsb szc to 3:0 */
1317 #define	TSBREG_VAMASK_SHIFT	13		/* set up VA mask */
1318 
1319 #define	BIGKTSB_SZ_MASK		0xf
1320 #define	TSB_SOFTSZ_MASK		BIGKTSB_SZ_MASK
1321 #define	MIN_BIGKTSB_SZCODE	9	/* 256k entries */
1322 #define	MAX_BIGKTSB_SZCODE	11	/* 1024k entries */
1323 #define	MAX_BIGKTSB_TTES	(TSB_BYTES(MAX_BIGKTSB_SZCODE) / MMU_PAGESIZE4M)
1324 
1325 #define	TAG_VALO_SHIFT		22		/* tag's va are bits 63-22 */
1326 /*
1327  * sw bits used on tsb_tag - bit masks used only in assembly
1328  * use only a sethi for these fields.
1329  */
1330 #define	TSBTAG_INVALID	0x00008000		/* tsb_tag.tag_invalid */
1331 #define	TSBTAG_LOCKED	0x00004000		/* tsb_tag.tag_locked */
1332 
1333 #ifdef	_ASM
1334 
1335 /*
1336  * Marker to indicate that this instruction will be hot patched at runtime
1337  * to some other value.
1338  * This value must be zero since it fills in the imm bits of the target
1339  * instructions to be patched
1340  */
1341 #define	RUNTIME_PATCH	(0)
1342 
1343 /*
1344  * V9 defines nop instruction as the following, which we use
1345  * at runtime to nullify some instructions we don't want to
1346  * execute in the trap handlers on certain platforms.
1347  */
1348 #define	MAKE_NOP_INSTR(reg)	\
1349 	sethi	%hi(0x1000000), reg
1350 
1351 /*
1352  * Macro to get this CPU's tsbmiss area.
1353  */
1354 #define	CPU_TSBMISS_AREA(tsbmiss, tmp1)					\
1355 	CPU_INDEX(tmp1, tsbmiss);		/* tmp1 = cpu idx */	\
1356 	sethi	%hi(tsbmiss_area), tsbmiss;	/* tsbmiss base ptr */	\
1357 	sllx    tmp1, TSBMISS_SHIFT, tmp1;	/* byte offset */	\
1358 	or	tsbmiss, %lo(tsbmiss_area), tsbmiss;			\
1359 	add	tsbmiss, tmp1, tsbmiss		/* tsbmiss area of CPU */
1360 
1361 
1362 /*
1363  * Macro to set kernel context + page size codes in DMMU primary context
1364  * register. It is only necessary for sun4u because sun4v does not need
1365  * page size codes
1366  */
1367 #ifdef sun4v
1368 
1369 #define	SET_KCONTEXTREG(reg0, reg1, reg2, reg3, reg4, label1, label2, label3)
1370 
1371 #else
1372 
1373 #define	SET_KCONTEXTREG(reg0, reg1, reg2, reg3, reg4, label1, label2, label3) \
1374 	sethi	%hi(kcontextreg), reg0;					\
1375 	ldx	[reg0 + %lo(kcontextreg)], reg0;			\
1376 	mov	MMU_PCONTEXT, reg1;					\
1377 	ldxa	[reg1]ASI_MMU_CTX, reg2;				\
1378 	xor	reg0, reg2, reg2;					\
1379 	brz	reg2, label3;						\
1380 	srlx	reg2, CTXREG_NEXT_SHIFT, reg2;				\
1381 	rdpr	%pstate, reg3;		/* disable interrupts */	\
1382 	btst	PSTATE_IE, reg3;					\
1383 /*CSTYLED*/								\
1384 	bnz,a,pt %icc, label1;						\
1385 	wrpr	reg3, PSTATE_IE, %pstate;				\
1386 /*CSTYLED*/								\
1387 label1:;								\
1388 	brz	reg2, label2;	   /* need demap if N_pgsz0/1 change */	\
1389 	sethi	%hi(FLUSH_ADDR), reg4;					\
1390 	mov	DEMAP_ALL_TYPE, reg2;					\
1391 	stxa	%g0, [reg2]ASI_DTLB_DEMAP;				\
1392 	stxa	%g0, [reg2]ASI_ITLB_DEMAP;				\
1393 /*CSTYLED*/								\
1394 label2:;								\
1395 	stxa	reg0, [reg1]ASI_MMU_CTX;				\
1396 	flush	reg4;							\
1397 	btst	PSTATE_IE, reg3;					\
1398 /*CSTYLED*/								\
1399 	bnz,a,pt %icc, label3;						\
1400 	wrpr	%g0, reg3, %pstate;	/* restore interrupt state */	\
1401 label3:;
1402 
1403 #endif
1404 
1405 #endif	/* _ASM */
1406 
1407 #ifndef _ASM
1408 
1409 /*
1410  * Page coloring
1411  * The p_vcolor field of the page struct (1 byte) is used to store the
1412  * virtual page color.  This provides for 255 colors.  The value zero is
1413  * used to mean the page has no color - never been mapped or somehow
1414  * purified.
1415  */
1416 
1417 #define	PP_GET_VCOLOR(pp)	(((pp)->p_vcolor) - 1)
1418 #define	PP_NEWPAGE(pp)		(!(pp)->p_vcolor)
1419 #define	PP_SET_VCOLOR(pp, color)                                          \
1420 	((pp)->p_vcolor = ((color) + 1))
1421 
1422 /*
1423  * As mentioned p_vcolor == 0 means there is no color for this page.
1424  * But PP_SET_VCOLOR(pp, color) expects 'color' to be real color minus
1425  * one so we define this constant.
1426  */
1427 #define	NO_VCOLOR	(-1)
1428 
1429 #define	addr_to_vcolor(addr) \
1430 	(((uint_t)(uintptr_t)(addr) >> MMU_PAGESHIFT) & vac_colors_mask)
1431 
1432 /*
1433  * The field p_index in the psm page structure is for large pages support.
1434  * P_index is a bit-vector of the different mapping sizes that a given page
1435  * is part of. An hme structure for a large mapping is only added in the
1436  * group leader page (first page). All pages covered by a given large mapping
1437  * have the corrosponding mapping bit set in their p_index field. This allows
1438  * us to only store an explicit hme structure in the leading page which
1439  * simplifies the mapping link list management. Furthermore, it provides us
1440  * a fast mechanism for determining the largest mapping a page is part of. For
1441  * exmaple, a page with a 64K and a 4M mappings has a p_index value of 0x0A.
1442  *
1443  * Implementation note: even though the first bit in p_index is reserved
1444  * for 8K mappings, it is NOT USED by the code and SHOULD NOT be set.
1445  * In addition, the upper four bits of the p_index field are used by the
1446  * code as temporaries
1447  */
1448 
1449 /*
1450  * Defines for psm page struct fields and large page support
1451  */
1452 #define	SFMMU_INDEX_SHIFT		6
1453 #define	SFMMU_INDEX_MASK		((1 << SFMMU_INDEX_SHIFT) - 1)
1454 
1455 /* Return the mapping index */
1456 #define	PP_MAPINDEX(pp)	((pp)->p_index & SFMMU_INDEX_MASK)
1457 
1458 /*
1459  * These macros rely on the following property:
1460  * All pages constituting a large page are covered by a virtually
1461  * contiguous set of page_t's.
1462  */
1463 
1464 /* Return the leader for this mapping size */
1465 #define	PP_GROUPLEADER(pp, sz) \
1466 	(&(pp)[-(int)(pp->p_pagenum & (TTEPAGES(sz)-1))])
1467 
1468 /* Return the root page for this page based on p_szc */
1469 #define	PP_PAGEROOT(pp)	((pp)->p_szc == 0 ? (pp) : \
1470 	PP_GROUPLEADER((pp), (pp)->p_szc))
1471 
1472 #define	PP_PAGENEXT_N(pp, n)	((pp) + (n))
1473 #define	PP_PAGENEXT(pp)		PP_PAGENEXT_N((pp), 1)
1474 
1475 #define	PP_PAGEPREV_N(pp, n)	((pp) - (n))
1476 #define	PP_PAGEPREV(pp)		PP_PAGEPREV_N((pp), 1)
1477 
1478 #define	PP_ISMAPPED_LARGE(pp)	(PP_MAPINDEX(pp) != 0)
1479 
1480 /* Need function to test the page mappping which takes p_index into account */
1481 #define	PP_ISMAPPED(pp)	((pp)->p_mapping || PP_ISMAPPED_LARGE(pp))
1482 
1483 /*
1484  * Don't call this macro with sz equal to zero. 8K mappings SHOULD NOT
1485  * set p_index field.
1486  */
1487 #define	PAGESZ_TO_INDEX(sz)	(1 << (sz))
1488 
1489 
1490 /*
1491  * prototypes for hat assembly routines.  Some of these are
1492  * known to machine dependent VM code.
1493  */
1494 extern uint64_t sfmmu_make_tsbtag(caddr_t);
1495 extern struct tsbe *
1496 		sfmmu_get_tsbe(uint64_t, caddr_t, int, int);
1497 extern void	sfmmu_load_tsbe(struct tsbe *, uint64_t, tte_t *, int);
1498 extern void	sfmmu_unload_tsbe(struct tsbe *, uint64_t, int);
1499 extern void	sfmmu_load_mmustate(sfmmu_t *);
1500 extern void	sfmmu_ctx_steal_tl1(uint64_t, uint64_t);
1501 extern void	sfmmu_raise_tsb_exception(uint64_t, uint64_t);
1502 #ifndef sun4v
1503 extern void	sfmmu_itlb_ld(caddr_t, int, tte_t *);
1504 extern void	sfmmu_dtlb_ld(caddr_t, int, tte_t *);
1505 #endif /* sun4v */
1506 extern void	sfmmu_copytte(tte_t *, tte_t *);
1507 extern int	sfmmu_modifytte(tte_t *, tte_t *, tte_t *);
1508 extern int	sfmmu_modifytte_try(tte_t *, tte_t *, tte_t *);
1509 extern pfn_t	sfmmu_ttetopfn(tte_t *, caddr_t);
1510 extern void	sfmmu_hblk_hash_rm(struct hmehash_bucket *,
1511 			struct hme_blk *, uint64_t, struct hme_blk *);
1512 extern void	sfmmu_hblk_hash_add(struct hmehash_bucket *, struct hme_blk *,
1513 			uint64_t);
1514 
1515 /*
1516  * functions exported to machine dependent VM code
1517  */
1518 extern void	sfmmu_patch_ktsb(void);
1519 #ifndef sun4v
1520 extern void	sfmmu_patch_utsb(void);
1521 #endif /* sun4v */
1522 extern pfn_t	sfmmu_vatopfn(caddr_t, sfmmu_t *, tte_t *);
1523 extern void	sfmmu_vatopfn_suspended(caddr_t, sfmmu_t *, tte_t *);
1524 #ifdef	DEBUG
1525 extern void	sfmmu_check_kpfn(pfn_t);
1526 #else
1527 #define		sfmmu_check_kpfn(pfn)	/* disabled */
1528 #endif	/* DEBUG */
1529 extern void	sfmmu_memtte(tte_t *, pfn_t, uint_t, int);
1530 extern void	sfmmu_tteload(struct hat *, tte_t *, caddr_t, page_t *,	uint_t);
1531 extern void	sfmmu_tsbmiss_exception(struct regs *, uintptr_t, uint_t);
1532 extern void	sfmmu_init_tsbs(void);
1533 extern caddr_t  sfmmu_ktsb_alloc(caddr_t);
1534 extern int	sfmmu_getctx_pri(void);
1535 extern int	sfmmu_getctx_sec(void);
1536 extern void	sfmmu_setctx_sec(int);
1537 extern void	sfmmu_inv_tsb(caddr_t, uint_t);
1538 extern void	sfmmu_init_ktsbinfo(void);
1539 extern int	sfmmu_setup_4lp(void);
1540 extern void	sfmmu_patch_mmu_asi(int);
1541 extern void	sfmmu_init_nucleus_hblks(caddr_t, size_t, int, int);
1542 extern void	sfmmu_cache_flushall(void);
1543 extern pgcnt_t  sfmmu_tte_cnt(sfmmu_t *, uint_t);
1544 extern void	*sfmmu_tsb_segkmem_alloc(vmem_t *, size_t, int);
1545 extern void	sfmmu_tsb_segkmem_free(vmem_t *, void *, size_t);
1546 extern void	sfmmu_steal_context(sfmmu_t *, uint8_t *);
1547 
1548 extern void	hat_kern_setup(void);
1549 extern int	hat_page_relocate(page_t **, page_t **, spgcnt_t *);
1550 extern uint_t	hat_preferred_pgsz(struct hat *, caddr_t, size_t, int);
1551 extern int	sfmmu_get_ppvcolor(struct page *);
1552 extern int	sfmmu_get_addrvcolor(caddr_t);
1553 extern int	sfmmu_hat_lock_held(sfmmu_t *);
1554 
1555 /*
1556  * Functions exported to xhat_sfmmu.c
1557  */
1558 extern kmutex_t *sfmmu_mlist_enter(page_t *);
1559 extern void	sfmmu_mlist_exit(kmutex_t *);
1560 extern int	sfmmu_mlist_held(struct page *);
1561 extern struct hme_blk *sfmmu_hmetohblk(struct sf_hment *);
1562 
1563 /*
1564  * MMU-specific functions optionally imported from the CPU module
1565  */
1566 #pragma weak mmu_large_pages_disabled
1567 #pragma weak mmu_set_ctx_page_sizes
1568 #pragma weak mmu_preferred_pgsz
1569 #pragma weak mmu_check_page_sizes
1570 
1571 extern int mmu_large_pages_disabled(uint_t);
1572 extern void mmu_set_ctx_page_sizes(sfmmu_t *);
1573 extern uint_t mmu_preferred_pgsz(sfmmu_t *, caddr_t, size_t);
1574 extern void mmu_check_page_sizes(sfmmu_t *, uint64_t *);
1575 
1576 extern sfmmu_t 		*ksfmmup;
1577 extern struct ctx	*ctxs;
1578 extern uint_t		nctxs;
1579 extern caddr_t		ktsb_base;
1580 extern uint64_t		ktsb_pbase;
1581 extern int		ktsb_sz;
1582 extern int		ktsb_szcode;
1583 extern caddr_t		ktsb4m_base;
1584 extern uint64_t		ktsb4m_pbase;
1585 extern int		ktsb4m_sz;
1586 extern int		ktsb4m_szcode;
1587 extern uint64_t		kpm_tsbbase;
1588 extern int		kpm_tsbsz;
1589 extern int		ktsb_phys;
1590 extern int		enable_bigktsb;
1591 #ifndef sun4v
1592 extern int		utsb_dtlb_ttenum;
1593 extern int		utsb4m_dtlb_ttenum;
1594 #endif /* sun4v */
1595 extern int		uhmehash_num;
1596 extern int		khmehash_num;
1597 extern struct hmehash_bucket *uhme_hash;
1598 extern struct hmehash_bucket *khme_hash;
1599 extern kmutex_t		*mml_table;
1600 extern uint_t		mml_table_sz;
1601 extern uint_t		mml_shift;
1602 extern uint_t		hblk_alloc_dynamic;
1603 extern struct tsbmiss	tsbmiss_area[NCPU];
1604 extern struct kpmtsbm	kpmtsbm_area[NCPU];
1605 extern int		tsb_max_growsize;
1606 #ifndef sun4v
1607 extern int		dtlb_resv_ttenum;
1608 extern caddr_t		utsb_vabase;
1609 extern caddr_t		utsb4m_vabase;
1610 #endif /* sun4v */
1611 extern vmem_t		*kmem_tsb_default_arena[];
1612 extern int		tsb_lgrp_affinity;
1613 
1614 /* kpm externals */
1615 extern pfn_t		sfmmu_kpm_vatopfn(caddr_t);
1616 extern void		sfmmu_kpm_patch_tlbm(void);
1617 extern void		sfmmu_kpm_patch_tsbm(void);
1618 extern void		sfmmu_kpm_load_tsb(caddr_t, tte_t *, int);
1619 extern void		sfmmu_kpm_unload_tsb(caddr_t, int);
1620 extern void		sfmmu_kpm_tsbmtl(short *, uint_t *, int);
1621 extern int		sfmmu_kpm_stsbmtl(char *, uint_t *, int);
1622 extern caddr_t		kpm_vbase;
1623 extern size_t		kpm_size;
1624 extern struct memseg	*memseg_hash[];
1625 extern uint64_t		memseg_phash[];
1626 extern kpm_hlk_t	*kpmp_table;
1627 extern kpm_shlk_t	*kpmp_stable;
1628 extern uint_t		kpmp_table_sz;
1629 extern uint_t		kpmp_stable_sz;
1630 extern uchar_t		kpmp_shift;
1631 
1632 #define	PP_ISMAPPED_KPM(pp)	((pp)->p_kpmref > 0)
1633 
1634 #define	IS_KPM_ALIAS_RANGE(vaddr)					\
1635 	(((vaddr) - kpm_vbase) >> (uintptr_t)kpm_size_shift > 0)
1636 
1637 #endif /* !_ASM */
1638 
1639 /* sfmmu_kpm_tsbmtl flags */
1640 #define	KPMTSBM_STOP		0
1641 #define	KPMTSBM_START		1
1642 
1643 /* kpm_smallpages kp_mapped values */
1644 #define	KPM_MAPPEDS		-1	/* small mapping valid, no conflict */
1645 #define	KPM_MAPPEDSC		1	/* small mapping valid, conflict */
1646 
1647 /* Physical memseg address NULL marker */
1648 #define	MSEG_NULLPTR_PA		-1
1649 
1650 /*
1651  * Memseg hash defines for kpm trap level tsbmiss handler.
1652  * Must be in sync w/ page.h .
1653  */
1654 #define	SFMMU_MEM_HASH_SHIFT		0x9
1655 #define	SFMMU_N_MEM_SLOTS		0x200
1656 #define	SFMMU_MEM_HASH_ENTRY_SHIFT	3
1657 
1658 #ifndef	_ASM
1659 #if (SFMMU_MEM_HASH_SHIFT != MEM_HASH_SHIFT)
1660 #error SFMMU_MEM_HASH_SHIFT != MEM_HASH_SHIFT
1661 #endif
1662 #if (SFMMU_N_MEM_SLOTS != N_MEM_SLOTS)
1663 #error SFMMU_N_MEM_SLOTS != N_MEM_SLOTS
1664 #endif
1665 
1666 /* Physical memseg address NULL marker */
1667 #define	SFMMU_MEMSEG_NULLPTR_PA		-1
1668 
1669 /*
1670  * Check KCONTEXT to be zero, asm parts depend on that assumption.
1671  */
1672 #if (KCONTEXT != 0)
1673 #error KCONTEXT != 0
1674 #endif
1675 #endif	/* !_ASM */
1676 
1677 
1678 #endif /* _KERNEL */
1679 
1680 #ifndef _ASM
1681 /*
1682  * ctx, hmeblk, mlistlock and other stats for sfmmu
1683  */
1684 struct sfmmu_global_stat {
1685 	int		sf_tsb_exceptions;	/* # of tsb exceptions */
1686 	int		sf_tsb_raise_exception;	/* # tsb exc. w/o TLB flush */
1687 
1688 	int		sf_pagefaults;		/* # of pagefaults */
1689 
1690 	int		sf_uhash_searches;	/* # of user hash searches */
1691 	int		sf_uhash_links;		/* # of user hash links */
1692 	int		sf_khash_searches;	/* # of kernel hash searches */
1693 	int		sf_khash_links;		/* # of kernel hash links */
1694 
1695 	int		sf_swapout;		/* # times hat swapped out */
1696 
1697 	int		sf_ctxfree;		/* ctx alloc from free list */
1698 	int		sf_ctxdirty;		/* ctx alloc from dirty list */
1699 	int		sf_ctxsteal;		/* ctx allocated by steal */
1700 
1701 	int		sf_tsb_alloc;		/* # TSB allocations */
1702 	int		sf_tsb_allocfail;	/* # times TSB alloc fail */
1703 	int		sf_tsb_sectsb_create;	/* # times second TSB added */
1704 
1705 	int		sf_tteload8k;		/* calls to sfmmu_tteload */
1706 	int		sf_tteload64k;		/* calls to sfmmu_tteload */
1707 	int		sf_tteload512k;		/* calls to sfmmu_tteload */
1708 	int		sf_tteload4m;		/* calls to sfmmu_tteload */
1709 	int		sf_tteload32m;		/* calls to sfmmu_tteload */
1710 	int		sf_tteload256m;		/* calls to sfmmu_tteload */
1711 
1712 	int		sf_tsb_load8k;		/* # times loaded 8K tsbent */
1713 	int		sf_tsb_load4m;		/* # times loaded 4M tsbent */
1714 
1715 	int		sf_hblk_hit;		/* found hblk during tteload */
1716 	int		sf_hblk8_ncreate;	/* static hblk8's created */
1717 	int		sf_hblk8_nalloc;	/* static hblk8's allocated */
1718 	int		sf_hblk1_ncreate;	/* static hblk1's created */
1719 	int		sf_hblk1_nalloc;	/* static hblk1's allocated */
1720 	int		sf_hblk_slab_cnt;	/* sfmmu8_cache slab creates */
1721 	int		sf_hblk_reserve_cnt;	/* hblk_reserve usage */
1722 	int		sf_hblk_recurse_cnt;	/* hblk_reserve	owner reqs */
1723 	int		sf_hblk_reserve_hit;	/* hblk_reserve hash hits */
1724 	int		sf_get_free_success;	/* reserve list allocs */
1725 	int		sf_get_free_throttle;	/* fails due to throttling */
1726 	int		sf_get_free_fail;	/* fails due to empty list */
1727 	int		sf_put_free_success;	/* reserve list frees */
1728 	int		sf_put_free_fail;	/* fails due to full list */
1729 
1730 	int		sf_pgcolor_conflict;	/* VAC conflict resolution */
1731 	int		sf_uncache_conflict;	/* VAC conflict resolution */
1732 	int		sf_unload_conflict;	/* VAC unload resolution */
1733 	int		sf_ism_uncache;		/* VAC conflict resolution */
1734 	int		sf_ism_recache;		/* VAC conflict resolution */
1735 	int		sf_recache;		/* VAC conflict resolution */
1736 
1737 	int		sf_steal_count;		/* # of hblks stolen */
1738 
1739 	int		sf_pagesync;		/* # of pagesyncs */
1740 	int		sf_clrwrt;		/* # of clear write perms */
1741 	int		sf_pagesync_invalid;	/* pagesync with inv tte */
1742 
1743 	int		sf_kernel_xcalls;	/* # of kernel cross calls */
1744 	int		sf_user_xcalls;		/* # of user cross calls */
1745 
1746 	int		sf_tsb_grow;		/* # of user tsb grows */
1747 	int		sf_tsb_shrink;		/* # of user tsb shrinks */
1748 	int		sf_tsb_resize_failures;	/* # of user tsb resize */
1749 	int		sf_tsb_reloc;		/* # of user tsb relocations */
1750 
1751 	int		sf_user_vtop;		/* # of user vatopfn calls */
1752 
1753 	int		sf_ctx_swap;		/* # times switched MMU ctxs */
1754 	int		sf_tlbflush_all;	/* # times flush all TLBs */
1755 	int		sf_tlbflush_ctx;	/* # times flush TLB ctx */
1756 	int		sf_tlbflush_deferred;	/* # times !flush ctx imm. */
1757 
1758 	int		sf_tlb_reprog_pgsz;	/* # times switch TLB pgsz */
1759 };
1760 
1761 struct sfmmu_tsbsize_stat {
1762 	int		sf_tsbsz_8k;
1763 	int		sf_tsbsz_16k;
1764 	int		sf_tsbsz_32k;
1765 	int		sf_tsbsz_64k;
1766 	int		sf_tsbsz_128k;
1767 	int		sf_tsbsz_256k;
1768 	int		sf_tsbsz_512k;
1769 	int		sf_tsbsz_1m;
1770 	int		sf_tsbsz_2m;
1771 	int		sf_tsbsz_4m;
1772 };
1773 
1774 struct sfmmu_percpu_stat {
1775 	int	sf_itlb_misses;		/* # of itlb misses */
1776 	int	sf_dtlb_misses;		/* # of dtlb misses */
1777 	int	sf_utsb_misses;		/* # of user tsb misses */
1778 	int	sf_ktsb_misses;		/* # of kernel tsb misses */
1779 	int	sf_tsb_hits;		/* # of tsb hits */
1780 	int	sf_umod_faults;		/* # of mod (prot viol) flts */
1781 	int	sf_kmod_faults;		/* # of mod (prot viol) flts */
1782 };
1783 
1784 #define	SFMMU_STAT(stat)		sfmmu_global_stat.stat++;
1785 #define	SFMMU_STAT_ADD(stat, amount)	sfmmu_global_stat.stat += amount;
1786 #define	SFMMU_STAT_SET(stat, count)	sfmmu_global_stat.stat = count;
1787 
1788 #endif /* !_ASM */
1789 
1790 #ifdef	__cplusplus
1791 }
1792 #endif
1793 
1794 #endif	/* _VM_HAT_SFMMU_H */
1795