xref: /freebsd/sys/powerpc/aim/moea64_native.c (revision 5bf5ca772c6de2d53344a78cf461447cc322ccea)
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/md_var.h>
119 #include <machine/mmuvar.h>
120 
121 #include "mmu_oea64.h"
122 #include "mmu_if.h"
123 #include "moea64_if.h"
124 
125 #define	PTESYNC()	__asm __volatile("ptesync");
126 #define	TLBSYNC()	__asm __volatile("tlbsync; ptesync");
127 #define	SYNC()		__asm __volatile("sync");
128 #define	EIEIO()		__asm __volatile("eieio");
129 
130 #define	VSID_HASH_MASK	0x0000007fffffffffULL
131 
132 static __inline void
133 TLBIE(uint64_t vpn) {
134 #ifndef __powerpc64__
135 	register_t vpn_hi, vpn_lo;
136 	register_t msr;
137 	register_t scratch, intr;
138 #endif
139 
140 	static volatile u_int tlbie_lock = 0;
141 
142 	vpn <<= ADDR_PIDX_SHFT;
143 	vpn &= ~(0xffffULL << 48);
144 
145 	/* Hobo spinlock: we need stronger guarantees than mutexes provide */
146 	while (!atomic_cmpset_int(&tlbie_lock, 0, 1));
147 	isync(); /* Flush instruction queue once lock acquired */
148 
149 #ifdef __powerpc64__
150 	__asm __volatile("tlbie %0" :: "r"(vpn) : "memory");
151 	__asm __volatile("eieio; tlbsync; ptesync" ::: "memory");
152 #else
153 	vpn_hi = (uint32_t)(vpn >> 32);
154 	vpn_lo = (uint32_t)vpn;
155 
156 	intr = intr_disable();
157 	__asm __volatile("\
158 	    mfmsr %0; \
159 	    mr %1, %0; \
160 	    insrdi %1,%5,1,0; \
161 	    mtmsrd %1; isync; \
162 	    \
163 	    sld %1,%2,%4; \
164 	    or %1,%1,%3; \
165 	    tlbie %1; \
166 	    \
167 	    mtmsrd %0; isync; \
168 	    eieio; \
169 	    tlbsync; \
170 	    ptesync;"
171 	: "=r"(msr), "=r"(scratch) : "r"(vpn_hi), "r"(vpn_lo), "r"(32), "r"(1)
172 	    : "memory");
173 	intr_restore(intr);
174 #endif
175 
176 	/* No barriers or special ops -- taken care of by ptesync above */
177 	tlbie_lock = 0;
178 }
179 
180 #define DISABLE_TRANS(msr)	msr = mfmsr(); mtmsr(msr & ~PSL_DR)
181 #define ENABLE_TRANS(msr)	mtmsr(msr)
182 
183 /*
184  * PTEG data.
185  */
186 static volatile struct lpte *moea64_pteg_table;
187 static struct rwlock moea64_eviction_lock;
188 
189 /*
190  * PTE calls.
191  */
192 static int	moea64_pte_insert_native(mmu_t, struct pvo_entry *);
193 static int64_t	moea64_pte_synch_native(mmu_t, struct pvo_entry *);
194 static int64_t	moea64_pte_clear_native(mmu_t, struct pvo_entry *, uint64_t);
195 static int64_t	moea64_pte_replace_native(mmu_t, struct pvo_entry *, int);
196 static int64_t	moea64_pte_unset_native(mmu_t mmu, struct pvo_entry *);
197 
198 /*
199  * Utility routines.
200  */
201 static void	moea64_bootstrap_native(mmu_t mmup,
202 		    vm_offset_t kernelstart, vm_offset_t kernelend);
203 static void	moea64_cpu_bootstrap_native(mmu_t, int ap);
204 static void	tlbia(void);
205 
206 static mmu_method_t moea64_native_methods[] = {
207 	/* Internal interfaces */
208 	MMUMETHOD(mmu_bootstrap,	moea64_bootstrap_native),
209 	MMUMETHOD(mmu_cpu_bootstrap,	moea64_cpu_bootstrap_native),
210 
211 	MMUMETHOD(moea64_pte_synch,	moea64_pte_synch_native),
212 	MMUMETHOD(moea64_pte_clear,	moea64_pte_clear_native),
213 	MMUMETHOD(moea64_pte_unset,	moea64_pte_unset_native),
214 	MMUMETHOD(moea64_pte_replace,	moea64_pte_replace_native),
215 	MMUMETHOD(moea64_pte_insert,	moea64_pte_insert_native),
216 
217 	{ 0, 0 }
218 };
219 
220 MMU_DEF_INHERIT(oea64_mmu_native, MMU_TYPE_G5, moea64_native_methods,
221     0, oea64_mmu);
222 
223 static int64_t
224 moea64_pte_synch_native(mmu_t mmu, struct pvo_entry *pvo)
225 {
226 	volatile struct lpte *pt = moea64_pteg_table + pvo->pvo_pte.slot;
227 	struct lpte properpt;
228 	uint64_t ptelo;
229 
230 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
231 
232 	moea64_pte_from_pvo(pvo, &properpt);
233 
234 	rw_rlock(&moea64_eviction_lock);
235 	if ((be64toh(pt->pte_hi) & LPTE_AVPN_MASK) !=
236 	    (properpt.pte_hi & LPTE_AVPN_MASK)) {
237 		/* Evicted */
238 		rw_runlock(&moea64_eviction_lock);
239 		return (-1);
240 	}
241 
242 	PTESYNC();
243 	ptelo = be64toh(pt->pte_lo);
244 
245 	rw_runlock(&moea64_eviction_lock);
246 
247 	return (ptelo & (LPTE_REF | LPTE_CHG));
248 }
249 
250 static int64_t
251 moea64_pte_clear_native(mmu_t mmu, struct pvo_entry *pvo, uint64_t ptebit)
252 {
253 	volatile struct lpte *pt = moea64_pteg_table + pvo->pvo_pte.slot;
254 	struct lpte properpt;
255 	uint64_t ptelo;
256 
257 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
258 
259 	moea64_pte_from_pvo(pvo, &properpt);
260 
261 	rw_rlock(&moea64_eviction_lock);
262 	if ((be64toh(pt->pte_hi) & LPTE_AVPN_MASK) !=
263 	    (properpt.pte_hi & LPTE_AVPN_MASK)) {
264 		/* Evicted */
265 		rw_runlock(&moea64_eviction_lock);
266 		return (-1);
267 	}
268 
269 	if (ptebit == LPTE_REF) {
270 		/* See "Resetting the Reference Bit" in arch manual */
271 		PTESYNC();
272 		/* 2-step here safe: precision is not guaranteed */
273 		ptelo = be64toh(pt->pte_lo);
274 
275 		/* One-byte store to avoid touching the C bit */
276 		((volatile uint8_t *)(&pt->pte_lo))[6] =
277 #if BYTE_ORDER == BIG_ENDIAN
278 		    ((uint8_t *)(&properpt.pte_lo))[6];
279 #else
280 		    ((uint8_t *)(&properpt.pte_lo))[1];
281 #endif
282 		rw_runlock(&moea64_eviction_lock);
283 
284 		critical_enter();
285 		TLBIE(pvo->pvo_vpn);
286 		critical_exit();
287 	} else {
288 		rw_runlock(&moea64_eviction_lock);
289 		ptelo = moea64_pte_unset_native(mmu, pvo);
290 		moea64_pte_insert_native(mmu, pvo);
291 	}
292 
293 	return (ptelo & (LPTE_REF | LPTE_CHG));
294 }
295 
296 static int64_t
297 moea64_pte_unset_native(mmu_t mmu, struct pvo_entry *pvo)
298 {
299 	volatile struct lpte *pt = moea64_pteg_table + pvo->pvo_pte.slot;
300 	struct lpte properpt;
301 	uint64_t ptelo;
302 
303 	moea64_pte_from_pvo(pvo, &properpt);
304 
305 	rw_rlock(&moea64_eviction_lock);
306 	if ((be64toh(pt->pte_hi & LPTE_AVPN_MASK)) !=
307 	    (properpt.pte_hi & LPTE_AVPN_MASK)) {
308 		/* Evicted */
309 		moea64_pte_overflow--;
310 		rw_runlock(&moea64_eviction_lock);
311 		return (-1);
312 	}
313 
314 	/*
315 	 * Invalidate the pte, briefly locking it to collect RC bits. No
316 	 * atomics needed since this is protected against eviction by the lock.
317 	 */
318 	isync();
319 	critical_enter();
320 	pt->pte_hi = be64toh((pt->pte_hi & ~LPTE_VALID) | LPTE_LOCKED);
321 	PTESYNC();
322 	TLBIE(pvo->pvo_vpn);
323 	ptelo = be64toh(pt->pte_lo);
324 	*((volatile int32_t *)(&pt->pte_hi) + 1) = 0; /* Release lock */
325 	critical_exit();
326 	rw_runlock(&moea64_eviction_lock);
327 
328 	/* Keep statistics */
329 	moea64_pte_valid--;
330 
331 	return (ptelo & (LPTE_CHG | LPTE_REF));
332 }
333 
334 static int64_t
335 moea64_pte_replace_native(mmu_t mmu, struct pvo_entry *pvo, int flags)
336 {
337 	volatile struct lpte *pt = moea64_pteg_table + pvo->pvo_pte.slot;
338 	struct lpte properpt;
339 	int64_t ptelo;
340 
341 	if (flags == 0) {
342 		/* Just some software bits changing. */
343 		moea64_pte_from_pvo(pvo, &properpt);
344 
345 		rw_rlock(&moea64_eviction_lock);
346 		if ((be64toh(pt->pte_hi) & LPTE_AVPN_MASK) !=
347 		    (properpt.pte_hi & LPTE_AVPN_MASK)) {
348 			rw_runlock(&moea64_eviction_lock);
349 			return (-1);
350 		}
351 		pt->pte_hi = htobe64(properpt.pte_hi);
352 		ptelo = be64toh(pt->pte_lo);
353 		rw_runlock(&moea64_eviction_lock);
354 	} else {
355 		/* Otherwise, need reinsertion and deletion */
356 		ptelo = moea64_pte_unset_native(mmu, pvo);
357 		moea64_pte_insert_native(mmu, pvo);
358 	}
359 
360 	return (ptelo);
361 }
362 
363 static void
364 moea64_cpu_bootstrap_native(mmu_t mmup, int ap)
365 {
366 	int i = 0;
367 	#ifdef __powerpc64__
368 	struct slb *slb = PCPU_GET(aim.slb);
369 	register_t seg0;
370 	#endif
371 
372 	/*
373 	 * Initialize segment registers and MMU
374 	 */
375 
376 	mtmsr(mfmsr() & ~PSL_DR & ~PSL_IR);
377 
378 	/*
379 	 * Install kernel SLB entries
380 	 */
381 
382 	#ifdef __powerpc64__
383 		__asm __volatile ("slbia");
384 		__asm __volatile ("slbmfee %0,%1; slbie %0;" : "=r"(seg0) :
385 		    "r"(0));
386 
387 		for (i = 0; i < n_slbs; i++) {
388 			if (!(slb[i].slbe & SLBE_VALID))
389 				continue;
390 
391 			__asm __volatile ("slbmte %0, %1" ::
392 			    "r"(slb[i].slbv), "r"(slb[i].slbe));
393 		}
394 	#else
395 		for (i = 0; i < 16; i++)
396 			mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]);
397 	#endif
398 
399 	/*
400 	 * Install page table
401 	 */
402 
403 	__asm __volatile ("ptesync; mtsdr1 %0; isync"
404 	    :: "r"((uintptr_t)moea64_pteg_table
405 		     | (uintptr_t)(flsl(moea64_pteg_mask >> 11))));
406 	tlbia();
407 }
408 
409 static void
410 moea64_bootstrap_native(mmu_t mmup, vm_offset_t kernelstart,
411     vm_offset_t kernelend)
412 {
413 	vm_size_t	size;
414 	vm_offset_t	off;
415 	vm_paddr_t	pa;
416 	register_t	msr;
417 
418 	moea64_early_bootstrap(mmup, kernelstart, kernelend);
419 
420 	/*
421 	 * Allocate PTEG table.
422 	 */
423 
424 	size = moea64_pteg_count * sizeof(struct lpteg);
425 	CTR2(KTR_PMAP, "moea64_bootstrap: %d PTEGs, %d bytes",
426 	    moea64_pteg_count, size);
427 	rw_init(&moea64_eviction_lock, "pte eviction");
428 
429 	/*
430 	 * We now need to allocate memory. This memory, to be allocated,
431 	 * has to reside in a page table. The page table we are about to
432 	 * allocate. We don't have BAT. So drop to data real mode for a minute
433 	 * as a measure of last resort. We do this a couple times.
434 	 */
435 
436 	moea64_pteg_table = (struct lpte *)moea64_bootstrap_alloc(size, size);
437 	DISABLE_TRANS(msr);
438 	bzero(__DEVOLATILE(void *, moea64_pteg_table), moea64_pteg_count *
439 	    sizeof(struct lpteg));
440 	ENABLE_TRANS(msr);
441 
442 	CTR1(KTR_PMAP, "moea64_bootstrap: PTEG table at %p", moea64_pteg_table);
443 
444 	moea64_mid_bootstrap(mmup, kernelstart, kernelend);
445 
446 	/*
447 	 * Add a mapping for the page table itself if there is no direct map.
448 	 */
449 	if (!hw_direct_map) {
450 		size = moea64_pteg_count * sizeof(struct lpteg);
451 		off = (vm_offset_t)(moea64_pteg_table);
452 		DISABLE_TRANS(msr);
453 		for (pa = off; pa < off + size; pa += PAGE_SIZE)
454 			pmap_kenter(pa, pa);
455 		ENABLE_TRANS(msr);
456 	}
457 
458 	/* Bring up virtual memory */
459 	moea64_late_bootstrap(mmup, kernelstart, kernelend);
460 }
461 
462 static void
463 tlbia(void)
464 {
465 	vm_offset_t i;
466 	#ifndef __powerpc64__
467 	register_t msr, scratch;
468 	#endif
469 
470 	i = 0xc00; /* IS = 11 */
471 	switch (mfpvr() >> 16) {
472 	case IBM970:
473 	case IBM970FX:
474 	case IBM970MP:
475 	case IBM970GX:
476 	case IBMPOWER4:
477 	case IBMPOWER4PLUS:
478 	case IBMPOWER5:
479 	case IBMPOWER5PLUS:
480 		i = 0; /* IS not supported */
481 		break;
482 	}
483 
484 	TLBSYNC();
485 
486 	for (; i < 0x200000; i += 0x00001000) {
487 		#ifdef __powerpc64__
488 		__asm __volatile("tlbiel %0" :: "r"(i));
489 		#else
490 		__asm __volatile("\
491 		    mfmsr %0; \
492 		    mr %1, %0; \
493 		    insrdi %1,%3,1,0; \
494 		    mtmsrd %1; \
495 		    isync; \
496 		    \
497 		    tlbiel %2; \
498 		    \
499 		    mtmsrd %0; \
500 		    isync;"
501 		: "=r"(msr), "=r"(scratch) : "r"(i), "r"(1));
502 		#endif
503 	}
504 
505 	EIEIO();
506 	TLBSYNC();
507 }
508 
509 static int
510 atomic_pte_lock(volatile struct lpte *pte, uint64_t bitmask, uint64_t *oldhi)
511 {
512 	int	ret;
513 	uint32_t oldhihalf;
514 
515 	/*
516 	 * Note: in principle, if just the locked bit were set here, we
517 	 * could avoid needing the eviction lock. However, eviction occurs
518 	 * so rarely that it isn't worth bothering about in practice.
519 	 */
520 
521 	__asm __volatile (
522 		"1:\tlwarx %1, 0, %3\n\t"	/* load old value */
523 		"and. %0,%1,%4\n\t"		/* check if any bits set */
524 		"bne 2f\n\t"			/* exit if any set */
525 		"stwcx. %5, 0, %3\n\t"      	/* attempt to store */
526 		"bne- 1b\n\t"			/* spin if failed */
527 		"li %0, 1\n\t"			/* success - retval = 1 */
528 		"b 3f\n\t"			/* we've succeeded */
529 		"2:\n\t"
530 		"stwcx. %1, 0, %3\n\t"       	/* clear reservation (74xx) */
531 		"li %0, 0\n\t"			/* failure - retval = 0 */
532 		"3:\n\t"
533 		: "=&r" (ret), "=&r"(oldhihalf), "=m" (pte->pte_hi)
534 		: "r" ((volatile char *)&pte->pte_hi + 4),
535 		  "r" ((uint32_t)bitmask), "r" ((uint32_t)LPTE_LOCKED),
536 		  "m" (pte->pte_hi)
537 		: "cr0", "cr1", "cr2", "memory");
538 
539 	*oldhi = (pte->pte_hi & 0xffffffff00000000ULL) | oldhihalf;
540 
541 	return (ret);
542 }
543 
544 static uintptr_t
545 moea64_insert_to_pteg_native(struct lpte *pvo_pt, uintptr_t slotbase,
546     uint64_t mask)
547 {
548 	volatile struct lpte *pt;
549 	uint64_t oldptehi, va;
550 	uintptr_t k;
551 	int i, j;
552 
553 	/* Start at a random slot */
554 	i = mftb() % 8;
555 	for (j = 0; j < 8; j++) {
556 		k = slotbase + (i + j) % 8;
557 		pt = &moea64_pteg_table[k];
558 		/* Invalidate and seize lock only if no bits in mask set */
559 		if (atomic_pte_lock(pt, mask, &oldptehi)) /* Lock obtained */
560 			break;
561 	}
562 
563 	if (j == 8)
564 		return (-1);
565 
566 	if (oldptehi & LPTE_VALID) {
567 		KASSERT(!(oldptehi & LPTE_WIRED), ("Unmapped wired entry"));
568 		/*
569 		 * Need to invalidate old entry completely: see
570 		 * "Modifying a Page Table Entry". Need to reconstruct
571 		 * the virtual address for the outgoing entry to do that.
572 		 */
573 		if (oldptehi & LPTE_BIG)
574 			va = oldptehi >> moea64_large_page_shift;
575 		else
576 			va = oldptehi >> ADDR_PIDX_SHFT;
577 		if (oldptehi & LPTE_HID)
578 			va = (((k >> 3) ^ moea64_pteg_mask) ^ va) &
579 			    VSID_HASH_MASK;
580 		else
581 			va = ((k >> 3) ^ va) & VSID_HASH_MASK;
582 		va |= (oldptehi & LPTE_AVPN_MASK) <<
583 		    (ADDR_API_SHFT64 - ADDR_PIDX_SHFT);
584 		PTESYNC();
585 		TLBIE(va);
586 		moea64_pte_valid--;
587 		moea64_pte_overflow++;
588 	}
589 
590 	/*
591 	 * Update the PTE as per "Adding a Page Table Entry". Lock is released
592 	 * by setting the high doubleworld.
593 	 */
594 	pt->pte_lo = htobe64(pvo_pt->pte_lo);
595 	EIEIO();
596 	pt->pte_hi = htobe64(pvo_pt->pte_hi);
597 	PTESYNC();
598 
599 	/* Keep statistics */
600 	moea64_pte_valid++;
601 
602 	return (k);
603 }
604 
605 static int
606 moea64_pte_insert_native(mmu_t mmu, struct pvo_entry *pvo)
607 {
608 	struct lpte insertpt;
609 	uintptr_t slot;
610 
611 	/* Initialize PTE */
612 	moea64_pte_from_pvo(pvo, &insertpt);
613 
614 	/* Make sure further insertion is locked out during evictions */
615 	rw_rlock(&moea64_eviction_lock);
616 
617 	/*
618 	 * First try primary hash.
619 	 */
620 	pvo->pvo_pte.slot &= ~7ULL; /* Base slot address */
621 	slot = moea64_insert_to_pteg_native(&insertpt, pvo->pvo_pte.slot,
622 	    LPTE_VALID | LPTE_WIRED | LPTE_LOCKED);
623 	if (slot != -1) {
624 		rw_runlock(&moea64_eviction_lock);
625 		pvo->pvo_pte.slot = slot;
626 		return (0);
627 	}
628 
629 	/*
630 	 * Now try secondary hash.
631 	 */
632 	pvo->pvo_vaddr ^= PVO_HID;
633 	insertpt.pte_hi ^= LPTE_HID;
634 	pvo->pvo_pte.slot ^= (moea64_pteg_mask << 3);
635 	slot = moea64_insert_to_pteg_native(&insertpt, pvo->pvo_pte.slot,
636 	    LPTE_VALID | LPTE_WIRED | LPTE_LOCKED);
637 	if (slot != -1) {
638 		rw_runlock(&moea64_eviction_lock);
639 		pvo->pvo_pte.slot = slot;
640 		return (0);
641 	}
642 
643 	/*
644 	 * Out of luck. Find a PTE to sacrifice.
645 	 */
646 
647 	/* Lock out all insertions for a bit */
648 	if (!rw_try_upgrade(&moea64_eviction_lock)) {
649 		rw_runlock(&moea64_eviction_lock);
650 		rw_wlock(&moea64_eviction_lock);
651 	}
652 
653 	slot = moea64_insert_to_pteg_native(&insertpt, pvo->pvo_pte.slot,
654 	    LPTE_WIRED | LPTE_LOCKED);
655 	if (slot != -1) {
656 		rw_wunlock(&moea64_eviction_lock);
657 		pvo->pvo_pte.slot = slot;
658 		return (0);
659 	}
660 
661 	/* Try other hash table. Now we're getting desperate... */
662 	pvo->pvo_vaddr ^= PVO_HID;
663 	insertpt.pte_hi ^= LPTE_HID;
664 	pvo->pvo_pte.slot ^= (moea64_pteg_mask << 3);
665 	slot = moea64_insert_to_pteg_native(&insertpt, pvo->pvo_pte.slot,
666 	    LPTE_WIRED | LPTE_LOCKED);
667 	if (slot != -1) {
668 		rw_wunlock(&moea64_eviction_lock);
669 		pvo->pvo_pte.slot = slot;
670 		return (0);
671 	}
672 
673 	/* No freeable slots in either PTEG? We're hosed. */
674 	rw_wunlock(&moea64_eviction_lock);
675 	panic("moea64_pte_insert: overflow");
676 	return (-1);
677 }
678 
679