xref: /freebsd/sys/powerpc/aim/moea64_native.c (revision f993ed2fbd3c307200ed9a6351e649f0904b39c5)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD AND 4-Clause-BSD
3  *
4  * Copyright (c) 2001 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Matt Thomas <matt@3am-software.com> of Allegro Networks, Inc.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 /*-
32  * Copyright (C) 1995, 1996 Wolfgang Solfrank.
33  * Copyright (C) 1995, 1996 TooLs GmbH.
34  * All rights reserved.
35  *
36  * Redistribution and use in source and binary forms, with or without
37  * modification, are permitted provided that the following conditions
38  * are met:
39  * 1. Redistributions of source code must retain the above copyright
40  *    notice, this list of conditions and the following disclaimer.
41  * 2. Redistributions in binary form must reproduce the above copyright
42  *    notice, this list of conditions and the following disclaimer in the
43  *    documentation and/or other materials provided with the distribution.
44  * 3. All advertising materials mentioning features or use of this software
45  *    must display the following acknowledgement:
46  *	This product includes software developed by TooLs GmbH.
47  * 4. The name of TooLs GmbH may not be used to endorse or promote products
48  *    derived from this software without specific prior written permission.
49  *
50  * THIS SOFTWARE IS PROVIDED BY TOOLS GMBH ``AS IS'' AND ANY EXPRESS OR
51  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
52  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
53  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
54  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
55  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
56  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
57  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
58  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
59  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
60  *
61  * $NetBSD: pmap.c,v 1.28 2000/03/26 20:42:36 kleink Exp $
62  */
63 /*-
64  * Copyright (C) 2001 Benno Rice.
65  * All rights reserved.
66  *
67  * Redistribution and use in source and binary forms, with or without
68  * modification, are permitted provided that the following conditions
69  * are met:
70  * 1. Redistributions of source code must retain the above copyright
71  *    notice, this list of conditions and the following disclaimer.
72  * 2. Redistributions in binary form must reproduce the above copyright
73  *    notice, this list of conditions and the following disclaimer in the
74  *    documentation and/or other materials provided with the distribution.
75  *
76  * THIS SOFTWARE IS PROVIDED BY Benno Rice ``AS IS'' AND ANY EXPRESS OR
77  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
78  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
79  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
80  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
81  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
82  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
83  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
84  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
85  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
86  */
87 
88 #include <sys/cdefs.h>
89 __FBSDID("$FreeBSD$");
90 
91 /*
92  * Native 64-bit page table operations for running without a hypervisor.
93  */
94 
95 #include <sys/param.h>
96 #include <sys/kernel.h>
97 #include <sys/ktr.h>
98 #include <sys/lock.h>
99 #include <sys/mutex.h>
100 #include <sys/proc.h>
101 #include <sys/sched.h>
102 #include <sys/sysctl.h>
103 #include <sys/systm.h>
104 #include <sys/rwlock.h>
105 #include <sys/endian.h>
106 
107 #include <sys/kdb.h>
108 
109 #include <vm/vm.h>
110 #include <vm/vm_param.h>
111 #include <vm/vm_kern.h>
112 #include <vm/vm_page.h>
113 #include <vm/vm_map.h>
114 #include <vm/vm_object.h>
115 #include <vm/vm_extern.h>
116 #include <vm/vm_pageout.h>
117 
118 #include <machine/cpu.h>
119 #include <machine/hid.h>
120 #include <machine/md_var.h>
121 #include <machine/mmuvar.h>
122 
123 #include "mmu_oea64.h"
124 #include "mmu_if.h"
125 #include "moea64_if.h"
126 
127 #define	PTESYNC()	__asm __volatile("ptesync");
128 #define	TLBSYNC()	__asm __volatile("tlbsync; ptesync");
129 #define	SYNC()		__asm __volatile("sync");
130 #define	EIEIO()		__asm __volatile("eieio");
131 
132 #define	VSID_HASH_MASK	0x0000007fffffffffULL
133 
134 /* POWER9 only permits a 64k partition table size. */
135 #define	PART_SIZE	0x10000
136 
137 static bool moea64_crop_tlbie;
138 static bool moea64_need_lock;
139 
140 static __inline void
141 TLBIE(uint64_t vpn) {
142 #ifndef __powerpc64__
143 	register_t vpn_hi, vpn_lo;
144 	register_t msr;
145 	register_t scratch, intr;
146 #endif
147 
148 	static volatile u_int tlbie_lock = 0;
149 	bool need_lock = moea64_need_lock;
150 
151 	vpn <<= ADDR_PIDX_SHFT;
152 
153 	/* Hobo spinlock: we need stronger guarantees than mutexes provide */
154 	if (need_lock) {
155 		while (!atomic_cmpset_int(&tlbie_lock, 0, 1));
156 		isync(); /* Flush instruction queue once lock acquired */
157 
158 		if (moea64_crop_tlbie)
159 			vpn &= ~(0xffffULL << 48);
160 	}
161 
162 #ifdef __powerpc64__
163 	/*
164 	 * Explicitly clobber r0.  The tlbie instruction has two forms: an old
165 	 * one used by PowerISA 2.03 and prior, and a newer one used by PowerISA
166 	 * 2.06 (maybe 2.05?) and later.  We need to support both, and it just
167 	 * so happens that since we use 4k pages we can simply zero out r0, and
168 	 * clobber it, and the assembler will interpret the single-operand form
169 	 * of tlbie as having RB set, and everything else as 0.  The RS operand
170 	 * in the newer form is in the same position as the L(page size) bit of
171 	 * the old form, so a slong as RS is 0, we're good on both sides.
172 	 */
173 	__asm __volatile("li 0, 0 \n tlbie %0" :: "r"(vpn) : "r0", "memory");
174 	__asm __volatile("eieio; tlbsync; ptesync" ::: "memory");
175 #else
176 	vpn_hi = (uint32_t)(vpn >> 32);
177 	vpn_lo = (uint32_t)vpn;
178 
179 	intr = intr_disable();
180 	__asm __volatile("\
181 	    mfmsr %0; \
182 	    mr %1, %0; \
183 	    insrdi %1,%5,1,0; \
184 	    mtmsrd %1; isync; \
185 	    \
186 	    sld %1,%2,%4; \
187 	    or %1,%1,%3; \
188 	    tlbie %1; \
189 	    \
190 	    mtmsrd %0; isync; \
191 	    eieio; \
192 	    tlbsync; \
193 	    ptesync;"
194 	: "=r"(msr), "=r"(scratch) : "r"(vpn_hi), "r"(vpn_lo), "r"(32), "r"(1)
195 	    : "memory");
196 	intr_restore(intr);
197 #endif
198 
199 	/* No barriers or special ops -- taken care of by ptesync above */
200 	if (need_lock)
201 		tlbie_lock = 0;
202 }
203 
204 #define DISABLE_TRANS(msr)	msr = mfmsr(); mtmsr(msr & ~PSL_DR)
205 #define ENABLE_TRANS(msr)	mtmsr(msr)
206 
207 /*
208  * PTEG data.
209  */
210 static volatile struct lpte *moea64_pteg_table;
211 static struct rwlock moea64_eviction_lock;
212 
213 static volatile struct pate *moea64_part_table;
214 
215 /*
216  * PTE calls.
217  */
218 static int	moea64_pte_insert_native(mmu_t, struct pvo_entry *);
219 static int64_t	moea64_pte_synch_native(mmu_t, struct pvo_entry *);
220 static int64_t	moea64_pte_clear_native(mmu_t, struct pvo_entry *, uint64_t);
221 static int64_t	moea64_pte_replace_native(mmu_t, struct pvo_entry *, int);
222 static int64_t	moea64_pte_unset_native(mmu_t mmu, struct pvo_entry *);
223 
224 /*
225  * Utility routines.
226  */
227 static void	moea64_bootstrap_native(mmu_t mmup,
228 		    vm_offset_t kernelstart, vm_offset_t kernelend);
229 static void	moea64_cpu_bootstrap_native(mmu_t, int ap);
230 static void	tlbia(void);
231 
232 static mmu_method_t moea64_native_methods[] = {
233 	/* Internal interfaces */
234 	MMUMETHOD(mmu_bootstrap,	moea64_bootstrap_native),
235 	MMUMETHOD(mmu_cpu_bootstrap,	moea64_cpu_bootstrap_native),
236 
237 	MMUMETHOD(moea64_pte_synch,	moea64_pte_synch_native),
238 	MMUMETHOD(moea64_pte_clear,	moea64_pte_clear_native),
239 	MMUMETHOD(moea64_pte_unset,	moea64_pte_unset_native),
240 	MMUMETHOD(moea64_pte_replace,	moea64_pte_replace_native),
241 	MMUMETHOD(moea64_pte_insert,	moea64_pte_insert_native),
242 
243 	{ 0, 0 }
244 };
245 
246 MMU_DEF_INHERIT(oea64_mmu_native, MMU_TYPE_G5, moea64_native_methods,
247     0, oea64_mmu);
248 
249 static int64_t
250 moea64_pte_synch_native(mmu_t mmu, struct pvo_entry *pvo)
251 {
252 	volatile struct lpte *pt = moea64_pteg_table + pvo->pvo_pte.slot;
253 	struct lpte properpt;
254 	uint64_t ptelo;
255 
256 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
257 
258 	moea64_pte_from_pvo(pvo, &properpt);
259 
260 	rw_rlock(&moea64_eviction_lock);
261 	if ((be64toh(pt->pte_hi) & LPTE_AVPN_MASK) !=
262 	    (properpt.pte_hi & LPTE_AVPN_MASK)) {
263 		/* Evicted */
264 		rw_runlock(&moea64_eviction_lock);
265 		return (-1);
266 	}
267 
268 	PTESYNC();
269 	ptelo = be64toh(pt->pte_lo);
270 
271 	rw_runlock(&moea64_eviction_lock);
272 
273 	return (ptelo & (LPTE_REF | LPTE_CHG));
274 }
275 
276 static int64_t
277 moea64_pte_clear_native(mmu_t mmu, struct pvo_entry *pvo, uint64_t ptebit)
278 {
279 	volatile struct lpte *pt = moea64_pteg_table + pvo->pvo_pte.slot;
280 	struct lpte properpt;
281 	uint64_t ptelo;
282 
283 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
284 
285 	moea64_pte_from_pvo(pvo, &properpt);
286 
287 	rw_rlock(&moea64_eviction_lock);
288 	if ((be64toh(pt->pte_hi) & LPTE_AVPN_MASK) !=
289 	    (properpt.pte_hi & LPTE_AVPN_MASK)) {
290 		/* Evicted */
291 		rw_runlock(&moea64_eviction_lock);
292 		return (-1);
293 	}
294 
295 	if (ptebit == LPTE_REF) {
296 		/* See "Resetting the Reference Bit" in arch manual */
297 		PTESYNC();
298 		/* 2-step here safe: precision is not guaranteed */
299 		ptelo = be64toh(pt->pte_lo);
300 
301 		/* One-byte store to avoid touching the C bit */
302 		((volatile uint8_t *)(&pt->pte_lo))[6] =
303 #if BYTE_ORDER == BIG_ENDIAN
304 		    ((uint8_t *)(&properpt.pte_lo))[6];
305 #else
306 		    ((uint8_t *)(&properpt.pte_lo))[1];
307 #endif
308 		rw_runlock(&moea64_eviction_lock);
309 
310 		critical_enter();
311 		TLBIE(pvo->pvo_vpn);
312 		critical_exit();
313 	} else {
314 		rw_runlock(&moea64_eviction_lock);
315 		ptelo = moea64_pte_unset_native(mmu, pvo);
316 		moea64_pte_insert_native(mmu, pvo);
317 	}
318 
319 	return (ptelo & (LPTE_REF | LPTE_CHG));
320 }
321 
322 static int64_t
323 moea64_pte_unset_native(mmu_t mmu, struct pvo_entry *pvo)
324 {
325 	volatile struct lpte *pt = moea64_pteg_table + pvo->pvo_pte.slot;
326 	struct lpte properpt;
327 	uint64_t ptelo;
328 
329 	moea64_pte_from_pvo(pvo, &properpt);
330 
331 	rw_rlock(&moea64_eviction_lock);
332 	if ((be64toh(pt->pte_hi & LPTE_AVPN_MASK)) !=
333 	    (properpt.pte_hi & LPTE_AVPN_MASK)) {
334 		/* Evicted */
335 		STAT_MOEA64(moea64_pte_overflow--);
336 		rw_runlock(&moea64_eviction_lock);
337 		return (-1);
338 	}
339 
340 	/*
341 	 * Invalidate the pte, briefly locking it to collect RC bits. No
342 	 * atomics needed since this is protected against eviction by the lock.
343 	 */
344 	isync();
345 	critical_enter();
346 	pt->pte_hi = be64toh((pt->pte_hi & ~LPTE_VALID) | LPTE_LOCKED);
347 	PTESYNC();
348 	TLBIE(pvo->pvo_vpn);
349 	ptelo = be64toh(pt->pte_lo);
350 	*((volatile int32_t *)(&pt->pte_hi) + 1) = 0; /* Release lock */
351 	critical_exit();
352 	rw_runlock(&moea64_eviction_lock);
353 
354 	/* Keep statistics */
355 	STAT_MOEA64(moea64_pte_valid--);
356 
357 	return (ptelo & (LPTE_CHG | LPTE_REF));
358 }
359 
360 static int64_t
361 moea64_pte_replace_inval_native(mmu_t mmu, struct pvo_entry *pvo,
362     volatile struct lpte *pt)
363 {
364 	struct lpte properpt;
365 	uint64_t ptelo;
366 
367 	moea64_pte_from_pvo(pvo, &properpt);
368 
369 	rw_rlock(&moea64_eviction_lock);
370 	if ((be64toh(pt->pte_hi & LPTE_AVPN_MASK)) !=
371 	    (properpt.pte_hi & LPTE_AVPN_MASK)) {
372 		/* Evicted */
373 		STAT_MOEA64(moea64_pte_overflow--);
374 		rw_runlock(&moea64_eviction_lock);
375 		return (-1);
376 	}
377 
378 	/*
379 	 * Replace the pte, briefly locking it to collect RC bits. No
380 	 * atomics needed since this is protected against eviction by the lock.
381 	 */
382 	isync();
383 	critical_enter();
384 	pt->pte_hi = be64toh((pt->pte_hi & ~LPTE_VALID) | LPTE_LOCKED);
385 	PTESYNC();
386 	TLBIE(pvo->pvo_vpn);
387 	ptelo = be64toh(pt->pte_lo);
388 	EIEIO();
389 	pt->pte_lo = htobe64(properpt.pte_lo);
390 	EIEIO();
391 	pt->pte_hi = htobe64(properpt.pte_hi); /* Release lock */
392 	PTESYNC();
393 	critical_exit();
394 	rw_runlock(&moea64_eviction_lock);
395 
396 	return (ptelo & (LPTE_CHG | LPTE_REF));
397 }
398 
399 static int64_t
400 moea64_pte_replace_native(mmu_t mmu, struct pvo_entry *pvo, int flags)
401 {
402 	volatile struct lpte *pt = moea64_pteg_table + pvo->pvo_pte.slot;
403 	struct lpte properpt;
404 	int64_t ptelo;
405 
406 	if (flags == 0) {
407 		/* Just some software bits changing. */
408 		moea64_pte_from_pvo(pvo, &properpt);
409 
410 		rw_rlock(&moea64_eviction_lock);
411 		if ((be64toh(pt->pte_hi) & LPTE_AVPN_MASK) !=
412 		    (properpt.pte_hi & LPTE_AVPN_MASK)) {
413 			rw_runlock(&moea64_eviction_lock);
414 			return (-1);
415 		}
416 		pt->pte_hi = htobe64(properpt.pte_hi);
417 		ptelo = be64toh(pt->pte_lo);
418 		rw_runlock(&moea64_eviction_lock);
419 	} else {
420 		/* Otherwise, need reinsertion and deletion */
421 		ptelo = moea64_pte_replace_inval_native(mmu, pvo, pt);
422 	}
423 
424 	return (ptelo);
425 }
426 
427 static void
428 moea64_cpu_bootstrap_native(mmu_t mmup, int ap)
429 {
430 	int i = 0;
431 	#ifdef __powerpc64__
432 	struct slb *slb = PCPU_GET(aim.slb);
433 	register_t seg0;
434 	#endif
435 
436 	/*
437 	 * Initialize segment registers and MMU
438 	 */
439 
440 	mtmsr(mfmsr() & ~PSL_DR & ~PSL_IR);
441 
442 	switch(mfpvr() >> 16) {
443 	case IBMPOWER9:
444 		mtspr(SPR_HID0, mfspr(SPR_HID0) & ~HID0_RADIX);
445 		break;
446 	}
447 
448 	/*
449 	 * Install kernel SLB entries
450 	 */
451 
452 	#ifdef __powerpc64__
453 		__asm __volatile ("slbia");
454 		__asm __volatile ("slbmfee %0,%1; slbie %0;" : "=r"(seg0) :
455 		    "r"(0));
456 
457 		for (i = 0; i < n_slbs; i++) {
458 			if (!(slb[i].slbe & SLBE_VALID))
459 				continue;
460 
461 			__asm __volatile ("slbmte %0, %1" ::
462 			    "r"(slb[i].slbv), "r"(slb[i].slbe));
463 		}
464 	#else
465 		for (i = 0; i < 16; i++)
466 			mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]);
467 	#endif
468 
469 	/*
470 	 * Install page table
471 	 */
472 
473 	if (cpu_features2 & PPC_FEATURE2_ARCH_3_00)
474 		mtspr(SPR_PTCR,
475 		    ((uintptr_t)moea64_part_table & ~DMAP_BASE_ADDRESS) |
476 		     flsl((PART_SIZE >> 12) - 1));
477 	else
478 		__asm __volatile ("ptesync; mtsdr1 %0; isync"
479 		    :: "r"(((uintptr_t)moea64_pteg_table & ~DMAP_BASE_ADDRESS)
480 			     | (uintptr_t)(flsl(moea64_pteg_mask >> 11))));
481 	tlbia();
482 }
483 
484 static void
485 moea64_bootstrap_native(mmu_t mmup, vm_offset_t kernelstart,
486     vm_offset_t kernelend)
487 {
488 	vm_size_t	size;
489 	vm_offset_t	off;
490 	vm_paddr_t	pa;
491 	register_t	msr;
492 
493 	moea64_early_bootstrap(mmup, kernelstart, kernelend);
494 
495 	switch (mfpvr() >> 16) {
496 	case IBMPOWER9:
497 		moea64_need_lock = false;
498 		break;
499 	case IBMPOWER4:
500 	case IBMPOWER4PLUS:
501 	case IBM970:
502 	case IBM970FX:
503 	case IBM970GX:
504 	case IBM970MP:
505 		moea64_crop_tlbie = true;
506 	default:
507 		moea64_need_lock = true;
508 	}
509 	/*
510 	 * Allocate PTEG table.
511 	 */
512 
513 	size = moea64_pteg_count * sizeof(struct lpteg);
514 	CTR2(KTR_PMAP, "moea64_bootstrap: %lu PTEGs, %lu bytes",
515 	    moea64_pteg_count, size);
516 	rw_init(&moea64_eviction_lock, "pte eviction");
517 
518 	/*
519 	 * We now need to allocate memory. This memory, to be allocated,
520 	 * has to reside in a page table. The page table we are about to
521 	 * allocate. We don't have BAT. So drop to data real mode for a minute
522 	 * as a measure of last resort. We do this a couple times.
523 	 */
524 	/*
525 	 * PTEG table must be aligned on a 256k boundary, but can be placed
526 	 * anywhere with that alignment on POWER ISA 3+ systems. On earlier
527 	 * systems, offset addition is done by the CPU with bitwise OR rather
528 	 * than addition, so the table must also be aligned on a boundary of
529 	 * its own size. Pick the larger of the two, which works on all
530 	 * systems.
531 	 */
532 	moea64_pteg_table = (struct lpte *)moea64_bootstrap_alloc(size,
533 	    MAX(256*1024, size));
534 	if (hw_direct_map)
535 		moea64_pteg_table =
536 		    (struct lpte *)PHYS_TO_DMAP((vm_offset_t)moea64_pteg_table);
537 	/* Allocate partition table (ISA 3.0). */
538 	if (cpu_features2 & PPC_FEATURE2_ARCH_3_00) {
539 		moea64_part_table =
540 		    (struct pate *)moea64_bootstrap_alloc(PART_SIZE, PART_SIZE);
541 		if (hw_direct_map)
542 			moea64_part_table =
543 			    (struct pate *)PHYS_TO_DMAP((vm_offset_t)moea64_part_table);
544 	}
545 	DISABLE_TRANS(msr);
546 	bzero(__DEVOLATILE(void *, moea64_pteg_table), moea64_pteg_count *
547 	    sizeof(struct lpteg));
548 	if (cpu_features2 & PPC_FEATURE2_ARCH_3_00) {
549 		bzero(__DEVOLATILE(void *, moea64_part_table), PART_SIZE);
550 		moea64_part_table[0].pagetab =
551 			(DMAP_TO_PHYS((vm_offset_t)moea64_pteg_table)) |
552 			(uintptr_t)(flsl((moea64_pteg_count - 1) >> 11));
553 	}
554 	ENABLE_TRANS(msr);
555 
556 	CTR1(KTR_PMAP, "moea64_bootstrap: PTEG table at %p", moea64_pteg_table);
557 
558 	moea64_mid_bootstrap(mmup, kernelstart, kernelend);
559 
560 	/*
561 	 * Add a mapping for the page table itself if there is no direct map.
562 	 */
563 	if (!hw_direct_map) {
564 		size = moea64_pteg_count * sizeof(struct lpteg);
565 		off = (vm_offset_t)(moea64_pteg_table);
566 		DISABLE_TRANS(msr);
567 		for (pa = off; pa < off + size; pa += PAGE_SIZE)
568 			pmap_kenter(pa, pa);
569 		ENABLE_TRANS(msr);
570 	}
571 
572 	/* Bring up virtual memory */
573 	moea64_late_bootstrap(mmup, kernelstart, kernelend);
574 }
575 
576 static void
577 tlbia(void)
578 {
579 	vm_offset_t i;
580 	#ifndef __powerpc64__
581 	register_t msr, scratch;
582 	#endif
583 
584 	i = 0xc00; /* IS = 11 */
585 	switch (mfpvr() >> 16) {
586 	case IBM970:
587 	case IBM970FX:
588 	case IBM970MP:
589 	case IBM970GX:
590 	case IBMPOWER4:
591 	case IBMPOWER4PLUS:
592 	case IBMPOWER5:
593 	case IBMPOWER5PLUS:
594 		i = 0; /* IS not supported */
595 		break;
596 	}
597 
598 	TLBSYNC();
599 
600 	for (; i < 0x400000; i += 0x00001000) {
601 		#ifdef __powerpc64__
602 		__asm __volatile("tlbiel %0" :: "r"(i));
603 		#else
604 		__asm __volatile("\
605 		    mfmsr %0; \
606 		    mr %1, %0; \
607 		    insrdi %1,%3,1,0; \
608 		    mtmsrd %1; \
609 		    isync; \
610 		    \
611 		    tlbiel %2; \
612 		    \
613 		    mtmsrd %0; \
614 		    isync;"
615 		: "=r"(msr), "=r"(scratch) : "r"(i), "r"(1));
616 		#endif
617 	}
618 
619 	EIEIO();
620 	TLBSYNC();
621 }
622 
623 static int
624 atomic_pte_lock(volatile struct lpte *pte, uint64_t bitmask, uint64_t *oldhi)
625 {
626 	int	ret;
627 	uint32_t oldhihalf;
628 
629 	/*
630 	 * Note: in principle, if just the locked bit were set here, we
631 	 * could avoid needing the eviction lock. However, eviction occurs
632 	 * so rarely that it isn't worth bothering about in practice.
633 	 */
634 
635 	__asm __volatile (
636 		"1:\tlwarx %1, 0, %3\n\t"	/* load old value */
637 		"and. %0,%1,%4\n\t"		/* check if any bits set */
638 		"bne 2f\n\t"			/* exit if any set */
639 		"stwcx. %5, 0, %3\n\t"      	/* attempt to store */
640 		"bne- 1b\n\t"			/* spin if failed */
641 		"li %0, 1\n\t"			/* success - retval = 1 */
642 		"b 3f\n\t"			/* we've succeeded */
643 		"2:\n\t"
644 		"stwcx. %1, 0, %3\n\t"       	/* clear reservation (74xx) */
645 		"li %0, 0\n\t"			/* failure - retval = 0 */
646 		"3:\n\t"
647 		: "=&r" (ret), "=&r"(oldhihalf), "=m" (pte->pte_hi)
648 		: "r" ((volatile char *)&pte->pte_hi + 4),
649 		  "r" ((uint32_t)bitmask), "r" ((uint32_t)LPTE_LOCKED),
650 		  "m" (pte->pte_hi)
651 		: "cr0", "cr1", "cr2", "memory");
652 
653 	*oldhi = (pte->pte_hi & 0xffffffff00000000ULL) | oldhihalf;
654 
655 	return (ret);
656 }
657 
658 static uintptr_t
659 moea64_insert_to_pteg_native(struct lpte *pvo_pt, uintptr_t slotbase,
660     uint64_t mask)
661 {
662 	volatile struct lpte *pt;
663 	uint64_t oldptehi, va;
664 	uintptr_t k;
665 	int i, j;
666 
667 	/* Start at a random slot */
668 	i = mftb() % 8;
669 	for (j = 0; j < 8; j++) {
670 		k = slotbase + (i + j) % 8;
671 		pt = &moea64_pteg_table[k];
672 		/* Invalidate and seize lock only if no bits in mask set */
673 		if (atomic_pte_lock(pt, mask, &oldptehi)) /* Lock obtained */
674 			break;
675 	}
676 
677 	if (j == 8)
678 		return (-1);
679 
680 	if (oldptehi & LPTE_VALID) {
681 		KASSERT(!(oldptehi & LPTE_WIRED), ("Unmapped wired entry"));
682 		/*
683 		 * Need to invalidate old entry completely: see
684 		 * "Modifying a Page Table Entry". Need to reconstruct
685 		 * the virtual address for the outgoing entry to do that.
686 		 */
687 		va = oldptehi >> (ADDR_SR_SHFT - ADDR_API_SHFT64);
688 		if (oldptehi & LPTE_HID)
689 			va = (((k >> 3) ^ moea64_pteg_mask) ^ va) &
690 			    (ADDR_PIDX >> ADDR_PIDX_SHFT);
691 		else
692 			va = ((k >> 3) ^ va) & (ADDR_PIDX >> ADDR_PIDX_SHFT);
693 		va |= (oldptehi & LPTE_AVPN_MASK) <<
694 		    (ADDR_API_SHFT64 - ADDR_PIDX_SHFT);
695 		PTESYNC();
696 		TLBIE(va);
697 		STAT_MOEA64(moea64_pte_valid--);
698 		STAT_MOEA64(moea64_pte_overflow++);
699 	}
700 
701 	/*
702 	 * Update the PTE as per "Adding a Page Table Entry". Lock is released
703 	 * by setting the high doubleworld.
704 	 */
705 	pt->pte_lo = htobe64(pvo_pt->pte_lo);
706 	EIEIO();
707 	pt->pte_hi = htobe64(pvo_pt->pte_hi);
708 	PTESYNC();
709 
710 	/* Keep statistics */
711 	STAT_MOEA64(moea64_pte_valid++);
712 
713 	return (k);
714 }
715 
716 static int
717 moea64_pte_insert_native(mmu_t mmu, struct pvo_entry *pvo)
718 {
719 	struct lpte insertpt;
720 	uintptr_t slot;
721 
722 	/* Initialize PTE */
723 	moea64_pte_from_pvo(pvo, &insertpt);
724 
725 	/* Make sure further insertion is locked out during evictions */
726 	rw_rlock(&moea64_eviction_lock);
727 
728 	/*
729 	 * First try primary hash.
730 	 */
731 	pvo->pvo_pte.slot &= ~7ULL; /* Base slot address */
732 	slot = moea64_insert_to_pteg_native(&insertpt, pvo->pvo_pte.slot,
733 	    LPTE_VALID | LPTE_WIRED | LPTE_LOCKED);
734 	if (slot != -1) {
735 		rw_runlock(&moea64_eviction_lock);
736 		pvo->pvo_pte.slot = slot;
737 		return (0);
738 	}
739 
740 	/*
741 	 * Now try secondary hash.
742 	 */
743 	pvo->pvo_vaddr ^= PVO_HID;
744 	insertpt.pte_hi ^= LPTE_HID;
745 	pvo->pvo_pte.slot ^= (moea64_pteg_mask << 3);
746 	slot = moea64_insert_to_pteg_native(&insertpt, pvo->pvo_pte.slot,
747 	    LPTE_VALID | LPTE_WIRED | LPTE_LOCKED);
748 	if (slot != -1) {
749 		rw_runlock(&moea64_eviction_lock);
750 		pvo->pvo_pte.slot = slot;
751 		return (0);
752 	}
753 
754 	/*
755 	 * Out of luck. Find a PTE to sacrifice.
756 	 */
757 
758 	/* Lock out all insertions for a bit */
759 	if (!rw_try_upgrade(&moea64_eviction_lock)) {
760 		rw_runlock(&moea64_eviction_lock);
761 		rw_wlock(&moea64_eviction_lock);
762 	}
763 
764 	slot = moea64_insert_to_pteg_native(&insertpt, pvo->pvo_pte.slot,
765 	    LPTE_WIRED | LPTE_LOCKED);
766 	if (slot != -1) {
767 		rw_wunlock(&moea64_eviction_lock);
768 		pvo->pvo_pte.slot = slot;
769 		return (0);
770 	}
771 
772 	/* Try other hash table. Now we're getting desperate... */
773 	pvo->pvo_vaddr ^= PVO_HID;
774 	insertpt.pte_hi ^= LPTE_HID;
775 	pvo->pvo_pte.slot ^= (moea64_pteg_mask << 3);
776 	slot = moea64_insert_to_pteg_native(&insertpt, pvo->pvo_pte.slot,
777 	    LPTE_WIRED | LPTE_LOCKED);
778 	if (slot != -1) {
779 		rw_wunlock(&moea64_eviction_lock);
780 		pvo->pvo_pte.slot = slot;
781 		return (0);
782 	}
783 
784 	/* No freeable slots in either PTEG? We're hosed. */
785 	rw_wunlock(&moea64_eviction_lock);
786 	panic("moea64_pte_insert: overflow");
787 	return (-1);
788 }
789 
790